Modeling and

Simulation for
Automatic Control

Olav Egeland and Jan Tommy Gravdahl

Norwegian University of Science and Technology
Trondheim, Norway

MARINE CYBERNETICS http://www.marinecybernetics.com

Trondheim, Norway
 c 2002 by Marine Cybernetics AS.
Copyright °

All rights reserved.

For ordering see URL: http://www.marinecybernetics.com. The book can also be

ordered by sending an e-mail to:

[email protected]

or via fax:

Marine Cybernetics AS
P.O. Box 4607, NO-7451 Trondheim, Norway
fax: [+47] 72 81 00 18

No parts of this publication may be reproduced by any means, transmitted, or trans-

lated into machine language without the written permission of the author. Requests
for permisson to reproduce parts of the book should be adressed directly to Professor
Olav Egeland, Department of Engineering Cybernetics, Norwegian University of Science
and Technology, NO-7491 Trondheim, Norway; E-mail: [email protected], fax:
[+47] 73594399.

ISBN 82-92356-01-0

Corrected second printing June 2003

Produced from camera-ready copy supplied by the author using Scientific WorkPlace.

Printed and bound by Tapir Trykkeri, Trondheim, Norway.

Preface

Modeling and simulation of dynamic processes are very important subjects in control
systems design. Most processes that are encountered in practical controller design are
very well described in the engineering literature, and it is important that the control
engineer is able to take advantage of this information. It is a problem that several books
must be used to get the relevant modeling information of a particular process, and it may
take a long time to go through all the necessary material. The idea of this book is to
supply the control engineer with a sufficient modeling background to design controllers
for a wide range of processes. In addition, the book provides a good starting point for
going into the specialist literature of different engineering disciplines. In this connection
the references indicate where to start. The book also contains more material than what
will normally be covered in the lectures of a typical course, so that students may return
to the book at a later stage and find additional information about a particular subject.
This will be more efficient than to extract the required information from a series of other
books. In this sense the book will be of great value for practicing control engineers.
The development of new products and systems is often done in a team of experts
with different backgrounds. It is hoped that this book will help control engineers to
communicate with other experts in this type of team. To achieve this we have been careful
to use standard terminology and notation from the different engineering disciplines in
question. Here we deliberately break the tradition evident in many books in the control
literature where the emphasis is on having a unified formulation specific to automatic
control.
The selection of the material is based on the experience of the authors in teaching
and research at the Norwegian University of Science and Technology. In addition to this,
material has been selected on the basis of extensive industrial activity through research
programs between university and industry, and product development in industry. In this
activity there has been close cooperation with experts from other disciplines, and this
has given useful experience on how to approach different topics, and on how to interact
with other specialists.
The style of modeling used in this book is inspired from the field of robotics where
modeling is presented in a precise style based on equations. In addition, quite detailed
results and optimized algorithms are included in standard textbook in robotics. As a
result of this, the development in our book relies on many equations, but it is our expe-
rience that this is well appreciated by most students, as they do no have to waste time
on trying to understand long written descriptions on subjects that are easily understood
in terms of a series of equations. Moreover, we have experienced that the material pre-
sented in this book is suited both for newcomers and for students with prior courses in
the topics of the book. In particular we have seen that students with virtually no back-
ground in dynamics have been able to master rigid body dynamics after going through

i
ii

the dynamics chapters of this book. At the same time, students who have taken courses
in dynamics also find the material in this book to be useful.
Parts of this book have been taught as a one-semester course at the Norwegian Uni-
versity of Science and Technology. The students are in the third year of their study
in electrical engineering with specialization in automatic control, and have taken a ba-
sic course in automatic control theory. Standard undergraduate courses in engineering
mathematics give a sufficient background in mathematics.
The results presented in this book have been developed and accumulated over a
period of 15 years. The first author would like to thank all of his doctoral students over
this period for their contributions. Also our colleagues and friends abroad have been
important in this work. Thanks are due to Rolf Johansson, Henk Nijmeijer, Rogelio
Lozano and Atul Kelkar for discussions on this book. In the writing of the book and
in the selection of the material we have benefited from the availability of the lecture
notes by Steinar Sælid, Rolf Henriksen and Torleif Iversen that have been used in earlier
versions of the course.
We would like to thank doctoral candidate Erlend Kristiansen for his work on simula-
tions, figures and proofreading. We would also like to thank Thor I. Fossen who has been
writing a book in parallel, and we have enjoyed all the discussions on writing in general
and modeling in particular. Thanks are also due to our colleagues Kristin Y. Pettersen,
Tor Arne Johansen and Asgeir Sørensen. We would also like to thank our colleagues at
the Department of Engineering Cybernetics for contributing to the stimulating working
environment that allowed us to write this book. Also the support from the Norwegian
Research Council has been important, as this has made it possible to have a large group of
PhD students and Post Docs at the Department. In particular, the Strategic University
Program in Marine Control Systems has given us very good working conditions.

Olav Egeland
Jan Tommy Gravdahl
December 2002
Contents

I Modeling 1
1 Model representation 3
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 State space methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.1 State space models . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.2 Second order models of mechanical systems . . . . . . . . . . . . . 5
1.2.3 Linearization of state space models . . . . . . . . . . . . . . . . . . 5
1.2.4 Linearization of second order systems . . . . . . . . . . . . . . . . 7
1.2.5 Stability with zero input . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.6 Stability of linear systems . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.7 Stability analysis using a linearized model . . . . . . . . . . . . . . 9
1.3 Transfer function models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.2 The transfer function of a state-space model . . . . . . . . . . . . . 10
1.3.3 Rational transfer functions . . . . . . . . . . . . . . . . . . . . . . 11
1.3.4 Impulse response and step response . . . . . . . . . . . . . . . . . 12
1.3.5 Loop transfer function . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.6 Example: Actuator with dynamic compensation . . . . . . . . . . 15
1.3.7 Stability of transfer functions . . . . . . . . . . . . . . . . . . . . . 16
1.3.8 Stability of closed loop systems . . . . . . . . . . . . . . . . . . . . 17
1.3.9 Partial differential equations . . . . . . . . . . . . . . . . . . . . . 17
1.4 Network description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.4.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.4.3 Multiport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.4.4 Example: DC motor with flexible load . . . . . . . . . . . . . . . . 21
1.4.5 Example: Voltage controlled DC motor . . . . . . . . . . . . . . . 22
1.4.6 Example: Diesel engine with turbocharger . . . . . . . . . . . . . . 23
1.4.7 Assigning computational inputs and outputs . . . . . . . . . . . . 24
1.4.8 Bond graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.5 Linear network theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.5.1 Driving point impedance . . . . . . . . . . . . . . . . . . . . . . . . 26
1.5.2 Linear two-ports . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.5.3 Impedance of two-port with termination . . . . . . . . . . . . . . . 28
1.5.4 Example: Passive mechanical two-port . . . . . . . . . . . . . . . . 29
1.5.5 Mechanical analog of PD controller . . . . . . . . . . . . . . . . . . 31
1.6 Example: Transmission line model . . . . . . . . . . . . . . . . . . . . . . 33
1.6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

iii
iv CONTENTS

1.6.2 Introductory example . . . . . . . . . . . . . . . . . . . . . . . . . 33

1.6.3 Effort and flow model . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.6.4 Transfer functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.6.5 Transfer function for terminated transmission line . . . . . . . . . 36
1.6.6 Wave variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
1.6.7 Lossless transmission line . . . . . . . . . . . . . . . . . . . . . . . 38
1.6.8 Line termination . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

2 Model analysis tools 41

2.1 Frequency response methods . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.1.1 The frequency response of a system . . . . . . . . . . . . . . . . . 41
2.1.2 Second order oscillatory system . . . . . . . . . . . . . . . . . . . . 42
2.1.3 Performance of a closed loop system . . . . . . . . . . . . . . . . . 43
2.1.4 Stability margins . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.2 Elimination of fast dynamics . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.2.1 Example: The electrical time constant in a DC motor . . . . . . . 45
2.2.2 Nonlinear system . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.3 Energy-based methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.3.2 The energy function . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.3.3 Second-order systems . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.3.4 Example: Mass-spring-damper . . . . . . . . . . . . . . . . . . . . 48
2.3.5 Lyapunov methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.3.6 Contraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.3.7 Energy flow in a turbocharged diesel engine . . . . . . . . . . . . . 51
2.4 Passivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.4.2 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.4.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.4.4 Energy considerations . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.4.5 Positive real transfer functions . . . . . . . . . . . . . . . . . . . . 56
2.4.6 Positive real rational transfer functions . . . . . . . . . . . . . . . 56
2.4.7 Positive realness of irrational transfer functions . . . . . . . . . . . 58
2.4.8 Passivity and positive real transfer functions . . . . . . . . . . . . 59
2.4.9 No poles on the imaginary axis . . . . . . . . . . . . . . . . . . . . 60
2.4.10 Single poles at the imaginary axis . . . . . . . . . . . . . . . . . . 60
2.4.11 Bounded real transfer functions . . . . . . . . . . . . . . . . . . . . 61
2.4.12 Passivity of PID controllers . . . . . . . . . . . . . . . . . . . . . . 62
2.4.13 Closed loop stability of positive real systems . . . . . . . . . . . . 62
2.4.14 Storage function formulation . . . . . . . . . . . . . . . . . . . . . 63
2.4.15 Interconnections of passive systems . . . . . . . . . . . . . . . . . . 64
2.4.16 Storage function for PID controller . . . . . . . . . . . . . . . . . . 65
2.4.17 Passive plant with PID controller . . . . . . . . . . . . . . . . . . . 65
2.4.18 Example: Control of mass-spring-damper system . . . . . . . . . . 66
2.4.19 Example: Active vibration damping . . . . . . . . . . . . . . . . . 66
2.4.20 Passive electrical one-port . . . . . . . . . . . . . . . . . . . . . . . 67
2.4.21 Electrical analog of PID controller . . . . . . . . . . . . . . . . . . 68
2.4.22 Passive electrical two-port . . . . . . . . . . . . . . . . . . . . . . . 69
2.4.23 Termination of electrical two-port . . . . . . . . . . . . . . . . . . 69
2.4.24 Passive electrical n-ports . . . . . . . . . . . . . . . . . . . . . . . . 70
CONTENTS v

2.4.25 Example: Telemanipulation . . . . . . . . . . . . . . . . . . . . . . 70

2.4.26 Passivity and gain . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
2.5 Uncertainty in modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
2.5.1 General state space models . . . . . . . . . . . . . . . . . . . . . . 74
2.5.2 Exact kinematic models . . . . . . . . . . . . . . . . . . . . . . . . 75
2.5.3 Balance equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.5.4 Passivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

II Motors and actuators 77

3 Electromechanical systems 79
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.2 Electrical motors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.2.2 Basic equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.2.3 Gear model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.2.4 Motor and gear . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.2.5 Transformation of rotation to translation . . . . . . . . . . . . . . 82
3.2.6 Torque characteristics . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.2.7 The four quadrants of the motor . . . . . . . . . . . . . . . . . . . 84
3.3 The DC motor with constant field . . . . . . . . . . . . . . . . . . . . . . 85
3.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.3.2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.3.3 Energy function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.3.4 Laplace transformed model . . . . . . . . . . . . . . . . . . . . . . 87
3.4 DC motor control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.4.2 Current controlled DC motor . . . . . . . . . . . . . . . . . . . . . 89
3.4.3 Velocity controlled DC motor . . . . . . . . . . . . . . . . . . . . . 90
3.4.4 Position controlled DC motor . . . . . . . . . . . . . . . . . . . . . 90
3.5 Motor and load with elastic transmission . . . . . . . . . . . . . . . . . . . 91
3.5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.5.2 Equations of motion . . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.5.3 Transfer functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.5.4 Zeros of the transfer function . . . . . . . . . . . . . . . . . . . . . 95
3.5.5 Energy analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.5.6 Motor with several resonances in the load . . . . . . . . . . . . . . 95
3.5.7 Two motors driving an elastic load . . . . . . . . . . . . . . . . . . 96
3.5.8 Energy analysis of two motors and load . . . . . . . . . . . . . . . 97
3.6 Motor and load with deadzone in the gear . . . . . . . . . . . . . . . . . . 97
3.6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.6.2 Elastic gear with deadzone . . . . . . . . . . . . . . . . . . . . . . 98
3.6.3 Rigid gear with deadzone . . . . . . . . . . . . . . . . . . . . . . . 98
3.6.4 Two motors with deadzone and load . . . . . . . . . . . . . . . . . 99
3.7 Electromechanical energy conversion . . . . . . . . . . . . . . . . . . . . . 100
3.7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
3.7.2 Inductive circuit elements . . . . . . . . . . . . . . . . . . . . . . . 101
3.7.3 Capacitive circuit elements . . . . . . . . . . . . . . . . . . . . . . 102
3.7.4 Magnetic energy of a linear inductive element . . . . . . . . . . . . 103
vi CONTENTS

3.7.5 Stored energy of a linear capacitive element . . . . . . . . . . . . . 103

3.7.6 Energy and coenergy . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.7.7 Electromechanical two-port with inductive element . . . . . . . . . 105
3.7.8 Electromechanical two-port with linear flux linkage . . . . . . . . . 107
3.7.9 Magnetic levitation . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.7.10 Voice coil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
3.7.11 Electromagnetic three-port . . . . . . . . . . . . . . . . . . . . . . 110
3.7.12 Electromechanical capacitive element . . . . . . . . . . . . . . . . . 111
3.7.13 Electromechanical two-port with linear charge . . . . . . . . . . . . 112
3.7.14 Example: Capacitive microphone . . . . . . . . . . . . . . . . . . . 113
3.7.15 Piezoelectric actuator . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.7.16 Actuator configuration . . . . . . . . . . . . . . . . . . . . . . . . . 115
3.8 DC motor with externally controlled field . . . . . . . . . . . . . . . . . . 116
3.8.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
3.8.2 Network description . . . . . . . . . . . . . . . . . . . . . . . . . . 118
3.8.3 DC motor with field weakening . . . . . . . . . . . . . . . . . . . . 119
3.9 Dynamic model of the general AC motor . . . . . . . . . . . . . . . . . . . 120
3.9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
3.9.2 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
3.9.3 Dynamic model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
3.10 Induction motors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
3.10.1 Basic dynamic model . . . . . . . . . . . . . . . . . . . . . . . . . . 126
3.10.2 Induction motor model in stator frame . . . . . . . . . . . . . . . . 126
3.10.3 Dynamic model in the flux frame . . . . . . . . . . . . . . . . . . . 128
3.11 Lagrangian description of electromechanical systems . . . . . . . . . . . . 131
3.11.1 Generalized coordinates . . . . . . . . . . . . . . . . . . . . . . . . 131
3.11.2 Energy and coenergy . . . . . . . . . . . . . . . . . . . . . . . . . . 131
3.11.3 Analogy of electrical and mechanical systems . . . . . . . . . . . . 132
3.11.4 The Lagrangian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
3.11.5 Electromechanical systems . . . . . . . . . . . . . . . . . . . . . . . 134
3.11.6 Lagrange formulation of general AC motor . . . . . . . . . . . . . 136
3.11.7 Lagrange formulation of induction motor . . . . . . . . . . . . . . 138
3.11.8 Lagrange formulation of DC motor . . . . . . . . . . . . . . . . . . 138

4 Hydraulic motors 141

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
4.2 Valves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
4.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
4.2.2 Flow through a restriction . . . . . . . . . . . . . . . . . . . . . . . 141
4.2.3 Regularization of turbulent orifice flow . . . . . . . . . . . . . . . . 142
4.2.4 Four-way valve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
4.2.5 Matched and symmetrical four-way valve . . . . . . . . . . . . . . 145
4.2.6 Symmetric motor and valve with critical spool . . . . . . . . . . . 145
4.2.7 Symmetric motor and valve with open spool . . . . . . . . . . . . . 148
4.2.8 Flow control using pressure compensated valves . . . . . . . . . . . 148
4.2.9 Balance valve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
4.3 Motor models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
4.3.1 Mass balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
4.3.2 Rotational motors . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
4.3.3 Elastic modes in the load . . . . . . . . . . . . . . . . . . . . . . . 154
CONTENTS vii

4.3.4 Hydraulic cylinder . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

4.4 Models for transfer function analysis . . . . . . . . . . . . . . . . . . . . . 156
4.4.1 Matched and symmetric valve and symmetric motor . . . . . . . . 156
4.4.2 Valve controlled motor: Transfer function . . . . . . . . . . . . . . 157
4.4.3 Hydraulic motor with P controller . . . . . . . . . . . . . . . . . . 159
4.4.4 Symmetric cylinder with matched and symmetric valve . . . . . . 161
4.4.5 Pump controlled hydraulic drive with P controller . . . . . . . . . 162
4.4.6 Transfer functions for elastic modes . . . . . . . . . . . . . . . . . 162
4.4.7 Mechanical analog . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
4.5 Hydraulic transmission lines . . . . . . . . . . . . . . . . . . . . . . . . . . 165
4.5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
4.5.2 PDE Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
4.5.3 Laplace transformed model . . . . . . . . . . . . . . . . . . . . . . 167
4.5.4 Lossless model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
4.5.5 Linear friction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
4.5.6 Nonlinear friction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
4.5.7 Wave variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
4.5.8 Example: Lossless pipe . . . . . . . . . . . . . . . . . . . . . . . . 171
4.5.9 Linear network models of transmission lines . . . . . . . . . . . . . 172
4.5.10 Rational approximations of transfer function models . . . . . . . . 172
4.5.11 Rational series expansion of impedance model . . . . . . . . . . . . 173
4.5.12 Rational series expansion of admittance model . . . . . . . . . . . 174
4.5.13 Galerkin derivation of impedance model . . . . . . . . . . . . . . . 175
4.5.14 Galerkin derivation of the admittance model . . . . . . . . . . . . 176
4.5.15 Galerkin derivation of the hybrid model . . . . . . . . . . . . . . . 177
4.5.16 Rational simulation models . . . . . . . . . . . . . . . . . . . . . . 178
4.6 Lumped parameter model of hydraulic line . . . . . . . . . . . . . . . . . 180
4.6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
4.6.2 Helmholtz resonator model . . . . . . . . . . . . . . . . . . . . . . 181
4.6.3 Model formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
4.6.4 Admittance model . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
4.6.5 Impedance model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
4.6.6 Hybrid model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
4.6.7 Natural frequencies . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
4.7 Object oriented simulation models . . . . . . . . . . . . . . . . . . . . . . 185
4.7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
4.7.2 Pump controlled hydraulic motor . . . . . . . . . . . . . . . . . . . 185
4.7.3 Cylinder with balance valve . . . . . . . . . . . . . . . . . . . . . . 188

5 Friction 191
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
5.1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
5.2 Static friction models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
5.2.1 Models for the individual phenomena . . . . . . . . . . . . . . . . 192
5.2.2 Combination of individual models . . . . . . . . . . . . . . . . . . 195
5.2.3 Problems with the static models . . . . . . . . . . . . . . . . . . . 196
5.2.4 Problems with signum terms at zero velocity . . . . . . . . . . . . 197
5.2.5 Karnopp’s model of Coulomb friction . . . . . . . . . . . . . . . . 198
5.2.6 More on Karnopp’s friction model . . . . . . . . . . . . . . . . . . 198
5.2.7 Passivity of static models . . . . . . . . . . . . . . . . . . . . . . . 199
viii CONTENTS

5.3 Dynamic friction models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200

5.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
5.3.2 The Dahl model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
5.3.3 Passivity of the Dahl model . . . . . . . . . . . . . . . . . . . . . . 202
5.3.4 The Bristle and LuGre model . . . . . . . . . . . . . . . . . . . . . 202
5.3.5 Passivity of the LuGre model . . . . . . . . . . . . . . . . . . . . . 204
5.3.6 The Elasto-Plastic model . . . . . . . . . . . . . . . . . . . . . . . 205
5.3.7 Passivity of the Elasto-Plastic model . . . . . . . . . . . . . . . . . 206

III Dynamics 207

6 Rigid body kinematics 209
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
6.2 Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
6.2.1 Vector description . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
6.2.2 The scalar product . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
6.2.3 The vector cross product . . . . . . . . . . . . . . . . . . . . . . . 211
6.3 Dyadics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
6.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
6.3.2 Introductory example: The inertia dyadic . . . . . . . . . . . . . . 214
6.3.3 Matrix representation of dyadics . . . . . . . . . . . . . . . . . . . 215
6.4 The rotation matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
6.4.1 Coordinate transformations for vectors . . . . . . . . . . . . . . . . 218
6.4.2 Properties of the rotation matrix . . . . . . . . . . . . . . . . . . . 219
6.4.3 Composite rotations . . . . . . . . . . . . . . . . . . . . . . . . . . 220
6.4.4 Simple rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
6.4.5 Coordinate transformations for dyadics . . . . . . . . . . . . . . . 222
6.4.6 Homogeneous transformation matrices . . . . . . . . . . . . . . . . 223
6.5 Euler angles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
6.5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
6.5.2 Roll-pitch-yaw . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
6.5.3 Classical Euler angles . . . . . . . . . . . . . . . . . . . . . . . . . 226
6.6 Angle-axis description of rotation . . . . . . . . . . . . . . . . . . . . . . . 226
6.6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
6.6.2 Angle-axis parameters . . . . . . . . . . . . . . . . . . . . . . . . . 227
6.6.3 Derivation of rotation dyadic . . . . . . . . . . . . . . . . . . . . . 227
6.6.4 The rotation dyadic . . . . . . . . . . . . . . . . . . . . . . . . . . 228
6.6.5 Rotation matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
6.7 Euler parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
6.7.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
6.7.2 Quaternions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
6.7.3 Unit quaternions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
6.7.4 The quaternion product for unit quaternions . . . . . . . . . . . . 234
6.7.5 Rotation by the quaternion product . . . . . . . . . . . . . . . . . 235
6.7.6 Euler parameters from the rotation matrix . . . . . . . . . . . . . 236
6.7.7 The Euler rotation vector . . . . . . . . . . . . . . . . . . . . . . . 237
6.7.8 Euler-Rodrigues parameters . . . . . . . . . . . . . . . . . . . . . . 238
6.8 Angular velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
6.8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
CONTENTS ix

6.8.2 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240

6.8.3 Simple rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
6.8.4 Composite rotations . . . . . . . . . . . . . . . . . . . . . . . . . . 241
6.8.5 Differentiation of coordinate vectors . . . . . . . . . . . . . . . . . 242
6.8.6 Differentiation of vectors . . . . . . . . . . . . . . . . . . . . . . . . 242
6.9 Kinematic differential equations . . . . . . . . . . . . . . . . . . . . . . . . 244
6.9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
6.9.2 Attitude deviation . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
6.9.3 Homogeneous transformation matrices . . . . . . . . . . . . . . . . 245
6.9.4 Euler angles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
6.9.5 Euler parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
6.9.6 Normalization for numerical integration . . . . . . . . . . . . . . . 249
6.9.7 Euler rotation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
6.9.8 Euler-Rodrigues parameters . . . . . . . . . . . . . . . . . . . . . . 250
6.9.9 Passivity of kinematic differential equations . . . . . . . . . . . . . 251
6.9.10 Angle-axis representation . . . . . . . . . . . . . . . . . . . . . . . 252
6.10 The Serret-Frenet frame . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
6.10.1 Kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
6.10.2 Control deviation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
6.11 Navigational kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
6.11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
6.11.2 Coordinate frames . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
6.11.3 Acceleration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
6.12 Kinematics of a rigid body . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
6.12.1 Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
6.12.2 Velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
6.12.3 Acceleration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
6.13 The center of mass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
6.13.1 System of particles . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
6.13.2 Rigid body . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261

7 Newton-Euler equations of motion 263

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
7.2 Forces and torques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
7.2.1 Resultant force . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
7.2.2 Torque . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
7.2.3 Equivalent force and torque . . . . . . . . . . . . . . . . . . . . . . 265
7.2.4 Forces and torques on a rigid body . . . . . . . . . . . . . . . . . . 266
7.2.5 Example: Robotic link . . . . . . . . . . . . . . . . . . . . . . . . . 268
7.3 Newton-Euler equations for rigid bodies . . . . . . . . . . . . . . . . . . . 268
7.3.1 Equations of motion for a system of particles . . . . . . . . . . . . 268
7.3.2 Equations of motion for a rigid body . . . . . . . . . . . . . . . . . 269
7.3.3 Equations of motion about a point . . . . . . . . . . . . . . . . . . 271
7.3.4 The inertia dyadic . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
7.3.5 The inertia matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
7.3.6 Expressions for the inertia matrix . . . . . . . . . . . . . . . . . . 275
7.3.7 The parallel axes theorem . . . . . . . . . . . . . . . . . . . . . . . 275
7.3.8 The equations of motion for a rigid body . . . . . . . . . . . . . . 276
7.3.9 Satellite attitude dynamics . . . . . . . . . . . . . . . . . . . . . . 277
7.4 Example: Ball and beam dynamics . . . . . . . . . . . . . . . . . . . . . . 278
x CONTENTS

7.5 Example: Inverted pendulum . . . . . . . . . . . . . . . . . . . . . . . . . 281

7.5.1 Equations of motion . . . . . . . . . . . . . . . . . . . . . . . . . . 281
7.5.2 Double inverted pendulum . . . . . . . . . . . . . . . . . . . . . . . 284
7.6 Example: The Furuta pendulum . . . . . . . . . . . . . . . . . . . . . . . 285
7.7 Principle of virtual work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
7.7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
7.7.2 Generalized coordinates . . . . . . . . . . . . . . . . . . . . . . . . 289
7.7.3 Virtual displacements . . . . . . . . . . . . . . . . . . . . . . . . . 290
7.7.4 d’Alembert’s principle . . . . . . . . . . . . . . . . . . . . . . . . . 290
7.8 Principle of virtual work for a rigid body . . . . . . . . . . . . . . . . . . . 293
7.8.1 Virtual displacements for a rigid body . . . . . . . . . . . . . . . . 293
7.8.2 Force and torque of constraint . . . . . . . . . . . . . . . . . . . . 294
7.9 Multi-body dynamics and virtual work . . . . . . . . . . . . . . . . . . . . 295
7.9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
7.9.2 Equations of motion . . . . . . . . . . . . . . . . . . . . . . . . . . 295
7.9.3 Equations of motion about a point . . . . . . . . . . . . . . . . . . 297
7.9.4 Ball and beam . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
7.9.5 Single and double inverted pendulum . . . . . . . . . . . . . . . . . 300
7.9.6 Furuta pendulum . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
7.9.7 Planar two-link manipulator: Derivation 1 . . . . . . . . . . . . . . 302
7.9.8 Planar two-link manipulator: Derivation 2 . . . . . . . . . . . . . . 304
7.9.9 Kane’s computational scheme for two-link manipulator . . . . . . . 306
7.9.10 Manipulator dynamics in coordinate form . . . . . . . . . . . . . . 308
7.9.11 Spacecraft and manipulator . . . . . . . . . . . . . . . . . . . . . . 309
7.10 Recursive Newton-Euler . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
7.10.1 Inverse dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
7.10.2 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311

8 Analytical mechanics 313

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
8.2 Lagrangian dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
8.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
8.2.2 Lagrange’s equation of motion . . . . . . . . . . . . . . . . . . . . 314
8.2.3 Generalized coordinates and generalized forces . . . . . . . . . . . 316
8.2.4 Pendulum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
8.2.5 Mass-spring system . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
8.2.6 Ball and beam . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
8.2.7 Furuta pendulum . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
8.2.8 Manipulator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
8.2.9 Passivity of the manipulator dynamics . . . . . . . . . . . . . . . . 321
8.2.10 Example: Planar two-link manipulator 1 . . . . . . . . . . . . . . . 322
8.2.11 Example: Planar two-link manipulator 2 . . . . . . . . . . . . . . . 323
8.2.12 Limitations of Lagrange’s equation of motion . . . . . . . . . . . . 324
8.3 Calculus of variations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
8.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
8.3.2 Variations versus differentials . . . . . . . . . . . . . . . . . . . . . 325
8.3.3 The variation of a function . . . . . . . . . . . . . . . . . . . . . . 325
8.3.4 The Euler-Lagrange equation for a general integral . . . . . . . . . 327
8.3.5 The variation of the rotation matrix . . . . . . . . . . . . . . . . . 328
8.3.6 The variation of the homogeneous transformation matrix . . . . . 330
CONTENTS xi

8.4 The adjoint formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331

8.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
8.4.2 Rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
8.4.3 Rigid motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
8.5 The Euler-Poincaré equation . . . . . . . . . . . . . . . . . . . . . . . . . 334
8.5.1 A central equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
8.5.2 Rotating rigid body . . . . . . . . . . . . . . . . . . . . . . . . . . 336
8.5.3 Free-floating rigid body . . . . . . . . . . . . . . . . . . . . . . . . 337
8.5.4 Mechanism with n degrees of freedom . . . . . . . . . . . . . . . . 339
8.6 Hamilton’s principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
8.6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
8.6.2 The extended Hamilton principle . . . . . . . . . . . . . . . . . . . 340
8.6.3 Derivation of Lagrange’s equation of motion . . . . . . . . . . . . . 341
8.6.4 Hamilton’s principle . . . . . . . . . . . . . . . . . . . . . . . . . . 342
8.6.5 Rotations with the Euler-Poincaré equation . . . . . . . . . . . . . 343
8.6.6 Rigid motion with the Euler-Poincaré equation . . . . . . . . . . . 343
8.7 Lagrangian dynamics for PDE’s . . . . . . . . . . . . . . . . . . . . . . . . 344
8.7.1 Flexible beam dynamics . . . . . . . . . . . . . . . . . . . . . . . . 344
8.7.2 Euler-Bernoulli beam . . . . . . . . . . . . . . . . . . . . . . . . . 346
8.7.3 Lateral vibrations in a string . . . . . . . . . . . . . . . . . . . . . 346
8.8 Hamilton’s equations of motion . . . . . . . . . . . . . . . . . . . . . . . . 347
8.8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
8.8.2 Hamilton’s equation of motion . . . . . . . . . . . . . . . . . . . . 347
8.8.3 The energy function . . . . . . . . . . . . . . . . . . . . . . . . . . 349
8.8.4 Change of generalized coordinates . . . . . . . . . . . . . . . . . . 350
8.9 Control aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
8.9.1 Passivity of Hamilton’s equation of motion . . . . . . . . . . . . . 351
8.9.2 Example: Manipulator dynamics . . . . . . . . . . . . . . . . . . . 352
8.9.3 Example: The restricted three-body problem . . . . . . . . . . . . 353
8.9.4 Example: Attitude dynamics for a satellite . . . . . . . . . . . . . 355
8.9.5 Example: Gravity gradient stabilization . . . . . . . . . . . . . . . 356
8.10 The Hamilton-Jacobi equation . . . . . . . . . . . . . . . . . . . . . . . . 358

9 Mechanical vibrations 361

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
9.2 Lumped elastic two-ports . . . . . . . . . . . . . . . . . . . . . . . . . . . 362
9.2.1 Hybrid two-port . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362
9.2.2 Displacement two-port . . . . . . . . . . . . . . . . . . . . . . . . . 362
9.2.3 Three masses in the hybrid formulation . . . . . . . . . . . . . . . 363
9.2.4 Three masses in the displacement formulation . . . . . . . . . . . . 364
9.2.5 Four masses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
9.3 Vibrating string . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
9.3.1 Linearized model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
9.3.2 Orthogonal shape functions . . . . . . . . . . . . . . . . . . . . . . 366
9.3.3 Galerkin’s method for orthogonal shape functions . . . . . . . . . . 367
9.3.4 Finite element shape functions . . . . . . . . . . . . . . . . . . . . 368
9.3.5 String element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
9.3.6 Assembling string elements . . . . . . . . . . . . . . . . . . . . . . 370
9.4 Nonlinear string dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
9.4.1 Kirchhoff’s nonlinear string model . . . . . . . . . . . . . . . . . . 371
xii CONTENTS

9.4.2 Marine cables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371

9.5 Euler Bernoulli beam . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
9.5.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
9.5.2 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . 375
9.5.3 Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
9.5.4 Orthogonal shape functions . . . . . . . . . . . . . . . . . . . . . . 376
9.5.5 Clamped-free Euler Bernoulli beam . . . . . . . . . . . . . . . . . . 378
9.5.6 Beam fixed to an inertia and a mass . . . . . . . . . . . . . . . . . 380
9.5.7 Orthogonality of the eigenfunctions . . . . . . . . . . . . . . . . . . 381
9.5.8 Galerkin’s method for orthogonal mode shapes . . . . . . . . . . . 382
9.6 Finite element model of Euler Bernoulli beam . . . . . . . . . . . . . . . . 385
9.6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
9.6.2 Beam element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
9.6.3 Assembling a structure . . . . . . . . . . . . . . . . . . . . . . . . . 386
9.6.4 Finite element model and Galerkin’s method . . . . . . . . . . . . 388
9.7 Motor and Euler Bernoulli beam . . . . . . . . . . . . . . . . . . . . . . . 390
9.7.1 Equations of motion . . . . . . . . . . . . . . . . . . . . . . . . . . 390
9.7.2 Assumed mode shapes . . . . . . . . . . . . . . . . . . . . . . . . . 391
9.7.3 Finite elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
9.8 Irrational transfer functions for beam dynamics . . . . . . . . . . . . . . . 393
9.8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
9.8.2 Clamped-free beam . . . . . . . . . . . . . . . . . . . . . . . . . . . 394
9.8.3 Motor and beam . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395

IV Balance equations 399

10 Kinematics of Flow 401
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
10.2 Kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
10.2.1 The material derivative . . . . . . . . . . . . . . . . . . . . . . . . 401
10.2.2 The nabla operator . . . . . . . . . . . . . . . . . . . . . . . . . . . 402
10.2.3 Divergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403
10.2.4 Curl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404
10.2.5 Material coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . 406
10.2.6 The dilation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
10.3 Orthogonal curvilinear coordinates . . . . . . . . . . . . . . . . . . . . . . 408
10.3.1 General results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 408
10.3.2 Cylindrical coordinates . . . . . . . . . . . . . . . . . . . . . . . . 411
10.4 Reynolds’ transport theorem . . . . . . . . . . . . . . . . . . . . . . . . . 413
10.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
10.4.2 Basic transport theorem . . . . . . . . . . . . . . . . . . . . . . . . 413
10.4.3 The transport theorem for a material volume . . . . . . . . . . . . 414
10.4.4 The transport theorem and balance laws . . . . . . . . . . . . . . . 414

11 Mass, momentum and energy balances 417

11.1 The mass balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
11.1.1 Differential form . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
11.1.2 Integral form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 418
11.1.3 Control volume with compressible fluid . . . . . . . . . . . . . . . 418
CONTENTS xiii

11.1.4 Mass flow through a pipe . . . . . . . . . . . . . . . . . . . . . . . 419

11.1.5 Continuity equation and Reynolds’ transport theorem . . . . . . . 420
11.1.6 Multi-component systems . . . . . . . . . . . . . . . . . . . . . . . 422
11.2 The momentum balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
11.2.1 Euler’s equation of motion . . . . . . . . . . . . . . . . . . . . . . . 423
11.2.2 The momentum equation for a control volume . . . . . . . . . . . 425
11.2.3 Example: Waterjet . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
11.2.4 Example: Sand dispenser and conveyor . . . . . . . . . . . . . . . 426
11.2.5 Irrotational Bernoulli equation . . . . . . . . . . . . . . . . . . . . 427
11.2.6 Bernoulli’s equation along a streamline . . . . . . . . . . . . . . . . 428
11.2.7 Example: Transmission line . . . . . . . . . . . . . . . . . . . . . . 429
11.2.8 Liquid mass flow through a restriction . . . . . . . . . . . . . . . . 431
11.2.9 Example: Water turbine . . . . . . . . . . . . . . . . . . . . . . . . 432
11.2.10 Example: Waterhammer . . . . . . . . . . . . . . . . . . . . . . . . 436
11.3 Angular momentum balance . . . . . . . . . . . . . . . . . . . . . . . . . . 437
11.3.1 General expression . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
11.3.2 Centrifugal pump with radial blades . . . . . . . . . . . . . . . . . 437
11.3.3 Euler’s turbomachinery equation . . . . . . . . . . . . . . . . . . . 439
11.3.4 Pump instability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
11.4 The energy balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441
11.4.1 Material volume . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441
11.4.2 Fixed volume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
11.4.3 General control volume . . . . . . . . . . . . . . . . . . . . . . . . 445
11.4.4 The heat equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
11.4.5 Transfer function for the heat equation . . . . . . . . . . . . . . . . 447
11.5 Viscous flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448
11.5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448
11.5.2 Tensor notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448
11.5.3 The velocity gradient tensor . . . . . . . . . . . . . . . . . . . . . . 451
11.5.4 Example: The velocity gradient for a rigid body . . . . . . . . . . 452
11.5.5 The stress tensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
11.5.6 Cauchy’s equation of motion . . . . . . . . . . . . . . . . . . . . . 454
11.5.7 Newtonian fluids . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
11.5.8 The Navier-Stokes equation . . . . . . . . . . . . . . . . . . . . . . 458
11.5.9 The Reynolds number . . . . . . . . . . . . . . . . . . . . . . . . . 459
11.5.10 The equation of kinetic energy . . . . . . . . . . . . . . . . . . . . 460
11.5.11 The energy balance for a viscous fluid . . . . . . . . . . . . . . . . 462
11.5.12 Fixed volume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
11.5.13 General control volume . . . . . . . . . . . . . . . . . . . . . . . . 463

12 Gas dynamics 465

12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
12.2 Energy, enthalpy and entropy . . . . . . . . . . . . . . . . . . . . . . . . . 465
12.2.1 Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
12.2.2 Enthalpy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
12.2.3 Specific heats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466
12.2.4 Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
12.2.5 The entropy equation . . . . . . . . . . . . . . . . . . . . . . . . . 467
12.2.6 Internal energy equation in terms of temperature . . . . . . . . . . 470
12.2.7 Energy balance in terms of pressure . . . . . . . . . . . . . . . . . 471
xiv CONTENTS

12.2.8 Piston motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472

12.3 Isentropic conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
12.3.1 Isentropic processes . . . . . . . . . . . . . . . . . . . . . . . . . . 472
12.3.2 Stagnation state . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474
12.3.3 Energy balance for isentropic processes . . . . . . . . . . . . . . . . 474
12.3.4 The speed of sound . . . . . . . . . . . . . . . . . . . . . . . . . . . 475
12.3.5 Helmholtz resonator . . . . . . . . . . . . . . . . . . . . . . . . . . 476
12.4 Acoustic resonances in pipes . . . . . . . . . . . . . . . . . . . . . . . . . . 477
12.4.1 Dynamic model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 477
12.4.2 Pipe closed at both ends . . . . . . . . . . . . . . . . . . . . . . . . 478
12.4.3 Pipe closed at one end . . . . . . . . . . . . . . . . . . . . . . . . . 479
12.4.4 Pressure measurement in diesel engine cylinder . . . . . . . . . . . 480
12.5 Gas flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 480
12.5.1 Gas flow through a restriction . . . . . . . . . . . . . . . . . . . . . 480
12.5.2 Example: Discharge of gas from tank . . . . . . . . . . . . . . . . . 482
12.5.3 The Euler equation around sonic speed . . . . . . . . . . . . . . . 483

13 Compressor dynamics 485

13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
13.1.1 Compressors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
13.1.2 Surge and rotating stall . . . . . . . . . . . . . . . . . . . . . . . . 486
13.2 Centrifugal Compressors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 486
13.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 486
13.2.2 Shaft dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 487
13.2.3 Compressor system . . . . . . . . . . . . . . . . . . . . . . . . . . . 489
13.2.4 Mass balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489
13.2.5 Momentum equation . . . . . . . . . . . . . . . . . . . . . . . . . . 490
13.3 Compressor characteristic . . . . . . . . . . . . . . . . . . . . . . . . . . . 491
13.3.1 Derivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491
13.3.2 The compressor characteristic at zero mass flow . . . . . . . . . . . 494
13.4 Compressor surge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 497
13.4.1 The Greitzer surge model . . . . . . . . . . . . . . . . . . . . . . . 497
13.4.2 Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499
13.4.3 Passivity of the Greitzer surge model . . . . . . . . . . . . . . . . . 500
13.4.4 Curvefitting of compressor characteristic . . . . . . . . . . . . . . . 501
13.4.5 Compression systems with recycle . . . . . . . . . . . . . . . . . . 504

V Simulation 507
14 Simulation 509
14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509
14.1.1 The use of simulation in automatic control . . . . . . . . . . . . . 509
14.1.2 The Moore Greitzer model . . . . . . . . . . . . . . . . . . . . . . 510
14.1.3 The restricted three-body problem . . . . . . . . . . . . . . . . . . 513
14.1.4 Mass balance of chemical reactor . . . . . . . . . . . . . . . . . . . 516
14.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
14.2.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
14.2.2 Computation error . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
14.2.3 The order of a one-step method . . . . . . . . . . . . . . . . . . . . 517
CONTENTS xv

14.2.4 Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 518

14.2.5 The linear test function . . . . . . . . . . . . . . . . . . . . . . . . 520
14.3 Euler methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521
14.3.1 Euler’s method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521
14.3.2 The improved Euler method . . . . . . . . . . . . . . . . . . . . . . 523
14.3.3 The modified Euler method . . . . . . . . . . . . . . . . . . . . . . 525
14.4 Explicit Runge-Kutta methods . . . . . . . . . . . . . . . . . . . . . . . . 526
14.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
14.4.2 Numerical scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
14.4.3 Order conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
14.4.4 Some explicit Runge-Kutta methods . . . . . . . . . . . . . . . . . 528
14.4.5 Case study: Pneumatic spring . . . . . . . . . . . . . . . . . . . . 528
14.4.6 Stability function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 531
14.4.7 FSAL methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534
14.5 Implicit Runge-Kutta methods . . . . . . . . . . . . . . . . . . . . . . . . 534
14.5.1 Stiff systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534
14.5.2 Implicit Runge-Kutta methods . . . . . . . . . . . . . . . . . . . . 535
14.5.3 Implicit Euler method . . . . . . . . . . . . . . . . . . . . . . . . . 535
14.5.4 Trapezoidal rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536
14.5.5 Implicit midpoint rule . . . . . . . . . . . . . . . . . . . . . . . . . 537
14.5.6 The theta method . . . . . . . . . . . . . . . . . . . . . . . . . . . 538
14.5.7 Stability function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 538
14.5.8 Some implicit Runge-Kutta methods . . . . . . . . . . . . . . . . . 539
14.5.9 Case study: Pneumatic spring revisited . . . . . . . . . . . . . . . 540
14.6 Stability of Runge-Kutta methods . . . . . . . . . . . . . . . . . . . . . . 544
14.6.1 Aliasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 544
14.6.2 A-stability, L-stability . . . . . . . . . . . . . . . . . . . . . . . . . 544
14.6.3 Stiffly accurate methods . . . . . . . . . . . . . . . . . . . . . . . . 546
14.6.4 Padé approximations . . . . . . . . . . . . . . . . . . . . . . . . . . 548
14.6.5 Stability for Padé approximations . . . . . . . . . . . . . . . . . . 550
14.6.6 Example: Mechanical vibrations . . . . . . . . . . . . . . . . . . . 551
14.6.7 Frequency response . . . . . . . . . . . . . . . . . . . . . . . . . . . 551
14.6.8 AN-stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 555
14.6.9 B-stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
14.6.10 Algebraic stability . . . . . . . . . . . . . . . . . . . . . . . . . . . 558
14.6.11 Properties of Runge-Kutta methods . . . . . . . . . . . . . . . . . 560
14.7 Automatic adjustment of step size . . . . . . . . . . . . . . . . . . . . . . 560
14.7.1 Estimation of the local error for Runge-Kutta methods . . . . . . . 560
14.7.2 Adjustment algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 563
14.8 Implementation aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . 563
14.8.1 Solution of implicit equations . . . . . . . . . . . . . . . . . . . . . 563
14.8.2 Dense outputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565
14.8.3 Event detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 566
14.8.4 Systems with inertia matrix . . . . . . . . . . . . . . . . . . . . . . 566
14.9 Invariants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
14.9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
14.9.2 Linear invariants . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
14.9.3 Quadratic functions . . . . . . . . . . . . . . . . . . . . . . . . . . 568
14.9.4 Quadratic invariants . . . . . . . . . . . . . . . . . . . . . . . . . . 569
xvi CONTENTS

14.9.5 Symplectic Runge-Kutta methods . . . . . . . . . . . . . . . . . . 571

14.10Rosenbrock methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 574
14.11Multistep methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 576
14.11.1 Explicit Adams methods . . . . . . . . . . . . . . . . . . . . . . . . 576
14.11.2 Implicit Adams methods . . . . . . . . . . . . . . . . . . . . . . . . 578
14.11.3 Predictor-Corrector implementation . . . . . . . . . . . . . . . . . 579
14.11.4 Backwards differentiation methods . . . . . . . . . . . . . . . . . . 580
14.11.5 Linear stability analysis . . . . . . . . . . . . . . . . . . . . . . . . 581
14.11.6 Stability of Adams methods . . . . . . . . . . . . . . . . . . . . . . 582
14.11.7 Stability of BDF methods . . . . . . . . . . . . . . . . . . . . . . . 583
14.11.8 Frequency response . . . . . . . . . . . . . . . . . . . . . . . . . . . 583
14.11.9 Adams methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585
14.11.10BDF methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585
14.12Differential-algebraic equations . . . . . . . . . . . . . . . . . . . . . . . . 585
14.12.1 Implicit Runge-Kutta methods for index 1 problems . . . . . . . . 587
14.12.2 Multistep methods for index 1 problems . . . . . . . . . . . . . . . 589

15 Computational fluid dynamics 591

15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 591
15.2 Governing equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 591
15.3 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 592
15.3.1 Hyperbolic equations . . . . . . . . . . . . . . . . . . . . . . . . . . 595
15.3.2 Parabolic equations . . . . . . . . . . . . . . . . . . . . . . . . . . 596
15.3.3 Elliptic equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 597
15.4 Diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
15.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
15.4.2 Finite volume method for stationary diffusion . . . . . . . . . . . . 599
15.5 Solution of equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 603
15.5.1 Worked example on stationary diffusion . . . . . . . . . . . . . . . 603
15.6 Stability issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 604
15.7 Finite volume method for diffusion dynamics . . . . . . . . . . . . . . . . 605
15.8 Finite volumes for Convection-Diffusion . . . . . . . . . . . . . . . . . . . 610
15.8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 610
15.8.2 Finite volume method for 1D diffusion and convection dynamics . 611
15.9 Pressure-velocity coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . 614
15.9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 614
15.9.2 The staggered grid . . . . . . . . . . . . . . . . . . . . . . . . . . . 615
15.9.3 The momentum equations . . . . . . . . . . . . . . . . . . . . . . . 617
15.9.4 The transient SIMPLE algorithm . . . . . . . . . . . . . . . . . . . 618
15.10Von Neuman stability method . . . . . . . . . . . . . . . . . . . . . . . . . 620
 Part I

Modeling

1
Chapter 1

Model representation

1.1 Introduction
In this chapter we will present model formulations for use in controller analysis and
design. The usual type of models in control problems are based on ordinary differential
equations with time as the free variable. The two main representations of such models are
state-space descriptions, where the model is given as a system of first-order differential
equations, and transfer function models using the Laplace transformation. In this setting
the signal-flow representation of models is used where each model has a defined set of
input variables and a set of output variables. Control theory offers a wide variety of
tools and techniques for controller analysis and design based on state-space models and
transfer function models.
The signal-flow description has been very successful in control applications. However,
in energy-based control analysis and in the development of simulation systems, there is an
alternative formulation which is based on an energy-flow description. This formulation
is of great use in the development of large simulation systems as it opens up for object-
oriented modeling. This is an approach where a model is developed for each physical
subsystem, and where the model of the total system is obtained by interconnecting the
models of the subsystems using energy-flow variables.
Throughout the book models are developed from physics. This includes physical
principles like Newton’s laws and balance equations, which are typically based on the
conservation of mass, momentum, energy, and electrical charge. In addition, results
are derived using the purely mathematical field of kinematics, which is the geometric
description of motion. Finally, empirically established constitutive equations are needed
to describe material properties like the relation between the force and deformation of a
spring, the relation between velocity gradients and viscous tension of a fluid, and the
relation between charge and voltage of a capacitive element.
In contrast to this, models may be obtained as black-box model where transfer func-
tions or state space models between inputs and outputs are established from identification
experiments . This approach will not be discussed in this book.
This chapter starts with a presentation of state-space models and transfer function
models in the signal-flow description, which is the usual formulation in automatic con-
trol. Then the energy-flow description is presented. Material on second order me-
chanical systems and systems described by partial differential equations is also dis-
cussed. Background material on control is found in (Kuo 1995), (Chen 1999) and

3
4 CHAPTER 1. MODEL REPRESENTATION

(Dorf and Bishop 2000), while additional material on linear systems theory is covered by
(Rugh 1996) and (Antsaklis and Michel 1997).

1.2 State space methods

1.2.1 State space models
A state space model

ẋ1 = f1 (x1 , . . . , xn , u1 , . . . up , t) (1.1)

..
. (1.2)
ẋn = fn (x1 , . . . , xn , u1 , . . . up , t) (1.3)

which in vector form is written

ẋ = f (x, u, t) (1.4)
is a set of first order differential equations describing the dynamics of the state vector
x = (x1 , . . . , xn )T under the action of the control or input vector is u = (u1 , . . . , up )T .
The measurement or output vector y is is often included in the model formulation, and
the state-space model is written

ẋ = f (x, u, t) (1.5)
y = h (x, t) (1.6)

An important class of systems for controller design is linear time-invariant systems

which are written in the form
ẋ = Ax + Bu
(1.7)
y = Cx + Du
A block diagram is shown in Figure 1.1.

D
u x y
B C

Figure 1.1: Linear time-invariant state space model

Example 1 Systems that can be written in the form

ẋ = f (x) + G (x) u (1.8)

are said to be affine in the control u, which means that when x is given, then the right side
of (1.8) is a constant plus a term that is linear in u. This type of system is important in
nonlinear control theory where methods are available for this type of model (Isidori 1989),
(Nijmeijer and der Schaft 1990).
1.2. STATE SPACE METHODS 5

1.2.2 Second order models of mechanical systems

Mechanical systems are often described as second order systems in the form
M (q) q̈ + f (q, q̇) = u (1.9)
where q is the vector of generalized coordinates and u is the generalized input force.
The matrix M (q) may be called the mass matrix. Intuitively, this can be regarded as
a generalization of Newton’s law which states that mass times acceleration is equal to
force. This second order model may be written in state space form by defining x1 = q,
x2 = q̇ which gives
µ ¶ µ ¶
ẋ1 x2
= (1.10)
ẋ2 M−1 (x1 ) [−f (x1 , x2 ) + u]
Some mechanical systems have models of a special structure due to the physical
properties of the systems. In particular, this is true for vibration problems and for
robotics. The model of a robot manipulators is written (Spong and Vidyasagar 1989),
(Sciavicco and Siciliano 2000)
M (q) q̈ + C (q, q̇) q̇ + g (q) = τ (1.11)
where M (q) is the symmetric and positive definite mass matrix, C (q, q̇) is the Coriolis
matrix, g (q) is the generalized force of gravity, q = (q1 , . . . , q6 ) is the vector of generalized
coordinates, and τ = (τ 1 , . . . , τ 6 ) is the vector of generalized actuator forces. The model
is usually left in the second order formulation, as the usual control techniques used for
manipulators rely on this formulation.

1.2.3 Linearization of state space models

Many methods and control techniques are available for linear systems. In particular,
control methods based on frequency response require a linear model. Therefore, if the
modeling of a system results in a nonlinear system
ẋ = f (x, u, t)
(1.12)
y = h (x, u, t)
it may be useful to linearize the system. Linearization is done around a solution of the
system. A solution of the system is a function (x0 (t), u0 (t)) that satisfies the system
equation
ẋ0 = f [x0 (t), u0 (t), t] (1.13)
We define the perturbations ∆x, ∆u and ∆y from the solution by
x(t) = x0 (t) + ∆x(t) (1.14)
u(t) = u0 (t) + ∆u(t) (1.15)
y(t) = h [x0 (t), u0 (t), t] + ∆y(t) (1.16)
Standard Taylor series linearization around the solution (x0 (t), u0 (t)) gives
¯ ¯
∂f ¯¯ ∂f ¯¯
ẋ = f [x0 (t), u0 (t), t] + ∆x + ∆u (1.17)
∂x ¯x0 (t),u0 (t) ∂u ¯x0 (t),u0 (t)
¯ ¯
∂h ¯¯ ∂h ¯¯
y = h [x0 (t), u0 (t), t] + ∆x + ∆u (1.18)
∂x ¯ x0 (t),u0 (t)∂u ¯ x0 (t),u0 (t)
6 CHAPTER 1. MODEL REPRESENTATION

where the matrices appearing from the differentiations have elements given by
½ ¾ ½ ¾ ½ ¾ ½ ¾
∂f ∂fi ∂f ∂fi ∂h ∂hi ∂h ∂hi
= , = , = , = (1.19)
∂x ∂xj ∂u ∂uj ∂x ∂xj ∂u ∂uj

Insertion of (1.13) into (1.17), and insertion of (1.16) into (1.18) gives the following
linearized system:

¯ ¯
∂f ¯¯ ∂f ¯¯
∆ẋ = ∆x + ∆u (1.20)
∂x ¯x0 (t),u0 (t) ∂u ¯x0 (t),u0 (t)
¯ ¯
∂h ¯¯ ∂h ¯¯
∆y = ∆x + ∆u (1.21)
∂x ¯x0 (t),u0 (t) ∂u ¯ x0 (t),u0 (t)

It is noted that this model is of the same form as (1.7). The matrices A, B, C and
D are seen to be given by
¯ ¯
∂f ¯¯ ∂f ¯¯
A(t) = , B(t) = (1.22)
∂x ¯x0 (t),u0 (t) ∂u ¯x0 (t),u0 (t)
¯ ¯
∂h ¯¯ ∂h ¯¯
C(t) = , D(t) = (1.23)
∂x ¯x0 (t),u0 (t) ∂u ¯x0 (t),u0 (t)

Example 2 A simplified model for the design of a cruise control system for a car is
obtained from Newton’s law. Suppose that the forces acting on the car are the air resis-
tance, which is proportional to the square of the velocity, and the motor force, which is
assumed to be proportional to the throttle input. Then the model is
1
mv̇ = − CD ρAv 2 + Kt u (1.24)
2
where v is the velocity and m is the mass of the car, CD is the drag coefficient, ρ is
the density of air, A is the projected area of the car when seen from the front, Kt is the
throttle constant, and u is the throttle input. The control input u0 corresponding to a
constant speed v0 is found by inserting mv̇0 = 0 in the model, which gives
1
u0 = CD ρAv02 (1.25)
2Kt
We define the perturbations ∆v = v − v0 and ∆u = u − u0 and find the linearized model

m∆v̇ = −CD ρAv0 ∆v + Kt ∆u (1.26)

Example 3 A standard laboratory demonstration of feedback control is the magnetic

levitation experiment where an electromagnet is used to control the vertical position of a
steel ball. The equation of motion for the ball is derived in Section 3.7.9 to be

i2
mz̈ = −C + mg (1.27)
z2
where m is the mass, z is the vertical position of the ball in the downwards direction, C is
a constant, i is the control input, which is the current of the electromagnet, and g is the
1.2. STATE SPACE METHODS 7

acceleration of gravity. Let zd be the constant desired position of the ball. The solution
(zd , id ) is found by inserting mz̈d = 0 in the model, which gives the constant current
r
mg
id = zd (1.28)
C

which will give a lifting force that can hold the ball stationary at position zd . We define
the perturbations ∆z = z − zd and ∆i = i − id and get the linearized model

i2d id
m∆z̈ = 2C ∆z − 2C 2 ∆i (1.29)
zd3 zd

1.2.4 Linearization of second order systems

A second order system
q̈ = f (q, q̇, u) (1.30)
T
may be linearized by reformulating it as a state-space model (1.4) with x = (qT , q̇T ) .
However, we may also linearize the system in the second order formulation, which may
be advantageous for some systems. Then the system is linearized around a solution
(q0 (t), q̇0 (t), u0 (t)) which satisfies

q̈0 = f (q0 , q̇0 , u0 ) . (1.31)

Taylor series expansion of the model around the solution gives

∂f ∂f ∂f
q̈0 + ∆q̈ = f (q0 , q̇0 , u0 ) + ∆q + ∆q̇ + ∆u (1.32)
∂q ∂ q̇ ∂u

and combination with (1.31) gives the linearized model

∂f ∂f ∂f
∆q̈ = ∆q + ∆q̇ + ∆u (1.33)
∂q ∂ q̇ ∂u

Example 4 Consider a pendulum with a point mass m on a massless beam of length

L as shown in Figure 1.2. The angle of rotation θ is set to zero when the pendulum is
hanging downwards. The equation of motion for the pendulum is

mL2 θ̈ + mLg sin θ = 0 (1.34)

which can be written

g
θ̈ = − sin θ (1.35)
L
Linearization around the solution (θ, θ̇) = (0, 0) gives the linear model
g
θ̈ = − θ (1.36)
L
8 CHAPTER 1. MODEL REPRESENTATION

L
g


Figure 1.2: Pendulum

1.2.5 Stability with zero input

The concept of stability is of fundamental importance in control theory, and it is highly
relevant in connection with modelling as we can highlight the stability properties of a
system by selecting an appropriate model formulation. We will therefore present results
on the stability of a state-space model that are important in connection with modeling.
We will focus on the stability of a system with zero input around an equilibrium state xe .
A state xe is an equilibrium state if the system is at rest in this equilibrium state. With
this we mean that if the system state starts in x = xe , then the state vector will remain
in xe . The equilibrium is said to be stable if it has the property that the state will stay
close to the equilibrium whenever the state starts near the equilibrium. If an equilibrium
of a system is not stable, then it is said to be unstable. If an equilibrium is stable and
the state converges to the equilibrium, then the equilibrium is said to be asymptotically
stable.

Example 5 Consider the state space model

µ ¶ µ ¶
d θ θ̇
= (1.37)
dt θ̇ − Lg sin θ

of a pendulum. The states of the pendulum are selected to be θ and θ̇ so that the state
vector is µ ¶
θ
x= (1.38)
θ̇

We see that ẋ = 0 whenever θ̇ = 0 and sin θ = 0, which is the case for θ = 0 and θ = π.
This means that the system has equilibrium states at
µ ¶ µ ¶
0 π
xe1 = and xe2 = (1.39)
0 0

Here x = xe1 is the equilibrium where the pendulum is hanging downwards, and x = xe2
is the equilibrium where the pendulum is raised upwards with the mass on the top. We
know from experience that if the pendulum starts from a state close to the downwards
configuration xe1 , and with a speed close to zero, then the pendulum will stay close to
the downwards configuration. Therefore, the system is stable around the equilibrium
1.2. STATE SPACE METHODS 9

xe1 . Our experience with the equilibrium xe2 is that for any small deviation from the
equilibrium state the pendulum will fall down and move far away from the equilibrium
state. Therefore, the system is unstable around the equilibrium xe2 .

1.2.6 Stability of linear systems

Consider the linear time-invariant system

ẋ = Ax + Bu (1.40)
y = Cx + Du (1.41)

where u is input, y is output, and x ∈ Rn is the state vector. The solution of the state
equation are known from basic textbooks in automatic control to be
Z t
A(t−t0 )
x (t) = e x (t0 ) + eA(t−τ ) Bu (τ ) dτ (1.42)
t0

The eigenvalues of the n × n matrix A are denoted λi , i = 1, . . . , n. It is assumed that A

has m ≤ n simple eigenvalues λi , i = 1, . . . , m, and that the remaining n − m eigenvalues
λm+1 = . . . = λn are coincident. Then, if the input is zero, that is u = 0, the solution of
the state equation is
Ãm n−m−1
!
X X
λi t i λn t
x(t) = Ki e + Km+i t e x (0) (1.43)
i=1 i=0

where Ki , i = 1, . . . , n are constant matrices depending on A and B. We see that when

the input is zero, then the system is

• Stable whenever all simple eigenvalues have real parts that are not positive, and
all multiple eigenvalues have real parts that are negative, that is, if

Re [λi ] ≤ 0, λi simple eigenvalue

(1.44)
Re [λi ] < 0, λi multiple eigenvalue

• Asymptotically stable if all eigenvalues are negative, that is, if

Re [λi ] , < 0 i = 1, . . . , n (1.45)

1.2.7 Stability analysis using a linearized model

The stability of a nonlinear system around an equilibrium can be studied by analyzing the
linearization of the system around the equilibrium. Then (Slotine 1991), (Khalil 1996)

• If the linearized system is asymptotically stable, then the nonlinear system is also
asymptotically stable.

• If the linearized system is stable, but with at least one pole at the imaginary axis,
then the nonlinear system may be stable or unstable.

• If the linearized system is unstable, then the nonlinear system is unstable.

10 CHAPTER 1. MODEL REPRESENTATION

Example 6 We will demonstrate this for a pendulum. The nonlinear system is

θ̈ + 2ζω 0 θ̇ + ω 20 sin θ = 0 (1.46)

where ω 20 = g/L and 2ζω 0 θ̇ is a viscous damping term where 0 ≤ ζ. First, linearization
around (θ, θ̇) = (0, 0) gives
θ̈ + 2ζω 0 θ̇ + ω 20 θ = 0 (1.47)
If ζ > 0 then the linearized system is asymptotically stable. This implies that the non-
linear system is asymptotically stable around (θ, θ̇) = (0, 0) for ζ > 0. If ζ = 0, then
the eigenvalues of the systems are at the imaginary axis, and we cannot conclude on
the stability of the nonlinear system by analyzing the linear system. Next, consider the
equilibrium point (θ, θ̇) = (π, 0). The linearized system is

θ̈ + 2ζω 0 θ̇ − ω 20 θ = 0 (1.48)

which has one pole in the right half plane. The linearized system is therefore unstable, and
we may conclude that the nonlinear system is unstable around the equilibrium (θ, θ̇) =
(π, 0).

1.3 Transfer function models

1.3.1 Introduction
Linear time-invariant systems may be represented by transfer functions based on the use
of the Laplace transform. This makes it possible to use important analysis and design
methods in the Laplace description, and it serves as a good starting point for using
frequency response techniques. The Laplace transformation is easier to use than the
Fourier transformation, and it is a more general and powerful tool in controller design
and analysis. Moreover, if the Fourier transformation exists, it is obtained as a special
case of the Laplace transformation by using s = jω for the complex variable s.

1.3.2 The transfer function of a state-space model

In this section we will derive the transfer function corresponding to a linear time-invariant
state-space model, and present some useful results. A linear time-invariant system

ẋ (t) = Ax (t) + Bu (t) (1.49)

y (t) = Cx (t) + Du (t) (1.50)

where x = (x1 , . . . , xn )T can be described by a transfer function using the Laplace

transformation. We use the notation

x (s) = L{x (t)}, u (s) = L{u (t)} and y (s) = L{y (t)} (1.51)

The Laplace transform of the time derivative of the state x is given by

L{ẋ (t)} = sL{x (t)} − x (t = 0) (1.52)

In the development of transfer function models the initial conditions x (t = 0) are always
set to zero. This can be done as the system is linear and superposition applies. Therefore
we set ẋ (t = 0) = 0 and get
L{ẋ (t)} = sx (s) (1.53)
1.3. TRANSFER FUNCTION MODELS 11

Then the Laplace transformed state-space model is found to be

sx (s) = Ax (s) + Bu (s) (1.54)
y (s) = Cx (s) + Du (s) (1.55)
We eliminate x (s) using the first equation and insert the expression into the second
equation. This gives h i
y (s) = C (sI − A)−1 B+D u (s) (1.56)
We define the transfer function H (s) from u (s) to y (s) as
h i
H (s) = C (sI − A)−1 B+D (1.57)

and write
y
(s) = H (s) . (1.58)
u
A block diagram is shown in Figure 1.3.

us ys
Hs

Figure 1.3: Transfer function representaion of system

1.3.3 Rational transfer functions

A transfer function is said to be rational if it can be written in the form
P (s)
H(s) = K (1.59)
Q(s)
where the scalar K is called the gain, P (s) is a polynomial in the complex variable s of
degree m, and Q (s) is a polynomial in s of degree n. A rational transfer function can be
factored in the form
(s + z1 ) (s + z2 ) . . . (s + zm )
H (s) = K (1.60)
(s + p1 ) (s + p2 ) . . . (s + pn )
The transfer function is said to have m zeros at s = −zi and n poles at s = −pi . The
poles and the zeros may be real, or they can appear as complex conjugated pairs. We see
that the transfer function is defined and continuous for all s except for the poles, which
are the singularities of a rational transfer function. The transfer function is said to be
proper if there are at least as many poles as zeros, that is, if n ≥ m, and it is said to
be strictly proper if there are more poles than zeros, that is, if n > m. If m > n, then
the transfer function is said to have m − n poles at infinity. In n > m, then the transfer
function is said to have n − m zeros at infinity.
It is noted that the transfer function of an n-dimensional state space model is a
proper rational transfer function with n poles under the assumption that all states are
controllable and observable.
Example 7 The transfer function H1 (s) = s is not proper, and has one pole at infinity,
while the transfer function H2 (s) = 1/s is a strictly proper transfer function with one
zero at infinity.
12 CHAPTER 1. MODEL REPRESENTATION

1.3.4 Impulse response and step response

The Dirac delta function δ (t), which is referred to as a unit impulse function, has the
Laplace transform L{δ (t)} = 1. Thus, the response of the system when the input is a
unit impulse is
y (s) = H (s) · 1 = H (s) (1.61)
which corresponds to the time function

y (t) = h(t) := L−1 {H (s)} (1.62)

Therefore, h (t) is referred to as the impulse response of the system.

We define the unit step function
½
0, t < 0
us (t) = (1.63)
1, 0 ≤ t

which has the Laplace transform

1
us (s) = L{us (t)} = (1.64)
s
The step response of the system, which is the response y (t) resulting from an initial value
y (t = 0) = 0 and the input u (t) = us (t), is

1
y (s) = H (s) (1.65)
s

u 1 x y
T
-

us ys
1
1Ts

Figure 1.4: A time constant

Example 8 The dynamic system

1
ẋ = (−x + u) (1.66)
T
y = x (1.67)

is referred to as a time constant. A block diagram is shown in Figure 1.4. The Laplace
transformation gives the transfer function
y 1
(s) = H (s) = (1.68)
u 1 + Ts
1.3. TRANSFER FUNCTION MODELS 13

The impulse response of the system is

t
h (t) = L−1 {H (s)} = e− T (1.69)

while the step response of the system is

½ ¾
−1 H (s) t
y (t) = L = 1 − e− T (1.70)
s

di
Example 9 By setting all initial values dti x (t) = 0, i = 1, . . . , n we find that
½ ¾
dn
L x (t) = sn X (s) (1.71)
dtn

u x x y
-

2 0

 20

us 1
ys
2
s 2 2 0  0

Figure 1.5: Second order oscillatory system

Example 10 The model

ẍ (t) = −2ζω 0 ẋ (t) − ω 20 x (t) + u (t) (1.72)

y (t) = x (t) (1.73)

given as a block diagram in Figure 1.5 is Laplace transformed to

s2 x (s) = −2ζω 0 sX (s) − ω 20 x (s) + u (s) (1.74)

which is solved for x (s) to give

1
x (s) = u (s) (1.75)
s2 + 2ζω 0 s + ω 20

The transfer function is

y 1
(s) = H (s) = 2 (1.76)
u s + 2ζω 0 s + ω 20
14 CHAPTER 1. MODEL REPRESENTATION

y d s es u s  ys

Ks Hs
-

Ls
Gs

Figure 1.6: Plant H (s) with a series compensation controller K (s) in feedback compen-
sation controller G (s)

1.3.5 Loop transfer function

We consider a plant H (s) with a series compensation controller K (s) and feedback
compensation controller G (s) as shown in Figure 1.6. The plant is given by

y (s) = H (s) u (s) (1.77)

and the controller is given by

u (s) = K (s) e (s) , e (s) = yd (s) − G (s) y (s) (1.78)

where yd is the input signal to the closed-loop system. Define the loop transfer function
by
L (s) = K (s) H (s) G (s) . (1.79)
From
e (s) = yd (s) − G (s) y (s) = yd (s) − G (s) H (s) K (s) e (s) (1.80)
it is seen that the transfer function from the input signal yd (s) to the error signal e (s)
is given by
e 1
S (s) := (s) = (1.81)
yd 1 + L (s)
where S (s) is called the sensitivity function of the closed loop system.
The closed-loop transfer function T (s) is defined as the transfer function from the
closed-loop input yd (s) to the output y (s). From the expression

y (s) = K (s) H (s) e (s) = K (s) H (s) [yd (s) − G (s) y (s)] (1.82)

it is possible to solve for y (s) as a function of yd (s), and the closed-loop transfer function
T (s) is found to be
y K (s) H (s)
T (s) := (s) = (1.83)
yd 1 + L (s)
The closed-loop transfer function can be written
½ 1
1 L (s) G(s) |L (s)| À 1
T (s) = ≈
G (s) 1 + L (s) K (s) H (s) |L (s)| ¿ 1

This means that if the loop-transfer function is large, that is, if |L (s)| À 1, then T (s) =
1/G (s).
1.3. TRANSFER FUNCTION MODELS 15

Example 11 If unity feedback is used, that is, if G (s) = 1, then

L (s)
T (s) = = 1 − S (s)
1 + L (s)

and T (s) is called the complementary sensitivity function.

1.3.6 Example: Actuator with dynamic compensation

u 0 s u a s us ys

1 Hs
Ka s
-

T is
1T is

Figure 1.7: System with dynamic feedback control of the actuator.

In many control systems there is a servomotor that acts as an actuator for the main
plant. In this case it can be useful to control the servomotor with an inner actuator loop
with dynamic feedback to achieve a suitable transfer function in the outer loop. Suppose
that the main plant is described by the transfer function model

y (s) = H (s) u (s) (1.84)

while the control input is obtained using a velocity controlled servomotor. The model
for the actuator is assumed to be given by
1
u (s) = ua (s) (1.85)
s
where ua (s) is the velocity command to the actuator. The actuator is here modeled by
an integration which is the transfer function of a motor with a velocity loop where ua is
the desired velocity input. The feedback for the actuator loop is given by

ua (s) = Ka [u0 (s) − Ga (s) u (s)] (1.86)

where the dynamic feedback is the high-pass filter

Ti s
Ga (s) = (1.87)
1 + Ti s
as shown in Figure 1.7. Then the loop transfer function of the actuator loop is

Ka Ti
La (s) = (1.88)
1 + Ti s
and we find that the closed loop transfer function of the actuator loop is given by

u 1 + Ti s 1
(s) = Kp (1.89)
u0 Ti s 1 + T1 s
16 CHAPTER 1. MODEL REPRESENTATION

where
Ka Ti Ti
Kp = , T1 = (1.90)
1 + Ka Ti 1 + Ka Ti
This is illustrated in Figure 1.8. If Kp and Ti are selected so that break frequency 1/T1
is much higher than the crossover frequency of the outer loop, then the actuator loop
will introduce integral action in the outer loop according to
y 1 + Ti s
(s) ≈ Kp H (s) (1.91)
u0 Ti s

u 0 s us ys

1Ti s 1
Kp Ti s 1T1 s Hs

Figure 1.8: System with the closed loop dynamics of the actuator with controller.

1.3.7 Stability of transfer functions

Consider the system
y(s) = H(s)u(s) (1.92)
where H(s) = C(sI − A)−1 B+D is the transfer function of the system. This system is
bounded-input-bonded-output stable, which is termed BIBO stable, if and only if all the
poles λi of H (s) have real parts that are less than zero, that is, Re λi < 0, i = 1, . . . , n.
This is shown as follows: The impulse response corresponding to the transfer function
H (s) is denoted h (t). Assume that the input is bounded according to

|u (t)| ≤ U for all t (1.93)

where U > 0 is a constant. The output is given by

Z ∞
y (t) = h (τ ) u (t − τ ) dτ (1.94)
0

Taking the absolute values on both sides, we find that

¯Z ∞ ¯
¯ ¯
|y (t)| = ¯¯ h (τ ) u (t − τ ) dτ ¯¯
Z 0∞
≤ |h (τ )| |u (t − τ )| dτ
0
Z ∞
≤ U |h (τ )| dτ (1.95)
0

Suppose that all the poles are to the left of −α. Then there is a constant k ≥ 0 so
that
|h (t)| ≤ ke−αt (1.96)
R ∞ −αt −1
and, using the fact that for α > 0 we have 0 e dt = α , it follows that
Z ∞
k
|h (τ )| dτ ≤ , if α > 0 (1.97)
0 α
1.3. TRANSFER FUNCTION MODELS 17

We find that
k2
|y (t)| ≤ U when α > 0 (1.98)
α
which shows that when Re [λi ] < 0, then y (t) is bounded whenever u (t) is bounded.
Suppose that all poles have real parts less than zero, except one pole in s = 0. Then
the control input u (t) = 1 will for large t give y (t) ∝ t → ∞. Next, suppose that all
poles have real parts less than zero, except a complex conjugated pole pair in s = ±jω 0 .
Then the control input u (t) = cos ω 0 t will give a response which for large t satisfies
y (t) ∝ t cos ω 0 t → ∞ where ∝ denotes proportional to. If there are poles with real part
larger than zero, than y (t) will be unbounded for a unit step input. This shows that an
unbounded output may occur for bounded input when there are poles on the imaginary
axis or to the right of the imaginary axis.

1.3.8 Stability of closed loop systems

The stability of a closed loop system can be analyzed by studying the sensitivity function
S (s). The closed loop system will be stable if the poles of S (s) have real parts that are
negative. This can be checked using one of the standard sufficient conditions on the
loop transfer function L (s), which are available from automatic control theory. Note,
however, that the conditions on L (s) are typically derived under certain assumptions on
the properties of L (s). The fundamental requirement for stability is that the poles of
S (s) do not have positive real part, and that multiple poles must have real parts that
are less than zero.

Example 12 Large tankers may be unstable, and the transfer function H (s) from the
rudder angle δ to the course angle ψ will then include a pole in the right half plane. An
example of this is the following model of a tanker (Blanke 1981), (Fossen 1994)

ψ 1 + Ta s
H(s) = (s) = K (1.99)
δ s(1 + T1 s)(T2 s − 1)
where K = 0.022, Ta = 38 s, T1 = 16 s and T2 = 192 s. The integration represented by
the factor s in the denominator is due to the integration from angular velocity around
the vertical axis to the course angle ψ. The transfer function has a pole at s = 1/T2 . An
autopilot with a PD controller
1 + T1 s
δ(s) = Kp (ψ − ψ) (1.100)
1 + 0.1T1 s 0
gives the characteristic equation

s(1 + 0.1T1 s)(T2 s − 1) + KKp (1 + Ta s) = 0 (1.101)

for the closed loop system. The closed loop poles are therefore at s = −0.5609 and
s = −0.0295 ± j0.0303. This is found using the MATLAB command
roots(conv([1.6 1 0],[192 -1]) + 20*0.022*[0 0 38 1])

1.3.9 Partial differential equations

Systems described by partial differential equations will typically lead to irrational trans-
fer functions. Irrational transfer functions can be approximated by a rational transfer
function with infinitely high order, and because of this such systems may be referred to
18 CHAPTER 1. MODEL REPRESENTATION

as infinite dimensional systems. An irrational transfer function is said to be analytic in

a region if it is defined and continuous in that region. The points where an irrational
transfer function ceases to be analytic are called the singularities of the transfer function.
We recall that for a rational transfer function the singularities are called poles.
We will demonstrate the appearance of irrational transfer functions for systems de-
scribed by partial differential equations by studying the partial differential equation
∂v (x, t) ∂v (x, t)
c =− , v (0, t) = v1 (t) (1.102)
∂x ∂t
This is the first order wave equation which describes the propagation of a wave-front
with velocity c. The variable v(x, t) has the Laplace transform v (x, s) = L {v (x, t)}, and
the time derivative has the transform
½ ¾
∂v (x, t)
L = sL {v (x, t)} = sv (x, s) (1.103)
∂t
From this it follows that the partial differential equation has the Laplace transform
∂v (x, s)
c = −sv (x, s) , v (0, s) = v1 (s) (1.104)
∂x
This is an ordinary differential equation of the first order in s, which has the solution
x
v (x, s) = v (0, s) exp(− s) (1.105)
c
The transfer function from v1 (s) to v2 (s) := v (L, s) at x = L is then found to be the
irrational transfer function
v2
(s) = e−T s (1.106)
v1
where T = L/c is the propagation time. We see that the solution at x = L is equal to
the solution at x = 0 with a time delay T .

Example 13 The time delay in (1.106) can be approximated by a rational Padé approx-
imation Pkk (−T s) of order k where (Golub and van Loan 1989, p. 557)
Qkk (s)
Pkk (s) = (1.107)
Qkk (−s)
Xk
k! (2k − i)! si
Qkk (s) = 1 + (1.108)
i=1
(k − i)! (2k)! i!

A third order Padé approximation is found to be given by

Ts (T s)2 (T s)3
−T s 1− + −
e ≈ P33 (−T s) = 2 10
(T s)2
120
(T s)3
(1.109)
Ts
1+ 2 + 10 + 120

By letting k go to infinity we can represent the time delay by a rational transfer function
of infinite dimension.

Example 14 Transmission line dynamics are described by the second order wave equa-
tion. A hydraulic transmission line where the outlet is open has the irrational transfer
function
sinh s
tanh s = (1.110)
cosh s
1.4. NETWORK DESCRIPTION 19

from the input flow to the input pressure. The transfer functions has zeros when the
numerator is zero, which is the case when
1¡ s ¢
sinh s = e − e−s = 0 ⇒ e−2s = 1 = ej2kπ (1.111)
2
This occurs for
s = jkπ (1.112)
where k = 0, ±1, ±2, . . .. In the same way we find that the singularities appear for
µ ¶
1
cosh s = 0 ⇒ s = ±j k + π (1.113)
2

It can be shown that the numerator and the denominator can be represented by infinite
dimensional polynomials in the complex variable s, and this gives the following infinite
dimensional representation of the transfer function
³ ¡ ¢2 ´ ³ ¡ s ¢2 ´ ³ ¡ s ¢2 ´
s 1 + πs 1 + 2π 1 + 3π ...
tanh s = ³ ¡ 2s ¢2 ´ ³ ¡ 2s ¢2 ´ ³ ¡ 2s ¢2 ´ (1.114)
1 + 3π 1 + 5π 1 + 7π ...

We see that there are infinitely many zeros and singularities along the imaginary axis.
Moreover, we see that the zeros and singularities alternate along the imaginary axes.
This implies that the phase of tanh jω is between -90◦ and +90◦ .

1.4 Network description

1.4.1 Introduction
The automatic control literature relies to large extent on the use of models that are
based on a signal-flow formulation. This means that different blocks of the model are
connected with signals that considered to flow in the direction of the signal arrow. We
might say that signal-flow description has unilateral interconnections. The reliance on
the signal-flow description is obviously due to the many control techniques based on a
signal-flow description of the physical plant in the form of state-space models and transfer
functions. Because of this, it is clear that modeling techniques for use in controller design
and analysis should provide methods for developing signal-flow models. However, there
are good reasons for deviating from a strict reliance on signal flow in the development
of mathematical models of physical systems. We will mention some arguments for this,
and then discuss what the consequences are.
Many physical systems that are important in control applications are conveniently
represented in an energy-flow description. In this case different blocks of the model are
connected so that energy flows in both directions, and we say that the formulation relies
on bilateral interconnections. The signals flowing between the blocks will then typically be
voltage and current in electrical systems, force and velocity in translational mechanical
systems, torque and angular velocity in rotational mechanical systems, pressure and
volumetric flow in isothermal flow problems, and enthalpy and mass flow in thermal flow
problems. Note that in this case it is not clearly defined in which direction a signal
propagates. The main advantage of an energy-flow formulation is that it well suited for
energy-based controller design using Lyapunov techniques and passivity. Moreover, it
20 CHAPTER 1. MODEL REPRESENTATION

Translation Rotation Thermal flow Hydrostatic flow Electrical

2
Effort e F (N) τ (Nm) h (J/kg) p (N/m ) u (V)
Flow f v (m/s) ω (1/s) w (kg/s) q (m3 /s) i (A)
Power P F v (W) τ ω (W) wh (W) pq (W) vi (W)

Table 1.1: Efforts and flows for physical systems.

leads naturally to object-oriented modeling, which is of great use in simulation. Energy-

flow models can be assigned signal flow directions so that the model can be described
in state-space, or with transfer functions. This means that the use of an energy-flow
description in the modeling phase goes well together with the use of signal-flow methods
in the controller design and analysis phase.
Object-oriented modeling is an approach where a model is developed for each physical
subsystem, and where the model of the total system is obtained by interconnecting the
models of the subsystems. To make this interconnection possible, it is necessary that
a suitable interface is defined between the subsystem models. As mentioned above,
such an interface has been established in the form of energy-flow variables. The main
advantage of this approach is that a library of models can be developed for different
physical subsystems, and models for a total system can then be established by simply
interconnecting these library models. Moreover, additional subsystems can easily be
attached to the systems, and subsystems can be upgraded or changed by simply changing
the relevant library module. This leads to re-use of models and straightforward updating
of subsystem models. This approach is very useful in the development of simulation
systems.
This chapter will present methods for this modeling technique. In particular we will
focus on the network description that to a large extent originates from electrical circuit
theory. We will also comment on the bond-graph formulation, which was developed for
energy-flow modeling of systems consisting of subsystems like electrical circuits, electri-
cal motors, hydraulic motors and mechanical parts. We will show that object-oriented
modeling leads to models that are well suited for use in control techniques based on state
space and transfer functions.

1.4.2 Background

The network description has been very successful in the analysis and design of electrical
circuits (Anderson and Vongpanitlerd 1973), (Nilsson 1983). The underlying properties
of an electrical circuit that make the network description efficient are seen also in other
types of physical systems like mechanical translation, mechanical rotation, thermal flow
and hydrostatic flow. To make the discussion general it is advantageous to define flow
variables f and effort variables e for these systems. The flow f corresponds to the
current i in an electrical network, while the effort variable e corresponds to the voltage
u in and electrical network. The efforts and flows for typical physical systems is shown
in Table 1.1.
1.4. NETWORK DESCRIPTION 21

1.4.3 Multiport
A multiport, which is also called an n-port, is a system with n ports. To port has an
effort ek and a flow fk so that the net power flowing into the n-port is given by
n
X
P = ek fk (1.115)
k=1

At each port of the n-port it is possible to connect a port of another m-port system
as long as the effort and flow variables are compatible, that is, an electrical port can be
connected with an electrical port, a translational port can be connected to a translational
port and so on.

1.4.4 Example: DC motor with flexible load

Transfer function models give a representation of a control system which is modular
with respect to the signal flow. It gives a clear model structure with well defined inputs
and outputs of the physical system and the controller. Therefore, when the signal-
flow description is used, it is convenient to connect the plant output to a controller,
and to connect the output of the controller as the input to the plant. Moreover, it is
straightforward to change controller type and controller structure in this setting.
However, signal flow models may be cumbersome to modify when the physical system
is modified. To illustrate this we may consider a DC motor with inertia Jm with no load.
We let the output be the motor velocity ω m , while the input is the motor torque T . The
transfer function is found from the equation of motion

Jm ω̇ m (t) = T (t) (1.116)

which gives the transfer function model

ωm 1
(s) = (1.117)
T Jm s
Suppose that there is a need to modify the model to account for a load with inertia J1
that is driven through a transmission with a spring and a damper. Then the elasticity
of the transmission has to be included in the transfer function. The transfer function
changes to
³ ´2
s s
ωm 1 1 + 2ζ a ωa + ωa
(s) = ³ ´2 (1.118)
T Js
1 + 2ζ 1 ωs1 + ωs1

where the parameters are given by

Jm J1
Je = and J = Jm + J1 (1.119)
J
r r r
K D 1 Jm Jm
ω1 = , ζ1 = √ , ωa = ω1 and ζ a = ζ (1.120)
Je 2 Je K J J 1
The derivation of this transfer function requires some work, but it is possible to do by
hand. A critical observation is that the model parameters are functions of both the motor
parameters and the load parameters.
22 CHAPTER 1. MODEL REPRESENTATION

If one more elastic transmission and an inertia J2 is added, then the transfer function
becomes
µ ³ ´2 ¶ µ ³ ´2 ¶
s s
1 + 2ζ 2a ω2a + ω2a s
1 + 2ζ 2b ω2b + ωs2b
ωm 1
(s) = µ ³ ´2 ¶ µ ³ ´2 ¶ (1.121)
T (Jm + J1 + J2 ) s s s
1 + 2ζ 21 ω21 + ω21 s
1 + 2ζ 22 ω22 + ωs22

At this level the modeling becomes quite complicated, and the inclusion of even more
degrees of freedom will lead to very extensive modeling efforts. We conclude that the
inclusion of new physical objects in the model may be quite complicated to account for
in the transfer function setting, and that the complexity increases dramatically when the
order of the system increases. In particular, we note that in the present formulation we
do not have a transfer function library of motors and elastic transmissions that can easily
be combined.
In contrast to this the network approach to modeling makes it possible to assemble
the model from the models of the physical objects of the system. In that case the modules
are connected through the ports. The motor model is
Jm ω̇ m = Tm − TL (1.122)
If the motor is running alone, then TL is set to zero. The inertia J1 and the flexible
transmission is described by
J1 ω̇ 1 = TL − T1 (1.123)
d
(θm − θ1 ) = (ω m − ω 1 ) (1.124)
dt
TL = D1 (ω m − ω 1 ) + K1 (θm − θ1 ) (1.125)
The motor model (1.122) and the load model (1.123—1.125) are then connected through
the port variables TL and ω m , while T1 is set to zero. Moreover, a second flexibility
described by
J2 ω̇ 2 = T1 − T2 (1.126)
d
(θ1 − θ2 ) = (ω 1 − ω 2 ) (1.127)
dt
T1 = D2 (ω 1 − ω 2 ) + K2 (θ1 − θ2 ) (1.128)
can be connected to the model with the port variables T1 and ω m with T2 = 0. We see
that this leads to what may be called a object-oriented approach where each physical
object has a model, and where a model of an interconnection of physical objects is
obtained by simply interconnecting the models of the objects through port variables.
This approach scales well in the sense that the complexity of the modeling does not
increase with the order of the model. The key to this is the careful selection of the
interconnection variables. Note that once the model has been established, a state-space
model is available, and a signal-flow model can be obtained by selecting input and output.
This makes it possible to apply controller design based on signal flow representation.

1.4.5 Example: Voltage controlled DC motor

To illustrate the combination of electrical and mechanical ports in the network setting
we discuss the model structure of a voltage controlled DC electrical motor with an in-
ertial load that is driven over an elastic shaft. The port interconnections are shown in
1.4. NETWORK DESCRIPTION 23

u a ,i a T L , 1 T 1 , 1
DC Elastic
Load
motor shaft

Figure 1.9: DC motor with elastic shaft and inertial load

Figure 1.9. The DC motor with constant field is connected to the electric supply by the
terminals of the armature circuit and the motor shaft. The motor may be described as
a two-port where one port is electrical and one port is mechanical. The electrical port
is the armature port where the effort variable is the armature voltage ua and the flow
variable is the armature current ia . The mechanical port is the motor shaft where the
effort is the load torque TL and the flow is the angular velocity ω 1 . The dynamic model
of the motor is given in state-space form as
dia
La = −Ra ia − KE ω m + ua (1.129)
dt
Jm ω̇ m = KT ia − TL (1.130)
θ̇m = ωm (1.131)

Suppose that the motor shaft is connected to a mechanical two-port describing a spring
and a damper. The port on the motor side has port variables TL and ω m , and the port
on the load side has variables T1 and ω 1 . The model is
d
(θm − θ1 ) = ω m − ω 1 (1.132)
dt
TL = K(θm − θ1 ) + B(ω m − ω 1 ) (1.133)
T1 = TL (1.134)

Finally, the spring and the damper is connected to an inertial load described as a one-port
with port variables T1 and ω 1 . The model is

Jm ω̇ 1 = TL (1.135)
θ̇1 = ω1 (1.136)

1.4.6 Example: Diesel engine with turbocharger

A model of a diesel engine with turbocharger can be described as a system of three
multiports, namely, the compressor, the turbine, and the diesel engine. The port inter-
connections are shown in Figure 1.10. The turbine and the compressor have a common
shaft, and it is convenient to include the inertia of the shaft in the turbine model.
The compressor has one port for the air intake, one port for the air outlet, and one for
the turbocharger shaft. The port variables for the air intake is the specific enthalpy hci
of the air, which is the effort, and the mass flow wci , which is the flow. The port variables
of the air outlet is the specific enthalpy hco and the mass flow wco . The port variables
for the turbocharger shaft is the compressor torque Tc and the angular velocity ω tc . The
turbine has an air intake port with port variables hti and wti , and air outlet port with
port variables hto and wto , and the turbine shaft has a port with variables Tc and the
angular velocity ω tc . The diesel engine has an air intake port connected to the air outlet
24 CHAPTER 1. MODEL REPRESENTATION

jQ Tm ,  m

Diesel
h co , w co engine h ti , w ti

T c ,  tc
Compressor Turbine

h ci , w ci h to , w to

Figure 1.10: Diesel engine with turbocharger in network description

port of the compressor with port variables hco and the mass flow wco . The engine has an
exhaust outlet port connected to the air intake port of the turbine with port variables
hti and wti . In addition, the motor drives the motor shaft which can be described as a
port with variables being the motor torque Tm and the motor shaft speed ω m . We note
that the power delivered from the engine is Pm = Tm ω m . Finally, the engine has a fuel
injector where the power added is the heat rate jQ of the fuel. However, this is not a
port in the usual sense as there is no bilateral energy flow.

1.4.7 Assigning computational inputs and outputs

In the network description using multiports the signal flow directions are not specified.
This agrees with our intuition that energy flows both ways, and that the multiports
interact through the port interconnections. However, in simulation systems where the
multiport models are used in computations, it must be specified which of the port vari-
ables that is the input to the computation and which of the port variables that is the
output from the computation. This means that a signal flow structure is assigned to the
model. This must be done so that:

1. Differential equations can be evaluated by integration and not differentiation.

2. Computational inputs and computational outputs must be compatible when port

interconnections are made so that signal flow directions agree.

Example 15 Signal directions can be assigned according to Figure 1.11 for the DC motor
with load shown in Figure 1.9 so that integrations are used in the computations. This is
done by using ua and TL as inputs to the DC motor so that dia /dt, ω̇ m and θ̇m can be
evaluated from (1.129—1.131). For the spring ω m and ω 1 are used as inputs so that TL
and T1 can be computed from (1.132—1.134), and for the load the input is T1 so that ω̇ 1
and θ̇1 can be computed from (1.135—1.136).
1.4. NETWORK DESCRIPTION 25

ua m T1
DC Elastic
ia motor TL shaft
1 Load

Figure 1.11: DC motor and load in network descriptions with signal directions assigned

Example 16 Signal directions are indicated in Figure 1.12 for the diesel engine with
turbocharger as shown in Figure 1.10. Given the turbocharger speed ω tc and the specific
enthalpies hci and hco the compressor will produce a mass flow wc to the engine, and
a compressor torque Tc on the shaft. Given the compressor torque Tc and the specific
enthalpies hti and hto , the turbine dynamics will give the turbocharger speed ω tc and the
exhaust mass flow wt . The gas contained in the diesel engine will have as inputs the
mass flow wc into the engine, the mass flow wt out of the engine, the motor shaft speed
ω m , and the heat flow jQ . The resulting outputs are the specific enthalpies hco and hti ,
and the motor torque Tm .

jQ m Tm

Diesel
wc h co engine wt h ti

Tc

Compressor  tc Turbine

h ci wc h to wt

Figure 1.12: Turbocharged diesel engine with signal directions assigned for the port
variables

Example 17 Karnopp’s friction model extends the basic Coulomb friction model for
dry friction to be valid also for zero velocity (Karnopp 1985). Karnopp’s model can be
explained by considering a mass m with velocity v that is pushed on a flat surface with
an actuator force Fa . The friction force on the mass is Ff so that the equation of motion
is
mv̇ = Fa − Ff (1.137)

The Karnopp model can be formulated as a two-port where the input port has effort Fa
26 CHAPTER 1. MODEL REPRESENTATION

and flow v, while the output port has effort Fa − Ff and flow v, and where
½
Fa v = 0 and |Fa | ≤ Fc
Ff = (1.138)
Fc sgn(v) else

Note that the computational input of the model at the input port is Fa under the condition
v = 0 and |Fa | ≤ Fc , and that the computational input changes to v when the condition
does no longer hold. Further discussion on this model is found in Chapter 5.

1.4.8 Bond graphs

In the basic definition of network models there is no assignment of the direction of signal
flow. However, to use network models in simulation the signal flow directions must be
specified. This gives a network with specified signal flow structure. By adding a special
graph representation of the network and the of signal flow, a formalism called bond graphs
is obtained (Karnopp, Margolis and Rosenberg 2000). Bond graph theory provides us
with interesting tools for connecting network modules when the signal flow structure is
important, as is the case if the model is to be used for simulations. Note that bond-graph
models combines well with state-space formulations of networks, where the bond graph
techniques can be used to assign signal-flow directions. The terminology used in bond
graphs is easily related to the network terminology. In particular, a bond is the same as
a port, and computational causality is related to the selection of inputs and outputs to a
computation.
In the bond graph setting the interconnection between multiports are given by bonds.
Each bond transmits an effort signal e and a flow signal f . The effort and flow are
selected so that the product ef has dimension power. We note that the effort and flow
for a bond corresponds to the port variables in the network description. A bond is
assigned a direction, which is the direction of positive power flow. From the outset we
are free to choose the direction of the effort signal for a bond, although in interconnections
of bonds there are certain rules that must be obeyed. The topic of bond graphs will not
be further discussed in this book, but the methods that are used in this book are closely
related to the concept of bond graphs.

1.5 Linear network theory

1.5.1 Driving point impedance
An electrical multiport is an electrical circuit which can be connected to the outside via
n ports. Each port has one positive and one negative electrical terminal. The voltage
over the terminals of port k is uk , while the current into the positive terminal and out
of the negative terminal is ik . This means that the power flowing into the circuit from
port k is Pk = uk ik . The port of one electrical multiport can be connected to a port of
another electrical multiport by connecting the terminals of the two ports.
A linear time-invariant electrical one-port with voltage u and current i can be de-
scribed by its driving-point impedance Z(s) which is the impedance of the port so that

u (s) = Z (s) i (s) (1.139)

In the same way the driving-point admittance Y (s) = Z(s)−1 satisfies

i (s) = Y (s) u (s) (1.140)

1.5. LINEAR NETWORK THEORY 27

1 i
+
u C
-

2 i
+
u R C
-

3 i
+ R L C
u
-

4 i
+ R1
u L C R
-

Figure 1.13: Passive eletrical one-ports

To become more familiar with the notion of the on the driving point impedance of
passive electrical one-ports we present the driving-point impedance of the one-ports in
Figure 1.13. Circuit 1 is a capacitor, circuit 2 is a parallel interconnection of a resistor
and a capacitor, circuit 3 is a RLC series connection, and circuit 4 is a resistor in series
with a parallel RLC interconnection. The driving point impedances are
1
Z1 (s) = (1.141)
Cs
R
Z2 (s) = (1.142)
1 + RCs
1 + RCs + LCs2
Z3 (s) = (1.143)
Cs
³ ´
L L
1 + R1 + R s + LCs2
Z4 (s) = R1 L
(1.144)
1+ Rs + LCs2
We see that Z1 (s) and Z2 (s) are strictly proper, Z3 (s) is not proper, and Z4 (s) is proper.
It is straightforward to verify that the poles of the impedances appear in the left half
plane, or as simple poles on the imaginary axis, and that the phase of the impedances is
between -90◦ and +90◦ .

We define the driving point impedance Z(s) and the driving point admittance Y (s) of a
general one-port with effort e and flow f according to

e (s) = Z (s) f (s) , f (s) = Y (s) e (s) (1.145)

28 CHAPTER 1. MODEL REPRESENTATION

1.5.2 Linear two-ports

A linear two-port can be described with a transfer function matrix. The convention is
then that the flows are positive when they are directed into the two-port. The transfer
function matrices give the relation between two input variables and two output variables.
Different transfer function matrices result depending on which variables that are selected
as inputs and outputs variables. In the impedance description the flows are inputs and
the efforts are outputs, and the transfer function model of the two-port is written
µ ¶ µ ¶µ ¶
e1 (s) z11 (s) z12 (s) f1 (s)
= (1.146)
e2 (s) z21 (s) z22 (s) f2 (s)
In the admittance form the efforts are inputs and the flows are outputs, and the model
is written µ ¶ µ ¶µ ¶
f1 (s) y11 (s) y12 (s) e1 (s)
= (1.147)
f2 (s) y21 (s) y22 (s) e2 (s)
A cascade formulation can be used where the variables of port two are inputs and the
variables of port one are outputs, or vice versa, which leads to the two model represen-
tations µ ¶ µ ¶µ ¶
e1 (s) a11 (s) a12 (s) e2 (s)
= (1.148)
f1 (s) a21 (s) a22 (s) −f2 (s)
and µ ¶ µ ¶µ ¶
e2 (s) b11 (s) b12 (s) e1 (s)
= (1.149)
f2 (s) b21 (s) b22 (s) −f1 (s)
The hybrid formulation is based on having the effort of one port and the flow of the other
port as inputs, and then having the remaining flow and the remaining effort as outputs.
This gives the two alternative model formulations
µ ¶ µ ¶µ ¶
e1 (s) h11 (s) h12 (s) e2 (s)
= (1.150)
f2 (s) h21 (s) h22 (s) f1 (s)
and µ ¶ µ ¶µ ¶
f1 (s) g11 (s) g12 (s) f2 (s)
= (1.151)
e2 (s) g21 (s) g22 (s) e1 (s)
In (Nilsson 1983) this is discussed in further detail, and formulas for transforming the
parameters of one formulation to the parameters of anther formulation are given.

1.5.3 Impedance of two-port with termination

Consider the case where a linear two-port is terminated with the impedance ZL (s) at
port 2. This means that port two is connected to a one-port
eL (s) = ZL (s) fL (s) (1.152)
with impedance ZL (s). The interconnection is ensured with the conditions
e2 = eL , f2 = −fL (1.153)
From the impedance formulation (1.146) of the two-port it is seen that the termination
gives µ ¶ µ ¶µ ¶
e1 (s) z11 (s) z12 (s) f1 (s)
= (1.154)
−ZL (s) f2 z21 (s) z22 (s) f2 (s)
1.5. LINEAR NETWORK THEORY 29

This set of equations can be solved to find the driving-point impedance description

e1 (s) = Z1 (s)f1 (s) (1.155)

of the resulting one-port. From the second row of (1.154) it is found that
z21
f2 = − f1 (1.156)
z22 + ZL

Insertion in the first row of (1.154) gives

µ ¶
z12 z21
e1 = z11 − f1 (1.157)
z22 + ZL

Therefore, the terminated one-port has driving-point impedance

z12 z21
Z1 = z11 − (1.158)
z22 + ZL

1.5.4 Example: Passive mechanical two-port

K1
F1 F2
m1

v1 D1 v2

Figure 1.14: Mechanical two-port

The mechanical system in Figure 1.14 can be described as a two-port. A mass m1 is

connected to a parallel interconnection of a spring with stiffness K1 and a damper with
coefficient D1 . Port 1 is connected to the spring and the damper, and has flow equal to
the velocity v1 and effort equal to the force F1 . Port 2 is connected to the mass, and has
flow v2 and effort F2 . The equation of motion for the mass is given by Newton’s law to
be
m1 ẍ2 = F1 − F2 (1.159)
where the force from the spring and the damper is

F1 = K1 (x1 − x2 ) + D1 (v1 − v2 ) (1.160)

Laplace transformation gives

m1 sv2 (s) = F1 (s) − F2 (s) (1.161)

and
K1 + D1 s
F1 (s) = [v1 (s) − v2 (s)] (1.162)
s
30 CHAPTER 1. MODEL REPRESENTATION

as sx1 (s) = v1 (s) and sx2 (s) = v2 (s). By solving for F2 (s) and v2 (s) we get the cascade
description
µ ¶ Ã m1 s2 +D1 s+K1 !µ ¶
F2 (s) D1 s+K1 −m 1 s F1 (s)
= s (1.163)
v2 (s) − D1 s+K 1
1 v1 (s)
Alternatively, we may develop the impedance description
µ ¶ Ã K1 +D1 s K1 +D1 s
!µ ¶
F1 (s) s s v1 (s)
= K1 +D1 s m1 s2 +D1 s+K1 (1.164)
F2 (s) s s
−v2 (s)

or the hybrid description

µ ¶ Ã s
!µ ¶
v1 (s) K1 +D1 s −1 F1 (s)
= (1.165)
F2 (s) 1 m1 s −v2 (s)

We note that the mechanical two-port has an electrical analog as shown in Figure 1.15.

i1 i2
m1 s
+ +
u1 D1 u2
K1
- s -

Figure 1.15: Electrical analog of mechanical two-port. The voltages u1 and u2 are the
analogs of the forces F1 and F2 , while the currents i1 and i2 are the analogs of the
velocities v1 and v2 .

K1 F2 K2
F1
m1

v1 D1 v2 D2

Figure 1.16: Termination of mechanical two-port with mechanical one-port with spring
and damper.

Suppose that the two-port of the previous section is terminated at port 2 by a me-
chanical one-port consisting of a parallel interconnection of a spring with stiffness K2
and a damper with coefficient D2 connected to a fixed point. The resulting mechanical
one-port is shown in Figure 1.16. The impedance of the termination is
D
K2 + D2 s 1 + K22 s
ZL (s) = = K2 (1.166)
s s
so that
F2 (s) = ZL (s)v2 (s) (1.167)
1.5. LINEAR NETWORK THEORY 31

This gives the impedance

F1 z12 z21
(s) = Z1 = z11 − (1.168)
v1 z22 + ZL

After some calculation we find that

D2 m1 2
F1 K1 + D1 s K2 (1 + K 2
s+ K2 s )
(s) = D +D m1
(1.169)
v1 s K1 + K2 (1 + K +K s +
1 2 2
K1 +K2 s )
1 2

1.5.5 Mechanical analog of PD controller

We will now present mechanical analogs for PD controllers for position control when the
control input is the applied force. Note that this is a PI controller from velocity.

Kp
u

x x0
D

Figure 1.17: Mechanical analog of PD controller.

We consider a mass m with position x and velocity v = ẋ. The equation of motion is
mẍ = u where the applied force u is the control input. The desired position is xd , and
the desired velocity is vd = ẋd . A PD controller u = K(1 + Td s)(xd − x) is used. The
control law can be written

u = K(xd − x) + D(vd − v) (1.170)

where D = KTd . The mechanical analog is found from the observation that the controller
force is the same force as the force that would appear if the mass m at position x was
connected to a point of position xd with a parallel interconnection of a spring with
stiffness K and a damper with coefficient D, as shown in Figure 1.17.

Kp
u

x x0
D

Figure 1.18: Mechanical analog of PD controller without velocity reference.

An alternative control law which is used when vd is not available is

u = Kp (x0 − x) − Dv (1.171)
32 CHAPTER 1. MODEL REPRESENTATION

K2
K1
m

xd x1 D2

Figure 1.19: Mechanical analog mass controlled with lead-lag controller.

This is the force that would result if the mass m was connected by a spring of stiffness
K to a point xd and by a damper to a fixed point as shown in Figure 1.18.
If the velocity is not measured it is possible to use a PD controller with limited
derivative action. Then the control is
1 + Td s
u(s) = Kp [x0 (s) − x(s)] (1.172)
1 + αTd s

where 0 ≤ α ≤ 1. This control law gives the same force as a mechanical analog where
the mass m is connected to a point with position xd by a spring of stiffness K1 in series
with a parallel interconnection of a spring of stiffness K2 and a damper with coefficient
D2 as shown in Figure 1.19.
This is shown by letting x1 be the position of the interconnection point between the
spring K1 and the parallel interconnection. Then the spring force is u = K1 (x1 − x), and
Laplace transformation gives x1 (s) = x(s) + u(s)/K1 . As there is no mass in the point
x1 , this force must be equal to the force from the parallel interconnection, so that

u(s) = K2 [xd (s) − x1 (s)] + D2 [vd (s) − v1 (s)] = (K2 + Ds)[xd (s) − x1 (s)] (1.173)

Insertion of x1 (s) gives

1
u(s) = (K2 + D2 s)[x0 (s) − x(s) − u(s)] (1.174)
K1

We solve for u(s) and get

K2 + D2 s
u(s) = K1 [x0 (s) − x(s)]
K1 + K2 + D2 s
D2
K1 K2 1+ K 2
s
= K D2
[x0 (s) − x(s)] (1.175)
K1 + K2 1 + K +K
2
s
1 2 K2

We see that this is a PD controller with limited derivative action where

K1 K2
Kp = (1.176)
K1 + K2
D2
Td = (1.177)
K2
K2
α = ∈ [0, 1) (1.178)
K1 + K2
1.6. EXAMPLE: TRANSMISSION LINE MODEL 33

i Rdx Lsdx i  di

+ +
u 1
Cs
dx Gdx u  du
- -

Figure 1.20: Length element of electric transmission line.

1.6 Example: Transmission line model

1.6.1 Introduction
The dynamics of transmission lines are important in several applications including elec-
trical lines for signal transmission, water pipes for hydroelectric power plants, gas and
oil pipelines, and systems with hydraulic drives. It is by no means obvious that these
systems have dynamic properties in common, however, it turns out that a comprehensive
theory of transmission line dynamics can be developed to describe important dynamic
properties for all these systems in spite of their diverse nature. We will first derive
the partial differential equations describing a transmission line, and we will then study
techniques of analysis.

1.6.2 Introductory example

In the following we will present the model for a general transmission line of length L
which is described as a two-port where one port is at the the input side with effort e1
and flow f1 , and the other port at the output side with effort e2 and the flow f2 . The
model will be presented in the form of a partial differential equation with boundary
conditions depending on the port variables. Then this model will be used to derive
transfer functions. Also wave variables and impedance matching be discussed. Then, in
a later chapter the results will be specialized for hydraulic transmission lines will, where
the effort variable is the pressure and the flow variable is the volumetric flow.
To make the ideas of the following presentation clear we start with the equations of
an electrical transmission line as an introductory example. To derive the model of an
electric transmission line we consider a length element dx of the line which is described
with the length coordinate x. The length element is modelled as by a series impedance
consisting of a resistor Rdx and an inductor M dx, and a parallel admittance consisting
of an admittance Gdx and a capacitor Cdx. The voltage and current laws give

∂i
u(x + dx, t) − u(x, t) = −Rdxi(x, t) − M dx (x, t) (1.179)
∂t
∂u
i(x + dx, t) − i(x, t) = −Gdxu(x + dx, t) − Cdx (x + dx, t) (1.180)
∂t
Dividing by dx we get

∂u ∂i
(x, t) = −Ri(x, t) − M (x, t) (1.181)
∂x ∂t
∂i ∂u
(x, t) = −Gu(x, t) − C (x, t) (1.182)
∂x ∂t
34 CHAPTER 1. MODEL REPRESENTATION

Laplace transformation of (1.181) and (1.182) gives the transmission line equations
∂u
(x, s) = −(R + M s)i(x, s) (1.183)
∂x
∂i
(x, s) = − (G + Cs) u(x, s) (1.184)
∂x
Note that these equations have the form
∂u
(x, s) = −Z(s)i(x, s) (1.185)
∂x
∂i
(x, s) = −Y (s)u(x, s) (1.186)
∂x
where Z(s) = R + M s is the series impedance, and Y (s) = G + Cs is the parallel
admittance.

1.6.3 Effort and flow model

A general transmission line is characterized in terms of the line length L, the series
impedance Z(s), the parallel admittance Y (s), and the two first-order differential equa-
tions
∂e
(x, s) = −Z(s)f (x, s) (1.187)
∂x
∂f
(x, s) = −Y (s)e(x, s) (1.188)
∂x
in the Laplace domain, where e is the effort variable and f is the flow variable.

From this characterization we arrive at the following second order differential equation
for the effort
∂2e
L2 2 (x, s) = Γ2 (s)e(x, s) (1.189)
∂x
where we have introduced the propagation operator Γ(s) defined by
p
Γ(s) = L Z(s)Y (s) (1.190)

and the additional requirement that Re[Γ(s)] ≥ 0. The solution of this differential
equation is ³ x´ ³ x´
e(x, s) = C1 cosh Γ + C2 sinh Γ (1.191)
L L
where C1 and C2 are given by the boundary conditions. From (1.191) we see that the
flow is given by
1 ³ ³ x´ ³ x ´´
f (x, s) = − C1 sinh Γ + C2 cosh Γ (1.192)
Zc (s) L L
where the characteristic impedance Zc (s) is defined by
s
Z(s)
Zc (s) = (1.193)
Y (s)
1.6. EXAMPLE: TRANSMISSION LINE MODEL 35

1.6.4 Transfer functions

A transmission line of length L can be regarded as a two-port with the input port at
x = 0 and the output port at x = L. At the input port the effort is e(0, t) and the flow is
f (0, t), while at the output port the effort is e(L, t) and the flow is f (L, t).The transfer
functions between the port variables can be derived from (1.191, 1.192). We denote the
Laplace transformed port variables e(0, s) = e1 (s), f (0, s) = f1 (s), e(L, s) = e2 (s) and
f (L, s) = f2 (s). From (1.191, 1.192) it is seen that we may then write the solutions in
the form
µ ¶ µ ¶µ ¶
e1 (s) 1 0 C1 (s)
= (1.194)
Zc f1 (s) 0 −1 C2 (s)
µ ¶ µ ¶µ ¶
e2 (s) cosh Γ sinh Γ C1 (s)
= (1.195)
Zc f2 (s) − sinh Γ − cosh Γ C2 (s)

From these two equations we can eliminate the constants C1 C2 and find the cascade
form of the transfer functions.

The cascade form of the transfer functions is given by

µ ¶ Ã !µ ¶
e2 (s) cosh Γ −Zc sinh Γ e1 (s)
= (1.196)
f2 (s) − sinh
Zc
Γ
cosh Γ f1 (s)

µ ¶ Ã !µ ¶
e1 (s) cosh Γ Zc sinh Γ e2 (s)
= sinh Γ (1.197)
f1 (s) Zc cosh Γ f2 (s)

The cascade form can be used for analysis purposes like the derivation transfer func-
tions when the line is terminated with a load impedance. However, the cascade form
is an ill-posed boundary value problem, and numerical models derived from the cascade
form may be ill-conditioned (Mäkinen, Piché and Ellman 2000).
By straightforward manipulation of the equations (1.194, 1.195) we can find the trans-
fer functions between different combinations of port variables. First we consider the im-
mittance forms which are the impedance form and the admittance form. These forms
are found from
µ ¶ µ ¶µ ¶
e1 (s) 1 0 C1 (s)
= (1.198)
e2 (s) cosh Γ sinh Γ C2 (s)
| {z } | {z }| {z }
e G(s) c
µ ¶ µ ¶ ¶µ
f1 (s) 0 −1 C1 (s)
Zc = (1.199)
−f2 (s) sinh Γ cosh Γ C2 (s)
| {z } | {z }| {z }
f H(s) c

Elimination of the vector c gives the expressions

1
e = Zc G(s)H(s)−1 f , f = H(s)G(s)−1 e (1.200)
Zc
36 CHAPTER 1. MODEL REPRESENTATION

This gives the impedance form

µ ¶ µ cosh Γ 1
¶µ ¶
e1 (s) sinh Γ sinh Γ f1 (s)
= Zc 1 cosh Γ (1.201)
e2 (s) sinh Γ sinh Γ −f2 (s)

and the admittance form

µ ¶ µ cosh Γ 1
¶µ ¶
f1 (s) 1 sinh Γ − sinh Γ e1 (s)
= 1 cosh Γ (1.202)
−f2 (s) Zc − sinh Γ sinh Γ e2 (s)

The hybrid form of the transfer functions are found from

µ ¶ Ã !µ ¶
e1 (s) 1 0 C1 (s)
= sinh Γ cosh Γ (1.203)
−f2 (s) Zc Zc C2 (s)
| {z } | {z }| {z }
u E(s) c
µ ¶ Ã ¶ !µ
f1 (s) 0 − Z1c C1 (s)
= (1.204)
e2 (s) cosh Γ sinh Γ C2 (s)
| {z } | {z }| {z }
y F(s) c

Proceeding as in the impedance and admittance case we eliminate the c vector and get
y = F(s)E(s)−1 u, u = E(s)F(s)−1 y (1.205)

The hybrid forms are

µ ¶ Ã !µ ¶
1 1
f1 (s) Zc tanh Γ − cosh Γ e1 (s)
= 1 (1.206)
e2 (s) cosh Γ Zc tanh Γ −f2 (s)
µ ¶ Ã !µ ¶
1
e1 (s) Zc tanh Γ cosh Γ f1 (s)
= 1 1 (1.207)
−f2 (s) − cosh Γ Zc tanh Γ e2 (s)

1.6.5 Transfer function for terminated transmission line

Suppose that a termination in the form of an impedance ZL (s) is used at port 2, which
means that the effort at port 2 is given by
e2 = ZL f2 (1.208)
Insertion of the termination equation (1.208) in the transmission form (1.197) gives
µ ¶ Ã !Ã !
e1 (s) cosh Γ Zc sinh Γ 1
= sinh Γ 1 e2 (s) (1.209)
f1 (s) Zc cosh Γ ZL

and we find the transfer functions

e2 1
(s) = Zc
(1.210)
e1 ZL sinh Γ + cosh Γ
e2 (s) 1
= Zc Zc
(1.211)
f1 (s) ZL cosh Γ + sinh Γ
1.6. EXAMPLE: TRANSMISSION LINE MODEL 37

which can be combined to give

Zc
f1 1 ZL cosh Γ + sinh Γ
(s) = Zc
(1.212)
e1 Zc ZL sinh Γ + cosh Γ

1.6.6 Wave variables

The system (1.187, 1.188) can be written
µ ¶ µ ¶µ ¶
∂ e(x, s) 0 −Z(s) e(x, s)
= (1.213)
∂x f (x, s) −Y (s) 0 f (x, s)
| {z }
A

This system can be made diagonal by finding the eigenvalues and the eigenvectors of
the matrix A. The eigenvalues of A are found to be ±Γ(s)/L, and the system is made
diagonal with the following change of variables:

1
e(x, s) = (a(x, s) + b(x, s)) (1.214)
2
1
f (x, s) = (a(x, s) − b(x, s)) (1.215)
2Zc (s)

which correspond to

a(x, s) = e(x, s) + Zc (s)f (x, s) (1.216)

b(x, s) = e(x, s) − Zc (s)f (x, s) (1.217)

This leads to

∂a Γ(s)
(x, s) = − a(x, s) (1.218)
∂x L
∂b Γ(s)
(x, s) = b(x, s) (1.219)
∂x L
The variables a and b are called the wave variables of the transmission line. The wave
variables have the solutions
h xi
a(x, s) = a(0, s) exp −Γ(s) (1.220)
· L ¸
(L − x)
b(x, s) = b(L, s) exp −Γ(s) (1.221)
L

We define the wave variables at the end-points of the line to be a1 (s) = a(0, s), a2 =
a(L, s), b2 (s) = b(L, s) and b2 (s) = b(L, s). Then we get

a2 (s)
= exp [−Γ(s)] (1.222)
a1 (s)
b1 (s)
= exp [−Γ(s)] (1.223)
b2 (s)
38 CHAPTER 1. MODEL REPRESENTATION

a(0,s) a(l,s)
exp(-Ts)

gL (s)
b(0,s) b(l,s)
exp(−Ts)

Figure 1.21: Scattering representation of lossless transmission line with load gL (s) cor-
responding to load impedance zL (s) at x = .

1.6.7 Lossless transmission line

The transmission line is said to be lossless if the impedance Z(s) is purely inductive and
the admittance Y (s) is purely capacitive. This means that there are real and positive
constants M and C so that Z(s) = M s andpY (s) = Cs. In this case the characteristic
impedance is real and given by Zc = Z0 := M/C while the propagation operator Γ(s)
becomes √
Γ(s) = L M Cs = T s (1.224)
where
1
c := √ (1.225)
MC
is the wave propagation velocity and T = L/c is the propagation time. This gives the
transfer functions
a2 (s)
= exp (−T s) (1.226)
a1 (s)
b1 (s)
= exp (−T s) (1.227)
b2 (s)

which are pure time delays of T = L/c. It is seen that the a(x, t) propagates in the
positive direction with velocity c, while b(x, t) propagates in the negative direction with
velocity −c. Because of this the variables a and b can be seen as wave variables.

b1 (s)
= exp (−2T s) GL (s) (1.228)
a1 (s)

1.6.8 Line termination

Consider a transmission line with a load impedance ZL connected at x = L. Then

e(L, s) = ZL (s)f (L, s) (1.229)

and
b2 (s) ZL (s) − Zc
= =: GL (s) (1.230)
a2 (s) ZL (s) + Zc
1.6. EXAMPLE: TRANSMISSION LINE MODEL 39

z0
+ +
u1 u2 zL
- -

Figure 1.22: Electrical tranmission line with load impedance ZL (s) at x = .

The boundary conditions in terms of the wave variables a and b are

a(0, s) = a1 (s) (1.231)

b(L, s) = GL (s)a(L, s) (1.232)

The transfer function from a1 (s) to b1 (s) is given by

b1 (s) b1 (s) b2 (s) a2 (s)
= = exp (−2Γ(s)) GL (s) (1.233)
a1 (s) b2 (s) a2 (s) a1 (s)

Example 18 Consider a short circuit as the load, so that zL = 0. Then gL = −1, and
b1 (s)
= − exp (−2Γ(s)) (1.234)
a1 (s)
which in the lossless case is a pure time delay of 2T together with a change of sign. The
effort transfer function is obviously
e(L, s)
=0 (1.235)
e(0, s)
Example 19 Consider an open circuit as the load. In this case i(L, s) = 0, zL = ∞,
and gL = 1. The transfer function becomes
b1 (s)
= exp (−2Γ(s)) (1.236)
a1 (s)
which in the lossless case is a pure time delay of 2T . The effort transfer function is
e(L, s) 2 1
= = (1.237)
e(0, s) [exp (Γ(s)) + exp (−Γ(s))] cosh(Γ(s))
Example 20 Impedance matching is achieved with zL = Zc . Then gL = 0, and
b1 (s)
=0 (1.238)
a1 (s)
which means that no wave is reflected. The voltage transfer function is
e(L, s)
= exp (−Γ(s)) (1.239)
e(0, s)
which in the lossless case is a pure time delay.
40 CHAPTER 1. MODEL REPRESENTATION

Example 21 An electrical transmission line is lossless if R = 0 and G = 0. Then the

series impedance is Z(s) = M s is the series impedance, and the parallel admittance
p is
Y (s) = Cs. In this case the characteristic impedance is real and given by Zc = M/C
while the function Γ(s) becomes
√
Γ(s) = L M Cs = sT (1.240)

Here we have used the wave propagation time T = L/c, and the wave propagation velocity
1
c := √ (1.241)
MC
For a standard coaxial cable the wave velocity is 75 % of the speed of light, which gives
c = 2.25 × 108 m/s.
Chapter 2

Model analysis tools

2.1 Frequency response methods

2.1.1 The frequency response of a system
The frequency response of a system can be studied by investigating the properties of
the transfer function on the imaginary axis, that is, for s = jω. The starting point for
frequency response analysis is the transfer function description
y (s) = H (s) u (s) (2.1)
Suppose that H (s) is strictly proper and rational, and that all the poles of H (s) have
real parts less than zero. Then we find the frequency response function H (jω) from the
transfer function by inserting s = jω, which is shown in the following:
The impulse response function corresponding to the transfer function H (s) is h (t) =
L−1 {H (s)}. Physical system are causal, which means that they do not give any response
to an impulse before the impulse is applied. Therefore, for a causal system the impulse
response function h (t) is zero for t < 0. Suppose that H (s) is strictly proper and rational,
and that all the poles of H (s) have real parts less than
R ∞ zero. Then the impulse response
h (t) will decay exponentially, which implies that 0 |h (t)| dt exists. The frequency
response Z Z
∞ ∞
H (jω) := F{h (t)} = h (t) e−jωt dt = h (t) e−jωt dt (2.2)
−∞ 0
will exist as
¯Z ¯ Z Z
¯ ∞ ¯ ∞ ¯ ¯ ∞
¯ h (t) e −jωt
dt¯¯ ≤ |h (t)| ¯e−jωt ¯ dt ≤ |h (t)| dt (2.3)
¯
0 0 0

Moreover, we see that the Fourier transform is given by

H (jω) = H (s)|s=jω (2.4)
where H (s) is the Laplace transform of h (t) defined by
Z ∞
H (s) := L{h (t)} = h (t) e−st dt (2.5)
0

It turns out to be a great advantage to work with the Laplace transform, which contains
the Fourier transform as a special case.

41
42 CHAPTER 2. MODEL ANALYSIS TOOLS

Example 22 The frequency response of a time constant H (s) = (1 + T s)−1 is

1
H (jω) = (2.6)
1 + jωT
with magnitude and phase given by
1
|H (jω)| = q and ∠H (jω) = − arctan ωT. (2.7)
2
1 + (ωT )

2.1.2 Second order oscillatory system

k
d m F
x

Figure 2.1: Mass-spring-damper system

A mass-spring-damper system (Figure 2.1) has the equation of motion

mẍ + dẋ + kx = F (2.8)

where x is the position of the mass m, d is the viscous friction coefficient and k is the
spring constant. The input is the force F . The model can be normalized in to the form
1
ẍ + 2ζω 0 ẋ + ω 20 x = F (2.9)
m
where the undamped natural frequency is
r
k
ω0 = (2.10)
m
and the relative damping is
1 d d
ζ= = √ (2.11)
2ω 0 m 2 km
The transfer function is found by inserting

L{ẍ (t)} = s2 L{x (t)}, L{ẋ (t)} = sL{x (t)} (2.12)

which leads to
x 1 1
H (s) = (s) =
F m s2 + 2ζω 0 s + ω 20
1 1
= ³ ´2 (2.13)
k
1 + 2ζ ωs0 + ωs0
2.1. FREQUENCY RESPONSE METHODS 43

We assume that ζ < 1 which implies that the poles of the transfer function are complex
conjugated and given by µ q ¶
2
λ = −ζ ± j 1 − ζ ω 0 (2.14)

The frequency response is

1 1
H (jω) = ³ ´2 (2.15)
k ω
1− ω0 + j2ζ ωω0

In particular we find that

1 1 1
H (jω 0 ) = = −j (2.16)
k j2ζ 2ζk
This shows that the phase of the frequency response at ω = ω 0 is ∠H (jω 0 ) = −90◦ ,
and moreover, that the magnitude is
1
|H (jω 0 )| = (2.17)
2ζk
which is inversely proportional to the relative damping ζ.

2.1.3 Performance of a closed loop system

|Lj |
Gain for
required Slope of -1
accuracy around
crossover 0 dB 
Damping of
noise, and
robustness to
unmodelled
dynamics

Closed loop Open loop

control control

Figure 2.2: Performance requirements on the loop transfer function in a Bode diagram.

Frequency response techniques are well suited to specify the performance of a control
system. This involves specifications on the loop transfer function L (jω) in a Bode
diagram. We recall that the magnitude of the sensitivity function S (jω) will satisfy the
approximations
¯ ¯ (
¯ 1 ¯ |L (jω)|−1 , |L (jω)| À 1
|S (jω)| = ¯¯ ¯= (2.18)
1 + L (jω) ¯ 1, 1 À |L (jω)|
44 CHAPTER 2. MODEL ANALYSIS TOOLS

and that for unity feedback the magnitude of the closed-loop transfer function T (jω)
can be approximated by
¯ ¯ ½
¯ L (jω) ¯ 1 |L (jω)| À 1
|T (jω)| = ¯¯ ¯= (2.19)
1 + L (jω) ¯ |L (jω)| 1 À |L (jω)|
Typically, we would like the sensitivity |S (jω)| to be small for low frequencies to
reduce the effect of disturbances on the system in the low-frequency region. In addition,
we would like |T (jω)| to be small for high frequencies to reduce the influence of measure-
ment noise and the influence of unmodeled dynamics. This implies that |L (jω)| should
be large for low frequencies, and small for high frequencies. In addition, there has to be
a significant interval around the crossover frequency ω c , defined by |L (jω c )| = 1, where
the phase should be around ∠L (jω) ≈ −90◦ to ensure a sufficient phase margin. These
requirements on the loop transfer function L (jω) are indicated in Figure 2.2.

2.1.4 Stability margins

If the loop transfer function L (s) is rational and has no poles with real part larger than
zero, then the Bode-Nyquist criterion states that the system is stable if
|L (jω 180 )| < 1 and ∠L (jω c ) > −180◦ . (2.20)
This can be expressed as conditions on the gain margin ∆K and the phase margin φ as
1
∆K := >1 (= 0 dB) (2.21)
|L (jω 180 )|
φ := 180◦ + ∠L (jω c ) > 0◦ (2.22)
where ω c is the crossover frequency defined by |L (jω c )| = 1, and ω 180 is defined by
∠L (jω 180 ) = −180◦ . It is possible to specify the performance of a control system around
the crossover frequency in terms of the stability margins. This is often done by specifying
a phase margin φ = 45◦ and a gain margin ∆K = 6 dB.
Example 23 We note that for any system
1
∆K = 6 dB ⇒ L (jω 180 ) = −
2
⇒ S (jω 180 ) = 2 and T (jω 180 ) = −1 (2.23)
We also note that for any system we have
1√
φ = 45◦ ⇒ L (jω c ) = − 2(1 + j)
2
⇒ |S (jω c )| = |T (jω c )| = 1.3 (2.24)
This may be acceptable if the desired value yd and the disturbances are of low frequency.
However, in high performance motion control like in robotics, the desired value yd will
have a significant frequency content close to the crossover frequency, and bearing in mind
that
|y (jω)| = |T (jω)| |yd (jω)| (2.25)
we see that we will have an amplification of the desired value by a factor of 1.3 close to
the crossover frequency ω c . In robotics this could cause serious problems.
The lesson to be learned from this is that performance specifications on the gain mar-
gins are not directly related to the closed loop performance. Thus, for high performance
systems it may be useful to study the functions S (jω) and T (jω) directly.
2.2. ELIMINATION OF FAST DYNAMICS 45

2.2 Elimination of fast dynamics

2.2.1 Example: The electrical time constant in a DC motor
Consider the following model of a DC motor:

dω m Ra
Tm = ia (2.26)
dt K
dia K 1
Ta = −ia − ωm + ua (2.27)
dt Ra Ra
where ω m is the motor speed, ia is the armature current, ua is the armature voltage, Ta
is the electrical time constant, and Tm is the mechanical time constant defined by

La JRa
Ta = , Tm = (2.28)
Ra K2

The transfer function H(s) from ua to ω m is found from the Laplace-transformed model

Ra
Tm sω m (s) = ia (s) (2.29)
K
K 1
(1 + Ta s) ia (s) = − ω m (s) + ua (s) (2.30)
Ra Ra
to be
ωm 1 1
H(s) = (s) = (2.31)
ua K 1 + Tm s + Ta Tm s
Suppose that Ta ¿ Tm so that Tm ≈ Tm + Ta . Then the transfer function can be written
1 1
H(s) = (2.32)
K (1 + Tm s)(1 + Ta s)

Suppose that Ta is small, so that the break frequency 1/Ta is much higher than the
frequency range where the model will be used. The the transfer function can be approx-
imated with
1 1
H(s) = (2.33)
K (1 + Tm s)
which is obtained using the approximation

1 + Ta s ≈ 1 (2.34)

We will now discuss how the state-space model (2.26, 2.27) should be modified to re-
flect this approximation. This is best seen from (2.30) where the approximation (2.34)
amounts to writing
K 1
ia (s) = − ω m (s) + ua (s) (2.35)
Ra Ra
This give the approximated state-space model

dω m Ra
Tm = ia (2.36)
dt K
K 1
0 = −ia − ωm + ua (2.37)
Ra Ra
46 CHAPTER 2. MODEL ANALYSIS TOOLS

We can use the second equation to eliminate i using

K 1
ia = − ωm + ua (2.38)
Ra Ra
and the first equation becomes
dω m 1
Tm = −ω m + ua (2.39)
dt K
which is consistent with the simplified transfer function given by (2.33).

2.2.2 Nonlinear system

In the nonlinear case it is not possible to use frequency arguments to eliminate high
frequency dynamics. In that case the corresponding model formulation is

ẋ = f (x, z, t, ) (2.40)
ż = g(x, z, t, ) (2.41)

If is small so that ż ¿ g(x, z, t, ), then it may be possible to approximate the system

by inserting = 0. This gives

ẋ = f (x, z, t, 0) (2.42)
0 = g(x, z, t, 0) (2.43)

which is a differential-algebraic system.

The differential-algebraic system may be written as a ordinary differential equation
if
z = z(x, t) (2.44)
is a solution to 0 = g(x, z, t, 0). In this case the system can be represented by the model

ẋ = f (x, z(x), t, 0) (2.45)

This topic is discussed in great detail in (Khalil 1996)

2.3 Energy-based methods

2.3.1 Introduction
So far controller design based on state-space methods and frequency response has been
discussed. These methods form the basis of any fundamental course on automatic control,
and there is a wide range of methods, algorithms and software packages. An important
formulation that complements state-space and frequency response methods is based on
the use of balance laws, and in particular on the use of energy functions. To give the
reader an indication on why this can be useful, we briefly consider the following examples:
If we are studying robot control around a constant desired position, then the kinetic
energy of the robot will increase if the robot is unstable, and the kinetic energy will be
reduced if the robot is asymptotically stable. If we are applying active vibration control
on a mechanical structure, then the vibration energy increases if the controlled system
is unstable, while the energy decreases if the system is asymptotically stable.
2.3. ENERGY-BASED METHODS 47

On background of this it seems reasonable that the stability properties of a system

may be related to the time derivative of some energy function of the system. In this
section we will present results that are motivated from energy considerations for me-
chanical, electrical, and hydrostatic systems, and systems with and thermal flow. These
model formulations are very useful in controller design and in analysis of control systems.

2.3.2 The energy function

We define an energy function V (x, t) ≥ 0 for the system

ẋ = f (x, u, t). (2.46)

The function V (x, t) may be the total energy of the system, or it may be some other
function, usually related to energy. When the system evolves the time derivative of the
energy function V (x, t) is

∂V ∂V
V̇ = + f (x, u, t). (2.47)
∂t ∂x

which follows from the standard rules of time differentiation of a function of two variables.
We say that V̇ is the time derivative along the solutions of the system. Information about
the time derivative of the energy of a system may give valuable insight into properties
of the dynamics of the system. In particular, if V̇ ≤ 0, then the energy is monotonically
decreasing, which may be important in connection with stability considerations. The
analysis of energy functions and their time derivatives along the solutions of the system
forms the basis for Lyapunov’s stability theory (Slotine 1991), (Khalil 1996). This is an
important tool in nonlinear control theory.

Example 24 Consider the system

ẋ1 = x2 (2.48)
ẋ2 = −ω 20 x1 − 2ζω 0 x2 (2.49)

and the energy function

1 2 2 1 2
V = ω x + x (2.50)
2 0 1 2 2
The time derivative of V along the solutions of the system is
¡ ¢
V̇ = ω 20 x1 x2 + x2 −ω 20 x1 − 2ζω 0 x2 (2.51)

which gives
V̇ = −2ζω 0 x22 ≤ 0 (2.52)
Note that the energy of the system decreases proportionally with the relative damping ζ
whenever x2 6= 0.

2.3.3 Second-order systems

If a system is given as a second order system

ẍ = f (x, ẋ, t) (2.53)

48 CHAPTER 2. MODEL ANALYSIS TOOLS

then the time derivative of an energy function V (x, ẋ, t) along the solutions of the system
is
∂V ∂V ∂V
V̇ = + ẋ + f (x, ẋ, t) (2.54)
∂t ∂x ∂ ẋ
For the system in the Example 24 we could arrive at a second-order description by writing
x1 = x and x2 = ẋ. The dynamics could then be presented as the second order system

ẍ = −ω 20 x − 2ζω 0 ẋ (2.55)

and we find the time derivative of V = 12 ω 20 x2 + 12 ẋ2 along the solutions of the system to
be ¡ ¢
V̇ = ω 20 xẋ + ẋẍ = ω 20 xẋ + ẋ −ω 20 x − 2ζω 0 ẋ = −2ζω 0 ẋ2 (2.56)
This result is the same as the result in (2.52).

2.3.4 Example: Mass-spring-damper

Energy function
A mass m with position x is connected to a fixed point by a spring with spring constant k
and a damper with damping constant d as shown in Figure 2.1. The equation of motion
is
mẍ + dẋ + kx = 0 (2.57)
The potential energy of this system is U = 12 kx2 , while the kinetic energy is T = 12 mẋ2 .
The total energy is
1 1
V = T + U = mẋ2 + kx2 (2.58)
2 2
The time derivative of the energy function is

V̇ = mẋẍ + kxẋ (2.59)

The time derivative for solutions of the system is found by inserting the equation of
motion (2.57). This gives

V̇ = ẋ (−dẋ − kx) + kxẋ

= −dẋ2 (2.60)

Two observations are important at this point. First, V̇ < 0, which means that the energy
is not increasing, and second, the energy decreases because of power dissipated in the
damper.
From the expression of the energy we see that
r r
2 2
x(t) ≤ V , ẋ ≤ V (2.61)
k m
This means that if the energy V decreases to zero, then also the position x and the
velocity ẋ will decrease to zero. Moreover, because V decreases, V (t) will be less than
the initial value V0 . Therefore
r r
2 2
x (t) ≤ V0 , ẋ (t) ≤ V0 (2.62)
k m
This means that if the initial energy is small, then also the position and velocity will
remain small.
2.3. ENERGY-BASED METHODS 49

Friction
Now, consider the mass-spring-damper system with a friction force, so that

mẍ + dẋ + kx = −Ff (2.63)

Then the time derivative of the energy function is

V̇ = −Ff ẋ − dẋ2 (2.64)

which has two terms, the power Ff ẋ dissipated by the system by the friction and the
power dẋ2 dissipated in the damper. In its very physical nature, friction work transfers
kinetic energy to heat energy. This means that the friction work will decrease the total
energy of the system, and it follows that

Ff ẋ ≥ 0 (2.65)

and, accordingly,
V̇ ≤ −dẋ2 (2.66)

External force
Suppose that the mass-spring-damper system is actuated with an input force F . Then
the equation of motion is
mẍ + dẋ + kx = F (2.67)
The time derivative of the energy function is found to be

V̇ = F ẋ − dẋ2 (2.68)

Here the term F ẋ is the power that is supplied to the system due to the force F . We see
that if F ẋ < 0, then the energy V will be decreasing.

Example 25 If F is supplied from a controller, we see that negative velocity feedback

F = −Kd ẋ (2.69)

will give
V̇ = − (Kd + d) ẋ2 (2.70)
It is seen that the feedback gain Kd appears as a damping coefficient.

2.3.5 Lyapunov methods

In Lyapunov design for the plant
ẋ = f (x, u) (2.71)
the main idea is to select a suitable energy function V (x), called a Lyapunov function
candidate, which is positive definite in the state vector x in the sense that V (x) = 0 for
x = 0, and V (x) > 0 for x 6= 0. A typical Lyapunov function candidate is

1 T
V (x) = x Px (2.72)
2
50 CHAPTER 2. MODEL ANALYSIS TOOLS

where P is a positive definite and symmetric matrix. Then

λmin (P) T λmax (P) T

x x ≤V (x) ≤ x x (2.73)
2 2
where λmin (P) > 0 is the smallest eigenvalue of P, and λmax (P) is the largest eigenvalue
of P. A control u is sought to ensure that the function V (x) decreases to zero, which
implies that the state vector converges to zero. There is no general method for selecting
the Lyapunov function candidate V (x), but for many applications it is possible to select
V (x) from some type of energy function.

Example 26 We investigate the problem of controlling the position of a mass m with

position x. The mass is actuated by a force u, and the desired position is xd = 0. The
kinetic energy of the mass is T = 12 mẋ2 . We may reason as follows: Suppose that
a spring with spring constant kp was fixed to the mass, and, in addition, that a viscous
damper with damping constant kd was fixed to the mass. Then the system would obviously
be stable. The potential energy of this spring would be 12 kp x2 . Now, if a spring and a
damper will stabilize the mass, why not let the force input u set up the same force as the
spring and the damper? We accordingly select the control to be the PD controller

u = −kp x − kd ẋ (2.74)

and get the closed loop dynamics

mẍ + kd ẋ + kp x = 0. (2.75)

This means that the controller defines a virtual spring and a virtual damper. We define
the Lyapunov function candidate to be the sum of the kinetic energy of the mass and the
potential energy of the virtual spring:
1 1
V = mẋ2 + kp x2 (2.76)
2 2
The time derivative of V along the solutions of the system (2.75) is

V̇ = ẋmẍ + kp xẋ
= −ẋ (kd ẋ + kp x) + kp xẋ
= −kd ẋ2 (2.77)

We see that whenever ẋ 6= 0, then V will decrease, and it can be shown that V will tend
to zero. Then, because m and kp are positive constants, this implies that x and ẋ will
tend to zero.

2.3.6 Contraction
We have introduced energy functions to study the stability of nonlinear system around an
equilibrium point by calculating the time derivative of the energy function for solutions
of the system. A slightly different view is taken in contraction analysis (Hartman 1982,
p. 537), (Lohmiller and Slotine 1998) where the convergence of different solutions to each
other is studied. We will look at two different solutions x1 (t) and x2 (t) for the system

ẋ = f (x)
2.3. ENERGY-BASED METHODS 51

where the initial conditions are x1 (t0 ) = x10 and x2 (t0 ) = x20 . We consider the energy
function
1
(x1 − x2 )T (x1 − x2 )
V =
2
which can be seen as a measure of how far the two solutions are from each other. The
time derivative of V along the solutions of the system is

V̇ = (x1 − x2 )T [f (x1 ) − f (x2 )]

Define the Jacobian matrix

∂f (x)
J(x) =
∂x
and consider the line αx1 (t) + (1 − α) x2 (t), 0 ≤ α ≤ 1 between x1 (t) and x2 (t). On
this line we have
d
f [αx1 + (1 − α) x2 ] = J [αx1 + (1 − α) x2 ] (x1 − x2 )
dα
and we may write
Z 1
f (x1 ) − f (x2 ) = J [αx1 + (1 − α) x2 ] dα(x1 − x2 )
0

Because of this, the time derivative of V can be expressed as

V̇ = −(x1 − x2 )T Q(x1 − x2 ) (2.78)

where Z 1 © ª
1
Q=− J [αx1 + (1 − α) x2 ] + JT [αx1 + (1 − α) x2 ] dα (2.79)
2 0
We see that whenever £ ¡ ¢¤
Re λi J + JT < 0 (2.80)
then Q is positive definite, and it follows that V̇ ≤ 0. This means that the two solutions
x1 (t) and x2 (t) will converge to each other as time goes to infinity. As the two solutions
were selected freely, this implies that any two solutions will converge to each other as
time goes to infinity.

2.3.7 Energy flow in a turbocharged diesel engine

Diesel engines are usually equipped with turbochargers (Heywood 1988), (Kiencke and
Nielsen 2000) as shown in Figure 2.3. A turbocharger has a turbine that is driven by the
exhaust, and, on the same shaft, a compressor that increases the pressure of the inlet air
to the motor. The purpose of this arrangement is to increase the pressure and thereby
increase the density of the air into the cylinder. The benefit of this is that by increasing
the mass of fresh air into the cylinder, it is possible to increase the amount of injected
diesel fuel while still having sufficient oxygen to achieve satisfactory combustion of the
fuel. This increases the energy that can be processed in a fixed cylinder volume.
A diesel engine with a turbocharger is a complicated and nonlinear system, still,
the energy flow is easy to model and important for understanding the dynamics of the
system. The required modeling tools for this is presented in Chapter 12. Energy is
52 CHAPTER 2. MODEL ANALYSIS TOOLS

Diesel
cylinder
Wc Wt

h c C T h t Wastegate
J tc , tc

air intake exhaust

Figure 2.3: Diesel engine with turbocharger

exchanged partly as thermal energy with power P = wh where w is mass flow and h
is specific enthalpy, and partly as rotation power P = τ ω where τ is torque and ω is
angular velocity. The speed ω tc of the turbocharger shaft depends on how much energy
that is absorbed by the turbine compared to how much energy that is used to compress
the air in the compressor. In other words, the time derivative of the kinetic energy of the
turbocharger shaft is equal to the thermal power which is converted to kinetic energy in
the turbine, minus the thermal power that is required in the compressor. This is written
· ¸
d 1
Jtc ω 2tc = wt ∆ht − wc ∆hc (2.81)
dt 2

where Jtc is the moment of inertia of the combined shaft of the turbine and the compres-
sor, wt is the mass flow through the turbine, ∆ht is the reduction in specific enthalpy
over the turbine, wc is the mass flow through the compressor, and ∆hc is the increase in
specific enthalpy over the compressor.
If the energy carried by the exhaust becomes to high, then the turbocompressor may
over-speed. This problem can be avoided by opening a wastegate valve for dumping power
from the system. The effect of this is to reduce wt . Alternatively, inlet guidevanes on
the turbine can be used to control the amount of energy that is absorbed by the turbine.
The effect of this is to reduce the change of specific enthalpy ∆ht in the turbine. During
acceleration it may be desirable to speed up the turbocharger to achieve sufficiently high
pressure of the injected air. This can be done by closing the wastegate, or by adjusting
the inlet guide vanes for a higher ∆ht . The use of inlet guide vanes gives a faster response
than the use of a wastegate, and has become the preferred solution in car engines.

2.4 Passivity
2.4.1 Introduction
The concept of passivity is very useful in control systems analysis and design. Passivity
theory provides a set of analysis tools that can be used for a wide range of physical systems
with commonly used controllers like the PID controller, and, in addition, for nonlinear
2.4. PASSIVITY 53

controllers based on adaptive techniques and backstepping. The main observation in

passivity theory is: If a system can be described as a parallel or feedback interconnection
of passive subsystems, then the total system will be passive, and it will not generate
energy.
One practical application of the theory of passivity is within stability analysis based
on energy-flow considerations for interconnections of physical subsystems. In this setting,
stability can be related to a decreasing total energy of the system. Now, suppose that
each physical subsystem is passive in the sense that it can store and dissipate energy,
but it cannot produce energy. Then it can be concluded that the total energy of the
system will decrease, which under certain assumptions implies that the system is stable.
The theory of passivity also provides a generalization of the pure energy-based analysis of
interconnected physical subsystems to the analysis of interconnections of general dynamic
subsystems, where the storage and dissipation of other functions than energy are studied.
Methods based on passivity can be used for linear time-invariant systems where useful
properties of certain transfer functions can be established from simple energy consider-
ations. This leads to very efficient tools for stability analysis. Moreover, passivity is
very useful for nonlinear systems as an extension of Lyapunov analysis in the sense that
Lyapunov results for an interconnection of subsystems can be inferred from the passivity
properties of the individual subsystems.

2.4.2 Definition
The following definition of passivity will be used:

Definition 1 Consider a system with input u and output y. Suppose that that there is
a constant E0 ≥ 0 so that for all control time histories u and all T ≥ 0 the integral of
y (t) u (t) satisfies
Z T
y (t) u (t) dt ≥ −E0 (2.82)
0
Then the system is said to be passive.

Some remarks can be made to this definition.

1. The definition is based on an input-output description where the input is u and
the output is y. There is no reference to the state of the system.
2. Note that it is the system with a specified input u and a specified output y that is
passive. Passivity cannot be defined without defining input and output.
3. The definition is valid for both linear and nonlinear systems, and for time-varying
and time-invariant systems.
4. If the system is passive with input u and output y, then it will also be passive with
input y and output u.

2.4.3 Examples
Time constant
A time constant given by
ẏ = −ay + u (2.83)
54 CHAPTER 2. MODEL ANALYSIS TOOLS

is passive if a ≥ 0. This is seen by inserting u = ẏ + ay into the integral of yu, which

gives
Z T Z T
yudt = y(ẏ + ay)dt
0 0
Z y(T ) Z T
= ydy + a y 2 dt
y(0) 0
Z T
1 2 1
= y (T ) − y 2 (0) + a y 2 dt (2.84)
2 2 0

The first and last term on the right side are positive. It follows that
Z T
1
yudt ≥ − y 2 (0) (2.85)
0 2

and the system has been shown to be passive with E0 := (1/2)y 2 (0).

Mass, spring and damper

A mass m with position x is connected with a spring and a damper to a fixed point. The
equation of motion is
mẍ + B ẋ + Kx = F (2.86)
where F is the control force acting on the mass. Then the system with input F and
output ẋ is passive. This is seen from
Z T Z T
F ẋdt = (mẍ + B ẋ + Kx) ẋdt
0 0
Z T Z T Z T
= mv̇vdt + Bv 2 dt + Kxẋdt
0 0 0
Z T
1 £ 2 ¤ 1 £ ¤
= m v (T ) − v 2 (0) + B v 2 dt + K x2 (T ) − x2 (0) (2.87)
2 0 2
Z T
≥ −E0 + B v 2 dt (2.88)
0

where
1 1
mv 2 (0) + Kx2 (0)
E0 = (2.89)
2 2
is the initial energy in the form of kinetic and potential energy. It follows that
Z T
F ẋdt ≥ −E0 (2.90)
0

which shows that the system is passive when F is input and ẋ is output. Note that the
product F ẋ between the input and the output is the power supplied to the mass because
of the control input F . Moreover, note that the constant E0 in the passivity inequality
(2.90) is the initial energy in the system. It seems intuitively right to describe this type
of system as passive as the systems has only passive components in the form of a mass,
a spring and a damper. In particular, there is no active element in the system that can
produce energy.
2.4. PASSIVITY 55

Electrical circuit
Consider an electrical circuit with input voltage u, which is the control input, and current
i, which is the system output. The circuit is a serial interconnection of a resistor R, a
capacitor C and an inductance L. The voltage law gives
1 di
u = Ri + q+L (2.91)
C dt
where q is the capacitor charge which satisfies q̇ = i. Then passivity from u to i is shown
by the calculation
Z T Z T Z Z t
1 t di
uidt = R i2 dt + q̇qdt + L i dt
0 0 C 0 0 dt
Z T
1 £ ¤ L£2 ¤
= R i2 dt + q 2 (T ) − q 2 (0) + i (T ) − i2 (0)
0 2C 2
Z T
≥ −E0 + R i2 dt (2.92)
0

where
1 2 L
E0 = q (0) + i2 (0) (2.93)
2C 2
is the initially stored energy in the circuit. Note that the product ui between the input
and the output is the power supplied to the circuit from the control input u. It is also
interesting to note that if the current had been take to be the control variable, and the
voltage had been the measurement, then the system would still be passive.

2.4.4 Energy considerations

We may think of passivity as a property related to balance equations, and in particular
to energy conservation, which we will use to illustrate the meaning of the concept of
passivity. Consider a system with input u and output y. Suppose that the product
RT
u (t) y (t) has the physical dimension of power, and 0 u (t) y (t) dt is the energy that is
supplied to the system due to the control action u. A critical observation is:
RT
• If 0 u (t) y (t) dt ≥ 0 for all control histories u and for all T ≥ 0, then energy is
absorbed by the system, and the system cannot supply any energy to the outside.
In this case the system is passive according to Definition 1.
RT
• If there exists some E0 > 0 so that the integral 0 u (t) y (t) dt ≥ −E0 for all con-
trol histories u and for all T ≥ 0, then the system may supply a limited quantity of
energy to the outside. The energy will typically be energy due to the initial con-
ditions of energy storage elements, which in mechanical systems may be potential
energy in springs and kinetic energy of masses, while in electrical systems it will be
energy stored in capacitors and inductances. According to Definition 1 the system
is passive, which makes sense as the system can only store energy received from
the outside, but it cannot produce energy.
RT
• If it is possible to find a control history u so that the integral 0 u (t) y (t) dt may
tend to −∞ for some T ≥ 0, then this means that the system may supply an
unlimited amount of energy to the outside. This is only possible if there is an
inexhaustible source of energy inside the system. This agrees with the fact that
the system is not passive according to Definition 1.
56 CHAPTER 2. MODEL ANALYSIS TOOLS

2.4.5 Positive real transfer functions

It turns out that a system is passive if and only if the transfer function from input
to output is positive real. This result, which is very useful, will be developed in the
following. First the concept of positive real transfer functions will be defined, and a
special conditions for rational transfer functions is presented. Then the connection to
passivity will be demonstrated. We start by defining positive real transfer functions.

Definition 2 The rational or irrational function H(s) is positive real if

1. H(s) is analytic for all Re [s] > 0.
2. H(s) is real for all positive and real s.
3. Re [H(s)] ≥ 0 for all Re [s] > 0.

It is emphasized that at this stage in the presentation there is no obvious physical

interpretation of this definition. Note that the definition is based on the properties for
the transfer function H(s) for Re [s] > 0, which is to the right of the imaginary axis.

2.4.6 Positive real rational transfer functions

In the case of rational transfer functions it is convenient to investigate the properties
of the transfer function on the imaginary axis by working with H(jω). This makes it
easier to check if a transfer function is positive real, and it leads to a more intuitive
understanding of the positive real property. Now, suppose that the transfer function
H(s) is rational. In this case Statement 1 of Definition 2 implies that there are no poles
to the right of the imaginary axis. Concerning Statement 3, the result can be formulated
on the imaginary axis by observing that the transfer function will be continuous at all
s except at the poles. This means that as long as jω is not a pole of H(s) the transfer
function, then
H(jω) = lim H(σ + jω) (2.94)
σ→0
σ>0

This implies that Re[H(jω)] ≥ 0 as long as jω is not a pole of H(s). These arguments
provide a partial explanation of the following important result:

The rational function H(s) is positive real if and only if

1. All the poles of H(s) have real parts less than or equal to zero.
2. Re H (jω) ≥ 0 for all ω so that jω is not a pole of H(s).
3. If jω 0 is pole in H(s), then it is a simple pole, and

Ress=jω0 [H(s)] = lim (s − jω 0 )H(s) (2.95)

s→jω 0

is real and positive. If H(s) has a pole at infinity, then it is a simple pole, and

H(jω)
R∞ = lim (2.96)
ω→∞ jω
exists, and is real and positive.

The derivation of this result is found in (Anderson and Vongpanitlerd 1973) and
(Lozano, Brogliato, Egeland and Maschke 2000).
2.4. PASSIVITY 57

Example 27 A time constant has transfer function

1
H (s) = (2.97)
1 + Ts
The frequency response is
1 1 − jωT
H (jω) = = 2 (2.98)
1 + jωT 1 + (ωT )
and we see that
1
Re H (jω) = 2 >0 (2.99)
1 + (ωT )
In addition, the only pole has a negative real part, and it follows that the transfer function
of a time constant is positive real.

Example 28 Consider the transfer function

s+c
H(s) = (2.100)
(s + a)(s + b)
where a, b and c are constants greater than zero. Both poles of the transfer function have
negative real parts. Then
jω + c (c + jω)(a − jω)(b − jω)
H(jω) = =
(jω + a)(jω + b) (a2 + ω 2 )(b2 + ω 2 )
abc + ω 2 (a + b − c) + j[ω(ab − ac − bc) − ω 3 ]
= (2.101)
(a2 + ω 2 )(b2 + ω 2 )
We find that if c ≤ a + b, then Re[h2 (jω)] > 0 for all ω, and the transfer function
is positive
p real. If c > a + b, then h2 (s) is not positive real as Re[h2 (jω)] < 0 for
ω > abc/(c − a − b).

Example 29 The transfer function H(s) = Ls where L > 0 has the frequency response
H(jω) = jωL, so that Re[H(jω)] = 0. The transfer function has only one pole, which is
at infinity. As
jωL
R∞ = lim =L (2.102)
ω→∞ jω

is real and positive, it follows that the transfer function is positive real.

Example 30 Consider the transfer function

s
H (s) = , ω0 > 0 (2.103)
s2 + ω 20
All the poles are simple poles on the imaginary axis in s = ±jω 0 . The frequency response
is
jω
H (jω) = 2 (2.104)
ω0 − ω2
so that Re[H (jω)] = 0. The residuals at the poles on the imaginary axis are
s 1
Ress=±jω0 H(s) = Ress=±jω0 = (2.105)
(s + jω 0 ) (s − jω 0 ) 2
The residuals are real and positive. The transfer function is therefore positive real.
58 CHAPTER 2. MODEL ANALYSIS TOOLS

Example 31 Consider the transfer function

s2 + a2
H (s) = , a > 0, ω 0 > 0 (2.106)
s (s2 + ω 20 )

All the poles are simple poles on the imaginary axis in s = 0 and s = ±jω 0 . The
frequency response is
a2 − ω 2
H (jω) = −j (2.107)
ω (ω 20 − ω 2 )
so that Re[H (jω)] = 0. The residuals at the poles on the imaginary axis are

a2 ω 20 − a2
Ress=0 H(s) = , Ress=±jω0 H(s) = (2.108)
ω 20 2ω 20

The residuals are real and positive and the transfer function is positive real if and only
if a < ω 0 .

Example 32 Consider a proper and rational transfer function

(s + z1 ) (s + z2 ) . . .
H(s) = (2.109)
s (s + p1 ) (s + p2 ) . . .

where Re[pi ] > 0 and Re[zi ] > 0. Then, H(s) is positive real if and only if Re[H(jω)] ≥ 0
for all ω 6= 0. This follows from
z1 z2 . . .
Ress=0 H(s) = >0 (2.110)
p1 p2 . . .

and from the observation that the H(s) has one pole in the origin while the remaining
poles have negative real parts.

2.4.7 Positive realness of irrational transfer functions

Introduction
Irrational transfer functions are obtained for systems described by partial differential
equations. Consider a linear system y(s) = H(s)u(s) with a irrational transfer function
H(s). As for rational transfer functions the system with input u and output y is passive if
and only if the transfer function H (s) is positive real (Anderson and Vongpanitlerd 1973).
For irrational transfer functions we have to study the properties of the transfer function
in the right half plane.

Example: Transmission line

To actuate a valve on the seafloor a hydraulic transmission line can be used. This is a
pipe of length L filled with oil. The output side of the transmission line is connected
to the valve at the seafloor, while the pressure is controlled on the input side of the
pipe. Then the control variable is the volumetric flow q1 at the input side, and the
measurement is the pressure p1 on the input side. The system can be described with the
transfer function
p1 sinh s es − e−s
(s) = H1 (s) = tanh s = = s (2.111)
q1 cosh s e + e−s
2.4. PASSIVITY 59

First we check if the transfer function is analytic in the right half plane. We find that

es + e−s = 0 ⇒ e2s = −1
π + 2kπ
⇒ s=j , k = 0, ±1, ±2, (2.112)
2

This means that h(s) is analytic for all Re [s] > 0. It is trivial that H1 (s) is real for all
positive and real s. We then have to establish that Re [H1 (s)] ≥ 0 for all Re [s] > 0 to
show that H1 (s) is positive real. To do this we introduce σ = Re [s] and ω = Im [s] , so
that s = σ + jω, and, using

sinh(σ + jω) = sinh σ cos ω + j cosh σ sin ω (2.113)

cosh(σ + jω) = cosh σ cos ω + j sinh σ sin ω. (2.114)

we find that
sinh σ cosh σ
Re [tanh s] = 2 > 0 for all σ > 0. (2.115)
|cosh s|
We have then showed that H1 (s) = tanh s is a positive real transfer function.
The transfer function from the pressure p1 on the input side to the pressure p2 on
the output side is
p2 1
(s) = H2 (s) = (2.116)
p1 cosh s
This transfer function is not positive real as the real part of the transfer function is
· ¸
1 cosh σ cos ω
Re = 2 (2.117)
cosh s |cosh s|

It follows that for all ω so that cos ω < 0 the real part will be negative for all σ > 0.

2.4.8 Passivity and positive real transfer functions

We have the following result:

A linear time-invariant system with input u and output y described with the transfer
function model y(s) = H(s)u(s) is passive if and only if the transfer function H(s) is
positive real.

To demonstrate that passivity and positive realness are related, we consider the linear
system
y(s) = H(s)u(s) (2.118)

with rational and strictly proper transfer function

(s + z1 ) (s + z2 ) . . . (s + zm )
H(s) = K (2.119)
(s + p1 ) (s + p2 ) . . . (s + pn )

We will now show that passivity is related to the positive realness of the transfer function
H(s).
60 CHAPTER 2. MODEL ANALYSIS TOOLS

2.4.9 No poles on the imaginary axis

First it is assumed that Re[pi ] > 0, which means that the system is stable, and that all
poles are to the left of the imaginary axis. Suppose that the input is
u(t) = U sin ω 0 t (2.120)
Then the output is
y(t) = U |H (jω 0 )| sin [ω 0 t + ∠H (jω 0 )] + yt (t) (2.121)
where yt (t) is the transient part of the output. Then the product yu is found to be
U2 U2
y (t) u(t) = Re H (jω 0 )− |H (jω 0 )| cos [2ω 0 t + ∠H (jω 0 )]+yt (t)U sin ω 0 t (2.122)
2 2
Integration gives
Z T
U 2T U2
y (t) u(t)dt = Re H (jω 0 ) + |H (jω 0 )| sin [2ω 0 t + ∠H (jω 0 )]
0 2 4ω 0
Z T
+ yt (t) u(t)dt (2.123)
0

The second term on the right side will be a sinusoidal signal that is bounded by its ampli-
tude. In the third term on the right side the transient signal yt (t) will tend exponentially
to zero. This leads to
¯ Z T ¯
¯ U2 ¯
¯ ¯
¯+ |H (jω 0 )| sin [2ω 0 t + ∠H (jω 0 )] + yt (t) u(t)dt¯ ≤ E0 (2.124)
¯ 4ω 0 0 ¯

for some constant E0 ≥ 0. This implies that

¯Z ¯
¯ T U 2T ¯
¯ ¯
¯ y (t) u(t)dt − Re H (jω 0 )¯ ≤ E0 (2.125)
¯ 0 2 ¯

for all T ≥ 0. From this result it is seen that the system is passive if and only if
Re H (jω) ≥ 0 for all ω. The if part is obvious. Concerning the only if part, it is
RT
seen that if Re H (jω 0 ) < 0 for some ω 0 , then there is no lower bound on 0 y (t) u(t)dt
as |U 2 T Re H (jω 0 ) /2| can be made arbitrarily large by selecting T sufficiently large.

2.4.10 Single poles at the imaginary axis

Assume that the system has all poles to the left of the imaginary axis except two simple
poles at s = ±ja on the imaginary axis. A partial fraction expansion gives
Ress=±ja H(s) Ress=±ja H(s)
H(s) = + + G(s) (2.126)
s + ja s − ja
sRess=±ja H(s)
= 2 + G(s) (2.127)
s2 + a2
where G(s) is due to the poles to the left of the imaginary axis. Then, if ω 0 6= a, the
results (2.121) and (2.125) are still valid. If ω 0 = a, then
sω 0 Ress=±jω0 H(s)
y(s) = H(s)u(s) = 2 2 + ... (2.128)
(s2 + ω 20 )
2.4. PASSIVITY 61

which corresponds to the time function

y(t) = t sin (ω 0 t) Ress=±jω0 H(s) + . . . (2.129)
and it follows that
Z T Z T
y (t) u(t)dt = tU sin2 (ω 0 t) Ress=±jω0 H(s)dt + . . . (2.130)
0 0

Finally, assume that the system has a simple pole at the origin s = 0. Then
Ress=0 H(s)
H(s) = + G0 (s) (2.131)
s
where G(s) is due to the poles to the left of the imaginary axis. If ω 0 6= 0, the results
(2.121) and (2.125) are still valid. If u(t) = U , then
U Ress=±jω0 H(s)
y(s) = H(s)u(s) = + ... (2.132)
s2
and the time function is
y(t) = tRess=0 H(s) + . . . (2.133)
This gives
Z T Z T
y (t) u(t)dt = t2 Ress=0 H(s)dt + . . . (2.134)
0 0
It may then be concluded that if the system has a simple pole in s = jω 0 at the imaginary
axis, then the system is passive if and only if the residual Ress=jω0 [H(s)] is real and
positive.

2.4.11 Bounded real transfer functions

Definition 3 The rational or irrational function B(s) is bonded real if

1. B(s) is analytic for all Re [s] > 0.
2. |B(s)| ≤ 1 for all positive and real s

The transfer function

H(s) − 1
B(s) = (2.135)
H(s) + 1
is bounded real if and only if H(s) is positive real. This is shown as follows:
Because H(s) is analytic and Re [H(s)] ≥ 0 in Re [s] > 0 it follows that B(s) is
analytic in Re [s] > 0. It is assumed that B(s) 6= ±1 in Re [s] > 0. Solving for H(s) we
get
1 + B(s)
H(s) = (2.136)
1 − B(s)
Then, as H(s) is analytic in Re [s] > 0, it follows that B(s) 6= 1 in Re[s] > 0. Consider
the following calculation:
1 1 1 + B(s) 1 1 + B ∗ (s)
Re H(s) = [H(s) + H ∗ (s)] = + (2.137)
2 2 1 − B(s) 2 1 − B ∗ (s)
1 − B(s)B ∗ (s)
= (2.138)
[1 − B(s)] [1 − B ∗ (s)]
62 CHAPTER 2. MODEL ANALYSIS TOOLS

From this computation it is seen that Re [H(s)] ≥ 0 for all Re [s] > 0 if and only if
|B(s)| ≤ 1 in Re [s] > 0.

Example 33 Suppose that the transfer function H(s) is positive real. Then the inverse
H −1 (s) is positive real. Statement 2 of Definition 2 is trivial in this case. Statements
1 and 3 are shown as follows: It is the possible to conclude from the maximum modulus
theorem that B(s) 6= −1 in Re [s] > 0, and we may express the inverse of H(s) as
1 − B(s)
G(s) = H −1 (s) = (2.139)
1 + B(s)
which is analytic in Re [s] > 0 because B(s) is analytic and B(s) 6= −1 in this region.
This proves Statement 1 for G(s). Finally, statement 3 for G(s) is verified from
1 1 − B(s) 1 1 − B ∗ (s)
Re G(s) = +
2 1 + B(s) 2 1 + B ∗ (s)
1 − B(s)B ∗ (s)
= (2.140)
[1 + B(s)] [1 + B ∗ (s)]

2.4.12 Passivity of PID controllers

A PID controller
1 + Ti s 1 + Td s
Hr (s) = K (2.141)
Ti s 1 + αTd s
where K > 0 and 0 ≤ α < 1 has phase
π
∠Hr (jω) = − + arctan Ti ω + arctan Td ω − arctan αTd ω. (2.142)
2
From this equation it is seen that the phase must satisfy
π π
− ≤ ∠Hr (jω) ≤ . (2.143)
2 2
It follows that Re [H(jω)] ≥ 0 for all ω 6= 0. The transfer function has no poles to the
right of the imaginary axis, and one single pole at the imaginary axis at s = 0. The
residual of this pole is found to be
K
Ress=0 Hr (s) = lim [sHr (s)] = (2.144)
s→0 Ti
which is real and positive, and it follows that a PID controller is positive real. This
implies the following result

A PID controller u(s) = Hr (s)e(s) is a passive system with input e and output u, and
the transfer function Hr (s) is positive real.

2.4.13 Closed loop stability of positive real systems

Stability properties can easily be established from passivity arguments for a feedback
interconnection of passive systems. Suppose that the system with input u and output y
is passive, and that it is given by

y(s) = H(s)u(s) (2.145)

2.4. PASSIVITY 63

The transfer function H(s) is then positive real due to the passivity of the system. The
input u is generated by the system

u(s) = G(s) [yd (s) − y(s)] (2.146)

where the transfer function G(s) is supposed to be positive real. Then the loop transfer
function is L(s) = G(s)H(s). Note the positive realness implies that two transfer func-
tions G(s) and H(s) will have no poles to the right of the imaginary axis. Moreover, the
magnitude of the phase of the of G(jω) and H(jω) will be less than or equal to 90◦ . This
implies that L(jω) has phase that is greater than -180◦ . This means that the system is
at least marginally stable.

2.4.14 Storage function formulation

Passivity can be described using storage functions which are closely related to energy
functions and Lyapunov functions. In the passivity setting systems can be interconnected
in parallel and feedback interconnections, and the resulting system can be analyzed using
passivity theory or Lyapunov theory.
We consider the state space model

ẋ = f (x, u) (2.147)
y = h (x) (2.148)

Suppose that there is a storage function V (x) ≥ 0 and a dissipation function g(x) ≥ 0
so that the time derivative of V for solutions of the system satisfies
∂V
V̇ = f (x, u) = uT y − g(x) (2.149)
∂x
for all control inputs u. Then the system with input u and output y is said to be passive.

The result follows from the calculation

Z T Z T
yT (t) u (t) dt = V (T ) − V (0) + g [x(t)] dt
0 0
≥ −V (0) (2.150)

Example 34 We consider again the mass-spring-damper system with input force F .

The model is
mẍ + dẋ + kx = F (2.151)
The total energy
1 1
V = mẋ2 + kx2 (2.152)
2 2
is a candidate for being a storage function. The time derivative of the energy function is
found to be
V̇ = F ẋ − dẋ2 (2.153)
where F ẋ is the power that is supplied to the system due to the force F . We see that
if the input is u = F and the output is selected to be y = ẋ then V̇ = yu − dẋ2 , which
means that the system with input F and output ẋ is passive.
64 CHAPTER 2. MODEL ANALYSIS TOOLS

Remark 1 Actually, it is sufficient that

Z T
g [x(t)] dt ≥ −Eg for all T ≥ 0 (2.154)
0

for some constant Eg ≥ 0.

2.4.15 Interconnections of passive systems

u1 y1 u u1 y1 y
System 1 System 1
u y

u2 y2 y2 u2
System 2 System 2

Figure 2.4: Parallel interconnection and feedback interconnection of two passive systems.

We now consider two systems

ẋ1 = f1 (x1 , u1 ), y1 = h1 (x1 ) (2.155)

ẋ2 = f2 (x2 , u2 ), y2 = h2 (x2 ) (2.156)

which are passive in the sense that there are functions V1 (x1 ) ≥ 0 and V2 (x2 ) ≥ 0 so
that
∂V1
V̇1 = f1 (x1 , u1 ) ≤ uT1 y1 − g1 (x1 ) (2.157)
∂x1
∂V2
V̇2 = f2 (x2 , u2 ) ≤ uT2 y2 − g2 (x2 ) (2.158)
∂x2
where g1 (x1 ) ≥ −Eg1 and g2 (x2 ) ≥ −Eg2 . We will now show that the parallel intercon-
nection and the feedback interconnection shown in Figure 2.4 are passive.
A parallel interconnection

u1 = u2 = u, y = y1 + y2 (2.159)

implies that the function V defined by

V := V1 + V2 ≥ 0 (2.160)

satisfies

V̇ = V̇1 + V̇2
= uT1 y1 − g1 (x1 ) + uT2 y2 − g2 (x2 )
= uT y−g1 (x1 ) − g2 (x2 ) (2.161)

which shows that the parallel interconnection with input u and output y is passive.
A feedback interconnection

y1 = u2 = y, u1 = u − y2 (2.162)
2.4. PASSIVITY 65

implies that

V̇ = V̇1 + V̇2
= uT1 y1 − g1 (x1 ) + uT2 y2 − g2 (x2 )
= uT y−g1 (x1 ) − g2 (x2 ) (2.163)

so that also the feedback interconnection with input u and output y is passive.

2.4.16 Storage function for PID controller

A PID controller
u(s) = Hpid (s)e(s) (2.164)
where e(s) is the input to the controller, u(s) is the output from the controller, and

1 + Ti s
Hpid (s) = K (1 + Td s)
Ts
µ i ¶
Td 1
= K 1+ + Td s + (2.165)
Ti Ti s

can be written in state-space form as

e
ż = (2.166)
Ti
·µ ¶ ¸
Td
u = K 1+ e + Td ė + z (2.167)
Ti

Consider the nonnegative function

1 1
Vpid = KTi z 2 + KTd e2 (2.168)
2 2
The time derivative of V along the solutions of the PID controller dynamics is

V̇pid = zKTi ż + eKTd ė

= eK (z + Td ė)
µ ¶
Td
= eu − K 1 + e2 (2.169)
Ti

It follows that the PID controller is passive.

2.4.17 Passive plant with PID controller

We consider a passive plant with a PID controller as shown in Figure 2.5. We assume
that the passive plant has input u and output y which satisfies

V̇p = yu − gp (2.170)

where Vp ≥ 0 and gp ≥ 0. Consider the nonnegative function

Vcl = Vp + Vpid (2.171)

66 CHAPTER 2. MODEL ANALYSIS TOOLS

yd e u Passive y
PID
system

Figure 2.5: Passive plant with PID controller.

The time derivative of Vcl along the solutions of the closed-loop system is
µ ¶
Td
V̇cl = yu − gp + eu − K 1 + e2
Ti
µ ¶
Td
= yd u − gp − K 1 + e2 (2.172)
Ti

where it is used that e = yd − y. We see that the closed-loop system with input yd and
output y is passive. In particular we have that if yd = 0, then
µ ¶
Td
V̇cl = −gp − K 1 + e2 ≤ 0 (2.173)
Ti

2.4.18 Example: Control of mass-spring-damper system

For the mass-spring-damper system with input force F we found that

V̇ = F ẋ − dẋ2 (2.174)

Suppose that the input F is generated by the P controller F = −K ẋ. Then, with the
same storage function we have

V̇ = −K ẋ2 − dẋ2 = − (Kp − d) ẋ2 ≤ 0 (2.175)

Suppose that the PID controller

F (s) = Hpid (s)e(s) (2.176)

is used where e(t) = ẋd (t) − ẋ(t). Then

V̇cl = V̇ + V̇pid (2.177)

µ ¶
Td
= ẋd F − dẋ2 − K 1 + e2 (2.178)
Ti

In particular, we see that if ẋd = 0, then V̇cl ≤ 0. Note that the controller is a PID
controller from velocity, which corresponds to a PD2 controller from position.

2.4.19 Example: Active vibration damping

Spacecraft and large space installations are design with lightweight structures, and the
lack of atmosphere gives little mechanical damping of vibrations. Even the effect of
temperature changes when a satellite passes from the shadow of the earth into the sunlight
2.4. PASSIVITY 67

may be sufficient to cause unacceptable vibrations in the structure. Because of this

the use of feedback for active vibration damping is important (Kelkar and Joshi 1996).
Vibration models are usually in the form

Mq̈ + D (q, q̇) q̇ + Kq = Bf (2.179)

where q is the vector of generalized coordinates, which are the elastic deformations, M
is a symmetric mass matrix, K is a symmetric stiffness matrix, D (q, q̇) q̇ is a damping
term and f is the input control force. Detailed vibration models of high accuracy can
be derived from the method of finite elements (Bathe 1996). To give the reader an idea
richness of the structural properties of this model we briefly mention some issues that will
be addressed later in the book: The total energy of the vibration system is V = K + U ,
where
1
K = q̇T Mq̇ ≥ 0 (2.180)
2
is the kinetic energy, and
1
U = qT Kq ≥ 0 (2.181)
2
is the potential energy. The damping term is an energy dissipation term related to
friction, and will always satisfy

q̇T D (q, q̇) q̇ ≥ 0. (2.182)

The time derivative of the total energy along the solutions of the system is
d T
V = q̇T Mq̈ + q̇T Kq = q̇T Bf − q̇ Dq̇. (2.183)
dt
If the input force is set to the P controller f = −kBT q̇, which is a velocity feedback, then
the time derivative of the energy is
d ¡ ¢
V = −q̇T kBBT + D q̇ ≤ 0 (2.184)
dt
It is interesting to note that with this velocity feedback the energy will decrease whenever
q̇ 6= 0. It should be clear from this discussion that the vibration model reflects important
physical properties related to energy that may be important in controller design. These
properties may be obscured if the model is reformulated in state space. Therefore, when
energy-based methods are used, the model is usually kept in the second-order form.

2.4.20 Passive electrical one-port

A passive electrical one-port is a circuit with one port and passive components in the
form of resistors, capacitors and inductors. Resistors are elements that dissipate energy,
while capacitors and inductors are elements that store energy. There are no elements
that generate energy. The total energy stored in the circuit is denoted V (t). The stored
energy cannot be negative, so we can assume that V ≥ 0. The port voltage is denoted
u(t) and the current into the port is denoted i(t). The power flowing into the one-port
is P (t) = u(t)i(t), while the power dissipated in the resistors is Pr ≥ 0.
The time derivative of the energy V stored in the circuit will be the power ui supplied
at the port minus the power loss Pr in the circuit. This is written

V̇ = ui − Pr (2.185)
68 CHAPTER 2. MODEL ANALYSIS TOOLS

Integration of this equation gives

Z T Z T
i (t) u (t) dt = V (T ) − V (0) + Pr (t)dt (2.186)
0 0

Here V (T ) ≥ 0 and Pr (t) ≥ 0, and we find that

Z T
i (t) u (t) dt ≥ −V (0) (2.187)
0

This means that for a passive electrical one-port the energy that can be extracted from
the circuit over the terminals is less or equal to the energy V (0) that is initially stored
in the circuit. Note that (2.187) has the form of a passivity inequality, and that if u is
input and i is output, then the system is passive. This implies that the driving point
impedance Z(s) = u(s)/i(s) is positive real. On background of this we may conclude
that

The driving point impedance of a passive electrical one-port is positive real.

Example 35 From the passivity inequality (2.187) it is seen that if the current is taken
as input and the voltage is considered to be the output, then the system will still be passive,
which means that the driving point admittance Y (s) = i(s)/u(s) is passive.

Example 36 To illustrate this we consider a passive electrical one-port which is a par-

allel interconnection of a resistor and a capacitor. The current is given by
1
i= u + C u̇ (2.188)
R
The total energy stored in the circuit is the energy V = (1/2)Cu2 stored by the capacitor.
The time derivative of the energy is
1 2
V̇ = Cv v̇ = iu − u (2.189)
R
where the loss term is due to the energy dissipation in the resistor. Figure 2.6.

i
+
u R C
-

Figure 2.6: Passive electrical one-port.

2.4.21 Electrical analog of PID controller

A PID controller from current to voltage of a one-port is given by
µ ¶
Td 1
u (s) = −K 1 + + Td s + [id (s) − i (s)] (2.190)
Ti Ti s
2.4. PASSIVITY 69

id R L C u

Figure 2.7: Electrical analog of PID controller for an electrical one-port where the current
i is controlled with the voltage u.

where id is the desired current. This is an electrical one-port where a current source with
current id is placed in in parallel to a series connection with a resistor R = K(1 + Td /Ti ),
a capacitor C = Ti /K, and an inductor L = Td K (Figure 2.7). A PI controller is obtained
by setting Td to zero, and in that case there is no inductor in the one-port. In the case
that id = 0, the PID controller is a passive one-port with only passive elements.

2.4.22 Passive electrical two-port

An electrical two-port has two ports, one input port with voltage u1 and current i1 , and
one output port with voltage u2 and current i2 . The power flowing into the two-port is
P (t) = u1 (t) i1 (t) + u2 (t) i2 (t) (2.191)
As for the passive one-port, the energy that can be extracted from a passive n-port
cannot be larger than the energy Vn (0) that is initially stored in the capacitors and the
inductors. This gives
Z T
[u1 (t) i1 (t) + u2 (t) i2 (t)] dt ≥ −V2 (0) (2.192)
0

2.4.23 Termination of electrical two-port

An electrical two-port is said to be terminated if the output port is connected to a
one-port so that
u2 = u, i2 = −i (2.193)
Then the two-port with the one-port termination becomes a one-port with port variables
u1 and i1 .
Suppose that both the two-port and the one-port termination are passive. Then the
following energy equations are valid
Z T
[u1 (t) i1 (t) + u2 (t) i2 (t)] dt ≥ −V2 (0) (2.194)
0
Z T
i (t) u (t) dt ≥ −V1 (0) (2.195)
0

If we add these equations and insert the connection equations (2.193) we get
Z T
i1 (t) u1 (t) dt ≥ − [V1 (0) + V2 (0)] (2.196)
0
70 CHAPTER 2. MODEL ANALYSIS TOOLS

The physical interpretation is that the energy that can be extracted from the combined
circuits is equal to the sum of the initially stored energy in the two circuits. This result
shows clearly shows that if a passive two-port is terminated with a passive-one-port, then
the resulting one-port with port variables u1 and i1 is passive.

2.4.24 Passive electrical n-ports

An electrical n-port has n ports with voltage uk and current ik . The power flowing into
the n-port is
n
X
P (t) = uk (t) ik (t) = iT (t) u (t) (2.197)
k=1

where u= (u1 , . . . , un ) and i = (i1 , . . . , in )T . As for the passive one-port, the energy
T

that can be extracted from a passive n-port cannot be larger than the energy En that is
initially stored in the capacitors and the inductors
Z T
iT (t) u (t) dt ≥ −Vn (0) (2.198)
0

A general n-port with effort vector e and flow vector f is passive if the energy that can be
extracted from the n-port is limited by the energy that is initially stored in the n-port,
that is, if
Z T
f T (t) e (t) dt ≥ −E0 (2.199)
0
As in the electrical case this can be expressed in terms of conditions on the impedance
Z (s) for a one-port.

2.4.25 Example: Telemanipulation

In a telemanipulation system a manipulator is remotely controlled by a human operator.
Early telemanipulation systems were master-slave systems where the operator moved a
handle fixed to a master manipulator that was connected to an identical slave manipulator
through mechanical linkages. This was used to protect the operator from hazardous
environments due to radioactivity or danger of contamination from biological samples.
The operator would then typically watch the slave manipulator through a window, and as
there was a direct mechanical coupling between the master and the slave manipulators
the operator would feel contact forces that resulted when the slave came into contact
with a sample or hit against a table. This feature is called force reflection. At a later
stage such systems were equipped with computer control. This was done to make it
possible to perform telemanipulation in hostile environments for operations in space,
and for underwater operations at great depths. A more recent activity is telesurgery.
In telemanipulation systems with computer control the motion of the master is mea-
sured by sensors, and the position and velocity commands are transmitted to the slave
through a computer, and the slave is driven by DC motors. The opens up for advanced
control functions, but it turns out that the system becomes unstable if force reflection is
used with time delays of 40 ms or more. This problem was analyzed in an energy flow
setting in (Anderson and Moore 1989) and (Niemeyer and Slotine 1991), and it was pro-
posed to transmit wave variables to obtain a closed loop system where the transmission
between the master and the slave could be described in terms of passive two-ports. The
main idea of the solution is presented in the following.
2.4. PASSIVITY 71

A human operator that moves a telemanipulation system using a master-slave con-

figuration will expect that the master-slave system in itself will not generate energy that
is transferred to the handle that the operator is using to move the master. If the system
were to generate energy, then the operator might get the impression that the telemanip-
ulation system had a mind of its own, and the operator might have to struggle against
movements that are generated by the telemanipulation system. The operator might even
get injured by the master. This means that the telemanipulation system as felt by the
operator may store energy and dissipate energy, but it may not generate energy. This
means that the handle connected to the master of the telemanipulation system should
appear to the operator as a port to a passive mechanical system where the velocities of
the handle are the flow variables, and the forces on the handle are the effort variables. If
the master and slave are connected with mechanical linkages, then the system will be a
passive mechanical system. However, when computer control is added, then the control
algorithms must be selected with care so that the system still appears as a passive system
to the operator.
What the operator will expect is that the telemanipulation system appears as a passive
two-port that transfers the velocity commands from the operator to the slave, and that
returns the force from the slave to the operator. A mathematical formulation of this in
one dimension is that the master is a two-port
mm v̇m = Fh − Fm (2.200)

with effort Fh and flow vm on the input port, and effort Fm and flow vm on the output
port. The slave is a passive two-port

ms v̇s = Fs − Fe (2.201)

with effort Fs and flow vs is the input port and effort Fe and vs at the output port. The
key to a satisfactory system is to have a passive two-port with effort Fm and flow vm on
the input port, and effort Fs and flow vs is the output port to connect the output port
of the master to the input port of the slave. The total energy E of the system will then
be Z Z
T T
E(T ) = E(0) + Fh (t)vm (t)dt + Fe (t)vs (t)dt (2.202)
0 0
It is reasonable to assume that the total energy is positive, that is, E ≥ 0, which implies
that Z T Z T
Fh (t)vm (t)dt ≥ −E(0) − Fe (t)vs (t)dt (2.203)
0 0
The physical interpretation of this is that the energy that is returned to the operator
through the handle on the master is less that the energy initially stored in the system
plus the energy supplied to the slave from the environment.
First, suppose that the master is directly connected to the slave through a rigid
interconnection so that
Fs = Fm and vs = vm (2.204)
Then the interconnection between the master and the slave is certainly a passive two-port,
and the desired passivity of the system is ensured. Moreover, we see that

(mm + ms )v̇m = Fh − Fe (2.205)

which means that the operator experiences the environmental force Fe on the slave, and
moves the combined inertia of master and slave.
72 CHAPTER 2. MODEL ANALYSIS TOOLS

Next, consider the case where the slave is driven by DC motors, and that the com-
manded velocity vm from the master is measured by a sensor and transmitted electroni-
cally without any time delay. Then a possible solution is to control the slave with a PD
controller with desired velocity vd and desired position xd given by

vd (t) = vm (t) and xd (t) = xm (t) (2.206)

The slave with PD controller is then given by the passive two-port

ms v̇s = Fs − Fe (2.207)
Fs = Ks (xm − xs ) + Ds (vm − vs ) (2.208)

with effort Fs and flow vm at the input port and effort Fe and flow vs at the output port.
Force reflection to the master is achieved by setting up the force Fm (t) = Fs (t) in the
master, which gives the following two-port for the master:

mm v̇m = Fh − Fs (2.209)

The resulting telemanipulation system is passive with a mechanical analog as shown in

Figure 2.8 where the transmission between the master and the slave is a spring with
stiffness Ks in parallel with a damper with coefficient Ds .

1 Fh

mm ms Fe
vm vs
2 Fh
Ks
mm ms Fe
vm vs
Ds
3 Fh
Ks
mm z 0 , ms Fe
vm vs
Ds

Figure 2.8: Mechanical analogs of telemanipulation systems with force reflection. The
master is represented by a mass mm that is moved with a force Fh from the human
operator. The slave is represented by a mass ms which is exposed to a force Fe from
the environment. In 1) there is a direct mechanical coupling between the master and the
salve. In 2) the slave is controlled by a DC motor with PD control, and there is no time
delay in the signal transmission. In 3) the slave is controlled by a DC motor with PD
control, and the signals are transmitted with time delay using wave variables.
2.4. PASSIVITY 73

Finally, suppose that there is a time delay τ in the electronic transmission between
master and slave. Early attempts involved using the same solution as presented above,
but with transmission given by

vd (t) = vm (t − τ ) and Fm (t) = Fs (t − τ ) (2.210)

The resulting system is not passive, as there is no passive mechanical two-port that delays
the velocity in the forward direction and that delays the force in the opposite direction.
Stability problems where experienced for such solutions already at time delays of 40 ms.
A different solution must therefore be sought. An elegant solution to this problem is to
use a lossless transmission line to interconnect master and slave. The key to this solution
is that a lossless transmission line transmits the wave variables with a time delay that is
the transport time τ of the transmission line. We therefore introduce the wave variables

am = Fm + z0 vm and bm = Fm − z0 vm (2.211)

for the output port of the master, and the wave variables

as = Fs − z0 vd and bs = Fs + z0 vd (2.212)

for the input port of the slave where z0 is the characteristic impedance of the transmission
line. The wave variables are transmitted according to

bs (t) = am (t − τ ) and bm (t) = as (t − τ ) (2.213)

In terms of forces and velocities this gives the following description of the passive two-port
1
vd (t) = vm (t − τ ) + [Fm (t − τ ) − Fs (t)] (2.214)
z0
Fm (t) = Fs (t − τ ) + z0 [vm (t) − v0 (t − τ )] (2.215)
with effort Fm and vm at the input and effort Fs and flow vs at the output port. The
slave may then be controlled with the PD controller

Fs = Ks (xd − xs ) + Ds (vd − vs ) (2.216)

as in the case of no time delay. The mechanical analog is as when there is no time delay,
but with a transmission corresponding to a lossless hydraulic transmission line with a
compressible fluid shown in Figure 2.8.

2.4.26 Passivity and gain

Consider a system with input u and output y. Define the variable

r = u + λy (2.217)

Then Z Z Z Z
T T T T
r2 dt = u2 dt + 2λ uydt + λ2 y 2 dt (2.218)
0 0 0 0
From this equation it is seen that
Z T
uydt ≥ −E0 (2.219)
0
74 CHAPTER 2. MODEL ANALYSIS TOOLS

is equivalent for all λ > 0 to

Z T Z T Z T Z T
2 2 2 2
r dt + 2λE0 ≥ u dt and r dt + 2λE0 ≥ λ y 2 dt (2.220)
0 0 0 0

This shows that passivity of the system with input u and output y is equivalent to a small
gain condition in the L2 norm (Khalil 1996) from r to u, and from r to y. A related
result is used in semi-group theory (Pazy 1983, p. 14).

Example 37 This result was used in attitude control in (Egeland and Godhavn 1994)
where
r = ω + λ² (2.221)
was used. A controller was designed so that r ∈ L2 , and then (2.220) was used to show
that the passivity of the system with input ω and output ² implied that the mapping from
r to ω, and the mapping from r to ² were L2 stable.

Example 38 In the adaptive tracking controller in (Slotine and Li 1988), stability in

the variable
r = q̇ + λq (2.222)
was used to establish convergence in q and q̇. In (Kelly, Carelli and Ortega 1989) the
same controller was analyzed, and it was shown that r ∈ L2 , and linear theory was used
to show that this implied convergence in q and q̇.

2.5 Uncertainty in modeling

2.5.1 General state space models
In a general state space model

ẋ = f (x, u, t) (2.223)
y = h (x, t) (2.224)

or, in the case of linear models,

ẋ = Ax + Bu (2.225)
y = Cx + Du (2.226)

there may be uncertainties related to

• Parameter values

• Model structure

• System order

There are techniques for describing such uncertainties (Skogestad and Postlethwaite
1996), which may be based on additive uncertainties. Such additive terms may be related
to model properties.
2.5. UNCERTAINTY IN MODELING 75

2.5.2 Exact kinematic models

Some types of models or parts of models are exact. In particular this applies to kinemat-
ics, which is the geometric description of motion. Kinematics, which is discussed in great
detail in Chapter 6, play an important role in the models describing the motion of rigid
bodies such as planes, ships, robots, cars and mechanisms. The main role of kinematics
in this context is in the kinematic differential equations that give the time derivative
of the configuration as a function of velocity and angular velocity. It may be of great
importance that the kinematic part of a model is exact because the kinematics are often
nonlinear, and even complicated. Therefore it may be important to keep in mind which
part of the model that is kinematic and therefore exact.

Example 39 In inertial navigation the position of a plane is calculated from inertial

sensors (Titterton and Weston 1997). The inertial sensors are gyroscopes, which measure
the angular velocity of the plane, and accelerometers, which measure the acceleration. The
velocity and position are calculated by integrating the acceleration measurements, while
the angular velocity measurements are used to calculate the direction of the axes of the
accelerometers. The direction of the accelerometers are usually described in terms of
Euler parameters, also known as quaternions. When the angular velocity vector ω b of
the plane is given, the Euler parameters may be calculated by numerical integration of
the kinematic differential equations
1
η̇ = − ²T ω b (2.227)
2
1¡ ¢
²̇ = ηI + ²× ω b (2.228)
2
where ²× ω b denotes the vector cross product between ² and ω b . The rotation matrix is
Rη, = I + 2η²× + 2²× ²× , and the acceleration in a star-fixed coordinate frame is
¡ ¢
a = I + 2η²× + 2²× ²× ab (2.229)

where ab are the measured accelerations. We see that the equations for calculating the
acceleration a are relatively complicated and highly nonlinear. However, it is interesting
to note that there are no approximations or uncertainty involved in the development of
this model.

In addition, kinematics play an important role in fluid mechanics, where the concepts
of divergence and curl are kinematic, and moreover, the Laplace equation

∇2 φ = 0 (2.230)

in potential flow is purely kinematic. Also Reynolds’ transport theorem is a kinematic

result.

2.5.3 Balance equations

Balance equations are based on the physical phenomenon that certain properties like
mass, momentum and energy are conserved. Therefore, the balance equations are exact
models from the outset. As an example of this, the Cauchy equation of motion for a fluid
is exact. This equation, which is discussed in detail in Section 11.2.1, is written
Dvi
ρ = Tji,j + ρfi (2.231)
Dt
76 CHAPTER 2. MODEL ANALYSIS TOOLS

where ρ is density, vi is the velocity in direction i, Tji,j is the gradient of the stress tensor,
and fi is the body force. The modeling assumptions enter in the constitutive equations
where the assumptions on the functional dependence of the stress tensor on fluid motion
enters. The mass balance is important in many physical systems, and perhaps even more
important is the energy balance. The model
d
[total energy in a V ] = [energy flow into V ]
dt
− [energy flow out of V ]
+ [energy generated in V ] (2.232)

where V is a fixed volume, is exact, and is not influenced by changes in parameter

values. Such insight may be especially interesting in relation to energy-based methods
for controller design.

2.5.4 Passivity
In several important applications the plant may have passivity as a physical property
for a given combination of control variable and output variable. This means that the
plant will be passive when plant parameters undergo large variations in numerical values,
and even under changes in model order. Then with a properly selected PID controller
the closed loop system will always be stable although there may by a high degree of
uncertainty in numerical parameters and detailed modeling.

Example 40 A DC motor is used to control a joint in a robot arm (Slotine 1991),

(Lozano et al. 2000). The DC motor is current controlled, and delivers a specified torque
u. The robot arm has several mechanical resonances, and it may pick up loads of unknown
inertia. Moreover, the robot may come into contact with a rigid environment. However,
in spite of all the uncertainty, the arm seen from the motor will be passive because
d
[total energy in robot] = q̇u − [energy dissipated by friction] (2.233)
dt
where q̇ is the shaft velocity of the robot joint, and u is the torque from the DC motor.
Therefore, the closed loop system is stable as long as the controller is passive. This is
the case if a PI controller from q̇ is used, which is the same as a PD controller from the
motor angle q.
 Part II

Motors and actuators

77
Chapter 3

Electromechanical systems

3.1 Introduction
This chapter deals with mathematical models of electrical motors, and models of elec-
tromechanical sensors and actuators. These electromechanical systems are based on
energy conversion between electrical energy and mechanical energy due to capacitive and
inductive effects. This type of electromechanical systems are important, as they are vital
components in most control systems. Special attention is given to the DC motor with
constant field, which is a basic building block in many control systems. This motor is
described by a simple model, and it is possible to control the motor torque directly. Be-
cause of its importance and simplicity the chapter starts with the model of a DC motor,
and presents typical load configurations for the DC motor. Then selected topics from the
general theory of electromechanical energy conversion is presented with emphasis on en-
ergy functions. This provides us with the necessary background to derive more advanced
models of electrical motors. This includes the model of a DC motor with externally
controlled field, the model of a general AC motor, and models for induction motors.

3.2 Electrical motors

3.2.1 Introduction
An electrical motor with rotary motion has a stationary part called the stator. The
rotary part of the motor is called the rotor. The rotor is fixed to the motor shaft
which drives the load. The motion of the rotor is due to the motor torque which is
set up by electromagnetic Lorentz forces acting on the rotor. There are many different
ways of setting up an appropriate Lorentz force, and electrical motors are characterized
depending on how this is done. Electrical motors are divided into DC motors and AC
motors. DC motors are well suited for control applications, as the torque of the motor
can be accurately controlled. The recent development in power electronics, however, has
made it possible to control the torque also for AC motors, and, consequently, AC motors
are now used for accurate control. A basic reference on electrical motors is (Fitzgerald,
Kingsley and Umans 1983), while a more advanced textbook including control methods
is (Leonhard 1996).

79
80 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

3.2.2 Basic equations

A rotary motor has a motor shaft that rotates with angular velocity ω m , and it has some
device for setting up a motor torque T so that the motor shaft has the following equation
of motion:
Jm ω̇ m = T − TL (3.1)
Here TL is the load torque acting on the shaft. The mechanical power delivered from the
motor to the shaft is
Pm = T ω m (3.2)
while the mechanical power delivered to the load is

PL = TL ω m (3.3)

The motor shaft dynamics can be described as a two-port with effort T and flow ω m at
the input port, and effort TL and flow ω m at the output port. Different types of motors
are characterized according to how the motor torque T is generated. In electrical motors
the torque is due to electromagnetic forces, in a hydraulic motor of the hydrostatic type
it is due to the pressure force from a pressurized fluid, while in a turbine the torque is
set up by the forces that result from the change of momentum in the flowing fluid.
The speed of a motor is commonly given in revolutions per minute (rev/min). The
relation to the SI unit rad/s is
rev 2π rad rad
1 = = 0.105 (3.4)
min 60 s s

3.2.3 Gear model

 in ,T in  out ,T out

1:n

Figure 3.1: Reduction gear

An electrical motor will typically have a speed range from zero up to about 3000
rev/min. Specially designed electrical motors may run up to 12000 rev/min. In compar-
ison to this, car engines run from 800—6000 rev/min. For many applications the required
speed range of the load is significantly less than the speed range of the motor, and a re-
duction gear must be used. This gives a lower speed of the load, and, more importantly,
it gives a higher torque.
A reduction gear with gear ratio n (Figure 3.1) is described by

ω out = nω in (3.5)

where ω in is the angular velocity of the shaft on the input side of the gear, and ω out is
the angular velocity of the shaft on the output side of the gear. For a reduction gear
n < 1, and a gear is said to have a gear ratio of, say, 10 if n = 1/10.
3.2. ELECTRICAL MOTORS 81

The relation between the input torque Tin and the output torque Tout is found by
comparing power in and power out for the gear. Suppose that the gear is lossless. Then
power in is equal to power out, that is,

Tin ω in = Tout ω out (3.6)

Inserting the expression for ω out we find that

1
Tout = Tin (3.7)
n
This means that a reduction gear reduces the speed by a factor n, while it amplifies the
torque by a factor 1/n.

A gear with gear ratio n may be described as a two-port

ω out = nω in (3.8)
1
Tout = Tin (3.9)
n
with variables Tin and ω in at the input port, and variables Tout and ω out at the output
port.

3.2.4 Motor and gear

 m ,T L L

Jm JL Te

1 :n

Figure 3.2: Motor and gear.

Consider a motor with equation of motion

Jm ω̇ m = T − TL (3.10)

that drives a load over a reduction gear with gear ratio n (Figure 3.2). Then the load
has a shaft speed ω L = nω m , and is driven by the output torque of the gear, which is
TL /n. The inertia of the load is JL , and it is assumed that an external torque Te acts on
the load. Then the equation of motion for the load is
1
JL ω̇ L = TL − Te (3.11)
n
If the load equation (3.11) is multiplied by n and added to the equation of motion of
the motor (3.10), then the result is the equation of motion for the system referred to the
motor side. Alternatively, the motor equation (3.10) can be divided by n and added to
the load equation (3.11). This will give the equation of motion of the system referred to
the load side. To sum up:
82 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

The equation of motion for motor, gear and load referred to the motor side is

(Jm + n2 JL )ω̇ m = T − nTe (3.12)

The equation of motion for motor, gear and load referred to the load side is
µ ¶
1 1
2
Jm + J L ω̇ L = T − Te (3.13)
n n

3.2.5 Transformation of rotation to translation

 m ,TL

r
Jm Fe
m
v

Figure 3.3: Transmission from rotation to translation.

Rotational motion of a shaft can be transformed to translational motion and vice

versa by mounting a wheel that rolls on a surface as shown in Figure 3.3. This type
of transmission is commonly seen in rack-and-pinion drives, friction gears, pulleys, and
between car wheels and the road. Suppose that the wheel has radius r, shaft speed ω m ,
and torque TL . Then the translational velocity will be v = rω m . Denote the force acting
from the wheel on the translating part by F . Then the input power will be ω m TL and the
output power will be vF . The gear does not store energy, and it follows that F = TL /r.
This shows that:

A rotation to translation transmission can be described by the two-port

v = rω m (3.14)
1
F = TL (3.15)
r
with variables TL and ω m at the input port, and variables F and v at the output port.

Consider a motor which drives a mass m in translational motion over a wheel with
radius r. The load is assumed to have equation of motion

mv̇ = F − Fe (3.16)

where Fe is an external force acting on the load. A motor described by

Jm ω̇ m = T − TL (3.17)

is used. By combining these two equations the following result is found:

3.2. ELECTRICAL MOTORS 83

The equation of motion for motor and load referred to the motor side is

(Jm + mr2 )ω̇ m = T − rFe (3.18)

The equation of motion for motor and load referred to the load side is
1 1
( Jm + m)v̇ = T − Fe (3.19)
r2 r

3.2.6 Torque characteristics

TL 1

2
TL
T
T

m m

Figure 3.4: To the left is shown a stable system where the load torque TL is increasing
for increasing motor velocity ω m . To the right is shown a system with two equilibrium
points. Equilibrium 1 is stable, while equilibrium 2 is unstable as the load torque TL
decreases faster than the motor torque T when the motor velocity ω m increases.

In many applications the load torque TL will depend on the motor speed. An example
of this is shown in the left diagram of Figure 3.4, where the load torque increases with
increasing speed. This will be the case for systems where the friction increases with the
velocity, like the air resistance of a car or a bicycle. Moreover, the motor torque will
typically be a decreasing function of the motor shaft speed ω m due to increasing energy
loss in the motor. It turns out that if both the motor torque and the load torque are
functions of the motor speed so that T = T (ω m ) and TL = TL (ω m ), then the stability
of the motor and load can be investigated in a torque-speed diagram. This is done by
linearization of the motor model (3.1), which gives

Jm ∆ω̇ m = k∆ω m (3.20)

where µ ¶¯
∂T ∂TL ¯¯
k= − (3.21)
∂ω m ∂ω m ¯ωmo

is a linearization constant. From linear stability theory we see that the system is stable
if and only if k ≤ 0. This can be investigated graphically in a torque-speed diagram as
shown in Figure 3.4.

Example 41 Suppose that a motor is connected to a load which is simply a friction

84 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

torque TL . The friction is given by

( " µ ¶2 #)
ωm
TL (ω m ) = Tc + (Ts − Tc ) exp − sgn(ω m ) + Bω m (3.22)
ωs
where Tc is the Coulomb friction and Ts is the static friction and
½
−1 ω m < 0
sgn(ω m ) = (3.23)
1 0 < ωm
The constant ω s is the characteristic velocity of the Stribeck effect, and B is the coeffi-
cient of the viscous friction. For further details on this friction characteristic, see the
Chapter 5. The motor torque is directly controlled, so that T is a constant. The equation
of motion is then
( " µ ¶2 #)
ωm
Jm ω̇ m = T − Tc + (Ts − Tc ) exp − sgn(ω m ) − Bω m (3.24)
ωs
For simplicity it is assumed that ω m ≥ 0 so that sgn(ω m ) = 1. Linearization gives
à " µ ¶2 # !
ωm ωm
Jm ∆ω̇ m = 2 (Ts − Tc ) exp − − B ∆ω m (3.25)
ωs ωs
This shows that the system is unstable for constant motor torque T at the speed ω m if
" µ ¶2 #
ωm ωm
B<2 (Ts − Tc ) exp − (3.26)
ωs ωs

3.2.7 The four quadrants of the motor

m ia

Generator Motor Generator Motor

T ua

Motor Generator Motor Generator

Figure 3.5: The four quadrants of the motor (to the left), and the four quadrants of the
power amplifier (to the right).

A motor delivers the mechanical power T ω m through the motor shaft, where T is
the motor torque and ω m is the motor speed. If T and ω m has the same signs under
stationary operation, then the motor delivers power. In this case the motor transforms
electrical power to mechanical power, and is said to work as a motor. If T and ω m has
opposite signs under stationary operation, then the motor receives mechanical power and
transforms it into electrical power, and is said to work as a generator. This is illustrated
in Figure 3.5.
3.3. THE DC MOTOR WITH CONSTANT FIELD 85

3.3 The DC motor with constant field

3.3.1 Introduction
The DC motor with constant field has a simple dynamic model, and has been controlled
accurately with simple electronics from the early period of automatic control. Because
of this it has been a very important component in servomechanisms, which are control
systems involving fast and accurate control of motion. In modern servomechanisms, the
DC motor will always be used as a current controlled DC motor, where a high gain
current loop is integrated with the motor. This makes it possible to control the motor
torque directly, and this is one of the reasons for the success of the motor. More recently,
advanced power electronics has made it possible to control other types of electrical motors
with the same fast response as the DC motor. This will typically involve some method
to control the motor torque, which leads to the same dynamic model as for the current
controlled DC motor. Therefore, the models and the analysis results presented for the
current controlled DC motor in this section is also valid for other types of electrical
motors where the motor torque can be controlled.

3.3.2 Model
ia
Ra La
T, m ,  m

+ +
TL
ua ea

- - Jm

Figure 3.6: Armature circuit of DC motor with constant field.

A DC motor with constant field is described by an armature circuit with current

ia and input voltage ua . The armature circuit is a serial connection of the armature
resistance Ra , the armature inductance La , and the electromechanical energy conversion
unit with induced voltage ea . This voltage ea is induced by the motor speed ω m in
combination with a constant electromagnetic field that is set up either by a field circuit
with a constant field current ie , or by a permanent magnet which replaces the field circuit.
An important characteristic of the DC motor with constant field is that the motor torque
is proportional to the armature current, and is given by

T = KT ia (3.27)

where KT is the torque constant.

The DC motor with constant field can be described as a serial interconnection of three
passive two-ports. The first two-port is the armature circuit where the input port has
variables ua and ia , and the output port has variables ea and ia . The second two-port is
86 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

the electromechanical energy conversion unit with an electrical port with port variables
ea and ia , and a mechanical port with port variables T and ω m . Finally, the third
two-port is the motor shaft with input port with variables T and ω m and output port
with variables TL and ω m . There is no energy storage in the electromechanical energy
conversion unit, which implies that the power ea ia of the electrical port equals the power
T ω m of the mechanical port. This gives

ea ia = T ω m = KT ia ω m ⇒ ea = KE ω m (3.28)

where KE = KT is the field constant. The dynamic model can then be found from the
voltage law of the armature circuit and the equation of motion for the motor shaft:

A DC motor with constant field has the dynamic model

d
La ia = −Ra ia − KE ω m + ua (3.29)
dt
Jm ω̇ m = KT ia − TL (3.30)
θ̇m = ωm (3.31)

The block diagram is shown in Figure 3.7.

TL

ua
1
ia - 1
m m
La KT Jm
- -

Ra

Ke

Figure 3.7: Voltage controlled DC motor.

3.3.3 Energy function

The total energy V of the motor is the sum of inductive energy stored in the armature
inductance La and the kinetic energy of the motor shaft. This gives
1 1
V = La i2a + Jm ω 2m ≥ 0 (3.32)
2 2
The time derivative of the energy function V along the solutions of the system is
dia
V̇ = ia La + ω m Jm ω̇ m
dt
= ia (−Ra ia − KE ω m + ua ) + ω m (KT ia − TL )
= ia ua − ω m TL − Ra i2a
≤ ia ua − ω m TL (3.33)
3.3. THE DC MOTOR WITH CONSTANT FIELD 87

Now, suppose that the load model with input ω m and output TL is passive, and that
there is a storage function VL ≥ 0 so that

V̇L ≤ ω m TL (3.34)

Then the total energy Vt := V + VL is greater that or equal to zero, and the time
derivative of Vt along the solution of the system is

V̇t ≤ ia ua (3.35)

We have then established the following result:

The DC motor model (3.29—3.31) with input ua and output ia is passive if the load with
input ω m and output TL is passive.

Example 42 Suppose that the load is simply a damper with viscous friction so that
TL = BL ω m . Then the system with input ω m and output TL is passive with storage
function VL = 0 as this gives

V̇L = 0 ≤ BL ω 2m = ω m TL (3.36)

It follows that the DC motor with input ua and output ia is passive with this load.

Example 43 Suppose that the load is an inertia JL with shaft angle θL connected to the
motor shaft by a spring with torque

TL = K(θm − θL ) + D(θ̇m − θ̇L ) (3.37)

The equation of motion for the inertia is

JL θ̈L = TL (3.38)

Then the system with input ω m and output TL is passive with storage function equal to
the total energy
1 2 1
VL = JL θ̇L + K(θm − θL )2 ≥ 0 (3.39)
2 2
because the time derivative of VL along the solutions of the system is

V̇L = JL θ̈L θ̇L + K(θm − θL )(θ̇m − θ̇L )

= TL θ̇L − K(θm − θL )(θ̇m − θ̇L )
= TL θ̇m + TL (θ̇L − θ̇m ) − K(θm − θL )(θ̇m − θ̇L )
= TL θ̇m − D(θ̇m − θ̇L )2 (3.40)

3.3.4 Laplace transformed model

Laplace transformation of the DC motor model (3.29—3.31) gives
1
sia (s) = [−Ra ia (s) − KE ω m (s) + ua (s)] (3.41)
La
KT 1
sω m (s) = ia (s) − TL (s) (3.42)
Jm Jm
sθm (s) = ω m (s) (3.43)
88 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

The equation of motion gives

KT 1
s2 θm (s) = ia (s) − TL (s) (3.44)
Jm Jm
while the armature equation gives

(La s + Ra ) ia (s) = −KE sθm (s) + ua (s) (3.45)

Insertion of (3.45) in (3.44) gives

KT 1 1
s2 θm (s) = (−KE sθm (s) + ua (s)) − TL (s) (3.46)
Jm La s + Ra Jm
and finally ³ ´
1 Ra La
KE ua (s) − KE KT 1+ Ra s
TL (s)
θm (s) = ³ ´ (3.47)
s KJm La 2
E KT
s + Jm Ra
KE KT s + 1

This can be written

1 Ra
KE ua (s) −KT KE (1 + Ta s)TL (s)
θm (s) = (3.48)
s(Ta Tm s2 + Tm s + 1)
where
La
Ta = (3.49)
Ra
is the electrical time constant of the motor, and
Jm Ra
Tm = (3.50)
KE KT
is the mechanical time constant. Usually, one may assume that the electrical time con-
stant is much less than the mechanical time constant so that the model can be written
1 Ra
KE ua (s) KE KT TL (s)
θm (s) = − (3.51)
s(1 + Tm s)(1 + Ta s) s(1 + Tm s)
This leads to the following result:

The transfer function from the input ua to the angle θm is

1
θm KE
Hp (s) = (s) = (3.52)
ua s(1 + Tm s)(1 + Ta s)

3.4 DC motor control

3.4.1 Introduction
A DC motor used in a servomechanism will more or less always have a current control
loop integrated in the motor, and it will normally have a speed loop outside of the current
loop. These feedback loops are often seen as an integrated part of the DC motor, and
3.4. DC MOTOR CONTROL 89

it is therefore useful to present models of the DC motor with current control and with
speed control. The advantage of these feedback loops is that the current loop will have
very high bandwidth, and it will therefore suppress nonlinearities in the power amplifier.
The velocity loop can also be given a high bandwidth, and will tend to eliminate the
effect of friction on the motor. The outer position control loop will normally have to
be slower than the first mechanical resonance, and this limits the gain in the position
loop. Usually a PI controller will be used in the current loop, and a PI controller with
limited integral action will used in the velocity loop. In the presentation here we use P
controllers to simplify the expressions. The main results will still be valid.

3.4.2 Current controlled DC motor

The transfer function from the input ua (s) to the current ia (s) of the armature circuit
can be found from
ia ia θm
(s) = (s) (s) (3.53)
ua θm ua
From the Laplace transformed model we see that

KT ia Jm s2
s2 θm = ia ⇒ (s) = (3.54)
Jm θm KT

and we find that

Jm
ia KE KT s
Ha (s) := (s) = (3.55)
ua 1 + Tm s + Tm Ta s2
where it is used that KE = KT . The following current controller is used:

ua = Ki (id − ia ) (3.56)

This gives the closed loop dynamics

ia Ki Ha (s)
(s) = (3.57)
id 1 + Ki Ha (s)

In practice it is possible to select a very high gain Ki . This is a consequence of the

passivity of the system when ua is input and ia is output, which implies that the transfer
function Ha (s) is positive real with phase satisfying |∠Ha (jω)| ≤ 90◦ . Therefore we may
let Ki approach infinity in the expression, which gives the approximation

ia (s) = id (s) (3.58)

Insertion in (3.44) gives the following result:

The model of a current controlled DC motor is given by the double integrator model
1
θm (s) = [KT id (s) − TL (s)] (3.59)
Jm s2
where the input is the desired current id .

The block diagram is shown in Figure 3.8.

90 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

TL

id - 1
m m
KT Jm

Figure 3.8: Current controlled DC motor.

Example 44 Consider a PI controller

1 + Ti s
ua = Ki Ti (i0 − ia ) (3.60)
Ti s
for armature current control where Ti = Tm . Assume that Ta ¿ Tm so that the denom-
inator of L(s) can be factored as (1 + Ta s)(1 + Tm s). Then the loop transfer function
is
Jm 1
L(s) = Ki (3.61)
KE KT 1 + Ta s
which shows that the controller is effective also at low frequencies. Also in this case the
model (3.59) results for realistic gains Ki . This is the controller that is used in practice.

3.4.3 Velocity controlled DC motor

The speed ω m of the motor satisfies sθm = ω m , and it follows that

ωm ωm θm KT
(s) = (s) (s) = (3.62)
id θm id Jm s

This transfer function is the product of a gain KT /Jm and an integrator 1/s. The velocity
controller
id = Kω (ω d − ω m ) (3.63)
gives gives the closed-loop dynamics
Ka
ωm s 1
(s) = = (3.64)
ωd 1+ sKa 1 + Ksa

where
KT Kω
Ka = (3.65)
Jm
is the acceleration constant of the system. We see that the velocity loop is stable as it
has only one pole, which is at s = −Ka .

3.4.4 Position controlled DC motor

A position feedback loop is closed around the velocity loop as shown in Figure 3.9. The
transfer function from the velocity ω m to the angle θm is an integrator, which leads to

θm 1
(s) = ³ ´ (3.66)
ωd s 1+ s
Ka
3.5. MOTOR AND LOAD WITH ELASTIC TRANSMISSION 91

d d m m
Kv Ka
- -

 ca

 cv

Figure 3.9: Current controlled DC motor with velocity loop and position loop. The
crossover frequency of the velocity loop is ω ca = Ka , and the crossover frequency of the
position loop is ω cv = Kv provided that Ka À Kv .

The position controller

ω d = Kv (θd − θm ) (3.67)
where Kv is the velocity constant gives the closed-loop dynamics
θm 1
(s) = 1 1 2
(3.68)
θd 1+ Kv s + Kv Ka s

Usually, Ka can be selected to be several hundred rad/s, while Kv is usually limited by

the first resonance in the system, which will typically occur in the range 10—100 rad/s.
Therefore, we may assume that Ka À Kv , and we get
θm 1
(s) = ³ ´³ ´ (3.69)
θd 1+ s
1+ s
Kv Ka

3.5 Motor and load with elastic transmission

3.5.1 Introduction
A situation that is often seen in control applications is that a motor is used to move some
inertial load using an elastic interconnection. The elasticity may be due to a flexibility in
shaft or in a gearbox, or it may be that the motor and load are interconnected by wires
or with a crane that is not completely rigid. This type of system will be modelled and
analyzed in this section. It turns out that the transfer functions of the system have very
interesting properties that have great significance in the selection of controller structure
for such systems. It will be shown that some of these properties can be explained from
passivity arguments where the energy formulation can be used efficiently. The results are
useful both from a practical perspective, and, in addition, the results provide valuable
insight into passivity-based controller design.

3.5.2 Equations of motion

We consider a motor driving a load through an elastic transmission as shown in Fig-
ure 3.10. The equation of motion for the motor and load are given by
Jm θ̈m = Tm − TL (3.70)
JL θ̈L = TL (3.71)
92 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

T m , m , m  L , L
Motor Load
Jm JL
K, D

Figure 3.10: Motor with elastic transmission

where θm is the motor shaft angle, θL is the load shaft angle, Tm is the motor torque, and
TL is the load torque from the transmission on the motor shaft. The elastic transmission
and the load inertia is modelled as a torsional spring with spring constant K in parallel
with a torsional damper with damping coefficient D. The resulting torque is
TL = −Kθe − Dθ̇e (3.72)
where
θe = θL − θm (3.73)
is the elastic deflection of the the transmission. In the derivation of the transfer functions
it is helpful to introduce the variable
JL
θr = θm + θL . (3.74)
Jm
and derive the model in terms of the variables θe and θr .
The equations of motion for θe and θr are found by combining the equations (3.70—
3.72). This gives
D K 1
θ̈e + θ̇e + θe = − Tm (3.75)
Je Je Jm
Tm
θ̈r = (3.76)
Jm
where
Jm JL
Je = , J = Jm + JL (3.77)
J

3.5.3 Transfer functions

We see from (3.75) and (3.76) that the dynamic model of the elastic deflection θe is a
second order oscillatory system, while the rigid motion θr results from a double integrator
from the motor torque Tm . The transfer functions from the input Tm to θe and θr are
found to given by
³ ´2
1
θe 1 ω1
(s) = − ³ ´2 (3.78)
Tm Jm
1 + 2ζ 1 ωs1 + ωs1
θr 1
(s) = (3.79)
Tm Jm s2
where r
K D 1
ω1 = and ζ 1 = √ (3.80)
Je 2 Je K
3.5. MOTOR AND LOAD WITH ELASTIC TRANSMISSION 93

The transfer functions for the original variables are found by solving (3.73) and (3.74),
which gives
µ ¶
Jm JL
θm = θr − θe (3.81)
J Jm
Jm
θL = (θr + θe ) (3.82)
J

This gives
· ¸
θm Jm θr JL θe
(s) = (s) − (s)
Tm J Tm Jm Tm
 ³ ´2 
JL 1
11 Jm ω 1 
=  2+ ³ ´2  (3.83)
J s s s
1 + 2ζ 1 ω1 + ω1

and
· ¸
θL Jm θr θe
(s) = (s) + (s)
Tm J Tm Tm
 ³ ´2 
1
Jm  1 1 ω1 
=  − ³ ´2  (3.84)
J Jm s2 Jm
1 + 2ζ 1 ωs1 + ωs1

After some work the following result is found:

The motor and elastic load with elastic transmission is described by the two transfer
functions
³ ´2
s s
θm 1 1 + 2ζ a ωa + ωa
Hθm (s) : = (s) = ³ ´2 (3.85)
Tm Js2
1 + 2ζ 1 ωs1 + ωs1
θL 1 1 + 2ζ 1 ωs1
HθL (s) : = (s) = ³ ´2 (3.86)
Tm Js2
1 + 2ζ 1 ωs1 + ωs1

where the parameters are given by

r
D 1 K
ζ1 = √ , ω1 = (3.87)
2 Je K Je
r r
Jm Jm
ζa = ζ , ωa = ω1 < ω1 (3.88)
J 1 J

The transfer functions are often formulated in terms of the shaft speeds ω m (s) =
94 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

sθm (s) and ω L (s) = sθL (s). Then the transfer functions are
³ ´2
s s
ωm 1 1 + 2ζ a ω a
+ ωa
Hωm (jω) : = (s) = ³ ´2 (3.89)
Tm Js
1 + 2ζ 1 ωs1 + ωs1
ωL 1 1 + 2ζ 1 ωs1
HωL (jω) : = (s) = ³ ´2 (3.90)
Tm Js
1 + 2ζ 1 ωs1 + ωs1

Bode Diagram θ /T Bode Diagram θ /T

m m L m

40 50

30

20
0
10
Magnitude

Magnitude
0

-10
-50
-20

-30

-40 -100

0 -150

-200
-50

-250
Phase

Phase

-100
-300

-150
-350

-200 -400
-1 0 1 -1 0 1
10 10 10 10 10 10

Frequency (rad/s) Frequency (rad/s)

Figure 3.11: Frequency response from the motor torque Tm to the motor angle θm (to
the left), and frequency response from the motor torque Tm to the load angle θL (to the
right).

An important observation is that ω a < ω 1 , which means that the break frequency
of the zeros in is smaller than the break frequency of the poles in θm /Tm (jω). The
frequency responses are shown in Figure 3.11 for K1 = 0.5, Jm = J1 = 1 and D1 = 0.01.
Note that the frequency response Hθm (jω) of the motor angle does not have any negative
phase contribution from the elasticity, whereas the frequency response HθL (jω) of the
load angle drops 180◦ because of the resonance. Obviously, this has serious consequences
for controller design, and for the achievable bandwidth when feedback is taken from
either θ̇m or θ̇L . In practice it means that when feedback is taken from θ̇L the crossover
frequency must be less than ω 1 . In contrast to this, the crossover may be selected above
ω 1 when feedback is taken from θ̇m . Experience shows that feedback loops from θ̇m are
very robust, and can be given a very high crossover frequency. Feedback from θ̇L gives
an upper limit on the crossover frequency at ω 1 .
3.5. MOTOR AND LOAD WITH ELASTIC TRANSMISSION 95

3.5.4 Zeros of the transfer function

The zeros of the transfer function Hθm (s) are the roots of
µ ¶2
s s
1 + 2ζ a + =0 (3.91)
ωa ωa

Under the assumption that ζ a ¿ 1 the transfer function Hθm (s) will have zeros close to
±jω a . This means that a nonzero torque input Tm (jω a ) with frequency ω a will give a
small θm (jω a ) as ±jω a are close to the zeros of Hθm (s).
From (3.85) and (3.86) it is seen that the transfer function from the motor angle to
the load angle is given by

θL θL Tm 1 + 2ζ 1 ωs1
(s) = (s) (s) = ³ ´2 (3.92)
θm Tm θm
1 + 2ζ a ωsa + ωsa

This means that poles of θL (s) /θm (s) are equal to the zeros of Hθm (s) which are close
to ±jω a . This means that the system θL (s) /θm (s) will have resonances near ±jω a , so
that a large amplitude in θL (jω a ) can occur with a small θm (jω a ). This agrees with the
fact that the zeros of Hθm (s) which are close to ±jω a .

3.5.5 Energy analysis

The sum of kinetic and potential energy for the motor, transmission and load is
1 1 1
V = Jm ω 2m + JL ω 2L + Kθ2e (3.93)
2 2 2

The time derivative of the energy as the system evolves will be the power θ̇m Tm supplied
2
by the input Tm minus the power Dθ̇e dissipated in the rotational damper. This is
written
2
V̇ = θ̇m Tm − Dθ̇e (3.94)
This implies that the system with input Tm and output θ̇m is passive. This again implies
that the transfer function Hωm (s) from the input Tm to ω m = θ̇m is positive real, which
means that
Re[Hωm (jω)] ≥ 0 for all ω. (3.95)
Thus, from simple energy arguments we can establish that

|∠Hωm (jω)| ≤ 90◦ (3.96)

which implies that

−180◦ ≤ ∠Hθm (jω) ≤ 0◦ . (3.97)
This result in agreement with the plot in Figure 3.11

3.5.6 Motor with several resonances in the load

We may connect an additional degree of freedom in the load by modifying the load into
a two-port with dynamics
JL ω̇ L = TL − T1 (3.98)
96 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

and by adding the mechanical two-port

J2 ω̇ 2 = T1 − T2 (3.99)
d
(θ1 − θ2 ) = (ω 1 − ω 2 ) (3.100)
dt
T1 = D1 (ω 1 − ω 2 ) + K1 (θ1 − θ2 ) . (3.101)

where ω 2 is the shaft speed, and J2 is the inertia. The transmission is modelled as a
torsional spring with spring constant K1 in parallel with a torsional damper with damping
coefficient D1 . The input port has effort T1 and flow ω 1 , while the output port has effort
T2 and flow ω 2 . We may add on any number of additional degrees of freedom as two-ports

Ji ω̇ i = Ti−1 − Ti (3.102)
d
(θi−1 − θi ) = (ω i−1 − ω i ) (3.103)
dt
Ti−1 = Di−1 (ω i−1 − ω i ) + Ki−1 (θi−1 − θi ) (3.104)

with port variables Ti−1 and ω i−1 at the input and Ti and ω i at the output. In a
computational setting the inputs are ω i−1 and Ti , while the outputs are ω i and Ti−1 .
The sum of kinetic and potential energy for a motor with n degrees of freedom in the
load is
Xn n−1
X1
1 1 1
V = Jm ω 2m + Ji ω 2i + Kθ2e + Ki (θi − θi+1 )2 (3.105)
2 i=1
2 2 i=1
2

The time derivative of the energy for the solutions of the system will be the power ω m Tm
supplied by the input Tm minus the power dissipated in the rotational dampers. This is
written
n−1
X
2
V̇ = ω m Tm − Dθ̇e − Di (ω i − ω i+1 )2 (3.106)
i=1

This implies that the system with input Tm and output ω m will still be passive with n
degrees of freedom in the load.

3.5.7 Two motors driving an elastic load

Consider an inertia JL with rotation angle θL that is driven by two motors. Motor 1 has
shaft angle θ1 , inertia J1 and motor torque Tm1 , while motor 2 has shaft angle θ2 , inertia
J2 and motor torque Tm2 . The motors are connected to the load using gears with gear
ratio n.
The model of the system is derived by first establishing the equations of motion for
the two motors and for the load, and then connecting the motors and the load by deriving
expressions for the connecting torques. The equations of motion for the motors are

J1 θ̈1 = Tm1 − Tg1 (3.107)

J2 θ̈2 = Tm2 − Tg2 (3.108)

where Tg1 is the torque from gear 1 on motor 1, and Tg2 is the torque from gear 2 on
motor 2. The equation of motion for the load is
1
JL θ̈L = (Tg1 + Tg2 ) − Te (3.109)
n
3.6. MOTOR AND LOAD WITH DEADZONE IN THE GEAR 97

where Te is an external disturbance torque.

The elastic deformation of the gears referenced to the motor side are given by
1 1
φ1 = θ1 − θL , φ2 = θ2 − θL
n n
The gears can then be modeled as springs and dampers with torques

Tg1 = K1 φ1 + D1 φ̇1 , Tg2 = K2 φ2 + D1 φ̇2 (3.110)

3.5.8 Energy analysis of two motors and load

The system of two motors and a load can be regarded as an interconnection of three
two-ports, where the load is a two-port connected to motor 1 through a port with input
θ̇1 and output Tg1 , and to motor 2 through a port with input θ̇2 and output Tg2 . The
total energy of the system is
1³ 2 2 2
´ 1¡ ¢
V = J1 θ̇1 + J2 θ̇2 + JL θ̇L + K1 φ21 + K2 φ22 ≥ 0 (3.111)
2 2
The time derivative of the energy will be the power suppled by the motor torques minus
the power dissipated in the dampers. This gives
2 2
V̇ = Tm1 θ̇1 + Tm2 θ̇2 − D1 φ̇1 − D2 φ̇2 (3.112)

This shows that if a passive controller from θ̇2 to Tm2 is used for motor 2, then the system
with input Tm1 and output θ̇1 will be passive.

Example 45 Suppose that a PD controller

Tm2 = −Kp2 θ2 − Kd2 θ̇2 (3.113)

is used for motor 2. This controller is passive when θ̇2 is considered to be the input and
Tm2 is the output. In agreement with this, the controller has a mechanical analog which
is a spring with stiffness Kp2 and a damper with coefficient Kd2 . The system can then be
analyzed using the energy function of the system including the mechanical analog. The
energy function for this system is
1³ 2 2 2
´ 1¡ ¢
Va = J1 θ̇1 + J2 θ̇2 + JL θ̇L + K1 φ21 + K2 φ22 + Kp2 θ22 ≥ 0 (3.114)
2 2
which will have time derivative along the solutions of the system given by
2 2 2
V̇a = Tm1 θ̇1 − D1 φ̇1 − D2 φ̇2 − Kd2 θ̇2 (3.115)

This shows that the system with input Tm1 and output θ̇1 is passive when the PD controller
(3.113) is used.

3.6 Motor and load with deadzone in the gear

3.6.1 Introduction
In this section we will study the problem of a motor that drives a load through a gear
with a deadzone. In the deadzone there is no physical contact between the input axis and
98 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

the output axis of the gear, and as a consequence of this there is no torque transmitted
through the gear in the deadzone. The modeling of a gear with deadzone requires some
care. In the following it will be seen that if it is assumed that there is elasticity in the
gear, then the modeling is simplified. In case of a rigid gear with deadzone it is necessary
to switch between two models of that have a different number of states.

3.6.2 Elastic gear with deadzone

The equations of motion for the motor and load are given by

Jm θ̈m = Tm − Tgm (3.116)

JL θ̈L = TgL (3.117)

where Tgm is the gear torque on the motor side, and TgL is the gear torque on the load
side. The gear ratio is n. The deflection between the motor and the load is given by
1
φ = θm − θL (3.118)
n
The gear has a deadzone δ. This means that the gear torque is zero when |φ| < δ.
Suppose that the gear is elastic, and that is can be described by a spring with stiffness
K outside of the deadzone. Then the gear torques Tgm and TgL are given as functions
of the gear deflection φ according to

 K (φ + δ) , φ ≤ −δ 1
Tgm (φ) = 0, −δ ≤ φ ≤ δ , TgL (φ) = Tgm (φ) (3.119)
 n
K (φ − δ) , δ ≤ φ

The gear is then a mechanical two-port in impedance form where port 1 has input θ̇m
and output Tgm , and port 2 with input θ̇L and output TgL . Port 1 of the gear will then
be compatible with the motor port, which has output θ̇m and input Tgm , and in the same
way port 2 of the gear can be connected with the port of the load. The interconnection
of the motor, gear and load is then straightforward, and a simulation model is given by
the equations (3.116—3.119).

3.6.3 Rigid gear with deadzone

If it is assumed that the gear is rigid, then the system will have two independent degrees
of freedom θ̇L and θ̇m in the deadzone, and only one degree of freedom θ̇L = θ̇m outside
of the deadzone. This means that the system changes the number of degrees of freedom
from two to one when it leaves the dead-zone. In this case the gear torques are functions
of the deflection when the system is inside the deadzone as
¾
Tgm (φ) = 0
, |φ| < δ (3.120)
TgL (φ) = 0

This can be regarded as an impedance model with inputs θ̇m and θ̇L and outputs Tgm
and TgL . Outside the deadzone the gear is defined by the usual gear equations
¾
θ̇L = nθ̇m
, |φ| = δ (3.121)
Tgm = nTgL
3.6. MOTOR AND LOAD WITH DEADZONE IN THE GEAR 99

In terms of inputs and outputs this can be regarded as a hybrid model with inputs θ̇m
and TgL and outputs θ̇L and Tgm , or it can be seen as a cascade model with inputs θ̇m
and Tgm and outputs θ̇L and TgL . Note, however, that the gear model for |φ| = δ cannot
be put in impedance form.
The impedance model (3.120) that is valid for |φ| < δ is a two-port with inputs and
outputs that are compatible with the two-port formulation of the motor and load where
shaft speed is output and torque is output. In contrast to this, the hybrid model (3.121)
does not have inputs and output that can be connected to the motor and load. In fact,
the system of motor, gear and load has only one degree of freedom in this case, and the
models of the motor and load must be combined into one model.
The way to handle this in a simulation system is to switch between three models,
where one model is valid inside the deadzone, and there is one model on each side of
the deadzone. At the negative side of the deadzone where φ = −δ the load angle is
θL = n(θm + δ), and JL θ̈L = TgL must be negative if the system is to stay at φ = −δ.
This implies that θ̈m must be negative, which is the case if Tm < 0, while the systems
enters the deadzone if Tm > 0 At the positive side of the deadzone where φ = δ then
Tm > 0 will give positive acceleration, and the system will stay at φ = δ. If Tm < 0,
then the system enters the deadzone. The three models are therefore given by
¡ ¢ ¾
Jm + n2 JL θ̈m = Tm
when φ = −δ and Tm < 0 (3.122)
θL = n(θm + δ)
¡ ¢ ¾
Jm + n2 JL θ̈m = Tm
when φ = δ and Tm > 0 (3.123)
θL = n(θm − δ)
¾
Jm θ̈m = Tm
otherwise (3.124)
JL θ̈L = 0

In simulations it is necessary to have an event-detection method to determine when the

system enters and leaves the deadzone.

3.6.4 Two motors with deadzone and load

Large space antennas need to be rotated with high accuracy at a very low speed. This
will normally require a reduction gear between the motor and the antenna, where gear
will typically have a deadzone. Because of this, the motor and antenna may oscillate
because of the deadzone, and this may prevent the system from achieving the specified
accuracy. A typical configuration for such systems is to use two motors that are connected
with gears to the antenna. With this type of solution the chattering may be eliminated
by pretensioning the motors in opposite directions so that both gears are loaded for
moderate control torques. The equations of motion for the motors are

J1 θ̈1 = Tm1 − Tg1 (3.125)

J2 θ̈2 = Tm2 − Tg2 (3.126)

where Tg1 is the torque from gear 1 on motor 1, and Tg2 is the torque from gear 2 on
motor 2. The equation of motion for the load is
1
JL θ̈L = (Tg1 + Tg2 ) − Te (3.127)
n
where Te is an external disturbance torque.
100 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

The gears are supposed to have a spring constant K and a deadzone δ. The deviation
angle of the gear referenced to the motor side are given by

1 1
φ1 = θ1 − θL , φ2 = θ2 − θL
n n

The gears can then be modeled as a spring with a deadzone, which gives the gear torques

 K (φi + δ) , φi ≤ −δ
Tgi = 0, −δ ≤ φi ≤ δ , i = 1, 2 (3.128)

K (φi − δ) , δ ≤ φi

The motors can then be connected to the load with the gear equations. The system is
shown in Figure 3.12.

Motor 1 Dead-zone
Torque
reference 1 Torque 1 1
loop
J 1s s Load
Load
torque 1 1
J ps s

Torque
reference 2 Torque 1 1
loop J 2s s
Motor 2 Dead-zone

Figure 3.12: Block diagram of two motors driving a load through elastic gears with
deadzone.

Due to the deadzone and the lack of damping in the gear the load may oscillate.
This can be eliminated by pretensioning the gears by controlling the motors with an
offset torque in opposite directions. Alternatively, damping of φ1 and φ2 can be achieved
by including rate feedback from φ̇1 and φ̇2 (Leonhard 1996), (Gavronski, Beech-Brandt,
Ahlstrom and Manieri 2000).

3.7 Electromechanical energy conversion

3.7.1 Introduction
Electrical motors and various electrical sensors and actuators are based on energy conver-
sion between electrical energy and mechanical energy. This energy conversion typically
takes place due to inductive and capacitive effects. The presentation that follows starts
with a presentation of energy functions for inductive and capacitive circuit elements, and
proceeds by extending these results to electromechanical systems.
3.7. ELECTROMECHANICAL ENERGY CONVERSION 101

3.7.2 Inductive circuit elements

The magnetic field intensity H and the magnet flux density B satisfy the equations

∇×H = J (3.129)
∇·B = 0 (3.130)

where J is the current density. The relation between the two vector fields H and B is
given by
B = µH (3.131)
where the permeability µ is a material constant given by the constitutive law

µ = (1 + χm ) µ0 (3.132)

of the magnetic material, where µ0 is the permeability of free space.

v h


H

Figure 3.13: Magnetic circiut with airgap

We consider the magnetic circuit in Figure 3.13. The magnetic circuit has an iron
core of cross section A and length . The iron core has a small airgap of length h and a
winding of N turns with current i. Then the integral form of (3.129) is
I
H · ds = N i =: F (3.133)

where F is the magnetomotive force (mmf), which is also called the ampere-turns, and
ds is a differential length which is tangent to the path of integration. The magnetic flux
φ of the circuit is
φ = AB (3.134)
The iron core has much higher magnetic permeability than the surrounding air, and as
the airgap is small, the iron core and the airgap form a closed path for the magnetic
flux φ of length in the core and length h in air. The magnetic flux density B has zero
divergence, which means that the flux φ is the same for all cross sections of the iron core
and for the airgap. Therefore we may consider the iron core and the airgap to form a
magnetic circuit where the magnetomotive force N i sets up a magnetic flux φ that flows
102 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

through the circuit. The magnitude of the magnetic field intensity is Hc = B/µ in the
core and Ha = B/µ0 in the airgap. Then (3.133) gives
µ ¶ µ ¶
h h φ
N i = Hc + Ha h = + B= + (3.135)
µ µ0 µ µ0 A
The magnetic circuit has a reluctance R defined so that the relation between the mag-
netomotive force and the flux is
N i = Rφ (3.136)
We see that the reluctance of the magnetic circuit is
µ ¶
Ni 1 h h
R= = + ≈ (3.137)
φ A µ µ0 Aµ0
where the approximation can be done as the magnetic permeability µ0 in air is much
smaller than the permeability µ in iron.
We now turn to the electrical circuit with the coil. Define the magnetic flux linkage
λ of the coil which is defined by
λ := N φ (3.138)
In the example in this section, the flux linkage is a linear function of the current, which
is given by
λ = Li (3.139)
where
µ0 N 2 A
L= (3.140)
h
is the inductance of the coil.

The current is given by the constitutive equation

i = i(λ) (3.141)

The voltage set up in the winding is the time derivative of the state λ by Faraday’s law:

u = λ̇ (3.142)

3.7.3 Capacitive circuit elements

The state of a capacitive circuit element is given by the electrical charge q, and the
current is defined by
i = q̇ (3.143)
The voltage is given by the constitutive equation

u = u(q) (3.144)

In the case of the linear element the charge is given by

q = Cu (3.145)
where C is the capacitance.
3.7. ELECTROMECHANICAL ENERGY CONVERSION 103

3.7.4 Magnetic energy of a linear inductive element

In this section we will derive an expression for the stored energy of a linear inductive
element with current i, flux linkage λ = Li, and voltage u. The state of the element is
given by the flux linkage λ. The voltage is given by Faraday’s law to be u = λ̇. The power
supplied to the inductive element is P = iu, and the stored energy has the differential

dλ
dWm = P dt = iudt = i dt = idλ (3.146)
dt
We can then calculate the stored energy corresponding to the state λ from
Z Z
λ
0 0
λ
λ0 0 λ2
Wm (λ) = i(λ )dλ = dλ = (3.147)
0 0 L 2L

In this expression the variable λ0 is introduced as a integration variable because λ is the

upper limit of the integration, and at the same time the current i(λ) is a function of λ.

3.7.5 Stored energy of a linear capacitive element

Consider a linear capacitive element with charge q, current i and voltage u. The state of
the capacitive element can be given by the charge q. The constitutive equation for the
element is q = Cu where C is the capacitance of the element. The current is by definition
given by i = q̇. Proceeding as for the inductive element we find that the stored energy
Wc has differential
dq
dWc = uidt = u dt = udq (3.148)
dt
The stored energy corresponding to the state q is found to be
Z q Z q
q0 0 q2
Wc (q) = u(q 0 )dq 0 = dq = (3.149)
0 0 C 2C

3.7.6 Energy and coenergy

We have seen that the stored energy of an inductive element is given as a function of
the flux linkage λ, while the stored energy of a capacitive element can be given as a
function of the charge q. In the derivation of models for electromechanical systems it is
convenient to have energy functions given by the current i and the voltage u, and this is
the motivation for introducing the concept of the coenergy. It is important to notice the
coenergy does not have a clear physical interpretation, it is merely a mathematical tool
that is useful in the to calculate the forces acting in electromechanical systems.
The energy of an inductive element is
Z λ ¡ ¢
Wm (λ) = i λ0 dλ0 (3.150)
0

where the energy is given as a function of the flux linkage λ. To change the variable to
the current i we define the coenergy of an inductive element by
Z i
∗
Wm (i) = λ (i0 ) di0 (3.151)
0
104 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

i
Characteristic curve
P
i

Coenergy stored
Wm i Energy stored
Wm i  i  Wm  
Wm  Wm    0 i d
 

 

Figure 3.14: Energy and coenergy of an inductive element.

From Figure 3.14 it is seen that

∗
Wm (λ) + Wm (i) = λi (3.152)

which means that the coenergy is found from the energy through a Legendre transfor-
mation
∗
Wm (i) = λi − Wm (λ) (3.153)
Again we stress that the coenergy is not necessarily a meaningful physical quantity, but
rather a mathematical tool that is useful in the modeling of electromechanical systems.
The differentials of the energy and the coenergy is given by
∗
dWm (λ) = idλ, dWm (i) = idλ + λdi − dWm (λ) = λdi (3.154)

This implies
∗
∂Wm (λ) ∂Wm (i)
= i and =λ (3.155)
∂λ ∂i
Example 46 If the flux linkage is given by λ = Li where the impedance L is a constant,
then the coenergy is
∗ λ2 1
Wm (i) = λi − = Li2 (3.156)
2L 2
∗
Then the numerical value of the energy Wm (λ) and the coenergy Wm (i) is the same,
note however, that the energy is a function of λ while the coenergy is a function of i.

The energy of a capacitive element is

Z q
Wc (q) = u (q 0 ) dq 0 (3.157)
0

We note that the energy is given in terms of the stored charge q. It may be desirable to
use the voltage u as a free variable instead of q. We therefore define the coenergy by
Z u
∗
Wc (u) = q (u0 ) du0 (3.158)
0
3.7. ELECTROMECHANICAL ENERGY CONVERSION 105

q
Characteristic curve

P
q

Energy stored Wc q

q
Wc q   uq  dq  Coenergy stored
0
Wc u  qu  Wc q 
Wc u

u u

Figure 3.15: Energy and coenergy of a capacitive element.

From Figure 3.15 it is seen that

Wc (q) + Wc∗ (u) = qu (3.159)

it it may be concluded that the coenergy is found from the energy through the Legendre
transformation
Wc∗ (u) = qu − Wc (q) (3.160)
The differential of the energy is
dWc (q) = udq (3.161)
while the differential of coenergy is

dWc∗ (u) = qdu + udq − dWc (q) = qdu (3.162)

We see that
∂Wc (q) ∂Wc∗ (u)
= u and =q (3.163)
∂q ∂u
Example 47 If q = Cu and the capacitance C is a constant, then
q2 1
Wc∗ (u) = qu − = Cu2 (3.164)
2C 2
and we see that the numerical value of the energy and the coenergy is the same. In this
case the characteristic curve of Figure 3.15 is a straight line.

3.7.7 Electromechanical two-port with inductive element

In this section we will consider an important energy-conversion system which will be
described as a two-port where input port is electrical with voltage u and current i, and
the output port is mechanical with effort F and flow ẋ. The conversion from electrical
energy to mechanical energy is done through an inductive element with flux linkage λ.
106 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

The electrical port has current i and voltage given by u = λ̇. The constitutive
equations are assumed to be given by

i = i(λ, x) (3.165)
F = F (λ, x) (3.166)

The power supplied to the two-port is

P = iλ̇ − F ẋ (3.167)

The energy stored in the magnetic system is assumed to be given by Wm (λ, x), and we
get
dWm (λ, x) = P dt = idλ − F dx (3.168)

Note that dWm (λ, x) is the absolute differential of the function Wm (λ, x). The absolute
differential may also be written

∂Wm (λ, x) ∂Wm (λ, x)

dWm (λ, x) = dλ + dx (3.169)
∂λ ∂x
Comparing the two expressions for the absolute differential we find that

∂Wm (λ, x)
i (λ, x) = (3.170)
∂λ
∂Wm (λ, x)
F (λ, x) = − (3.171)
∂x

The coenergy of the system is

∗
Wm (i, x) = λi − Wm (λ, x) (3.172)

The absolute differential of the coenergy is

∗
dWm (i, x) = idλ + λdi − dWm (λ, x) = λdi + F dx (3.173)

Then, from the general expression

∗ ∗
∗ ∂Wm (i, x) ∂Wm (i, x)
dWm (i, x) = di + dx (3.174)
∂i ∂x
for the absolute differential it follows that

The flux linkage λ and the force F of an electromechanical inductive element are given
∗
from the coenergy Wm (i, x) by
∗
∂Wm (i, x)
λ (i, x) = (3.175)
∂i
∗
∂Wm (i, x)
F (i, x) = (3.176)
∂x
3.7. ELECTROMECHANICAL ENERGY CONVERSION 107

3.7.8 Electromechanical two-port with linear flux linkage

If the flux linkage is linear in the current so that λ = L(x)i, and if the force is zero when
the flux linkage is zero so that F (0, x) = 0, then the energy Wm (λ, x) can be found by
first integrating (3.168) from (0, 0) to (0, x), and then from (0, x) to (λ, x). This gives
Z x Z λ Z λ 0
¡ 0 ¢ 0 λ 1 λ2
Wm (λ, x) = − 0 0
F (0, x )dx + i λ , x dλ = dλ0 = (3.177)
0 0 0 L(x) 2 L(x)
The coenergy is then given by

∗ λ2 1
Wm (i, x) = λi − = L(x)i2 (3.178)
2L(x) 2

An electromechanical two-port with linear flux linkage λ = L(x)i has force given by

∂Wm (λ, x) 1 ∂L (x) λ2

F (λ, x) = − = (3.179)
∂x 2 ∂x L(x)2

or alternatively, by
∗
∂Wm (i, x) 1 ∂L (x) 2
F (i, x) = = i (3.180)
∂x 2 ∂x

The voltage is given by

d di ∂L (x)
u = λ̇ = [L (x) i] = L + ẋi (3.181)
dt dt ∂x
A convenient way of describing this is to write
di
u=L +e (3.182)
dt
where the first term is the self inductance of the coil, and
∂L (x)
e := i ẋ (3.183)
∂x
is the voltage induced because of the velocity ẋ. The system is then regarded as a two-
port with input port variables u and i and output port variables e and i. The output port
is connected to a two-port describing the electromechanical conversion. This two-port has
input port with variables e and i, and output variables F and ẋ. This electromechanical
conversion unit does not store energy as ei = F ẋ, which means the electrical power ei in
is equal to mechanical power F ẋ out.

3.7.9 Magnetic levitation

In this section we will derive the model of the magnetic levitation experiment which is
used for teaching in many control laboratories. An iron ball of radius R is lifted by a
magnet with a coil of N turns and a current i around a core of length lc and cross section
A = πR2 . The vertical position of the ball is z, which is positive in the downwards
vertical direction. This coil has a magnetic circuit with magnetomotive force N i. The
flux φ flows through the iron core, then over the airgap, through the ball, and finally
along the return path through the open air as shown in Figure 3.16. This gives
108 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

i
+

Figure 3.16: Magnetic levitation experiment.

N i = φ (Ra + Rc + Rb + Rr ) (3.184)

where
z
Ra = (3.185)
Aµ0
is the reluctance of the airgap, Rc is the reluctance of the core, Rb is the reluctance of the
ball, and Rr is the reluctance of the return path through the open air. The reluctances
Rc and Rb are negligible. Moreover, if the ball is sufficiently close to the core, then the
return path through the open air will not change significantly as the ball moves, and we
may assume that Rr is a constant. Therefore we may write

z + z0
Ni = φ (3.186)
Aµ0

where z0 = Rr Aµ0 is a constant. The flux linkage is

N 2 Aµ0
λ = Nφ = i, (3.187)
z + z0

and we find that the inductance is

N 2 Aµ0
L (z) = (3.188)
z + z0

The magnetic force on the ball is found from (3.180) to be

i2 ∂L (z) µ AN 2 i2
F = =− 0 2 (3.189)
2 ∂z 2 (z + z0 )

The dynamics of the electrical circuit are given by

µ ¶
d di
u = Ri + λ̇ = Ri + L (z) i + L(z) (3.190)
dt dt
3.7. ELECTROMECHANICAL ENERGY CONVERSION 109

The model for the magnetic levitation experiment is then found from the equation of
motion and the circuit dynamics to be

1 i2
mz̈ = − AN 2 µ0 + mg (3.191)
2 (z + z0 )2
di µ0 AN 2 dz
L = −Ri + 2 i+u (3.192)
dt (z + z0 ) dt

Let zd be the constant desired position of the ball. The solution (zd , id ) is found from

1 i2d
0 = mz̈d = − AN 2 µ0 + mg (3.193)
2 (zd + z0 )2

which gives the constant current

s
2mg
id = (zd + z0 ) (3.194)
AN 2 µ0

corresponding to zd . We define ∆z = z − zd and ∆i = i − id and get the linearized model

around z = zd in the form

AN 2 µ0 i2d id
m∆z̈ = 3 ∆z − AN 2 µ0 ∆i
(zd + z0 ) (zd + z0 )2
p
2mg 2AN 2 µ0 mg
= ∆z − ∆i (3.195)
zd + z0 zd + z0

3.7.10 Voice coil

Paper cone

r
Motion
S N

Figure 3.17: Voice coil actuator for loudspeaker.

110 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

In a loudspeaker the sound is created with a voice coil motor by setting up motion in a
membrane as shown in Figure 3.17. The voice coil is an inductive motor with translational
motion. In this section we will develop a simplified model for the loudspeaker dynamics
that gives some insight into the dynamics at low frequencies (Meisel 1966), (Crandall,
Karnopp, E.F. Kurtz and Pridmore-Brown 1968). More accurate models will depend on
acoustic properties of the loudspeaker, and on the modes of vibration of the membrane.
The membrane has mass m, the air resistance is modeled as a damper with coefficient
b, and the cone suspension is modeled as a spring with stiffness k. The electrical circuit
has input voltage u, resistance R and inductance L and induced voltage e.
The flux density B in the airgap is constant and set up by a permanent magnet. The
magnetic flux from the permanent magnet through the winding is

φ = 2πrBx (3.196)

where x is the position of winding which is wound around a cylinder that is fixed to the
membrane. The flux linkage for the coil with N windings can then be written

λ (i, x) = Li + Ke x (3.197)

where L is the self inductance of the coil and Ke = 2πrN B. The magnetic coenergy is
found by integrating (3.173) from (0, 0) to (0, x), and then from (0, x) to (i, x). Under
the assumption that the force is zero when i = 0, this gives
Z i
∗ 1 2
Wm (i, x) = λ(i0 , x)di0 = Li + Ke xi (3.198)
0 2

We find the equations

F = Ke i, e = Ke ẋ (3.199)
∗
from e = λ̇ and F = ∂Wm (i, x) /∂x.
The equation of motion of the membrane and the voltage law for the circuit give the
model

mẍ + bẋ + kx = Ke i (3.200)

di
L = −Ri − Ke ẋ + u (3.201)
dt

3.7.11 Electromagnetic three-port

Electrical motors may have two electrical ports and one mechanical port. In this section
we will study the energy flow of this type of electromechanical three-port.
We consider a three-port where port one has voltage u1 = λ̇1 and current i1 , port
two has voltage u2 = λ̇2 and current i2 , while the mechanical port has effort T and flow
θ̇. The power flowing into the system is

P = iT λ̇ − T θ̇ (3.202)

where i = (i1 , i2 )T and λ = (λ1 , λ2 )T . The absolute differential of the energy Wm (λ1 , θ)
stored in the system is found from dW = P dt to be

dWm (λ, θ) = iT dλ − T dθ (3.203)

3.7. ELECTROMECHANICAL ENERGY CONVERSION 111

and by comparing this to the defining expression for the absolute differential we find that

∂Wm (λ, θ) T ∂Wm (λ, θ)

i= , T =− (3.204)
∂λ ∂θ
If we assume that the flux linkage vector is a linear function of the current vector,
then the flux linkage can be written

λ = M(θ)i (3.205)

where M(θ) is the inductance matrix. The matrix M(θ) can be shown to be positive
definite and symmetric. We may then integrate dWm (λ, θ) to get the energy function
Z λ ¡ 0 ¢T −1 1
Wm (λ, θ) = λ M (θ)dλ0 = λT M−1 (θ)λ (3.206)
0 2
The coenergy is found to be
1 1
∗
Wm (i, θ) = λT i− λT M−1 (θ)λ = iT M(θ)i − iT MT (θ)M−1 (θ)M(θ)i (3.207)
2 2
which gives
∗ 1
Wm (i, θ)= iT M(θ)i (3.208)
2
It follows that the torque is
∗
∂Wm (i, θ) 1 ∂M(θ)
T = = iT i (3.209)
∂θ 2 ∂θ

3.7.12 Electromechanical capacitive element

In this section we will study an electromechanical system with a capacitive element. The
system is described as a two-port where the input port is electrical with voltage u and
current i, and the output port is mechanical with effort F and flow ẋ. We assume that
the constitutive equations are given by

u = u(q, x) (3.210)
F = F (q, x) (3.211)

The power supplied to the two-port is

P = uq̇ − F ẋ (3.212)

The energy stored in the magnetic system is Wc (q, x). The rate of change of the stored
energy must be equal to the energy supplied due to the power P . This may be expressed

dWc (q, x)
= P ⇒ dWc (q, x) = P dt = udq − F dx (3.213)
dt
Note that dWc (q, x) is the absolute differential of the function Wc (q, x). The absolute
differential may also be written

∂Wc (q, x) ∂Wc (q, x)

dWc (q, x) = dq + dx (3.214)
∂q ∂x
112 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

Comparing the two expressions for the absolute differential we find that

∂Wc (q, x)
u(q, x) = (3.215)
∂q
∂Wc (q, x)
F (q, x) = − (3.216)
∂x

The coenergy of the system is defined by

Wc∗ (u, x) = uq − Wc (q, x) (3.217)

The absolute differential of the coenergy is

dWc∗ (u, x) = udq + qdu − dWc (q, x) = qdu + F dx (3.218)

The charge q (u, x) and the force F (u, x) of an electromechanical capacitive element with
constitutive equations (3.210, 3.211) are given from the coenergy Wc∗ (u, x) by

∂Wc∗ (u, x)
q (u, x) = (3.219)
∂u
∂Wc∗ (u, x)
F (u, x) = (3.220)
∂x

3.7.13 Electromechanical two-port with linear charge

If the charge q is linear in the voltage u so that q = C(x)u, and if the force is zero when
the charge is zero so that F (0, x) = 0, then the energy Wc (q, x) can be found by first
integrating (3.213) from (0, 0) to (0, x), and then from (0, x) to (q, x). This gives
Z x Z q Z q
q0 1 q2
Wc (q, x) = − F (0, x0 )dx0 + u(q 0 , x)dq 0 = dq 0 = (3.221)
0 0 0 C(x) 2 C(x)

The coenergy is then from (3.217) to be

1 C(x)2 u2 1
Wc∗ (v, x) = C(x)u2 − = C(x)u2 (3.222)
2 C(x) 2

An electromechanical system with linear charge equation q = C(x)u has force given by
the energy according to (3.216), which gives
µ ¶
1 d 1
F (q, x) = − q2 (3.223)
2 dx C(x)

Alternatively, the force is found from the coenergy expression (3.220) to be

1 dC(x) 2
F (v, x) = u (3.224)
2 dx
3.7. ELECTROMECHANICAL ENERGY CONVERSION 113

3.7.14 Example: Capacitive microphone

A capacitive microphone can be designed with a membrane that moves a capacitive
element so that capacitance is altered. The capacitive element is charged with a voltage
E0 , and the change in capacitance can then be picked up by measuring the capacitor
current using an amplifier with resistance R. In this way the sound that excites the
membrane can be recorded (Crandall et al. 1968).
The constitutive equation for the charge over the capacitor is q = C(x)u where the
capacitance C(x) can be described by the relation

d0
C(x) = C0 (3.225)
d0 + x
Here d0 is the nominal position of the moving plate with mass m and position x. More-
over, the force over the capacitor is zero when the charge is zero. The force over the
capacitor is then found from (3.223) to be
µ ¶
1 d d0 + x 2
F (q, x) = − q (3.226)
2 dx C0 d0

and the constitutive equations are found to be

q q2
u(q, x) = , F (q, x) = − (3.227)
C(x) 2C0 d0

The dynamic model is then

q2
mẍ + bẋ + kx + = −Fa (3.228)
2C0 d0
q (d0 + x)
Rq̇ + = u0 (3.229)
C0 d0

Here (3.228) is the equation of motion for the moving plate where b is the friction coeffi-
cient, k is the spring stiffness, the last term on the left side is the force from the capacitor,
and Fa is the acoustic excitation force due to the sound. The stationary value of this
force is zero. Equation (3.229) is the voltage law for the electric circuit with a resistor
R in series with the capacitor, and with a constant driving voltage u0 . We note that the
dynamics of the system are nonlinear. Before linearization the equilibrium points are
investigated. By setting the time derivatives of x and q to zero and eliminating q we find
the equilibrium values x0 and q0 , which are related to u0 by

q02 q0 (d0 + x0 )
kx0 + = 0, = u0 (3.230)
2C0 d0 C0 d0

Elimination of q0 by inserting the second equation of (3.230) into the first equation gives

C0 d0 u20
kx0 + 2 = 0, −d0 ≤ x0 ≤ 0 (3.231)
2 (d0 + x0 )

where the first term is the spring force of the mechanical spring and the second term can
be regarded as a nonlinear electrical spring. There are two equilibrium points where one
is stable, and the other is unstable.
114 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

Linearization and Laplace transformation of (3.228) and (3.229) around x = x0 ,

q = q0 and Fa = 0 gives
¡ 2 ¢ q0
ms + bs + k ∆x(s) + ∆q(s) = −Fa (s) (3.232)
C0 d0
q0 ∆x(s) + (C0 d0 Rs + d0 + x0 ) ∆q(s) = 0 (3.233)

By inserting ∆x from the second equation into the first equation we get
· ¸
¡ 2 ¢ q2 ∆q(s)
ms + bs + k (C0 d0 Rs + d0 + x0 ) − 0 = Fa (s) (3.234)
C0 d0 q0
The output from the microphone is the increment in voltage ∆uR over the resistor R,
which is ∆uR = R∆i = Rs∆q. The transfer function from the acoustic force to the
measurement is found from (3.234) to be

∆uR (s)(s) Rs∆q(s) Rsq0

= = q02
(3.235)
Fa (s) Fa (s) 2
(ms + bs + k) (C0 d0 Rs + d0 + x0 ) − C0 d0

For large R this expression can be approximated with

q0
∆uR (s) q0 kC0 d0
(s) = =µ ³ ´2 ¶ (3.236)
Fa (ms2 + bs + k) C0 d0 s
1 + 2ζ 0 ω0 + ωs0

where
k b
ω 20 = ζ0 = √
, (3.237)
m 2 km
We see that the measured voltage u0 is proportional to the acoustic excitation force Fa
in the frequency range up to ω 0 .

Example 48 The condition for the equilibrium point to be stable is given by

q02
k (d0 + x0 ) − ≥0 (3.238)
C0 d0
which is simplified to
d0
x0 ≥ − (3.239)
3
by inserting the equation (3.230) for the equilibrium point.

3.7.15 Piezoelectric actuator

Piezoelectric sensors or actuators are electromechanical systems where the energy con-
version is due to capacitive effects. An example of this is the axial piezoelectric stack
actuator (IEEE 1987), (Fuller, Elliott and Nilsen 1996), which is a translational capac-
itive actuator with small displacements that can exert large forces. These actuators are
of particular interest for vibration damping.
The constitutive equations for a piezoelectric actuator are given by

q (u, x) = Cp x + Cu (3.240)
Fpe (u, x) = −Kx + Cp u (3.241)
3.7. ELECTROMECHANICAL ENERGY CONVERSION 115

where x is the elongation of the piezoelectric material, Fpe is the actuator force, q is the
charge, u is the voltage, C is the capacitance, Cp is the piezoelectric constant, and K is
the mechanical stiffness. The coenergy Wc∗ (u, x) of the actuator is found by integrating
dWc∗ = qdu + Fpe dx from (0, 0) to (0, x), and then from (0, x) to (u, x). This gives
Z x Z u
∗ 0 0
Wc (u, x) = Fpe (0, x ) dx + q (u0 , x) du0 (3.242)
0
Z x Z q 0
= − Kx0 dx0 + (Cp x + Cu0 ) du0 (3.243)
0 0

The coenergy is then found to be

1 1
Wc∗ (u, x) = − Kx2 + Cp xu + Cu2 (3.244)
2 2
and it can be verified that
∂Wc∗ (u, x) ∂Wc∗ (u, x)
Fpe = , q= (3.245)
∂x ∂u
Moreover, the energy stored in the actuator is found to be
1 1
Wc (q, x) = qu − Wc∗ (u, x) = Kx2 + Cu2 (q, x) (3.246)
2 2

3.7.16 Actuator configuration

Ks

F F
Actuator force
Piezoelectric
Actuator Piezoelectric force
F pe F pe

Figure 3.18: Piezoelectric actuator pretensioned by a spring

A typical actuator configuration would be to have the actuator pretensioned by a

spring with stiffness Ks and equilibrium xs (Figure 3.18), and to input the voltage u to
the actuator. Then the actuator force will be F = Fpe − Ks (x − xs ), and the actuator is
described by the equation
F = −(K + Ks )(x − x0 ) + Cp u (3.247)
where x0 = xs Ks /(K + Ks ). In this case we view the piezoelectric actuator as a force
actuator where the actuator force F is controlled with the voltage u. Then F is seen as
an output while the deflection x is an input to the actuator (Figure 3.19). Alternatively,
the actuator can be regarded as a displacement actuator controlled by the voltage u
according to
Cp u − F
x = x0 + (3.248)
K + Ks
116 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

Voltage V - F Force
Cp
control input output

K  Ks

x0 x Deflection
- response

Figure 3.19: Piezoelectric actuator seen as a force actuator

In this case the deflection x is treated as an output while the force F on the actuator
is an input (Figure 3.20. The decision on whether to treat the piezoelectric actuator as

Voltage V 1
x Deflection
control input
Cp KK s output
x0
F Force
response

Figure 3.20: Piezoelectric actuator as a deflection actuator

a force or deflection actuator depends on how the actuator is connected to the system
to be controlled. If the actuator is connected to a spring, then deflection should be the
output of the actuator. If the actuator is connected to mass, then force should be the
actuator output.

3.8 DC motor with externally controlled field

3.8.1 Model
In this section the model for DC motor with externally controlled field is presented
(Figure 3.21). The field is assumed to be set up by a field winding. The field winding
is an electrical circuit which has a conductor that is wound around an iron core in the
stator. The rotor has a winding, which is called the armature winding, around an iron
core in the rotor. The current in the field windings in the stator is used to set up a
magnetic flux density Be through the rotor, and a Lorentz force F ∼ ia × Be on the rotor
result when a current ia is run through a conductor in the rotor windings.
Kirchhoff’s voltage law for the field circuit in the stator gives

Re ie + λ̇e = ue (3.249)

where ie is the field current, Re is the field resistance, Ne is the number of windings in
the field circuit, λe = λe (ie ) is the flux linkage of the field circuit, and ue is the field
3.8. DC MOTOR WITH EXTERNALLY CONTROLLED FIELD 117

ia
Ra La
T, m ,  m

+ +
ie TL
ua + e ea
ue Jm
- -
- Re

Figure 3.21: Armature circuit and field circuit of DC motor.

circuit voltage. The voltage law for the armature circuit in the rotor gives

d
Ra ia + La ia + ea = ua (3.250)
dt

where ia is the armature current, Ra is the armature resistance, La is the armature

inductance, ua is the armature voltage, ea is the induced voltage, ω m is the angular
velocity of the motor shaft and kf is a constant. The main feature of a DC motor is that
the motor is designed so that the field is set up so that the induced voltage

ea = ke λe ω m (3.251)

is proportional to the flux linkage λe and the shaft speed ω m . The electrical power ea ia
and the mechanical power T ω m of the electromechanical conversion unit must be equal,
and this implies that the motor torque

T = kT λe ia (3.252)

where kT = ke is a constant, is proportional to the product of the magnetizing flux

linkage λe and the armature current ia . Then the armature part of the DC motor may
be described by the two-port

d
Ra ia + La ia + ke λe ω m = ua (3.253)
dt
T = kT λe ia (3.254)

The input port variables are ua and ia and the output port has variables T and ω m .
The equation of motion for the motor shaft is

Jm ω̇ m = T − TL (3.255)

where Jm is the inertia of the motor shaft, and TL is the torque that acts on the motor
shaft from the load.
118 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

The dynamic model of a DC motor with externally controlled field is given by

λ̇e = −Re ie + ue (3.256)

d
La ia = −Ra ia − ke λe ω m + ua (3.257)
dt
Jm ω̇ m = kT λe ia − TL (3.258)
θ̇m = ωm (3.259)

The input variables to the DC motor are the field voltage ue and the armature voltage
ua . Note that the model is nonlinear due to the product λe ω m in the armature equation
and the product λe ia in the field equation.

3.8.2 Network description

The energy of the armature circuit and the motor shaft is
1 1
V = La i2a + Jω 2m (3.260)
2 2
The time derivative along solution trajectories is
dia dω m
V̇ = ia La + ω m Jm
dt dt
= ia (−Ra ia − ke λe ω m + ua ) + ω m (kT λe ia − TL )
= ia ua − ω m TL − Ra i2a (3.261)

This shows that if the load is passive, then the DC motor with input ua and output ia is
passive. Note in particular that this result does not rely on any assumptions on how the
field circuit is controlled. This means that a passive current controller like a PI controller
will give a stable armature current control independently of the field circuit.
Moreover, the energy of the field circuit is
1
Ve = Le i2e (3.262)
2
with time derivative
die
V̇e = ie Le = ie (−Re ie + ue ) (3.263)
dt
= ie ue − Re i2e (3.264)

along the solutions of the system. These calculations show that the field circuit is passive
with input ue and output ie . Moreover, it is seen that there is no energy exchange between
the field circuit and the rest of the system. This means that the field current ie can be
controlled with a passive controller of the PI type independently of how the armature
circuit is controlled.
In terms of a network description the field circuit is a passive electrical one-port
with input port with voltage ue and current ie . The dynamic model of the armature
circuit and the motor shaft can be described as in the case of a DC motor with constant
field with interconnection of three passive two-ports, namely, the armature circuit, the
electromechanical energy conversion unit, and the motor shaft.
3.8. DC MOTOR WITH EXTERNALLY CONTROLLED FIELD 119

3.8.3 DC motor with field weakening

Stationary conditions for a DC motor are found by setting the time derivatives to zero.
Then the following equations result:

Re ie = ue (3.265)
Ra ia + kf λe ω m = ua (3.266)
kT λe ia = TL . (3.267)

The armature equation (3.266) gives an expression for the stationary velocity
ua − Ra ia ua Ra TL
ωm = = − (3.268)
kf λe kf λe kT kf λ2e
where the torque equation (3.267) was used to eliminate ia . Suppose that the armature
voltage is limited by
−U ≤ ua ≤ U (3.269)
while the flux linkage is limited by

0 ≤ λe ≤ λ̄ (3.270)

Then, with a constant maximum flux linkage λe = λ̄ the maximum velocity will be
U Ra TL
ωm = − . (3.271)
kf λ̄ kT kf λ̄2

It turns out that the speed of the the motor may be increased over this value by reducing
the flux linkage λe . This is referred to as field weakening. To see this we note that for
ua = U then
∂ω m U Ra TL 2Ra TL
=− 2 +2 < 0 if λe > λmin = . (3.272)
∂λe kf λe kT kf λ3e kT U
This means that when the armature voltage ua has reached its maximum value U , an
additional increase in velocity can be achieved by weakening the field as long as the
load torque TL is sufficiently small, so that λe > λmin . Field weakening is typically
implemented so that the induced voltage ea = kf λe ω m is kept constant, that is, with a
desired field is set to (
λed = ¯ λ̄ ¯ |ea | < ea,max (3.273)
¯ ea,max ¯
¯ kf ωm ¯ otherwise
The induced voltage can be computed from the armature voltage and armature current
measurements using
dia
ea = ua − Ra ia − La (3.274)
dt
and a feedback controller can be used to achieve the desired flux linkage λed . In this case
the desired flux linkage λed is considered as an input to the system, and, assuming that
the field circuit is sufficiently fast, the model is
dia
La = −Ra ia − kf λed ω m + ua (3.275)
dt
Jm ω̇ m = kT λed ia − TL (3.276)
θ̇m = ωm (3.277)
120 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

Typically, the armature current ia is controlled with a PI controller where the armature
voltage ua is the input. Stability of the armature current control loop is ensured as the
system with input ua and output ia is passive independently of the dynamics of the field
circuit.

3.9 Dynamic model of the general AC motor

3.9.1 Introduction
DC motors with a constant field have been the usual electrical motors in control appli-
cations. The reason for this is that it is straightforward to control the armature current
of this type of motor, and as the motor torque of a DC motor is proportional to the
armature current when the field is constant, it has been possible to control the motor
torque. This is certainly an ideal situation in control applications. In contrast to this,
AC motors have mainly been used in steady-state drives where the transient performance
has not been the main objective. There is a large literature on steady-state dynamics
for AC motors. However, in the last two decades there has been a strong development in
power electronics (Mohan, Undeland and Robbins 1989), and it is now possible to achieve
servo control of AC motors (Leonhard 1996). In the context of servo control there is a
need for dynamic models of AC motors that account for the transient dynamics. The
inclusion of transient dynamics may at first be unfamiliar for readers with a background
in the steady-state dynamics of AC motors, but it is needed for servo designs for AC
motors. The dynamic models presented in the following will start with the general AC
motor with a rotor that has a circular cross section. Then, this result will be specialized
for induction motors, which are of great interest in control systems due to their rugged
and reliable design.

3.9.2 Notation
For the modeling of the general AC motor we will represent voltages, currents and flux
linkages in terms of two-dimensional coordinate vectors in the stator-fixed frame s, the
rotor-fixed frame r, and in the rotor flux frame f . The stator-fixed frame s is also referred
to as the ab frame, while the flux frame f is referred to as the dq frame. The tradition
in the literature of electrical machines is to write the coordinate vectors as complex
numbers. A voltage vector given in the coordinates of the stator frame s is written as
the complex number
us = ua + jub . (3.278)
The same vector in the flux frame is written

uf = ud + juq . (3.279)

The frame f is obtained by rotation of frame s by an angle ρ. The coordinate transfor-

mation is written
us = uf ejρ . (3.280)
The time derivative of the vector us in the s frame can be expressed by the time derivative
in the f frame according to
µ f ¶
dus d ¡ f jρ ¢ du
= u e = + j ρ̇uf ejρ (3.281)
dt dt dt
3.9. DYNAMIC MODEL OF THE GENERAL AC MOTOR 121

In the same way the vector can be given in the rotor frame r as ur .

The rotational angle from the stator to the rotor is denoted θ, so that

us = ur ejθ (3.282)

and µ ¶
dus dur
= + j θ̇ur ejθ (3.283)
dt dt

3.9.3 Dynamic model

iS


  120 °
 
i S3 i S2
i R2

i R3

i R1  S ,t 

i S1

iR

Figure 3.22: General AC machine.

We will here develop the dynamic model of a general alternating current (AC) mo-
tor based on the presentation in (Leonhard 1996) where complex numbers are used to
represent vectors in the plane. An alternative formulation based on vector notation is
found in (Vas 1990). The results of this section will later be specialized to the dynamic
model of the induction motor. The general AC motor has a stationary part called the
stator, which is assumed to be an iron cylinder containing a concentric rotor, which is
the moving part of the motor. Both the rotor and the stator are assumed to have circular
cross sections separated by a constant airgap h. Both stator and rotor have symmetrical
three-phase windings close to the airgap. Each stator phase has NS windings, while each
rotor phase has NR windings. The magnetic permeability of the fully laminated stator
and rotor iron is assumed to be infinite, and iron losses are left out of the analysis. A mo-
tor with two poles is considered. For details on this consult (Leonhard 1996) for control
122 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

issues, and (Fitzgerald et al. 1983), which is a basic textbook on electrical machines.
The geometry of the AC motor is given in Figure 3.22. The angle θ is the rotor angle,
α is the angular coordinate of the stator, and β = α − θ is the corresponding angular
coordinate of the rotor. The currents of the stator phases 1, 2 and 3 are denoted iS1 , iS2
and iS3 , respectively. The stator currents are balanced so that

iS1 (t) + iS2 (t) + iS3 (t) = 0 (3.284)

Note that the currents iS1 (t), iS2 (t) and iS3 (t) are real, and, moreover, they are general
time functions that need not have any specified wave-form. The radial magnetomotive
force (mmf) distribution from the stator windings at the angular coordinate α is by
design sinusoidal and given by

FS (α, t) = NS [iS1 cos α + iS2 cos (α − γ) + iS3 cos (α − 2γ)] , γ = 120◦ (3.285)

From the trigonometric identity

cos(x − y) = cos x cos y + sin x sin y (3.286)

we get

FS (α, t) = NS [cos α (iS1 + iS2 cos γ + iS3 cos 2γ) (3.287)

+ sin α (iS1 + iS2 sin γ + iS3 sin 2γ)] . (3.288)

Complex notation
1 ¡ jα ¢
cos α = e + e−iα (3.289)
2
1 ¡ jα ¢
sin α = e − e−jα (3.290)
2j

is introduced, and this makes it possible to write the magnetomotive force in the form
1 £ ¤ £ ¤
FS (α, t) = NS isS (t)e−jα + is∗
S (t) e
jα
= NS Re isS (t)e−jα (3.291)
2
where the complex variable isS and its complex conjugate is∗
S are defined by

isS = iS1 + iS2 ejγ + iS3 ej2γ (3.292)

is∗
S = iS1 + iS2 e−jγ + iS3 e−j2γ . (3.293)

To explain the introduction of the complex current isS further we note that in the gen-
eration of the magnetomotive force FS (α, t) the three stator phases are rotated in the
stator by an angle γ = 120◦ relative to each other. Because of this, the influence of a
current iS2 (t) in the second phase is the same as the influence of a current iS2 (t) ejγ in
the first phase. In the same way, the influence of a current iS3 (t) in the third phase is the
same as the influence of a current iS3 (t) ej2γ in the first phase. The combined influence
of the three phases iS1 (t), iS2 (t) and iS3 (t) is therefore the same as the single complex
current isS .
The polar form of isS is written

isS = |iS | ejζ (3.294)

3.9. DYNAMIC MODEL OF THE GENERAL AC MOTOR 123

p s s
where |iS | = isS is∗
S is the magnitude of iS and ζ is the angle of iS . Using the polar form
of the stator current we may write the magnetomotive force at the angular coordinate α
as h i
1
FS (α, t) = NS |iS | ej(ζ−α) + |iS | e−j(ζ−α) = NS |iS | cos (ζ − α) . (3.295)
2
The currents of the rotor phases 1, 2 and 3 are denoted iR1 , iR2 and iR3 , respectively.
The radial magnetomotive force from the three-phase rotor windings is given at the
angular coordinate β in the rotor as

FR (β, t) = NR [iR1 cos β + iR2 cos (β − γ) + iR3 cos (β − 2γ)] (3.296)

where
β = α − θ. (3.297)
The rotor currents iR1 , iR2 and iR3 are represented by the complex variable irR and its
complex conjugate ir∗
R defined by

irR = iR1 + iR2 ejγ + iR3 ej2γ = |iR | ejξ (3.298)

ir∗
R = iR1 + iR2 e−jγ + iR3 e−j2γ = |iR | e−jξ (3.299)

which gives
1 h i
NR irR e−j(α−θ) + ir∗
FR (α, θ, t) = Re
j(α−θ)
. (3.300)
2
The total radial magnetomotive force is then given by the sum

F = FS (α, t) + FR (α, θ, t) (3.301)

As the magnetic permeability of the iron is much higher than the magnetic permeability
for air, the magnetomotive force will be over the airgaps. The corresponding magnetic
flux density at the airgap is
µ0
BS (α, θ, t) = [FS (α, t) + FR (α, θ, t)] (3.302)
2h
where µ0 is the magnetic permeability of air.
The flux linkage of stator winding 1 is
Z π2 Z ψ+ π2
1
λS1 (t) = NS cos ψ lrBS (α, θ, t) dαdψ. (3.303)
2 −π 2 ψ− π
2

Integration gives, after some work,

1 1
λS1 (t) = LS [isS (t) + is∗ s s∗
S (t)] + M [iR (t) + iR (t)] (3.304)
3 3
where
NS2 lrπµ0 NS NR lrπµ0
LS = 3 , M =3 (3.305)
8h 8h
Proceeding in the same way with the flux linkages λS2 and λS3 it is found that
1 £ ¤ 1 £s ¤
λS2 (t) = LS isS (t)e−jγ + is∗
S (t) e
jγ
+ M iR (t)e−jγ + is∗
R (t) e
jγ
3 3
1 £s −j2γ s∗ j2γ
¤ 1 £s ¤
λS3 (t) = LS iS (t)e + iS (t) e + M iR (t)e−j2γ + is∗
R (t) e
j2γ
3 3
124 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

Then, we may define the complex stator flux linkage λsS by

λsS = λS1 + λS2 ejγ + λS3 ej2γ , (3.306)
and we find that the complex stator flux linkage λsS is given by the complex stator current
isS and the complex rotor current isR according to
λsS = LS isS + M isR (3.307)
Note that the complex rotor current isR = irR ejθ is given in the stator frame s.
In the same way the rotor flux linkages λR1 , λR2 and λR3 of the rotor phases 1, 2 and
3 are represented by the complex rotor flux variable λrR defined by
λrR = λR1 + λR2 ejγ + λR3 ej2γ (3.308)
The complex rotor flux linkage λrR is given by the complex current variables irR and irS
according to
λrR = LR irR + M irS (3.309)
where
NR2 lrπµ0
LS = 3 (3.310)
8h
and the complex stator current irS = isS e−jθ is given in the rotor frame r. The stator

i S1 i R1
RS  S1  R1 RR

i S2 i R2
RS  S2  R2 RR
+ +
i S3 i R3
u S1 + + u R1
RS  S3  R3 RR
- u S2 + u R2 -
+
- u S3 u R3 -
- -

Figure 3.23: Stator and rotor circuits

and rotor circuits are shown in Figure 3.23. For each of the stator phases 1, 2 and 3 the
voltage law leads to
dλS1
RS iS1 + = uS1 (3.311)
dt
dλS2
RS iS2 + = uS2 (3.312)
dt
dλS3
RS iS3 + = uS3 (3.313)
dt
Likewise, the voltage law for each of the rotor phases gives
dλR1
RR iR1 + = uR1 (3.314)
dt
dλR2
RR iR2 + = uR2 (3.315)
dt
dλR3
RR iR3 + = uR3 (3.316)
dt
3.9. DYNAMIC MODEL OF THE GENERAL AC MOTOR 125

Define the complex stator voltage

usS = uS1 + uS2 ejγ + uS3 ej2γ (3.317)
and the complex rotor voltage
urR = uR1 + uR2 ejγ + uR3 ej2γ (3.318)

The voltage laws for stator and rotor in terms of complex variables can be written

dλsS
RS isS + = usS (3.319)
dt
r dλrR
RR iR + = urR . (3.320)
dt
where
λsS = LS isS + M isR , λrR = LR irR + M irS (3.321)

Note that the dynamics of the stator circuits are given in the stator-fixed frame s,
while the dynamics of the rotor circuits are given in the rotor fixed frame r.
The motor torque is due to the Lorentz force on the rotor, which is caused by the
rotor current flowing in the magnetic flux from the stator. The magnetic flux from the
stator is
NS µ0 h s i
BRS (β, θ, t) = iS (t)e−j(β+θ) + is∗
S (t) e
j(β+θ)
(3.322)
4h
while the current density in the rotor is given by
1 ∂FR NR £ r −jβ ¤
aR (β, t) = = −j iR e − ir∗
Re
jβ
(3.323)
2r ∂β 4r
The resulting torque is
dT = −rBRS aR lrdβ (3.324)
j M h s −j(β+θ) i£ ¤
= i e + is∗
S e
j(β+θ)
irR e−jβ − ir∗
Re
jβ
dβ (3.325)
2π 3 S
h
j M s∗ r jθ i
= i i e − isS ir∗
Re
−jθ
+ isS irR e−j(2β+θ) − is∗ r∗ j(2β+θ)
S iR e dβ (3.326)
2π 3 S R
which is integrated to
· ¸
M isS ir∗
Re
−jθ
− is∗ r jθ
S (t) iR e 2
T = = M Im [isS isR ∗ ] (3.327)
3 2j 3

The dynamic model of a general AC motor is given by

disS d ¡ r jθ ¢
RS isS + LS +M i e = usS (3.328)
dt dt R
dir d ¡ s −jθ ¢
RR irR + LR R + M i e = urR (3.329)
dt dt S
dω m 2 h ¡ ¢∗ i
J = M Im isS irR ejθ − TL (3.330)
dt 3
dθ
= ωm (3.331)
dt
126 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

The general AC motor can be described as an electromechanical three-port that is

connected to a mechanical rotary two-port. The electromechanical three-port has a stator
port with effort usS and flow isS , a rotor port with effort urR and flow irR , and a mechanical
port with effort T and flow ω m . The mechanical port is connected to the mechanical
port which has effort T and flow ω m at the input port and effort TL and flow −ω m at
the input port.

3.10 Induction motors

3.10.1 Basic dynamic model

i R1
 R1 RR

i R2
 R2 RR

i R2
 R3 RR

Figure 3.24: Rotor circuit of indiction motor.

A type of induction motor that is widely used is the squirrel cage induction motor,
which is designed so that the rotor circuits are short-circuited (Figure 3.24). The dynamic
model is obtained by setting the rotor voltage urR to zero in the general AC motor model.

The model for an induction motor with short-circuited rotor is give by

disS d ¡ r jθ ¢
RS isS + LS +M i e = usS (3.332)
dt dt R
dir
d ¡ s −jθ ¢
RR irR + LR R + M i e = 0 (3.333)
dt dt S
dω m 2 h ¡ ¢∗ i
J = M Im isS irR ejθ − TL (3.334)
dt 3
dθ
= ωm (3.335)
dt

In a network description the induction motor model is obtained from the general AC
motor model by introducing a short circuit of the rotor port.

3.10.2 Induction motor model in stator frame

An alternative model formulation that is often seen for induction motors is obtained by a
change of variables from current irR to flux linkage λsR in the stator equation (3.332) and
the rotor equation (3.333). The resulting model, which is called the ab-model is given in
the stator frame in terms of stator currents isS and rotor flux λsR . The model is derived
from (3.319) and (3.320). First it is noted that (3.320) implies that
d ¡ s −jθ ¢ jθ
RR isR + λ e e =0 (3.336)
dt R
3.10. INDUCTION MOTORS 127

where the equation is expressed in the s frame, and where it is used that urR = 0.
Differentiation gives
dλs
RR isR + R − jω m λsR = 0 (3.337)
dt
Next, insertion of (3.307) into (3.319) gives
d
RS isS + (LS isS + M isR ) = usS (3.338)
dt
From (3.309) it follows that
1
isR = irR ejθ = (λs − M isS ) (3.339)
LR R
and, combining this with (3.337), it is found that
µ s ¶ µ ¶
disR 1 dλR disS 1 s s disS
= −M = jω m λR − RR iR − M
dt LR dt dt LR dt
·µ ¶ s
¸
1 RR R R di
= jω m − λsR + M isS − M S (3.340)
LR LR LR dt

Insertion of these results into (3.338) and (3.337) gives

·µ ¶ ¸
dis M RR RR dis
RS isS + LS S + jω m − λsR + M isS − M S = usS (3.341)
dt LR LR LR dt
RR s dλs
(λR − M isS ) + R − jω m λsR = 0 (3.342)
LR dt
which is simplified to
µ ¶ µ ¶
disS RR M 2 s M RR
LS σ = − RS + iS − jω m − λsR + usS (3.343)
dt L2R LR LR
dλsR RR s RR M s
= − λ + jω m λsR + i (3.344)
dt LR R LR S
where
M2
σ =1− (3.345)
LS LR
The torque can be expressed
2 h ¡ ¢∗ i 2 M h ¡ ¢∗ i
∗
T = M Im isS irR ejθ = Im isS λrR ejθ + M isS (isS )
3 3 LR
2M h ¡ ¢ ∗
i 2M
= Im isS λrR ejθ = Im [isS λsR ∗ ] (3.346)
3 LR 3 LR
£ ∗¤
where it is used that Im isS (isS ) = 0. The equation of motion is then

dω m 2M h ¡ ¢∗ i
J = Im isS λrR ejθ − TL
dt 3 LR
The coordinate form of this model is found by inserting

isS = ia + j ib , usS = ua + jub, λsR = λa + jλb , (3.347)

128 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

This gives the ab-model for an induction motor:

dia L2R RS + M 2 RR RR κ 1
= − ia + λa + κω m λb + ua (3.348)
dt L2R LS σ LR LS σ
dib L2 RS + M 2 RR Rr κ 1
= − R 2 ib + λb − κω m λa + ub (3.349)
dt LR LS σ LR LS σ
dλa RR RR M
= − λa − ω m λb + ia (3.350)
dt LR LR
dλb RR RR M
= − λb + ω m λa + ib (3.351)
dt LR LR
Jm ω̇ m = µ (λa ib − λb ia ) − TL (3.352)

where
M2 M 2M
σ =1− , κ= , µ= (3.353)
LS LR σLS LR 3 LR

This model has no trigonometric terms.

3.10.3 Dynamic model in the flux frame

b
q iS i Sq
Rotor axis
i Sb

d
i Sd Flux axis

R
 
a
i Sa

Figure 3.25: The flux frame dq.

A model formulation that is of great practical importance is the dq-model which is

the basis for field-oriented control of induction motors. This model has strong similarities
with the dynamic model of the DC motor, and is particularly suited for the design of
servo controllers for induction motors. To arrive at the dq-model the rotor flux linkage
is written in polar form
λsR = λrR ejθ = |λR | ejρ (3.354)
where |λR | is the magnitude of the rotor flux linkage. We define the rotor flux frame f as
the frame rotated by an angle ρ relative to the stator frame s (Figure 3.25). This frame
3.10. INDUCTION MOTORS 129

is also referred to as the dq frame. Here d refers to the direct axis, which is the xf axis,
and q refers to the quadrature axis, which is the yf axis. The rotor flux linkage in the
dq frame is then purely real, which is seen from

λfR = λrR ejθ e−jρ = |λR | (3.355)

The stator current and voltage in the dq frame are written

ifS = isS e−jρ = iSd + j iSq (3.356)

ufS = usS e−jρ = uSd + j uSq (3.357)

The torque is then

2M h ³ ´∗ i 2 M
T = Im ifS λfR = |λR | iSq . (3.358)
3 LR 3 LR

This expression for the torque is very interesting as it resembles the torque expression for
a DC motor. In particular, it is seen that if the field represented by the rotor flux linkage
|λR | can be controlled to a constant, say |λR | = Λ, then the torque will be proportional
to the stator current component iSq along the quadrature axis yf . To investigate this
further we develop a model of the induction motor in the dq frame. The rotor flux linkage
can be expressed by the rotor and stator currents by rotating (3.309) to the flux frame
f , which gives
λfR = LR ifR + M ifS (3.359)
This expression is combined with the voltage law

dλrR
RR irR + =0 (3.360)
dt
for the rotor windings. In the f frame this gives
³ ´
d λfR e−j(θ−ρ) RR ³ f ´
ej(θ−ρ) + λR − M ifS = 0, (3.361)
dt LR

which leads to the following differential equation for the rotor flux linkage:

dλfR RR ³ f ´
+ j (ρ̇ − ω m ) λfR + λR − M ifS = 0. (3.362)
dt LR

The real and imaginary part of this equation gives

LR d |λR |
+ |λR | = M iSd (3.363)
RR dt
M RR iSq
ρ̇ = ω m + (3.364)
LR |λR |

Note that the second equation is singular when the rotor flux is zero, that is, for |λR | = 0.
The stator flux equation is

d
RS isS + (LS isS + M isR ) = usS (3.365)
dt
130 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

The rotor current is eliminated, which gives

· ¸
d M s
RS isS + LS σisS + λR = usS (3.366)
dt LR
where
M2
σ =1− (3.367)
LS LR
In the f frame this gives
à !
difS M dλfR M f
RS ifS + LS σ + j ρ̇ifS + + j ρ̇ λ = ufS (3.368)
dt LR dt LR R
The real and imaginary part of this equation is
diSd M d |λR |
LS σ = −RS iSd + uSd − + LS σ ρ̇iSq (3.369)
dt LR dt
diSq M
LS σ = −RS iSq + uSq − ρ̇ |λR | − LS σ ρ̇iSd (3.370)
dt LR
A reasonable assumption is that the rotor flux linkage is slowly changing compared to
the stator current, and this expression is therefore not developed further by inserting
expressions for the derivative of the rotor flux linkage.

The dq model of an induction motor is given by

diSd M d |λR |
LS σ = −RS iSd + uSd − + LS σ ρ̇iSq (3.371)
dt LR dt
diSq M
LS σ = −RS iSq + uSq − ρ̇ |λR | − LS σ ρ̇iSd (3.372)
dt LR
LR d |λR |
+ |λR | = M iSd (3.373)
RR dt
M RR iSq
ρ̇ = ω m + (3.374)
LR |λR |
Jm ω̇ m = µ |λR | iSq − TL (3.375)

We can interpret this in the following way: The stator current component iSd along
the direct axis is controlled by the voltage uSd , and is used to magnetize the motor,
that is, to control the flux linkage |λR |. The stator current component iSq along the
quadrature axis is controlled by the quadrature voltage component uSq , and is used to
control the torque Tm . This is the underlying principle of field-oriented control. In field-
oriented control the magnitude |λR | of the rotor flux is controlled to a constant value Λ
with the controllers
iSd,ref = Kid (s) [Λ − |λR | (s)] (3.376)
uSd = Kud (u) [iSd,ref − iSd ] (3.377)
Then the torque can be controlled to a desired value τ d by controlling the quadrature
current
uSq = Kuq (u) [iSq,ref − iSq ] (3.378)
where iSq,ref = τ d / (µΛ). In addition, the flux angle ρ must be calculated in some way.
Solutions to this problem are discussed in great detail in (Leonhard 1996).
3.11. LAGRANGIAN DESCRIPTION OF ELECTROMECHANICAL SYSTEMS 131

Example 49 The dq model of the induction motor can be described as a electromechan-

ical three-port with a quadrature axis port with voltage uSq and current iSq , a direct axis
port with voltage uSd and current iSd , and a mechanical rotary port with effort T and
flow ω m . The direct port is used to control the flux in the machine, while the quadrature
port is used to control the torque.

3.11 Lagrangian description of electromechanical sys-

tems
3.11.1 Generalized coordinates
The dynamics of electrical motors involves the dynamics of an electrical system and a
mechanical system. The mechanical system can be modeled with the Lagrange’s equation
of motion, and this is quite useful for analysis and controller design. It turns out that also
the electrical system can be modelled using a Lagrange formalism, and this description
can be integrated with the Lagrangian description of the mechanical system. A detailed
discussion of this is found in (Meisel 1966) and (Crandall et al. 1968), and the material
in this section is based on these two references.
Lagrangian models for mechanical systems (Chapter 8 relies on the definition of gen-
eralized coordinates qi that may be positions or angles. In addition the generalized
momenta
∂L
pi = (3.379)
∂ q̇i
are used, where L is the Lagrangian. In electrical systems there are two alternatives for
generalized coordinates. One is the electrical charge q of a capacitive element, and the
other is the flux linkage λ of an inductive element. The time derivative of the electrical
charge is a current, and the time derivative of a flux linkage is a voltage, so we may write
q̇ = i and λ̇ = u (3.380)

3.11.2 Energy and coenergy

The power input to a circuit of n elements can be written
n
X
P = ik uk = q̇k λ̇k (3.381)
k=1

where q̇k = ik is the current through element k, and λ̇k = uk is the voltage over element
k. The energy stored in the circuit is
Xn Z t
W = q̇k λ̇k dt (3.382)
k=1 t0

Suppose that circuit element k is capacitive, and that the voltage λ̇k can be given as
a function of the charge qk according to
λ̇k = λ̇k (qk ) (3.383)
Then, using q̇k dt = dqk , the energy of element k can be written
Z qk
Wck (qk ) = λ̇k (qk0 ) dqk0 (3.384)
qk (t0 )
132 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

where qk0 is the dummy variable used in the integration. The coenergy of element k is
³ ´ Z λ̇k ³ 0´ 0
∗
Wck λ̇k = qk λ̇k dλ̇k (3.385)
λ̇k (t0 )

0
where λ̇k is the dummy variable in the integration. Then the following expressions hold
∗
∂Wck (qk ) ∂Wck (λ̇k )
= λ̇k and = qk (3.386)
∂qk ∂ λ̇k
Next, suppose that circuit element i is inductive, and that the current can be written
as a function
q̇i = q̇i (λi ) (3.387)
of the flux linkage λi . Then the energy is of an inductive element is
Z λi
¡ ¢
Wmi (λi ) = q̇i λ0i dλ0i (3.388)
λi (t0 )

Define the coenergy of an inductive element by

Z q̇i
∗
Wmi (q̇i ) = λi (q̇i0 ) dq̇i0 (3.389)
q̇i (t0 )

The flux linkage and the current are then given by

∗
∂Wmi (λi ) ∂Wmi (q̇i )
= λi and = q̇i (3.390)
∂ q̇i ∂λi
³ ´
The energy Wc (q) and the coenergy Wc∗ λ̇ of the capacitive elements of the circuit
are given by X ³ ´ X ³ ´
Wc (q) = Wck (qk ) , Wc∗ λ̇ = Wck∗
λ̇k (3.391)
k k

where the summation is done over the capacitive elements of the circuit. The energy
∗
Wm (λ) and the coenergy Wm (q̇) of the circuit are given by
X X
∗ ∗
Wm (λ) = Wmi (λi ) , Wm (q̇) = Wmi (q̇i ) (3.392)
i i

where the summation is done over the inductive elements of the circuit.

3.11.3 Analogy of electrical and mechanical systems

In a translational mechanical system the effort is the force F and the flow is the velocity
ẋ. In an electrical circuit the usual analog of a mechanical system is obtained using the
voltage u as the effort variable, and the current q̇ as the flow variable. This gives electrical
analog 1 where a mass corresponds to an inductance, a viscous damper corresponds to a
resistor, and a spring corresponds to a capacitor as in Table 3.1.
An alternative electrical analog appears if the current is used as the effort variable,
and the voltage is taken as the flow variable. This gives electrical analog 2 where a mass
corresponds to a capacitor, a viscous damper corresponds to a resistor, and a spring
corresponds to an inductance as in Table 3.2.
3.11. LAGRANGIAN DESCRIPTION OF ELECTROMECHANICAL SYSTEMS 133

Translational Electrical analog 1

Effort Energy Effort Energy
1
Mass F = mẍ K= 2 mẋ
2
u = Lq̈ ∗
Wm = 12 Lq̇ 2
Damper F = B ẋ 0 u = Rq̇ 0
1 2 1 2
Spring F = Kx V = 2 Kx u = Cq Wc = 2C q

Table 3.1: Electrical analog 1.

Translational Electrical analog 2

Effort Energy Effort Energy
1 1 2
Mass F = mẍ K= 2 mẋ
2
i= C λ̈ Wc∗ = 12 C λ̇
1
Damper F = B ẋ 0 i= R λ̇ 0
1 2 1 1 2
Spring F = Kx V = 2 Kx i= Lλ Wm = 2L λ

Table 3.2: Electrical analog 2.

Suppose that electrical analog 1 is used, which means that the charge q is taken as a
generalized coordinate of the electrical system. Then the energy stored in an inductance
is analog to the kinetic energy of a mass, and the energy stored in a capacitor is the
analog of the potential energy in a spring. On the equation level there is no distinction
between mechanical components and electrical components, and therefore Lagrangian
dynamics can be formulated also for electrical systems by using the analogies that have
been pointed out.

3.11.4 The Lagrangian

With electrical analog 1 the voltages are given as functions of the charge and current.
Then the generalized coordinate vector is the charge vector q, and the Lagrangian can
be defined by
∗
Le (q, q̇) = Wm (q̇) − Wc (q) , (3.393)
which is the coenergy of the inductive elements minus the energy of the capacitive ele-
ments. When electrical analog 2 is used the currents are given as functions of the flux
linkage and the voltages. Then the generalized coordinate vector is the flux linkage vector
λ, and the Lagrangian is
³ ´ ³ ´
Le λ, λ̇ = Wc∗ λ̇ − Wm (λ) , (3.394)

which is the coenergy of the capacitive elements minus the energy of the inductive ele-
ments.
The generalized forces Qi corresponding to the generalized coordinate in the form
of an electrical charge qi is found from a power consideration. The power a generalized
force Qi is Pi = Qi q̇i , which shows that the generalized force must be the voltage over
element i, that is,
Qi = ui . (3.395)
134 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

The generalized force Λk corresponding to a generalized coordinate in the form of a

flux linkage λk is found from Pk = Λk λ̇k , and it follows that the generalized force Λk
must be the current through element k, that is
Λk = ik . (3.396)

3.11.5 Electromechanical systems

For an electromechanical system the Lagrangian L for the complete system turns out
to be the sum of the Lagrangian for the electrical system and the Lagrangian of the
mechanical system. Suppose that the charge vector qe is the generalized coordinate
vector of the electrical system. The generalized coordinate vector q and the generalized
force vector Q for the total system are then taken to be
µ ¶ µ ¶
qe Qe
q= , Q= (3.397)
qm τ
where Qe are the generalized forces corresponding to the electrical coordinates qe , and
τ are the generalized forces corresponding to the mechanical coordinates qm . We define
the kinetic energy term to be
∗
T (q, q̇) = Tm (qm , q̇m ) + Wm (q̇e , qm ) (3.398)
which is the sum of the kinetic energy of the mechanical system and the coenergy of the
inductive elements of the electrical system. The potential energy term in the Lagrangian
is defined to be
V (q) = Vm (qm ) + Wc (qe , qm ) (3.399)
which is the sum of the potential energy of the mechanical system and the potential
energy of the capacitive elements of the electrical system. Then the Lagrangian for the
total system is
L (q, q̇) = T (q, q̇) − V (q) (3.400)
The dynamics are given by
d ∂L T ∂L
− = Q (3.401)
dt ∂ q̇e ∂qe
d ∂L T ∂L
− = τ (3.402)
dt ∂ q̇m ∂qm
These expressions can be further developed into
µ ¶T
d ∂Wm ∗
(q̇e , qm ) ∂Wc (qe , qm ) T
+ =Q (3.403)
dt ∂ q̇e ∂qe
and
µ ¶T
d ∂Tm (qm , q̇m ) ∂Tm (qm , q̇m ) T ∂Wm
∗
(q̇e , qm ) T
− −
dt ∂ q̇m ∂qm ∂qm
T T
∂Wc (qe , qm ) ∂Vm (qm )
+ + =τ (3.404)
∂qm ∂qm
∗ ∗
The term ∂Wm /∂qm is the change in the electrical coenergy Wm due to a change in the
mechanical configuration qm . This is the generalized interaction force on the mechanical
system due to electromagnetic forces.
3.11. LAGRANGIAN DESCRIPTION OF ELECTROMECHANICAL SYSTEMS 135

Example 50 We consider an electrical motor with shaft angle qm and inertia Jm . The
kinetic energy of the shaft is given by Tm (q̇m ) = 12 Jm q̇m
2
, and the potential energy is
V (q) = 0. The coenergy of the inductive elements of the motor is given by

∗ 1 T
Wm (q̇e , qm ) = q̇ Me (qm ) q̇e (3.405)
2 e
∗
and the Lagrangian is L = Tm (q̇m ) + Wm (q̇e , qm ). Then the dynamics of the motor is
given by

∂Me (qm )
Me (qm ) q̈e + q̇m q̇e = u (3.406)
∂qm
1 ∂Me (qm )
Jm q̈m − q̇Te q̇e = τ (3.407)
2 ∂qm

where the input voltage u is the generalized force corresponding to the generalized co-
ordinates, and τ is the external torque acting on the motor. The coupling between the
electrical part and the mechanical part gives rise to a torque

1 T ∂Me (qm )
τm = q̇ q̇e (3.408)
2 e ∂qm

delivered from the motor, and an induced voltage

∂Me (qm )
e = q̇m q̇e (3.409)
∂qm
∗
In terms of the energy T = Tm (q̇m ) + Wm (q̇e , qm ), the coupling terms balance the energy
exchange between the electrical part and the mechanical part along the solutions of the
system. This is seen from

1 ∂Me (qm )
Ṫ = q̇m Jm q̈m + q̇Te Me (qm ) q̈e + q̇m q̇Te q̇e
2 ∂qm
= q̇m τ +q̇Te u (3.410)

and we may conclude that if the mechanical port with effort τ m and flow q̇m is terminated
with a passive mechanical one-port, then the system with input u and output q̇e is passive.

Example 51 The electrical motor in the previous example can be described as a passive
electrical two-port in series with a passive electromechanical two-port. The input port of
the electrical two-port has variables u and q̇e and output port with variables e and −q̇e .
The electromechanical two-port has input port with variables e and −q̇e and output port
with variables τ m and q̇m .

Example 52 The flux linkage for the motor in the previous examples is given by
∗
∂Wm (q̇e , qm )
λ= = Me (qm ) q̇e (3.411)
∂ q̇e

Example 53 The Lagrangian of a mechanical system is actually defined in (Meisel 1966)

and (Crandall et al. 1968) as L = T ∗ − V where T ∗ is the kinetic coenergy. This is based
on the application of Newton’s law in the form F = ṗ where p = p(v) is the momentum.
136 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

The power is P = F v = ṗv, the differential of the kinetic energy is dT = P dt = vdp, and
the kinetic energy is Z p
T (p) = v(p0 )dp0 (3.412)
0
The kinetic coenergy is then Z v
T ∗ (v) = p(v 0 )dv 0 (3.413)
0
In the linear case where p = mv, the kinetic energy and the kinetic coenergy have same
numerical values as
1 p2 1
T (p) = = mv 2 = T ∗ (v) (3.414)
2m 2
Therefore, the kinetic coenergy T ∗ (v) is referred to as the kinetic
penergy and is denoted
by T (v). However, in a relativistic setting where m(v) = m0 / 1 − (v 2 /c2 ) there is a
difference between kinetic energy and kinetic coenergy.

3.11.6 Lagrange formulation of general AC motor

In this section we will derive the model of the general AC using Lagrange’s equation of
motion in vector notation. The charge vector qe , the current vector q̇e and the voltage
vector u are given by
µ s ¶ µ s ¶ µ s ¶
qS iS uS
qe = , q̇e = and u = (3.415)
qrR irR urR
Note that the stator charge vector qsS is given in the stator frame s, and that the rotor
charge vector qrR is given in the rotor frame r. The reason for this is that the stator
charge is in the stator, while the rotor charge rotates with the rotor. If the definition
of current as the time derivative of charge is to be meaningful, it is necessary to take
the time derivative of the charge in the frame where the charge is flowing. Therefore the
currents must be defined by
d s d
isS = (q ) = q̇sS , irR = (qrR ) =: q̇rR (3.416)
dt S dt
which is in agreement with (3.415).
We consider an electrical motor with shaft angle qm and inertia Jm . The kinetic
energy of the shaft is given by Tm (q̇m ) = 12 Jm q̇m
2
. The coenergy of the inductive elements
of the motor is given by
∗ 1 T
Wm (q̇e , qm ) = q̇ Me (qm ) q̇e (3.417)
2 e
where the inductance matrix Me (qm ) is
µ ¶
LS I2 M R2 (qm )
Me (qm ) = (3.418)
M R2 (−qm ) LR I2
and µ ¶
cos qm − sin qm
R2 (qm ) = (3.419)
sin qm cos qm
is the 2 × 2 rotation matrix corresponding to a rotation by and angle qm . We note that
∂R2 (qm )
= SR2 (qm ) , Ṙ2 (qm ) = q̇m SR2 (qm ) (3.420)
∂qm
3.11. LAGRANGIAN DESCRIPTION OF ELECTROMECHANICAL SYSTEMS 137

where µ ¶
0 −1
S= (3.421)
1 0
The partial derivative of the inductance matrix with respect to the shaft angle is
µ ¶
∂Me (qm ) 0 M SR2 (qm )
P (qm ) := = (3.422)
∂qm −M SR2 (−qm ) 0
The Lagrangian is equal to the sum of kinetic energy and electrical coenergy:
∗
L = T (q̇e , qm , q̇m ) = Tm (q̇m ) + Wm (q̇e , qm ) (3.423)
Then the dynamics of the motor is given by
Me (qm ) q̈e + P (qm ) q̇m q̇e + Re q̇e = u (3.424)
1 T
J q̈m − q̇e P (qm ) q̇e = −τ L (3.425)
2
Example 54 The dynamic equations for the electrical part can be written out as
disS dir
LS + M R2 (qm ) S + M q̇m SR2 (qm ) irR + RS isS = usS (3.426)
dt dt
disS dirR
M R2 (−qm ) + LR − M q̇m SR2 (−qm ) isS + RR irR = urR (3.427)
dt dt
This leads to the result
dis d
LS S + M [R2 (qm ) irR ] + RS isS = usS (3.428)
dt dt
dir d
LR R + M [R2 (−qm ) isS ] + RR irR = urR (3.429)
dt dt
dq̇m
J = τm − τL (3.430)
dt
where the motor torque is given by
T
τ m = M (isS ) SR2 (qm ) irR (3.431)
This result corresponds to the model (3.328—3.331) of the general AC motor except for
the factor 2/3 in the torque expression which is due to the three phase assumption used
in the previous derivation.
Example 55 The time derivative of the energy function T for solutions of the general
AC motor is given by
∂T ∂T
Ṫ = q̇ + q̈
∂q ∂ q̇
µ ¶T µ ¶T
0 Me (qm ) q̇e
= 1 T q̇+ q̈
2 q̇e P (qm ) q̇e J q̇m
1 T
= q̇ P (qm ) q̇e q̇m + q̇Te Me (qm ) q̈e +q̇m J q̈m
2 e
1 T
= q̇ P (qm ) q̇e q̇m + q̇Te [−P (qm ) q̇m q̇e − Re q̇e − ue ]
2 e · ¸
1 T
+q̇m q̇e P (qm ) q̇e − τ L
2
= q̇Te ue − q̇m τ L − q̇Te Re q̇e
= isT s rT r T
S uS + iR uR − q̇m τ L − q̇e Re q̇e
138 CHAPTER 3. ELECTROMECHANICAL SYSTEMS

3.11.7 Lagrange formulation of induction motor

The vector model of the induction motor is given by
µ ¶
usS
Me (qm ) q̈e + P (qm ) q̇m q̇e + Re q̇e = (3.432)
0
1
J q̈m − q̇Te P (qm ) q̇e = −τ L (3.433)
2
The total energy is
1 T 1 2
V =q̇ Me (qm ) q̇e + J q̇m (3.434)
2 e 2
and the time derivative of the energy is
V̇ = isT s T
S uS − q̇m τ L − q̇e Re q̇e (3.435)
It follows that if the load is passive, then the stator circuit dynamics with input usS
and output isS is passive. The passivity of this model was used for controller design in
(Nicklasson 1996) and (Ortega, Loria, Nicklasson and Sira-Ramirez 1998).

3.11.8 Lagrange formulation of DC motor

In this section the model for a DC motor with external field will be derived from the
model (3.428—3.430) for the general AC motor. In a DC motor with externally controlled
field the current vectors are given by
µ ¶ µ ¶
ie 0
isS = and isR = R2 (qm ) irR = (3.436)
0 −ia
This means that the field current is along the a axis of the stator frame s, while the
armature current in the rotor is along the negative b axis. Note that the angle between
the stator current vector and the rotor current vector is 90◦ . In the same way the voltage
vectors are given by
µ ¶ µ ¶
s ue s r 0
uS = and uR = R2 (qm ) uR = (3.437)
0 −ua
The dynamics of the motor is then
disS dis
LS + M R + RS isS = usS (3.438)
dt dt
d £ ¤ d
LR R2 (−qm ) iSR + M [R2 (−qm ) isS ] + RR irR = urR (3.439)
dt dt
dq̇m T
J = M (isS ) SisR − τ L(3.440)
dt
The rotor equation is transformed to the stator frame, and the product rule of differen-
tiation is used to arrive at
diSR dis
LR − LR q̇m SiSR + M S − M q̇m SisS + RR isR = usR (3.441)
dt dt
The stator flux linkage is
µ ¶ µ ¶
Le ie λe
λsS = LS isS + M isR = = (3.442)
−M ia −M ia
3.11. LAGRANGIAN DESCRIPTION OF ELECTROMECHANICAL SYSTEMS 139

where the field flux linkage λe = LS ie . In the torque expression (3.431) the stator current
can be expressed in terms of the stator flux linkage according to
1
isS = (λs − M isR ) (3.443)
LS S
which gives the torque expression
M M
τm = (λsS )T SisR = λe ia (3.444)
LS LS

where it is used that (isS )T SisR = 0.

Then the model is found from the first row of the stator equation (3.438), the second
row of the rotor equation (3.441), and the equation of motion (3.440). This gives

die
LS + RS ie = ue (3.445)
dt
dia M
LR + q̇m λe + RR ia = ua (3.446)
dt LS
dq̇m M
J = λe ia − τ L (3.447)
dt LS
140 CHAPTER 3. ELECTROMECHANICAL SYSTEMS
Chapter 4

Hydraulic motors

4.1 Introduction
Hydraulic motors are widely used because of the low weight and small size of hydraulic
motors compared to electrical motors with the same power. Typically a hydraulic motor
can have 10 times as high power as an electrical motor of the same dimensions. Hydraulic
systems can be divided into hydrostatic and hydrodynamic systems. Hydrostatic motors
are motors that are driven by pressure work of a flowing fluid. In contrast to this,
hydrodynamic motors are driven by the exchange of momentum of a fluid that flows past
the turbine blades. We will use the term hydraulic systems for hydrostatic systems in
the following. In this chapter dynamic models for hydraulic systems will be presented
and analyzed. The main reference for the material is (Merritt 1967). We mention the
following conversion rules between commonly used physical units:

• 1 bar = 105 Pa
• 1 atm = 1.01325 · 105 Pa
• 1 psi = 1 pound/inch2 = 6897 Pa = 0.068 bar
• 1 Pa = 1 N/m2

4.2 Valves
4.2.1 Introduction
Valves are important components of hydraulic systems, and are used to control flow. In
this section background material and models for typical valves will be developed.

4.2.2 Flow through a restriction

The flow through a restriction or orifice in a valve is generally turbulent and is given by
r
2
q = Cd A ∆p (4.1)
ρ
where A is the cross section of the orifice, and ∆p is the pressure drop over the orifice,
and ρ is the density of the fluid. The discharge coefficient Cd is a constant. Under

141
142 CHAPTER 4. HYDRAULIC MOTORS

the assumption of zero loss of energy, and that the flow area is not smaller than A, the
discharge coefficient is found to be Cd = 1 from the continuity equation and Bernoulli’s
equation in Section 11.2.8. In practice, there will be some loss of energy, and the cross
section of the flow will be somewhat smaller than the cross section A. This will reduce
the discharge coefficient Cd to be in the range 0.60 — 0.65 for orifices with sharp edges,
and in the range 0.8 — 0.9 when the edges are rounded.
The Reynolds number for flow through a restriction is given by

D
Re = q (4.2)
Aν
where D is the diameter of the restriction, A is the cross sectional area of the flow, and
ν is the kinematic viscosity. For hydraulic oil the kinematic viscosity is approximately
ν ≈ 30 × 10−6 m2 /s. The flow may be assumed to be turbulent and given by (4.1) for
Reynolds numbers larger than 1000.
For a narrow restriction with low volumetric flow q the Reynolds number becomes
small. If the Reynolds number becomes less than 10 the flow may be assumed to be
laminar and given by
q = Cl ∆p (4.3)
where Cl is a constant and ∆p is the pressure drop. This is the case for leakage flows
through narrow openings, and for typical restrictions in pressure feedback channels.

When Re > 1000 the flow through a restriction will be turbulent and proportional to the
square root of the pressure difference according to (4.1). When Re < 10 the flow will be
laminar and proportional to the pressure drop as in (4.3).

Example 56 The leakage flow coefficient of laminar flow through a circular tube (Hagen-
Poiseuille flow) is (Merritt 1967):

r2
Cl = A (4.4)
8µL

where µ = νρ is the absolute viscosity

4.2.3 Regularization of turbulent orifice flow

The turbulent flow characteristic in (4.1) is often used to describe the flow through
an orifice for all Reynolds numbers. This is not physically justified, and, moreover, it
creates problems in simulations as the derivative of the characteristic (4.1) is infinite at
the origin where the flow approaches zero. As discussed in Section 4.2.2, the Reynolds
number becomes small when the flow tends to zero, and this means that the flow is
actually laminar around zero flow. On background of this it is recommended that the
valve characteristic is modified so that the flow is modeled as laminar around zero flow
and turbulent for high Reynolds numbers. In the following it is shown how this can be
done.
First it is noted that the laminar flow characteristic (4.3) can be written in the same
form as the turbulent flow characteristic (4.2) by defining a threshold constant Retr for
the Reynolds number by
C 2 DA
Retr = 2 d (4.5)
Cl µ
4.2. VALVES 143

and by expressing the laminar flow characteristic (4.3) according to

r r
Re 2
q = Cd A ∆p (4.6)
Retr ρ
This result is verified by squaring (4.6) and inserting the Reynolds number from (4.2),
which gives
C 2 2DA
q= d ∆p (4.7)
Retr µ
Then (4.3) is recovered when Retr is defined by (4.5).
This means that we should seek a flow characteristic that satisfies
 q q
 Cd Re A 2 ∆p Re ¿ Retr
Retr ρ
q= q (4.8)
 2
Cd A ρ ∆p Retr ¿ Re
To obtain a solution which is defined for all ∆p a smooth transition between the laminar
and turbulent regimes was introduced in (Ellman and Piché 1999) using
 ³ ´
 3ν Retr A ∆p 3 − ∆p ∆p ≤ ptr
4 D ptr ptr
q= q (4.9)
 2
Cd A ρ ∆p ptr ≤ ∆p
Here the threshold pressure
9 Re2tr ρν 2 1
ptr = (4.10)
8Cd2 D2
corresponds to a given threshold Retr for the Reynolds number. Assuming a circular
orifice with diameter D, we have
√ √
πD2 2 4A A π
A= ⇒ D = ⇒ = A (4.11)
4 π D 2
Moreover, we define the constant
πD2 9 Re2tr ρν 2 π
Ftr = ptr A = ptr = (4.12)
4 8Cd2 4
Then the following result has been established:
The flow through a restriction can be described by the regularized flow characteristic
 √ √ ³ ´
 3ν Retr π A A∆p 3 − A∆p A∆p ≤ Ftr
4 2 Ftr Ftr
q(A, ∆p) = q (4.13)
 2
Cd A ρ ∆p Ftr ≤ A∆p

where
9 Re2tr ρν 2 π
Ftr = (4.14)
8Cd2 4
The regularized characteristic (4.13) describes the flow as laminar according to (4.3) for
low Reynolds numbers, and turbulent as given by (4.1) for high Reynolds numbers. There
is a smooth transition between the laminar and the turbulent flow regimes.

Numerical values for hydraulic oil are according to (Ellman and Piché 1999): Retr =
1000, ρ = 900 kg/m3 , ν = 30 × 10−6 m2 /s =30 cSt, Cd = 0.6. The regularized charac-
teristic (4.13) is well suited for numerical simulation as it it physically justified, and it
eliminates the problems that are experienced with the turbulent characteristic (4.1).
144 CHAPTER 4. HYDRAULIC MOTORS

q1
qs p1
ps
qa qb
q2
p2

qc qd
qr
pr

Figure 4.1: Four-way valve

4.2.4 Four-way valve

Typical flow control valves used in hydraulic systems have four orifices, and the flow is
controlled by varying the flow areas of the orifices. This type of valve is termed a four-
way valve. In this section we will derive the flow equations for the four-way valve shown
schematically in Figure 4.1. The valve is connected to the rest of the hydraulic system
through four ports, where each port has pressure as the effort variable and volumetric
flow as the flow variable. The supply port is connected to the pressure supply with
pressure ps and flow qs , the return port is connected to the return tank with pressure
pr = 0 and flow qr , port 1 with pressure p1 and flow q1 is connected to input side of the
load, and port 2 with pressure p2 and flow q2 is connected to the output side of the load.
The volumetric flows through the orifices a, b, c and d are given by the orifice equations

q
qa = Cd Aa (xv ) ρ2 (ps − p1 )
q
qb = Cd Ab (xv ) ρ2 (ps − p2 )
q (4.15)
qc = Cd Ac (xv ) ρ2 (p1 − pr )
q
qd = Cd Ad (xv ) ρ2 (p2 − pr )

where the opening areas Aa (xv ), Ab (xv ), Ac (xv ) and Ad (xv ) of the orifices are assumed
to be functions of the spool position xv , and the turbulent flow characteristic (4.1) has
been used to keep the equations simple. The more elaborate flow model (4.13) should be
used in simulations. The port flows are related to the orifice flows through the equations

qs = qa + qb , qr = qc + qd (4.16)
q1 = qa − qc , q2 = qd − qb (4.17)
4.2. VALVES 145

q1 q2

xv

qa qb
qc
qd

qs qr

Figure 4.2: A matched and symmetric four-way valve.

4.2.5 Matched and symmetrical four-way valve

In a spool-controlled four-way valve the port openings are controlled by controlling the
spool position xv (Figure 4.2). A matched and symmetrical valve is designed so that
Aa (xv ) = Ad (xv ) = Ab (−xv ) = Ac (−xv ) (4.18)

If a matched and symmetric valve is equipped with a critical spool and rectangular
orifices, then the port areas are given by
½
0, xv ≤ 0
Aa (xv ) = Ad (xv ) = (4.19)
bxv , xv ≥ 0
½
−bxv , xv ≤ 0
Ab (xv ) = Ac (xv ) = (4.20)
0, xv ≥ 0
A matched and symmetric valve with open center spool and rectangular orifices has port
openings given by
Aa (xv ) = Ad (xv ) = b(U + xv ), |xv | ≤ U (4.21)
Ab (xv ) = Ac (xv ) = b(U − xv ), |xv | ≤ U (4.22)

4.2.6 Symmetric motor and valve with critical spool

The characteristic of a four-way valve is given by the orifice equations (4.15). These
equations can be combined into one characteristic if the valve is assumed to be matched
and symmetric, and if it is assumed that the load is symmetric in the sense that
q1 = q2 (4.23)
The symmetric load assumption (4.23) implies that the load does not accumulate fluid,
which means that compressibility effects are not accounted for. This assumption is
146 CHAPTER 4. HYDRAULIC MOTORS

therefore not consistent with the assumptions that will used in the derivation of models of
hydraulic motors in the following. However, the assumption of a matched and symmetric
valve and a symmetric load leads to a very useful transfer function model for valve
controlled hydraulic motors, and in spite of the inconsistent assumptions introduced in
the modeling, the resulting transfer function model turns out to represent the dynamics
of the system with sufficient accuracy. In this connection it is it is interesting to note that
major textbooks on hydraulic control systems like (Merritt 1967) and (Watton 1989) rely
to a great extent on the use of the symmetric load assumption (4.23) in the analysis of
control systems for valve controlled motors and cylinders.
The symmetric load assumption (4.23) together with the orifice equations (4.15) and
the matching conditions (4.19, 4.20) imply the equations

qa = qd , qb = qc (4.24)

which in turn imply that

ps + pr = p1 + p2 (4.25)

In the symmetric load case it is convenient to define the load pressure

pL = p1 − p2 (4.26)

and the load flow

1
qL = (q1 + q2 ) (4.27)
2

We then find that the pressures p1 and p2 can be expressed as

ps + pL ps − pL
p1 = , p2 = (4.28)
2 2
and the load flow can be found to be
r r
1 1
qL = Cd Aa (xv ) (ps − pL ) − Cd Ab (xv ) (ps + pL ) (4.29)
ρ ρ
If a valve with a critical spool and rectangular ports is connected to a symmetric load,
then the port areas are given by
½ ½
0 xv ≤ 0 −bxv xv ≤ 0
Aa (xv ) = , Ab (xv ) = (4.30)
bxv xv ≥ 0 0 xv ≥ 0

This leads to the following result:

The load flow of a matched and symmetric valve with a symmetric load can be expressed
by valve characteristic r
1
qL = Cd bxv (ps − sgn(xv )pL ) (4.31)
ρ

The valve characteristic (4.31) is usually written in the nondimensional form

r
qL xv pL
p = 1 − sgn(xv ) (4.32)
Cd bxv max ps /ρ xv max ps
4.2. VALVES 147

1.5

*
x =1
v

0.5
L
*
normalized flow, q

*
x =0
v
0

-0.5

-1
*
x =-1
v

-1.5
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
*
normalized load pressure, p
L

Figure 4.3: Valve characteristic

which is commonly plotted using pressure-flow curves as in (Figure 4.3), where the nor-
malized values are
∗ qL xv pL
qL = q , x∗v = , p∗L = (4.33)
Cd bxv max ps xv max ps
ρ

Valve-controlled hydraulic motors are usually designed so the load pressure pL is

limited according to |p∗L | < 23 , and in this range the pressure-flow curves are close to
linear. The valve characteristic can be linearized to give

qL = Kq xv − Kc pL (4.34)

where r
∂qL 1
Kq = = Cd b (ps − sgn(xv )pL ) (4.35)
∂xv ρ
and p
∂qL Cd bxv (1/ρ)(ps − sgn(xv )pL )
Kc = − = (4.36)
∂pL 2(ps − sgn(xv )pL )
At zero flow, zero load pressure and zero spool position, that is, at qL = 0, pL = 0
and xv = 0 the constants of linearization are
r
ps
Kq0 = Cd b (4.37)
ρ

Kc0 = 0 (4.38)
148 CHAPTER 4. HYDRAULIC MOTORS

If the system is designed so that the load pressure satisfies the condition |pL | < 23 ps , then

2
|pL | < ps ⇒ 0.577Kq0 ≤ Kq ≤ 1.29Kq0 (4.39)
3
The calculated value for Kc is not consistent with what is found in practice. A more
realistic value for the constant Kc is obtained by setting the spool in its zero position
(xv = 0) and measuring the leakage flow ql as a function of the load pressure pL . The
flow-pressure coefficient Kc0 is then found from Kc0 = ql /pL .

The valve characteristic (4.31) is only valid when a matched and symmetrical valve with
critical spool is connected to a symmetric load as defined by (4.23). If the load is not
symmetric, then the valve must be modelled with the orifice equations (4.15).

Example 57 A regularization of the characteristic (4.31) for simulation is found from

(4.13) by defining
pe := ps − sgn(xv )pL (4.40)
and inserting ∆p = p̃/2 into the expression of (4.13). This gives
 √ √ ³ ´
 3ν Retr π A Aep 3 − Aep Aep ≤ 2Ftr
4 2 2Ftr 2Ftr
qL = q (4.41)
 1
Cd Av ρ pe 2Ftr ≤ Aep

where
9 Re2tr ρν 2 π
Ftr = , A = bxv , pe = ps − sgn(xv )pL (4.42)
8Cd2 4

4.2.7 Symmetric motor and valve with open spool

A matched and symmetric valve with open spool with rectangular ports and symmetric
load gives the load flow
³ r r
qL xv ´ pL ³ xv ´ pL
p = 1+ 1− − 1− 1+ , |xv | ≤ U (4.43)
Cd bU ps /ρ U p U p

4.2.8 Flow control using pressure compensated valves

Flow control valves can be designed with an additional pressure compensation spool
that is designed to keep the pressure across the main spool constant. This is done with
hydraulic feedback interconnections in the valve as shown in Figure 4.4. Let the valve
have an input port with pressure p1 and flow q1 , and an output port with pressure p2
and flow q2 . The motion of the pressure compensation spool is controlled by a spring
with force
Fc = −Kc xc + Fc0 (4.44)
and by two compensation chambers with pressures p3 and p4 acting on the spool cross
section Ac . Chamber 3 is connected to the pressure pc through a restriction with laminar
flow constant C3 , and chamber 4 is connected to the output pressure p2 with a restriction
with laminar flow constant C4 , and is connected to the output pressure p2 on the on the
4.2. VALVES 149

pc
p3
p 1, q 1 Ac xv
xc

p4

p 2, q 2

Figure 4.4: The figure shows the mechanical design and symbols for a pressure com-
pensated valve for volume flow control. The compensation spool with position xc which
is positive in the upwards direction will be positioned so that the pressure difference
p2 − pc over the main spool is approximately constant for varying input variables q1 , p1
and output variables q2 , p2 . The position xv of the main spool may be controlled with a
electric actuator or a pilot valve.

spring side, and to the internal pressure pc on the opposite side. The input and output
flows are given by the orifice equations
r
2
q1 = Cd bc xc (p1 − pc ) (4.45)
ρ
r
2
q2 = Cd bxv (pc − p2 ) (4.46)
ρ
where xc is the position of the spool in the pressure compensation valve, and xv is the
position of the main valve.
To analyze the dynamics of the pressure correction valve we use the equation of
motion for the pressure compensation spool
mc s2 xc (s) + Kc xc (s) = Ac [p4 (s) − p3 (s)] + Fc0 (4.47)
and the pressure dynamics of the compensation chambers
V3
ṗ3 = +Ac ẋc + C3 (pc − p3 ) (4.48)
β
V4
ṗ4 = −Ac ẋc + C4 (p2 − p4 ) (4.49)
β
The Laplace transformed pressure equations are
µ ¶
V3
C3 1 + s p3 = +Ac sxc + C3 pc (4.50)
C3 β
µ ¶
V4
C4 1 + s p4 = −Ac sxc + C4 p2 (4.51)
C4 β
150 CHAPTER 4. HYDRAULIC MOTORS

Under the assumption that the time constants V3 /(C3 β) and V4 /(C4 β) are sufficiently
small, we may use the approximations

Ac
C3 p3 = +Ac sxc + C3 pc ⇒ p3 = pc + sxc (4.52)
C3
Ac
C4 p4 = −Ac sxc + C4 p2 ⇒ p4 = p2 − sxc (4.53)
C4
Insertion in the equation of motion gives

C3 + C4
mc s2 xc (s) + A2c sxc (s) + Kc xc (s) = Ac [p2 (s) − pc (s)] + Fc0 (4.54)
C3 C4
which is Laplace transformed and rearranged to
µ ¶
Fc0 Kc mc 2 A2c C3 + C4
pc (s) − p2 (s) = + − s + s + 1 xc (s)
Ac Ac Kc Kc C3 C4
µ 2 ¶
Fc0 Kc s s
= + − + 2ζ c + 1 xc (s) (4.55)
Ac Ac ω 2c ωc

where ω 2c = Kc /mc . It can be seen that for frequencies well below ω c , the compensation
spool dynamics will satisfy
Fc0 Kc xc
pc − p2 = − (4.56)
Ac Ac
It follows that if Kc /Ac is sufficiently small, then the pressure difference pc − p2 over the
main spool will be approximately constant, and the flow (4.46) through the valve can be
approximated by s
2 Fc0
q1 = q2 = Cd bxv (4.57)
ρ Ac
This means that the use of an additional pressure compensated stage in the valve, the
flow through the valve becomes proportional to the orifice area bxv of the main spool.

4.2.9 Balance valve

xv
A
p 2, q 2
p 1, q 1

Ap

pp

Figure 4.5: Balance valve: Mechanical design and symbol

4.3. MOTOR MODELS 151

Balance valves are used in heavy lifting operations to ensure that a hanging load will
not fall down if the supply pressure is lost. In a balance valve a preset spring will push
the spool towards the closed position. The inlet pressure will set up a force that will
push the spool to the open position, while a high output pressure will tend to close the
spool. In addition, a pilot pressure is used to assist in the opening of the valve. Consider
a balance valve with inlet pressure p1 , outlet pressure p2 and pilot pressure pp . The
balance valve has a spool with cross section A at the spring end and cross section Ap at
the pilot end. Define the area
Ar = A − Ap (4.58)
and the pilot ratio
Ap
R=
Ar
The spring force on the spool is F = F0 + Ke xv where F0 is the pretensioning of the
spring, Ke is the spring stiffness, and xv is the spool position. We define xv = 0 in the
closed position, while the valve is open for xv > 0. A preset pressure p0 = F0 /Ar is
defined for convenience of notation. The equation of motion for the spool is
mv ẍv = pp Ap + (p1 − p0 ) Ar − Kxv − p2 A (4.59)
where mv is the mass of the spool. For a properly selected balance valve, the spool
dynamics will be stable, and in the frequency range of the rest of the system it can be
represented by the static characteristic
Ar
xv = [p1 − p0 + Rpp − p2 (R + 1)] , 0 ≤ xv ≤ xv,max
K
It is seen that the valve will open when the input pressure p1 and the pilot pressure pp
are sufficiently high in comparison to the preset pressure p0 and the outlet pressure p2 .
The influence of the pilot pressure increases when the area ration R increases. If p1 > p2
then there will be flow in the positive direction if the spool opens. If the pressures reverse
so that p2 > p1 , then the flow is lead through the relief which can be considered to have
a flow area Ac when it is open. The resulting flow is given by
 q
 Cd xv b 2 (p1 − p2 ) p1 > p2
q1 = qρ (4.60)
 −Cd Ac 2 (p2 − p1 ) p1 < p2
ρ

Again the regularized orifice flow model (4.13) should be used in simulations.

4.3 Motor models

4.3.1 Mass balance
The compressibility effect of the working fluid is significant for hydraulic motors. This
means that the density ρ is a function of the pressure p. A customary assumption is:

The relation between the differential dρ in density and the differential dp in pressure is
given by
dρ dp
= (4.61)
ρ β
where β is the bulk modulus.
152 CHAPTER 4. HYDRAULIC MOTORS

We see that the bulk modulus β has the physical dimension of pressure. Usually a
numerical value of β = 7 × 108 Pa =7000 bar (which corresponds to 105 psi) is assumed
for the bulk modulus, although the value can change by a factor of 10.
The mass balance for a volume V is given by

d
(ρV ) = win − wout (4.62)
dt
Here win = ρqin is the mass flow and qin is the volumetric flow into the volume, while
wout = ρqout is the mass flow and qout is the volumetric flow out of the volume. The
density is assumed to be a function of time only. This leads to

ρ̇V + ρV̇ = ρ(qin − qout ) (4.63)

and insertion of the expression (4.61) leads to the following result:

The mass balance of a hydraulic volume V is

V
ṗ + V̇ = qin − qout (4.64)
β

Example 58 The differential pressure work on a volume V of constant mass m is

µ ¶
m dρ pV
pdV = pd = −pV =− dp (4.65)
ρ ρ β

This means that the stored energy in a volume V due to a pressure p is

Z p
V 0 0 1V 2
Wp = p dp = p (4.66)
0 β 2β

4.3.2 Rotational motors

rotating cylinder barrel piston shoes

and drive shaft slide on plate

q 1 ,p 1 TL

q 2 ,p 2
m

stationary
valving plate stationary cam plate
case drain line at fixed angle

Figure 4.6: Hydraulic motor

4.3. MOTOR MODELS 153

q1 q2

c em p 1 c em p 2

T m

V 1 p1 V 2 p2

C im p 1  p 2 

Figure 4.7: Rotational hydraulic motor of the single vane type with limited travel.

Rotational hydraulic motors are available in many designs, and are made with limited
travel and continuous travel (Merritt 1967). A limited travel motor (Figure 4.7) will have
a maximum rotational angle that will be slightly less than 180◦ or slightly less than 360◦ ,
while a motor with continuous travel (Figure 4.6) there is no limit on the rotational angle.
A hydraulic motor may also run as a pump. The dynamic model is the same for a motor
and pump operation.
In this section the dynamic model for a motor with limited travel will be derived. A
schematic diagram of the motor is shown in Figure 4.7. The resulting model is equal
to the model for motors of continuous travel. A motor with limited travel has one inlet
chamber and one outlet chamber. The inlet chamber has volume V1 and pressure p1 ,
and the flow into the chamber is q1 . The outlet chamber has volume V2 and pressure
p2 , and the flow out of the chamber is q2 . A motor torque is set up by the pressure
difference between the two chambers, and the motor torque drives the motor shaft. A
dynamic model for a rotational hydraulic motor can be derived from the mass balances
of chambers 1 and 2, and the equation of motion for the motor shaft. The mass balance
for the inlet and outlet chambers are
V1
V̇1 + ṗ1 = −Cim (p1 − p2 ) − Cem p1 + q1 (4.67)
β
V2
V̇2 + ṗ2 = −Cim (p2 − p1 ) − Cem p2 − q2 (4.68)
β

where Cim is the coefficient for the internal leakage and Cem is the coefficient for leakage
out of the motor. β is the bulk modulus, and V1 and V2 are the volumes of the two
chambers. The rate of change of the chamber volumes are proportional to the angular
velocity ω m of the motor:
V̇1 = −V̇2 = Dm ω m (4.69)
The constant Dm is called the displacement. The shaft angle is denoted θm .
The motor torque T is proportional to the pressure difference, and by equating the
154 CHAPTER 4. HYDRAULIC MOTORS

qL
xv
Valve pL Motor

m
J1 J2 Jn

Figure 4.8: Valve controlled motor with elastic modes in the load.

power of the motor torque with the power of the working fluid we find that
T ω m = p1 V̇1 + p2 V̇2 = (p1 − p2 ) Dm ω m (4.70)
It follows that the motor torque is
T = Dm (p1 − p2 ) (4.71)
The equation of motion for the motor shaft is therefore
Jt ω̇ m = −Bm ω m + Dm (p1 − p2 ) − TL (4.72)
where Jt is the moment of inertia of the motor, Bm is the viscous friction coefficient, and
TL is the load torque. To sum up:

The model of a rotational hydraulic motor is given by

V1
ṗ1 = −Cim (p1 − p2 ) − Cem p1 − Dm ω m + q1 (4.73)
β
V2
ṗ2 = −Cim (p2 − p1 ) − Cem p2 + Dm ω m − q2 (4.74)
β
Jt ω̇ m = −Bm ω m + Dm (p1 − p2 ) − TL (4.75)

The rotational hydraulic motor can be described with a two-port for each chamber,
and a three-port for the shaft dynamics. Chamber 1 has one port with effort p1 and flow
q1 , and one port with effort T1 = Dm p1 and flow ω m . Chamber 2 has one port with effort
p2 and flow q2 , and one port with effort T2 = Dm p2 and flow ω m . The shaft model has
one port with effort T1 and flow ω m , one port with effort −T2 and flow ω m , and one port
with effort TL and flow ω m . In terms of computation the systems can be interconnected
if the variables q1 and ω m are inputs to chamber 1, q2 and ω m are inputs to chamber 2,
and T = T1 − T2 and TL are inputs to the shaft dynamics.

4.3.3 Elastic modes in the load

In many applications there will be elastic resonances in the load. If there is one resonance,
then this can be modelled with an elastic transmission and an inertia. This can be
modelled as a mechanical two-port
J1 ω̇ 1 = TL − T1 (4.76)
θ̇1 = ω1 (4.77)
TL = D1 (ω m − ω 1 ) + K1 (θm − θ1 ) (4.78)
4.3. MOTOR MODELS 155

where the input port has been connected to the motor shaft. The inputs to the two-port
are ω m and T1 , while TL and ω 1 are outputs. We may add on any number of additional
degrees of freedom as two-ports

Ji ω̇ i = Ti−1 − Ti (4.79)
Ti = Di (ω i−1 − ω i ) + Ki (θi−1 − θi ) (4.80)

with port variables Ti−1 and ω i−1 at the input and Ti and ω i at the output.

4.3.4 Hydraulic cylinder

q1
V 1 ,p 1 V 2 ,p 2

A1 A2
q2

Figure 4.9: Symmetric hydraulic cylinder

q1
V 1 ,p 1 V 2 ,p 2
A1
A1
q2

Figure 4.10: Single-rod hydraulic piston

The model of a hydraulic cylinder, which is a linear hydraulic motor, is found in

same way as for a rotational motor. The cylinder will have an inlet chamber with
volume V1 = V10 + A1 xp and pressure p1 , and an outlet chamber with volume with
V2 = V20 − A2 xp and pressure p2 . Here V10 and V20 the chamber volumes when the
piston position xp is zero. Suppose that the piston has cross sectional area Ap , and that
the piston is connected to a rod with cross section Ar .

1. If the rod goes through both chambers as in Figure 4.9, then the cylinder is said
to be symmetric, and the areas A1 and A2 are equal and given by

A1 = A2 = Ap − Ar (4.81)

2. If the rod goes through chamber 2 but not chamber 1 as in Figure 4.10, then the
cylinder is said to have a single-rod piston and the areas are given by

A1 = Ap , A2 = Ap − Ar (4.82)
156 CHAPTER 4. HYDRAULIC MOTORS

The motor force acting on the piston will be F = A1 p1 − A2 p2 . The mass balance for
the inlet and outlet chambers and the equation of motion for the piston will then give
the model.

The dynamic model for a hydraulic cylinder is

V10 + A1 xp
ṗ1 = −Cim (p1 − p2 ) − Cem p1 − A1 ẋp + q1 (4.83)
β
V20 − A2 xp
ṗ2 = −Cim (p2 − p1 ) − Cem p2 + A2 ẋp − q2 (4.84)
β
mt ẍp = −Bp ẋp + A1 p1 − A2 p2 − FL (4.85)

Here q1 is the flow into chamber 1, q2 is the flow out of chamber 2, Cim is the
coefficient for the internal leakage and Cem is the coefficient for leakage out of the motor,
mt is the mass of the piston and load, Bp is the viscous friction coefficient, FL is the load
force.

4.4 Models for transfer function analysis

4.4.1 Matched and symmetric valve and symmetric motor
Valve controlled hydraulic motors are used for servomechanisms where high accuracy and
high bandwidth are the primary objectives. The power efficiency is moderate or low for
such systems, so that for systems where power efficiency is important it is usual to have
pump controlled hydraulic motors, which will be addressed in a later section. If the load
is assumed to be symmetric in the sense that it satisfies the symmetric load condition
(4.23), and the valve is matched and symmetric and satisfies (4.31), then it is possible
to combine the two mass balances of the motor into one single mass balance, where the
load flow qL is input and the load pressure pL is output. This is very useful in transfer
function analysis of the valve controlled motor.
We consider the motor in Figure 4.6. It is assumed that when the shaft angle is zero,
then the volumes are both equal to V0 . The volumes may then be written
V1 = V0 + Dm θm , V2 = V0 − Dm θm (4.86)
Subtraction of the mass balance (4.74) for chamber 2 from the mass balance (4.73) for
chamber 1 gives
V0 Dm θm
2Dm ω m + (ṗ1 − ṗ2 ) + (ṗ1 + ṗ2 ) = q1 + q2 − 2Cim (p1 − p2 ) − Cem (p1 − p2 ) (4.87)
β β
In this expression we have the pressures and flows of the individual chambers. It is
recalled that according to (4.25) the sum of the chamber pressures p1 and p2 are equal
to the constant supply pressure ps , and it follows that ṗ1 + ṗ2 = 0. It is then possible
to reformulate (4.87) using the load pressure pL defined in (4.26) and the load flow qL
defined in (4.27). This gives
Vt
ṗL = −Ctm pL − Dm ω m + qL (4.88)
4β
where Vt = V1 + V2 = 2V0 is the total volume and Ctm = Cim + 12 Cem is the leakage
coefficient. Combining this with the equation of motion (4.72) we get the following result:
4.4. MODELS FOR TRANSFER FUNCTION ANALYSIS 157

The model of a symmetric hydraulic motor with a matched and symmetric valve is given
by
Vt
ṗL = −Ctm pL − Dm ω m + qL (4.89)
4β
Jt ω̇ m = −Bm ω m + Dm pL − TL (4.90)

Example 59 An energy function of the motor is

1 1V 2
V = Jt ω 2m + p (4.91)
2 2 4β L

The time derivative is

Vt
V̇ = ω m Jt ω̇ m + pL ṗL
4β
= −Bm ω 2m − ω m TL + ω m Dm pL − Ctm p2L − pL Dm ω m + pL qL
= pL qL − ω m TL − Bm ω 2m − Ctm p2L (4.92)

We see that is the load dynamics is passive, then the system with input qL and output pL
is passive. However, this does not have much relevance for the controller design for this
system.

4.4.2 Valve controlled motor: Transfer function

TL

xv 4 pL T -
1 m m
Kq Vt
Dm Jt
- - -

K ce Bm

Dm

Figure 4.11: Valve controlled hydraulic motor.

A linearized dynamic model for a valve controlled motor is found by inserting the
linearized valve characteristic (4.34) into the model (4.89, 4.90). The result is

Vt
ṗL = −Kce pL − Dm ω m + Kq xv (4.93)
4β
Jt ω̇ m = −Bm ω m + Dm pL − TL (4.94)
θ̇m = ωm (4.95)

where Kce = Kc + Ctm is the leakage coefficient for motor and valve, Bm is the viscous
friction coefficient, while θm is the angle of rotation of the motor shaft. A block diagram
is shown in Figure 4.11. Note the similarity to Figure 3.7.
158 CHAPTER 4. HYDRAULIC MOTORS

The Laplace transformed model is found by Laplace transformation of the model

(4.93, 4.94). This gives
µ ¶
Vt
Kce 1 + s pL = (−Dm sθm + Kq xv ) (4.96)
4βKce
¡ 2 ¢
Jt s + Bm s θm = Dm pL − TL (4.97)
Insertion of the mass balance (4.96) into the equation of motion (4.97) gives
µ ¶
Vt ¡ ¢
Kce 1 + s Jt s2 + Bm s θm = −Dm 2
sθm + Dm Kq xv
4βKce
µ ¶
Vt
−Kce 1 + s TL (4.98)
4βKce
which can be rearranged as
³ ´
Kq Kce Vt
Dm xv (s) − 2
Dm 1+ 4βKce s TL (s)
θm (s) = h ³ ´ ³ ´i (4.99)
Vt Jt 2 Kce Jt Bm Vt
s 4βDm 2 s + 2
Dm + 2
4βDm s + 1 + BmD
Kce
2
m

Under the assumption that Bm = 0 the standard formulation of this expression is ob-
tained:

The Laplace transformed model of a symmetric hydraulic motor with matched and sym-
metric valve and Bm = 0 is given by
³ ´
Kq Kce s
Dm xv (s) − 2
Dm 1 + ω t TL (s)
θm (s) = ³ ´ (4.100)
2
s 1 + 2ζ h ωsh + ωs 2
h

where ω h the hydraulic undamped natural frequency, ζ h is the the relative damping, and
ω t is the break frequency of the pressure dynamics defined by
2
r
2 4βDm Kce βJt 4βKce
ωh = , ζh = , ωt = (4.101)
Vt Jt Dm Vt Vt

We note that
4βKce
2ζ h ω h =
= ωt (4.102)
Vt
The transfer function from the spool position xv to the shaft angle θm is given by
Kq
θm Dm
Hm (s) = (s) = ³ ´ (4.103)
xv s 1 + 2ζ h ωsh + s2
ω 2h

The magnitude of the frequency response Hm (jω) is shown in Figure 4.12 with the
parameters Kq /Dm = 40, ω h = 400 rad/s and ζ h = 0.1.
The transfer function Hm (s) has a pole in s = 0, which corresponds to the integrator
from angular velocity to the valve angle. This means that for low frequencies where
s s2
(1 + 2ζ h + 2)≈1 (4.104)
ωh ωh
4.4. MODELS FOR TRANSFER FUNCTION ANALYSIS 159

2
10

1
10

|H (jω)|
m

0 dB
0
10
amplitude

-1
10

-2
10

-3
10
0 1 2 3
10 10 10 10

ω (rad/s)

Figure 4.12: Magnitude of the frequency response from the valve spool position xv to
the motor angle θ. Numerical values are Kq /Dm = 40, ω h = 400 rad/s and ζ h = 0.1.
The dashed line is drawn at -6 dB.

the motor velocity ω m will be proportional to the spool position xv . The gain Kq /Dm of
the transfer function is the flow gain Kq divided by the displacement Dm . The displace-
ment is given by the geometry of the motor, and will be available with high accuracy.
Thus, variations in the
√ gain will only depend on the flow gain Kq . The flow gain will
√
vary with the factor ps − pL / ps , and under the usual design rule |pL | < 23 ps the flow
gain will be between 57.7 % and 129 % of the nominal value.
The hydraulic undamped natural frequency ω h is an important parameter in the
design of electrohydraulic servomechanisms. The undamped natural frequency is given
by β, Dm and Jt . The parameters Jt and Dm can be found with high accuracy, while
the bulk modulus β may vary. However, the numerical value β = 7.0 · 108 Pa (= 105 psi)
(Merritt 1967) will in many cases be reasonably accurate when the working fluid is
hydraulic oil. It turns out that the leakage coefficient Kce will be dominated by the
valve.
The transfer function to the load torque TL to the shaft angle θm is
³ ´
θm −K ce
2
Dm 1 + ωst
(s) = ³ ´ (4.105)
TL 2
s 1 + 2ζ h ωsh + ωs 2
h

4.4.3 Hydraulic motor with P controller

With a proportional controller

xv = Kp (θd − θm ) (4.106)
160 CHAPTER 4. HYDRAULIC MOTORS

the loop transfer function for a hydraulic motor is

Kv
L(s) = Kp Hm (s) = ³ ´ (4.107)
s2
s 1 + 2ζ h ωsh + ω 2h

where
Kp Kq
Kv = (4.108)
Dm
is the velocity constant of the closed loop system. The loop transfer function L(s) has
a pole in s = 0 and two complex conjugated poles as the numerical value of the relative
damping is typically in the range 0.1 < ζ h < 0.5.
An important parameter in the control design is the gain of the loop transfer function
at ω 180 , which is the frequency where the frequency response L (jω h ) has a phase of 180◦ .
A closer inspection of the loop transfer function L(s) reveals that ω 180 = ω h , and that

Kv
|L (jω h )| = , ∠L (jω h ) = −180◦ (4.109)
2ζ h ω h

Using the expression for ζ h in (4.101) we find that

Kv
|L (jω 180 )| = (4.110)
2ζ h ω h

Thus, a gain margin of ∆K = 6 dB, which occurs for |L (jω 180 )| = 1/2, is achieved with

Dm
Kv = ζ h ω h ⇒ Kp = ζ ωh (4.111)
Kq h

For the numerical values in Figure 4.12 a gain margin of 6 dB will be obtained with
Kv = ζ h ω h = 40, which corresponds to a gain of Kp = Kv Dm /Kq = 1, and it follows
that in Figure 4.12 we have L(jω) = Hm (jω) if Kp = 1. The dashed line in the figure,
which is drawn at -6 dB, will therefore indicate |L (jω 180 ) | = Kv /(2ζ h ω h ) = 0.5.
Then, from Nyquist stability theory it may be concluded that:

A rotation motor with matched and symmetric valve that is controlled with a proportional
controller xv = Kp (θd − θm ) will be stable if the velocity constant satisfies

Kp Kq Dm
Kv = ≤ 2ζ h ω h ⇒ Kp ≤ 2 ζ ωh (4.112)
Dm Kq h

A gain margin of 6 dB is achieved with

Dm
Kv = ζ h ω h ⇒ Kp = ζ ωh (4.113)
Kq h

Example 60 Suppose that the leakage coefficient Kce is determined by the valve, which is
the typical situation as the leakage in motors are usually negligible. Then if two different
motors are used with the same valve and the same fluid, the constants Kq , Kce and β
will be unchanged. It follows that the stability limit will be proportional to Vt−1 .
4.4. MODELS FOR TRANSFER FUNCTION ANALYSIS 161

Example 61 If a nonzero Bm is used, the undamped natural frequency and the relative
damping are given by µ ¶
2
2 4βDm Bm Kce
ωh = 1+ 2
Vt Jt Dm
and à r s !µ ¶− 12
Kce βJt Bm Vt Bm Kce
ζh = + 1+ 2
Dm Vt 4Dm βJt Dm
We note that in this case
µ ¶ 2
Kce Jt Bm Vt 4βDm 4βKce Bm
2ζ h ω h = 2
+ 2
= + (4.114)
Dm 4βDm Vt Jt Vt Jt

4.4.4 Symmetric cylinder with matched and symmetric valve

A symmetric cylinder, which is a cylinder with a symmetric piston, has a dynamic model
that is similar to a rotary motor. Therefore, the model of a symmetric cylinder with
matched and symmetric valve is found by combining the two mass balances and using
the equation of motion for the mass. This gives
Vt
ṗL = −Ctp pL − Ap ẋp + qL (4.115)
4β
mt ẍp = −Bp ẋp + Ap pL − FL (4.116)

The Laplace transform of a symmetric cylinder with matched and symmetric valve is
³ ´
Kq Kce s
Ap xv (s) − 2
Ap 1 + ω t FL (s)
xp (s) = ³ ´ (4.117)
2
s 1 + 2ζ h ωsh + ωs 2
h

where r
4βA2p Kce βmt 4βKce
ω 2h = , ζh = , ωt = (4.118)
Vt mt Ap Vt Vt
if it is assumed that Bp = 0.

Example 62 With a proportional controller xv = Kp (xd − xp ), the stability limit for the
gain Kp is found in the same way as for the rotation motor with matched and symmetric
valve. Moreover, to have a gain margin of 6 dB the gain should be selected as
Ap
Kp = ζ ωh (4.119)
Kq h
Example 63 A hydraulic cylinder is to be selected so that it can generate a force F0
for a given supply pressure ps , and so that the position xp of the piston can be changed
between zero and x̄p . The cross sectional area of the cylinder must then be Ap = F0 /ps ,
and the volume is found from Vt = Ap x̄p = F0 x̄p /ps . Note that the required volume can
be found if the force F0 , the stroke x̄p and the supply pressure ps is given. Suppose that
a similar installation with the same valve has volume Vs and bandwidth ω s as defined by
the crossover frequency. Then the bandwidth of the system with volume Vt will be
Vs
ωc = ωs (4.120)
Vt
162 CHAPTER 4. HYDRAULIC MOTORS

4.4.5 Pump controlled hydraulic drive with P controller

The model of a pump controlled motor is derived in Section 4.7.2. At this point we simply
state that the Laplace transformed model of a pump controlled motor with constant
motor displacement Dm and constant pump speed ω p is
³ ´
kp ω p Ct
Dm φp (s) − Dm2 1 + ωs0 TL (s)
θm (s) = ³ ´ (4.121)
2
s 1 + 2ζ h ωsh + ωs 2
h

where r
2
βDm Ct βJt βCt
ω 2h = , ζh = , ω0 = (4.122)
V0 Jt 2Dm V0 V0
It is seen that the dynamic model of a pump controlled hydraulic motor has the same
structure as a valve controlled motor. The main differences are:

• The volume V0 includes the high pressure pipe and the high pressure chamber of
the motor and pump. Only the high pressure side is considered to be driving the
motor, and because of this the volume term in ω 2h is V0 for the pump controlled
system instead of the 4Vt term which appears for the valve controlled motor.

• The gain kp ω p /Dm does not vary and can be found with high accuracy.

• The relative damping of the system may be very small compared to a valve con-
trolled motor where the main leakage is in the valve. Additional leakage may be
introduced in the system to make it less oscillatory. This will give loss of power,
but it may be necessary to achieve satisfactory performance.

The usual controller for this system is a proportional feedback from the motor shaft
angle θm :
up = Kp (θd − θm ) (4.123)
The loop transfer function L (s) is seen to be

Kv
L (s) = ³ ´ (4.124)
s2
s 1 + 2ζ h ωsh + ω2h

where Kv = Kp kp ω p /Dm is the velocity constant. The system is seen to be stable if and
only if
Kv ≤ 2ζ h ω h (4.125)
where 2ζ h ω h = ω 0 . Typically, a gain margin equal to 2 will be used, in which case the
velocity constant is set to
Dm
Kv = ζ h ω h ⇒ Kp = ζ ωh (4.126)
kp ω p h

4.4.6 Transfer functions for elastic modes

Suppose that the load is driven by the motor through an elastic transmission as shown
in Figure 4.8. We restrict our analysis to one mechanical resonance in the load, which is
the case when an inertia is connected to the motor shaft through a spring and a damper.
4.4. MODELS FOR TRANSFER FUNCTION ANALYSIS 163

Pressure
dynamics TL
Valve
xv Kq
flow D 2m 1 T - m m
G 1
Dm K ce 1 s J ts s
- 

Displacement flow

Figure 4.13: Block diagram for valve controlled motor with elastic modes in the load.

Then the transfer function from the motor torque T to the motor angle θm is given by
(3.85) to be
Js2 θm (s) = G (s) T (s) (4.127)
where ³ ´2
1 + 2ζ a ωsa + s
ωa
G (s) = ³ ´2 , ωa < ω1 (4.128)
1+ 2ζ 1 ωs1 + s
ω1

The pressure dynamics will still be given by (4.96), while the equation of motion is found
from (4.127) and T = Dm pL . This gives
µ ¶
Vt
Kce 1 + s pL = (−Dm sθm + Kq xv ) (4.129)
4βKce
Jt s2 θm = G (s) Dm pL (4.130)

Insertion of the first equation into the second gives

µ ¶
Vt ¡ ¢
Kce 1 + s Jt s2 θm = G (s) −Dm
2
sθm + Dm Kq xv (4.131)
4βKce

This is more or less the same equation as (4.98) except for the appearance of G(s), and
the transfer function is found to be
K
θm G (s) Dmq
He (s) = (s) = ³ ´ (4.132)
xv s G (s) + 2ζ h ωsh + s2
ω 2h

which reduces to the transfer function Hm (s) given by (4.103) for the rigid case if G(s) =
1. The block diagram is given in Figure 4.13.

1. In the frequency ranges ω ¿ ω a and ω À ω 1 we will have G (jω) ≈ 1 and therefore

Hs (jω) ≈ Hm (jω) . This means that in the frequency range below ω a and above
ω 1 the frequency response is the same for the rigid and the elastic case.
2. If ω 1 ¿ ω h , then
Kq
Dm
He (s) ≈ ³ ´ (4.133)
s2
s 1 + 2ζ h ωsh + ω 2h
164 CHAPTER 4. HYDRAULIC MOTORS

3. If ω h ¿ ω a , then
K
G (s) Dmq
He (s) ≈ ³ ´ (4.134)
s2
s 1 + 2ζ h ωsh + ω 2h

4.4.7 Mechanical analog

K2
K1
m Fs

x D2 x0

Figure 4.14: Mechanical analog of hydraulic motor with proportional position controller.

We consider a mechanical analog of a valve controlled motor with P controller. The

analog is modelled as in Figure 4.14 with a spring S1 of stiffness K1 in series with a
parallel interconnection of a spring S2 of stiffness K2 and a damper with coefficient D2 .
The spring S1 is connected to a mass m in position x, while the spring S2 is connected
to a moving attachment of position x0 . The force from the spring S1 on the mass is
³ ´
1 + ωs1
F1 (s) = Kp ³ ´ [x0 (s) − x (s)] (4.135)
1 + ωs2

which clearly shows that this corresponds to a PD controller with limited derivative
action, and where the constants are given by.

K1 K2 K2 K1 + K2
Kp = , ω1 = , ω2 = (4.136)
K1 + K2 D2 D2

Suppose that a mass m with position x and friction coefficient B is actuated by the force
F1 from the mechanical interconnection, and that the mass is subject to the load force
FL . Then the equation of motion will be
³ ´
1 + ωs1
(ms2 + Bs)x(s) = Kp ³ ´ [x0 (s) − x (s)] − FL (4.137)
1 + ωs2

Consider a hydraulic motor with equation of motion given by (4.98)

Kq
¡ 2 ¢ D2 Dm xv − sθm
Jt s + Bm s θm = m s − TL
Kce 1+ ωt

where
4βKce
ωt = (4.138)
Vt
4.5. HYDRAULIC TRANSMISSION LINES 165

Then, with proportional feedback xv = Kp (θ0 − θm ) this becomes

¡ 2 ¢ 2 1+ s 2
Dm Kv Dm 1
Jt s + Bm s θm = −Kv s θ m + Kv θ0 − TL (4.139)
Kce 1 + ωt Kce 1 + ωst

where Kv = Kp Kq /Dm is the velocity constant. If we introduce the variable θd defined

by
1
θ0 (s) = θd (s) (4.140)
1 + Ksv
then equation (4.139) can be written
s
¡ 2 ¢ D2 1 + Kv
Jt s + Bm s θm = Kv m [θd (s) − θm (s)] − TL
Kce 1 + ωst

and we find that the dynamics are the same as for the mechanical analog if the constants
satisfy
D2 K1 K2 K2 K1 + K2
Kv m = , Kv = , ωt = (4.141)
Kce K1 + K2 D2 D2
We may solve for the parameters of the mechanical analog, which are found to be
2 2 2
4βDm Kv 4βDm 1 4βDm
K1 = , K2 = , D2 = (4.142)
Vt ω t − Kv Vt ω t − Kv Vt
Note that the mechanical analog is passive if and only if Kv ≤ ω t , which is also the
condition for stability of the closed loop system. It is interesting to see that in the
unstable case when Kv > ω t , then the mechanical analog is no longer passive as K2
and D2 become negative for Kv > ω t . This means that the closed loop system has a
passive mechanical analog if and only if the closed loop system is stable. This result also
applies to pump-controlled motors, which have the same transfer transfer functions as
valve controlled motors with minor adjustments in the parameters.

4.5 Hydraulic transmission lines

4.5.1 Introduction
The mass balance of a volume V was found in (4.64) to be given by

V
ṗ + V̇ = qin − qout (4.143)
β
In the derivation of this equation it was assumed that the pressure would be the same
over the volume, which means that the pressure p = p(t) is a function of time only.
Pressure changes will propagate with the speed of sound, which is about c = 1000 m/s
for hydraulic oil. If the volume is reasonably small so that the pressure only propagates
less than one meter, then pressure differences in the volume will disappear after 1 ms. It
is then normally justified to assume that the pressure is the same over the volume.
However, there are systems where the spatial variations of the pressure must be taken
into account by describing the pressure as a function of position and time. If the volume
V is a pipe of length L, then the time for a pressure change to propagate through the
pipe will be T = L/c. Long pipes are used in large hydraulic installations where pipes
166 CHAPTER 4. HYDRAULIC MOTORS

A
q q  dq
, p

x x  dx

Figure 4.15: Volume element for hydraulic transmission line.

of length up to 10 m are not uncommon. Moreover, in offshore oil and gas production
pipes of several hundred meters may be used. A propagation time of T = 10 ms will
result if L = 10 m. This introduces a time delay that may be significant if bandwidths
up to 100 rad/s (16 Hz) are required. The propagation time will increase to T = 0.5 s
for a pipe with length L = 500 m. In addition to problems associated with time delays,
a severe problem with long hydraulic pipes is that pressure pulses may be reflected at
the end of the pipe. This may cause strong pressure fluctuations in the system that will
limit bandwidth, and that will increase the risk of mechanical damage to the system.
On background of this there is a need to describe the pressure and flow dynamics of
long hydraulic pipes. It will be shown that the dynamics of such systems are described
by partial differential equations in the form of the wave equation, and that this is a
special case of the theory of transmission lines. The relevant models, analysis tools and
simulation algorithms will be presented in the following. Basic references are (Goodson
and Leonard 1972), (Stecki and Davis 1986a) and (Stecki and Davis 1986b).

4.5.2 PDE Model

A hydraulic transmission line is a pipe filled with a compressible liquid which may be
water or mineral oil. The pipe is of length L and has a cross section of area A, and the
length coordinate along the pipe is denoted x. The pressure of the liquid is p(x, t), the
volumetric flow is q(x, t), the density is ρ(x, t), and the bulk modulus is β.
The dynamic model is developed in detail in Section 11.2.7. The model is found from
the mass balance and momentum balance of a differential control volume Adx where A
is the cross sectional area of the pipe and x is the length coordinate along the pipe. The
velocity along the pipe is denoted v, and the volumetric flow is q = Av̄, where v̄ is the
velocity v averaged over the cross section A. The friction force on the volume element
is F dx where F = F (q) is assumed to be a function of the volumetric flow q. Then,
assuming that the velocity v̄ is small, and that the density can be considered to be a
constant ρ0 , the following model is found from the mass balance and the momentum
balance:

The model for a hydraulic transmission line can be written as the partial differential
equations
∂p(x, t) ∂q(x, t)
= −cZ0 (4.144)
∂t ∂x
∂q(x, t) c ∂p(x, t) F [q(x, t)]
= − − (4.145)
∂t Z0 ∂x ρ0
4.5. HYDRAULIC TRANSMISSION LINES 167

where the sonic velocity c and the line impedance Z0 are defined by
s p
β ρ0 c ρ0 β
c= , Z0 = = (4.146)
ρ0 A A

4.5.3 Laplace transformed model

The PDE model (4.144, 4.145) can be Laplace transformed to give
∂q(x, s) s
= − p(x, s) (4.147)
∂x cZ0
∂p(x, s) Z0 s Z0 F [q(x, s)]
= − q(x, s) − (4.148)
∂x c cρ0
The friction force F [q(x, s)] will depend on the volumetric flow q(x, s), and different
models will result depending on the friction model that is used. It is commonly assumed
that the friction F [q(x, s)] is a linear function of q(x, s). This makes it possible to define
the propagation operator Γ(s) according to

Z0 Γ(s)2 Z0 s Z0 F [q(x, s)]

q(x, s) = q(x, s) + (4.149)
LT s c cρ0
where T = L/c is the propagation time.

The transmission line model can be written

∂q(x, s) Ts
= − p(x, s) (4.150)
∂x LZ0
∂p(x, s) Z0 Γ(s)2
= − q(x, s) (4.151)
∂x LT s
where Γ(s) is the wave propagation operator, Z0 is the line impedance, and T is the
propagation time.

To complete the model the friction model F = F [q(x, s)] must be specified so that
the wave propagation operator Γ(s) can be found from (4.149). This will be done in the
following with three different friction models.

Example 64 The equations of the transmission line model (4.150) and (4.151) can be
combined so the Laplace transformed model can be written as a wave equation in pressure
or flow as given by the two equations
∂ 2 p(x, s)
L2 − Γ2 p(x, s) = 0 (4.152)
∂x2
∂ 2 q(x, s)
L2 − Γ2 q(x, s) = 0 (4.153)
∂x2
Example 65 The series impedance X(s) and the parallel admittance Y (s) are given by

Z0 Γ(s)2 Ts
X(s) = , Y (s) = (4.154)
LT s L(s)Z0
168 CHAPTER 4. HYDRAULIC MOTORS

The characteristic impedance Zc (s) is then found to be

s
X(s) Γ(s)
Zc (s) = = Z0 (4.155)
Y (s) Ts

4.5.4 Lossless model

First it is assumed that there is no friction in the pipe, which means that F = 0. The
transmission line model becomes
∂q(x, s) s
= − p(x, s) (4.156)
∂x cZ0
∂p(x, s) Z0 s
= − q(x, s) (4.157)
∂x c
Comparison with the general case (4.151) and (4.155) shows that:

In the lossless case the propagation operator Γ(s) and the characteristic impedance Zc (s)
are given by
Γ(s) = T s, Zc (s) = Z0 (4.158)

4.5.5 Linear friction

Loss terms in the form of friction in the pipe can be modelled using the Hagen-Poiseuille
equation (White 1999) by assuming laminar flow. Then the friction force is

F = ρ0 Bq (4.159)

where the friction coefficient B is

8ν 0
B= (4.160)
r02
where r0 is the radius of the pipe, and ν 0 is the kinematic viscosity. The model (4.147,
4.148) becomes

∂q(x, s) s
= − p(x, s) (4.161)
∂x cZ0
∂p(x, s) Z0
= − (s + B) q(x, s) (4.162)
∂x c
Then the propagation operator can be found according to (4.150) to be given by

Γ2 = T 2 s(s + B) (4.163)

From this result and (4.155) it is seen that:

With linear friction the propagation operator and the characteristic impedance are
r r
s+B s+B
Γ = Ts , Zc = Z0 (4.164)
s s
4.5. HYDRAULIC TRANSMISSION LINES 169

In this the case the wave equation from the lossless case is modified to
∂2p ∂p 2
2 ∂ p(x, s)
+ B − c =0 (4.165)
∂t2 ∂t ∂x2
or, using the Laplace transform,
∂ 2 p(x, s)
L2 = T 2 s (s + B) p(x, s) (4.166)
∂x2

4.5.6 Nonlinear friction

In the case of nonlinear friction the PDE model is (Goodson and Leonard 1972)
µ ¶
∂q A ∂p µ0 ∂ 2 q 1 ∂q
= − + + (4.167)
∂t ρ0 ∂x ρ0 ∂r2 r ∂r
µ ¶
∂ρ ρ ∂q ∂v v
= − 0 − ρ0 + (4.168)
∂t A ∂x ∂r r
µ 2 ¶
∂T ∂ T 1 ∂T
= α0 + (4.169)
∂t ∂r2 r ∂r
In this case viscosity and heat transfer effects has been added to the model. Without
further explanation we state that the propagation operator is
1
Γ2 = (T s)2 ¡ ps¢ (4.170)
N r ν
where the function N is given by the two alternative expressions
2J1 (jz) I2 (z)
N (z) = 1 − = (4.171)
jzJ0 (jz) I0 (z)
Here J0 and J1 are Bessel functions of the first kind of order 0 and 1, respectively, and
I0 and I2 are modified Bessel functions of the first kind of order 0 and 1, respectively.
Note that the propagation operator is irrational. Details are found in (Goodson and
Leonard 1972).

4.5.7 Wave variables

The transmission line model given by (4.150) and (4.151) can be written
µ ¶ Ã Ts
!µ ¶
∂ q(x, s) 0 − L(s)Z 0 q(x, s)
= 2 (4.172)
∂x p(x, s) − Z0LT
Γ(s)
s 0 p(x, s)
| {z }
A

By taking the eigenvectors of the matrix A it is found that the system can be made
diagonal a change of variables.

The transmission line can be modeled with the wave variables

a(x, s) = p(x, s) + Zc (s)q(x, s) (4.173)

b(x, s) = p(x, s) − Zc (s)q(x, s) (4.174)
170 CHAPTER 4. HYDRAULIC MOTORS

Here
Γ
Zc = Z0 (4.175)
Ts
is the characteristic impedance of the transmission line. From 4.172) the model for the
wave variables is found to be
∂a(x, s) Γ
= − a(x, s) (4.176)
∂x L
∂b(x, s) Γ
= b(x, s) (4.177)
∂x L
The solutions for the wave variables are given by
³ x´
a(x, s) = exp −Γ a(0, s) (4.178)
µ L ¶
L−x
b(x, s) = exp −Γ b(L, s) (4.179)
L

In the lossless case

ρ0 c
Γ = T s; Zc = Z0 = (4.180)
A
The solutions are then
³ x ´
a(x, s) = exp − T s a(0, s) (4.181)
µ L ¶
L−x
b(x, s) = exp − T s b(L, s) (4.182)
L

It is seen that a describes a wave moving in the positive x direction, and b describes a
wave moving in the negative x direction.

The inverse relations are

1
p(x, s) = [a(x, s) + b(x, s)] (4.183)
2
1
q((x, s) = [a(x, s) − b(x, s)] (4.184)
2Zc (s)

Suppose that the transmission line is terminated with an impedance ZL (s) so that

p(L, s) = ZL (s)q(L, s) (4.185)

The boundary conditions for the wave variables are given by

a(0, s) = a1 (s) (4.186)

b(L, s) = GL (s)a(L, s) (4.187)

where
ZL (s) − Zc
GL (s) = (4.188)
ZL (s) + Zc
The transfer function from a(0, s) to b(0, s) is then found to be

b(0, s)
= exp (−2Γ) GL (s) (4.189)
a(0, s)
4.5. HYDRAULIC TRANSMISSION LINES 171

In the lossless case this gives the transfer function

b(0, s)
= exp (−2T s) GL (s) (4.190)
a(0, s)
which is a time delay of 2T multiplied multiplied with GL (s).

4.5.8 Example: Lossless pipe

In the lossless case the transfer function from the inlet pressure p(0, s) to the outlet
pressure p(L, s) is given by (1.210), but an alternative expression can be found from
(4.183) to be
p(L, s) a(L, s) + b(L, s) [1 + GL (s)] exp (−T s)
= = (4.191)
p(0, s) a(0, s) + b(0, s) [1 + exp (−2T s) GL (s)]
while the transfer function from q(0, s) to p(0, s) is given by
b(0,s)
p(0, s) a(0, s) + b(0, s) 1+ a(0,s) exp (T s) + exp (−T s) GL (s)
= Zc = Zc b(0,s)
= Zc
q(0, s) a(0, s) − b(0, s) 1− exp (T s) − exp (−T s) GL (s)
a(0,s)
ZL cosh T s + Zc sinh T s
= Zc (4.192)
Zc cosh T s + ZL sinh T s
Consider a lossless pipe which is open at the outlet x = L. Then p(L, s) = 0 and
therefore ZL (s) = 0 and GL (s) = −1. It follows that
b(0, s)
= − exp (−2T s) (4.193)
a(0, s)
while
p(L, s)
=0 (4.194)
p(0, s)
and
p(0, s)
= Zc tanh T s (4.195)
q(0, s)
Next, consider a pipe which is closed at the outlet at x = L. Then q(L, s) = 0, ZL (s) = ∞
and GL (s) = 1 The transfer functions become
b(0, s)
= exp (−2T s) (4.196)
a(0, s)
and
p(L, s) 1
= (4.197)
p(0, s) cosh(T s)
and
p(0, s) 1
= Zc (4.198)
q(0, s) tanh T s
Finally, consider impedance matching which is achieved with a restriction giving p(L, s) =
Zc q(L, s), that is with ZL = Zc . Then GL (s) = 0, and
b(0, s)
=0 (4.199)
a(0, s)
172 CHAPTER 4. HYDRAULIC MOTORS

p(L, s)
= exp (−T s) (4.200)
p(0, s)
and
p(0, s)
= Zc (4.201)
q(0, s)

4.5.9 Linear network models of transmission lines

From an input-output perspective a transmission lines can be modeled as a passive system
with one inlet port at x = 0 with pressure p1 and flow q1 into the line, and one outlet
port at x = L with pressure p2 and flow q2 out of the volume. The dynamics can then
be described by transfer functions. If the ports are connected to valves at both sides,
then the valves will normally be describe with pressures as inputs and flows as outputs.
This means that the inputs to the transmission line model will be flows, and the transfer
function should be given in impedance form
µ ¶ µ ¶
p1 (s) q1 (s)
= Z(s) (4.202)
p2 (s) −q2 (s)
If the transmission line is connected to volumes at both sides, where the volumes may
be chambers in a pump, a hydraulic motor or a cylinder, then flows will be inputs and
pressures will be outputs of the mass balance models of the volumes. This means that
pressures will be input variables at the transmission line ports, and an admittance form
µ ¶ µ ¶
q1 (s) p1 (s)
= Y(s) (4.203)
−q2 (s) p2 (s)
is is the appropriate model formulation. If the transmission line is connected to a valve
at port 1 and a volume at port 2, then the flow q1 will be the input at port 1 and the
pressure p2 will be the input variable at port 2. The the model should be formulated as
a hybrid model µ ¶ µ ¶
p1 (s) q1 (s)
= H(s) (4.204)
q2 (s) p2 (s)
The impedance model, the admittance model and the hybrid models are well suited for
analysis and simulation models.
To describe a cascade of components it may seem to be a good idea to use the cascade
form µ ¶ µ ¶
p2 (s) p1 (s)
= B(s) (4.205)
q2 (s) q1 (s)
However, the cascade form leads to an ill-conditioned formulation. This is due to the
fact that the solution of the wave equation can be described as the sum of two waves
that travel in opposite directions. The cascade form is only suited to describe solutions
that propagate from port 1 to port 2.

4.5.10 Rational approximations of transfer function models

The dynamic model of a transmission line is given by partial differential equations. Be-
cause of this, the transfer function matrices Z(s), Y(s) and H(s) will have irrational
entries. It is necessary to find rational approximations to derive simulation models based
on the transfer function description. Methods for finding rational approximations of the
transfer functions are developed in the following sections. The material is taken from
(Piché and Ellman 1996) and (Mäkinen et al. 2000). Alternative solutions are presented
in (Yang and Tobler 1991).
4.5. HYDRAULIC TRANSMISSION LINES 173

4.5.11 Rational series expansion of impedance model

The dynamic model for a hydraulic transmission line is given by partial differential equa-
tions, and the transfer functions are therefore irrational. For use in simulation models
there is a need for an approximation in the form of ordinary differential equations cor-
responding to rational transfer functions. There are several ways of doing this. In this
section we will present a rational series expansion of the irrational transfer functions.
The results are taken from (Piché and Ellman 1996), and the point of departure is the
impedance form of the transfer functions where volumetric flows are inputs, and pressures
are outputs. The impedance model is given by (1.201) as
µ ¶ µ cosh Γ 1
¶µ ¶
p1 (s) sinh Γ sinh Γ q1 (s)
= Zc (s) 1 cosh Γ (4.206)
p2 (s) sinh Γ sinh Γ
−q2 (s)
The development is simplified by a change of variables into symmetric variables ps and
qs , and antisymmetric variables pa and qa according to
1 1
qs = (q1 − q2 ) , ps = (p1 + p2 ) (4.207)
2 2
1 1
qa = (q1 + q2 ) , pa = (p1 − p2 ) (4.208)
2 2
The transfer function model of a hydraulic transmission line can be written in the
impedance form

ps (s) = Zs (s)qs (s) (4.209)

pa (s) = Za (s)qa (s) (4.210)

where the transfer functions or impedance functions are found from (4.206—4.208) to be
given by
µ ¶
Z0 Γ(s) cosh Γ(s) + 1
Zs (s) = (4.211)
Ts sinh Γ(s)
µ ¶
Z0 Γ(s) cosh Γ(s) − 1
Za (s) = (4.212)
Ts sinh Γ(s)

The transfer functions Zs (s) and Za (s) both have singularities for
sinh Γ = 0 ⇔ e2Γ = 1 ⇔ Γ = jω sk (4.213)
where the natural frequencies are
ω sk = kπ, k = 0, ±1, ±2, . . . (4.214)
Note that there are infinitely many singularities with an even spacing of π along the
imaginary axis.

The following partial fraction expansions of the impedance functions can be used to arrive
at the rational series with infinitely many terms:
∞
X ∞
X
2Z0 4Z0 Γ2 4Z0 Γ2
Zs (s) = + , Za (s) = (4.215)
Ts T s (Γ2 + ω 2sk ) T s (Γ2 + ω 2sk )
k=2,4,... k=1,3,...

where ω sk = kπ
174 CHAPTER 4. HYDRAULIC MOTORS

It is possible to develop a rational transfer function model by using a truncated model

where terms up to k = N are included.
Example 66 The partial fraction expansion is done by expanding (cosh Γ + 1)/ sinh Γ
and (cosh Γ − 1)/ sinh Γ using the formula
f (s) A1 A2 f (ai )
= + + . . . ⇒ Ai = 0 (4.216)
g(s) s − a1 s − a2 g (ai )
We can then find the coefficients of the partial fraction expansions from
½
cosh Γ + 1 e2Γ + 2eΓ + 1 0 Γ = jω sk , k = odd
= = (4.217)
cosh Γ e2Γ + 1 2 Γ = jω sk , k = even
The coefficient for the second order terms are found for even k from
2 2 4Γ
+ = 2 (4.218)
Γ + jω sk Γ − jω sk Γ + ω 2sk
In the same way we calculate
½
cosh Γ − 1 e2Γ − 2eΓ + 1 2 Γ = jω sk , k = odd
= = (4.219)
cosh Γ e2Γ + 1 0 Γ = jω sk , k = even
and find that for odd k we have
2 2 4Γ
+ = 2 (4.220)
Γ + jω sk Γ − jω sk Γ + ω 2sk

4.5.12 Rational series expansion of admittance model

The admittance for of the transmission line model is given by (1.202) as
µ ¶ µ cosh Γ 1
¶µ ¶
q1 (s) 1 sinh Γ − sinh Γ p1 (s)
= 1 cosh Γ (4.221)
−q2 (s) Zc − sinh Γ sinh Γ
p2 (s)
Again the model is simplified by using symmetric and asymmetric variables defined in
(4.207) and (4.208). Then the transfer functions become
qs (s) = Ys (s)qs (s) (4.222)
qa (s) = Ya (s)qa (s) (4.223)
where the admittances are
T s cosh Γ + 1
Ys (s) = (4.224)
Z0 Γ sinh Γ
µ ¶
T s cosh Γ − 1
Ya (s) = (4.225)
Z0 Γ sinh Γ

Using partial fraction expansion in the same way as for the impedance model we find the
following rational representation of the infinite-dimensional admittances:
∞
X ∞
X
4T s 2T s 4T s
Ys (s) = , Ya (s) = + (4.226)
Z0 (Γ2 + ω 2sk ) Z0 Γ2 Z0 (Γ2 + ω 2sk )
k=1,3,... k=2,4,...
4.5. HYDRAULIC TRANSMISSION LINES 175

4.5.13 Galerkin derivation of impedance model

An alternative and more general approach to find a rational model of the transmission line
dynamics is based the use of Galerkin’s method (Mäkinen et al. 2000). Shape functions
φk (x) are then used to express the pressure as
N
X
p̄(s, x) = Pk (s)φk (x) (4.227)
k=0

This is used in combination with the transmission line model (4.157) which is
d2 p(s)
Γ2 p(s) − L2 =0 (4.228)
dx2
The boundary conditions are supposed to be
∂p(0, s) Z0 Γ(s)2 ∂p(L, s) Z0 Γ(s)2
=− q(0, s), =− q(L, s) (4.229)
∂x LT s ∂x LT s
where q(0, s) and q(L, s) are inputs to the model.
The pressure shape functions in the lossless case with zero flow at the end-points are
µ ¶
kπ
φk (x) = cos x (4.230)
L
which is a well-established result for the wave equation. These shape functions are
orthogonal in the sense that
Z L Z L µ ¶ µ ¶
kπ jπ L
φk (x)φj (x)dx = cos x cos x dx = δ kj (4.231)
0 0 L L 2
Moreover, the derivatives of the shape functions are orthogonal and satisfies
Z L Z L µ ¶ µ ¶ 2
0 0 kπ jπ (kπ)
φk (x)φj (x)dx = sin x sin x dx = δ kj (4.232)
0 0 L L 2
These shape functions will be used as assumed modes in a Ritz approximation as in
(Mäkinen et al. 2000) to derive a rational model with Galerkin’s method. This is done
by multiplying the shape function φk (x) with the model (4.228) using p = p̄, and then
integrating over the length of the transmission line. This gives
Z L µ ¶
d2 p̄(s, x)
I := φk (x) Γp̄(s, x) − L2 dx = 0 (4.233)
0 dx2
As usual in the Galerkin approach the expression for the integral is developed using
partial integration:
Z L µ ¶
dφ (x) dp̄(s, x)
I = φk Γ2 φk (x)p̄(s, x) + L2 k dx
0 dx dx
¯L
¯
2 ∂ p̄(s, x) ¯
+ φk (x)L
∂x ¯0
Z Lµ ¶
dφ (x) dp̄(s, x)
= Γ2 φk (x)p̄(s, x) + L2 k dx
0 dx dx
z0 Γ2 L
+ [φk (L)q(L) − φk (0)q(0)] (4.234)
Ts
176 CHAPTER 4. HYDRAULIC MOTORS

Using the orthogonality of φk (x) and φ0k (x) as stated in (4.231) and (4.232) we find that

L³ 2 ´ z Γ2 L £
0 ¤
Pk (s) Γ + (kπ)2 − (−1)k q(L) + q(0) = 0 (4.235)
2 Ts
which means that the pressure coefficients are given by

2z Γ2 £ ¤
Pk (s) = ³ 0 ´ q(0) + (−1)k q(L) (4.236)
2
T s Γ2 + (kπ)

Then the impedance functions Zs (s) and Za (s) in the model

ps (s) = Zs (s)qs (s) (4.237)

pa (s) = Za (s)qa (s) (4.238)

for the symmetric variables are found to be

∞
X ∞
X
2Z0 4Z Γ2 4Z Γ2
Zs (s) = + ³ 0 ´ , Za (s) = ³ 0 ´ (4.239)
Ts 2 2 2 2
k=2,4,... T s Γ + (kπ) k=1,3,... T s Γ + (kπ)

This result is the same as the result (4.215) that was obtained by series expansion of the
transfer functions.

4.5.14 Galerkin derivation of the admittance model

The Galerkin solution when the pressures are inputs is found in a similar way as in the
case where the flows are inputs. In this case the flow is represented by shape functions
φk (x) so that
∞
X
q̄(s, x) = Qk (s)φk (x) (4.240)
k=0

and the transmission line model is given by (4.157) as

d2 q(s)
Γ2 q(s) − L2 =0 (4.241)
dx2
with boundary conditions

∂q(0, s) s ∂q(L, s) s
=− p(0, s), =− p(L, s) (4.242)
∂x cZ0 ∂x cZ0

where p(0, s) and p(L, s) are inputs to the model. The shape functions are taken to be
the orthogonal eigenfunctions of the lossless wave equation when the pressures are zero
at both ends. This gives µ ¶
kπ
φk (x) = cos x (4.243)
L
The Galerkin approach gives
Z L µ ¶
d2 q̄(x)
I := φk (x) Γ2 q̄(x) − L2 dx = 0 (4.244)
0 dx2
4.5. HYDRAULIC TRANSMISSION LINES 177

and
Z Lµ ¶
dφ (x) dq̄(x)
I = φk Γ2 φk (x)q̄(s) + L2 k dx
0 dx dx
¯L
2 ∂q(x) ¯
¯
+ φk (x)L
∂x ¯0
Z Lµ ¶
dφ (x) dq̄(x)
= Γ2 φk (x)q̄(x) + L2 k dx
0 dx dx
LT s
+ [φk (L)p(L) − φk (0)p(0)] (4.245)
z0
and due to the orthogonality of φk (x) and φ0k (x), it follows that
L³ 2 2
´ LT s £ ¤
Qk (s) Γ + (kπ) = p(0) + (−1)k p(L) (4.246)
2 z0
so that
2T s £ ¤
Qk (s) = ³ ´ p(0) + (−1)k p(L) (4.247)
2
z0 Γ2 + (kπ)
The admittance functions of the symmetric and asymmetric variables
qs (s) = Ys (s)ps (s) (4.248)
qa (s) = Ya (s)pa (s) (4.249)
are found to be
∞
X ∞
X
4T s 2T s 4T s
Ys (s) = ³ ´ , Ya (s) = 2
+ ³ ´ (4.250)
2 2 Z0 Γ 2 2
k=1,3,... Z0 Γ + (kπ) k=2,4,... Z0 Γ + (kπ)

which is the same result as the result (4.226) that was found from the transfer functions.

4.5.15 Galerkin derivation of the hybrid model

Finally the hybrid case is investigated where the pressure is given at one end, and where
flow is given at the other end. Then the model is
d2 p(s)
Γ2 p(s) − L2 =0 (4.251)
dx2
with boundary condition
∂p(L, s) Z0 Γ(s)2
p(0, s) = p1 , =− q(L, s) (4.252)
∂x LT s
where p1 and q(L, s) inputs to the model. The pressure is represented by
∞
X
p̄(s, x) = p1 + Pk (s)φk (x) (4.253)
k=1

where the shape functions

·µ ¶ ¸
1 πx
φk (x) = sin k− (4.254)
2 L
178 CHAPTER 4. HYDRAULIC MOTORS

are the orthogonal eigenfunctions of the lossless wave equation when the pressure at the
input is zero and the flow at the output is zero. The Galerkin method gives
Z L µ ¶
d2 p̄(s, x)
I := φk (x) Γp̄(s, x) − L2 dx = 0 (4.255)
0 dx2
and
Z Lµ ¶
dφ (x) dp̄(s, x)
I = φk Γ2 φk (x)p̄(s, x) + L2 k dx
0 dx dx
¯L
¯
2 ∂ p̄(s, x) ¯
+ φk (x)L
∂x ¯0
Z Lµ ¶
dφ (x) dp̄(s, x)
= Γ2 φk (x)p̄(s, x) + L2 k dx
0 dx dx
Z0 Γ2 L h i
+ (−1)k q(L) (4.256)
Ts
and orthogonality of the shape functions gives
à ·µ ¶ ¸2 !
L 2 1 Γ2 L Z0 Γ2 L £ ¤
Pk (s) Γ + k− π +¡ 1
¢ p 1 − (−1)k q(L) = 0 (4.257)
2 2 k− 2 π Ts
and the pressure coefficients are found to be
à !
2Γ2 1 k Z0
Pk (s) = − £¡ ¢ ¤2 ¡ ¢ p1 + (−1) q2 (4.258)
Γ2 + k − 12 π k − 12 π Ts

The output variables q1 (s) and p2 (s) are then found to be

∞
X
p2 (s) = p1 + (−1)k+1 Pk (s) (4.259)
k=1
∞ µ ¶
Ts X 1
q1 (s) = − k − πPk (s) (4.260)
Z0 Γ2 2
k=1

4.5.16 Rational simulation models

To find a simulation model it is necessary to develop a model with finite dimension. This
can be done by truncating the infinite series (4.215). Then the model is
ps (s) = Zs (s)qs (s) (4.261)
pa (s) = Za (s)qa (s) (4.262)
where the impedances are given by the truncated versions
N
X N
X −1
2Z0 4Z0 Γ2 4Z0 Γ2
Zs = + , Za = (4.263)
Ts T s (Γ2 + ω 2sk ) T s (Γ2 + ω 2sk )
k=2,4,... k=1,3,...

of the rational transfer functions where N terms are included. It is assumed that N
is an even number. To find a state-space formulation for this model it is necessary to
investigate the terms of Zs (s) and Za (s) closer. In the lossless case Γ = T s and
4Z0 Γ2 4Z0 T s
= 2 2 (4.264)
T s (Γ2 + ω 2sk ) T s + ω 2sk
4.5. HYDRAULIC TRANSMISSION LINES 179

which is straightforward to represent as a second-order system. Moreover, with linear

friction Γ2 = T 2 s(s + B) and we find that

4Z0 Γ2 4Z0 T (s + B)
= 2 2 (4.265)
T s (Γ2 + ω 2sk ) T s + BT 2 s + ω 2sk

which is also a second-order system.

With nonlinear friction it is necessary to introduce a rational approximation of Γ2 .
An approximation due to (Woods 1983) is

(T s)2
Γ2 = ¡ ¢−1/2 (4.266)
1 − 1 + 2 TBs

In (Piché and Ellman 1996) this approximation is used to get a rational approximation
which is accurate at the natural frequencies. This is done with

4Z0 Γ2 4T (s + B)
2 2 = 2 (4.267)
T s (Γ + ω sk ) (T s) + Bk T 2 s + ω 2k

where
1p B Bk
Bk = ω sk B + , ω k = ω sk − (4.268)
2 8 2
Then the transfer function from qs (s) to ps (s) can be expressed as a parallel intercon-
nection of an integrator and N/2 second order systems that can be given a state-space
realization. In the same way the transfer function from qa (s) to pa (s) will be a parallel
interconnection of N/2 second order systems.
Solutions computed from such a truncated model will give spurious and non-physical
oscillations in the face of discontinuities like a step change in an input. This is known
as the Gibb’s phenomenon, and resembles problems that appear with data windows in
digital signal processing. A solution to the problem (Piché and Ellman 1996) is to use a
data window to modify the truncated transfer function to
N
X N
X −1
2Z0 4Z0 Γ2 σ k 4Z0 Γ2 σ k
Zs = + , Za = (4.269)
Ts T s (Γ2 + ω 2sk ) T s (Γ2 + ω 2sk )
k=2,4,... k=1,3,...

where
sin β k ω sk
σk = , βk = (4.270)
βk N +1
are the coefficients of a Riemann window. It is also possible to use a Hann window with
σ k = (1 − cos β k )/2 or a Hamming window with σ k = 0.54 + 0.46 cos β k . Moreover, a
steady-state correction is necessary to achieve correct steady-state pressure reduction,
which is
p2 = p1 − εZ0 q (4.271)
where it is assumed that q1 = q2 = q. To obtain this steady-state result with a truncated
approximation it is sufficient to insert bN B for B in Za where
 −1
N
X −1
σk 
bN = 8 (4.272)
ω 2k
k=1,3,...
180 CHAPTER 4. HYDRAULIC MOTORS

Γ2 Zs (s)/Z0 Za (s)/Z0
P PN −1
(T s)2 2
Ts + Nk=2,4,...
4σ k T s
(T s)2 +(kπ)2
4σ k T s
k=1,3,... (T s)2 +(kπ)2
2
PN 4σ k T (s+B) PN−1 4σ k T (s+bN B)
T 2 s (s + B) Ts + k=2,4,... (T s)2 +BT 2 s+(kπ)2 k=1,3,... (T s)2 +BT 2 s+(kπ)2
2 1 2
PN 4σ k T (s+B) PN−1 4σ k T (s+bN B)
(T s) √ Ts + k=2,4,... (T s)2 +Bk T 2 s+ω 2k k=1,3,... (T s)2 +Bk T 2 s+ω2k
N (r s
ν )

Table 4.1: Rational approximations of infinite dimensional transfer function with three
different friction models.

The rational truncated models are summarized in the following Table 4.1. SIMULINK
models are available on the web (Mäkinen et al. 2000).
The constants of the models are given by

L 8ν 0 1√ B Bk
T = , B = 2 , Bk = kπB + , ω 2k = kπ − (4.273)
c r0 2 8 2
 −1
N
X −1
sin β k ω sk σk 
σk = , βk = , bN = 8 (4.274)
βk N +1 ω 2k
k=1,3,...

Numerical values for a transmission line is L = 20 m, ρ = 870 kg/m3 , c = 1400 m/s,

ν 0 = 8 × 10−5 m2 /s and r0 = 6 × 10−3 m. This corresponds to a propagation time of
T = L/c = 14 ms.

4.6 Lumped parameter model of hydraulic line

4.6.1 Introduction
The hydraulic transmission line has been described by distributed parameter models,
which are models that are formulated by partial differential equations. Transfer functions
that describe distributed parameter models of transmission lines are irrational with terms
like cosh T s, sinh T s, tanh T s and exp (−T s). We recall that transfer functions with
irrational terms are called infinite dimensional as they can be expressed by a series
expansion in the complex variable s with an infinite number of terms. For analysis and
control design it may be desirable to obtain finite-dimensional models of transmission
lines. This can be done by some numerical discretization scheme or by truncating a
series expansion of an irrational model. In this section we will follow a different path.
We will reformulate the model by describing the physics of the system with a lumped
parameter model by describing the transmission line as a series of control volumes of
finite size instead of using infinitesimal control volumes. This type of model is based
on the same assumptions that are used in the Helmholtz resonator model, and we will
therefore briefly present the Helmholtz resonator, which is the physical system in fluid
flow that is analog to a mass-spring-damper system in flexible mechanical systems.
4.6. LUMPED PARAMETER MODEL OF HYDRAULIC LINE 181

4.6.2 Helmholtz resonator model

A Helmholtz resonator consists of a volume V that is connected to a pipe of length h
and cross section A (Figure 12.4). To develop the mathematical model of the system the
following assumptions are made:

1. The velocity of the fluid in the volume is sufficiently small to assume that the
pressure p is the same over the volume.

2. The compressibility effects in the pipe are negligible, so that the volumetric flow q
is the same along the pipe.

This means that the Helmholtz resonator is modeled by a pipe with incompressible
fluid flow that is connected to a volume with compressibility effects. The mass balance
of the volume is
V
ṗ = q (4.275)
β
while the momentum balance of the pipe is

hρ0 q̇ = −Ap (4.276)

where the inlet pressure of the pipe has been set to zero. By differentiating the mass
balance (4.275) with respect to time and inserting the momentum equation (4.276) the
harmonic oscillator
p̈ + ω 2H p = 0 (4.277)
is obtained, where
Aβ Ac20
ω 2H = = (4.278)
V hρ0 Vh
is the Helmholtz frequency. Here c20 = β/ρ0 is the sonic speed corresponding to the
constant density ρ0 .

4.6.3 Model formulation

In this section a chain of Helmholtz resonators will be used to model a hydraulic trans-
mission line. Consider a hydraulic transmission line of length L and cross section A. The
model is developed by connecting Helmholtz resonators where the model of Helmholtz
resonator i is established by a mass balance

Ah
ṗi = qi−1 − qi (4.279)
β

for a volume Vh = hA with pressure pi . Here qi−1 is the volumetric flow into the volume,
qi is the volumetric flow out of the volume, and β = c20 ρ0 is the bulk modulus. In
addition, the model includes a momentum balance

hρ0 q̇i−1 = A(pi−1 − pi ) − F h (4.280)

for an incompressible fluid with density ρ0 in a pipe of length h with volumetric flow
qi−1 . Here F is the friction force per unit length. This gives the following model for
Helmholtz resonator i:
182 CHAPTER 4. HYDRAULIC MOTORS

A hydraulic transmission line can be modeled by a chain of N Helmholtz resonators with

model
c2 ρ0
ṗi = (qi−1 − qi ) (4.281)
Ah
A F
q̇i−1 = (pi−1 − pi ) − (4.282)
hρ0 ρ0

Note that when h tends to zero, then the model will converge to the transmission line
model

∂p c2 ρ0 ∂q
= − (4.283)
∂t A ∂x
∂q A ∂p F
= − − (4.284)
∂t ρ0 ∂x ρ0

which equivalent to the transmission line model (4.144, 4.145). This means that the
lumped parameter model converges to the partial differential equation model when h
tends to zero.
The model (4.281, 4.282) describes a two-port with input variables qi and pi−1 and
output variables qi−1 and pi . This means that this is a model in hybrid form. We
denote the port variables of the transmission line at x = 0 as pin and qin , while the port
variables at the line end are x = L are pout and qout .Depending on which of the port
variables that are selected for inputs and outputs the model will be in admittance form,
impedance form, or hybrid form. Equations for these three cases will be presented in the
next sections.

4.6.4 Admittance model

If the input variables to the model are the pressures pin and pout , then the transmission
line can be modeled with an admittance model. The Helmholtz resonator model (4.144,
4.145) is in hybrid form, and because of this an extra pipe of length h and volumetric flow
qN must be connected to the outlet of a chain of N Helmholtz resonators i = 1, . . . , N .
This is shown in Figure 4.16 for N = 3. The model has N volumes so that volume i

q0 q1 q2 q3
V1 V2 V3
p1 p2 p3

h h h h

Figure 4.16: A chain of interconnected Helmholtz resonators with ducts of length h

representing a transmission line in the admittance form.

is centered at xi = ih for i = 1, . . . , N , and there are (N + 1) pipes where pipe i is

centered at xi+1/2 = (i + 1/2)h for i = 0, . . . , N as shown in Figure 4.17. We note that
V = (N + 1)Vh and L = (N + 1)h, so that the number of volumes N will tend to infinity
4.6. LUMPED PARAMETER MODEL OF HYDRAULIC LINE 183

q0 q1 q2 q3

V1 V2 V3
p1 p2 p3

0 h 2h 3h 4h

Figure 4.17: Spatial discretization of a transmission line with pressure inputs at both
sides to get a description in the form of a chain of Helmholtz resonators.

when h tends to zero. The model is

c2 ρ0
ṗi = (qi−1 − qi ) , i = 1, . . . , N (4.285)
Ah
A F
q̇i−1 = (pi−1 − pi ) − , i = 1, . . . , N + 1 (4.286)
hρ0 ρ0
p0 = pin , pN+1 = pout (4.287)

4.6.5 Impedance model

Suppose that the input variables at the line ends are the volumetric flows qin and qout . In
this case the model has N volumes, but the pipe of the first resonator must be removed
to have the right input variable. This means that volume 1 is at the start of the line
and volume N is at the end of the line. Volume i is centered at xi−1/2 = (i − 1/2)h for
i = 1, . . . , N , and is connected with pipes of length h and cross section A. There are
N − 1 pipes, where pipe i is centered at xi = ih for i = 1, . . . , N − 1. We note that
V = N Vh and L = N h. The model is
c2 ρ0
ṗi = (qi−1 − qi ) , i = 1, . . . , N (4.288)
Ah
A F
q̇i−1 = (pi−1 − pi ) − , i = 2, . . . , N (4.289)
hρ0 ρ0
q0 = qin , qN = qout (4.290)

4.6.6 Hybrid model

Suppose that the input variables to the transmission line model at the inlet side is the
pressures pin , and that the input variable at the outlet side is qout . Then the model is
in hybrid form. The Helmholtz resonator model (4.144, 4.145) is also in hybrid form,
and because of this the transmission line can be represented by a chain of N Helmholtz
resonators i = 1, . . . , N with model
c2 ρ0
ṗi = (qi−1 − qi ) , i = 1, . . . , N (4.291)
Ah
A F
q̇i−1 = (pi−1 − pi ) − , i = 1, . . . , N (4.292)
hρ0 ρ0
p0 = pin , qN = qout (4.293)
184 CHAPTER 4. HYDRAULIC MOTORS

In this case volume i will be centered at xi = ih for i = 1, . . . , N , and there are N

pipes where pipe i is centered at xi+1/2 = (i + 1/2)h for i = 0, . . . , N − 1. We note that
V = (N + 1/2)Vh and L = (N + 1/2)h.

4.6.7 Natural frequencies

We may think of each Helmholtz resonator as a two-port with port variables qi−1 and
pi−1 for port 1 and port variables qi and pi on port 2. Note that qi pi has the physical
dimension power. The system can be seen as a system with input variables qi and pi−1 and
output variables equal to the states qi−1 and pi . The hybrid transfer function model is
 A s

µ ¶ hρ0 ω 2H 1+ s2
− 1s2 µ ¶
1+
qi−1 (s)  ω2 ω2  pi−1 (s)
= 1
H
hρ0 s
H
 (4.294)
pi (s) s2 A s2
−qi (s)
1+ 1+
ω2 ω2
H H

where the Helmholtz frequency ω H is given by

Ac2 c
ω 2H = ⇒ ωH = (4.295)
Vh h h
Through the discretization we have introduced Helmholtz resonator i with oscillatory
poles at s = ±jω H .

Example 67 Consider a transmission line with both ends closed. This means that the
inlet and outlet flow are given as inputs, so that an impedance model should be used.
With two volumes, that is, with L = 2h, the dynamics of the model are given by

c2 ρ0
ṗ1 = − q (4.296)
Ah
c2 ρ0
ṗ2 = q (4.297)
Ah
A
q̇ = (p1 − p2 ) (4.298)
ρ0 h

Laplace transformation of the model, and insertion of the pressure equations in the mass
flow equation leads to µ ¶
c2
s2 + 2 2 q(s) = 0 (4.299)
h
This system has undamped natural frequency
√ c √ c c
ω0 = 2 = 2 2 = 2.82 (4.300)
h L L
while the exact value for the first resonance of the partial differential equation model is
found from the period T1 = 2L/c, which gives

2π c c
ω1 = = π = 3.14 (4.301)
T1 L L

This means that the resonance frequency of this simplified model is about 11% lower than
the exact resonance frequency.
4.7. OBJECT ORIENTED SIMULATION MODELS 185

Example 68 For N = 3 volumes we have L = 3h, and the state space model of the
impedance model is

Aṗi = ω 2H (hρ0 qi−1 − hρ0 qi ) , i = 1, 2, 3 (4.302)

hρ0 q̇i−1 = (Api−1 − Api ), i = 2, 3 (4.303)
    
Ap1 0 −ω 2H 0 0 0 Ap1
 hρ0 q1   0   hρ0 q1 
d    1 0 −1 0   
 Ap2   2
0 −ω 2H 0   
dt   =  0 ωH   Ap2  (4.304)
 hρ0 q2   0 0 1 0 −1   hρ0 q2 
2
Ap3 0 0 0 ωH 0 Ap3
The resonance frequencies of the system matrix are
√
c c √ 3 3c c
ωH = = 3 and 3ω H = = 5.2 (4.305)
h L L L
while the first and second resonance of the exact model are
c c c c
ω1 = π = 3.14 and ω 2 = 2π = 6.28 (4.306)
L L L L
Example 69 The input flow is zero and the outlet flow is zero, then a hybrid model
should be used. With N = 1, then L = 3h/2, and the model is
µ ¶ µ ¶µ ¶
d Ap1 0 −ω 2H Ap1
= (4.307)
dt hρ0 q1 1 0 hρ0 q1

The resonance is at
3c c
ω1 = = 1.5 (4.308)
2L L
while the exact value for the first resonance is
πc c
ω= = 1.57 (4.309)
2L L

4.7 Object oriented simulation models

4.7.1 Introduction
The final section of the chapter will present examples on how subsystem models can
be connected to a model, and how subsystem models may be added or changed without
causing extensive work. This will be done using subsystem models that are interconnected
with effort and flow variables.

4.7.2 Pump controlled hydraulic motor

Pump controlled hydraulic motors are used in applications where high power is required
as the power efficiency may be as high as 90% for such systems. Such systems are used
in vehicles and in cranes for heavy lifting operations.
We consider the system shown in Figure 4.18 which depicts an arrangement which
is called a hydrostatic gear or hydraulic gear . The system has a pump with a variable
displacement, which is driven by a motor with constant speed. The pump is connected
to a motor which may have a fixed or variable displacement. The speed and direction of
186 CHAPTER 4. HYDRAULIC MOTORS

p m
p1
Electric m
motor
p p2 m

Figure 4.18: Pump controlled hydraulic motor with variable displacement in the pump
and the motor. The arrangement is known as a hydrostatic gear.

rotation of the rotor can then be controlled with the displacement of the pump which is
varied by the stroke angle φp of the pump. The stroke of the pump is controlled with a
small servomechanism, which may use a valve controlled hydraulic motor or an electrical
actuator.
The dynamic model of the pump is given by (4.73—4.75). This is also the case for the
hydraulic motor. In this type of system the high pressure side with pressure p1 will be
driving the motor, while the pressure p2 of the return pipe from the motor to the pump
will be zero. Therefore, only the mass balances of the high pressure side are included in
the model. Then the model for the pump is

V1p
ṗ1p = Dp ω p − Cip p1p − Cep p1p − q1p (4.310)
β
Jp ω̇ p = −Bp ω p − Dp p1p + Tem (4.311)

where Dp is the pump displacement, ω p is the angular velocity of the pump, Vip is the
volume of chamber 1 of the pump, Cip and Cep are leakage coefficients, Jp is the inertia of
the pump, Tem is the torque from the electrical motor, and Bm is the friction coefficient
of the pump. The model for the hydraulic motor is

V1m
ṗ1m = −Dm ω m − Cim p1m − Cem p1m + q1m (4.312)
β
Jm ω̇ m = −Bm ω m + Dm p1m − TL (4.313)

where Vim is the volume of chamber 1 of the motor, Cim and Cem are leakage coefficients,
and Bm is the friction coefficient of the motor. The displacement of the pump and motor
are assumed to be given by

Dp = kp φp (4.314)
Dm = km φm (4.315)

where φp is the stroke angle of pump, φm is the stroke angle of motor, and kp and km
are displacement coefficients.
In contrast to the model of the valve controlled motor it is not possible to connect
the outputs of the pump model to the inputs of the hydraulic motor and vice versa. The
reason for this is the model are interconnected by imposing the conditions q1 = q1p = q1m
and p1 = p1p = p1m . However, q1 will then be input to both models, and p1 will be output
from both models. The solution to this problem is to combine the two mass balances
4.7. OBJECT ORIENTED SIMULATION MODELS 187

(4.310) and (4.312) by adding them. This gives the model

V1
ṗ1 = −Dm ω m + Dp ω p − C1 p1 (4.316)
β
Jp ω̇ p = Tem − Bp ω p − Dp p1 (4.317)
Jm ω̇ m = −Bm ω m + Dm p1 − TL (4.318)

whereV1 = V1p + V1m , Cit = Cip + Cim and C1 = Cit + Cep + Cem .

Example 70 Suppose that the motor displacement Dm and the pump speed ω p are con-
stants, and that the pump displacement is Dp = kp φp where the stroke angle φp of the
pump is the control input. Then the model is
V1
ṗ1 = −Dm ω m − C1 p1 + kp ω p φp (4.319)
β
Jm ω̇ m = −Bm ω m + Dm p1 − TL (4.320)

Example 71 If the pump and the hydraulic motor are connected with a transmission
line on the high pressure side, then there will be independent mass balances for the pump
and the motor. The high pressure port of the pump would then have pressure pp and flow
qp and the high pressure port of the motor would have pressure pm and flow qm . The
models of the pump and the motor would be
V1p
ṗ1p = Dp ω p − Cip p1p − Cep p1p − q1p (4.321)
β
Jp ω̇ p = −Bp ω p − Dp p1p + Tem (4.322)

and
V1m
ṗ1m = −Dm ω m − Cim p1m − Cem p1m + q1m (4.323)
β
Jm ω̇ m = −Bm ω m + Dm p1m − TL (4.324)

These model have flows as inputs and pressures as outputs. The transmission line model
of the high pressure line would be an admittance model
µ ¶ µ cosh Γ 1
¶µ ¶
q1p (s) 1 sinh Γ − sinh Γ p1p (s)
= 1 (4.325)
q1m (s) Zc sinh Γ − cosh Γ
sinh Γ
p1m (s)

with the pressures p1p and p1m as inputs and the flows q1p and q1m as outputs. The
admittance model is irrational due to the transcendental functions cosh Γ and sinh Γ.
Rational approximation of this irrational model must be used to develop a simulation
model.

Example 72 A load with inertia J1 and angular velocity ω 1 with elastic transmission
can be connected to the motor by formulating the load with a port with the motor speed
ω m as the input variable, and the load torque TL as the output variable. This is done by
the load model

J1 ω̇ 1 = TL − T1 (4.326)
θ̇1 = ω1 (4.327)
TL = D1 (ω m − ω 1 ) + K1 (θm − θ1 ) (4.328)
188 CHAPTER 4. HYDRAULIC MOTORS

We may add on any number of additional degrees of freedom as two-ports

Ji ω̇ i = Ti−1 − Ti (4.329)
θ̇i = ωi (4.330)
Ti = Di (ω i−1 − ω i ) + Ki (θi−1 − θi ) (4.331)

with port variables Ti−1 and ω i−1 at the input and Ti and ω i at the output.

Example 73 Consider the connection of a load with inertia J1 with a stiff connection
to the motor so that the load has the same shaft speed as the motor. Then there would
not be an extra state because of the load, and the load is included in the model by adding
the load inertia to the motor inertia in the equation of motion for the motor shaft, which
gives
(Jm + J1 ) ω̇ m = −Bm ω m + Dm pm − TL (4.332)

4.7.3 Cylinder with balance valve

Hydraulics for heavy lifting with a crane will often be implemented with a hydraulic
cylinder with a single-rod piston. To prevent problems with hanging loads the cylinder
may be connected to a balance valve on the lifting side of the piston. In this section
this type of system will be controlled with a valve. To establish a object oriented model
we first define the ports of the components, and then the ports will be connected, and
finally we present the models for each of the components.
The inlet port of the cylinder has pressure p1 and flow q1 into the chamber, the outlet
port has pressure p2 and flow q2 out of the chamber. The balance valve has input port
pressure pBV 1 and flow qBV into the valve, and outlet port with pressure pBV 2 and flow
qBV out of the valve. The control valve has one port A with pressure pA and flow qA
out of the valve, and one port B with pressure pB and flow qB into the valve. The pipes
from the control valve to the motor and balance valve are denoted A and B. The volume
of pipe A is VA and the volume of pipe B is VB . The inlet of pipe A has pressure pA and
flow qA1 into the pipe, and the outlet port has pressure pA and flow qA2 out of the pipe.
The inlet of pipe B has pressure pB and flow qA1 into the pipe, and the outlet port has
pressure pA and flow qA2 out of the pipe.
The control valve port A is connected to pipe A, and the valve port B is connected
to pipe B. Pipe A is connected to the inlet port of the balance valve so that pBV 1 = pA ,
and pipe B is connected to the outlet port of the motor so that pB = p2 . The outlet
port of the balance valve is connected to the inlet port of the motor so that pBV 2 = p1
and qBV = q1 .
The flows qA and qB are found from the orifice equations
q
qa = Cd Aa (xv ) ρ2 (ps − pA )
q
qb = Cd Ab (xv ) ρ2 (ps − pB )
q (4.333)
qc = Cd Ac (xv ) ρ2 (pA − pr )
q
qd = Cd Ad (xv ) ρ2 (pB − pr )

and
qA = qa − qc , qB = qd − qb (4.334)
where xv is an input to the model.
4.7. OBJECT ORIENTED SIMULATION MODELS 189

The pressures pA is found from the pipe model that is given by the mass balance
VA
ṗA = qA − q1 (4.335)
β
Note that there is no model for pipe B because pipe B is connected to the outlet chamber
of the motor so that pB = p2 . Instead, the volume of pipe B is included in the chamber
volume of chamber 2 in the motor.
The volumetric flow q1 is found from the flow characteristic of the the balance valve,
which is  q
 Cd xbv1 b 2 (p1 − pA ) pA > p1
q1 = qρ (4.336)
 −Cd Ac 2 (pA − p1 ) p1 < pA
ρ

where the spool position of the balance valve is

Ar
xbv1 = [p1 − p01 + Rp2 − pA (R + 1)] , 0 ≤ xbv1 ≤ xbv1,max
K
The piston moves in the vertical direction, and the position of the piston is denoted
by xp which is positive in the upwards direction. The cylinder has one chamber with area
A1 , pressure p1 and volume V1 = V10 + A1 xp . The other chamber has area A2 = A1 − Ar
where Ar is the area of the rod, pressure p2 and volume V2 = V20 − A2 xp . The model
for the cylinder is then
V10 + A1 xp
ṗ1 = −Cim (p1 − p2 ) − Cem p1 − A1 ẋp + q1 (4.337)
β
VB + V20 − A2 xp
ṗ2 = −Cim (p2 − p1 ) − Cem p2 + A2 ẋp − q2 (4.338)
β
mt ẍp = −Bp ẋp + A1 p1 − A2 p2 − FL (4.339)

where FL is an input to the model.

Example 74 If a balance valve had been added at the outlet port of the motor, then
pressure pB would have to be computed from a line model
VB
ṗB = q2 − qB (4.340)
β
and the mass balance of chamber 2 of the motor would have been
V20 − A2 xp
ṗ2 = −Cim (p2 − p1 ) − Cem p2 + A2 ẋp − q2 (4.341)
β
The flow q2 would then be found from the balance valve characteristic of the balance valve
at port 2 as  q
 −Cd xbv2 b 2 (p2 − pB ) pB > p2
q2 = q ρ (4.342)
 Cd Ac 2 (pB − p2 ) p2 < pB
ρ

where the spool position of the balance valve is

Ar
xbv2 = [p2 − p02 + Rp1 − pB (R + 1)] , 0 ≤ xbv2 ≤ xbv2,max
K
190 CHAPTER 4. HYDRAULIC MOTORS
Chapter 5

Friction

5.1 Introduction
5.1.1 Background
Friction is the tangential reaction force between two surfaces in contact. The friction
force is dependent on a number of factors, such as contact geometry, properties of the
surface materials, displacement, relative velocity and lubrications. Friction is a highly
complex phenomenon, composed of several physical phenomena in combination. As a
result of this, models of friction are to a large extent empirical, which means that the
models are constructed in order to reproduce effects observed in experiments. On the
other hand, some of the dynamic friction models aim at modelling the physics behind
the phenomenon.
A macroscopic smooth surface is far from smooth when viewed at a microscopical
scale. The small features of the surface are called asperities. When two surfaces are
brought into contact, the true contact occur between the asperities in what is called
asperity junctions. An example of this is shown in Figure 5.1. In engineering materials,
the slope of the asperities are typical in the range 5◦ − 10◦ , and the width is typical
10 µm. When two bodies in contact are brought into relative motion by an external
force, the asperities will behave like springs, and there will be an elastic deformation of
the asperities. This motion is referred to as pre-sliding displacement or the Dahl effect.

Asperity
junctions

Asperities

Figure 5.1: The asperities and junctions of two bodies in contact, viewed at a microscop-
ical scale

191
192 CHAPTER 5. FRICTION

The tangential force Ft (x) can in this regime be approximated by

Ft (x) = −kt x (5.1)

where kt is the stiffness of the contact and x is the relative displacement.

By increasing the external force the asperities will undergo plastic, irreversible defor-
mation and then rupture. The force needed to break the junctions is referred to as the
break-away force Fb , and the phenomena itself is called break-away. The stiffness kt can
be found from
Fb
kt = (5.2)
xb
where xb is the break-away displacement. Before beak-away the system is said to stick,
while after break-away the system is said to slip, and the term stick-slip friction is used
to characterize the phenomenon.

5.2 Static friction models

Static friction models present the friction force as a function of velocity. This function
can be characterized by the following for regimes:

I. Static friction Elastic deformation of the asperities, the Dahl effect.

II. Boundary lubrication For very low velocities, no fluid lubrication occur, and the
friction is dominated by shear forces in the solid boundary film.

III. Partial fluid lubrication The Stribeck effect

IV. Full fluid lubrication A lubricant film thicker than the size of the asperities is
maintained, and no solid contact occurs. The friction is purely viscous.

The resulting map is referred to as the generalized Stribeck curve, which is shown in
Figure 5.2. Static friction models will represent these regimes to a varying extent. A
selection of static friction models that are commonly used is shown in Figure 5.3.

5.2.1 Models for the individual phenomena

Coulomb friction
The classical model of friction where the friction force is proportional to load, opposes the
motion, and is independent of contact area is known as Coulomb friction. The friction
force in the Coulomb model is given by

Ff = Fc sgn(v), v 6= 0 (5.3)
where the Coulomb force Fc is given by

Fc = µFN (5.4)

Here µ is the friction coefficient and FN is the load. Equation (5.4) can be derived as
follows. It is assumed that there is no contamination, such as lubrication, of the contact
surfaces. The friction is then referred to as dry friction. In this context friction can
5.2. STATIC FRICTION MODELS 193

Regime I. No sliding.
Elastic deformation
F

Regime III. Partial fluid

Regime II. Boundary

Regime IV. Full fluid

lubrication

lubrication

lubrication

Figure 5.2: The generalized Stribeck curve, showing friction as a function of velocity for
low velocities, (Armstrong-Hélouvry et al. 1994).

F F F
a) b) c)

v v v

F F F
d) e) f)

v v v

Figure 5.3: Static friction models: a) Colomb friction b) Coulomb+stiction c)

Coulomb+stiction+viscous d) Stribeck effect e) Hess and Soom; Armstrong f) Karnopp
model
194 CHAPTER 5. FRICTION

be defined as the shear strength of the asperity junction areas, and the friction force is
proportional to the true area of contact Ac

Ff = Ac fs (5.5)

where fs is the shear force per unit area, a constant material property. The true area of
contact Ac can be found from
FN
Ac = (5.6)
py
where FN is the load, and py is the yield pressure, a constant material property. Com-
bining (5.6) with (5.5) gives
FN
Ff = fs = µFN (5.7)
py
where the friction coefficient µ is found as µ = fs /py , and thus (5.4) holds. From
the derivation, it is seen that Ac is cancelled out of the expression, and so friction is
independent of contact area.
According to Armstrong-Hélouvry et al. (1994), Fc is also dependent on lubricant
viscosity and contact geometry. The nature of Coulomb friction was known to Leonardo
Da Vinci, and his results were further developed by Coulomb. The friction force given by
(5.3) is not necessary symmetric in v, that is, Ff may take different values for different
directions of the velocity.

Static friction
Static friction is also known as stiction and models the fact that in some cases the friction
force is larger in magnitude for zero velocity than for a non-zero velocity. According to
the stiction model the system sticks if the velocity is zero and |Ff | < Fs , and it breaks
away if |Ff | = Fs where Fs > Fc is the stiction force, which is larger in magnitude than
the Coulomb force Fc .
During a pre-sliding displacement, some motion is possible even when a mechanism
is stuck in static friction. If the applied force returns to zero, the position returns to its
initial value, possibly after a transient of pre-sliding displacement.

Viscous friction
Viscous friction is present in fluid lubricated contacts between solids. The concept of vis-
cous friction was first introduced by Reynolds(1886). A viscous friction model takes into
account that due to hydrodynamic effects, the friction force depends on the magnitude
of the velocity, and not only its direction. The usual linear model is given by

Ff v = Fv v (5.8)

where the viscous friction is proportional to velocity. The constant of proportionality Fv

depends on lubricant viscosity, loading and contact geometry. Generally, viscous friction
exhibits a non-linear behavior, and a nonlinear version of (5.8) is

Ff v = Fv |v|δv sgn(v) (5.9)

where 0 < δ v ≤ 1 is a constant.

5.2. STATIC FRICTION MODELS 195

Figure 5.4: Frictional lag

Decreasing viscous friction: The Stribeck effect

Decreasing viscous friction is also known as Stribeck friction or the Stribeck effect, and
was first described in Stribeck (1902). The Stribeck effect has its background in partial
fluid lubrication. In this case some of the load is carried by the fluid film created by the
lubricant, and some by elastic and plastic deformation of the asperities. The fluid film
thickness increases with velocity, and the resulting tangential force decreases since the
shear forces of the film are smaller than the shear forces of the asperities. Several static
models have been proposed for the Stribeck effect. Armstrong-Hélouvry (1990) propose
to use h 2
i
Ff = Fc + (Fs − Fc )e−(v/vs ) sgn(v), v 6= 0 (5.10)
where Fs is the static friction, Fc is the Coulomb friction and vs is denoted the char-
acteristic velocity of the Stribeck friction. Equation (5.10) will model Coulomb friction,
stiction and the Stribeck effect. The model is not defined for zero velocity, but |Ff | → |Fc |
when the velocity tends to zero. Hess and Soom (1990) propose the expression
" #
(Fs − Fc )
Ff = Fc + sgn(v), v 6= 0 (5.11)
1 + (v/vs )2

for modeling the same effect. It is important to notice that models such as (5.10) or (5.11)
are not based on the physics of the phenomenon, but is rather a curve fit to experimental
data as shown in Figures 5.3 d) and f).

Other friction related phenomena

In the friction experiments of Hess and Soom (1990) it was observed that friction force
was lower for decreasing velocities than for increasing. This is a hysteresis effect, and is
referred to as frictional lag, or frictional memory, see Figure 5.4. Moreover, the break-
away force has been found to depend on the rate of change of externally applied force.
Larger rates gives smaller break-away force.

5.2.2 Combination of individual models

The static models presented above can be combined to produce models that take several
of these phenomena into account. For instance, the most commonly used model in
engineering, which is the Coulomb+viscous friction model can be found by adding (5.3)
and (5.8) together to form
F = Fc sgn (v) + Fv v (5.12)
196 CHAPTER 5. FRICTION

Fs
4 Fv
3

2 Fc

Friction force F 0

-1

-2

-3

-4

-5
-5 -4 -3 -2 -1 0 1 2 3 4 5
Velocity v

Figure 5.5: An example of a friction map

The Stribeck effect can also be included using (5.11), which gives (Hess and Soom 1990)
à !
(Fs − Fc )
F = Fc + sgn(v) + Fv v (5.13)
1 + (v/vs )2

A plot of the Friction curve given by (5.13) is shown in Figure 5.5

5.2.3 Problems with the static models

There are two main problems connected to the use of static friction models for simulation
and control applications:

1. They are dependent on the detection of zero velocity, as the model rely on switching
at zero velocity.

2. They do not describe all observed dynamic effects, such as pre-sliding displacement,
varying break-away force and frictional lag..

The zero velocity problem can be handled by a static model known as the Karnopp
model, Karnopp (1985), where a zero velocity interval, |v| < η is used. Outside this
interval,that is for |v| ≥ η, friction force is the usual function of velocity, but within the
interval, the velocity is considered to be zero and friction is a function of other forces in
the system: ½
Ff (v) , |v| ≥ η
Ff (v, F ) = (5.14)
Ff (F ) , |v| < η
where F represents the sum of other forces in the system and η > 0 is the small constant
defining the Karnopp zero interval. This is advantageous in simulations, but the zero
interval does not agree with real friction, and the model strongly depends on the rest of
the system through Ff (F ). A plot of the static Karnopp model is shown in Figure 5.3 f).
5.2. STATIC FRICTION MODELS 197

A friction model known as Armstrong’s 7-parameter model Armstrong-Hélouvry et al.

(1994) is capable of modelling some of the phenomena not included in the classical static
models. The model consists of two equations, one for sticking and one for sliding:

F (x) = σ 0 x, (5.15)

when sticking, and

à !
1
F = Fc + Fs (γ, td ) 2 sgn(v) + Fv v (5.16)
1 + (v(t − τ l )/vs )

where µ ¶
td
Fs (γ, td ) = Fs,a + Fs,∞ − Fs,a (5.17)
td + γ
when sliding. Fs,a is the Stribeck friction at the end of the previous sliding period and td
is the dwell time, the time in stick. Although the Armstrong 7-parameter model describes
more phenomena than the classical models, it still requires switching between different
equations. The problems with static friction models in connection with simulation and
control systems design has led to the use of dynamic friction models for high precision
friction modeling.

5.2.4 Problems with signum terms at zero velocity

In static friction models the friction term will typically include a term which more or
less looks like the Coulomb friction model Ff = Fc sgn(v). This model is not defined at
zero velocity, however, it is not unusual that the model is extended to be valid at zero
velocity using the model 
 −Fc , v < 0
Fc sgn(v) = 0 v=0 (5.18)

Fc , 0 < v
This is e.g. done in the Simulink block for Coulomb and viscous friction. It is not difficult
to see that this model does not reflect the physics of the problem. To makes this clear
we consider a mass m with position x, velocity v = ẋ and an active force Fa acting on
the mass. The friction force is Ff , and Newton’s law gives

mv̇ = Fa − Ff (5.19)

Now, suppose that v = 0. If the friction force is given by Ff = Fc sgn(v) as defined

by 5.18, then Ff = 0 for v = 0, and the system will not stick, but rather accelerate
according to mv̇ = Fa . If Fa is positive and nonzero, then the velocity becomes positive,
and the friction force will at the next time step be Ff = −Fc . If Fa < Fc this will cause
a reversal of the acceleration, and at the next step the velocity may have changed sign
so that Ff = Fc . It is clear that this will give strong oscillations in the system, and that
these oscillations have nothing to do with the physics of the system. In a simulation
with fixed time-step, there will be strong oscillations in the numerical solution, whereas
a simulation with variable time-step will more or less stop as the time step will be made
very small to in order to reach the specified accuracy.
The conclusion to this discussion is that models that contain a signum term like the
one in (5.18) will not give results that agree with the physics of the problem at zero
velocity. Moreover, serious problems are introduced in simulations.
198 CHAPTER 5. FRICTION

5.2.5 Karnopp’s model of Coulomb friction

Karnopp’s friction model extends the basic Coulomb friction model for dry friction to be
valid also for zero velocity. It is straightforward to extend this model to include stiction
and the Stribeck effect. Karnopp’s model can be explained by considering a mass m with
velocity v that is pushed on a flat surface with a force Fa . The friction force on the mass
is Ff so that the equation of motion is given by (5.19) According to the Coulomb friction
model the friction force Ff is of magnitude Fc in the opposite direction of the velocity
as long as the velocity is nonzero. This may be modelled as
½
Fa + Fc , v < 0
mv̇ = (5.20)
Fa − Fc , v > 0

In this formulation the friction force is a function of the velocity. Note, however, that
this model is undefined for zero velocity, so there is a need for refining the model. This is
done in Karnopp’s model by observing that the physical behavior of the system at zero
velocity is that the velocity remains equal to zero as long as the force Fa is less than Fc
in magnitude. This can be written

mv̇ = 0, v = 0 and |Fa | ≤ Fc (5.21)

Combining this with the equation of motion we see that Ff = Fa when v = 0 and
|Fa | ≤ Fc . Thus, for zero velocity, the friction force is a function of the force acting on
the mass. Define the saturation function sat(x, S) so that sat(x, S) = x when |x| ≤ S
and sat(x, S) = Ssgn(x) when |x| ≥ S.

The Karnopp friction model for Coulomb friction is given by

½
sat(Fa , Fc ) when v = 0
Ff = (5.22)
Fc sgn(v) else

Note that the computational input of the Karnopp model at the input port is Fa
when v = 0 and |Fa | ≤ Fc , and that the computational input changes to v when the
condition does no longer hold.

5.2.6 More on Karnopp’s friction model

The main contribution of Karnopp’s friction model is the handling of the sticking phe-
nomenon at zero speed. This can also be applied to other friction models with signum
terms. The model (5.10) which includes sticking and the Stribeck effect can be modeled
with Karnopp’s method with the friction force
(
sat(F
h a , Fc ) i when v = 0
Ff = 2 (5.23)
Fc + (Fs − Fc )e−(v/vs ) sgn(v) else

where Fa is the applied force and v is the velocity.

In simulations with Karnopp’s model there must be a switch between the two regimes
in (5.22) or (5.23). This requires some method for detecting that the velocity is zero.
Some simulation systems with variable-step integration methods will have event-detection
mechanisms that can be used for this purpose. This type of event detection is included
5.2. STATIC FRICTION MODELS 199

in MATLAB and Simulink. This method is used in friction models in Modelica, where
the computational inputs are switched at zero velocity.
Alternatively, a dead-zone around zero velocity can be used where the velocity is
treated as is it were zero in the computation of the friction force. Then the friction
model (5.22 will be modified to
½
sat(Fa , Fc ) when |v| ≤ δ
Ff = (5.24)
Fc sgn(v) else

where the magnitude δ of the dead-zone will have to be selected depending on the size
of the time-step, and on the maximum acceleration that can be expected in the system.
The model (5.24) is straightforward to implement in MATLAB and Simulink. If the
friction model is used in an observer in a control system , then the time-step will have to
be fixed, and a dead-zone must be used. It is clear that the introduction of a dead-zone
will be an approximation that will introduce some error. However, the performance of
this solution is vastly superior to the naive implementation in (5.18) of the signum term,
and Karnopp’s model should be the standard way of modeling friction unless dynamic
phenomena like pre-sliding and frictional hysteresis are the dominant physical effects.

5.2.7 Passivity of static models

Friction in its very nature is a dissipative phenomenon as the friction force cause dis-
sipation of energy whenever the velocity is nonzero. An exception to this is elastic
deformation in the pre-sliding region where energy is stored as in a spring, and the sys-
tem will still be passive although energy is not dissipated. Because of this a friction
model should be passive in the sense that the system with velocity as input and friction
force as output should be passive to reflect the physics of the system. In this section we
establish the passivity properties of the static friction models.
For any of the static friction models presented above, the friction force Ff (v) is a
sector nonlinearity, that is Ff (v) satisfies

Ff (v) ∈ sector(k1 , k2 ) ⇐⇒ k1 v 2 < Ff (v)v < k2 v 2 (5.25)

A plot of the static model (5.13) is shown in Figure 5.5, and as can be seen Ff (v) is
located in the first and third quadrants, that is Ff (v) ∈ sector[0, ∞). Moreover, by
studying Figure 5.5, it can be seen that

Ff (v) ∈ sector[k1 , ∞), where 0 < k1 ≤ Fv (5.26)

Calculating the power Ff (v)v for a static friction model where Ff (v) satisfies (5.26), and
integrating, we get
Z T Z T
Ff (v)vdt > k1 v 2 dt, where 0 < k1 ≤ Fv (5.27)
0 0

It follows from (5.27), that the system with input v and output Ff is passive. This
result can be generalized to any sector nonlinearity. Karnopp’s model is identical to the
RT
static models except at zero velocity. Therefore the integral 0 Ff (v)vdt will be the
same as for the static methods. This implies that Karnopp’s model is passive. When the
dead-zone is included passivity cannot be established.
200 CHAPTER 5. FRICTION

5.3 Dynamic friction models

5.3.1 Introduction
There are two problems connected to the use of static friction models for simulation and
control applications. First, static models are dependent on the detection of zero velocity,
as the models rely on switching between different models at zero velocity. Second, static
models do not describe dynamic effects such as pre-sliding displacement, varying break
away force and frictional lag. Because of this, dynamic modes of friction have been
proposed.

5.3.2 The Dahl model

The model of Dahl (1968) was developed for the purpose of simulating control systems
with friction. A dynamic friction model can be developed by differentiating the friction
force with respect to time:
dF ∂F ∂F dx
= + (5.28)
dt ∂t ∂x dt
Then if ∂F/∂t = 0, the expression

dF dF dx
= (5.29)
dt dx dt
is found. For small displacements the friction is determined by the pre-sliding elastic
deformation of the asperities. This can be modeled as a linear spring:

x ¿ 1 ⇒ |F | = |σx| ¿ Fc (5.30)

with σ being the spring stiffness. For large displacements, the model should behave like
a Coulomb model. Dahl (1976) found that a model that satisfies this is given by
¯ ¯α µ ¶
dF ¯ F dx ¯¯ F dx
¯
= σ ¯1 − sgn ¯ sgn 1 − sgn
dx Fc dt Fc dt
µ ¶α µ ¶
F dx α+1 F dx
= σ 1 − sgn sgn 1 − sgn (5.31)
Fc dt Fc dt
³ ´
By using the fact that F < Fc it follows that sgn 1 − FFc sgn dxdt > 0, and (5.31) simplifies
to µ ¶α
dF F dx
= σ 1 − sgn (5.32)
dx Fc dt
The constant α depends on the material of the solid, α ≥ 1 describe ductile type mate-
rials, while α < 1 describes brittle type materials. Applications of the model, however,
typically employ α = 1, so that
µ ¶
dF F dx
= σ 1 − sgn (5.33)
dx Fc dt

It follows from (5.33) that in steady state

dx
Fss = Fc sgn = Fc sgn(v) (5.34)
dt
5.3. DYNAMIC FRICTION MODELS 201

which can be compared to (5.3). The Dahl model includes the phenomena Coulomb
friction and pre-sliding displacement. From (5.33) it is seen that for small displacements
the friction force can be approximated by F ≈ σx, which is the model of a linear spring
with σ being the spring stiffness. For large displacements the friction for can be approx-
imated by F ≈ Fc as for a static Coulomb model. However, as the model (5.33) is rate
independent, it is not capable of describing such phenomena as Stribeck-effect.
Combination of (5.29) and (5.33) then leads to the following result

The Dahl friction model is given by the dynamic model

µ ¶
dF F
= σ v − |v| (5.35)
dt Fc

The Dahl model can be written in the form

dF |v|
=σ (Ff − F ) (5.36)
dt Fc
where Ff = Fc sgn(v). This resembles a low pass filter where the computed friction force
F tracks the Coulomb friction Fc sgn(v) with a pole at |v|/Fc which corresponds to a time
constant T = Fc /|v|. The pole of the low pass filter tends to zero when the velocity tends
to zero, and this is advantageous in simulation as the apparent gain around zero is small
in spite of the step due to the signum function. However, this advantage does not come
for free. The low gain of the model at low speeds may create problems in the modeling of
the sticking regime. In particular, if the system (5.19) is excited by an oscillatory force
Fa where |Fa | < Fc , then the system may drift if the friction force Ff is modelled with
the Dahl model, although sticking would be expected from the physics of the system.
To conclude, the Dahl model is very simple to implement compared to Karnopp’s
model as there is no switching in the Dahl model. Moreover, there are no problems with
oscillations around zero velocity with Dahl’s model. However, the model may give drift
in the sticking region, so for models where correct sticking behavior is important it is
recommended to use Karnopp’s model.
Example 75 It is interesting to note that the differential equation in (5.33) can be solved
explicitly. Assuming Fc to be a constant and considering only forward motion, that is
ẋ > 0, we find by integrating (5.33) that
Z Z
dF
³ ´ = dx
σ 1 − FFc

This gives µ ¶
Fc σ
− ln σ − F = x + C 0 (5.37)
σ Fc
where C 0 is a constant of integration, and where it has been assumed that F < Fc . Solving
for F we find that
Fc C 00 − Fσ x
F = Fc − e c
σ
where C 00 is another constant. Finally, by using the fact that F (0) = 0, we find that C 00 =
σ, and consequently Dahl’s friction model for forward speed can be written in the form
³ σ
´
F = Fc 1 − e− Fc x (5.38)
202 CHAPTER 5. FRICTION

Fc
2
x  0

1



Friction force F 0

-1

x  0

-2
F c

-3
0 1 2 3 4 5 6 7
Displacement x

Figure 5.6: The Dahl friction model for positive and negative velocity.

By doing the same calculations for ẋ < 0, we find that

³ σ
´
F = Fc sgn(v) 1 − e− Fc xsign(v) (5.39)

which is shown for Fc = 2 and σ = 0.5 in Figure 5.6.

5.3.3 Passivity of the Dahl model

For dynamic Dahl model the friction force is given by
µ ¶
dF F |v|
= σ0 v − , Fc > 0
dt Fc

we consider the storage function

F2
V =
2σ 0
The time derivative along the solutions of the system is

1 F 2 |v|
V̇ = F Ḟ = F v − (5.40)
σ0 Fc

It is seen that for the Dahl model the system with input v and output F is passive.

5.3.4 The Bristle and LuGre model

The Bristle model to friction was introduced by (Haessig, Jr. and Friedland 1991). The
basic assumption behind the Bristle friction model is that asperity junctions can be
modeled of elastic bristles as shown in Figure 5.7. As the surfaces move relative to each
5.3. DYNAMIC FRICTION MODELS 203

Figure 5.7: The asperity junctions of two bodies in contact are modeled as elastic bristles.
For simplicity, only the upper body is moving in this figure, and the bristles of the lower
body are assumed rigid.

other the strain in the bond increases and the bristles acts as springs giving rise to the
friction force. The force is given by
N
X
F = σ 0 (xi − bi ) (5.41)
i=1

where N is the number of bristles, σ 0 is the stiffness of the bristles, xi is the relative
position of the bristle and bi is the location where a bond was formed. In simulations, a
bond will snap when |xi − bi | = δ s , and then a new will be formed at a random location
relative to the previous location. The complexity of this model increases with N . The
stiffness of the bristles σ 0 can be made velocity dependent. An interesting property of
this model is that it attempts to capture the random nature of friction. However, it is
inefficient in simulations due to its complexity. Also it may give rise to oscillatory motion
in stick due to the lack of damping. The LuGre (Lund-Grenoble ) model (Canudas de
Wit, Olsson, Åström and Lischinsky 1995) is based on the same idea as the bristle
model, but the friction force is generated by a dynamic equation reminiscent of the Dahl
model describing the average deflection of the bristles, thereby reducing the complexity
introduced by the sum in (5.41).

The LuGre dynamic friction model is given by the dynamic system

|v|
ż = v − σ 0 z (5.42)
g(v)
where the function g(v) is selected to be
2
σ 0 g(v) = Fc + (Fs − Fc )e−(v/vs ) (5.43)
to account for stiction and the Stribeck effect as in (5.10). The friction force is given by
F = σ 0 z + σ1 ż + σ 2 v. (5.44)
204 CHAPTER 5. FRICTION

The following property of the model is noted: It is seen from (5.42) that the model
for z can be written
|v|
ż = [g(v)sgn(v) − σ 0 z] (5.45)
g(v)
From this formulation it is seen that if the initial value of z satisfies |σ 0 z(0)| ≤ gmax ,
where gmax is the maximum value of g(v), then the absolute value of the state z(t) is
upper bounded according to
|σ0 z(t)| ≤ gmax (5.46)
It is also noted that
Fc ≤ σ 0 g(v) ≤ Fs (5.47)
It is seen that the stationary solution of (5.42) is

zss = g(v)sgn(v). (5.48)

This shows that the therm σ2 v in (5.44) will tend to represent stiction and the Stribeck
effect, while σ2 v term will represent viscous friction. The is characterized by the six
parameters σ 0 , σ 1 , σ 2 , vs , Fs and Fc . By comparing the LuGre model with the Dahl
model as given by (5.35) it is clear that the LuGre model is a generalization of the Dahl
model, where the Dahl model appears in the case that σ 0 z = F , σ 1 = σ 2 = 0 and
g(v) = Fc . The LuGre model has the potential of being more accurate than the Dahl
model as it includes the Stribeck effect, and as it may represent frictional lag. As with
the Dahl model the LuGre model may drift in the sticking region.

5.3.5 Passivity of the LuGre model

To find out if the model is passive from velocity v to friction force F we follow the
procedure of (Barabanov and Ortega 2000) and investigate the integral
Z T Z T Z T µ ¶
dz
vF dt = σ0 vzdt + v σ1 + σ 2 v dt (5.49)
0 0 0 dt

The first term on the right side corresponds to the Dahl part of the LuGre model. We
find that this term is not problematic as
Z T Z T µ ¶
|v|
σ 0 vzdt = σ 0 z ż + σ0 z dt
0 0 g(v)
Z T
σ0 £ 2 2
¤ |v|
= z (T ) − z (0) + z2 σ0 dt
2 0 g(v)
σ0
≥ − z 2 (0) (5.50)
2
The second term on the right side of (5.49) is somewhat more involved. We find that
Z T µ ¶ Z T µ ¶
dz |v|
v σ1 + σ 2 v dt = v (σ1 + σ 2 ) v − σ 1 σ 0 z dt (5.51)
0 dt 0 g(v)

Using (5.46) we find that

Z T µ ¶ Z T µ ¶
dz gmax
v σ1 + σ 2 v dt ≥ v 2 (σ 1 + σ 2 ) − σ 1 dt (5.52)
0 dt 0 gmin
5.3. DYNAMIC FRICTION MODELS 205

This implies that the LuGre model is passive from v to F if

gmax gmin
(σ 1 + σ 2 ) − σ 1 ≥ 0 ⇒ σ1 ≤ σ2 (5.53)
gmin gmax − gmin

Insertion of the maximum and minimum values of g according to (5.47) leads to the
conclusion that the LuGre model is passive if
Fc
σ1 ≤ σ2 (5.54)
Fs − Fc
In (Barabanov and Ortega 2000) it was shown that this is a necessary and sufficient
condition for passivity from v to F .

5.3.6 The Elasto-Plastic model

The LuGre model does not render true stiction. Therefore a new dynamic friction model
was proposed in (Dupont, Hayward, Armstrong and Altpeter 2002) with the aim of
having a model that accounts for both true stiction and pre-sliding. The model is a
generalization of the LuGre model. The model is written
µ ¶i
σ0 sgn(v)
ż = v 1 − α(z, v) z (5.55)
g(v)

dz
F = σ0 z + σ1 + σ2 v (5.56)
dt
and the term α(z, v), which is the new feature of the model when compared to LuGre,
is used to render true stiction. The piecewise continuous function α(z, v) is defined as
 

  0 |z| ≤ zb

0 < α < 1 zb < |z| < zmax (v) , sgn(v) = sgn(z)
α(z, v) =  , (5.57)

 1 |z| ≥ zmax (v)

0 , sgn(v) 6= sgn(z)

where
g(v)
0 < zb < zmax (v) = , ∀v ∈ R. (5.58)
σ0
An example of the term α(z, v) is
 

 
 0 |z| ≤ zb

  µ ³
z +z
´¶
 z− max2 b
1 1 , sgn(v) = sgn(z)
α(z, v) =  2 sin π zmax −zb + 2 zb < |z| < zmax (v)
 


 1 |z| ≥ zmax (v)


0 , sgn(v) 6= sgn(z)

In the Elasto-Plastic model, the body displacement

x = z + w.

is decomposed into its elastic and plastic (inelastic) components z and w. Stiction cor-
responds to the existence of a breakaway displacement zb > 0 such that for |z| ≤ zb all
motion of the friction interface consists entirely of elastic displacement. In this context,
206 CHAPTER 5. FRICTION

elastic displacement z corresponds to pre-sliding displacement and plastic (inelastic) dis-

placement w corresponds to sliding displacement. The choice of α(z, v) = 0, |z| ≤ zb
in (5.57), directly implies that the Elasto-Plastic model has a true stiction phase. The
Elasto-Plastic model is relatively complicated to implement. Therefore it is recommended
to use Karnopp’s model if the objective is to have a model with true stiction, and accurate
modeling of pre-sliding is not important, which normally will be the case. However, for
problems of very high accuracy where pre-sliding is the dominant frictional phenomenon
the Elasto-Plastic model can be used.

5.3.7 Passivity of the Elasto-Plastic model

As the Elasto-Plastic model differs from the LuGre model only in the inclusion of the
term α(z, v) in (5.55), the passivity analysis of the LuGre applies. It follows that the
system with input v and output F will be passive provided that
Fc
σ1 ≤ σ2 (5.59)
Fs − Fc
 Part III

Dynamics

207
Chapter 6

Rigid body kinematics

6.1 Introduction
Rigid body dynamics is important for a wide range of control applications, and is essential
in robot control, ship control, the control of aircraft and satellites, and vehicle control
in automotive systems. The field of rigid body dynamics is old and is very rich in
results. Important results date back to Newton in the 17th century, Euler in the 18th
century and Lagrange, Hamilton and Rodrigues in the 19th century. Because of the
development in control applications like robotics, aerospace, and the development in
numerical simulation, the selection of topics and method to be presented in rigid body
dynamics has developed quite a lot the last two decades, and this text attempts to reflect
this change. The material is based on general texts like (Kane and Levinson 1985) and
(Robertson and Schwertassek 1988), texts on spacecraft dynamics like (Kane, Likins and
Levinson 1983) and (Hughes 1986), and robotics books like (Spong and Vidyasagar 1989)
and (Sciavicco and Siciliano 2000).

6.2 Vectors
6.2.1 Vector description
Forces, torques, velocities and accelerations are well-known entities that can be described
by vectors. A vector u can be described by its magnitude |u| and its direction. Note that

x3

a 3


a2 x2
a1
x1

Figure 6.1: The coordinate frame a.

209
210 CHAPTER 6. RIGID BODY KINEMATICS


n
v



u

Figure 6.2: Vectors u, v and n.

this description of a vector does not rely on the definition of any coordinate frame. In
this respect the description may be said to be coordinate-free. Alternatively, a Cartesian
coordinate frame can be introduced, and the vector can be described in terms of its
components in the Cartesian coordinate frame. Let the Cartesian coordinate frame a
be defined by three orthogonal unit vectors a1 , a2 and a3 that are unit vectors along
the x1 , x2 , x3 axes of a (Figure 6.1). Then the vector u can be expressed as a linear
combination of the orthogonal unit vectors a1 , a2 and a3 by

u = u1 a1 + u2 a2 + u3 a3 (6.1)

where
ui = u · ai , i ∈ {1, 2, 3} (6.2)
are the unique components or coordinates of u in a. A related description of the vector is
the coordinate vector form where the coordinates of the vector are written as a column
vector  
u1
u =  u2  (6.3)
u3

6.2.2 The scalar product

The scalar product between two vectors u and v is given in the coordinate-free description
by
u · v = |u||v| cos θ (6.4)
where θ is the angle between the two vectors (Figure 6.2). With reference to the frame
a we may then represent the vectors u and v by their coordinate vectors
   
u1 v1
u =  u2  , v =  v2  (6.5)
u3 v3

where ui = u · ai and vi = v · ai . The scalar product in terms of coordinate vectors is

u·v = (u1 a1 + u2 a2 + u3 a3 ) · (v1 a1 + v2 a2 + v3 a3 )

= u1 v1 + u2 v2 + u3 v3
= uT v

where it is used that ai · aj = δ ij which is equal to unity when i = j and zero otherwise.
6.2. VECTORS 211

The scalar product can be written in the three alternative forms

3
X
u·v = ui vi = uT v (6.6)
i=1

6.2.3 The vector cross product

The vector cross product is given in the coordinate-free form by
u × v = n|u||v| sin θ (6.7)
where 0 ≤ θ ≤ π and n is a unit vector that is orthogonal to both u and v and defined
so that (u, v, n) forms a right-hand system (Figure 6.2).
With reference to a Cartesian frame a the vector cross product can be evaluated from
¯ ¯
¯ a1 a2 a3 ¯
¯ ¯
w = u × v = ¯¯ u1 u2 u3 ¯¯ (6.8)
¯ v1 v2 v3 ¯

In component form this may alternatively be expressed by introducing the permutation

symbol 
 1 when i, j, k is a cyclic permutation
εijk = −1 when i, j, k is not a cyclic permutation (6.9)

0 when i = j, i = k or j = k
Here, as the indices {i, j, k} is a cyclic permutation if they are equal to {1, 2, 3}, {2, 3, 1}
or {3, 1, 2}, and not a cyclic permutation if they are {1, 3, 2}, {2, 1, 3} or {3, 2, 1}. It is
noted that the definition implies that
εijk = −εjik = −εikj (6.10)
εijk = εjki = εkij (6.11)
Then the components of w = u × v are given by
3 X
X 3
wi = εijk uj vk (6.12)
j=1 k=1

and the vector may be written

3 X
X 3 X
3
w= εijk ai uj vk (6.13)
i=1 j=1 k=1

In coordinate vector notation we introduce the skew-symmetric form of the coordinate

vector u defined by  
0 −u3 u2
u× :=  u3 0 −u1  (6.14)
−u2 u1 0
Then the vector cross product can be written in coordinate vector form as
 
u2 v3 − u3 v2
w = u× v =  u3 v1 − u1 v3 
u1 v2 − u2 v1
212 CHAPTER 6. RIGID BODY KINEMATICS

We sum up this result with the following three equivalent representations of the vector
cross product:

The vector cross product has the following three equivalent representations:
3 X
X 3 X
3
w = u × v ⇔ wi = εijk uj vk ⇔ w = u× v (6.15)
i=1 j=1 k=1

Example 76 Orthogonal unit vectors a1 , a2 , a3 satisfy

½
1, when i = j
ai · aj = δ ij = (6.16)
0, when i 6= j

and

a1 × a2 = a3 , a1 × a3 = −a2 (6.17)
a2 × a3 = a1 , a2 × a1 = −a3 (6.18)
a3 × a1 = a2 , a3 × a2 = −a1 (6.19)

Example 77 The relation between the components of the skew symmetric form u× and
the vector form of u can be expressed in terms of the permutation symbol as
¡ ×¢
u ik = εijk uj (6.20)
1 ¡ ¢
uj = εijk u× ik (6.21)
2

Example 78 For three arbitrary vectors a, b, c the vector cross product satisfies

a × (b × c) = ba · c − a · bc (6.22)

This can be shown by calculation of the components on both sides. Let a, b, and c be the
coordinate representations of a, b and c in some coordinate frame. The coordinate form
of (6.22) is
a× b× c = baT c − aT bc = (baT − aT bI)c (6.23)
which implies
a× b× = baT − aT bI (6.24)
In particular we note that
a× a× = aaT − aT aI. (6.25)

Example 79 From (6.25) it follows that

¡ ¢ ¡ ¢
a× a× a× = a× aaT − aT aI = − aT a a× (6.26)

where it is used that a× a = 0. In particular, if k is a unit vector, then kT k = 1, and

×
k× k× k = −k× (6.27)
6.3. DYADICS 213

Example 80 Let a, b and c be three arbitrary vectors. The Jacobi identity is written

a × (b × c) + b × (c × a) + c × (a × b) = 0 (6.28)

This identity is established from (6.22) which gives

ba · c − a · bc + cb · a − b · ca + ac · b − c · ab = 0 (6.29)

The coordinate form of the Jacobi identity is

a× b× c + b× c× a + c× a× b = 0 (6.30)

Example 81 The Jacobi identity implies that

(a × b) × c = a × (b × c) − b × (a × c) (6.31)

In coordinate form this is written

(a× b)× c = a× b× c − b× a× c (6.32)

which implies that

(a× b)× = a× b× − b× a× (6.33)

Example 82 The following problem is investigated: To what extent can the vector v be
determined when
w =u×v (6.34)
and w and u are given? In coordinate form this is written

w = u× v (6.35)

The skew symmetric matrix u× is singular, which is obvious from the identity u × u = 0
which implies that u× u = 0. This means that it is not possible to solve for v. However,
it is possible to find two equations for v. First, it is clear that w · v = 0, which means
that v is in the plane orthogonal to w. Second, it is found that

w |w|
w= sin θ|u||v| ⇒ |v| = (6.36)
|w| |u| sin θ

This shows that if the angle θ between u and v is selected to be some value, then the
length of v is given by (6.36).

6.3 Dyadics
6.3.1 Introduction
The idea of using vectors in mathematical modelling of physical systems is well known.
Also the use of column vectors to represent vectors is easy to accept. In analogy with
this it turns out that certain matrices can be the representation of physical quantities
described by pairs of vectors. Such matrices play an important role in rigid body dy-
namics and fluid mechanics, and it is worthwhile to invest some time in developing the
required formalism.
214 CHAPTER 6. RIGID BODY KINEMATICS

6.3.2 Introductory example: The inertia dyadic

The angular momentum of a rigid body about its center of mass, which will be discussed
in great detail in Section 7.3.2, can be written in coordinate-free form as a vector h, or it
may be written in terms of its coordinates as a column vector h or the generic component
hi , where  
X3 h1
h= hi ai , h =  h2  (6.37)
i=1 h3
Likewise the angular velocity can be represented by a vector ω, by a column vector ω,
or by the generic component ω i , where
 
X3 ω1
ω= ω i ai , ω =  ω 2  (6.38)
i=1 ω3
A standard result in rigid body dynamics (Section 7.3.2) is that the angular momentum
can be expressed by the angular velocity according to the two alternative formulations
3
X
h = Mω, hi = mij ω j (6.39)
j=1

where M = {mij } is the inertia matrix of the rigid body about its center of mass. The
first formulation gives the relation between the column vectors h and ω, and the other
formulation presents the relation between the generic components hi and ω j . At this
stage one might wonder: Is there a corresponding equation for the relation between h
and ω in coordinate-free form? This turns out to be the case, but to be able to do this
we need to introduce the concept of a dyadic, which is the sum of pairs of vectors. We
define the inertia dyadic by
X3 X3
M := mij ai aj (6.40)
i=1 j=1

Note that ai aj is a pair of vectors which should not be confused with the scalar product
ai · aj . Consider the following calculation:
3 X
X 3 3
X
M ·ω = mij ai aj · ω k ak
i=1 j=1 k=1
3 X
X 3 X
3
= mij ai (aj · ω k ak )
i=1 j=1 k=1
3 X
X 3
= mij ω j ai (6.41)
i=1 j=1

Here we have used the result aj · ak = δ jk . Comparing with (6.37) and (6.39) we see that
this implies that
h=M ·ω (6.42)
which is the relation between h and ω in vector notation. This result is equivalent to the
expressions in (6.39). We note that the dyadic M represents the same physical quantity
as the inertial matrix M and the components mij . We conclude that:
6.3. DYADICS 215

The angular momentum vector h can be expressed by the angular velocity vector ω with
the three equivalent formulations
3
X
h=M ·ω ⇔ h = Mω ⇔ hi = mij ω j (6.43)
j=1

where M is the inertia dyadic and M is the inertia matrix, which is the matrix represen-
tation of the inertia dyadic.

6.3.3 Matrix representation of dyadics

We define the dyadic D to be a linear combination of pairs of vectors ai aj given by
3 X
X 3
D= dij ai aj (6.44)
i=1 j=1

where
dij = ai · D · aj (6.45)
are the components of the dyadic D in frame a. The matrix

D = {dij } (6.46)

is said to be the the matrix representation of the dyadic D in frame a. Scalar premul-
tiplication with a vector, that is the scalar product of the vector u with the dyadic D
gives a vector according to
3
X 3 X
X 3
w = u·D = uk ak · dij ai aj
k=1 i=1 j=1
3 X
X 3
= dij ui aj (6.47)
i=1 j=1

Scalar postmultiplication with a vector, which is the scalar product of a dyadic with a
vector gives the vector
3 X
X 3 3
X
z = D·u= dij ai aj · uk ak (6.48)
i=1 j=1 k=1
3 X
X 3
= dij uj ai (6.49)
i=1 j=1

We define the column vectors w = (w1 , w2 , w3 )T and z = (z1 , z2 , z3 )T corresponding to

the vectors w and z. We may then write the equivalent expressions

w =u·D ⇔ wT = uT D (6.50)

z =D·u ⇔ z = Du (6.51)
216 CHAPTER 6. RIGID BODY KINEMATICS

The identity dyadic is defined by

3 X
X 3
I := δ ij ai aj = a1 a1 + a2 a2 + a3 a3 (6.52)
i=1 j=1

where δ ij is equal to unity when i = j, and zero otherwise. This implies that for any
vector u
I ·u=u·I =u (6.53)
and for any dyadic D we have
I ·D =D·I =D (6.54)
The equivalent matrix form of these equations are
¡ ¢T
Iu = uT I = u (6.55)
ID = DI = D (6.56)

Example 83 Let ω be a vector and let M be a dyadic. Define the scalar

1
K= ω·M ·ω (6.57)
2
which is defined independently of any coordinate frame. Let ω be given in the a frame by
ω = ω 1 a1 + ω 2 a2 + ω 3 a3 , and let the dyadic be given by
3 X
X 3
M= mij ai aj . (6.58)
i=1 j=1

Then the quadratic form is found to be

3 3
1 XX
K= ω i ω j mij (6.59)
2 i=1 j=1

The corresponding representation in matrix form is given by

1 T
K= ω Mω (6.60)
2

where M = {mij } is the matrix representation of the dyadic M .

Example 84 Consider the dyadic K := kk. Let w be an arbitrary vector. Then

w · K = (w · k)k and K · w = k(k · w) (6.61)

The coordinate form is

wT K = wT kkT and Kw = kkT w (6.62)

It follows that the matrix form of K = kk is

K = kkT (6.63)
6.3. DYADICS 217

Example 85 The dyadic form of the vector cross product is

u × v = u× · v = u · v × (6.64)
where
3 X
X 3 X
3
u× = εijk uj ai ak (6.65)
i=1 j=1 k=1

is the dyadic form of the skew symmetric form u× , and

X 3 X
3 X 3
v× = εijk vj ai ak (6.66)
i=1 j=1 k=1

We may then check that

X 3 X
3 X 3 3
X 3 X
X 3 X
3
u× · v = εijk uj ai ak · vp ap = εijk ai uj vk = u × v (6.67)
i=1 j=1 k=1 p=1 i=1 j=1 k=1

3
X 3 X
X 3 X
3 3 X
X 3 X
3
u · v× = up ap · εijk vj ai ak = εijk ak ui vj (6.68)
p=1 i=1 j=1 k=1 i=1 j=1 k=1

X 3 X
3 X 3 3 X
X 3 X
3
= εkij ak ui vj = εijk ai uj vk = u × v (6.69)
i=1 j=1 k=1 i=1 j=1 k=1

The skew symmetric form used in the coordinate vector form is consistent with (6.64) as
¡ ¢T ¡ ¢T
u× v = − v× u = v× u = uT v× (6.70)
This shows that the matrix representation of cross product dyadic u× is the skew sym-
metric form u× .
Example 86 The dyadic form of the triple cross product (6.22) is
h i
a× · b× · c = ba − (a · bI) · c (6.71)
and it follows that
a× · b× = ba − a · bI (6.72)
In particular, we note that
a× · a× = aa − a · aI (6.73)
Example 87 The triple scalar product satisfies
³ ´
d × a · w = d · (a × w)

for any vectors d, a and w, and it follows from (6.71) that

³ ´ ³ ´ ³ ³ ´´ ³ ´
d × a · b × c = d · a × b × c = d · a · cI − ca · b (6.74)
The same result could have been obtained using
³ ´ ³ ´ ³ ´ ³ ´ ¡ ¢
d × a · b × c = − d × a · c × b = −d · a× · c× · b = −d · a× · c× · b (6.75)
In the development of the kinetic energy for a rigid body the following special case will be
used: ³ ´ ¡ ¢
(ω × r) · (ω × r) = ω · r · rI − rr · ω = −ω · r× · r× · ω (6.76)
218 CHAPTER 6. RIGID BODY KINEMATICS

za zb

b3

b2
3
a yb

b1
xb

 2
a ya
a1
xa

Figure 6.3: Frames a and b.

6.4 The rotation matrix

6.4.1 Coordinate transformations for vectors
It was shown that a vector can be described in terms of its component in a coordinate
frame a with orthogonal unit vectors a1 , a2 , a3 . Dynamic models for use in robotics, car
dynamics, aerospace, marine systems, and navigation typically involve several Cartesian
frames, so that a vector may have to be described in more than one frame. To investigate
this we introduce a second coordinate frame b with orthogonal unit vectors b1 , b2 , b3 along
the axes. A vector v may then be represented with respect to any of the systems a and
b. We use the notation
X 3 X3
v= via ai and v = vib bi (6.77)
i=1 i=1

where
via = v · ai (6.78)
are the coordinates of v in a, and
vib = v · bi (6.79)
are the coordinates of v in b. To distinguish the column vectors of coordinates in frame
a from the column vector of coordinates in frame b we write
 a   b 
v1 v1
va =  v2a  and vb =  v2b  (6.80)
v3a v3b

where superscript a denotes that the vector is given by the the coordinates in a, and the
superscript b denotes that the vector is given by the coordinates in b.
To find the relation between the coordinate vectors va and vb in frames a and b the
following calculation is used:

via = v · ai = (v1b b1 + v2b b2 + v3b b3 ) · ai

X3 ³ ´
= vjb ai · bj (6.81)
j=1

This leads to the following result:

6.4. THE ROTATION MATRIX 219

The coordinate transformation from frame b to frame a is given by

va = Rab vb (6.82)

where
Rab = {ai · bj } (6.83)
is called the rotation matrix from a to b. The elements rij = ai · bj of the rotation matrix
Rab are called the direction cosines.

We see that the rotation matrix from a to b transforms a coordinate vector in b to

a coordinate vector in a. Because of this the matrix may also be called the coordinate
transformation matrix from b to a.

6.4.2 Properties of the rotation matrix

The rotation matrix has a number of useful properties that will be discussed in this
section. First it is noted that the rotation matrix from b to a can be found in the same
way as the rotation matrix from a to b by simply interchanging a and b in the expressions.
This gives
Rba = {bi · aj } (6.84)
For all vb we have
vb = Rba va = Rba Rab vb (6.85)
This implies that
Rba Rab = I, (6.86)
and it follows that
Rba = (Rab )−1 (6.87)
A comparison of the elements in the matrices in (6.83) and (6.84) leads to the conclusion
that Rba = (Rab )T . Combining these results we arrive at the first result:

The rotation matrix is orthogonal and satisfies

−1
Rba = (Rab ) = (Rab )T (6.88)

Consider a vector p with coordinate vector pa in frame a. Define the vector q defined
by its coordinate vector
qa = Rab pa (6.89)
Note that the vector q is defined by the vector p and the rotation matrix Rab . The
coordinate vector qb in b is according to the usual coordinate transformation rule

qb = Rba qa = Rba Rab pa = pa (6.90)

which means that the coordinates of q in b are equal to the coordinates of p in a. This
is the second result: The rotation matrix from a to b rotates the vector p to the vector
q so that qb = pa .
220 CHAPTER 6. RIGID BODY KINEMATICS

The rotation matrix Rab from a to b has two interpretations:

1. Let the vector v have coordinate vector vb in b and coordinate vector va in a. Then
the rotation matrix Rab transforms the coordinate vector in b to the coordinate
vector in a according to
va = Rab vb (6.91)
In this equation Rab acts as a coordinate transformation matrix.
2. The vector p with coordinate vector pa in a is rotated to the vector q with coordi-
nate vector qb = pa by
qa = Rab pa (6.92)
In this equation Rab acts as a rotation matrix.

As a special case of this the rotation matrix rotates the orthogonal unit vectors
a1 , a2 , a3 in a to the orthogonal unit vectors b1 , b2 , b3 in b which is seen from
     
1 0 0
aa1 = bb1 =  0  , aa2 = bb2 =  1  and aa3 = bb3 =  0  (6.93)
0 0 1

Moreover, from bai = Rab aai it follows that the columns of the rotation matrix are the
coordinate vectors bai of bi in frame a, that is
¡ ¢
Rab = ba1 ba2 ba3 (6.94)

which is the third result.

The determinant of the rotation matrix Rab is found by direct calculation to be

det Rab = r11 (r22 r33 − r32 r23 ) + r21 (r32 r13 − r12 r33 ) + r31 (r12 r23 − r22 r13 )
h i
= (ba1 )T (ba2 )× ba3 = (ba1 )T ba1 = 1

where it is used that (ba2 )× ba3 = ba1 , and that ba1 is a unit vector. We have then shown
the fourth result: The rotation matrix has a determinant equal to unity, that is

det Rab = 1 (6.95)

Finally, the set SO(3) is defined. We have established that the rotation matrix
is orthogonal and has a determinant equal to unity. The set of all matrices that are
orthogonal and with a determinant equal to unity is denoted by SO(3), that is,

SO(3) = {R|R ∈ R3×3 , RT R = I and det R = 1} (6.96)

Here R3×3 is the set of all 3 × 3 matrices with real elements. A matrix R is a rotation
matrix if and only if it is an element of the set SO(3).

6.4.3 Composite rotations

The rotation from frame a to a frame c may be described a a composite rotation made
up by a rotation from a to b, and then a rotation from b to c. The transformation of vc
6.4. THE ROTATION MATRIX 221


b3 
a3



b2



a2

Figure 6.4: A rotation by an angle φ around a1 .

to b and to a is given by

vb = Rbc vc
va = Rac vc

Combining these two equations we get

va = Rab vb = Rab Rbc vc

This shows that:

The rotation matrix of a composite rotation is the product of the rotation matrices:

Rac = Rab Rbc

This shows that the rotation matrix for the composite rotation Rac is simply the
product of the rotation matrices Rab from a to b and Rbc from b to c. It is straightforward
to extend this result to the composite rotation of three or more rotations. In the case of
three rotations we have
Rad = Rab Rbc Rcd (6.97)

6.4.4 Simple rotations

A rotation about a fixed axis is called a simple rotation. We will here derive the rotation
matrices corresponding to simple rotations about the x, y and z axes. Consider a rotation
by an angle φ about the xa axis from a frame a to a frame b. The resulting rotation
matrix is denoted Rx (φ). In the same way we define Ry (θ) to be the rotation by an
angle θ about the y axis, and Rz (ψ) to be the rotation by an angel ψ about the z axis.
For the rotation Rx (φ) we see from Figure 6.4 that a1 = b1 , so that a1 · b1 = 1, while

a1 · b2 = a1 · b3 = a2 · b1 = a3 · b1 = 0 (6.98)
a2 · b2 = cos φ, a3 · b3 = cos φ (6.99)
a3 · b2 = sin φ, a2 · b3 = − sin φ (6.100)
222 CHAPTER 6. RIGID BODY KINEMATICS

In the same way we can find the elements of the matrices Ry (θ) and Rz (ψ). This results
in
 
1 0 0
Rx (φ) =  0 cos φ − sin φ  (6.101)
0 sin φ cos φ
 
cos θ 0 sin θ
Ry (θ) =  0 1 0  (6.102)
− sin θ 0 cos θ
 
cos ψ − sin ψ 0
Rz (ψ) =  sin ψ cos ψ 0  (6.103)
0 0 1

6.4.5 Coordinate transformations for dyadics

A coordinate vector can be transformed from a frame a to a frame b through multipli-
cation with the rotation matrix. An important property of dyadics is that the matrix
representation transform from frame a to frame b with a similarity transformation using
the rotation matrix. The dyadic D can be expressed in frames a and b by
X 3
3 X 3 X
X 3
D= daij ai aj = dbpq bp bq (6.104)
i=1 j=1 p=1 q=1

where daij are the components in frame a and dbpq are the components in frame b. The
matrix representation in the two frames are denoted
Da = {daij }, Db = {dbij } (6.105)
Let the vector z be given by
z =D·u (6.106)
Then, in frames a and b this may be written in matrix form as
za = Da ua , zb = Db ub (6.107)
We then find that
Da ua = za = Rab zb = Rab Db ub = Rab Db Rba ua (6.108)
and, since ua is arbitrary, this implies that

The matrix representation of a dyadic transforms by a similarity transform with the

rotation matrix according to
Da = Rab Db Rba (6.109)

Example 88 In rigid body dynamics a frame b with orthogonal unit vectors b1 , b2 , b3 is

fixed in the rigid body. Then the inertia dyadic can be written
3 X
X 3
M= mbij bi bj (6.110)
i=1 j=1
6.4. THE ROTATION MATRIX 223

and the corresponding matrix representation is

© ª
Mb = mbij (6.111)

An important result in rigid body dynamics is that when frame b is fixed in the rigid body,
and therefore moves with the rigid boy, then Mb is a constant matrix. In contrast to this,
the matrix representation Ma in a stationary coordinate frame a will be given by

Ma = Rab Mb Rba (6.112)

Example 89 The relation between the skew symmetric forms of a vector is given by
¡ b ¢× b × ×
u v = wb = Rba wa = Rba (ua ) va = Rba (ua ) Rab vb (6.113)

which implies that

¡ b ¢× ×
u = Rba (ua ) Rab (6.114)

that is, the skew symmetric form of the vector ua transforms to frame b by a similarity
transformation. This is a consequence of the fact that (ua )× is the matrix representation
of the dyadic u× .

6.4.6 Homogeneous transformation matrices

By now we are familiar with the notion of a rotation matrix which specifies the orien-
tation of a coordinate frame with respect to some other frame. To extend our set of
mathematical tools we introduce the concept of a homogeneous transformation matrix
which is a matrix that describes the position and orientation of a coordinate frame with
respect to a reference frame. To be precise we consider a frame a and a frame b, and let
Rab be the rotation matrix from a to b, while raab is the position in a coordinates of the
origin of frame b relative to the origin of frame a.

The position and orientation of frame b relative to frame a is given by the homogeneous
transformation matrix µ a ¶
a Rb raab
Tb = ∈ SE(3) (6.115)
0T 1

Here the set SE(3) is the Special Euclidean Group of dimension 3 defined by
½ µ ¶ ¾
R r
SE(3) = T|T= , R ∈ SO(3), r ∈ R3 (6.116)
0T 1

The inverse of Tab is found by matrix inversion to be

µ ¶ µ ¶
(Rab )T − (Rab )T raab Rba rbba
(Tab )−1 = = = Tba (6.117)
0T 1 0T 1

This means that

(Tab )−1 = Tba (6.118)
224 CHAPTER 6. RIGID BODY KINEMATICS

Composite homogenous transformation matrices give

µ a ¶µ b ¶
Rb raab Rc rbbc
Tab Tbc =
0T 1 0T 1
µ a b a a b
¶
Rb Rc rab + Rb rbc
=
0T 1
µ a a
¶
Rc rac
=
0T 1
a
= Tc (6.119)
We conclude that
Tac = Tab Tbc (6.120)
Example 90 In the description of robotic manipulators, a coordinate frame is fixed to
each link of the arm. The manipulator is made up by rigid bodies called links that are
connected by joints. Each joint is assumed to have one degree of freedom that is either a
translation or a rotation. In a typical manipulator design with rotary joints we may think
of link 1 as the torso that is connected to the upper arm (link 2) by a shoulder joint with a
horizontal axis of rotation. The upper arm is in turn connected to the lower arm (link 3)
by an elbow joint. The lower arm is connected with the robot hand (joint 6) trough three
rotary joints that form the robotic wrist. In general, the base frame 0 is fixed to the floor,
frame 1 is fixed to the first link, frame 2 to the second link and so on, and for a six link
manipulator frame 6 is fixed to the robot hand. The position and orientation of frame
i + 1 relative to frame i can then be specified in terms of a homogeneous transformation
matrix µ i ¶
i Ri+1 rii,i+1
Ti+1 = (6.121)
0T 1
and the position and orientation of the hand is given by
µ 0 ¶
R6 r006
T06 = (6.122)
0T 1
which is computed from
T06 = T01 T12 . . . T56 . (6.123)
In the Denavit-Hartenberg convention the transformation from frame i to frame i + 1 is
given by the Denavit-Hartenberg parameters αi , ai , di , θi according to
µ ¶µ ¶µ ¶
Rz (θi ) 0 I ai e1 + di e3 Rx (αi ) 0
Tii+1 =
0T 1 0T 1 0T 1
µ ¶
Rz (θi ) Rx (αi ) ai Rz (θi ) e1 + di e3
= (6.124)
0T 1
T T
where e1 = (1, 0, 0) and e3 = (0, 0, 1) . The joint variable is di for translational joints
and θi for rotational joints.

6.5 Euler angles

6.5.1 Introduction
A rotation matrix describes the orientation of a frame b with respect to a frame a.
The rotation matrix is a 3 × 3 matrix with nine elements. The orthogonality of the
6.5. EULER ANGLES 225

  za
zb

yb




ya
xa 


xb

Figure 6.5: Roll-pitch-yaw Euler angles.

matrix gives six constraints on the elements of the matrix, so that there are only three
independent parameters that describes the rotation matrix. Therefore, it is of great
interest to investigate if it is possible to find three parameters that give a parameterization
of the rotation matrix.
A widely used set of parameters for the rotation matrix is the Euler angles. In this
description the rotation matrix is given as a composite rotation of selected combinations
of rotations about the x, y and z axes. There are many possible permutations of x, y and
z rotations, and a description of this is given in (Kane et al. 1983). Here we will present
the two sets of Euler angles that are the most often seen, namely the roll-pitch-yaw
angles, and the classical Euler angles.

6.5.2 Roll-pitch-yaw
The Euler angles of the roll-pitch yaw type are commonly used to describe the motion of
rigid bodies that move freely, like aeroplanes, spacecraft, ships and underwater vehicles.
The rotation from a to b is described as a rotation ψ about the za axis, then a rotation θ
about the current (rotated) y axis, and finally a rotation φ about the current (rotated)
x axis as shown in Figure 6.5. The resulting rotation matrix is

Rab = Rz (ψ)Ry (θ)Rx (φ) (6.125)

This formulation is very useful as it makes it possible to describe the rotation of

e.g. an airplane as a sequence of a roll rotation about the longitudinal axis of the plane,
then a pitch rotation about a lateral axis of the plane, and finally a yaw rotation about
the vertical axis of the plane. Obviously, it is easier to interpret this sequence of simple
rotations angle than a rotation matrix.
To derive and remember the expression (6.125) it is convenient to use the rotation
matrix interpretation of the simple rotations Rz (ψ), Ry (θ) and Rx (φ). In this interpre-
tation the rotation from a to b is in the same sequence as the matrices are written in
(6.125), namely, first Rz (ψ), then Ry (θ), and finally Rx (φ). Note that in a coordinate
226 CHAPTER 6. RIGID BODY KINEMATICS

 za
zb

 yb

 

ya
xa

 
xb

Figure 6.6: Classical Euler angles.

transformation interpretation the matrix Rab transforms a vector vb to a vector va ac-

cording to va = Rab vb = Rz (ψ)Ry (θ)Rx (φ)vb . Then the vector vb is first transformed
by Rx (φ), then by Ry (θ) and finally by Rz (ψ).

6.5.3 Classical Euler angles

The classical Euler angles are used to describe the rotation of rigid bodies that are
connected to a fixed base by three joints. Typically this involves robotic wrist joints,
platforms stabilized by gyroscopes in inertial navigation, and pointing devices. In this
description the rotation consists of a rotation ψ about the za axis, then a rotation θ
about the current (rotated) y axis, and finally a rotation φ about the current (rotated)
z axis as shown in Figure 6.6. The resulting rotation matrix is

Rab = Rz (ψ)Ry (θ)Rz (φ) (6.126)

6.6 Angle-axis description of rotation

6.6.1 Introduction
In the previous section it was shown that the rotation matrix can be represented by
Euler angles, which are very useful in some applications. In particular this is the case
for a robotic wrist joint where the hand is connected to the arm through three rotational
joints. Also in ship dynamics it is convenient to describe the rotation of the ship in
terms of the Euler angles roll, pitch and yaw. Likewise airplane dynamics rely on a
characterization based on the roll angle, the pitch angle and the sideslip angle, which are
the Euler angles from the wind frame to the airplane frame. The motivation for this is
that the forces acting on the plane are functions of these Euler angles. However, in many
other applications involving rotation there is no clear physical motivation for introducing
Euler angles. The use of Euler angles in the equations of motion may then introduce
complicated expressions with inherent singularities. There are alternative descriptions
6.6. ANGLE-AXIS DESCRIPTION OF ROTATION 227

of rotation that avoid these problems, and that are well suited for simulation as well as
for controller design and analysis. On background of this it may be argued that Euler
angles have been over-emphasized in the dynamics literature. In the following we will
study the angle-axis parametrization of the rotation, which is a very useful tool in the
development of kinematic models and equations of motion for use in control systems.

6.6.2 Angle-axis parameters

A rotation matrix Rab is orthogonal with determinant equal to unity. It can be shown
(Angeles 1988), (McCarthy 2000) that this implies that one of the eigenvalues to the
matrix is equal to one, and that the corresponding unit eigenvector k satisfies
Rab k = k (6.127)
This purely algebraic result can be given a geometric interpretation which is the basis for
the angle-axis parameterization of the rotation matrix Rab . The geometric interpretation
that will be used is that the eigenvector k is the coordinate vector of a unit vector k,
where k is defined by its coordinate vector
ka = k (6.128)
in frame a. The transformation rule
ka = Rab kb (6.129)
then implies that
ka = kb = k (6.130)
which means that k has the same coordinates in frames a and b. It is therefore possible
to describe the rotation from a to b as a simple rotation by an angle θ about the vector
k which is fixed in both a and b. On background of this (θ, k) is called the angle-axis
parameterization of the rotation matrix Rab . Note that this gives four parameters and
one constraint equation, namely the angle θ plus the three coordinates of the unit vector
k, and the constraint equation k · k = 1.

6.6.3 Derivation of rotation dyadic

We will here derive the expression for the rotation matrix given by the angle θ and the
vector k. The derivation is taken from (Kane et al. 1983). It was shown in Section 6.4.2
that the rotation matrix Rab rotates a vector p in a to a vector q in b so that the coordinates
of p in a are equal to the coordinates of q in b. This result is used to find an expression
for the rotation matrix Rab in terms of θ and k. We will do this by deriving an expression
where q is given by p, θ and k. To simplify the derivation two additional frames c and d
are used where Rac = Rbd .
Consider two frames c and d that initially coincide. Let c1 , c2 , c3 be the orthogonal
unit vectors in c, and let d1 , d2 , d3 be the orthogonal unit vectors in d. The frames are
selected so that c3 = d3 = k. Frame d is obtained by rotating frame c by an angle θ
about k as shown in Figure 6.7. Let the vector p be a fixed vector in frame c, and let the
vector q be a fixed vector in frame d so that p and q coincide before the rotation. Then
it is possible to express q after the rotation by p, θ and k.
The vectors p and q can be written
p = xc1 + yc2 + z k (6.131)
228 CHAPTER 6. RIGID BODY KINEMATICS

L

p

q

k


d2
 
d1
c1
c2

Figure 6.7: Rotation of the vector p by an angle θ around the k vector.

and
q = xd1 + y d2 + z k (6.132)
The unit vectors d1 and d2 can be written

d1 = cos θc1 + sin θc2 (6.133)

d2 = − sin θc1 + cos θc2 (6.134)

and insertion into (6.132) gives

q = (x cos θ − y sin θ)c1 + (x sin θ + y cos θ)c2 + z k (6.135)

Consider the calculation

cos θp + sin θk × p + (1 − cos θ)k k · p

= cos θ(xc1 + yc2 + z k) − sin θ(−xc2 + yc1 ) + (1 − cos θ)z k
= (x cos θ − y sin θ)c1 + (x sin θ + y cos θ)c2 + z k (6.136)

where it is used that c1 × k = −c2 , c2 × k = c1 , and that p· k = z. It follows by comparison

with (6.135) that

q = cos θp + sin θk × p + (1 − cos θ)k k · p

³ ´
= cos θI + sin θk × + (1 − cos θ)k k · p (6.137)

6.6.4 The rotation dyadic

The rotation of the vector p to the vector q as given by (6.137) can be written in the
dyadic form
q = Rk,θ · p (6.138)
6.6. ANGLE-AXIS DESCRIPTION OF ROTATION 229

where
Rk,θ = cos θI + sin θk× + (1 − cos θ)k k (6.139)
is the rotation dyadic of the angle-axis description. Now, recall that q is obtained by
rotating p according to qa = Rba pa , which in combination with (6.138) implies that the
rotation matrix Rba is the matrix representation of the rotation dyadic Rk,θ in a. This
leads to the result

The rotation matrix Rab can be described as a rotation by an angle θ about a unit vector
k where Rab is given by

Rab = cos θI + sin θ(ka )× + (1 − cos θ)ka (ka )T (6.140)

This is the angle-axis parameterization of the rotation matrix.

Using the standard transformation rule and the identities (ka )× ka = 0 and (ka )T ka =
1 gives
ka = Rab kb = kb (6.141)
which shows that the rotation vector k has the same coordinates in a and b.
Example 91 Inserting ka = (kx ky kz )T we get
 
kx2 vθ + cθ kx ky vθ − kz sθ kx kz vθ + ky sθ
Rb =  kx ky vθ + kz sθ
a
ky2 vθ + cθ ky kz vθ − kx sθ  (6.142)
kx kz vθ − ky sθ ky kz vθ + kx sθ kz2 vθ + cθ
where the notation sθ = sin θ, cθ = cos θ and vθ = 1−cθ is used to simplify the expression.
Example 92 Suppose that the rotation axis is given by k = a3 , which means that ka =
(0, 0, 1)T . Then the matrix representation of Rk,θ in a is
 
cos θ − sin θ 0
Rab =  sin θ cos θ 0  = Rz,θ (6.143)
0 0 1
which is to be expected as this is a rotation by an angle θ about the z axis.
Example 93 As Rk,θ is a dyadic and k = c3 , it follows from the transformation rule
(6.109) that
Rab = Rac Rz,θ Rca (6.144)

6.6.5 Rotation matrix

We use the notation Rk,θ = Rab and k = ka so that the rotation matrix is written
Rk,θ := cos θI + k× sin θ + kkT (1 − cos θ) (6.145)
An alternative expression is found by inserting the identity
k× k× = kkT − kT kI = kkT − I. (6.146)
which gives
Rk,θ = I + k× sin θ + k× k× (1 − cos θ) (6.147)
The inverse to the rotation matrix is
(Rab )T = Rba = Rk,−θ . (6.148)
230 CHAPTER 6. RIGID BODY KINEMATICS

Example 94 The derivative of Rk,θ with respect to θ is found from (6.147) to be

dRk,θ
= k× cos θ + k× k× sin θ (6.149)
dθ
Using (6.27) we find that

k× Rk,θ = k× + k× k× sin θ − k× (1 − cos θ) (6.150)

and we may conclude that

d
Rk,θ = k× Rk,θ (6.151)
dθ
d
Example 95 It is known that the differential equation dt x = Ax has the solution x(t) =
x(0) exp(At) when A is a constant matrix. When k is a constant vector the solution of
(6.151) is found in the same way to be

Rk,θ = exp[k× θ] (6.152)

as Rk,θ (θ = 0) = I.

Example 96 The matrix exponential exp(A) for a quadratic matrix A is defined by

1 2 1
exp(A) = I + A + A + A3 . . . (6.153)
2! 3!
The result (6.152) can be derived directly from (6.147). First we use (6.27) to establish
the identity
(k× )2n+1 = (−1)n k× (6.154)
by induction. This is done by noting that (6.27) implies that (6.154) is true for n = 1,
and moreover, for n = 1, 2, . . . we have

(k× )2n+1 = (−1)n k× ⇒ (k× )2(n+1)+1 = (−1)n k× (k× )2 = (−1)n (−1)k× . (6.155)

Then we may evaluate exp[k× θ] directly from the definition (6.153), and find that

θ2 θ3 θ4 θ5 θ6
exp[k× θ] = I + k× θ + (k× )2 + (k× )3 + (k× )4 + (k× )5 + (k× )6 . . .
2! 3! 4! 5! 6!
2 3 4 5 6
θ θ θ θ θ
= I + k× [θ + k× ] + (k× )3 [ + k× ] + (k× )5 [ + k× ] . . .
2! 3! 4! 5! 6!
2 3 4 5 6
θ θ θ θ θ
= I + k× [θ + k× ] − k× [ + k× ] + k× [ + k× ] . . .
2! 3! 4! 5! 6!
3 5 2 4 6
θ θ θ θ θ
= I + k× [θ − + . . .] − (k)2 [ − + . . .]
3! 5! 2! 4! 6!
¡ ¢ 2
= I + k× sin θ + k× (1 − cos θ) (6.156)

This shows in view of (6.147) that

Rk,θ = exp[k× θ] (6.157)

which is in agreement with the previously derived result (6.152).

6.7. EULER PARAMETERS 231

6.7 Euler parameters

6.7.1 Definition
The Euler parameters were introduced by Euler in 1770, and are essentially the same as
the unit quaternions that were devised by Hamilton in on October 16, 1843, and which
involved the definition of a complex number with one real part and three imaginary parts.
The Euler parameters have no singularities, and give rational expressions for the rotation
matrix as opposed to the angle/axis parameters, which lead to trigonometric terms in
the expressions for the rotation matrix. The Euler parameters are of particular use in the
numerical simulation of rotation, and in stability analysis of attitude control systems.
The Euler parameters are defined in terms of the angle-axis parameters θ and k, and
are given by the scalar η and the vector defined by
θ θ
η = cos , = k sin (6.158)
2 2
In coordinate form this is written
θ θ
η = cos , ² = k sin (6.159)
2 2
We note that
θ θ
η 2 + · = η 2 + ²T ² = cos2 + sin2 = 1 (6.160)
2 2
Insertion of the trigonometric identities
θ θ
sin θ = 2 sin cos (6.161)
2 2
2 θ θ θ θ
cos θ = cos − sin2 = 2 cos2 − 1 = 1 − 2 sin2 (6.162)
2 2 2 2
into (6.145) makes it possible to express the rotation matrix Rk,θ in terms of the Euler
parameters Rk,θ = Re (η, ²), where
Re (η, ²) = (η 2 − ²T ²)I + 2²²T + 2η²× (6.163)
= (2η 2 − 1)I + 2²²T + 2η²× (6.164)
= (1 − 2²T ²)I + 2²²T + 2η²× (6.165)
Here it is used that η 2 = cos2 2θ , ²T ² = sin2 θ2 and sin θ2 cos θ2 k× = η²× . From (6.165)
and (6.146) an alternative form of the rotation matrix is found.

The rotation matrix is given by the corresponding Euler parameters according to

Re (η, ²) = I + 2η²× + 2²× ²× (6.166)

A given rotation will correspond to two sets of Euler parameters (η, ) and (−η, − ) with
opposite signs as
Re (−η, −²) = Re (η, ²) (6.167)
The inverse of Re (η, ) given by

Re (η, ²)T = Re (η, −²) (6.168)

corresponds to the Euler parameters (η, − ).

232 CHAPTER 6. RIGID BODY KINEMATICS

Example 97 From (6.165) and (6.160) we find that

TraceR = 3(η 2 − ²T ²) + 2²T ² = 4η 2 − 1 (6.169)

6.7.2 Quaternions
The vector µ ¶
η
p= (6.170)
²
of Euler parameters can be treated as a unit quaternion vector. This makes it possible
to introduce a wealth of techniques and analysis tool from the theory of quaternions.
In the following, the necessary background on quaternions will be presented, and this
will be specialized to unit quaternions representing a rotation matrix through its Euler
parameters.
A quaternion is represented by a vector
µ ¶
α
q= (6.171)
β

of dimension 4 where α is the scalar part and β = (β 1 , β 2 , β 3 )T is the vector part.

The quaternion product between two quaternion vectors q1 = (α1 β T1 )T and q2 =

(α2 βT2 )T in R4 is defined by
µ ¶ µ ¶ µ ¶
α1 α2 α1 α2 − β T1 β2
⊗ = (6.172)
β1 β2 α1 β 2 + α2 β 1 + β ×1 β2

where α1 , α2 ∈ R and β1 , β 2 ∈ R3 .

Example 98 The commutator of the quaternion product is given by

µ ¶ µ ¶ µ ¶ µ ¶ µ ¶ µ ¶
α1 α2 α2 α1 0 0
⊗ − ⊗ =2 ⊗ (6.173)
β1 β2 β2 β1 β1 β2

where it is used that β× ×

1 β 2 = −β 2 β 1 .

Example 99 Define the matrices

µ ¶
α −βT
F(q) = ∈ R4×4 (6.174)
β αI + β×
µ ¶
α −βT
E(q) = ∈ R4×4 (6.175)
β αI − β×

The matrix F(q) represents quaternion pre-multiplication with q in the sense that for
any u ∈ R4
q ⊗ u = F(q)u (6.176)
while E(q) represents quaternion post-multiplication with q in the sense that

u ⊗ q = E(q)u (6.177)
6.7. EULER PARAMETERS 233

Example 100 The concept of quaternions was introduced by Hamilton who got the idea
on the 16th of October 1843 while he was walking with his wife to the Royal Irish Academy
(der Waerden 1976). In Hamilton’s formulation the quaternion was written

q = α + iβ 1 + jβ 2 + kβ 3 (6.178)

where i, j and k are imaginary units satisfying

i2 = j 2 = k2 = −1 (6.179)

ij = −ji = k, jk = −kj = i, ki = −ik = j (6.180)

It is interesting to see that the quaternion is actually an extension of complex numbers
z = a + ib where i2 = −1. The complex numbers form a division algebra, which means
that if z1 and z2 are complex numbers, then the sum z1 + z2 , the difference z1 − z2 ,
and the product z1 z2 are complex numbers, and likewise z1 /z2 is a complex number if
z2 6= 0. In addition, the magnitudes or moduli satisfy |z| = |z1 ||z2 |, which is referred to
as the law of the moduli. Hamilton had for a long time attempted to extend the theory
of complex numbers to triplets a + ib + jc where i and j are imaginary units, but he
was unable to achieve a division algebra that satisfy the law of the moduli. Hamilton’s
great idea was then to introduce one more complex unit, which resulted in the quaternion
a + ib + jc + dk, which lead to a division algebra where the law of the moduli was satisfied.
It is now established that this is possible for dimensions 1, 2, 4 and 8, so the attempt
to do this for triplets could not have succeeded. Hamilton’s formulation of the product
of quaternions is based on the rules for the imaginary units i, j and k. In this setting
the quaternion vectors q1 = (α1 , β 11 , β 12 , β 13 )T and q2 = (α2 , β 21 , β 22 , β 23 )T can be
represented by the quaternion numbers (Samson, Borgne and Espiau 1991)

q1 = α1 + iβ 11 + jβ 12 + kβ 13 (6.181)
q2 = α2 + iβ 21 + jβ 22 + kβ 23 (6.182)

Then the product of q1 and q2 is found to be

q1 q2 = (α1 + iβ 11 + jβ 12 + kβ 13 ) (α2 + iβ 21 + jβ 22 + kβ 23 )
= α1 α2 − β 11 β 21 − β 12 β 22 − β 13 β 23
+i (α1 β 21 + α2 β 11 + β 12 β 23 − β 13 β 22 )
+j (α1 β 22 + α2 β 12 + β 13 β 21 − β 11 β 23 )
+k (α1 β 23 + α2 β 13 + β 11 β 22 − β 12 β 21 ) (6.183)

We see that this product of quaternion numbers satisfy

q = q1 q2 (6.184)

where
q = α + iβ 1 + jβ 2 + kβ 3 (6.185)
corresponds to the quaternion vector q = q1 ⊗ q2 as defined in (6.172).

6.7.3 Unit quaternions

A unit quaternion µ ¶
η
p= (6.186)
²
234 CHAPTER 6. RIGID BODY KINEMATICS

is a quaternion with unit length, that is, a quaternion that satisfies

pT p = η 2 + ²T ² = 1 (6.187)
We see that if η and ² are Euler parameters, then p is a unit quaternion corresponding
to the rotation matrix Rη, . The unit quaternion corresponding to R−1η, = Rη,− is the
inverse unit quaternion p̄ defined by
µ ¶
η
p̄ = (6.188)
−²
The unit quaternion corresponding to the identity matrix R1,0 = I is the identity quater-
nion pid defined by µ ¶
1
pid = (6.189)
0

6.7.4 The quaternion product for unit quaternions

The quaternion product of two unit quaternions p1 and p2 is a unit quaternion
µ ¶
η 1 η 2 − ²T1 ²2
p := p1 ⊗ p2 = (6.190)
η 1 ²2 + η 2 ²1 + ²× 1 ²2

This is shown by direct computation of pT p which gives

pT p = η 21 η 22 − 2η 1 η 2 ²T1 ²2 + (²T1 ²2 )2 + η 21 ²T2 ²2 + 2η 1 η 2 ²T2 ²1 + η 22 ²T1 ²1 − ²T2 ²× ×
1 ²1 ²2
2 T 2 T
= (η 1 + ²1 ²1 )(η 2 + ²2 ²2 ) = 1
where it is used that ²× × × T T 2 T
1 ²2 is orthogonal to ²1 and ²2 , ²1 ²1 = ²1 ²1 −²1 ²1 I, η 1 +²1 ²1 = 1
2 T
and η 2 + ²2 ²1 = 1.
It is straightforward to check that the quaternion product of p and the inverse p̄ is
the identity quaternion, that is,
p ⊗ p̄ = p̄ ⊗ p = pid (6.191)
This follows from
µ ¶ µ ¶ µ ¶ µ ¶
η η η 2 + ²T ² 1
p ⊗ p̄ = ⊗ = = = pid (6.192)
² −² η² − η² − ²× ² 0
In the same way it can be shown that p̄ ⊗ p = pid .
We may also verify that
p ⊗ pid = pid ⊗ p = p (6.193)
Example 101 The unit quaternion satisfies
µ ¶ µ ¶ µ ¶
η̇ η 0
ṗ ⊗ p̄ = ⊗ = (6.194)
²̇ −² −η̇² + η ²̇ + ²× ²̇
where we have used
1 d ¡ 2 ¢
η̇η + ²̇T ² = η + ²T ² = 0 (6.195)
2 dt
This result is used in the derivation of the kinematic differential equation for unit quater-
nions.
Example 102 The time derivative of the inverse unit quaternion is found from
p ⊗ p̄ = pid ⇒ ṗ ⊗ p̄ + p ⊗ p̄˙ = 0 (6.196)
which implies
p̄˙ = −p̄ ⊗ ṗ ⊗ p̄ (6.197)
6.7. EULER PARAMETERS 235

6.7.5 Rotation by the quaternion product

Let R := Re (η, ²) be the rotation matrix corresponding to the Euler parameters η and
². Let v ∈R3 be an arbitrary vector. We are already familiar with the notion that Rv
is either the coordinate vector of the vector v in some other frame, or it is a rotation of
the vector v.
The transformation Rv can be achieved with the Euler parameters and the quaternion
product according to
µ ¶ µ ¶ µ ¶ µ ¶
0 η 0 η
= ⊗ ⊗ (6.198)
Rv ² v −²

This is shown by direct computation of the quaternion products:

µ ¶ µ ¶ µ ¶ µ ¶
η 0 η η²T v − η²T v − ²T ²× v
⊗ ⊗ =
² v −² η 2 v + 2η²× v + ²²T v + ²× ²× v
µ ¶
0
=
(I + 2η²× + 2²× ²× ) v
µ ¶
0
= (6.199)
Rv

where we have used (²× )2 = ²²T − ²T ²I and η 2 + ²T ² = 1.

We will now see how composite rotations can be expressed in terms of unit quater-
nions. Let
R1 = Re (η 1 , ²1 ) and R2 = Re (η 2 , ²2 ) (6.200)
and let
R = R1 R2 = Re (η, ²). (6.201)
be the composite rotation where p = (η ²T )T , p1 = (η 1 ²T1 )T and p2 = (η 2 ²T2 )T .
Let u be an arbitrary vector, and define v := R2 u and w := R1 v = Ru. Then
µ ¶ µ ¶
0 0
= p2 ⊗ ⊗ p2 (6.202)
v u
µ ¶ µ ¶ µ ¶
0 0 0
= p1 ⊗ ⊗ p1 = p1 ⊗ p2 ⊗ ⊗ p2 ⊗ p1 (6.203)
w v u

At the same time we have w = R1 R2 u = Ru which gives

µ ¶ µ ¶
0 0
=p⊗ ⊗p (6.204)
w u

Comparing these results we find that the Euler parameters η, ² corresponding to the
composite rotation R is given by
p = p1 ⊗ p2 (6.205)
which can also be written

η = η1 η 2 − ²T1 ²2 (6.206)
² = η 1 ²2 + η 2 ²1 + ²×
1 ²2 (6.207)
236 CHAPTER 6. RIGID BODY KINEMATICS

Example 103 We see that

RRT = I (6.208)
is consistent with
p ⊗ p̄ = pid (6.209)
Moreover,
RI = IR = R (6.210)
is seen to be consistent with

p ⊗ pid = pid ⊗ p = pid (6.211)

Example 104 Let F(·) and E(·) be the matrices corresponding to pre-multiplication and
post-multiplication, respectively, as defined in (6.174) and (6.175) Then
µ ¶ µ ¶
0 0
=p⊗ ⊗ p̄ (6.212)
Rv v
can be written µ ¶ µ ¶ µ ¶
0 0 0
= F(p) ⊗ p̄ = F(p)E(p̄) (6.213)
Rv v v
This leads to one more formula for the rotation matrix:
¡ ¢¡ ¢T
R = −² ηI + ²× −² ηI − ²× (6.214)

6.7.6 Euler parameters from the rotation matrix

The problem to be solved in this section is how to find the Euler parameters η, ² when
the rotation matrix R = {rij } is given. This is done using a method due to Shepperd
(Shepperd 1978).
The rotation matrix is given in terms of the Euler parameters by (6.165):
 2 
η + 21 − 22 − 23 2( 1 2 − η 3 ) 2( 1 3 + η 2 )
R = Re (η, ²) =  2( 1 2 + η 3 ) η 2 − 21 + 22 − 23 2( 2 3 − η 1 )  (6.215)
2( 1 3 − η 2 ) 2( 2 3 + η 1 ) η − 21 − 22 + 23
2

In addition, η 2 + 2
1 + 2
2 + 2
3 = 1. The following notation is introduced to simplify the
algorithms:    
z0 η
 z1   1 
z=   
 z2  := 2  2  (6.216)
z3 3

T := r11 + r22 + r33 = TraceR (6.217)

and
r00 := T (6.218)
This gives the symmetric set of equations

z02 = 1 + 2r00 − T (6.219)

z12 = 1 + 2r11 − T (6.220)
z22 = 1 + 2r22 − T (6.221)
z32 = 1 + 2r33 − T (6.222)
6.7. EULER PARAMETERS 237

that appear from the diagonal elements of R, while the off-diagonal terms give the
equations

z0 z1 = r32 − r23 z2 z3 = r32 + r23 (6.223)

z0 z2 = r13 − r31 z3 z1 = r13 + r31 (6.224)
z0 z3 = r21 − r12 z1 z2 = r21 + r12 (6.225)

The algorithm is as follows:

1. Find the largest element in {r00 , r11 , r22 , r33 }. This element is denoted rii .
2. Compute p
|zi | = 1 + 2rii − T (6.226)

3. Determine the sign of zi from some criterion, like continuity of solution, or η > 0.
4. Find the remaining zj from the three equations out of (6.223—6.225) that have as
the left side zj zi for all j 6= i. For example, if z0 was found under step 2 and 3,
then the remaining zj are found from

z1 = (r32 − r23 )/z0 (6.227)

z2 = (r13 − r31 )/z0 (6.228)
z3 = (r21 − r12 )/z0 (6.229)

5. Compute η = z0 /2 and i = zi /2.

Note that this algorithm avoids division by zero as the division is done with the zi
that has the largest absolute value.

6.7.7 The Euler rotation vector

The Euler rotation vector
e = k sin θ ∈ R3 (6.230)
is defined from the angle-axis parameters (k,θ) . From (6.145) it is seen that the rotation
matrix Rk,θ and its transpose RTk,θ are given by

Rk,θ = e× + cos θI + kkT (1 − cos θ) (6.231)

RTk,θ = −e× + cos θI + kkT (1 − cos θ) (6.232)

which implies that

1¡ ¢
e× = Rk,θ − RTk,θ (6.233)
2
From this we see that if Rk,θ = {rij }, then the Euler rotation vector can be found from
 
r32 − r23
1
e =  r13 − r31  (6.234)
2
r21 − r12

We note that if Rk,θ = Rab , then

e = ea = eb (6.235)
as Rk,θ k = k.
238 CHAPTER 6. RIGID BODY KINEMATICS

Example 105 In robot control the desired orientation of the robot hand may be specified
to be ¡ ¢
Rd = nd sd ad ∈ SO(3) (6.236)
Suppose that the actual orientation of the robot hand is
¡ ¢
R = n s a ∈ SO(3) (6.237)

where n is the normal vector, s is the slide vector and a is the approach vector of the
hand (Spong and Vidyasagar 1989), (Sciavicco and Siciliano 2000). Then the deviation
of R from Rd is given by the rotation matrix Re = {r̃ij } which is defined by

e := RRTd
R ⇒ e d
R = RR (6.238)

The component form of this equation is

r̃ij = ni ndj + si sdj + ai adj (6.239)

³ ´
If an angle-axis parameters of Re are k, e θ̃ and e e θ̃ is the associated Euler rotation
e = ksin
vector, then (6.234) gives
 
r̃32 − r̃23
1
e
e = r̃13 − r̃31 
2
r̃21 − r̃12
     
n3 nd2 − n2 nd3 s3 sd2 − s2 sd3 a3 ad2 − a2 ad3
1 1 1
= n1 nd3 − n3 nd1  +  s1 sd3 − s3 sd1  +  a1 ad3 − a3 ad1 
2 2 2
n2 nd1 − n1 nd2 s2 sd1 − s1 sd2 a2 ad1 − a1 ad2

Using the definition of the vector cross product the Euler rotation vector e
e corresponding
to the deviation Re can be written
1 ×
e
e= (n n + s× ×
d s + ad a) (6.240)
2 d

6.7.8 Euler-Rodrigues parameters

The Euler-Rodrigues parameters are defined by (Hughes 1986)
θ
ρ = k tan (6.241)
2
This can be expressed in terms of the Euler parameters according to
²
ρ= (6.242)
η
It is evident that the Euler-Rodrigues parameters are undefined when η = 0 ⇔ θ =
π + 2kπ where k ∈ {. . . − 1, 0, 1 . . .}.
The derivations that follows use the relation

1 = η 2 + ²T ² = η 2 (1 + ρT ρ) (6.243)

which implies
1
η2 = (6.244)
1 + ρT ρ
6.8. ANGULAR VELOCITY 239

Then
2
R=I+ [ρ× + ρ× ρ× ] (6.245)
1 + ρT ρ
is found from (6.166). We note that there are no trigonometric terms in (6.245).
The Euler-Rodrigues parameters can be found from the rotation matrix using
 
r32 − r23
² e 1  r13 − r31 
ρ= = 2 = (6.246)
η 2η TraceR + 1 r −r
21 12

where (6.234) and (6.169) are used.

Next we will derive Cayley’s formula (Angeles 1988) from equation (6.245). To do
this we need some algebraic manipulations. First we observe that ρ× ρ× = ρρT − ρT ρI
implies that ¡ ¢
ρ× ρ× ρ× = − ρT ρ ρ× (6.247)
Using this result and (6.245) we find that
· ¸
¡ ¢ 2 ¡ × ¢ ¡ ¢
R I − ρ× = I + T
ρ + ρ × ×
ρ I − ρ× = I + ρ× (6.248)
1+ρ ρ
This leads to the following result:

The rotation matrix can be given by Cayley’s formula

¡ ¢¡ ¢−1
R = I + ρ× I − ρ× (6.249)

where ρ is the vector of Euler-Rodrigues parameters corresponding to R.

The Cayley transformation cay(u) ∈ SO(3) maps a three-dimensional vector u into

a rotation matrix according to
· ¸· ¸−1
1 1
cay(u) : = I + u× I − u× ∈ SO(3) (6.250)
2 2

This transformation is used in numerical integrators in attitude problems (Lewis and

Simo 1994). In particular it is well suited for the implementation of the implicit mid-
point rule for the integration of the rotation matrix. We note that

R = cay (2ρ) (6.251)

and that
cay(kθ) ≈ Rk,θ , θ small (6.252)

6.8 Angular velocity

6.8.1 Introduction
If the position vector r is given in an inertial frame, then the velocity vector v = ṙ
is known to be the rate of change of the position vector r. In the same way we would
like to have some physical entity that describes the rate of change of a rotation matrix
Rab . This is not quite as simple as for the case of position and velocity. However, the
240 CHAPTER 6. RIGID BODY KINEMATICS

rotation matrix can be described by three independent variables, and this indicates that
there might be some entity that represents the time derivative of the rotation matrix
using three parameters. We will in the following analyze this problem, and arrive at the
definition of the angular velocity vector ω, which represents the time derivative of the
rotation matrix.

6.8.2 Definition
The rotation matrix Rab is orthogonal and satisfies

Rab (Rab )T = I (6.253)

Time differentiation of the matrix product gives

d £ a a T¤
Rb (Rb ) = Ṙab (Rab )T + Rab (Ṙab )T = 0 (6.254)
dt

From this equation it is seen that the matrix Ṙab (Rab )T is skew symmetric. Now, any skew
symmetric 3 × 3 matrix can be seen as the skew symmetric form of a column vector. This
means that it is possible to define a vector so that its skew symmetric form is equal to
Ṙab (Rab )T . It is quite remarkable that this vector can be given a physical interpretation,
and that it is of fundamental importance in dynamics.

Let the vector ω ab be defined by requiring that its coordinate form ω aab in frame a satisfies

(ω aab )× = Ṙab (Rab )T (6.255)

The vector ω ab is said to be the angular velocity vector of frame b relative to frame a.

A kinematic differential equation for the rotation matrix Rab appears from the defini-
tion of the angular velocity by post-multiplication of (6.255) with Rab . Moreover, using
the coordinate transformation rule (ω aab )× = Rab (ω bab )× Rba for the skew symmetric form
of a vector an alternative formulation of the kinematic differential equation is found.

The kinematic differential equation of the rotation matrix is given by the two alternative
forms
Ṙab = (ω aab )× Rab (6.256)
Ṙab = Rab (ω bab )× (6.257)

6.8.3 Simple rotations

Using the rotation matrices Rx (φ), Ry (θ) and Rz (ψ) of the simple rotations about the
x, y and z axes, we define the angular velocities
h i×
ω x (φ̇) : = Ṙx (φ)RTx (φ) (6.258)
h i×
ω y (θ̇) : = Ṙy (θ)RTy (θ) (6.259)
h i×
ω z (ψ̇) : = Ṙz (ψ)RTz (ψ) (6.260)
6.8. ANGULAR VELOCITY 241

From the definitions it is clear that ω x (φ̇) is the angular velocity of a rotation by an
angular rate φ̇ about the x axis, ω y (θ̇) is the angular velocity of a rotation by an angular
rate θ̇ about the y axis, and ω z (ψ̇) is the angular velocity of a rotation by an angular
rate ψ̇ about the z axis. From (6.101) we find that
  
h i× 0 0 0 1 0 0
ω x (φ̇) = φ̇  0 − sin φ − cos φ   0 cos φ sin φ  (6.261)
0 cos φ − sin φ 0 − sin φ cos φ
 
0 0 0
=  0 0 −φ̇  (6.262)
0 φ̇ 0
h i× h i×
In the same way we may compute ω y (θ̇) and ω z (ψ̇) . This results in
     
φ̇ 0 0
ω x (φ̇) =  0  , ω y (θ̇) =  θ̇  , and ω z (ψ̇) =  0  (6.263)
0 0 ψ̇
Consider the angle-axis parameterization when k = ka is a constant vector and
Rab = Rk,θ = I + k× sin θ + k× k× (1 − cos θ) (6.264)
Then the angular velocity is
¡ ¢¡ ¢
(ω aab )× = θ̇ k× cos θ + k× k× sin θ I − k× sin θ + k× k× (1 − cos θ)
£
= θ̇ k× cos θ + k× k× sin θ − k× k× cos θ sin θ + k× sin2 θ
¡ ¢ ¤
−k× cos θ − cos2 θ + k× k× cos θ sin θ − k× k× sin θ
= θ̇k× (6.265)
This shows that:

For a simple rotation, the angular velocity vector ω ab is along the axis of rotation k, and
is given by
ω ab = θ̇k (6.266)

This gives an intuitively appealing interpretation of the angular velocity. If the axis
of rotation is not constant, then the expressions become somewhat more involved.

6.8.4 Composite rotations

Consider the composite rotation Rad = Rab Rbc Rcd . The time derivative of Rad is, according
to the product rule,
Ṙad = Ṙab Rbc Rcd + Rab Ṙbc Rcd + Rab Rbc Ṙcd (6.267)
a T c T
¡ b ¢T a T
and the transpose is (Rd ) = (Rd ) Rc (Rb ) . Then the angular velocity of the
composite rotation is
³ ´
T T ¡ ¢T T
(ω aad )× = Ṙad (Rad ) = Ṙab Rbc Rcd + Rab Ṙbc Rcd + Rab Rbc Ṙcd (Rcd ) Rbc (Rab )
T ¡ ¢T T T T
= Ṙab (Rab ) + Rab Ṙbc Rbc (Rab ) + Rac Ṙcd (Rcd ) (Rac )
= (ω aab )× + Rab (ω bbc )× (Rab )T + Rac (ω ccd )× (Rac )T
= (ω aab )× + (ω abc )× + (ω acd )× (6.268)
242 CHAPTER 6. RIGID BODY KINEMATICS

This implies that the angular velocities ω ab , ω bc and ω cd can be added vectorially.

The angular velocity of the composite rotation matrix Rad = Rab Rbc Rcd is the sum of the
angular velocities according to

ω ad = ω ab + ω bc + ω cd (6.269)

Example 106 In a gimbal system for inertial navigation the rotation matrix from the
vehicle frame b to the instrumented platform frame p will be

Rbp = Rz (ψ)Ry (θ)Rx (φ) (6.270)

which corresponds to the angular velocity vector

ω bbp = ω z (ψ̇) + Rz (ψ)ω y (θ̇) + Rz (ψ)Ry (θ)ω x (φ̇) (6.271)

6.8.5 Differentiation of coordinate vectors

A coordinate vector is differentiated with respect to time by differentiating the compo-
nents of the vector with respect to time:
 a   a 
u u̇
a d a d  1a   1a 
u̇ := (u ) = u2 = u̇2 (6.272)
dt dt ua u̇a
3 3

The relation between the time derivative in frame a and the time derivative in frame
b is found by differentiating the equation

ua = Rab ub (6.273)

which gives
u̇a = Rab u̇b + Ṙab ub (6.274)
Insertion of Ṙab = Rab (ω bab )× gives the relation

£ ¤
u̇a = Rab u̇b + (ω bab )× ub (6.275)

6.8.6 Differentiation of vectors

Differentiation of a vector u must be done with reference to some reference frame. The
time derivative of the vector u = ua1 a1 + ua2 a2 + ua3 a3 referenced to frame a is defined by
a
d
u := u̇a1 a1 + u̇a2 a2 + u̇a3 a3 (6.276)
dt
where the leading superscript a on the time differentiation operator denotes that the
differentiation is taken with reference to frame a. The time derivative referenced to
frame b is
b
d
u = u̇b1 b1 + u̇b2 b2 + u̇b3 b3 (6.277)
dt
6.8. ANGULAR VELOCITY 243

where it is assumed that u = ub1 b1 + ub2 b2 + ub3 b3 . The corresponding column vector
representation is  a   b 
u̇1 u̇1
u̇a =  u̇a2  , u̇b =  u̇b2  (6.278)
u̇a3 u̇b3
From (6.275) we find that

a b
d d
u = u + ω ab × u (6.279)
dt dt
where
a
d
u = u̇a1 a1 + u̇a2 a2 + u̇a3 a3 (6.280)
dt
b
d
u = u̇b1 b1 + u̇b2 b2 + u̇b3 b3 (6.281)
dt

Example 107 In the same way partial differentiation with respect to some variable q in
frame a and b is defined by
a
∂u ∂ua1 ∂ua ∂ua
: = a1 + 2 a2 + 3 a3 (6.282)
∂q ∂q ∂q ∂q
b b b
∂u ∂u1 ∂u ∂ub
: = b1 + 2 b2 + 3 b3 (6.283)
∂q ∂q ∂q ∂q
Example 108 An alternative definition of the angular velocity vector is used in (Kane
and Levinson 1985):
à ! à ! à !
a a a
db2 db3 db1
ω ab = b1 · b3 + b2 · b1 + b3 · b2 (6.284)
dt dt dt

Here b1 , b2 , b3 are the orthogonal unit vectors of the frame b. We will now show that this
is in agreement by our definition (6.255). From (6.94) we have
¡ ¢
Rab = ba1 ba2 ba3 (6.285)
and from the definition of ω bab in (6.255) we get
 
0 baT a
1 ḃ2 baT a
1 ḃ3
b × T
(ω ab ) = R Ṙ =  aT
b2 ḃ1 a
0 baT a 
2 ḃ3 (6.286)
baT
3 ḃa
1 baT a
3 ḃ2 0
Before proceeding we show that this matrix is skew symmetric. We note that ba1 , ba2
d
and ba3 are orthogonal. Then baT a aT a aT a aT a
1 b2 = 0 and dt (b1 b2 ) = ḃ1 b2 + b1 ḃ2 = 0, which
implies that b1 ḃ2 = −b2 ḃ1 . In the same way it is found that b2 ḃ3 = −baT
aT a aT a aT a a
3 ḃ2 and
aT a aT a
b3 ḃ1 = −b1 ḃ3 , and the right side in (6.286) is seen to be skew symmetric. We write
ω bab in its vector form, and express the scalar products in terms of coordinate-free vectors
to get
 aT a   a

b3 ḃ2 b3 · ddtb2
ω bab =  baT a = a 
1 ḃ3  b1 · ddtb3  (6.287)
aT a a
b2 ḃ1 db1
b2 · dt
244 CHAPTER 6. RIGID BODY KINEMATICS

We see that this is indeed the coordinate form of the definition (6.284) of (Kane and
Levinson 1985).

6.9 Kinematic differential equations

6.9.1 Introduction
A model describing the rotation of a rigid body can be separated into the equation of mo-
tion, which is a differential equation for the angular velocity, and a kinematic differential
equation which gives the time derivative of some parameterization of the rotation matrix
as a function of the angular velocity. From a modeling perspective it is interesting to
note that kinematic differential equations are exact models with no uncertainty and no
approximations involved. In the following we will derive kinematic differential equations
for the different parametrizations of rotation that have been presented in the previous
sections.

6.9.2 Attitude deviation

We consider the problem where the rotation of a rigid body is to be controlled. This will
be the case in the attitude control of a satellite, or in the control of a robotic hand. Let
the frame a define a reference orientation, let the frame b be a frame fixed in the body.
Then the rotation matrix R := Rab will describe the orientation of the body. Suppose
that it is specified that the desired rotation of the body is given by a rotation matrix
Rd . The question is then how to represent the control deviation between the actual
value R and the desired value Rd . In a typical control setting we control some output
vector y to its desired value yd , and the control deviation ye = y − yd is simply obtained
by subtraction. In the case of rotation matrices it does not make sense to subtract Rd
from R as the result would not be a rotation matrix. Instead the deviation between the
two rotation matrices is described by the rotation matrix R e a ∈ SO(3) defined by

e a := RRTd
R ⇒ e a Rd
R=R (6.288)

We see that the rotation matrix R is described as the composite rotation defined by the
e a and Rd . To make this clear we introduce the intermediate frame c
rotation matrices R
so that
Re a = Rac , Rd = Rcb (6.289)
and
d e
Ra = Ṙac = (ω aac )× Rac (6.290)
dt
Ṙd = Ṙcb = Rcb (ω bcb )× (6.291)

Then, if we define the angular velocity vectors

e a := ω aac , ω bd := ω bcb
ω a = ω aab , ω (6.292)

e a and Rd are given by

we find that the kinematic differential equations for R, R

d e e a , Ṙd = Rd (ω b )×
Ṙ = (ω a )× R, ω a )× R
Ra = (e d (6.293)
dt
6.9. KINEMATIC DIFFERENTIAL EQUATIONS 245

It follows from
ω aab = ω aac + ω acb (6.294)
a
that ωe = ω a − ω ad , and we may sum up that the kinematic differential equations for the
attitude deviation is

e a := RRTd
R (6.295)

ea
ω = ω a − ω ad (6.296)
d e ea
Ra ω a )× R
= (e (6.297)
dt

We define an alternative representation of the deviation between the two rotations

e b ∈ SO(3) defined by
using the rotation matrix R
e b := RTd R
R ⇒ eb
R = Rd R (6.298)

The desired angular velocity ω ad is in this case defined by

Ṙd = (ω ad )× Rd (6.299)

Then, by introducing an intermediate frame as in the case above, we find that the kine-
matic differential equations referred to the b frame is

e b := RTd R
R (6.300)

eb
ω = ω b − ω bd (6.301)
d e e b (e b ×
Rb = R ω ) (6.302)
dt

6.9.3 Homogeneous transformation matrices

The time derivative of the homogeneous transformation matrix
µ a ¶
Rb raab
Tab = ∈ SE(3) (6.303)
0T 1
is found to be
µ ¶ µ a ¶µ ¶
Rab (ω bab )× ṙaab Rb raab (ω bab )× b
vab
Ṫab = =
0T 0 0T 1 0T 0
µ ¶
(ω bab )× vab b
= Tab T (6.304)
0 0
We see that the time derivative of a homogeneous transformation matrix has certain
similarities with the time derivative of a rotation matrix as expressed in (6.257). This
similarity becomes more evident if we introduce the vector
µ b ¶
vab
w= (6.305)
ω bab
246 CHAPTER 6. RIGID BODY KINEMATICS

which is the twist vector in the b frame. The twist vector w is a six-dimensional vector
containing the velocity and the angular velocity. In analogy with the angular velocity
¡ ¢×
ω bab and its matrix form ω bab , the twist vector w has a matrix form in the set se(3)
which is µ ¶
(ω bab )× vab
b
ŵ = ∈ se(3) (6.306)
0T 0

The time derivative of the homogeneous transformation matrix is given by

Ṫab = Tab ŵ (6.307)

This topic is treated in great detail in (Murray, Li and Sastry 1994).

Example 109 The transformation rule for a twist vector is not as straightforward as for
the angular velocity vector. This is seen in the time derivative of Tab when it expressed
in the a frame:
µ ¶
a (ω aab )× Rab vab
a
Ṫb =
0T 0
µ a ×
¶µ a ¶
(ω ab ) vab − (ω aab )× raab
a
Rb raab
= (6.308)
0T 0 0T 1
a
The physical interpretation of the velocity term vab −(ω aab )× raab is not as obvious as when
the coordinates of the b frame is used. A geometric interpretation is given in (Murray
et al. 1994).

6.9.4 Euler angles

When Euler angles are used the rotation matrix Rad from frame a to frame d is a composite
rotation involving three simple rotations. In the roll-pitch-yaw case the simple rotations
are
Rab = Rz (ψ), Rbc = Ry (θ) and Rcd = Rx (φ) (6.309)
We see that the angular velocities associated with the simple rotations are
     
0 0 φ̇
ω aab =  0  , ω bbc =  θ̇  and ω ccd =  0  (6.310)
ψ̇ 0 0

From (6.269) we have that the angular velocity of d relative to a is the sum of the angular
velocities resulting from each of the three simple rotations due to ψ, θ and φ:

ω ad = ω ab + ω bc + ω cd (6.311)

In the a frame this gives:

     
0 0 φ̇
ω aad =  0  + Rz,ψ  θ̇  + Rz,ψ Ry,θ  0 
ψ̇ 0 0
 
− sin ψ θ̇ + cos ψ cos θφ̇
=  cos ψ θ̇ + sin ψ cos θφ̇  (6.312)
ψ̇ − sin θφ̇
6.9. KINEMATIC DIFFERENTIAL EQUATIONS 247

In the d frame we find


    
0 0 φ̇
ω dad = Rx,−φ Ry,−θ  0  + Rx,−φ  θ̇  +  0  (6.313)
ψ̇ 0 0
 
− sin θψ̇ + φ̇
=  sin φ cos θψ̇ + cos φθ̇  (6.314)
cos φ cos θψ̇ − sin φθ̇
T
Define the vector φ = (φ, θ, ψ) . We can then write
 
cos ψ cos θ − sin ψ 0
ω aad = Ea (φ)φ̇ =  sin ψ cos θ cos ψ 0  φ̇ (6.315)
− sin θ 0 1
and  
1 0 − sin θ
ω dad = Ed (φ)φ̇ =  0 cos φ sin φ cos θ  φ̇. (6.316)
0 − sin φ cos φ cos θ
We note that det[Ea (φ)] = det[Ed (φ)] = cos θ which implies that the matrices are
singular for cos θ = 0.
We can solve for φ̇, and find that
 
cos ψ sin ψ 0
1  − sin ψ cos θ cos ψ cos θ
φ̇ = Ea (φ)−1 ω aad = 0  ω aad (6.317)
cos θ
cos ψ sin θ sin ψ sin θ cos θ

and
 
cos θ sin φ sin θ cos φ sin θ
1 
φ̇ = Ed (φ)−1 ω dad = 0 cos φ cos θ − sin φ cos θ  ω dad (6.318)
cos θ
0 sin φ cos φ

Let ai be the orthogonal unit vectors of the a frame, bi be the unit vectors of the b
frame, and let ci be the orthogonal unit vectors of the c frame. Then the roll-pitch-yaw
description gives the angular velocity ω ad as a sum of an angular velocity ω ab along a3 ,
an angular velocity ω bc along b2 , and an angular velocity ω cd along c1 . The physical
interpretation of the singularity of Ea (φ) and Ed (φ) at cos θ = 0 is due to the fact that
when cos θ = 0, then the rotation vectors a3 and c1 align so that both ω ab and ω cd are
along the a3 vector while ω bc is along the b2 vector. This means that it is not possible
to describe an angular velocity along a3 × b2 when cos θ = 0. This is the Euler-angle
singularity, which is a singularity due to the mathematical representation of the rotation
matrix.

6.9.5 Euler parameters

In this section we will derive the kinematic differential equations for the Euler parameters.
These differential equations give the time derivatives of the Euler parameters as functions
of the angular velocity. We let R := Rab and ω := ω ab so that Ṙ = (ω a )× R. Moreover
we let R = Re (η, ²) and p = (η ²T )T . We then have

Ṙ = (ω a )× R = R(ω b )× (6.319)
248 CHAPTER 6. RIGID BODY KINEMATICS

The derivation is based on the coordinate transformation rule using the quaternion
product. For an arbitrary vector u ∈ R3 we have
µ ¶ µ ¶
0 0
=p⊗ ⊗ p̄ (6.320)
Ru u
We take the time derivative of both sides and get
µ ¶ µ ¶ µ ¶ µ ¶ µ ¶
0 0 0 0 0
+ = ṗ ⊗ ⊗ p̄ + p ⊗ ⊗ p̄ + p ⊗ ⊗ p̄˙ (6.321)
Ṙu Ru̇ u u̇ u

Then, because the transformation rule in (6.320) is valid for any vector it is also valid
for u̇. This implies that
µ ¶ µ ¶ µ ¶
0 0 0
= ṗ ⊗ ⊗ p̄ + p ⊗ ⊗ p̄˙
Ṙu u u
µ ¶ µ ¶
0 0
= ṗ ⊗ p̄ ⊗ p ⊗ ⊗ p̄ − p ⊗ ⊗ p̄ ⊗ ṗ ⊗ p̄
u u
µ ¶ µ ¶
0 0
= (ṗ ⊗ p̄) ⊗ − ⊗ (ṗ ⊗ p̄)
Ru Ru
µ ¶
0
= 2 × (6.322)
(η ²̇ − η̇² + ²× ²̇) Ru

where we have used (6.197), (6.191), (6.193), (6.194) and (6.173). From Ṙ = (ω a )× R
we have µ ¶ µ ¶
0 0
= (6.323)
Ṙu (ω a )× Ru
Comparing this with (6.322) we find that the angular velocity ω a is given by
ω a = 2[η ²̇ − η̇² + ²× ²̇] (6.324)
From (6.194) it is seen that this can be written in quaternion form, and this leads to the
result

The angular velocity is given in frames a and b by

µ ¶ µ ¶
0 0
= 2ṗ ⊗ p̄, = 2p̄ ⊗ ṗ (6.325)
ωa ωb

and the kinematic differential equation for the quaternion vector is

µ ¶ µ ¶
1 0 1 0
ṗ = ⊗ p, ṗ = p ⊗ (6.326)
2 ωa 2 ωb

Here the transformation rule (6.320) has been used, and the kinematic differential
equations appear by postmultiplication with p for the expression in the a frame, and by
premultiplication with p for the expression in the b frame.
The component form of these last four equations gives the result
ωb = 2[η ²̇ − η̇² − ²× ²̇] (6.327)
ωa = 2[η ²̇ − η̇² + ²× ²̇] (6.328)
6.9. KINEMATIC DIFFERENTIAL EQUATIONS 249

1
η̇ = − ²T ω b (6.329)
2
1
²̇ = [ηI + ²× ]ω b (6.330)
2
and
1
η̇ = − ²T ω a (6.331)
2
1
²̇ = [ηI − ²× ]ω a (6.332)
2
Example 110 From (6.174), (6.175) and (6.329—6.332) it is seen that the kinematic
differential equations can be written in vector form as
µ Ã ¡ ¢T !
T ¶
1 0 − (ω a ) 1 0 − ωb
ṗ = × p= ¡ ¢× p (6.333)
2 ω a (ω a ) 2 ωb − ωb
or µ ¶µ ¶ µ ¶µ ¶
1 η −²T 0 1 η −²T 0
ṗ = = (6.334)
2 ² ηI + ²× ωa 2 ² ηI − ²× ωa

6.9.6 Normalization for numerical integration

From (6.334) it is seen that
µ ¶µ ¶
d ¡ T ¢ η −²T 0
p p = pT =0 (6.335)
dt ² ηI + ²× ωa

This shows that if p is initialized as a unit vector, then it will remain a unit vector, as
should be expected. Numerical integration of the quaternion vector p from the kinematic
differential equation will introduce numerical errors that will cause the length of p to
deviate from unity. To compensate for such errors a normalization term is added to the
kinematic differential equation. This can be done with the following modification of the
kinematic differential equation, which should be used in numerical integration:
µ ¶µ ¶
1 η −²T 0 λ
ṗ = × a + (1 − pT p)p (6.336)
2 ² ηI + ² ω 2
Here λ is a positive gain. Then
d ¡ T ¢ λ
p p = (1 − pT p)pT p (6.337)
dt 2
We see that this will give the desired result as pT p will increase whenever pT p <
1, and pT p will decrease whenever pT p > 1. When pT p = 1 the usual kinematic
differential equations are recovered. Linearization about pT p = 1 gives ė = −λe where
e = 1−pT p. This means that the normalization converges with a time constant T = λ−1 .
A Simulink toolbox has implemented this algorithm with λ = 100, which means that the
normalization converges with a time constant of 0.01 s.
Another alternative is to normalize directly after each time step using the normaliza-
tion assignment
p
p:= p (6.338)
pT p
250 CHAPTER 6. RIGID BODY KINEMATICS

6.9.7 Euler rotation

As sin θ = 2 sin θ2 cos θ2 , it is seen that e = k sin θ can be expressed by the Euler parame-
ters according to
e = 2η². (6.339)
The kinematic differential equation is then found from
ė = 2(η̇² + η ²̇). (6.340)
This gives
ė = [η 2 I − ²²T − η²× ]ω a (6.341)
= [η 2 I − ²²T + η²× ]ω b (6.342)
where (6.331) and (6.332) are used.
Alternative expressions are found from
R = (2η 2 − 1)I + 2²²T + 2η²× (6.343)
and
RT = (2η 2 − 1)I + 2²²T − 2η²× (6.344)
which leads to
1
ė = [Trace(R)I − R]ω a (6.345)
2
1
ė = [Trace(RT )I − RT ]ω b (6.346)
2
where we have used that 4η 2 − 1 = Trace(R).
Note that for θ = 0 we have
ė|θ=0 = ω a = ω b (6.347)

6.9.8 Euler-Rodrigues parameters

The kinematic differential equation is derived from
d ² η ²̇ − η̇²
ρ̇ = = (6.348)
dt η η2
which gives
1
ρ̇ =[I + ρ× + ρρT ] ω b . (6.349)
2
An equation for the angular velocity is found from
1 × η ²̇ − η̇² 1
ρ× ρ̇ = ² 2
= 2 ²× ²̇ (6.350)
η η η
as ²× ² = 0. From (6.327) it is seen that
µ ¶
η ²̇ − η̇² 1 ×
ω b = 2η 2 − ² ²̇ (6.351)
η2 η2
Insertion of (6.244), (6.348) and (6.350) gives
2
ωb = [I − ρ× ]ρ̇ (6.352)
1 + ρT ρ
6.9. KINEMATIC DIFFERENTIAL EQUATIONS 251

Example 111 Equation (6.245) can be written

1
R= [2I + 2ρ× + 2ρρT − (1 + ρT ρ)I] (6.353)
1 + ρT ρ

which implies that

1 + ρT ρ
I + ρ× + ρρT = (R + I) (6.354)
2
Then (6.349) can be written

1 + ρT ρ
ρ̇ = (R + I) ω b (6.355)
4
We recall from (6.169) that
TraceR + 1 = 4η 2 (6.356)
and, using (6.244), we arrive at

1
ρ̇ = (R + I) ω b (6.357)
TraceR + 1

6.9.9 Passivity of kinematic differential equations

In translational dynamics the integration from velocity v = ẋ to position x is a passive
dynamic system. In fact, the function Vx = 12 xT x ≥ 0 has time derivative

∂Vx
V̇x = ẋ = xT v (6.358)
∂x
so that the system with input v and output x is clearly passive. It is interesting to
investigate if similar results can be established for rotational dynamics.
The starting point for such an investigation (Egeland and Godhavn 1994) is the
differential equation
1
η̇ = − ²T ω (6.359)
2
¯ ¯
where |η| = ¯cos θ ¯ ≤ 1. Define
2

V = 2 (1 − η) ≥ 0 (6.360)

The time derivative for solutions of the kinematic differential equations of the Euler
parameters is
V̇ = −2η̇ = ²T ω (6.361)
It follows that the kinematic system with input ω and output ² is passive.
At this stage it is not very difficult to extend this result to other kinematic represen-
tations based on the Euler parameters. First we note that if we multiply the equation
for η̇ by η, we get
1 1
η η̇ = − η²T ω = − eT ω (6.362)
2 4
where e = 2η² is the Euler rotation vector. We are then lead to the function
¡ ¢
Ve = 2 1 − η 2 ≥ 0 (6.363)
252 CHAPTER 6. RIGID BODY KINEMATICS

which has time derivative

V̇e = −4η η̇ = eT ω (6.364)
and we have shown that the kinematic system with input ω and output e is passive.
Finally we note that
η̇ 1 ²T 1
=− ω = − ρT ω (6.365)
η 2η 2
Define
Vρ = −2 ln |η| ≥ 0, η 6= 0 (6.366)
which is defined for all η except for η = 0 where also ρ is undefined. Then
η̇
V̇ρ = −2 = ρT ω (6.367)
η
and it is seen that the kinematic system with input ω and output ρ is passive.

Example 112 It is interesting to note that

²T ² + (1 − η)2 = ²T ² + η 2 − 2η + 1 = 2 (1 − η) = V (6.368)

and that ¡ ¢
2²T ² = 2 1 − η 2 = Ve (6.369)

Example 113 Consider a system with equation of motion

Mω̇ + ω × Mω = τ (6.370)

where M is a constant, symmetric and positive definite matrix, and where the input is
selected to be
τ = −Kd ω − kp ² (6.371)
where Kd is a constant, symmetric and positive definite matrix, and kp is a positive
constant. The energy function
1 T
V = ω Mω + 2kp (1 − η) ≥ 0 (6.372)
2
has time derivative
¡ ¢
V̇ = ω T −ω × Mω − Kd ω − kp + kp ²T ω
= −ω T Kd ω (6.373)

along the solutions of the system. This means that the energy of the system decreases
whenever ω 6= 0. For further details see (Wen and Kreutz-Delgado 1991), where a cross-
term was added to the energy function, and (Egeland and Godhavn 1994).

6.9.10 Angle-axis representation

The kinematic differential equations for the angle θ and the unit vector k in the angle
axis representation of the rotation matrix are derived in this section. The derivation is
based on differentiation of the Euler parameters. To find the equation for θ̇ we observe
that µ ¶ µ ¶
1 θ 1 1 θ
− sin θ̇ = η̇ = − ²T ω = − sin kT ω
2 2 2 2 2
6.10. THE SERRET-FRENET FRAME 253

Whenever sin (θ/2) 6= 0, this implies

θ̇ = kT ω (6.374)

To find the equation for k̇ we note that

µ ¶ µ ¶
d θ θ 1 θ θ
²̇ = sin k+ sin k̇ = cos θ̇k+ sin k̇
dt 2 2 2 2 2

Combining this with the kinematic differential equations of the Euler parameters, we get
µ ¶
1£ ×
¤ 1 θ θ
ηI + ² ω = cos θ̇k+ sin k̇
2 2 2 2

which gives

θ ³ ´
2 sin k̇ = η I − kkT ω + ²× ω
2
θ³ ´ θ
= cos I − kkT ω+ sin k× ω
2 2

Then the kinematic differential equation for k is found using k× k× = kkT −I. Whenever
sin (θ/2) 6= 0 the result is
· ¸
1 × × × θ
k̇ = k −k k cot ω (6.375)
2 2
The equations (6.374) and (6.375) have a singularity at θ = 0, which is in agreement
with the fact that k is undefined for a zero rotation θ = 0.

6.10 The Serret-Frenet frame

6.10.1 Kinematics

f2 C

f1

f3

Figure 6.8: The Serret-Frenet frame for a curve C.

In aerospace, automotive steering and ship control the desired trajectory of the system
may be given as a curve in a fixed frame i. The control deviations from the desired curve
to the actual configuration of the system can then be calculated in the Serret-Frenet
frame f . This frame has axes along the tangent, the normal and the binormal of the
curve as shown in Figure 6.8. We will develop the equations for this frame in this section.
254 CHAPTER 6. RIGID BODY KINEMATICS

We let the curve C be given by r (s) where s is the length along the curve. We define
the unit tangent
i
dr
f1 = (6.376)
ds
The principal unit normal of the curve is defined by the unit vector

1 i df1
f2 = (6.377)
κ ds
where ¯ ¯
¯ i df ¯ ¯¯ i d2 r ¯¯
¯ 1¯ ¯ ¯
κ=¯ ¯= (6.378)
¯ ds ¯ ¯ ds2 ¯
is the curvature of the curve. Finally the unit binormal vector is defined by the unit
vector
f3 = f1 × f2 (6.379)
so that f1 , f2 , f3 forms a set of orthogonal unit vectors. The plane defined by f1 and f2
is called the osculating plane, the plane defined by f2 and f3 called the normal plane,
while the plane defined by f3 and f1 is called the rectifying plane.
From f3 · f3 = 1 it follows that

d ³ ´ i
df3
0= f3 · f3 = 2 · f3 (6.380)
ds ds
while (6.377) and f3 · f2 = 0 implies that
i
df1
f3 · = f3 · κf2 = 0 (6.381)
ds
Then, from f3 · f1 = 0 it follows that

d ³ ´ i df
3
i
df3
0= f3 · f1 = · f1 + f3 · κf2 = · f1 (6.382)
ds ds ds
From (6.380) and (6.382) it is seen that i df3 /ds is along f2 . This makes it possible to
write
i
df3
= τ f2 (6.383)
ds
where
i
df3
τ = f2 · (6.384)
ds
is the torsion of the curve. From f2 = f3 × f1 we find that
i i i
df2 df3 df1
= × f1 + f3 × = τ f2 × f1 + f3 × κf2 = −τ f3 − κf1 (6.385)
ds ds ds
The angular velocity ω if of the Serret-Frenet frame f relative to the frame i is seen from
(6.284) to be
à ! à ! à !
i
df2 i
df3 i
df1 ³ ´
ω if = f3 · f1 + f1 · f2 + f2 · f3 = ṡ τ f1 + κf3 (6.386)
dt dt dt

To sum up the results of this section:

6.11. NAVIGATIONAL KINEMATICS 255

The unit vectors f1 , f2 , f3 of the Serret-Frenet frame satisfies the kinematic differential
equations
i i i
df1 df2 df3
= κf2 , = −κf1 − τ f3 , = −τ f2 (6.387)
ds ds ds
where κ is the curvature and τ is the torsion of the curve. The angular velocity ω if of
the Serret-Frenet frame f relative to the frame i, and the velocity vf of the origin of
frame f are given by ³ ´
ω if = ṡ τ f1 + κf3 , vf = ṡf1 (6.388)

6.10.2 Control deviation

Suppose that a curve is to be followed by some body b, and that the origin of the Serret-
Frenet frame f is placed so that the origin of the body-fixed frame b has position

rbf = y f2 + z f3 (6.389)

or in other words, the frame f is placed so that the origin of frame b is in the normal
plane of f . Then the velocity of the origin of b is
i
d
vb = vf + rbf
dt
f
d
= vf + rbf + ω if × rbf (6.390)
dt
which in coordinate form in frame f gives
      
ṡ 0 0 −ṡκ 0 0
vbf =  0  +  ẏ  +  ṡκ 0 −ṡτ   y  (6.391)
0 ż 0 ṡτ 0 z

This gives the following relation between the time derivatives ṡ, ẏ, ż of the parameters
s, y, z and the body velocity vbb
 
ṡ − ṡκy
 ẏ − ṡτ z  = Rf vbb (6.392)
b
ż + ṡτ y

while the relative angular velocity between the frames b and f is given by
 
ṡτ
ω fbf = ω fif − ω fib =  0  − Rfb ω bib (6.393)
ṡκ

6.11 Navigational kinematics

6.11.1 Introduction
Navigation systems rely on GPS navigation using satellite signals, and on the use of in-
ertial navigation (Titterton and Weston 1997), (Farell and Barth 1999). In inertial nav-
igation the position and rotation of a ship, aeroplane or some other vehicle is computed
256 CHAPTER 6. RIGID BODY KINEMATICS

using measurements from gyroscopes and accelerometers. The accelerometers measure

the acceleration, and the gyroscopes measure the angular velocity ω of the gyroscopes
so the the rotation matrix R of the accelerometers can be computed by integration of
the kinematic differential equation Ṙ = Rω × provided that the initial rotation of the
accelerometers is known. In this section the kinematics of inertial navigation is presented.

6.11.2 Coordinate frames

Stella Polaris

 ie

 ie t
ye
xi yi

xe
Aries Point

Figure 6.9: The earth-centered star-fixed frame i and the earth-centered earth-fixed frame
e.

The frame i is star-fixed and earth-centered, and is considered to be an inertial frame

where Newton’s law is valid. The frame i has its origin in the center of the earth, and has
axes (xi , yi , zi ) that point towards certain fixed stars. In particular, the zi axis points
towards Stella Polaris, while the xi axis is pointing towards the Aries point in the Vernal
Equinox direction. The earth-fixed frame is denoted by e, and has its origin in the center
of the earth, and has axes (xe , ye , ze ) that rotate with the earth. The ze axis points
towards Stella Polaris through the North Pole. Frames i and e are shown in Figure 6.9.
The locally horizontal frame or geographic frame is denoted n. This frame has its origin
at some point with latitude L and longitude l at the surface of the earth, and has axes
(N, E, D) pointing north, east and down as shown in Figure 6.10. A spherical earth with
radius re is assumed. In high precision system the earth is described by an ellipsoid, but
to avoid a too complicated presentation this will not be done here.
The rotation of the earth frame e relative to the inertial frame i is given by the
rotation matrix
Rie = R(z, ω ie t)
6.11. NAVIGATIONAL KINEMATICS 257

ze

N
E
D
L
ye

l
xe

Figure 6.10: The local horizontal frame n with axes N, E, D pointing north, east and
down.

where the angular velocity is

 
0
ω iie = ω eie = 0 
ω ie

The scalar ω ie corresponds to 360◦ in 24 hours, which gives ω ie = 7. 27 × 10−5 rad/s.

The rotation from the inertial frame i to the geographic frame n is given by
h ³ π ´i
Rin = R(z, λ)R y, − L +
 2 
− sin L cos λ − sin λ − cos L cos λ
=  − sin L sin λ cos λ − cos L sin λ 
cos L 0 − sin L

where λ = l + ω ie t. The angular velocity of frame n is

     
0 0 L̇ sin λ
ω iin =  0  + R(z, λ)  −L̇  =  −L̇ cos λ 
λ̇ 0 λ̇

or in the geographic frame

     
0 0 λ̇ cos L
ω nin =  −L̇  + R(y, 90◦ )R(y, L)  0  =  −L̇ 
0 λ̇ −λ̇ sin L
258 CHAPTER 6. RIGID BODY KINEMATICS

6.11.3 Acceleration
An accelerometer measures the specific force
i 2
d
f= r − g0
dt2
where g0 is the acceleration of gravity. The gravitational field g of the earth is the
combined effect of the acceleration of gravity and the centripetal acceleration due to the
rotation of the earth:
g = g0 − ω ie × (ω ie × r)
The velocity
e i
d d
v := r = r − ω ie × r (6.394)
dt dt
is defined as the time derivative of the position vector in the earth frame. Note that v
i
is not defined as dtd r. In navigation algorithms the time derivative of the velocity v in
the geographic frame is needed. This is the position vector r differentiated first in the e
frame and then in the n frame. This gives
n i
d d
v = v − ω in × v
dt dt µ ¶
i
d id
= r − ω ie × r − ω in × v
dt dt
i 2 i i
d d d
= 2
r − ω ie × r − ω ie × r − ω in × v
dt dt dt
i 2
d
= r − (ω ie + ω in ) × v − ω ie × (ω ie × r)
dt2
i 2
d
= r − (2ω ie + ω en ) × v − ω ie × (ω ie × r) (6.395)
dt2
where it is used that ω ie is constant in the i frame, and that ω in = ω ie +ω en . The specific
force f is therefore
n
d
f= v + (2ω ie + ω en ) × v − g
dt
n
In the n frame the velocity v is found to be given by L, l, h and their derivatives:
   
vN (re + h) L̇
vn =  vE  =  (re + h) l˙ cos L 
vD −ḣ
where h is the height above the surface of the earth. The vector of gravity is assumed to
be
gn = (0, 0, g)T
Then
 ³ ´ 
    l˙ + 2ω ie vE sin L − L̇vD  
fN v̇N  ³ ´  0
   
f n =  fE  =  v̇E  +  − l˙ + 2ω ie (vN sin L + vD cos L) − 0 (6.396)
fD v̇D  ³ ´  g
l˙ + 2ω ie vE cos L + L̇vN

As vN , vE vD are function of L, l, h and their derivatives this clearly shows that f n is a

function of L, l, h and their derivatives.
6.12. KINEMATICS OF A RIGID BODY 259

6.12 Kinematics of a rigid body

6.12.1 Configuration

zi
o
r
zb
p
ro yb
rp
xb

yi
xi

Figure 6.11: Rigid body b with the fixed frame b and the fixed points o and p.

The configuration of a rigid body defines the position of all points in the rigid body.
For a rigid body the configuration can be specified in terms of the position ro of one
fixed point in the rigid body, and the rotation matrix Rib from a reference frame i to
a body-fixed frame b. Then the position of any point p in the rigid body, which is not
necessarily fixed in the rigid body, is given by
rp = ro + r (6.397)
as shown in Figure 6.11.
Here r is the vector from o to p with coordinate vector rb in the b frame. This vector
is given in the i frame by
ri = Rib rb (6.398)

6.12.2 Velocity
The frame i is assumed to be an inertial frame which is also referred to as a Newtonian
frame. The velocities of o and p are given by
i i
d d
vo := ro , vp := rp (6.399)
dt dt
From (6.397) and the rule for differentiation in moving frames it is seen that the velocity
of p can be expressed as
b
d
vp = vo + r + ω ib × r (6.400)
dt

6.12.3 Acceleration
The acceleration vectors are defined by
i 2 i 2
d d
ap := 2
rp , ao := 2 ro (6.401)
dt dt
260 CHAPTER 6. RIGID BODY KINEMATICS

while the angular acceleration vector is defined by

i b
d d
αib := ω ib = ω ib (6.402)
dt dt
where the second equality is a consequence of
i b b
d d d
ω ib = ω ib + ω ib × ω ib = ω ib (6.403)
dt dt dt
In the formulation of the equations of motion for rigid bodies, we need the following
result:
i 2 i 2 i
µ ¶ i 2 µ ¶
d d d id d i
d bd
rp = ro + r = 2 ro + r + ω ib × r
dt2 dt2 dt dt dt dt dt
i 2 b
µ ¶ µb ¶
d d bd d
= ro + r + ω ib × r + ω ib × r + ω ib × r
dt2 dt dt dt
i 2 b 2 b
µ b
¶
d d d d
= 2
ro + 2 r + 2ω ib × r + ω ib × r + ω ib × (ω ib × r) (6.404)
dt dt dt dt
In terms of acceleration, angular acceleration and velocities this is written

b 2
d
ap = ao + r
|{z} |{z} dt 2
| {z }
Acceleration Acceleration
Second derivative
of p of o
of r in b
b
d
+ 2ω ib × r + αib × r + ω ib × (ω ib × r) (6.405)
| {z } | {z }
| {z dt }
Transversal Centripetal
Coriolis
acceleration acceleration
acceleration

An alternative formulation is obtained by inserting the expression

i b
d d
ao = vo = vo + ω ib × vo (6.406)
dt dt
which gives

b b 2 b
d d d
ap = vo + ω ib × vo + 2 r + 2ω ib × r + αib × r + ω ib × (ω ib × r) (6.407)
dt dt dt

Note the difference between the term ω ib × vo which is related to the velocity of o,
b
and the Coriolis acceleration 2ω ib × dtd r which is related to the motion of p in the b frame
relative to o.
If the point p is fixed in the body b, then the vector r is constant in frame b so that
b i
d d
r = 0 ⇒ r = ω ib × r, r fixed in b (6.408)
dt dt
6.13. THE CENTER OF MASS 261

For vp this gives

vp = vo + ω ib × r, r fixed in b (6.409)
The acceleration is found to be
ap = ao + αib × r + ω ib × (ω ib × r), r fixed in b (6.410)
We see that in this case there is no Coriolis acceleration. Using (6.406) the acceleration
can be written
b
d
ap = vo + ω ib × vo + αib × r + ω ib × (ω ib × r), r fixed in b (6.411)
dt

6.13 The center of mass

6.13.1 System of particles
Consider a system of N particles each of mass mk and with position rk relative to the
origin of the inertial frame i. The center of mass is the point with position rc defined by
N
X
mrc = mk rk (6.412)
k=1
PN
where m = k=1 mk is the sum of the mass of the particles. The velocity vc and the
acceleration ac of the center of mass are defined by
N
i
drc X
mvc = m = mk vk (6.413)
dt
k=1
N
i 2
d rc X
mac = m = mk ak (6.414)
dt2
k=1

6.13.2 Rigid body

The position rc of the center of mass of a rigid body b is defined by
Z
mrc = rp dm (6.415)
b
R
where m = b dm is the mass of the rigid body, and rp is the position of a mass element
dm which is fixed in frame b. The position of the mass element relative to the center of
mass is given by r so that
rp = rc + r (6.416)
as shown in Figure 6.12. From the definition of the center of mass we see that
Z Z
rdm = rp dm − mrc = 0 (6.417)
b b

The velocity vc of the center of mass is given by

i i Z Z i Z
drc d drp
mvc = m = rp dm = dm = vp dm (6.418)
dt dt b b dt b

while the acceleration ac of the center of mass is found in the same way. We conclude
that
262 CHAPTER 6. RIGID BODY KINEMATICS

zi
dm
r zb
c

rp yb

rc
xb

yi
xi

Figure 6.12: Mass element dm relative to the center of mass c.

The motion of the mass center in the rigid body b satisfies the equations
Z Z Z
mrc = rp dm, mvc = vp dm, mac = ap dm (6.419)
b b b
Chapter 7

Newton-Euler equations of
motion

7.1 Introduction
The development of the equations of motions for rigid bodies and systems of rigid bodies
is the topic of this chapter. The equations of motion are differential equations for the
velocity and angular velocity. The derivations in this chapter are based on Newton’s
law and its extension to rotational dynamics, which is usually attributed to Euler. This
provides the motivation for the term Newton-Euler equation of motion. The derivations
rely on vector operations. The presentation starts with some results on forces and torques
on rigid bodies. Then the basic Newton-Euler equations of motion are presented and used
to derive the equations of motion for the ball-and-beam system, the Furuta pendulum
and the inverted pendulum. Then the principle of virtual work is presented, and its use is
demonstrated for multi-body systems. The use of recursive computations in manipulator
dynamics is also discussed.

7.2 Forces and torques

To derive the equations of motions for rigid bodies we need some results on resultant
forces and moments, which are presented in this section. The material is taken from
(Kane and Levinson 1985).

7.2.1 Resultant force

A force vector F will have a line of action, which means that the moment of F about a
point P is r × F where r is the position vector from P to some arbitrary point on the line
of action as shown in Figure 7.1. A vector with a line of action is called a bound vector .
(r)
Consider a set S of nF forces Fj . The resultant force FS of the set S is the vector

nF
X
(r)
FS = Fj (7.1)
j=1

263
264 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION


F

r
line of action

Figure 7.1: A force F acting on a rigid body. The line of action of the force is indicated
as a dashed line, and the distance r from a point P is shown.

while the moment about P of the set S of forces is

nF
X
NS/P = rP j × Fj (7.2)
j=1

where rP j is the position vector from P to an arbitrary point on the line of action of
(r)
Fj . Note that in this description the resultant FS is a sum of forces, and can not be
considered to be a force with a line of action. Note in particular that the resultant force
does not appear in the expression for the moment NS/P . The moment NS/Q about some
other point Q is found from

nF
X nF
X
NS/Q = rQj × Fj = (rP j + rQP ) × Fj
j=1 j=1
nF
X nF
X
= rP j × Fj + rQP × Fj (7.3)
j=1 j=1

The moment NS/Q of the set S about a point Q is the moment NS/P of the set S about
(r)
the point P plus the moment about Q that would have resulted if the resultant FS had
line of action through P :
(r)
NS/Q = NS/P + rQP × FS (7.4)

This result is straightforward to apply, however, the resultant force does not have a
(r)
line of action, so the procedure of pretending that FS has its line of action through
P is not completely satisfying. Therefore we will follow the approach of (Kane and
Levinson 1985) and introduce an equivalent representation with a bound vector and a
torque. To do this it is necessary to introduce the concept of a torque.
7.2. FORCES AND TORQUES 265

7.2.2 Torque
(r)
A couple is a set C of forces with zero resultant force, that is FC = 0. From (7.4) it
is seen that this implies that NC/P = NC/Q , which means that the moment of a couple
will be the same about any point, and it is therefore meaningful to define the moment of
the couple without reference to any point.

(r)
The torque TC is defined as the moment of the couple C. The resultant FC of a couple
is by definition zero. Therefore, the moment of the couple C is the same about any point,
which means that
TC := NC/P = NC/Q (7.5)
for arbitrary points Q and P .

Example 114 Consider a couple with two forces F1 and F2 that have zero resultant
force, which implies F2 = −F1 . Define the position vector r21 between an arbitrary point
on the line of action of F2 and the line of action of F1 . The torque T of the couple, which
is the moment of the two forces about an arbitrary point P , is then found from

T = r1 × F1 + r2 × F2 = r1 × F1 − (r1 − r21 ) × F1 = r21 × F1 (7.6)

We see that the torque does not depend on the selection of the point P .

Example 115 In this example we will derive force and torque expressions for a satellite
with six gas jet actuators and three momentum wheels. The gas jet actuators set up
forces Fj , and the momentum wheels set up torques Tj . The resultant force and the total
moment about the center of mass are then
6
X 3
X 6
X
F (r) = Fj , Nc = Tj + rj × Fj (7.7)
j=1 j=1 j=1

In the control of the attitude of the satellite it would make sense to arrange the gas jet
actuators in pairs that produce torques in the form of couples. This is done by requiring
F1 = −F4 , F2 = −F5 , F3 = −F6 , r1 = −r4 , r2 = −r5 and r3 = −r6 . This implies
that the resultant force is zero, that is, F (r) = 0. Therefore, the set of forces constitute
a couple, and because of this the moment about the center of mass is actually a torque
Tc = Nc given by
X3 X3
Tc = Tj + 2rj × Fj (7.8)
j=1 j=1

7.2.3 Equivalent force and torque

Two sets S and Σ of force vectors are said to be equivalent if they have equal resultant
and equal moment about any point. Consider a set S of nF forces with resultant force
(r)
FS and moment NS/P about a point P . An equivalent set Σ of forces can then be
defined with a single force and a torque due to a couple. To do this we let the set Σ to
be the force FΣ with line of action through the point P , and the torque TΣ so that
(r)
FΣ = FS , TΣ = NS/P (7.9)
266 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

To see that the sets S and Σ will be equivalent we observe that the set Σ will have
(r) (r)
resultant FΣ = FΣ = FS , and the moments about an arbitrary point Q will be equal,
which is confirmed by comparing the expression for NS/Q in (7.4) with the moment NΣ/Q ,
which is the torque TΣ plus the moment of FΣ about Q, that is,

NΣ/Q = TΣ + rQP × FΣ (7.10)

that result from (7.4).

We conclude that the following sets of forces are equivalent:

(r)
1. A set S with resultant FS and with moment NS/P about P , where the resultant
(r)
FS does not have a line of action, and where the moment NS/Q about a point Q
is found from the rule (7.4).
(r)
2. A force FΣ = FS with line of action through P in combination with a torque
TΣ = NS/P . Then the moment about a point Q is found from FΣ and TΣ according
to equation (7.10).

The main difference between the two equivalent representations S and Σ is that when
S is used the resultant is not a true force vector as it is not a bound vector, and the
additional rule (7.4) must be used to find the moment about some other point Q. In
contrast to this, when the set Σ is used, the force FΣ can be treated as a force vector and
the torque TΣ can be treated as a torque, and hence the usual definition of a moment
about a point can be used to calculate the moment about a point Q.

7.2.4 Forces and torques on a rigid body

A mass force is a force f dm that acts on a mass element dm = ρdV at position rp . An
example of this is the gravity force gdm acting on dm. The resultant gravity force on a
body is Z
G= gdm = mg (7.11)
b

The moment of the gravity force about the origin of frame i is

Z Z
NG/i = rp × gdm = rp dm × g = rc × mg = rc × G (7.12)
b b

The interpretation of this is that the gravity forces gdm will set up a moment equal to the
moment of the resultant gravity force G would give if G had line of action through the
center of mass. For this reason the center of mass is also called the center of gravity. In
this connection it may be argued that the concept of a center of mass is more fundamental
than a center of gravity which requires the presence of a field of gravity. From (7.4) it
follows that the moment of gravity about the center of mass is zero, that is, NG/c = 0.
The resultant forces acting on a body b will be
nF
X
(r)
Fb =G+ Fj (7.13)
j=1
7.2. FORCES AND TORQUES 267

where Fj are nF contact forces acting on the body. The moment on the body b about
its center of mass c is
nF
X
Nb/c = Tb + rcj × Fj (7.14)
j=1

where rcj = rF j − rc is the vector from the center of mass c to the line of action of the
forces Fj , and Tb is the contact torque due to couples acting the body. Typically this
would be motor torques. There is no moment from gravity as the moment is about the
center of mass. The moment about some other point o is found from the rule (7.4), which
gives
(r)
Nb/o = Nb/c + rg × Fb (7.15)

where
rg := roc (7.16)

is the vector from o to the center of mass c.

Equivalent descriptions of the forces and moments on a body b are (Kane and Levinson
1985)
(r)
1. The resultant force Fb without specification of line of action, the moment Nb/c
about the center of mass, and, in addition, the rule (7.15) for calculating the
moment about some other point o.
(r)
2. The force Fbc = Fb with line of action through the center of mass c in combination
with the torque Tbc = Nb/c .
(r)
3. The force Fbo = Fb with line of action through the point o in combination with
the torque Tbo = Nb/o . Then Fbo and Tbo can be found from Fbc and Tbc with

Fbo = Fbc (7.17)

Tbo = Tbc + rg × Fbc (7.18)

The resultant force and the moment are represented using Descriptions 2 and 3 is used
in the software package Autolev for multibody simulation based on Kane’s formulation
of the equations of motion.

Example 116 Suppose that Description 2 is used, and that Fbc and Tbc are given. Then
the moment on b about a point o is found from

Nb/o = Tbc + rg × Fbc (7.19)

If Description 3 is used and Fbo and Tbo are given, then the moment on b about c is found
from
Nb/c = Tbo − rg × Fbo (7.20)
268 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

7.2.5 Example: Robotic link

Consider a robot manipulator with 6 rigid bodies, called links, which are connected with
rotary joints. The forces acting on a link k are the contact force Fk−1,k on link k from
link k − 1, the contact force Fk+1,k from link k + 1 on link k, and the gravity force Gk .
The line of action of Fk−1,k passes through a point of position rk−1 , and the line of action
of Fk+1,k goes through a point of position rk . The center of mass has position rkc . We
note that due to the principle of action and reaction Fk+1,k = −Fk,k+1 where Fk,k+1 is
the force acting on link k + 1 from link k. The torques in the form of couples that act
on the link are the contact torque Tk−1,k on link k from link k − 1, and the contact force
Tk+1,k = −Tk,k+1 from link k + 1 on link k. This gives the following expression for the
resultant forces on link k
(r)
Fk = Fk−1,k + Fk+1,k + Gk (7.21)
The moment about the center of mass kc with position rkc is

Nk/kc = Tk−1,k + Tk+1,k + (rk−1 − rkc ) × Fk−1,k + (rk − rkc ) × Fk+1,k (7.22)

An equivalent description is possible with the force

(r)
Fkc := Fk (7.23)

with line of action through the center of mass, and the torque

Tkc := Nk/kc (7.24)

To calculate the moment Nk/k−1 on link k about the point k − 1 with position vector
rk−1 , it is used that the force Fkc has line of action through c, and the the moment is
found to be the torque Tkc plus the moment of Fkc about k − 1, which gives

Nk/k−1 = Tkc + (rkc − rk−1 ) × Fkc

We may check that this makes sense by inserting of the expressions for Tkc and Fkc ,
which gives

Nk/k−1 = Tk−1,k + Tk+1,k + (rk − rk−1 ) × Fk+1,k + (rkc − rk−1 ) × Gk (7.25)

7.3 Newton-Euler equations for rigid bodies

7.3.1 Equations of motion for a system of particles
Consider a system of N particles each of mass mk and with position rk relative to the
origin of the inertial frame i. By setting up Newton’s law for each particle and summing
up we get the result (Goldstein 1980), (Kane and Levinson 1985)

mac = F (r) (7.26)

where F (r) is the resultant force on the system of particles, and ac is the acceleration of
the center of mass.
The angular momentum of particle k about the center of mass is

hk/c = (rk − rc ) × mk vk (7.27)

7.3. NEWTON-EULER EQUATIONS FOR RIGID BODIES 269

zi
m 
F

rp

yi
xi

Figure 7.2: Mass point subject to a force F .

where vk is the velocity of particle k and rc is the position of the center of mass. With
reference to frame i, the time derivative of hk/c is
i
d
hk/c = (vk − vc ) × mk vk + rck × mk ak
dt
= −vc × mk vk + rck × mk ak (7.28)

Summation over all particles leads to

i
d
hc = Nc (7.29)
dt
where (6.413) and (7.26) is used, and where
N
X
hc = rck × mk vk (7.30)
k=1

is the angular momentum of the system about the center of mass, and
N
X
Nc = rck × Fk (7.31)
k=1

is the moment of the forces about the center of mass. This means that the time derivative
of the angular momentum about the center of mass is equal to the moment of the forces
about the center of mass.

7.3.2 Equations of motion for a rigid body

This result (7.26, 7.29) in the previous section was derived for a system of N particles.
The result can be generalized to a rigid body b by summing up the equations of motion for
mass elements dm of position rp , velocity vp and acceleration ap . To simplify expressions
the set of forces and torques acting on the rigid body is represented by the equivalent
270 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

set with a force Fbc with line of action through the center of mass and magnitude equal
to the resultant force, and a torque Tbc = Nb/c that equals the moment about the center
of the mass. The equations of motion for a rigid body are then found from (7.26, 7.29)
to be

Fbc = mac (7.32)

i
d
Tbc = hb/c (7.33)
dt
where Z
hb/c = r × vp dm (7.34)
b
is the angular momentum of the body b about the center of mass, and

r = rp − rc (7.35)

is the position of the mass element relative to the center of mass. Using vp = vc + ω ib × r,
this can be written
Z Z
hb/c = rdm × vc + r × (ω ib × r)dm
Zb b

= r × (ω ib × r)dm
b
Z
= − r × (r × ω ib )dm (7.36)
b

where we have used (6.417). By introducing the dyadic representation of the vector cross
product we may write this in the form
Z Z
hb/c = − r× · (r× · ω ib )dm = − r× · r× dm · ω ib (7.37)
b b

This expression motivates the definition of the inertia dyadic of b about c as

Z
Mb/c = − r× · r× dm (7.38)
b

The angular momentum about c can then be written

hb/c = Mb/c · ω ib (7.39)

Insertion in (7.33) gives

i
d³ ´ bd
Tbc = Mb/c · ω ib = (Mb/c · ω ib ) + ω ib × (Mb/c · ω ib ) (7.40)
dt dt
Finally it is used that Mb/c is constant in b. This leads to the following result:

When referenced to the center of mass the equation of motion for a rigid body can be
written

Fbc = mac (7.41)

Tbc = Mb/c · αib + ω ib × (Mb/c · ω ib ) (7.42)
7.3. NEWTON-EULER EQUATIONS FOR RIGID BODIES 271

dm
r r´

c o
rg

Figure 7.3: Definition of the vectors r, r0 and rg .

7.3.3 Equations of motion about a point

In important applications like ship dynamics and aeroplane dynamics the motion of the
body b is described in terms of translation of a fixed point o which is not the mass center,
and the rotation about the point o. In this case it is convenient to represent the forces
and the moments acting on a rigid body by an equivalent set with a force Fbo = Fbc
with line of action through the point o and magnitude equal to the resultant force, and
a torque
Tbo = Tbc + rg × Fbc (7.43)

where rg = rc − ro is the vector from o to c as shown in Figure 7.3. We mention

the following result before proceeding towards a description where the dynamics are
referenced to a point o:

A mixed formulation of the equations of motion where the force and torque are referenced
to the point o and the acceleration and inertia dyadics are referenced to the mass center
is given by

Fbo = mac (7.44)

Tbo = rg × mac + Mb/c · αib + ω ib × (Mb/c · ω ib ) (7.45)

Then the force equation can be referenced to the point o by combining (7.32) and
(6.410). This gives

Fbo = m [ao + αib × rg + ω ib × (ω ib × rg )] . (7.46)

The torque equation involves the angular momentum about c, while we would like to
have an expression involving the angular momentum about o, which is given by
Z
hb/o = r 0 × vp dm (7.47)
b

where

rp = ro + r 0 (7.48)
vp = vo + ω ib × r 0 (7.49)
272 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

Insertion of (7.49) into (7.47) leads to

Z Z
hb/o = r dm × vo + r 0 × (ω ib × r 0 ) dm
0
b b
Z Z
= (rp − ro ) dm × vo − r 0 × (r 0 × ω ib ) dm
b b
= rg × mvo + Mb/o · ω ib (7.50)

where the inertia dyadic Mb/o of b about o is defined by

Z
Mb/o = − (r 0 )× · (r 0 )× dm (7.51)
b

Note that Mb/o is constant in frame b. Time differentiation with respect to reference to
frame i gives
i
d d³ b ´
hb/o = (vc − vo ) × mvo + rg × mao + Mb/o · ω ib + ω ib × (Mb/o · ω ib )
dt dt
= vc × mvo + rg × mao + Mb/o · αib + ω ib × (Mb/o · ω ib ) (7.52)

We may also express the angular momentum about o with the angular momentum
about c by combining (7.34), (7.47) and r 0 = r + rg . This gives
Z Z
hb/o = (r + rg ) × vp dm = hb/c + rg × vp dm (7.53)
b b
= hb/c + rg × mvc (7.54)

which implies that

i i
d d
hb/o = hb/c + rg × mac − vo × mvc (7.55)
dt dt
From this equation and (7.52) it follows that
i
d
h = rg × m (ao − ac ) + Mb/o · αib + ω ib × (Mb/o · ω ib ) (7.56)
dt b/c
This result in combination with equations (7.32), (7.33) and (7.43) gives

Tbo = rg × mao + Mb/o · αib + ω ib × (Mb/o · ω ib ) (7.57)

With reference to a point o the equation of motion for a rigid body can be written

Fbo = m [ao + αib × rg + ω ib × (ω ib × rg )] (7.58)

Tbo = rg × mao + Mb/o · αib + ω ib × (Mb/o · ω ib ) (7.59)

7.3.4 The inertia dyadic

The inertia dyadic of b about the center of the mass was defined as
Z
Mb/c = − r× · r× dm (7.60)
b
7.3. NEWTON-EULER EQUATIONS FOR RIGID BODIES 273

From (6.73) the alternative expression

Z
Mb/c = (r 2 I − rr)dm (7.61)
b

is found. The dyadic can be evaluated in the b frame and is written

3 X
X 3
Mb/c = mbij bi bj (7.62)
i=1 j=1

Example 117 Consider a rigid body with a fixed coordinate frame b with orthogonal unit
vectors b1 , b2 and b3 that coincide with the main axes of inertia of the body b. Then the
inertia dyadic is
Mb/c = m11 b1 b1 + m22 b2 b2 + m33 b3 b3 . (7.63)
where m11 , m22 and m33 are constants. The angular velocity is written

ω ib = ω 1 b1 + ω 2 b2 + ω 3 b3 (7.64)

where ω i = ω ib · bi is the component of ω ib along bi . The angular momentum is then

hb/c = (m11 b1 b1 + m22 b2 b2 + m33 b3 b3 ) · (ω 1 b1 + ω 2 b2 + ω 3 b3 ). (7.65)

As the unit vectors are orthogonal, it follows that bi · bj = 0 for i 6= j, and bi · bi = 1.

This gives
hb/c = m11 ω 1 b1 + m22 ω 2 b2 + m33 ω 3 b3 (7.66)

Example 118 The kinetic energy of a rigid body is

Z
1
K= vp · vp dm. (7.67)
2 b

Insertion of vp = vc + ω ib × r gives
Z
1 1
K= mvc 2 + (ω ib × r) · (ω ib × r)dm (7.68)
2 2 b
R
as vc · ω ib × b
rdm = 0. The last term on the right side is simplified using
Z Z
1 1
(ω ib × r) · (ω ib × r)dm = − (ω ib × r) · (r × ω ib )dm
2 b 2 b
Z
1
= − ω ib · r× · r× · ω ib dm
2 b
Z
1
= − ω ib · r× · r× dm · ω ib
2 b
1
= ω ib · Mb/c · ω ib (7.69)
2
This leads to the following expression for the kinetic energy:
1 1
K= mvc 2 + ω ib · Mb/c · ω ib (7.70)
2 2
274 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

Example 119 Using (7.49) the kinetic energy is found from the computation
Z
1
K = vp · vp dm
2 b
µ Z ¶ Z
1 0 1
= mvo · vo + vo · ω ib × r dm + (ω ib × r 0 ) · (ω ib × r 0 )dm
2 b 2 b
1 1
= mvo · vo + vo · (ω ib × mrg ) + ω ib · Mb/o · ω ib (7.71)
2 2
to have the form
1 1
K = mvo · vo − vo · mrg× · ω ib + ω ib · Mb/o · ω ib (7.72)
2 2

7.3.5 The inertia matrix

The matrix representation of the inertia dyadic Mb/c in frame b is the inertia matrix
Z
¡ b ¢× ¡ b ¢×
Mbb/c = − r r dm (7.73)
b
From (6.25) the more usual expression
Z h
b
¡ b ¢T b ¡ ¢T i
Mb/c = r r I − rb rb dm (7.74)
b
is found. The angular momentum of b about c can then be written in column vector
form as
hbb/c = Mbb/c ω bib (7.75)
Equation (7.75) can be transformed to frame i using the transformation rule:
Rib hbib = Rib Mbb/c ω bib (7.76)
Insertion of ω bib = Rbi ω iib gives
hiib = Rib Mbb/c Rbi ω iib (7.77)
Moreover, the inertia dyadic can also be represented by a matrix frame i using
Z h
¡ i ¢T i ¡ ¢T i
Mib/c = r r I − ri ri dm (7.78)
b
which satisfies
hib/c = Mib/c ω iib (7.79)
Comparison of the two expressions (7.77) and (7.79) leads to the conclusion
Mib/c = Rib Mbb/c Rbi (7.80)
We see that the inertia matrix transforms from frame b to frame i by a similarity
transformation. This is to be expected as it is the matrix representation of a dyadic.
The generic element mbij of the dyadic is a second order tensor. Because of this the
inertia matrix is often referred to as the inertia tensor .
Example 120 The kinetic energy as given by (7.70) can be expressed in coordinate form
as
1 ¡ ¢T 1 ¡ b ¢T b
K = m vcb vcb + ω Mb/c ω bib (7.81)
2 2 ib
The second term is the kinetic energy due to rotation. As the kinetic energy is greater or
equal to zero, it follows that Mbb/c is a positive definite matrix.
7.3. NEWTON-EULER EQUATIONS FOR RIGID BODIES 275

o c

l
2 l xb

Figure 7.4: Slender beam of length l.

7.3.6 Expressions for the inertia matrix

The inertia matrix is given by
Z
£ b 2 ¤
Mbb/c = (r ) I − rb (rb )T dm (7.82)
b

in the body-fixed frame b. Let rb = (x, y, z)T . The the inertia matrix is then found to be
 2 
Z y + z2 −xy −xz
Mbb/c =  −xy x2 + z 2 −yz  dm (7.83)
b −xz −yz x + y2
2

Under the assumption that the b frame is fixed in the body b, the inertia matrix Mbb/c
in the b frame is a constant matrix. In frame i we have Mib/c = Rib Mbb/c Rbi which will
not be constant if frame b is rotating relative to frame i.

7.3.7 The parallel axes theorem

The inertia dyadic about the point o is
Z Z
Mb/o = − (r 0 )× · (r 0 )× dm = − (r + rg )× · (r + rg )× dm
b b
Z µZ ¶× µZ ¶× Z
× ×
= − r · r dm − rg× · rdm − rdm × × ×
· rg − rg · rg dm
b b b b

Using (6.417) we find the following result:

The inertia dyadic of b about o is related to the inertia dyadic of b about c according to
h i
Mb/o = Mb/c − mrg × rg × = Mb/c + m (rg · rg )I − rg rg (7.84)

The corresponding matrix expressions is

¡ ¢× ¡ b ¢× £ ¤
Mbb/o = Mbb/c − m rbg rg = Mbb/c + m (rbg )2 I − rbg (rbg )T (7.85)

This is the parallel axes theorem.

Example 121 A slender beam has length and mass m. A coordinate frame b is fixed
in the beam with the x axis in the length axis of the beam as shown in Figure 7.4. The
mass element is set to be dm = (m/ )dx. The inertia matrix about the center of mass is
 
0 0 0
2
Mbb/c =  0 m12 0  (7.86)
m 2
0 0 12
276 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

The inertia matrix about the endpoint o of the beam is found from the parallel axes
theorem to be
     
0 0 0 µ ¶2 m( 2 )2 0 0 0 0 0
2 2
b
Mb/o =  m
0 12 0  +m I−  0 0 0  =  0 m
0 
2 3
m 2 0 0 0 m 2
0 0 12 0 0 3

Example 122 The kinetic energy of a rigid body can be expressed with reference to a
point o by

1 ¡ b ¢T b 1 ¡ b ¢T b
K = m vc vc + ω Mb/c ω bib
2 2 ib
1 ¡ b ¢T ¡ b ¢ 1 ¡ b ¢T b
= m vo + (ω bib )× rbg vo + (ω bib )× rbg + ω Mb/c ω bib
2 2 ib
1 ¡ b ¢T b ¡ ¢T ¡ ¢T
= m vc vc + m vob (rbg )×T ω bib + m (ω bib )× rbg vob
2
1 ¡ ¢T ³ b ¡ ¢× ¡ b ¢× ´ b
+ ω bib Mb/c − m rbg rg ω ib
2 Ã
µ b ¶T ¡ ¢×T ! µ ¶
1 vo mI m rbg vob
= ¡ ¢ × (7.87)
2 ω bib m rbg Mbb/o ω bib

when the description is referenced to a fixed point o in the body. In the derivation the
T T
rules a× b = −b× a and (a× b) c = (c× b) a are used.

7.3.8 The equations of motion for a rigid body

In this section we will sum up with different versions of the equations of motion for
(r)
a rigid body b where the resultant force is Fb and the total moment on b about the
center of mass is Nb/c . First we represent the forces and the moments by the equivalent
(r)
representation with a force Fbc = Fb with line of action through the center of mass c
in combination with a torque Tbc = Nb/c . The equations of motion are

Fbc = mac (7.88)

Tbc = Mb/c · αib + ω ib × (Mb/c · ω ib ) (7.89)

In the b frame the coordinate form is written in matrix form as

µ ¶µ ¶ à ! µ ¶
mI 0 abc 0
¡ b ¢× b Fbbc
+ = (7.90)
0 Mbb/c αbib ω ib Mb/c ω bib Tbb/c

b
d
In view of ac = dt vc + ω ib × vc the equations of motion can be written
b
d
Fbc = m vc + mω ib × vc (7.91)
dt
Tbc = Mb/c · αib + ω ib × (Mb/c · ω ib ) (7.92)

or
7.3. NEWTON-EULER EQUATIONS FOR RIGID BODIES 277

µ ¶µ ¶ à ¡ ¢× ! µ ¶
mI 0 v̇cb m ω bib vcb Fbbc
+ ¡ b ¢× b = (7.93)
0 Mbb/c αbib ω ib Mb/c ω bib Tbbc

The representation of the forces and torques is changed to an equivalent representation

(r)
with a force Fbo = Fb with line of action through o in combination with a torque
Tbo = Tbc + rg × Fbc , where rg is the vector from o to c. The equations of motion are
then

Fbo = m[ao + αib × rg + ω ib × (ω ib × rg )] (7.94)

Tb/o = rg × mao + Mb/o · αib + ω ib × (Mb/o · ω ib ) (7.95)

In the special case where o is the center of mass, then rg = 0, and the result is the same
as in (7.91, 7.92).
The coordinate form in the b frame is written in matrix form as

µ ¶µ ¶ à ¡ ¢× ! µ ¶
mI m(rbg )×T abo m(ω bib )× ω bib rbg Fbbo
+ = (7.96)
m(rbg )× Mbb/o αbib (ω bib )× Mbb/o ω bib Tbbo

Here it is used that a × b = −b × a for any two vectors a and b, and that (·)× =
T
− [(·)× ] . Note that the leading matrix on the left side is symmetric and positive definite.
This matrix can be regarded as a mass matrix.
An alternative formulation is
·b ¸
d
Fbo = m vo + ω ib × vo + αib × rg + ω ib × (ω ib × rg ) (7.97)
dt
b
d
Tbo = mrg × vo + mrg × (ω ib × vo ) + Mb/o · αib + ω ib × (Mb/o · ω ib ) (7.98)
dt
with matrix form

µ ¶µ ¶ µ £ ¤ ¶ µ b ¶
mI m(rbg )×T v̇ob m(ω bib )× (ω bib )× rbg + vob Fbo
+ =
m(rbg )× Mbb/o αbib (ω bib )× Mbb/o ω bib + m(rbg )× (ω bib )× vob Tbbo

7.3.9 Satellite attitude dynamics

Suppose that the inertia matrix in the body-fixed frame b is

Mbb/c = diag(m11 , m22 , m33 ). (7.99)

Then the angular momentum is

 
m11 ω 1
hbb/c = Mbb/c ω bib =  m22 ω 2  . (7.100)
m33 ω 3
278 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

The torque Tbbc = (T1 , T2 , T3 )T is acting on the body. The torque law is then
i i ³ ´ bd ³ ´ ³ ´
d d
Tbc = hb/c = Mb/c · ω ib = Mb/c · ω ib + ω ib × Mb/c · ω ib (7.101)
dt dt dt
The inertia dyadic is constant in the b frame, therefore
³ ´
Mb/c · αib + ω ib × Mb/c · ω ib = Tbc (7.102)

where (6.402) is used. With coordinate vectors this is written

Mbb/c ω̇ bib + (ω bib )× Mbb/c ω bib = Tbbc . (7.103)
Written out in components the model is
m11 ω̇ 1 + (m33 − m22 )ω 2 ω 3 = T1 (7.104)
m22 ω̇ 2 + (m11 − m33 )ω 3 ω 1 = T2 (7.105)
m33 ω̇ 3 + (m22 − m11 )ω 1 ω 2 = T3 (7.106)

7.4 Example: Ball and beam dynamics

g R
 m

b 1 z b ,b 3 

b2
yb
xb

Figure 7.5: Ball and beam system.

In this section we will derive the equations of motion for a ball-and-beam system with
a ball that rolls in a track on a beam. To do this we start with the kinematics of the
system, and then combine the equations of motion for the ball and for the beam. We
fix a coordinate system b in the beam so that the xb axis is along the track, and the zb
axis is along the motor shaft. The orthogonal unit vectors b1 , b2 , b3 are placed along the
xb , yb , zb axes. According to (6.103) the scalar products of the unit vectors of frames n
and b are related by
n1 · b1 = cos θ, n1 · b2 = − sin θ (7.107)
n2 · b1 = sin θ, n2 · b2 = cos θ (7.108)
n3 · b1 = n3 · b2 = 0, n3 · b3 = 1 (7.109)
which implies that
n1 = cos θb1 − sin θb2 (7.110)
n2 = sin θb1 + cos θb2 , (7.111)
n3 = b3 (7.112)
7.4. EXAMPLE: BALL AND BEAM DYNAMICS 279

We note that

n1 × b1 = sin θb3 , n1 × b2 = cos θb3 (7.113)

n2 × b1 = − cos θb3 , n2 × b2 = sin θb3 (7.114)

Note that n2 is pointing vertically downwards so that the acceleration of gravity is

g = gn2 . The beam is rotated with angular velocity ω 1 = θ̇b3 by a motor so that
the track can be given an angle θ relative to the horizontal line, and the ball can be
made to roll along the beam. The system is shown in Figure 7.5.
The radius of the ball is R, and the position of the ball along the track is denoted by
x. The position of the center of the ball is

r2 = xb1 − Rb2 .

The velocity is
b
d ³ ´ ³ ´
v2 = r + ω 1 × r = ẋb1 + θ̇b3 × xb1 − Rb2 = ẋ + θ̇R b1 + θ̇xb2 , (7.115)
dt
and the acceleration is
b
d
a2 = v2 + ω 1 × v2
dt
³ ´ ³ ´ h³ ´ i
= ẍ + θ̈R b1 + θ̈x + θ̇ẋ b2 + θ̇b3 × ẋ + θ̇R b1 + θ̇xb2
³ 2
´ ³ 2
´
= ẍ + θ̈R − θ̇ x b1 + θ̈x + 2θ̇ẋ + θ̇ R b2 . (7.116)

It follows that the ball rolls along the track with an angular velocity given by
µ ¶
ẋ
ω 2 = θ̇ + b3 (7.117)
R

as it is assumed that the ball does not slide.

The kinematic equations have now been established, and we will now develop the
equations of motion. The mass of the ball is m, and the moment of inertia of the ball
about its center of inertia is
2
J2 = m2 R2 (7.118)
5
which is tabulated in textbooks on dynamics. The contact force acting from the beam
on the ball is
F = Fx b1 + Fy b2 (7.119)
while the gravitational force on the ball is
³ ´
G = m2 gn2 = m2 g sin θb1 + cos θb2 . (7.120)

It is noted that the contact torque between the ball and the beam is zero.
The angular momentum equation for the ball is (7.32)
n
µ ¶
d ẍ
T2c = J2 ω 2 = J2 θ̈ + b3 (7.121)
dt R
280 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

It is convenient to use the moment

³ ´
N2/o = −Rb2 × G2 = m2 Rg sin θb3 (7.122)

about the contact point between the ball and the beam in the equation of motion. The
reason for this is that the unknown constraint force Fx will not show up in the torque
in this case. From (7.45) the moment N2/o and the torque T2c are related through the
expression ³ ´
N2/o = T2c + −Rb2 × m2 a (7.123)
which gives
µ ¶ ³ ´
ẍ 2
m2 Rg sin θ = J2 θ̈ + + m2 R ẍ + θ̈R − θ̇ x (7.124)
R
µ ¶
¡ ¢ J2 2
= J2 + m2 R2 θ̈ + + m2 R ẍ − m2 Rxθ̇ (7.125)
R
This is written
¡ ¢ 1 ¡ ¢ 2
J2 + m2 R2 θ̈ + J2 + m2 R2 ẍ = m2 Rxθ̇ + Rm2 g sin θ (7.126)
R
By inserting the value of J2 from (7.118), we get
¡ ¢ 7
J2 + m2 R2 = m2 R2 (7.127)
5
The Newton’s law for the ball is
m2 a2 = F + G2 (7.128)
In the yb direction this gives
³ 2
´
m2 θ̈x + 2θ̇ẋ + θ̇ R = Fy + m2 g cos θ (7.129)

To proceed an expression for the contact force Fy from the beam on the ball is needed.
This can be found from the equation of motion for the beam. The contact force from
the ball on the beam in the yb direction is −Fy . In the equation of motion for the beam
this gives
J1 θ̈b3 = xb1 × (−Fy b2 ) + T b3 (7.130)
which leads to
J1 θ̈ = −xFy + T (7.131)
This equation is combined with (7.129), and the result is
¡ ¢ 2
J1 + m2 x2 θ̈ = T + m2 gx cos θ − 2m2 xθ̇ẋ − m2 θ̇ xR (7.132)
where T is the motor torque and Fy is the contact force in the yb direction.

The model of the ball and beam is given by

¡ ¢ 2
J1 + m2 x2 θ̈ = T + m2 gx cos θ − 2m2 xθ̇ẋ − m2 θ̇ xR (7.133)
¡ ¢ 1 ¡ ¢ 2
J2 + m2 R2 θ̈ + J2 + m2 R2 ẍ = m2 Rxθ̇ + Rm2 g sin θ (7.134)
R
7.5. EXAMPLE: INVERTED PENDULUM 281

Example 123 The rate of change of the energy of the ball and beam system will be equal
to the power θ̇T supplied by the torque T . If the model does not satisfy this condition,
then the model is not correct, which provides us with a method to check the validity of
the model. The total energy of the system is
1 1 1
V = J1 ω 1 · ω 1 + J2 ω 2 · ω 2 + m2 v2 · v2 + m2 g (−x sin θ + R cos θ)
2 2 2
µ ¶2 µ³ ´2 ³ ´2 ¶
1 2 1 ẋ 1
= J1 θ̇ + J2 θ̇ + + m2 ẋ + θ̇R + θ̇x
2 2 R 2
+m2 g (−x sin θ + R cos θ) (7.135)
The time derivative along the solutions of the system is
µ ¶ µ ¶ ³ ´ ³ ´
ẋ ẍ
V̇ = θ̇J1 θ̈ + θ̇ + J2 θ̈ + + ẋ + θ̇R m2 ẍ + θ̈R
R R
³ ´
+θ̇xm2 θ̈x + θ̇xm2 θ̇ẋ − m2 g ẋ sin θ + xθ̇ cos θ + Rθ̇ sin θ
µ ¶ µ ¶
¡ ¢ ẋ ¡ ¢ ẍ 2
= θ̇ J1 + m2 x2 θ̈ + θ̇ + J2 + m2 R2 θ̈ + + m2 xẋθ̇
R R
³ ´
−m2 g ẋ sin θ + xθ̇ cos θ + Rθ̇ sin θ
= θ̇T (7.136)
This result shows that the model is consistent with the energy flow in the system.
Example 124 Insertion of ω 2 = θ̇ + ẋ/R gives a diagonal mass matrix:
¡ ¢ 2
J1 + m2 x2 θ̈ = T + m2 gx cos θ − 2m2 xθ̇ẋ − m2 θ̇ xR (7.137)
7 2
m2 R2 ω̇ 2 = Rm2 θ̇ x + Rm2 g sin θ (7.138)
5
Example 125 If the radius of the ball becomes small, that is, when R → 0, then the
model becomes
¡ ¢
J1 + mvx2 θ̈ = T + m2 gx cos θ − 2mvxθ̇ẋ (7.139)
7 2
m2 ẍ = m2 θ̇ x + m2 g sin θ (7.140)
5
Example 126 Linearization about θ̇ = 0, θ = 0, ẋ = 0 and x = 0 gives
J1 θ̈ = T + m2 gx (7.141)
5
ẍ = gθ (7.142)
7
which gives
d4 1 5 g
x= m2 gx + T (7.143)
dt4 J1 7 J1

7.5 Example: Inverted pendulum

7.5.1 Equations of motion
Consider a pendulum on a cart as shown in Figure 7.6. The mass of the cart is mv , the
position of the cart is x, and the force on the cart is F . The pendulum is a point mass
282 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

mp
p

mv

m

Figure 7.6: Inverted pendulum.

mb at the end of a massless rod of length Lb . The angle of the pendulum is denoted θb ,
which is zero at the upright position.
To derive the equations of motion for the system we will first describe the kinematics
of the system by assigning coordinate frames n and b to describe the motion of the cart
and the pendulum, and then derive kinematic equations for the unit vectors of frames
n and b. A non-moving coordinate frame n is defined with unit vector n1 along the
motion of the cart, with n2 in the vertical downwards direction and n3 along the axis of
rotation. A frame b is fixed to the pendulum. The rotation matrix is Rnb = Rz,θb which
is a rotation by an angle θb about the axis defined by n3 = b3 . The angular velocity of
the pendulum is ω b = θ̇b n3 = θ̇b b3 . The relation between the unit vectors in frames n
and b is as given in Section 7.4.
The next step in the derivation of the equations of motion is to derive kinematic
equations for position, velocity and acceleration. The position of the point mass m1 is

rb = xn1 − Lb b2 (7.144)

The velocity is found to be

n
drb
vb = = ẋn1 − θ̇b b3 × Lb b2 = ẋn1 + θ̇b Lb b1 (7.145)
dt
and acceleration is
2
ab = ẍn1 + θ̈b Lb b1 + θ̇b b3 × θ̇b Lb1 = ẍn1 + θ̈b Lb b1 + θ̇b Lb b2 (7.146)

At this stage, the kinematic model has been established, and the equations of motion
can be derived. This is done by combining Newton’s law for the point mass and for the
cart, and with the torque law for the pendulum. Newton’s law for the point mass gives

mb ab = Fb + mgn2 (7.147)

where g is the acceleration of gravity. In the n1 direction this gives

³ 2
´
mb ẍ + θ̈b Lb cos θb − θ̇b Lb sin θb = Fb · n1 (7.148)
7.5. EXAMPLE: INVERTED PENDULUM 283

Newton’s law for the cart gives

mv ẍ = F − Fb · n1 (7.149)

Combination of the two equations gives

2
(mv + mb )ẍ + mb Lb θ̈b cos θb − mb θ̇b Lb sin θb = F (7.150)

The torque law for the pendulum about the connection point is according to (7.45)
³ ´
−Lb b2 × mb gn2 = −Lb b2 × mb ẍn1 + θ̈b Lb b1 − θ̇b Lb b2 (7.151)

where rg = −Lb b2 in the notation of (7.45). The component of this equation in the n3
direction is
mb Lb g sin θb = mb Lb ẍ cos θb + mb L2b θ̈b (7.152)

The model for the cart and pendulum has then been found to be
2
(mv + mb )ẍ + mb Lb θ̈b cos θb = mb θ̇b Lb sin θb + F (7.153)
mb L2b θ̈b + mb Lb ẍ cos θb = mb Lb g sin θb (7.154)

where F is the external force acting on the cart.

Example 127 The rate of change of the energy in the system is equal to the power F ẋ
supplied by the external force F . We will now check if the model is consistent with this
observation. The total energy of the system is
1 1
V = mv ẋ2 + mb vb · vb + mb gLb cos θb
2 2
1 1 2
= mv ẋ2 + mb (ẋ2 + 2Lb cos θb ẋθ̇b + θ̇b L2b ) + mb gLb cos θb (7.155)
2 2
The time derivative of the energy along the solutions of the system is
h i ³ ´
V̇ = ẋ (mv + mb )ẍ + mb Lb cos θb θ̈b + θ̇b mb L2b θ̈b + mb Lb cos θb ẍ
2
−Lb mb sin θb ẋθ̇b − mb gLb θ̇b sin θb
³ 2
´ 2
= ẋ mb θ̇b Lb sin θb + F + θ̇b mb Lb g sin θb − Lb mb sin θb ẋθ̇b − mb gLb θ̇b sin θb
= F ẋ (7.156)

This shows that the model is consistent with the energy flow of the system.

Next we combine the cart and pendulum model with the motor model. The cart is
controlled with a current controlled DC motor with dynamics given by

Jm θ̈m = KT u − TL (7.157)

where θm is the motor angle, u is the input, KT is the torque constant, Jm is the inertial
of the motor, and TL is the load torque from the cart. The motor is connected to the cart
with a string that runs over a pulley fixed to the motor axis. The radius of the pulley is
r, and it follows that
TL = rF, ẋ = rθ̇m (7.158)
284 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

which gives
Jm KT
ẍ = u−F (7.159)
r2 r
It is observed that the cart and pendulum is driven by the motor through a port with
effort F and and flow ẋ. The effort F is input to the cart and pendulum model, and the
flow ẋ is output. At the same time the motor model has input F and output ẋ. This
means that the inputs and the outputs of the port interconnection are incompatible, so
that the equations must be combined by adding equations (7.153) and (7.159). This gives
2 KT
(m + mb )ẍ + mb Lb cos θb θ̈b − mb θ̇b Lb sin θb = u (7.160)
r
where m = mv + Jm /r2 .

The model of cart, pendulum and motor is

2 KT
(m + mb )ẍ + mb Lb cos θb θ̈b − mb θ̇b Lb sin θb = u (7.161)
r
mb L2b θ̈b + mb Lb ẍ cos θb = mb Lb g sin θb (7.162)

7.5.2 Double inverted pendulum

mc
c

b mb

mv

m

Figure 7.7: Double inverted pendulum.

A double pendulum system is obtained by adding one more pendulum to the cart
and pendulum system as shown in Figure 7.7. The variables of the second pendulum are
denoted with a subscript c. The position of the point mass mc of the second pendulum
is
rc = xn1 − Lc c2 (7.163)
The velocity is
vc = ẋn1 + θ̇c Lc c1 (7.164)
7.6. EXAMPLE: THE FURUTA PENDULUM 285

and acceleration is
2
ac = ẍn1 + θ̈c Lc1 + θ̇c Lc c2 (7.165)
Newton’s law for the point mass of the second pendulum gives
³ 2
´
mc ẍ + θ̈c Lc cos θc − θ̇c Lc sin θc = Fc · n1 (7.166)

Newton’s law for the cart is modified by one additional term, which is due to the contact
force from the second pendulum. This gives

mv ẍ = F − Fb · n1 − Fc · n1 (7.167)

The torque law for the second pendulum about the connection point is

mc Lc g sin θc = mc Lc ẍ cos θc + mc L2c θ̈c (7.168)

The model for a cart and two pendulums then is found to be

(mv + mb + mc )ẍ + mb Lb θ̈b cos θb + mc Lc θ̈c cos θc

2 2
−mb θ̇b Lb sin θb − mc θ̇c Lc sin θc = F (7.169)
mb L2 θ̈b + mb Lẍ cos θb = mb Lb g sin θb (7.170)
mc L2c θ̈c + mc Lc ẍ cos θc = mc Lc g sin θc (7.171)

The motor model is included by inserting

KT Jm
F = u − 2 ẍ (7.172)
r r

7.6 Example: The Furuta pendulum

The Furuta pendulum is a laboratory example where a rotational joint with vertical axis
of rotation is used to balance an inverted pendulum (Aström and Furuta 2000). The
inertial frame n is defined with the n3 axis vertically upwards. The frame b is obtained
by a rotation θ1 about the n3 vector, and the frame c is obtained by a rotation θ2 about
the b2 axis (Figure 7.8). According to (6.103) the frames n and b have direction cosines

n1 · b1 = cos θ1 , n1 · b2 = − sin θ1 (7.173)

n2 · b1 = sin θ1 , n2 · b2 = cos θ1 (7.174)
n1 · b3 = n2 · b3 = n3 · b1 = n3 · b2 = 0, n3 · b3 = 1 (7.175)

and the unit vectors of frame b and frame c have direction cosines

b1 · c1 = cos θ2 , b1 · c3 = sin θ2 (7.176)

b1 · c2 = b3 · c2 = b2 · c1 = b2 · c3 = 0, b2 · c2 = 1 (7.177)
b3 · c1 = − sin θ2 , b3 · c3 = cos θ2 (7.178)

It is noted that

b1 = cos θ2 c1 + sin θ2 c3 (7.179)

b3 = − sin θ2 c1 + cos θ2 c3 (7.180)
286 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

c3
m
c2
 c1
n3

n1 L2


n2 1 2 
b3
L1
T

b2

b1

Figure 7.8: Coordinate frames n, b and c used in the description of the Furuta pendulum.

and that

b1 × c3 = − cos θ2 c2 (7.181)
b3 × c1 = cos θ2 c2 , b3 × c2 = − cos θ2 c1 − sin θ2 c3 , b3 × c3 = sin θ2 c2 (7.182)

The acceleration of gravity is g = −gn3 . The first link is rotated with angular velocity

ω 1 = θ̇1 b3 (7.183)

and the second link is rotated with angular velocity

ω 2 = θ̇1 b3 + θ̇2 c2 (7.184)

The position of the mass

r = L1 b2 + L2 c3
The velocity is
³ ´
v = L1 θ̇1 b3 × b2 + L2 θ̇1 b3 + θ̇2 c2 × c3

= −L1 θ̇1 b1 + L2 θ̇1 sin θ2 c2 + L2 θ̇2 c1 (7.185)

and the acceleration is

a = −L1 θ̈1 b1 + L2 θ̈1 sin θ2 c2 + L2 θ̇1 θ̇2 cos θ2 c2 + L2 θ̈2 c1

³ ´ ³ ´
−L1 θ̇1 θ̇1 b3 × b1 + θ̇1 b3 + θ̇2 c2 × L2 θ̇1 sin θ2 c2 + L2 θ̇2 c1

= −L1 θ̈1 b1 + L2 θ̈1 sin θ2 c2 + L2 θ̇1 θ̇2 cos θ2 c2 + L2 θ̈2 c1

2 2 2
−L1 θ̇1 b2 − L2 θ̇1 sin θ2 (cos θ2 c1 + sin θ2 c3 ) + L2 θ̇1 θ̇2 cos θ2 c2 − L2 θ̇2 c3
7.6. EXAMPLE: THE FURUTA PENDULUM 287

which gives
³ 2
´
a = −L1 θ̈1 b1 + L2 θ̈1 sin θ2 + 2L2 θ̇1 θ̇2 cos θ2 − L1 θ̇1 b2
³ 2
´ ³ 2 2
´
+ L2 θ̈2 − L2 θ̇1 sin θ2 cos θ2 c1 − L2 θ̇1 sin2 θ2 + L2 θ̇2 c3 (7.186)

The kinematic equations have now been established, and we will develop the equations
of motion. Newton’s law for the mass gives

F = ma

where F is the force on link 2 from link 1. The the torque law for the first angle is

T = J1 θ̈1 + (r × ma) · b3 (7.187)

After some relative extensive vector calculations we may find that

h³ ´ i
(r × ma) · b3 = L1 b2 + L2 c3 × ma · b3
2
= mL21 θ̈1 − mL1 L2 cos θ2 θ̈2 + mL1 L2 θ̇2 sin θ2
+mL22 θ̈1 sin2 θ2 + 2mL22 θ̇1 θ̇2 sin θ2 cos θ2 (7.188)

This gives the equation of motion

J1 θ̈1 + mL21 θ̈1 + mL22 sin2 θ2 θ̈1 − mL1 L2 cos θ2 θ̈2

2
= T − mL1 L2 θ̇2 sin θ2 − 2mL22 θ̇1 θ̇2 sin θ2 cos θ2 (7.189)

The equation of motion for the second link is found from

L2 c3 × (−mg b3 ) = L2 c3 × ma (7.190)

where rg = L2 c3 in the notation of (7.95). The component of (7.190) in the c2 direction

is
³ 2
´
L2 mg sin θ2 = mL2 c3 × −L1 θ̈1 b1 + L2 θ̈2 c1 − L2 θ̇1 sin θ2 cos θ2 c1 · c2 (7.191)

which is simplified to
2
L2 mg sin θ2 = mL22 θ̈2 − mL1 L2 cos θ2 θ̈1 − mL22 θ̇1 sin θ2 cos θ2 (7.192)

We may then conclude as follows:

The dynamic model of the Furuta pendulum is

(J1 + mL21 + mL22 sin2 θ2 )θ̈1 − mL1 L2 cos θ2 θ̈2

2
= T − mL1 L2 θ̇2 sin θ2 − 2mL22 θ̇1 θ̇2 sin θ2 cos θ2 (7.193)
2
−mL1 L2 cos θ2 θ̈1 + mL22 θ̈2 = mL22 θ̇1 sin θ2 cos θ2 + mL2 g sin θ2 (7.194)
288 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

Example 128 The derivation of the dynamic model of the Furuta pendulum is quite
complicated, and it is important to check for errors in the model. This can be done by
investigating if the model satisfies the energy flow requirement that the time derivative
of the total energy is equal to the power θ̇1 T supplied by the motor torque T . The total
energy is the kinetic energy of the first rotational link, the kinetic energy of the mass,
and the potential energy due to gravity. This gives
1 2 1
V = J1 θ̇1 + mv · v + mgL2 cos θ2
2 2
1 2 1 ³ 2 2 2 2
´
= J1 θ̇1 + m L1 θ̇1 + L22 sin2 θ2 θ̇1 + L22 θ̇2 − 2L1 L2 θ̇1 θ̇2 cos θ2 + mgL2 cos θ2
2 2
1 2 1 2
= (J1 + mL21 + mL22 sin2 θ2 )θ̇1 + mL22 θ̇2
2 2
−mL1 L2 θ̇1 θ̇2 cos θ2 + mgL2 cos θ2 (7.195)

The time derivative for the solutions of the system is

h i
V̇ = θ̇1 (J1 + mL21 + mL22 sin2 θ2 )θ̈1 − mL1 L2 θ̈2 cos θ2
³ ´
+θ̇2 mL22 θ̈2 − mL1 L2 θ̈1 cos θ2
2 2
+mL1 L2 θ̇1 θ̇2 sin θ2 + mL22 θ̇1 θ̇2 sin θ2 cos θ2 − θ̇2 mgL2 sin θ2
³ 2
´
= θ̇1 T − mL1 L2 θ̇2 sin θ2 − 2mL22 θ̇1 θ̇2 sin θ2 cos θ2
³ 2
´
+θ̇2 mL22 θ̇1 sin θ2 cos θ2 + mL2 g sin θ2
2 2
+mL1 L2 θ̇1 θ̇2 sin θ2 + mL22 θ̇1 θ̇2 sin θ2 cos θ2 − θ̇2 mgL2 sin θ2
= θ̇1 T (7.196)

This shows that the model is consistent with the energy flow dynamics.

7.7 Principle of virtual work

7.7.1 Introduction
The equations of motion give the relation between the forces and torques acting on the
system and the resulting accelerations. There are two classes of forces that are important
in this connection: The active forces, which are also termed actuator forces, and the
forces of constraint. In the design of a control system we are mainly concerned with the
actuator forces, which can be command to achieve a specified motion. In contrast to this,
the main concern in a mechanical design will be the forces of constraint, which are forces
that ensure that the mechanical system is not damaged, and which ensure the system
does not break into parts. The following examples illustrate the contrast between the
two classes of forces:

• In the design of a robot control system we are interested in the motor torques
required for a desired acceleration. In the mechanical design of a robot it is different,
then it is important that the forces of constraint that appear in the bearings of the
joints are within acceptable limits so that the joint is not damaged. Note that as
long as the robot joints are intact, the forces of constraint are not relevant in the
control systems design.
7.7. PRINCIPLE OF VIRTUAL WORK 289

• In speed control of a train the control problem is to set up an engine force that
give a desired acceleration. The mechanical design problem is to ensure that the
tracks and the wheels can support the forces of constraint, which in this case are
the forces required to keep the train on the track.
• For a football player the motion control problem is to use the muscles of the leg
to set up active forces that result in a desired motion. For the knee the muscles
will provide the active forces that rotate the knee about its axis of rotation. The
forces of constraint will keep the knee join together so that is not damaged. As
long as the joint is strong enough, the football player need not be concerned about
the forces of constraint.

From this we get the idea that in the design of control systems we need not know the
forces of constraint to get the solutions we are seeking. It turns out that it may be quite
complicated to derive the forces of constraint, and therefore it seems to be attractive
to find a way to eliminate the forces of constraint from the equations of motion. The
principle of virtual work is a tool that allows us to do this, but first we have to introduce
generalized coordinates and the concept of virtual displacements.

7.7.2 Generalized coordinates

Consider N particles numbered by k = 1, . . . , N . Each particle is of mass mk and has
position
rk = xk i1 + yk i2 + zk i3 (7.197)
in an Newtonian coordinate frame i with orthogonal unit vectors i1 , i2 , i3 along the axes.
The position vectors rk define the configuration of the system. The resultant force on
(r)
each particle is Fk . Newton’s law for each particle is given by

d2 (r)
mk rk = Fk (7.198)
dt2
Note that all differentiations of vectors are done in the Newtonian frame i in this section.
Adding over all particles gives
N
X N
X (r)
d2
mk 2
rk = Fk (7.199)
dt
k=1 k=1

Suppose that there is an n-dimensional column vector q = (q1 , . . . , qn )T so that the

position rk of all particles are given as functions of q and t, that is,

rk = rk [q(t), t] (7.200)

Then the variables q1 , . . . , qn are called the generalized coordinates of the system. If
n is the minimum number of generalized coordinates that define the configuration of
the system, then q1 , . . . , qn will in addition be termed the minimal coordinates. The
n-dimensional space described by the generalized coordinates is called the configuration
space of the system.
The velocity of particle k can be expressed in terms of the generalized coordinates
according to
Xn
d ∂rk ∂rk
vk = rk = q̇i + (7.201)
dt i=1
∂qi ∂t
290 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

Example 129 For later use we note that (7.201) implies that partial differentiation of
vk with respect to q̇i gives
∂vk ∂rk
= (7.202)
∂ q̇i ∂qi
Moreover, we find that by interchanging derivation with respect to qi and t that

∂vk ∂ drk d ∂rk

= = (7.203)
∂qi ∂qi dt dt ∂qi

7.7.3 Virtual displacements

We now introduce the concept of virtual displacements which is very important in dy-
namics. The virtual displacement δrk of particle k is defined by
n
X ∂rk
δrk = δqi (7.204)
i=1
∂qi

where δqi is the virtual displacement in the generalized coordinate qi . If the time deriva-
tives q̇i of the generalized coordinates are independent, then the virtual displacements
δqi are linearly independent, and there are n independent virtual displacements δrk , and
the system is said to have n degrees of freedom.
If there is a linear constraint on the generalized velocities q̇i given by

A(q)q̇ = 0 (7.205)

then the virtual displacements of the generalized coordinates will satisfy

A(q)δq = 0 (7.206)

where δq = (δq1 , . . . , δqn )T . If the null-space of A(q) has dimension ndof ≤ n, which
means that there are ndof independent generalized velocities q̇i , then there are ndof
independent virtual displacements δrk and the system is said to have ndof degrees of
freedom.

7.7.4 d’Alembert’s principle

From the outset there are N particles, each with three coordinates, hence, if the particles
are moving independently of each other, then a system of N particles will have 3N
degrees of freedom. However, to satisfy constraints of the form rk = rk [q(t), t] where the
velocities q̇i are independent, the system will have only n degrees of freedom. To make
these constraints hold there must be certain forces acting on the particles. Such forces
can be characterized in a number of ways, but it turns out to appropriate to define forces
(c)
of constraints Fk that satisfy the principle of virtual work which is given by
N
X (c)
δrk · Fk =0 (7.207)
k=1

(c)
Here Fk is the force of constraint acting on particle k. Then the resultant force on
particle k is given by
(r) (c)
Fk = Fk + Fk (7.208)
7.7. PRINCIPLE OF VIRTUAL WORK 291

k
F  c
F k
Constraint
Actuators d System d
r
dt k
r
dt k
mechanism

Figure 7.9: A mechanical system is driven by actuators that set up active forces Fk acting
on the system. In addition there is a constraint mechanism that set up constraint forces
(c)
Fk that ensures that the system does not break in parts. It simplifies the equation of
motion greatly if the forces of constraint are formulated so that they do not influence on
the action of the active forces, which is achieved by the principle of virtual work. The
forces of constraint can then be eliminated from the equations of motion.

where Fk is the active force on particle k.

The principle of virtual work can now be used to eliminate the forces of constraint
(c)
Fk from the equation of motion. This is done by taking the scalar product between the
equation of motion for particle k and the virtual displacement δrk , and then summing
over all particles. This gives
N
X N
X N
X
d2 rk (c)
δrk · mk = δrk · Fk + δrk · Fk
dt2
k=1 k=1 k=1
N
X
= δrk · Fk (7.209)
k=1

and we arrive at the following formulation of the equation of motion

N
X µ ¶
d2 rk
δrk · mk 2 − Fk = 0 (7.210)
dt
k=1

which is called d’Alembert’s principle. Note that the only forces appearing in this for-
mulation are the externally applied forces Fk .
If we insert the expression for δrk from (7.204) and change the order of the summation,
we find that
Xn N
X µ ¶
∂rk d2 rk
δqi · mk 2 − Fk = 0 (7.211)
i=1
∂qi dt
k=1

If the virtual displacements δqi in the generalized coordinates are independent, then
d’Alembert’s principle can be written
N
X µ ¶
∂rk d2 rk
· mk 2 − Fk = 0 (7.212)
∂qi dt
k=1

Example 130 A train is running along a railway. The generalized coordinate of the
292 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

r
t
 c
F k

q
r

v
rq,t r
q
q i i

Figure 7.10: Train running along a track r (q, t) where the track has a velocity ∂r/∂t
due to the rotation of the earth. The virtual displacement δr is along the track, and the
force of constraint F(c) is perpedicular to the track in accordance with the principle of
virtual work.

train is the position coordinate along the railway track, which is denoted q. The position
of the train in a Newtonian coordinate frame is r (q, t), and the velocity of the train is
dr ∂r ∂r
v= = q̇ + (7.213)
dt ∂q ∂t
Here the first term on the right side is due to the motion along the track, and the other
term is due to the rotation of the earth. The virtual displacement of the train is defined
as
∂r
δr = δq (7.214)
∂q
which is a vector tangent to the track. The train is subjected to the resultant force
F (r) =F + F (c) where F (c) is the constraint force that keeps the train on the track, and
F is the motor and braking force that controls the velocity of the train. The principle of
virtual work then simply states that
δr · F (c) = 0 (7.215)

The physical interpretation of this is that F (c) is normal to the track because δr is tangent
to the track. This is illustrated in Figure 7.10.
Example 131 In coordinate form we have
δrk = Jk δq (7.216)
The principle of virtual work states that for all δq we have
N
X N
X
(c) (c)
0= δrTk Fk T
= δq JTk Fk (7.217)
k=1 k=1
7.8. PRINCIPLE OF VIRTUAL WORK FOR A RIGID BODY 293

(c)
This means that the constraint force Fk satisfies
N
X (c)
JTk Fk = 0 (7.218)
k=1

(c)
which shows that Fk is in the null-space of JT , while the active force Fk is the part of
(r) (c)
the resultant force Fk = Fk + Fk that is in the range space of Jk . Extensive treatment
of null-spaces and range space is found in (Strang 1988).

7.8 Principle of virtual work for a rigid body

7.8.1 Virtual displacements for a rigid body
The configuration of a rigid body b can be given by a rotation matrix Rib and a position
rc of the center of the mass. The velocity is given by the velocity vc of the center of
mass and the angular velocity ω ib of frame b relative to frame i. The forces and moments
acting on the rigid body are represented by a force Fbc with line of action through the
(a)
mass center, and a torque Tbc . The force Fbc can be split into an active force Fbc and
(c)
a constraint force Fbc . In the same way the Tbc can be described as a sum of an active
(a) (c)
torque Tbc and a constraint torque Tbc so that
(a) (c) (a) (c)
Fbc = Fbc + Fbc , Tbc = Tbc + Tbc (7.219)
(c) (c)
The constraint force Fb and the constraint torque Tb can be eliminated with the princi-
ple of virtual work. To do this we will have to define virtual displacements corresponding
to the velocity vc and the angular velocity ω ib .
It is assumed that the configuration of the body is described by n ≤ 6 generalized
coordinates qj . Then the velocity and angular velocity are given by
n
X
vc = vc,j q̇j + vt (7.220)
j=1
Xn
ω ib = ω ib,j q̇j + ω t (7.221)
j=1

where
∂rc ∂vc ∂ω ib
vc,j = = , ω ib,j = (7.222)
∂qj ∂ q̇j ∂ q̇j
Following the terminology of (Kane and Levinson 1985), vc,j is called partial velocity
j and ω ib,j is called partial angular velocity j. The virtual displacements may then be
defined by
n
X
δrc = vc,j δqj (7.223)
j=1
Xn
σ ib = ω ib,j δqj (7.224)
j=1
294 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

Example 132 If the configuration is given partly in terms of a rotation matrix, then it
may not be convenient to use generalized coordinates. In this case the use of generalized
velocities uj will serve our purpose. Then the velocity and angular velocity is given by
n
X
vc = vc,j uj + vt (7.225)
j=1
Xn
ω ib = ω ib,j uj + ω t (7.226)
j=1

where the partial velocities and the partial angular velocities are given by

∂vc ∂ω ib
vc,j = , ω ib,j = (7.227)
∂uj ∂uj

In this case the virtual displacements can be defined by

n
X
δrc = vc,j ζ j (7.228)
j=1
Xn
σ ib = ω ib,j ζ j (7.229)
j=1

where ζ j are independent variables.

Example 133 In a description using the velocity vo of a point o in the rigid body the
partial velocities vo,j = ∂vo /∂ q̇j will be needed in the virtual displacements
n
X
δro = vo,j δ q̇j (7.230)
j=1

of the point o. The virtual displacements in rotation are the same as when the velocity
of the center of mass is used.

7.8.2 Force and torque of constraint

The mechanical power of a mass force f dm acting on the point p in a rigid body b is
dPm = vp · df . The resulting power Pm from the mass forces on the body b is found by
integrating dPm over the body b using the expression vp = vc + ω ib × r. The result is
Z Z
Pm = vp · df = (vc + ω ib × r) · df = vc · F (r) + ω ib · Nb/c (7.231)
b b

where F (r) is the resultant force on the body b, and Nb/c is the moment on the body b
about the center of the mass. As usual we will represent F (r) and Nb/c by the equivalent
description where Fbc is a force vector equal to the resultant force and that has line of
action through the center of mass, and Tbc is a torque that is equal to the moment of the
forces about the center of mass. The power supplied to a rigid body is then given by

P = vc · Fbc + ω ib · Tbc (7.232)

7.9. MULTI-BODY DYNAMICS AND VIRTUAL WORK 295

(a) (c)
The force Fbc is given as a sum of an active force Fbc and a constraint force Fbc , and in
(a)
the same way the torque Tbc is the sum of an active torque Fbc and a constraint torque
(c)
Tbc . This is written
(a) (c) (a) (c)
Fbc = Fbc + Fbc , Tbc = Tbc + Tbc (7.233)
(c)
According to the principle of virtual work the force of constraint Fbc and the torque of
(c)
constraint Tbc on a rigid body satisfy
(c) (c)
δrc · Fbc + σ ib · Tbc = 0 (7.234)

where δrc and σ ib are the virtual displacements defined by (7.223 and (7.224).

7.9 Multi-body dynamics and virtual work

7.9.1 Introduction
The principle of virtual work can be used to derive the equations of motion for a wide
range of mechanical systems. The method is more systematic than a straightforward
application of the Newton-Euler equations as it provides a systematic procedure for elim-
inating the forces and torques of constraint. The physical interpretation of the method is
that the equations of motion are projected into the directions associated with the gener-
alized speeds of the system. This may give significant simplification in the derivation of
the equations of motion as it leaves out expressions pertaining to the directions associated
with the forces of constraint. In the following the principle of virtual work will be used
to derive the equations of motion for multi-body systems. This includes manipulators,
the ball and beam system, inverted pendulums and the Furuta pendulum.
The principle of virtual work in the derivation of equations of motion has a long
tradition in mechanics that stems back to d’Alembert and Euler. Recently the this
method has been treated extensively in (Kane et al. 1983) and (Kane and Levinson 1985),
and the method presented in this section is based on these references.

7.9.2 Equations of motion

In this section we will investigate the dynamics of a multi-body system in the form of nb
interconnected rigid bodies. The system has n generalized coordinates q1 , . . . , qn that may
be angles or translations. The generalized coordinates are assumed to be independent of
each other, which implies that the virtual displacements δqj are independent. Frame k
is assumed to be fixed in rigid body k of the system. Rigid body k has mass mk , inertia
dyadic Mk/c about the center of mass, and angular velocity ω 0k relative to the stationary
base frame 0. The center of mass in body k has velocity vck and acceleration ack .
The velocity and the angular velocity of body k are given in terms of the generalized
coordinates by
Xn n
X
vck = vck,j q̇j , ω 0k = ω 0k,j q̇j (7.235)
j=1 j=1

and the virtual displacements are then given by

n
X n
X
δrck = vck,j δqj , σ ok = ω 0k,j δqj (7.236)
j=1 j=1
296 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

where δqj are independent variables that can be selected arbitrarily.

The forces and moments acting on body k are represented by the equivalent descrip-
tion with a force Fkc acting through the center of mass, and a torque Tkc . The force and
the torque are split into an active part and a constraint part according to
(a) (c) (a) (c)
Fkc = Fbc + Fbc , Tkc = Tbc + Tbc (7.237)

The equations of motion for the rigid body can then be written
(a) (c)
mk ack − Fkc − Fkc = 0 (7.238)
(a) (c)
Mk/c · α0k + ω 0k × (Mk/c · ω 0k ) − Tkc − Tkc = 0 (7.239)

From the equations of motion it follows that

³ ´
(a) (c)
0 = δrck · mack − Fkc − Fkc
³ ´
(a) (c)
+σ 0k · Mk/c · α0k + ω 0k × (Mk/c · ω 0k ) − Tkc − Tkc (7.240)

for body k.
The power supplied to the multi-body system from the forces Fkc and torques Tkc is
n ³
X ´
P = vck · Fkc + ω 0k · Tkc (7.241)
k=1

where the result is obtained by summing up the power supplied to each rigid body.
According to the principle of virtual work the forces and torques of constraint satisfy
n ³
X ´
(c) (c)
δrck · Fkc + σ 0k · Tkc = 0 (7.242)
k=1

From this result it is seen that the forces and torques of constraint can be eliminated by
summing up the equations (7.240) for all k, which gives
n h
X ³ ´ ³ ´i
(a) (a)
δrck · mk ack − Fkc + σ 0k · Mk/c · α0k + ω 0k × (Mk/c · ω 0k ) − Tkc =0
k=1

Insertion of the expression for the virtual displacements δrck and σ 0k from (7.236) gives
n h
X ³ ´ ³ ´i
(a) (a)
vck,j · mk ack − Fkc + ω 0k,j · Mk/c · α0k + ω 0k × (Mk/c · ω 0k ) − Tkc δqj = 0
k=1

Note that this is a sum over the bodies numbered by k, while the index j of the generalized
coordinate is held constant. To complete the derivation we define the generalized forces
n
X (a) (a)
τj = vck,j · Fkc + ω 0k,j · Tkc (7.243)
k=1

and note that the variations δqj of the generalized coordinates are assumed to be arbi-
trary. Then the following result appears:
7.9. MULTI-BODY DYNAMICS AND VIRTUAL WORK 297

The equations of motion for a multi-body system with generalized coordinates qj can be
written
n h
X ³ ´i
vck,j · mk ack + ω 0k,j · Mk/c · α0k + ω 0k × (Mk/c · ω 0k ) = τ j (7.244)
k=1

where n
X (a) (a)
τj = vck,j · Fkc + ω 0k,j · Tkc (7.245)
k=1

This is Kane’s formulation of the equation of motion for a multi-body system.

7.9.3 Equations of motion about a point

In the same way as in the previous section the equations of motion can be found using a
description where the velocity of a point o is used. In this case the force Fbo is assumed
to have line of action through o, and the torque is Tbo = Tbc + rg × Fbc where rg is the
vector from o to c. The equations of motion are in this case given by

Fbo = mac (7.246)

Tbo = rg × mao + Mb/o · αib + ω ib × (Mb/o · ω ib ). (7.247)

The force and torque are written as the sum of the active and constraint part according
(a) (c) (a) (c)
to Fbo = Fbo + Fbo and Tbo = Tbo + Tbo . Again the constrain force and torque is
eliminated with the principle of virtual work, which gives
n h
X ³ ´ ³
(a)
δrok · mk ack − Fko + σ0k · rgk × mk aok + Mk/o · α0k
k=1
´i
(a)
+ω 0k × (Mk/o · ω 0k ) − Tko ) = 0 (7.248)

n
X (a) (a)
τj = vok,j · Fko + ω 0k,j · Tko (7.249)
k=1

The equations of motion for a multi-body system with generalized coordinates qj can be
written
n h
X ³
vok,j · mk aok + ω 0k,j · rgk × mk aok + Mk/o · α0k
k=1
´i
+ω 0k × (Mk/o · ω 0k ) = τj (7.250)

where n
X (a) (a)
τj = vok,j · Fko + ω 0k,j · Tko (7.251)
k=1

This is Kane’s formulation about a point o.

Example 134 A simple example is presented to illustrate the use of the equation of
motion in the form (7.250). It is clear that this formulation is overly complicated for
298 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

this system, however, it may be a worthwhile exercise to understand the concept of this
section. Consider a thin beam with homogeneous mass distribution and length L. The
end of the beam denoted o. The beam is driven by a motor that is attached at o and
rotates the beam by a angle θ and acts with a torque T = τ b3 on the beam. The velocity
of o is zero. A frame b is fixed to the beam so that b1 is along the length of the beam,
and b3 is the axis of rotation of the motor. A fixed frame i has axes that coincide with
the axes of b when θ = 0. The rotation matrix, the angular velocity and the angular
acceleration from i to b are

Rib = Rz,θ , ω ib = θ̇b3 , αib = θ̈b3 (7.252)

The position of the point o is ro = 0, and it follows that vo = 0 and ao = 0. The inertia
dyadic about o is
mL2 ³ ´
Mo = b2 b2 + b3 b3 (7.253)
3
The angle θ is selected as the generalized coordinate of the system. The associated general
force is τ . Then
vo,1 = 0, ω ib,1 = b3 (7.254)
The principle of virtual work gives
³ ´
0 = ω ib,1 · Mo · αib + ω ib × (Mo · ω ib ) − T (7.255)

and insertion of the relevant expressions gives

µ ¶
mL2 ³ ´ mL2 ³ ´
τ = b3 · b2 b2 + b3 b3 · θ̈b3 + b3 · θ̇b3 × b2 b2 + b3 b3 · θ̇b3
3 3
mL2
= θ̈ (7.256)
3
This leads to the equation of motion, which is simply

mL2
θ̈ = τ (7.257)
3

7.9.4 Ball and beam

In this section the equations of motion will be derived for the ball and beam shown in
Figure 7.5 with Kane’s equations (7.244). In this problem the force of constraint is the
contact force Fx b1 + Fy b2 from the beam on the ball. The normal component Fy will be
the contact force from the beam on the ball, while Fx is the friction force that makes the
ball roll without slipping. The magnitude of these two forces need not be known, and it
is convenient that the principle of virtual work cancel out these forces.
The generalized coordinates are q1 = θ and q2 = x. The angular velocities and the
velocity of the ball is given by

ω1 = θ̇b3 (7.258)
b3
ω2 = θ̇b3 + ẋ (7.259)
³ R ´
v2 = θ̇ Rb1 + xb2 + ẋb1 (7.260)
7.9. MULTI-BODY DYNAMICS AND VIRTUAL WORK 299

The acceleration of the center of mass of the ball is

³ 2
´ ³ 2
´
a2 = ẍ + θ̈R − θ̇ x b1 + θ̈x + 2θ̇ẋ + θ̇ R b2 (7.261)

The angular accelerations are

α1 = θ̈b3 (7.262)
µ ¶
ẍ
α2 = θ̈ + b3 (7.263)
R
The partial angular velocities and partial velocities are then found to be

v2,1 = Rb1 + xb2 , v2,2 = b1 (7.264)

b3
ω 1,1 = b3 , ω 2,1 = b3 , ω 2,2 = (7.265)
R
The inertia dyadics are
³ ´ ³ ´
M1/c = J1 b2 b2 + b3 b3 , M2/c = J2 b1 b1 + b2 b2 + b3 b3 (7.266)

The forces and torques acting on the beam and ball are represented by the forces and
torques

F1 = 0 (7.267)
F2 = Fx b1 + Fy b2 + m2 gn2 (7.268)
T1/c = T b3 + xb1 × (−F ) = T b3 − xFy b3 (7.269)
T2/c = Rb2 × F2 = −RFx b3 (7.270)

Here the constraint forces Fx and Fy should have been left out according to the procedure
of the virtual work, however, but we keep them in the derivation to see that they are
actually cancelled.
The equation of motion (7.244) gives

v2,1 · m2 a2 + ω 1,1 · M1/c · α1 + ω 2,1 · M2/c · α2 = v2,1 · F2 + ω 1,1 · T1/c + ω 2,1 · T2/c
v2,2 · m2 a2 + ω 2,2 · M2/c · α2 = v2,2 · F2 + ω 2,2 · T2/c

which give
³ ´ ³³ 2
´ ³ 2
´ ´
Rb1 + xb2 · m2 ẍ + θ̈R − θ̇ x b1 + θ̈x + 2θ̇ẋ + θ̇ R b2
µ ¶
ẍ
+b3 · J1 θ̈b3 + b3 · J2 θ̈ + b3
R
³ ´ ³ ´ ³ ´
= Rb1 + xb2 · Fx b1 + Fy b2 + m2 gn2 + b3 · T b3 − xFy b3 − b3 · RFx b3

and
³³ ´ ³ ´ ´ b µ ¶
2 2 3 ẍ
b1 · m2 ẍ + θ̈R − θ̇ x b1 + θ̈x + 2θ̇ẋ + θ̇ R b2 + · J2 θ̈ + b3
R R
³ ´ b ³ ´
3
= b1 · Fx b1 + Fy b2 + m2 gn2 + · −RFx b3
R
300 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

Note that at this point the forces of constraint Fx and Fy are eliminated from the
equations. Evaluation of the scalar products gives
³ 2
´ ³ 2
´
Rm2 ẍ + θ̈R − θ̇ x + xm2 θ̈x + 2θ̇ẋ + θ̇ R
µ ¶
ẍ
+J1 θ̈ + J2 θ̈ + = m2 g (R sin θ + x cos θ) + T
R

and ³ ´ J µ ¶
2 2 ẍ
m2 ẍ + θ̈R − θ̇ x + θ̈ + = m2 g sin θ
R R
By rearranging the equations we find that

The equations for a ball and beam system can be written

£ ¡ ¢¤ 1 ¡ ¢
J1 + J2 + m2 x2 + R2 θ̈ + J2 + m2 R2 ẍ + 2m2 xẋθ̇
R
= T + m2 g (R sin θ + x cos θ) (7.271)
1 ¡ ¢ 1 ¡ ¢ 2
J2 + m2 R2 θ̈ + 2 J2 + m2 R2 ẍ − m2 θ̇ x = m2 g sin θ (7.272)
R R

This model is found to be equivalent to the model (7.133, 7.134) derived with the
Newton-Euler method as (7.272) is equal to (7.134), and (7.271) is obtained by adding
7.134 to (7.133).

7.9.5 Single and double inverted pendulum

In this section the equations of motion will be derived using Kane’s equations (7.244)
for a single and double inverted pendulum as shown in Figures 7.6 and 7.7. First the
single pendulum will be treated. In this problem the force of constraint is the contact
force Fb between the pendulum and the cart. The velocity and acceleration of the cart
are denoted
vv = ẋn1 , av = ẍn1 (7.273)
while the velocity and the acceleration of the point mass at the end of the pendulum is

vb = ẋn1 + θ̇1 Lb b1 (7.274)

2
ab = ẍn1 + θ̈b Lb b1 + θ̇b Lb b2 (7.275)

The angular velocity and angular acceleration of the pendulum are

ω b = θ̇b b3 , αb = θ̈b b3 (7.276)

The force of the cart is F n1 , while the force on the pendulum is the gravity force mb gn2 .
The pendulum is assumed to be made with a massless rod and a point mass, and it
follows that the inertia dyadic is zero.
The nonzero partial velocities and partial angular velocities are

vv,x = n1 , vb,x = n1 (7.277)

vb,θb = Lb b1 , ω b,θb = b3 (7.278)
7.9. MULTI-BODY DYNAMICS AND VIRTUAL WORK 301

Kane’s equations (7.244, 7.245) give

³ 2
´
n1 · mv ẍn1 + n1 · mb ẍn1 + θ̈b Lb b1 + θ̇b Lb b2 = F (7.279)
³ 2
´
Lb b1 · mb ẍn1 + θ̈b Lb b1 + θ̇b Lb b2 = Lb b1 · gmb n2 (7.280)

where we have left out the force of constraint Fb that would have been cancelled anyway
if it had been included. By evaluating the scalar products, we find that the equations of
motion are given by
2
(mv + mb ) ẍ + mb Lb cos θb θ̈b − mb Lb sin θb θ̇b = F (7.281)
mb Lb cos θb ẍ + mb L2b θ̈b = mb Lb g sin θb (7.282)

which is the same result as (7.153, 7.154).

Next we add one more pendulum with point mass mc at the end of a rod of length
Lc . The velocity and acceleration of the point mass of the second pendulum are

vc = ẋn1 + θ̇1 Lc c1 (7.283)

2
ac = ẍn1 + θ̈c Lc c1 + θ̇c Lc c2 (7.284)

while the angular velocity and angular acceleration are

ω c = θ̇b b3 , αc = θ̈b b3 (7.285)

The gravity force on the second pendulum is mc gn2 . The nonzero partial velocities and
partial angular velocities are

vv,x = n1 , vb,x = n1 , vc,x = n1 (7.286)

vb,θb = Lb b1 , ω b,θb = b3 (7.287)
vc,θc = Lc c1 , ω c,θc = b3 (7.288)

Kane’s equations give

³ 2
´
n1 · mv ẍn1 + n1 · mb ẍn1 + θ̈b Lb b1 + θ̇b Lb b2
³ 2
´
+n1 · mc ẍn1 + θ̈c Lc c1 + θ̇c Lc c2 = F (7.289)
³ 2
´
Lb b1 · mb ẍn1 + θ̈b Lb b1 + θ̇b Lb b2 = Lb b1 · gmb n2 (7.290)
³ 2
´
Lc c1 · mc ẍn1 + θ̈c Lc c1 + θ̇c Lc c2 = Lc c1 · gmc n2 (7.291)

and the equations of motion are found by evaluation of the scalar products to be

(mv + mb + mc ) ẍ + mb Lb cos θb θ̈b + mc Lc cos θc θ̈c

2 2
−mb Lb sin θb θ̇b − mc Lc sin θc θ̇c = F (7.292)
mb Lb cos θb ẍ + mb L2b θ̈b = mb Lb g sin θb (7.293)
mc Lc cos θc ẍ + mc L2c θ̈c = mc Lc g sin θc (7.294)

which is in agreement with (7.169—7.171).

302 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

7.9.6 Furuta pendulum

In this section we show how Kane’s equations of motion can be used for the Furuta
pendulum. The equations of motion were derived using the Newton-Euler formulation
in Section 7.6.
With reference to Section 7.6 the angular velocities are

ω1 = θ̇1 b3 (7.295)
ω2 = θ̇1 b3 + θ̇2 c2 (7.296)

The accelerations are

ac1 = 0 (7.297)
³ 2
´
ac2 = −L1 θ̈1 b1 + L2 θ̈1 sin θ2 + 2L2 θ̇1 θ̇2 cos θ2 − L1 θ̇1 b2
³ 2
´ ³ 2 2
´
+ L2 θ̈2 − L2 θ̇1 sin θ2 cos θ2 c1 − L2 θ̇1 sin2 θ2 + L2 θ̇2 c3 (7.298)

while the partial velocities and partial angular velocities are

vc2,1 = −L1 b1 + L2 sin θ2 c2 , ω 1,1 = b3 , ω 2,1 = b3 (7.299)

vc2,2 = L2 c1 , ω 2,2 = c2 (7.300)

The inertia dyadics are M1/c = J1 (n2 n2 + n3 n3 ) and Mc2 = 0. Then Kane’s equations
of motion give
³ ´
vc2,1 · m2 ac2 + ω 1,1 · M1/c · α1 = τ1 (7.301)
vc2,2 · m2 ac2 = vc2,2 · (−m2 gn3 ) (7.302)

These two equations are identical to equations (7.187) and (7.190) that were used in the
development of the model with the Newton-Euler approach. The rest of the derivation
is therefore as in Section 7.6.

7.9.7 Planar two-link manipulator: Derivation 1

To demonstrate the use of Kane’s equations of motion we consider the planar two-link
manipulator in Figure 7.11. Frame 0 with orthogonal unit vectors i0 , j0 , k0 is a fixed
Newtonian frame with i0 along the x0 axis, j0 along the y0 axis, and k0 along the z0
axis. Frame 1 with orthogonal unit vectors i1 , j1 , k1 is fixed in link 1, and is obtained
from frame 0 by a rotation q1 about k0 . Frame 2 with orthogonal unit vectors i2 , j2 , k2 is
fixed in link 2, and is obtained by a rotation q2 about k1 . This means that k0 = k1 = k2 .
The direction cosines are given by

i0 · i1 = cos q1 , i0 · j1 = − sin q1 (7.303)

j0 · i1 = sin q1 , j0 · j1 = cos q1 (7.304)

and

i1 · j2 = cos q2 , i1 · j2 = − sin q2 (7.305)

i1 · j2 = sin q2 , i1 · j2 = cos q2 (7.306)
7.9. MULTI-BODY DYNAMICS AND VIRTUAL WORK 303

c2

 q2
b1

b2

c1

a2 
a1
q1
x

Figure 7.11: Planar two-link manipulator.

The angular velocity of the two links are

ω1 = q̇1 k1 (7.307)
ω2 = (q̇1 + q̇2 ) k2 (7.308)

and the angular acceleration is

α1 = q̈1 k1 (7.309)
α2 = (q̈1 + q̈2 ) k2 (7.310)

The position of the centers of mass are

rc1 = Lc1 i1 (7.311)

rc2 = L1 i1 + Lc2 i2 (7.312)

The velocity of the centers of mass are

vc1 = Lc1 q̇1 k1 × i1

= q̇1 Lc1 j1 (7.313)
vc2 = L1 q̇1 k1 × i1 + (q̇1 + q̇2 ) k2 × Lc2 i2
= q̇1 L1 j1 + (q̇1 + q̇2 ) Lc2 j2 (7.314)

and the acceleration of the centers of mass are

ac1 = q̈1 Lc1 j1 + q̇1 Lc1 ω 1 × j1

= q̈1 Lc1 j1 − q̇12 Lc1 i1 (7.315)
ac2 = q̈1 L1 j1 + (q̈1 + q̈2 ) Lc2 j2
+q̇1 L1 ω 1 × j1 + (q̇1 + q̇2 ) Lc2 ω 2 × j2
= q̈1 L1 j1 + (q̈1 + q̈2 ) Lc2 j2 − q̇12 L1 i1 − (q̇1 + q̇2 )2 Lc2 i2 (7.316)
304 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

This gives the following partial velocities and partial angular velocities:

vc1,1 = Lc1 j1 , vc2,1 = L1 j1 + Lc2 j2 , ω 1,1 = k1 , ω 2,1 = k2 (7.317)

vc1,2 = 0, vc2,2 = Lc2 j2 , ω 1,2 = 0 ω 2,2 = k2 (7.318)

The mass of link 1 is m1 , the mass of link 2 is m2 , and the inertia dyadics about the
centers of mass are

M1/c = I1x i1 i1 + I1y j1 j1 + I1z k1 k1 (7.319)

M2/c = I2x i2 i2 + I2y j2 j2 + I2z k2 k2 (7.320)

Then the equation of motion (7.244) gives

vc1,1 · m1 ac1 + vc2,1 · m2 ac2

³ ´
+ω 1,1 · M1/c · α1 + ω 1 × (M1/c · ω 1 )
³ ´
+ω 2,1 · M2/c · α2 + ω 2 × (M2/c · ω 2 ) = τ 1 (7.321)

³ ´
vc2,2 · m2 ac2 + ω 2,2 · M2/c · α2 + ω 2 × (M2/c · ω 2 ) = τ 2 (7.322)

This gives
³ ´
Lc1 j1 · m1 q̈1 Lc1 j1 − q̇12 Lc1 i1
³ ´ ³
+ L1 j1 + Lc2 j2 · m2 q̈1 L1 j1 + (q̈1 + q̈2 ) Lc2 j2
´
−q̇12 L1 i1 − (q̇1 + q̇2 )2 Lc2 i2

+k1 · I1z q̈1 k1 + k2 · I2z (q̈1 + q̈2 ) k2 = τ 1 (7.323)

³ ´
2
Lc2 j2 · m2 q̈1 L1 j1 + (q̈1 + q̈2 ) Lc2 j2 − q̇12 L1 i1 − (q̇1 + q̇2 ) Lc2 i2

+k2 · Ib3 (q̈1 + q̈2 ) k2 = τ 2 (7.324)

which leads to the following conclusion:

The equations of motions for a planar two-link manipulator are given by

£ ¡ ¢¤
I1z + I2z + m1 L2c1 + m2 L2c2 + L21 + 2L1 Lc2 cos q2 q̈1
£ ¡ ¢¤ £ ¤
+ I2z + m2 L2c2 + L1 Lc2 cos q2 q̈2 − m2 L1 Lc2 sin q2 2q̇1 q̇2 + q̇22 = τ 1(7.325)
£ ¡ ¢¤ ¡ ¢
I2z + m2 L2c2 + L1 Lc2 cos q2 q̈1 + I2z + m2 L2c2 q̈2 + m2 L1 Lc2 sin q2 q̇12 = τ 2 (7.326)

7.9.8 Planar two-link manipulator: Derivation 2

In this section the equations of motion for the planar two-link manipulator in Figure 7.11
will be derived about the origins of the coordinate frames. The direction cosines, the
7.9. MULTI-BODY DYNAMICS AND VIRTUAL WORK 305

angular velocities and the angular accelerations are as in the previous section. The
velocity of the origin of frame 1 and 2 are given by

vo1 = 0 (7.327)
vo2 = q̇1 L1 j1 (7.328)

This gives the following partial velocities and partial angular velocities:

vo1,1 = 0, vo2,1 = L1 j1 , ω 1,1 = k1 , ω 2,1 = k2 (7.329)

vo1,2 = 0, vo2,2 = 0, ω 1,2 = 0 ω 2,2 = k2 (7.330)

The accelerations are

ao1 = 0 (7.331)
ao2 = q̈1 L1 j1 + q̇1 L1 ω 1 × j1
= q̈1 L1 j1 − q̇12 L1 i1 (7.332)

The inertia dyadic about the origins are found from the parallel axes theorem (7.84) to
be
¡ ¢ ¡ ¢
M1/o = I1x i1 i1 + I1y + m1 L2c1 j1 j1 + I1z + m1 L2c1 k1 k1 (7.333)
¡ 2
¢ ¡ 2
¢
M2/o = I2x i2 i2 + I2y + m2 Lc2 j2 j2 + I2z + m2 Lc2 k2 k2 (7.334)

Then the equation of motion (7.250) gives

vo2,1 · m2 ac2
³ ´
+ω 1,1 · M1/o · α1 + ω 1 × (M1/o · ω 1 )
³ ´
+ω 2,1 · rg2 × mao2 + M2/o · α2 + ω 2 × (M2/o · ω 2 ) = τ 1 (7.335)
³ ´
ω 2,2 · rg2 × ma2 + M2/o · α2 + ω 2 × (M2/o · ω 2 ) = τ 2 (7.336)
which gives
³ ´
L1 j1 · m2 q̈1 L1 j1 + (q̈1 + q̈2 ) Lc2 j2 − q̇12 L1 i1 − (q̇1 + q̇2 )2 Lc2 i2
¡ ¢ ³ ´
+ I1z + m1 L2c1 q̈1 + k2 · Lc2 i2 × m2 q̈1 L1 j1 − q̇12 L1 i1
¡ ¢
+ I2z + m2 L2c2 (q̈1 + q̈2 ) = τ 1 (7.337)
³ ´ ¡ ¢
k2 · Lc2 i2 × m2 q̈1 L1 j1 − q̇12 L1 i1 + I2z + m2 L2c2 (q̈1 + q̈2 ) k2 = τ 2 (7.338)
and finally the equations of motions are found to be
¡ ¢
I1z + I2z + m1 L2c1 + m2 L2c2 + m2 L21 + 2m2 L1 Lc2 cos q2 q̈1
¡ ¢
+ I2z + m2 L2c2 + m2 L1 Lc2 cos q2 q̈2
£ ¤
−m2 L1 Lc2 sin q2 2q̇1 q̇2 + q̇22 = τ 1 (7.339)
¡ ¢
[I2z + m2 (L2c2 + L1 Lc2 cos q2 )]q̈1 + I2z + m2 L2c2 q̈2 + m2 L1 Lc2 sin q2 q̇12 = τ 2 (7.340)
This is the same result as the result of the previous section.
306 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

7.9.9 Kane’s computational scheme for two-link manipulator

In this section we will see that a more efficient derivation is possible by using the method
proposed in (Kane and Levinson 1983), where the equations of motion for a six-link
robot were derived. We will use the equation of motion in the form (7.244), but we use
a change of coordinates so that the generalized speeds are

u1 = q̇1 , u2 = q̇1 + q̇2 (7.341)

The corresponding generalized forces are

K1 = τ 1 − τ 2 , K2 = τ 2 (7.342)

We note that this implies that

K1 u1 + K2 u2 = (τ 1 − τ 2 ) q̇1 + τ 2 (q̇1 + q̇2 ) = τ 1 q̇1 + τ 2 q̇2 (7.343)

The equations of motion for the two link manipulator are then

K1 = vc1,u1 · m1 ac1 + vc2,1 · m2 ac2

³ ´
+ω 1,u1 · M1/c · α1 + ω 1 × (M1/c · ω 1 )
³ ´
+ω 2,u1 · M2/c · α2 + ω 2 × (M2/c · ω 2 ) (7.344)
K2 = vc1,u1 · m1 ac1 + vc2,u2 · m2 ac2
³ ´
+ω 1,u2 · M1/c · α1 + ω 1 × (M1/c · ω 1 )
³ ´
+ω 2,u2 · M2/c · α2 + ω 2 × (M2/c · ω 2 ) (7.345)

where the partial velocities are referenced to the generalized speeds.

The expressions for the angular velocities are

ω1 = u1 k1 (7.346)
ω2 = u2 k2 (7.347)

Following the method of (Kane and Levinson 1983) we introduce intermediate variables
Zi for simplifying the derivation. The velocities are written

vc1 = Z1 u1 j1 (7.348)
vc2 = Z2 u1 j1 + Z3 u2 j2 (7.349)

and

vc1 = Z4 j1 (7.350)
vc2 = Z5 j1 + Z6 j2 (7.351)

where the intermediate Zi variables are defined by

Z1 = Lc1 , Z2 = L1 , Z3 = Lc2 (7.352)

Z4 = u1 Z1 , Z5 = u1 Z2 , Z6 = u2 Z3 (7.353)
7.9. MULTI-BODY DYNAMICS AND VIRTUAL WORK 307

The nonzero partial velocities and partial angular velocities with respect to u1 and u2
are
vc1,u1 = Z1 j1 , vc2,u1 = Z2 j1 , ω 1,u1 = k1 (7.354)
vc2,u2 = Z3 j2 , ω 2,u2 = k2 (7.355)
The angular accelerations are
α1 = u̇1 k1 (7.356)
α2 = u̇2 k2 (7.357)
and the accelerations of the centers of mass are found to be
ac1 = Z1 u̇1 j1 + Z1 u1 ω 1 × j1
= Z1 u̇1 j1 − Z1 u21 i1 (7.358)
ac2 = Z2 u̇1 j1 + Z3 u̇2 j2 + Z2 u1 ω 1 × j1 + Z3 u2 ω 2 × j2
= Z2 u̇1 j1 + Z3 u̇2 j2 − Z2 u21 i1 − Z3 u22 i2 (7.359)
Then the equation of motion (7.344) and (7.345) are found to be
³ ´
K1 = Z1 j1 · m1 Z1 u̇1 j1 − Z1 u21 i1 (7.360)
³ ´ ³ ´
+ Z2 j1 + Z3 j2 · m2 Z2 u̇1 j1 + Z3 u̇2 j2 − Z2 u21 i1 − Z3 u22 i2 (7.361)

+k1 · Ia3 u̇1 k1 (7.362)

³ ´
K2 = Z3 j2 · m2 Z2 u̇1 j1 + Z3 u̇2 j2 − Z2 u21 i1 − Z3 u22 i2 (7.363)

+k2 · Ib3 u̇2 k2 (7.364)

This can be written
K1 = X11 u̇1 + X12 u̇2 + Z7 (7.365)
K2 = X21 u̇1 + X22 u̇2 + Z8 (7.366)
where the coefficients are given by
X11 = m1 Z12 + m2 Z22 + Ia3 (7.367)
X12 = X21 = m2 Z2 Z3 cos q1 (7.368)
X22 = m2 Z32 + Ib3 (7.369)
Z7 = −m2 Z2 Z3 sin q1 u22 (7.370)
Z8 = m2 Z2 Z3 sin q1 u21 (7.371)
To obtain the original variables qi and τ i we use that
q̇1 = u1 , q̇2 = u2 − u1 (7.372)
τ1 = K1 + K2 , τ 2 = K2 (7.373)
To have a model in a form suitable for simulation the model is written
u̇1 = Y11 (K1 − Z7 ) + Y12 (K2 − Z8 ) (7.374)
u̇2 = Y21 (K1 − Z7 ) + Y22 (K2 − Z8 ) (7.375)
308 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

where µ ¶ µ ¶−1
Y11 Y12 X11 X12
= (7.376)
Y21 Y22 X21 X22

Example 135 Insertion of the Zi and Xij constants gives the equations of motion on
the form
¡ ¢
m1 L2c1 + m2 L21 + I1z u̇1 + (m2 L1 Lc2 cos q1 ) u̇2
−m2 L1 Lc2 sin q1 u22 = K1 (7.377)
¡ ¢
m2 L1 Lc2 cos q1 u̇1 + m2 L2c2 + I2z u̇2 + m2 L1 Lc2 sin q1 u21 = K2 (7.378)

A change of variables by insertion of (7.341) and (7.342), and the addition of the second
equation to the first equation gives the equations of motion as given by (7.325) and
(7.326). Note that the equations of motion have a simpler form when the variables uj
and Kj are used.

7.9.10 Manipulator dynamics in coordinate form

In this section Kane’s formulation (7.244) of the equations of motion for a multi-body
system will be used to derive coordinate vector form of the equations of motion for a
manipulator. The manipulator has n joints that connect the n links of the manipulator.
The generalized coordinates in the case of rotational joints are the joint angles qj . It is
assumed that the joint angles are independent.
In coordinate vector form the equations of motion for link k are written
à ¡ ¢(a) ¡ k ¢(c) !
mk akck − Fkk − Fk
¡ ¢× ¡ ¢(a) ¡ k ¢(c) =0 (7.379)
Mkk/c αk0k + ω k0k Mkk/c ω k0k − Tkkc − Tkc

The velocity and angular velocity of link k are given by

µ k ¶
vck
= Jk (q) q̇. (7.380)
ω k0k

where Jk (q) is the Jacobian of link k. The virtual displacements of rigid body k are
given by µ ¶
δrk
= Jk (q) δq. (7.381)
σk
Note that the velocities q̇j are assumed to be independent, and this implies that the
virtual displacements δqj are arbitrary.
The principle of virtual work can be written
n µ
X ¶T µ ¶ Xn µ ¶
δrk (Fkk )(c) T T (Fkk )(c)
0= = δq Jk (7.382)
σk (Tkkc )(c) (Tkkc )(c)
k=1 k=1

The virtual displacements δq are independent, and it follows that

n
X µ ¶
(Fkk )(c)
JTk =0 (7.383)
(Tkkc )(c)
k=1
7.9. MULTI-BODY DYNAMICS AND VIRTUAL WORK 309

The input generalized forces are defined by

n
X µ ¶
Fkk
τ = JTk (7.384)
Tkkc
k=1

which are the joint torque in the case of rotary joints. Then the equation of motion can
be written à !
Xn
makck
T
Jk ¡ ¢× =τ (7.385)
Mkk/c αk0k + ω k0k Mkk/c ω k0k
k=1

From this equation it is possible to see that row j of the equations of motion is the sum
of the projections of the equations of motion for link k along column j of the Jacobian
Jk .

7.9.11 Spacecraft and manipulator

In this section we will derive the equations of motion for a spacecraft with a manipulator.
The spacecraft is described as link 0, and the manipulator links are denoted by link 1 to
link 6. The mass of link k is mk , the inertia dyadic of link k about the center of mass
of link k is Mk/c , the position is rck , the velocity is vck , and the acceleration is ack , the
angular velocity is with respect to an inertial frame i is ω ik , and the angular acceleration
is αik . The links of the manipulator are connected with rotational joints so that joint k
with joint angle qk connects link k − 1 and k. The motor torque applied at joint k is τ k .
The control forces and torques applied to the spacecraft are represented by a force F0c
with line of action through the center of mass, and a torque T0c .
The position of the center of mass of link k, k ≥ 1, is
k
X
rck = rc0 + dj − dk,kc (7.386)
j=1

where dj is the position of the origin of frame k, k ≥ 1, relative to the origin of frame 0,
and dk,kc is the position of the center of mass of link k relative to the origin of frame k.
The generalized speed vector and the generalized force vector are defined by

vc0 = u1 i0 + u2 j0 + u3 k0 (7.387)
ω ik = u4 i0 + u5 j0 + u6 k0 (7.388)
uk+6 = q̇k (7.389)

and the generalized forces are

F0c = F1 i0 + F2 j0 + F3 k0 (7.390)
T0c = F4 i0 + F5 j0 + F6 k0 (7.391)
Fk+6 = τk (7.392)

The partial velocities vck,j and partial angular velocities ω ik,j are then defined by
12
X 12
X
vck = vck,j uj , ω ik = ω ik,j uj (7.393)
k=0 k=0
310 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

y k1 yk

x k1 d k1,k
k xk
d k1,kc
d k,kc
k
z k1
zk  k1  k1

Figure 7.12: Link k.

The equation of motion is then given by

12 h
X ³ ´i
vck,j · mk ack + ω ik,j · Mk/c · αk + ω k × (Mk/c · ω k ) = Fj (7.394)
k=0

Note that this formulation allows us to develop the equations of motion without intro-
ducing generalized coordinates in the form of Euler angles for the spacecraft.

7.10 Recursive Newton-Euler

7.10.1 Inverse dynamics
The inverse dynamics problem for manipulators is the problem of computing τ given q,
q̇ and q̈. A computational algorithm for this is the recursive Newton-Euler scheme of
(Luh, Walker and Paul 1980). The main idea of the recursive Newton-Euler scheme is to
compute velocities, angular velocities, accelerations and angular accelerations recursively
from the base to the tip of the arm. Then the equations of motion are used for each
link to compute the required resultant forces and torques on each link, and finally the
contact forces and torques between the links are found by recursive computation from
the tip of the arm to the base. Finally, the motor torques are found by projecting the
contact torque onto the rotational axis of the joint.
Consider the link shown in Figure 7.12. The acceleration of the origin of Denavit-
Hartenberg frame k is denoted ak . The contact force and torque from link (k − 1) on
link k are denoted Fk−1,k and Tk−1,k , respectively. The distance from the origin of frame
(k − 1) to the origin of frame k is denoted dk−1,k . The distance from the origin in frame
(k − 1) to the mass center kc is denoted dk−1,kc , while dk,kc = dk−1,kc − dk−1,k is the
distance from the origin of frame k to kc . We note that dk−1,k , dk−1,kc and dk,kc are
constant vectors in frame k. We define z = (0, 0, 1)T as the unit vector in the z direction.
Note that in accordance with the Denavit-Hartenberg convention the rotational axis of
the joint between links k and (k + 1) is the z axis of frame k.
The forces and moments acting on link k are represented by the force Fkkc with
magnitude and direction equal to the resultant force and line of action through the
center of mass kc , and by the torque Tkkc which is equal in magnitude and direction to
the moment about the center of mass. The acceleration of frame 0 is set to be minus
the acceleration of gravity as gravity always enters in the expressions together with
acceleration as ak − g.
7.10. RECURSIVE NEWTON-EULER 311

Recursive Newton-Euler:
Initialization: ω 00 = 0, α00 = 0, a00 = − g0 where g0 is the acceleration of gravity.
Outwards recursion (k = 1 . . . 6):
³ ´
ω kk = Rkk−1 ω k−1 k−1 + zθ̇ k (7.395)
h ¡ k−1 ¢× i
αkk = Rkk−1 αk−1 k−1 + ω k−1 zθ̇k + zθ̈k (7.396)
akk = Rkk−1 ak−1
k−1
¡ k ¢× k ¡ ¢× ¡ k ¢× k
+ αk dk−1,k + ω kk ω k dk−1,k (7.397)

The equations of motion referred to the centers of mass (k = 1 . . . 6):

¡ ¢× ¡ ¢× ¡ k ¢× k
Fkkc = mk [akk + αkk dkk,kc + ω kk ω k dk,kc ] (7.398)
¡ k ¢× k
Tkkc k k
= Mk/c αk + ω k Mk/c ω k k
(7.399)

Inwards recursion (k = 5 . . . 0):

Fkk−1,k = Rkk+1 Fk+1 k

k,k+1 + Fkc (7.400)
tkk−1,k = Rkk+1 tk+1
k,k+1
¡ k ¢× ¡ ¢×
+ dk−1,k Rkk+1 Fk+1
k,k+1 + dkk−1,kc Fkkc + Tkkc (7.401)

Motor torques (k = 1 . . . 6):

τ k = zT Rkk−1 tkk−1,k (7.402)

If joint (k +1) is prismatic, then ω k = ω k−1 and ḋk−1 ˙

k,k = dk z, where dk is the Denavit-
Hartenberg parameter that specifies translation along the z axis. Then the outwards
recursion is

ω kk = Rkk−1 ω k−1
k−1 (7.403)
αkk = Rkk−1 αk−1
k−1 (7.404)
¡ k ¢× k
akk = Rkk−1 (ak−1 ¨
k−1 + zdk ) + αk dk−1,k
¡ k ¢× ¡ k ¢× k ¡ k ¢× k
+ ωk ωk dk−1,k + 2 ω k Rk−1 zd˙k (7.405)

while the motor force is

τ k = zT Rk−1
k Fkk−1,k (7.406)
The algorithm requires 117n − 24 multiplications and 103n − 21 additions for a ma-
nipulator with n rotational joints, which gives 678 multiplications and 597 additions for
a manipulator with six rotational joints.

7.10.2 Simulation
The recursive Newton-Euler scheme computes the required generalized forces τ (t) when
q(t), q̇(t) and q̈(t) are given. In contrast to this the simulation problem is to calculate the
state vector given by [q(t), q̇(t)] when the initial state [q(0), q̇(0)] and the generalized
forces τ (t) are given. This is done by numerical integration of the acceleration q̈(t).
Therefore, in the simulation of manipulator dynamics the acceleration q̈(t) must be
312 CHAPTER 7. NEWTON-EULER EQUATIONS OF MOTION

computed when [q(t), q̇(t)] and τ (t) are given. In the following a method based on the
use of the recursive Newton-Euler algorithm to establish the model in the Lagrangian
form. This is a convenient solution when the RNE algorithm is available.
The simulation problem involves the computation of q̈ given q, q̇ and τ . This can
be done using recursive Newton-Euler with the method of Walker and Orin (Walker and
Orin 1982), (Sciavicco and Siciliano 2000). The method is based on the fact that the
equation of motion for a manipulator can be written

M(q)q̈ + C(q, q̇)q̇ + g(q) = τ (7.407)

This result will be derived in Section 8.2.8.

We denote the recursive Newton-Euler as a function RNE (·) which takes q, q̇ and q̈
as inputs and outputs τ . This is written

τ = RNE(q, q̇, q̈) (7.408)

Simulation using the recursive Newton-Euler scheme:

¡ ¢
1. Compute column j of M(q) = m1 m2 . . . m6 (j = 1 . . . 6):

mj = RNE(q, 0, ej ) (7.409)

with 0 a0 = 0 where e1 = (1, 0, . . . , 0)T , e2 = (0, 1, . . . , 0)T , . . . , e6 = (0, 0, . . . , 1)T .

2. Compute
C(q, q̇)q̇ + g(q) = RNE(q, q̇, 0) (7.410)

3. Compute the accelerations from

q̈ = M−1 (q)(−C(q, q̇)q̇ − g(q) + τ ) (7.411)

We see that the RNE (·) function is used n + 1 times, and that a Gaussian elimination
is required to compute q̈. This means that the computational requirements for the
simulation is higher than for the inverse dynamics problem.
Chapter 8

Analytical mechanics

8.1 Introduction
The term analytical mechanics was introduced by Lagrange with his work Mécanique
Analytique which was published in 1788. In this work Lagrange emphasized the use of
algebraic operations in the derivation and analysis of equations of motion as opposed to
the earlier works of Newton and Euler which relied on vector operations. In our presen-
tation of analytical mechanics we will first explore Lagrangian dynamics, which is based
on the use of generalized coordinates, generalized forces and energy functions. Then we
will present a related formulation based on the Euler-Poincaré equation, where dynam-
ics on SO(3) and SE(3) can be described using energy functions without the reliance
on generalized coordinates. Finally the extended Hamilton’s principle and Hamilton’s
equations of motion will be presented. These methods are energy-based, and quite useful
as they provide a systematic way of deriving energy functions that are potential Lya-
punov function candidates. Moreover, Hamilton’s principle and Hamilton’s equations of
motion provide the basis for the Hamilton-Jacobi equation which is important in opti-
mal control theory. The material in this chapter is based on classical texts on dynamics
like (Goldstein 1980) and (Lovelock and Rund 1989), more recent text on dynamics
like (Arnold 1989) and (Marsden and Ratiu 1994), and robotics books like (Spong and
Vidyasagar 1989), (Sciavicco and Siciliano 2000) and (Murray et al. 1994). The results
that will be presented in this chapter are well established in the dynamics literature.
However, a control engineer will have to consult a great number of books, some of which
are quite advanced, to find the selection of analysis tools that will be presented here. Note
that although some of the material may seem to be abstract at a first reading, the meth-
ods are of great use in practical controller design and analysis, and in the development
of simulation systems.

8.2 Lagrangian dynamics

8.2.1 Introduction
The equations of motion for a mechanical system can be derived in the Newton-Euler
formulation, which is based on Newton’s second law in a vector formulation. It has been
documented in robotics that the Newton-Euler equations lead to an efficient formulation
suited for computations in real-time control and simulation (Luh et al. 1980). An alter-

313
314 CHAPTER 8. ANALYTICAL MECHANICS

native way of deriving the equations of motion is to use Lagrange’s formulation which is
based on algebraic operations on energy expressions using generalized coordinates and
generalized forces. Lagrange’s formulation may be better suited to derive results re-
lated to energy conservation and passivity, as it is based on the expressions for kinetic
and potential energy. This is becoming even more important in control theory as many
new controller designs are energy-based using Lyapunov designs or passivity (Slotine
1991), (Krstić, Kanellakopoulos and Kokotović 1995), (Khalil 1996), (Arimoto 1996),
(Sepulchre, Janković and Kokotović 1997), (Lozano et al. 2000). Well-known examples
in robotics is the independent-joint controller (Takegaki and Arimoto 1981), and the
adaptive tracking controller (Slotine and Li 1988), and related results have appeared in
other applications like attitude control (Wen and Kreutz-Delgado 1991) and vibration
damping (Kelkar and Joshi 1996). It is therefore of great interest to study Lagrange’s
equation of motion and related concepts of analytical dynamics for use in controller
design and analysis.

8.2.2 Lagrange’s equation of motion

Lagrange’s equations of motion for a mechanical system are equivalent to the Newton-
Euler equations of motion, although the methods derive the equations of motion in
two different ways. We have already presented Newton-Euler formulations, and we will
now show how to derive Lagrange’s equation of motion from d’Alembert’s principle as
presented in Section 7.7 for a system of particles (Goldstein 1980). We consider N
particles, where particle k has mass mk and position rk (q1 , . . . , qn , t), where q1 , . . . , qn
are the generalized coordinates of the system. The velocity of particle k is vk = drk /dt,
and the acceleration is ak = dvk /dt. Time differentiation and partial differentiation of
vectors are in a Newtonian frame in this section.
The starting point for our derivation of Lagrange’s equation of motion is d’Alembert’s
principle in the form (7.211)
"N #
Xn X ∂rk ³ ´
· mk ak − Fk δqi = 0 (8.1)
i=1
∂qi
k=1

To proceed we introduce the kinetic energy T of the system, which is

N
X 1
T = mk vk · vk (8.2)
2
k=1

We find that
ÃN ! N N
∂T ∂ X1 X ∂vk X ∂rk
= mk vk · vk = · mk vk = · mk vk (8.3)
∂ q̇i ∂ q̇i 2 ∂ q̇i ∂qi
k=1 k=1 k=1
ÃN ! N N
∂T ∂ X1 X ∂vk X d ∂rk
= mk vk · vk = · mk vk = · mk vk (8.4)
∂qi ∂qi 2 ∂qi dt ∂qi
k=1 k=1 k=1

where (7.202) and (7.203) are used. The following calculation can then be done:
XN µ ¶ X N µ ¶
d ∂T d ∂rk d ∂rk ∂rk
= · mk vk = · mk vk + · mk ak
dt ∂ q̇i dt ∂qi dt ∂qi ∂qi
k=1 k=1
N
X
∂T ∂rk
= + · mk ak (8.5)
∂qi ∂qi
k=1
8.2. LAGRANGIAN DYNAMICS 315

This result combined with (8.1) leads to

n
" µ ¶ N
#
X d ∂T ∂T X ∂rk
− − · Fk δqi = 0 (8.6)
i=1
dt ∂ q̇i ∂qi ∂qi
k=1

The third term in the bracket is defined to be the generalized force

N
X ∂rk
Qi := · Fk (8.7)
∂qi
k=1

associated with the generalized coordinate qi . This gives

Xn · µ ¶ ¸
d ∂T ∂T
− − Qi δqi = 0 (8.8)
i=1
dt ∂ q̇i ∂qi

Then, under the assumption that the time derivatives q̇i of the generalized coordinates
are independent, the virtual displacements δqi are arbitrary, and it follows that
µ ¶
d ∂T ∂T
− = Qi (8.9)
dt ∂ q̇i ∂qi
The generalized force Qi is assumed to be given by a conservative force −∂U/∂qi due to
a potential U = U (q) plus the generalized actuator force τ i . This is written
∂U
Qi = − + τi (8.10)
∂qi
Then the equation of motion becomes
µ ¶
d ∂T ∂T ∂U
− + = τi (8.11)
dt ∂ q̇i ∂qi ∂qi
From this result, Lagrange’s equation of motion is found:

Lagrange’s equation of motion is formulated using the Lagrangian

L (q, q̇, t) = T (q, q̇, t) − U (q) (8.12)

The equation of motion is µ ¶

d ∂L ∂L
− = τi (8.13)
dt ∂ q̇i ∂qi

Example 136 For use in the part on Hamiltonian dynamics we derive the following
result: Time differentiation of the Lagrangian gives
dL (q, q̇, t) ∂L (q, q̇, t) ∂L (q, q̇, t) ∂L (q, q̇, t)
= q̇ + q̈ +
dt ∂q ∂ q̇ ∂t
µ ¶
d ∂L (q, q̇, t) ∂L (q, q̇, t) ∂L (q, q̇, t)
= − τ T q̇ + q̈ + (8.14)
dt ∂ q̇ ∂ q̇ ∂t
where Lagrange’s equation of motion (8.13) has been inserted. This gives
µ ¶
dL (q, q̇, t) d ∂L (q, q̇, t) ∂L (q, q̇, t)
= q̇ + − τ T q̇ (8.15)
dt dt ∂ q̇ ∂t
316 CHAPTER 8. ANALYTICAL MECHANICS

l g
q

Figure 8.1: Mathematical pendulum.

8.2.3 Generalized coordinates and generalized forces

The power supplied to the system with rk = rk (q) from the forces Fk is
N N
à n ! n
ÃN !
X drk X X ∂rk X X ∂rk
· Fk = q̇i · Fk = · Fk q̇i
dt i=1
∂qi i=1
∂qi
k=1 k=1 k=1
Xn
= Qi q̇i (8.16)
i=1

This shows that the product Qi q̇i between the generalized force Qi and the generalized
speed q̇i has dimension power. This means that a system with two degrees of freedom
with q1 = x is a position and q2 = θ is an angle, then Q1 must be a force and Q2 must
be a torque.

8.2.4 Pendulum
A mathematical pendulum is a mass point of mass m in the gravity field which is con-
nected by a massless rod of length L to a frictionless joint with angle q. The pendulum
is shown in Figure 8.1. The kinetic energy is
1 1
T = mv 2 = m 2 q̇ 2 (8.17)
2 2
The potential energy is
U = mg (1 − cos q) (8.18)
The resulting Lagrangian is
1 2 2
L= m q̇ − mg (1 − cos q) (8.19)
2
and the equation of motion is
d ¡ 2 ¢
m q̇ + mg sin q = 0 (8.20)
dt
which gives r
g
q̈ + ω 20 sin q = 0, ω0 = (8.21)
8.2. LAGRANGIAN DYNAMICS 317

8.2.5 Mass-spring system

A mass-spring system with mass m and spring stiffness k will have Lagrangian
1 2 1 2
L=T −U = mq̇ − kq (8.22)
2 2
Lagrange’s equation of motion is then found to be
d
(mq̇) + kq = τ (8.23)
dt
which can be written in the familiar form
mq̈ + kq = τ (8.24)

8.2.6 Ball and beam

The ball and beam system presented in Section 7.4 has kinetic energy
µ ¶2 ·³ ´2 ³ ´2 ¸
1 2 1 ẋ 1
T = J1 θ̇ + J2 θ̇ + + m ẋ + θ̇R + θ̇x
2 2 R 2
µ ¶T µ ¡ 2 ¢ 1¡ ¢ ¶µ ¶
1 θ̇ J1 + J2¡ + m x +¢ R2 R J2 + mR2 θ̇
= 1 2 J2 (8.25)
2 ẋ R J2 + mR m+ R 2 ẋ
and potential energy
U = mg (R cos θ − x sin θ) (8.26)
The generalized coordinates are selected as
q1 = θ and q2 = x (8.27)
Then, with L = T − U , we have the following partial derivatives
µ ¶ h³ ´ ³ ´ i
∂L ẋ
= J1 θ̇ + J2 θ̇ + + m ẋ + θ̇R R + θ̇x x (8.28)
∂ θ̇ R
µ ¶ ³ ´
∂L ẋ 1
= J2 θ̇ + + m ẋ + θ̇R (8.29)
∂ ẋ R R
∂L
= mg (R sin θ + x cos θ) (8.30)
∂θ
∂L 2
= mθ̇ x + mg sin θ (8.31)
∂x
and the equations of motion can be written
£ ¡ ¢¤
J1 + J2 + m x2 + R2 θ̈
1 ¡ ¢
+ J2 + mR2 ẍ + 2mxẋθ̇ = mg (R sin θ + x cos θ) + τ (8.32)
R
µ ¶
1 ¡ ¢ J2 2
J2 + mR2 θ̈ + m + 2 ẍ − mθ̇ x = mg sin θ (8.33)
R R
We note that these equations of motion have the same form as (7.271, 7.272) which were
found using the formulation of Kane. We note that the matrix formulation
µ ¶ Ã ! µ ¶
θ̈ −2mxẋθ̇ + mg (R sin θ + x cos θ) τ
M = 2 + (8.34)
ẍ mθ̇ x + mg sin θ 0
318 CHAPTER 8. ANALYTICAL MECHANICS

has a positive definite and symmetric mass matrix

µ ¡ ¢ 1
¡ ¢ ¶
J1 + J2¡ + m x2 +¢ R2 R J2 + mR2
M= 1 2 J2 (8.35)
R J2 + mR m+ R 2

8.2.7 Furuta pendulum

The kinetic energy T and the potential energy U of the Furuta pendulum are given by
1 2 1 2
T = (J1 + mL21 + mL22 sin2 θ2 )θ̇1 + mL22 θ̇2 − mL1 L2 θ̇1 θ̇2 cos θ2 (8.36)
2 2
U = mgL2 cos θ2 (8.37)

which give the Lagrangian

1 2 1 2
L= (J1 + mL21 + mL22 sin2 θ2 )θ̇1 + mL22 θ̇2 − mL1 L2 θ̇1 θ̇2 cos θ2 − mgL2 cos θ2 (8.38)
2 2
The partial derivatives are
∂L
= (J1 + mL21 + mL22 sin2 θ2 )θ̇1 − mL1 L2 θ̇2 cos θ2 (8.39)
∂ θ̇1
∂L
= mL22 θ̇2 − mL1 L2 θ̇1 cos θ2 (8.40)
∂ θ̇2
∂L
= 0 (8.41)
∂θ1
∂L 2
= mL22 sin θ2 cos θ2 θ̇1 + mL1 L2 θ̇1 θ̇2 sin θ2 + mgL2 sin θ2 (8.42)
∂θ2
and the equations of motion are found by evaluation
µ ¶
d ∂L ∂L
− = (J1 + mL21 + mL22 sin2 θ2 )θ̈1 − mL1 L2 θ̈2 cos θ2
dt ∂ θ̇1 ∂θ1
2
+2mL22 θ̇1 θ̇2 sin θ2 cos θ2 + mL1 L2 θ̇2 sin θ2 (8.43)

µ ¶
d ∂L ∂L
− = mL22 θ̈2 − mL1 L2 θ̈1 cos θ2 + mL1 L2 θ̇1 θ̇2 sin θ2
dt ∂ θ̇2 ∂θ2
2
−mL22 sin θ2 cos θ2 θ̇1 − mL1 L2 θ̇1 θ̇2 sin θ2
−mgL2 sin θ2 (8.44)

The equations of motion of the Furuta pendulum are

(J1 + mL21 + mL22 sin2 θ2 )θ̈1 − mL1 L2 θ̈2 cos θ2

2
+2mL22 θ̇1 θ̇2 sin θ2 cos θ2 + mL1 L2 θ̇2 sin θ2 = τ (8.45)
2
mL22 θ̈2 − mL1 L2 θ̈1 cos θ2 − mL22 sin θ2 cos θ2 θ̇1 − mgL2 sin θ2 = 0 (8.46)

This result is in agreement with the result derived with the Newton-Euler approach.
The Lagrange derivation is much simpler for this system.
8.2. LAGRANGIAN DYNAMICS 319

8.2.8 Manipulator
In this section we will derive the Lagrangian equations of motion for a manipulator
(Spong and Vidyasagar 1989), (Sciavicco and Siciliano 2000). The manipulator has n
links which are rigid bodies. The links are assumed to be connected with rotary joints
of one degree of freedom. The joint angle of joint i is denoted qi . The joint angles are
the generalized coordinates of the manipulator. The vector of generalized coordinates is
T
denoted q = (q1 . . . qn ) . At each joint there is a motor torque τ i which are the input
T
generalized forces. The vector of generalized forces is denoted τ = (τ 1 . . . τ n ) .
The kinetic energy of link i is
1 i T i 1
Ti = mi (vci ) (vci ) + (ω i0i )T Mici ω i0i (8.47)
2 2
i
where mi is the mass, vci is the velocity of the center of mass, ω i0i is the angular velocity,
i i
and Mci is the inertia matrix around the center of mass. The velocity vci and the angular
i
velocity ω 0i are linear combinations of the time derivatives of the generalized coordinates,
and we may write
i
X
i i
vci = vci,j (q) q̇j = Jvci (q) q̇ (8.48)
j=1
i
X
ω i0i = ω i0i,j (q) q̇j = Jω0i (q) q̇ (8.49)
j=1

Then the kinetic energy of link i can be written

1 1
Ti = mi q̇T JTvci (q) Jvci (q) q̇+ q̇T JTω0i (q) Mici Jω0i (q) q̇ (8.50)
2 2
and the total kinetic energy for the manipulator is

1 T Xh i
n
T = q̇ mi JTvci (q) Jvci (q) +JTωi (q) Mici Jω0i (q) (8.51)
2 i=1

This shows that the kinetic energy of the manipulator can be written as the quadratic
form
1
T = q̇T M (q) q̇ (8.52)
2
where the n × n mass matrix M (q) given by
n
X £ ¤
M (q) = mi JTvci (q) Jvci (q) + JTω0i (q) Mici Jω0i (q) (8.53)
i=1

is symmetric. Moreover, the kinetic energy is nonnegative, which implies that M (q) is
positive definite. The potential energy is due to the gravity potential, and is written
n
X n
X
U (q) = Ui (q) = mi gT rci (q) (8.54)
i=1 i=1

The Lagrangian of the manipulator is therefore

1 T
L= q̇ M (q) q̇ − U (q) (8.55)
2
320 CHAPTER 8. ANALYTICAL MECHANICS

The derivation of the Lagrangian equation of motion is a relatively complicated exercise,

and we therefore state the main results first and present derivation afterwards.

The equations of motion for a manipulator can be written

M (q) q̈ + C (q, q̇) q̇ + g (q) = τ (8.56)

where M (q) = MT (q) is positive definite and g (q) is the gradient of the gravity poten-
tial. The matrix C (q, q̇) can be selected to be
( n )
X
C (q, q̇) = {ckj } = cijk q̇i (8.57)
i=1

where µ ¶
1 ∂mkj ∂mik ∂mij
cijk := + − (8.58)
2 ∂qi ∂qj ∂qk
are the Christoffel symbols of the first kind. In this case the matrix Ṁ − 2C is skew
symmetric.

To derive Lagrange’s equation of motion it is convenient to use the component form

n n
1 XX
T = mij (q) q̇i q̇j (8.59)
2 i=1 j=1

for the kinetic energy, which gives the Lagrangian

n n
1 XX
L= mij (q) q̇i q̇j − U (q) (8.60)
2 i=1 j=1

We find that
 
n n
d ∂L d 1 X 1X
= mkj q̇j + mik q̇i 
dt ∂ q̇k dt 2 j=1 2 i=1
Xn n n µ ¶
1 X X ∂mkj ∂mik
= mkj (q) q̈j + + q̇i q̇j (8.61)
j=1
2 i=1 j=1 ∂qi ∂qj

and that
n n
∂L 1 X X ∂mij ∂U
= q̇i q̇j − (8.62)
∂qk 2 i=1 j=1 ∂qk ∂qk

The resulting equation of motion is

n
X n X
X n
mkj (q) q̈j + cijk (q) q̇i q̇j + gk (q) = τ k (8.63)
j=1 i=1 j=1

where cijk are the Christoffel symbols of the first kind as defined by (8.58), and

∂U
gk := (8.64)
∂qk
8.2. LAGRANGIAN DYNAMICS 321

Then the equation of motion (8.56) appears by defining the matrix

n
X
C (q, q̇) = {ckj (q, q̇)} , ckj (q, q̇) = cijk (q) q̇i (8.65)
i=1

and the gravity vector

∂U T
g (q) = (8.66)
∂q
Finally, we will show that the matrix

N = Ṁ − 2C (8.67)

is skew symmetric. This is shown by considering element

nkj = ṁkj − 2ckj (8.68)

of the matrix. We find that

n µ
X ¶
∂mkj ∂mkj ∂mik ∂mij
ṁkj − 2ckj = − − + q̇i
i=1
∂qi ∂qi ∂qj ∂qk
n µ
X ∂mij ¶
∂mik
= − q̇i (8.69)
i=1
∂qk ∂qj

This implies
nkj = −njk (8.70)
which shows that N is skew symmetric.

8.2.9 Passivity of the manipulator dynamics

The time derivative of the energy E = T + U is found by the chain rule to be
µ ¶
d 1 T ∂U
Ė(q, q̇) = q̇ M (q) q̇ + q̇
dt 2 ∂q
1 ∂U
= q̇T M (q) q̈+ q̇T Ṁ (q) q̇ + q̇ (8.71)
2 ∂q

The time derivative along the solutions of the system is found by inserting the equation
of motion (8.56) and (8.66). This gives

1 T
Ė(q) = q̇T [−C (q, q̇) q̇ − g (q) + τ ] + q̇T Ṁ (q) q̇ + g (q) q̇
2
1 h i
= q̇T τ + q̇T T Ṁ (q) −2C (q, q̇) q̇ (8.72)
2

Finally, the skew symmetry of Ṁ − 2C gives the result

T
Ė(q) = q̇ τ (8.73)

The kinetic energy is always nonnegative.

322 CHAPTER 8. ANALYTICAL MECHANICS

If there is a constant Umin so that the potential energy is lower bounded according to
U ≥ Umin , then the storage function V = T + U − Umin ≥ 0 will have time derivative

V = q̇T τ (8.74)

along the solutions of the system. This implies that the manipulator dynamics (8.56)
with input τ and output q̇ is passive.

8.2.10 Example: Planar two-link manipulator 1

The planar manipulator from Section 7.9.7 has kinetic energy
1 1 1 1
T = m1 vc1 · vc1 + m2 vc2 · vc2 + ω 1 · M1/c · ω 1 + ω 2 · M2/c · ω 2 (8.75)
2 2 2 2
This can be written
1 1
T = m11 q̇12 + m12 q̇1 q̇2 + m22 q̇22 (8.76)
2 2
where
¡ ¢
m11 = I1z + I2z + m1 L2c1 + m2 L21 + L2c2 + 2L1 Lc2 cos q2 (8.77)
m12 = m21 = I2z + m2 L2c2 + m2 L1 Lc2 cos q2 (8.78)
m22 = I2z + m2 L2c2 (8.79)
are the elements of the inertia matrix. The potential energy is
U = (m1 gLc1 + m2 gL1 ) sin q1 + m2 gLc2 sin(q1 + q2 ) (8.80)
Then, from L = T − U the partial derivatives are found to be
∂L ∂T
= = m11 q̇1 + m12 q̇2 (8.81)
∂ q̇1 ∂ q̇1
∂L ∂T
= = m21 q̇1 + m22 q̇2 (8.82)
∂ q̇2 ∂ q̇2

∂L ∂T ∂U
= − = − (m1 Lc1 + m2 L1 ) g cos q1 − m2 Lc2 g cos(q1 + q2 ) (8.83)
∂q1 ∂q1 ∂q1
∂L ∂T ∂U 1 ∂m11 2 ∂m21
= − = q̇ + q̇1 q̇2 − m2 gLc2 g cos(q1 + q2 ) (8.84)
∂q2 ∂q2 ∂q2 2 ∂q2 1 ∂q2
The equations of motion are then found from (8.13) to be
µ ¶ µ ¶
∂m11 ∂m12 ∂U
m11 q̈1 + m12 q̈2 + q̇2 q̇1 + q̇2 q̇2 + = τ 1 (8.85)
∂q2 ∂q2 ∂q1
µ ¶
∂m21 1 ∂m11 2 ∂m21 ∂U
m21 q̈1 + m22 q̈2 + q̇2 q̇1 − q̇ − q̇1 q̇2 + = τ 2 (8.86)
∂q2 2 ∂q2 1 ∂q2 ∂q2
which gives the equations of motion in the form
¡ ¡ ¢¢
I1z + I2z + m1 L2c1 + m2 L21 + L2c2 + 2L1 Lc2 cos q2 q̈1
¡ ¢
+ I2z + m2 L2c2 + m2 L1 Lc2 cos q2 q̈2
¡ ¢
−m2 L1 Lc2 sin q2 2q̇1 q̇2 + q̇22
+ (m1 Lc1 + m2 L1 ) g cos q1 + m2 Lc2 g cos(q1 + q2 ) = τ 1 (8.87)
8.2. LAGRANGIAN DYNAMICS 323
¡ ¢ ¡ ¢
I2z + m2 L2c2 + m2 L1 Lc2 cos q2 q̈1 + I2z + m2 L2c2 q̈2
+m2 L1 Lc2 q̇12 sin q2 + m2 Lc2 g cos(q1 + q2 ) = τ 2 (8.88)

8.2.11 Example: Planar two-link manipulator 2

In this section we will see that the equations of motion will be simplified by introducing
a following change of generalized coordinates to
µ ¶ µ ¶µ ¶
φ1 1 0 q1
φ= = = Aq (8.89)
φ2 1 1 q2

with associated generalized forces

µ ¶ µ ¶µ ¶
K1 1 −1 τ1
K= = = A−T τ (8.90)
K2 0 1 τ2

as this gives
T
KT φ̇ = τ A−1 Aq̇ = τ T q̇ (8.91)
Note that φ̇1 = ω 1 and φ̇2 = ω 2 .
With the new set of generalized coordinates the kinetic energy is
1 T 1 T 1 T
T = q̇ M(q)q̇ = φ̇ A−T M(q)A−1 φ̇ = φ̇ D(φ)φ̇ (8.92)
2 2 2
where the mass matrix D(φ) = {dij (φ)} corresponding to the new coordinates φ is found
to be
µ ¶µ ¶µ ¶
1 −1 m11 m12 1 0
D(φ) = A−T M(q)A−1 =
0 1 m21 m22 −1 1
µ ¶
m11 − 2m12 + m22 m12 − m22
= (8.93)
m12 − m22 m22

which gives
µ ¶
I1z + m1 L2c1 + m2 L21 m2 L1 Lc2 cos (φ2 − φ1 )
D(φ) = (8.94)
m2 L1 Lc2 cos (φ2 − φ1 ) I2z + m2 L2c2

The equations of motion are then found from (8.13) to be

µ ¶ µ ¶
∂d12 ∂d12 ∂d12 ∂U
d11 φ̈1 + d22 φ̈2 + φ̇ + φ̇ φ̇2 − φ̇1 φ̇2 + = K1
∂φ 1 ∂φ2 2 ∂φ ∂φ1
µ 1 ¶ µ 1¶
∂d21 ∂d21 ∂d21 ∂U
d21 φ̈1 + d22 φ̈2 + φ̇ + φ̇ φ̇1 − φ̇1 φ̇2 + = K2
∂φ1 1 ∂φ2 2 ∂φ2 ∂φ2

to be
¡ ¢
I1z + m1 L2c1 + m2 L21 φ̈1 + m2 L1 Lc2 cos (φ2 − φ1 ) φ̈2
2
−m2 L1 Lc2 sin (φ2 − φ1 ) φ̇2 + (m1 Lc1 + m2 L1 ) g cos φ1 = K1 (8.95)
¡ ¢
m2 L1 Lc2 cos (φ2 − φ1 ) φ̈1 + I2z + m2 L2c2 φ̈2
2
+m2 L1 Lc2 sin (φ2 − φ1 ) φ̇1 + m2 Lc2 g cos(φ2 ) = K2 (8.96)
324 CHAPTER 8. ANALYTICAL MECHANICS

8.2.12 Limitations of Lagrange’s equation of motion

Lagrange’s equation of motion is based on the use of a set of generalized coordinates.
For many systems the use of generalized coordinates is convenient. Typically, this is the
case for robotic manipulators where the joint angles are suitable candidates for the use
as generalized coordinates. However, there are other systems which are more efficiently
described in terms of the rotation matrix and the angular velocity, and for such systems
the use of generalized coordinates may introduce complicated expressions.
To illustrate this we use the rotational dynamics of a rigid body as an example. The
kinetic energy is
1
T = ω T Mω (8.97)
2
where ω is the angular velocity in body-fixed coordinates, and M is the constant inertia
matrix in body coordinates. The configuration of the rotational dynamics is given by the
rotation matrix R. To derive Lagrange’s equation of motion for this system we have to
select a set of generalized coordinates. The usual set of generalized coordinates for this
T
system is the roll-pitch-yaw angles ψ, θ and φ, that is, q = (φ, θ, ψ) . Then the kinetic
energy is found to be
1
T = q̇T ETd (q)MEd (q)q̇ (8.98)
2
where  
1 0 − sin θ
Ed (q) =  0 cos φ sin φ cos θ  (8.99)
0 − sin φ cos φ cos θ
Lagrange’s equation of motion is
µ ¶
d ¡ T ¢ ∂ 1 T T
Ed (q)MEd (q)q̇ − q̇ Ed (q)MEd (q)q̇ = 0 (8.100)
dt ∂q 2
Here, we clearly see that the use of the generalized coordinate vector q has introduced
kinematic terms in the form of the matrix Ed (q) in the equation of motion. This causes an
unnecessary complication of the expressions, and moreover, the matrix Ed (q) is singular
for cos θ = 0, which introduces a singularity in the mathematical model which is due to
the mathematical representation. A great deal of patience is required to arrive at the
result
Mω̇ + ω × Mω = 0 (8.101)
which is straightforward to derive in the Newton-Euler formulation.
Still, it would be useful if there were some energy-based formulation that resembled
Lagrange’s equation, but where the definition of generalized coordinates was not required.
The form that we will use is the Euler-Poincaré equation of motion, but before we
can present it, we need some background material on the calculus of variations. This
includes a quite interesting development of the variation of the rotation matrix, and of
the homogeneous transformation matrix.

8.3 Calculus of variations

8.3.1 Introduction
In the following the concept of variations in dynamics is discussed, and standard results
on the variation of a function is presented. In addition, variational tools on SO(3) and
SE(3) are presented.
8.3. CALCULUS OF VARIATIONS 325

8.3.2 Variations versus differentials

We consider a continuous and differentiable function f (x1 , x2 , x3 ), and we would like to
investigate the extremal points of the function, which are the points (x1 , x2 , x3 ) where
f has its minima or maxima. This will be done by finding the stationary values of the
function f, which are the values f (x1 , x2 , x3 ) at points (x1 , x2 , x3 ) where the rate of
change of f is zero. The usual technique in calculus is to find the differential

∂f ∂f ∂f
df = dx1 + dx2 + dx3 (8.102)
∂x1 ∂x2 ∂x3
and locate the stationary points as the points where df = 0. In mechanical systems it
is customary to associate the differentials dxi with the actual infinitesimal change of
the variables xi , and in accordance with this df denotes the infinitesimal change in the
function f due to the infinitesimal change dxi . The time derivative of f is found by
dividing with dt, giving
df ∂f ∂f ∂f
= ẋ1 + ẋ2 + ẋ3 (8.103)
dt ∂x1 ∂x2 ∂x3
where ẋi = dxdt are the velocities.
i

As opposed to this, Lagrange introduced the concept variations or virtual changes

δxi in the variables xi (Lanczos 1986), (Lovelock and Rund 1989). The variation δxi
is to be considered as a mathematical experiment without any change in the physical
variable xi . This makes sense as it should be possible to decide if a point (x1 , x2 , x3 )
is an extremal point without moving the system around. The variation in f associated
with the variation δxi is

∂f ∂f ∂f
δf = δx1 + δx2 + δx3 (8.104)
∂x1 ∂x2 ∂x3
which reflects an infinitesimal change in f due to the mathematical experiment δxi
without any change being done in the variables xi . Stationary values are then found
when δf = 0, as would be expected. This distinction between the operator d and δ may
seem strange at first, but it turns out that this provides us with a number of very useful
techniques for analysis of mechanical systems. Prominent examples of such techniques
are d’Alembert’s principle and Hamilton’s principle. Extensive discussion on the concept
of variations in dynamics is found in (Lanczos 1986).

8.3.3 The variation of a function

In this section we will present a number of useful results on variations. We start with
the following definition:

Consider a continuous and differentiable function f (x). Define the perturbed function

f (x, α) = f (x) + αφ(x) (8.105)

where φ(x) is an arbitrary continuous and differentiable function. The variation of f at

x is then defined as ¯
df (x, α) ¯¯
δf (x) = = φ(x) (8.106)
dα ¯α=0
326 CHAPTER 8. ANALYTICAL MECHANICS

fx,  fx

f df
dx

Figure 8.2: The variation δf and the differential df of a function f (x).

The difference between df and δf (x), which are both infinitesimal changes in f , is clear
in this case as df is the infinitesimal change in the function f (x) due to an infinitesimal
change dx, while δf (x) is the infinitesimal change due to the infinitesimal changes in
f (x) to f (x, ) for the same x. This is shown in Figure 8.2.
Using the definition (8.106) it is straightforward to establish several results for the
variation of a function. The variation of the derivative of the function is found from the
derivative of f (x, α). Consider the function
f 0 (x, α) = f 0 (x) + αφ0 (x) (8.107)
0
where (·) denotes the derivative with respect to x. The variation is
¯
df 0 (x, α) ¯¯
δf 0 (x) = = φ0 (x) (8.108)
dα ¯α=0
We see that · ¸
d d
[δf (x)] = δ f (x) (8.109)
dx dx
that is, the derivative of the variation is equal to the variation of the derivative.
The variation of the definite integral
Z b
I= f (x)dx (8.110)
a

where a and b are constants is found from the function

Z b
I (α) = f (x, α)dx (8.111)
a

The variation is given by

¯ Z b ¯ Z b ¯
dI (α) ¯¯ d ¯ df (x, α) ¯¯
¯
δI = = f (x, α)dx¯ = dx
dα ¯α=0 dα a ¯ a dα ¯α=0
α=0
Z b
= δf (x)dx (8.112)
a

which shows that the variation of the integral is equal to the integral of the variation.
We sum up that
8.3. CALCULUS OF VARIATIONS 327

The variation operation commutes with differentiation and integration in the sense that
· ¸
d d
[δf (x)] = δ f (x) (8.113)
dx dx
Z b Z b
δ f (x, α)dx = δf (x)dx (8.114)
a a

8.3.4 The Euler-Lagrange equation for a general integral

To derive the Euler-Lagrange equation we consider the definite integral
Z b
I= f (y, y 0 , x)dx (8.115)
a

where
dy
y = y(x), y0 = (8.116)
dx
with boundary conditions
y(a) = ya , y(b) = yb (8.117)
At this point this is a purely mathematical exercise without any physical interpretation.
The integral I will depend on the curve that is defined by the function y = y(x). We
want to find the path where the integral has a stationary value with respect to curves
that result from infinitesimal changes in the function y(x). To this end we define the
function
y(x, α) = y(x) + αφ(x) (8.118)
where φ(x) is an arbitrary function that vanishes at the end-points, that is, φ(a) = φ(b) =
0. The variation in y is then defined as
¯
dy(x, α) ¯¯
δy(x) = = φ(x) (8.119)
dα ¯α=0

We note that derivation and variation commute, that is,

µ ¶
d dy
(δy) = δ (8.120)
dx dx

There is no point in introducing a variation in the variable x. Therefore δx = 0 will

always be used in this type of problem.
The variation of the definite integral is
Z b
δI = δ f (y, y 0 , x)dx
a
Z b ¯
d ¯
0 ¯
= f [y(x, α), y (x, α), x] dx¯
dα a ¯
α=0
Z b
= δf [y(x, α), y 0 (x, α), x] dx (8.121)
a
328 CHAPTER 8. ANALYTICAL MECHANICS

where
¯
d ¯
0
δf [y(x), y (x), x] = f [y(x, ), y (x, ), x]¯¯
0
dα α=0
∂f ∂f 0
= δy + 0 δy (8.122)
∂y ∂y
This gives
Z b µ ¶
∂f ∂f
δI = δy + 0 δy 0 dx (8.123)
a ∂y ∂y
Then the standard technique is to use partial integration of the second term in the
integrand. We see that
Z µ ¶ ¯b Z µ ¶ Z µ ¶
b
∂f 0 ∂f ¯ b
d ∂f b
d ∂f
δy dx = δy ¯¯ − δydx = − δydx (8.124)
a ∂y 0 ∂y 0 a a dx ∂y 0 a dx ∂y 0

because the variation vanishes at the end-points. This gives

Z b· µ ¶¸
∂f d ∂f
δI = − δydx (8.125)
a ∂y dx ∂y 0

Since δy(x) is arbitrary for all x, we see that δI = 0 implies the Euler-Lagrange equation
µ ¶
d ∂f ∂f
− =0 (8.126)
dx ∂y 0 ∂y

8.3.5 The variation of the rotation matrix

It is not immediately obvious how to define the variation of a rotation matrix. However,
we will see in this section that the definition can be formulated in analogy with the
definition (8.106) for a function. A rigid body has orientation given by R ∈ SO(3). The
time derivative of R is
Ṙ = Rω × (8.127)
where ω is the angular velocity in body coordinates. The differential of the rotation
matrix can be written
dR = Rω × dt (8.128)
The variation δR of the rotation matrix is found by considering the perturbation
(Marsden and Ratiu 1994, p. 390)

R(α) = R exp(ασ × ) (8.129)

where σ is an arbitrary three-dimensional vector in body coordinates, and σ × is the

corresponding skew-symmetric form. This means that R(α) is the composite rotation of
R and a rotation about the axis defined by σ. The variation in R is then defined as
¯
d × ¯
¯
δR = R exp(ασ )¯ (8.130)
dα α=0

which gives the result

δR = Rσ × (8.131)
8.3. CALCULUS OF VARIATIONS 329

We note that variation and time differentiation commutes as

· ¯ ¸ · ¸¯
d d d £ ¤¯ d d ¡ ¢ ¯
(δR) = R exp(ασ × ) ¯¯ = R exp(ασ × ) ¯¯ = δ Ṙ (8.132)
dt dt dα α=0 dα dt α=0

We will now derive the relation between the variation δω × of the angular velocity
ω = RT Ṙ and σ × = RT δR. First we note that since RT R = I we have
×

¡ ¢ ¡ ¢ ¡ ¢
0 = δI = δ RT R = δ RT R + RT δR ⇒δ RT = −RT δRRT (8.133)

In the same way we find that

d ¡ T¢ T
R = −RT ṘR (8.134)
dt
Consider
¡ ¢ T T T ×
δω × = δ RT Ṙ + R δ Ṙ = −R δRRT Ṙ + R δ Ṙ = −σ ω × +RT δ Ṙ

and
d × d ¡ T¢ d T
σ = R δR + RT (δR) = −RT ṘR δR + RT δ Ṙ = −ω × σ × +RT δ Ṙ
dt dt dt
This gives
d × × d ¡ ¢×
δω × = σ + ω× σ× − σ× ω = σ× + ω× σ (8.135)
dt dt
where (6.33) is used. The vector form of this is
d
δω = σ + ω×σ (8.136)
dt
To sum up:

The variation of the rotation matrix can be defined by

δR = Rσ × (8.137)

where σ × is the skew symmetric form of a vector σ in body coordinates. The variation
of the angular velocity in body coordinates is
d
δω = σ + ω×σ (8.138)
dt

Example 137 The time derivative of the rotation matrix R at the identity I is
¯
dR ¯¯
= ω × ∈ so(3) (8.139)
dt ¯R=I

The set so (3) of skew symmetric matrices is the Lie algebra of SO(3). Thus the matrix
forms ω × and σ × are both in the Lie algebra so (3). The Lie bracket in so (3) is (Arnold
1989), (Marsden and Ratiu 1994), (Murray et al. 1994)
£ × ×¤ × ¡ ¢×
ω , σ := ω × σ × − σ × ω = ω × σ (8.140)
330 CHAPTER 8. ANALYTICAL MECHANICS

The matrix adω is defined from

£ × ×¤ ×
ω , σ = (adω σ) (8.141)

From these two equations it is seen that

adω = ω × (8.142)

which gives
d
δω = σ + adω σ (8.143)
dt

8.3.6 The variation of the homogeneous transformation matrix

A rigid body has configuration given by
µ ¶
R r
T= ∈ SE(3) (8.144)
0 1

The time derivative of T can be written is the form

Ṫ = Tŵ (8.145)

where µ ¶
ω× v
ŵ = (8.146)
0 0
T
is the 4 × 4 matrix representation of the vector w = (v , ω T )T .
The variation of T is µ ¶
δR δr
δT = (8.147)
0 0
This can be written
δT = Tη̂ (8.148)
where µ ¶
σ× δr
η̂ = (8.149)
0 0
is the matrix form of the vector η= (δrT , σ T )T .
In analogy with the derivation for SO(3) we find that
¡ ¢ −1 −1 −1 −1
δ ŵ = δ T−1 Ṫ + T δ Ṫ = −T δTT−1 Ṫ + T δ Ṫ = −η̂ŵ + T δ Ṫ

and
d d ¡ −1 ¢ d −1 −1
η̂ = T δT + T−1 (δT) = −T−1 ṪT δT + T−1 δ Ṫ = −ŵη̂ + T δ Ṫ
dt dt dt
and we may conclude that the variation of the six-dimensional velocity vector w is related
to the time derivative of the variation η by

d
δ ŵ = η̂ + ŵη̂ − η̂ ŵ (8.150)
dt
8.4. THE ADJOINT FORMULATION 331

The last two terms are evaluated by the computation

µ × ¶µ × ¶ µ × ¶µ × ¶
ω v σ δr σ δr ω v
ŵη̂ − η̂ ŵ = −
0 0 0 0 0 0 0 0
µ × × × × × ×
¶
ω σ −σ ω ω δr − σ v
=
0 0
µ × ¶
× × ×
(ω σ) ω δr + v σ
=
0 0
\
= ad wη (8.151)

where adw is the matrix µ ¶

ω× v×
adw := (8.152)
0 ω×
The vector form corresponding to (8.150) is therefore

d
δw = η + adw η (8.153)
dt

The variation of the homogeneous transformation matrix can be defined by

δT = Tη̂ (8.154)

where the vector η is in body coordinates. Then the variation of the velocity w in body
coordinates is
d
δw = η + adw η (8.155)
dt

Remark 2 The time derivative of the homogeneous transformation matrix T at the

identity I is ¯
dT ¯¯
= ŵ ∈ se(3) (8.156)
dt ¯T=I
The set se (3) is the set of matrices of the form ŵ as defined in (8.146). The set se (3) is
the Lie algebra of SE(3). The matrix forms ŵ and η̂ are in the Lie algebra se (3). The
Lie bracket in se (3) is (Arnold 1989), (Marsden and Ratiu 1994), (Murray et al. 1994)

[ŵ, η̂] = ŵη̂ − η̂ ŵ (8.157)

The matrix adw is defined by

\
[ŵ, η̂] = adwη (8.158)
We see from (8.150) that this agrees with the expression (8.152) for adw .

8.4 The adjoint formulation

8.4.1 Introduction
In the previous section the ad operator was used in SO(3) and SE(3). To make the
presentation complete we include a brief presentation of the Ad and ad operators and
the Lie bracket on SO(3) and SE(3) in this section.
332 CHAPTER 8. ANALYTICAL MECHANICS

8.4.2 Rotations
The configuration in SO(3) is given by R ∈ SO(3). The time derivative is Ṙ = Rω × =
Ω× R where Ω = Rω. The skew-symmetric forms of ω and Ω are related by
Ω× = Rω × RT (8.159)
The mapping from ω to Ω can also be written in terms of vectors in the adjoint repre-
sentation
Ω = AdR ω (8.160)
where the matrix AdR is the adjoint transformation on SO(3). It is clear that
AdR = R (8.161)
Consider a vector v in the moving coordinate frame, and let V = Rv be the vector in
the fixed frame. Then
V× = Rv× RT , V = AdR v (8.162)
The time derivative is
µ ׶
dV× dv ×
= R RT + Ṙv RT + Rv× ṘT
dt dt
µ ׶
dv
= R RT + Rω × v× RT − Rv× ω × ṘT
dt
µ ׶
dv ¡ ¢
= R RT + R ω × v× − v× ω × ṘT
dt
µ ׶
dv £ ¤
= R RT + R ω × , v× ṘT (8.163)
dt
where £ × ×¤ ¡ ¢×
ω , v = ω × v× − v × ω × = ω × v (8.164)
which is the Lie bracket on SO(3). In vector form the equation (8.163) is written
dV dv
=R + Rω × v (8.165)
dt dt
However, time differentiation of V = AdR v gives
dV dv d (AdR )
= AdR + v (8.166)
dt dt dt
which implies
d (AdR )
= Rω × = AdR ω × (8.167)
dt
We now define the operator adω according to
d (AdR )
= AdR adω (8.168)
dt
and it follows that in SO(3) we have
adω = ω × (8.169)
This implies that £ ¤
(adω v)× = ω × , v× (8.170)
× ×
which means that adω v is a vector representation of the Lie bracket [ω , v ] . Intuitively
we can think of adω v as the rate of change in v due to the motion due to the angular
velocity ω, which is the directional derivative of v in the direction of ω.
8.4. THE ADJOINT FORMULATION 333

8.4.3 Rigid motion

The configuration in SE(3) is given by
µ ¶
R r
T= ∈ SE(3) (8.171)
0 1
with inverse µ ¶
−1 RT −RT r
T = ∈ SE(3) (8.172)
0 1
Consider the vector µ ¶
V + r× Ω
W= (8.173)
Ω
in inertial coordinates and µ ¶
v
w= (8.174)
ω
in body-fixed coordinates. The corresponding matrix forms in se(3) are given by
µ × ¶ µ × ¶
Ω V+r× Ω ω v
Ŵ = ∈ se(3), ŵ = ∈ se(3) (8.175)
0 0 0 0

The derivative of T is
µ ¶ µ ¶µ ¶
Rω × Rv R r ω× v
Ṫ = = = Tŵ (8.176)
0 0 0 1 0 0
or µ ¶ µ ¶µ ¶
Ω× R v Ω× V+r× Ω R r
Ṫ = = = ŴT (8.177)
0 0 0 0 0 1
We see that
Ŵ = TŵT−1 (8.178)
In vector form we get
W = AdT w (8.179)
where the adjoint transformation on SE(3) is given by
µ ¶
R r× R
AdT = (8.180)
0 R
with inverse µ ¶
−1 RT −RT r×
(AdT ) = (8.181)
0 RT
The time derivative of AdT is
µ × ¶
d Rω ×
Rv× + r Rω ×
(AdT ) =
dt Rω ×0
µ ¶ µ ¶
R r× R ω × v×
=
0 R 0 ω×
= AdT adw (8.182)

where adw is defined by

d
(AdT ) = AdT adw (8.183)
dt
334 CHAPTER 8. ANALYTICAL MECHANICS

We see that µ ¶
ω× v×
adw = (8.184)
0 ω×
Let w be an a six dimensional vector, and let W = AdT w. Then

Ŵ = TŵT−1 (8.185)

and time differentiation gives

dŴ dŵ −1 −1 −1
= Ṫ T + ṪŵT + TŵṪ
dt dt
dŵ −1
= Ṫ T + TûŵT−1 − TŵûT−1
dt
dŵ −1
= Ṫ T + T (ûŵ − ŵû) T−1
dt
dŵ −1
= Ṫ T + T [û, ŵ] T−1 (8.186)
dt
where
[û, ŵ] = ûŵ − ŵû (8.187)
is the Lie bracket in se(3). In adjoint form the time derivative is
µ ¶
dW dW
= AdT + adu w (8.188)
dt dt

Comparing the two expressions, we find that

\
adu w= [û, ŵ] (8.189)

As in SO(3) we see that adu w is a vector form of the Lie bracket [û, ŵ] . Much in the
same way as in SO(3) the intuitive interpretation of adu w is that it is the rate of change
in w due to the motion induced by the velocity vector u.
¡ ¢T
Remark 3 The physical interpretation of U = AdT u where u = v ω is not quite
straightforward. However, this is not a problem as this vector, which is called the spatial
¡ ¢T
velocity vector, is not widely used. We see from U = Rv−Ω× r Rω that the
velocity Rv−Ω× r is the velocity of a point of the rigid body which is at the origin of
the fixed frame. The angular velocity is simply the angular velocity of the rigid body in
the coordinates of the fixed frame.

8.5 The Euler-Poincaré equation

8.5.1 A central equation
Consider a rigid body b where the mass element dm has position r in inertial coordinates.
The externally applied force on mass element dm is df , and the force of constraint on
the mass element is denoted df (c) . Newton’s law for a mass element dm is

r̈dm − df −df (c) = 0 (8.190)

8.5. THE EULER-POINCARÉ EQUATION 335

The virtual displacement of the mass element dm is denoted δr. We take the scalar
product between the virtual displacement δr and Newton’s law, and integrate the result
over the rigid body. This gives the following equation of motion:
Z ³ ´T
r̈dm − df −df (c) δr = 0 (8.191)
b

The total kinetic energy of body b is

Z
1
T = ṙT dmṙ (8.192)
2 b

We note that the variation of the kinetic energy expressed in terms of ṙ is

Z
δT = ṙT dmδ ṙ (8.193)
b

Moreover, we define the virtual work

Z
Wδ := df T δr (8.194)
b

To simplify the expression the product rule for differentiation is used to arrive at
Z Z Z Z
d d
r̈T dmδr = ṙT dmδr− ṙT dmδ ṙ = ṙT dmδr−δT (8.195)
b dt b b dt b
This results in the equation of motion in the form
Z Z ³ ´T
d
ṙT dmδr − δT − Wδ − df (c) δr = 0 (8.196)
dt b b

where the forces of constraint df (c) still appears.

Next a change of variables is introduced. The main difference from the usual La-
grange formulation is that we do not necessarily use generalized coordinates. Instead a
generalized speed vector u is introduced, where the velocity ṙ is affine in the components
of a vector u. The motivation for this is that this allows us to work with the rotation
matrices and the angular velocity in SO(3) and SE(3). The velocity is expressed by
∂ ṙ ∂r
ṙ = u+ (8.197)
∂u ∂t
We define the variation associated with u to be ξ, so that the virtual displacement δr is
given by
∂ ṙ
δr = ξ (8.198)
∂u
The first term on the left side of (8.196) can then be written
Z Z µ ¶
d d ∂ ṙT dmṙ ∂ ṙ
ṙT dmδr = ξ
dt b dt b ∂ ṙ 2 ∂u
· ¸
d ∂T
= ξ (8.199)
dt ∂u
This leads to the following result:
336 CHAPTER 8. ANALYTICAL MECHANICS

The equation of motion can be written

µ ¶ Z ³ ´T
d ∂T
ξ − δT − Wδ − df (c) δr = 0 (8.200)
dt ∂u b

R ¡ ¢T
Suppose that T = T (u), Wδ = τ T ξ and that b df (c) δr = 0. Then δT = (∂T /∂u) δu,
and the equation of motion is found to be
µ ¶
d ∂T ∂T ³ ´
ξ− δu − ξ̇ − τ T ξ = 0 (8.201)
dt ∂u ∂u

The equation of motion (8.200) was presented in (Bremer 1988) where it was termed
a central equation as it forms a basis from which related results like Lagrange’s equation
of motion, Hamel-Boltzmann’s equation and the Euler-Poincaré equation can be derived
with a reasonable effort. In the following we will use the equation of motion in the form
(8.201) to derive the Euler-Poincaré equation in SO(3) and SE(3).
Example 138 It is noted that if generalized coordinates are available so that u = q̇,
then the usual equations
∂r ∂r ∂r
ṙ = q̇+ , δr = δq (8.202)
∂q ∂t ∂q
are recovered in place of (8.197) and (8.198). Moreover, if T = T (q, q̇), Wδ = τ T ξ and
R ¡ (c) ¢T
b
df δr = 0, then ξ = δq and
∂T ∂T
δT = δq+ δ q̇ (8.203)
∂q ∂ q̇
This gives the familiar result
· µ ¶ ¸T
d ∂T ∂T
− −τ δq = 0 (8.204)
dt ∂ q̇ ∂q
which gives Lagrange’s equation of motion if the elements of δq are independent.

8.5.2 Rotating rigid body

A rotating rigid body b has configuration R and angular velocity ω in the body-fixed b
frame. The generalized speed is taken to be u = ω where ω × = RT Ṙ. The corresponding
variation vector is σ × = RT δR. The kinetic energy is
1 T
T = ω Mω (8.205)
2
where M is constant, positive definite and symmetric. Moreover, suppose that the gen-
eralized forces are denoted τ , so that the virtual work is Wδ = τ T σ. Then
∂T
= ωT M (8.206)
∂ω
and the equation of motion can then be found from (8.201) to be
µ ¶
d ∂T ∂T
σ− (δω − σ̇) − τ T σ = 0 (8.207)
dt ∂ω ∂ω
8.5. THE EULER-POINCARÉ EQUATION 337

Application of (8.138) gives

µ µ ¶ ¶
d ∂T ∂T ×
− ω − τT σ=0 (8.208)
dt ∂ω ∂ω

If the rigid body undergoes a free rotation, then the components of σ are independent.
This leads to:

The Euler-Poincaré equation for a rotating rigid body is

µ ¶T µ ¶T
d ∂T × ∂T
+ω =τ (8.209)
dt ∂ω ∂ω

Insertion of (8.206) gives the familiar equation

Mω̇ + ω × (Mω) = τ (8.210)

8.5.3 Free-floating rigid body

A free-floating rigid body b has configuration given by the homogeneous transformation
matrix µ ¶
R rp
T= ∈ SE(3) (8.211)
0 1
where R is the rotation matrix and rp is the position of some fixed point p in b. The
generalized speed is selected to be
µ b ¶
vp
u=w= (8.212)
ωb

which is given in the body-fixed frame b, and which has a 4 × 4 matrix form ŵ = T−1 Ṫ.
The associated variation vector is η which is defined by its 4 × 4 matrix form η̂= T−1 δT
To find an expression for the kinetic energy we need to find expressions for

∂ ṙ ∂ ṙ
ṙ = w, δr = η (8.213)
∂w ∂w
¡ ¢
This is found by observing that the velocity ṙ of the mass element is ṙ = R vpb + ω b× rbpm
where r = rp + rpm and ṙp = Rvbp . This gives

∂ ṙ ³ ´
= R I ... −rb× (8.214)
∂w pq

The kinetic energy is then found to be

Z
1
T = ṙT dmṙ (8.215)
2 b
Z µ ¶ ³ ´
1 I
= wT b× RT R I ... −rb× wdm (8.216)
2 b rpq pq

1 T b
= w Dp w (8.217)
2
338 CHAPTER 8. ANALYTICAL MECHANICS

where
µ R b× ¶ µ ¶
R mI −R r dm mI mrb×
Dbp = b×
b pq
b× b× = b×
g
(8.218)
r
b pq
dm − r r
b pq pq
dm −mrg Mbp

For a free-floating rigid body the principle of virtual work states that the forces of con-
straint does no virtual work, that is,
Z ³ ´T
df (c) δr = 0 (8.219)
b

The virtual work is Z µ ¶T

T Fb
Wδ = df δr = η (8.220)
b Lbp
where Fb is the total force on the rigid body with line of action through the point P ,
and Lbp is the total torque. The vectors Fb and Lbp are given in the frame b.
In this setting, the kinetic energy will be a function of velocity and angular velocity,
which is written T = T (w). The equation of motion is then found from (8.201) to be
µ ¶ µ ¶T
d ∂T ∂T Fb
η− (δw − η̇) − η=0 (8.221)
dt ∂w ∂w Lbp

Then, by noting from (8.153) that δw − η̇ = adw η we find that

" µ ¶ µ b ¶T #
d ∂T ∂T F
− adw − η=0 (8.222)
dt ∂w ∂w Lbp

Since η is arbitrary this gives

µ ¶T µ ¶T µ ¶
d ∂T ∂T Fb
− adTw = (8.223)
dt ∂w ∂w Lbp

and using (8.152) we find that

µ ¶T µ ¶µ ¶T µ ¶
d ∂T ω× 0 ∂T Fb
+ = (8.224)
dt ∂w v× ω× ∂w Lbp

This equation can be expanded to Kirchhoff’s equation of motion:

Euler-Poincaré’s equation for a free-floating rigid body gives Kirchhoff’s equations of

motion
µ ¶T µ ¶T
d ∂T × ∂T
+ω = Fb (8.225)
dt ∂v ∂v
µ ¶T µ ¶T µ ¶T
d ∂T ∂T ∂T
+ v× + ω× = Lbp (8.226)
dt ∂w ∂v ∂ω

Kirchhoff’s equations of motion are important in the modeling of ship motion, where
also the added inertia effects may be represented in this setting (Lamb 1945), (Sagatun
and Fossen 1991), (Leonard 1997), (Fossen 2002).
8.5. THE EULER-POINCARÉ EQUATION 339

Example 139 If it is been assumed that P is the center of mass of the rigid body, then
µ ¶T µ ¶
∂T mvcb
= (8.227)
∂w Mbc ω bc
and the equations of motion are found to be
mv̇b + ω b× mvcb = Fb (8.228)
Mbc ω̇ b + ω b× Mbc ω b = Lbc (8.229)

8.5.4 Mechanism with n degrees of freedom

We will now study a mechanism with n degrees of freedom. The velocity vectors of the
rigid body k are then
à k !
∂vkp
∂wk ∂wk ∂wk ∂u
wk = u, η k = η, = ∂ω k
(8.230)
∂u ∂u ∂u k
∂u
T
where u = (u1 , . . . , un ) is the vector of generalized velocities. Typically, for a robot
arm we will have u = q̇ where q is the n-dimensional vector of generalized coordinates.
In this case the principle of virtual work is used to eliminate the forces of constraint.
The principle of virtual work states that the total virtual work of the constraint forces is
zero. Therefore we need to sum up the virtual work done by the constraint forces for the
whole system to eliminate the constraint forces from the equation of motion. Note that
the constraint forces includes two types of constraint forces: Internal constraint forces in
each body which makes the body rigid, and interconnecting constraint forces that hold
the mechanism together.
The principle of virtual work for a mechanism with k interconnected rigid bodies can
be written n Z ³
X ´T
df (c) δr = 0 (8.231)
k=1 bk

where bk denotes body k. Therefore the forces of constraint can be eliminated by summing
up the equations of motions in the form (8.200). This gives
Xn · µ ¶ Z ³ ´T ¸
d ∂Tk
ηk − δTk − Wkδ − df (c) δr = 0 (8.232)
dt ∂wk b
k=1

we are able to eliminate the constraint forces, and get

Xn · µ ¶ ¸
d ∂Tk
η k − δTk − Wkδ = 0 (8.233)
dt ∂wk
k=1

If we proceed as for the free-floating rigid body, but keep the virtual displacement η k in
the expression we arrive at

Xn · µ ¶¸
T
k
mk v̇kc + ω k×
k mk vkck
− Fkk
0 = ηk
Mkk ω̇ kk + ω k× k k
k Mk ω k − Lk
k
k=1
 Ã !T µ 
Xn k
∂vkp k k× k k
¶
η T ∂u
m v̇
k kc + ω k m v
k kc − Fk 
= (8.234)
∂ω k
k Mk ω̇ kk + ω k× k k
k Mk ω k − Lkc
k=1 ∂u
340 CHAPTER 8. ANALYTICAL MECHANICS

Here η is arbitrary, and it follows that

n
"µ ¶T
X k
∂vkc ¡ k ¢
mv̇kc + ω × k
k mvkc − Fk
k
∂u
k=1
µ k ¶T ³ #
∂ω k ´
k k× k k
+ Mkc ω̇ k + ω k Mkc ω k − Lkc = 0 (8.235)
∂u

We note that Fkk is the applied force to body k, and Lkc is the applied torque to body
k around point center of mass of body k, and that the forces of constraint has been
eliminated in the derivation. This form of the equation of motion was called the Newton-
Euler equation of motion with eliminated constraint forces in (Bremer 1988). Written out
in component form it was called Kane’s equation of motion in (Kane and Levinson 1985).

8.6 Hamilton’s principle

8.6.1 Introduction
Hamilton’s principle is based on the use of the time integral of certain energy functions.
Hamilton’s principle can be used to derive Lagrange’s equation of motion for a system
described by n generalized coordinates q1 , . . . , qn . The motivation for introducing Hamil-
ton’s principle is that it is the starting point for the Hamilton-Jacobi equation, and that
it is used for systems described by partial differential equations. Moreover, it can be used
to derive the Euler-Poincaré equation. The Euler-Lagrange equation for the integral of
a function is the starting point for the development.

8.6.2 The extended Hamilton principle

The presentation starts with the extended Hamilton principle, which will be derived in
the following. Consider a system with N particles, where particle k has mass mk and
position rk (q1 , . . . , qn , t) where qi are the generalized coordinates of the system. The
velocity of particle k is vk = drk /dt, and the acceleration is ak = dvk /dt. The starting
point is again d’Alembert’s principle
N µ
X ¶
dvk
mk − Fk · δrk = 0 (8.236)
dt
k=1

The virtual work of the forces Fk satisfy

N
X
fδ
Fk · δrk = W (8.237)
k=1

fδ is defined by
where the function W
Xn µ ¶
f ∂U
Wδ = τj − δqj = Wδ − δU (8.238)
j=1
∂qj

Here
n
X
Wδ = τ j δqj (8.239)
j=1
8.6. HAMILTON’S PRINCIPLE 341

is the virtual work of the active generalized forces τ j , and

Xn
∂U
δU = δqj (8.240)
j=1
∂qj

is the variation of the potential energy U . The kinetic energy

N
1X
T = mk vk · vk (8.241)
2
k=1

has the variation

à N
! N
1X X
δT = δ mk vk · vk = mk vk · δvk
2
k=1 k=1
N
X N
X
d dvk
= (mk vk · δrk ) − mk · δrk (8.242)
dt dt
k=1 k=1

This result in combination with (8.236) and (8.237) leads to the equation
N
X
fδ − d
δT − W (mk vk · δrk ) = 0 (8.243)
dt
k=1

A critical observation for the next step in the derivation is the fact that if rk (t1 ) and
rk (t2 ) are fixed, then
Z ¯t=t2
N
t2 X
d
N
X ¯
¯
(mk vk · δrk ) dt = (mk vk · δrk )¯ =0 (8.244)
t1 dt ¯
k=1 k=1 t=t1

This means that we can eliminate the last term of (8.243) by integrating the expression
in (8.243) from t1 to t2 . The leads to the following result:

The extended Hamilton principle is given by

Z t2 ³ ´
δT + Wfδ dt = 0 (8.245)
t1

fδ = Wδ − δU , or, alternatively, by
where the endpoints are fixed and W
Z t2
(δL + Wδ ) dt = 0 (8.246)
t1

where the endpoints are fixed.

8.6.3 Derivation of Lagrange’s equation of motion

We consider a mechanical system with generalized coordinates q and Lagrangian L. We
study a trajectory C given by q(t) on the time interval t1 ≤ t ≤ t2 with the boundary
conditions that q(t1 ) and q(t2 ) are given. A variation δq(t) = ψ(t), δ q̇(t) = ψ̇(t) is
342 CHAPTER 8. ANALYTICAL MECHANICS

considered for the trajectory C, where ψ(t) is an arbitrary function. The boundary con-
ditions imply that ψ(t1 ) = ψ(t2 ) = 0. The corresponding variation δL in the Lagrangian
is µ ¶
∂L ∂L ∂L ∂L
δL = δq+ δ q̇ = ψ+ ψ̇ (8.247)
∂q ∂ q̇ ∂q ∂ q̇
The extended Hamilton principle (8.246) gives
Z t2 Z t2 µ ¶
∂L ∂L
(δL + Wδ ) dt = ψ+ ψ̇ + τ T ψ dt = 0 (8.248)
t1 t1 ∂q ∂ q̇
Partial integration gives
Z t2 · ¸t2 Z t2 µ ¶ Z t2 µ ¶
∂L ∂L d ∂L d ∂L
ψ̇dt = ψ − ψdt = − ψdt (8.249)
t1 ∂ q̇ ∂ q̇ t1 t1 dt ∂ q̇ t1 dt ∂ q̇
where it is used that ψ(t1 ) = ψ(t2 ) = 0. The variation of the integral is then found to
be Z t2 · µ ¶ ¸
∂L d ∂L
− + τ T ψdt = 0 (8.250)
t1 ∂q dt ∂ q̇
Since ψ(t) is arbitrary, this implies that
µ ¶T µ ¶T
d ∂L ∂L
− =τ (8.251)
dt ∂ q̇ ∂q
which is Lagrange’s equation of motion.

8.6.4 Hamilton’s principle

Suppose that Wδ = 0, which means that there are no active forces τ acting on the system.
In this case the extended Hamilton’s principle gives
Z t2
δLdt = 0 (8.252)
t1
which is known as Hamilton’s principle. The system will then follow some trajectory
(q, q̇) which is denoted C, where the trajectory C depends on the initial conditions.
Define the action integral of a trajectory C by
Z t2
A(C) = Ldt (8.253)
t1
which is the integral of the Lagrangian. The action integral does not have a clear phys-
ical interpretation, it is merely a mathematical tool. Hamilton’s principle can then be
reformulated to state that the variation of the action integral is zero for the trajectory
C, that is,
δA(C) = 0 (8.254)
The variation of the action integral is
Z t2 Z t2
δA(C) = δ Ldt = δLdt (8.255)
t1 t1
and from the derivation in the previous section it is seen that Hamilton’s principle implies
that µ ¶
d ∂L ∂L
− =0 (8.256)
dt ∂ q̇ ∂q
which is Lagrange’s equation of motion when the applied force is zero
8.6. HAMILTON’S PRINCIPLE 343

8.6.5 Rotations with the Euler-Poincaré equation

Consider a rigid body with rotation matrix R, angular velocity ω in body-fixed coordi-
nates defined by Ṙ = Rω × , and variation δR = Rσ × where σ is an arbitrary vector.
Suppose that the kinetic energy can be written
1 T
T = ω Mω (8.257)
2
where M is a constant matrix, and that the generalized force vector acting on the body
is τ so that Wδ = τ T σ. Assume that the potential energy is zero. Then
Z t2 Z t2 µ ¶
∂T
(δT + Wδ ) dt = δω + τ T σ dt
t1 t1 ∂ω
Z t2 · ¸
∂T ¡ ×
¢ T
= σ̇ + ω σ + τ σ dt
t1 ∂ω
Z t2 · µ ¶ ¸
d ∂T ∂T ×
= − + ω + τ T σdt (8.258)
t1 dt ∂ω ∂ω

where (8.138) and partial integration was used. Then as the components σ are indepen-
dent, Hamilton’s extended principle (8.245) gives the following result

The equation of motion for a rigid body can be written

µ ¶T µ ¶T
d ∂T ∂T
+ ω× =τ (8.259)
dt ∂ω ∂ω

which is the Euler-Poincaré equation in SO(3).

The Euler-Poincaré can be written out as Euler’s equation

Mω̇+ω × Mω = τ (8.260)

8.6.6 Rigid motion with the Euler-Poincaré equation

Consider a rigid body with homogeneous transformation matrix T and velocity vector
w = (v, ω)T in body coordinates so that Ṫ = Tŵ where ŵ is defined in (8.146). Let the
variation of T be given by δT = Tη̂. Suppose that the kinetic energy is given by
1 T
T = w Mw (8.261)
2
where M is constant. Assume that the virtual work is Wδ = τ T η, and that the potential
energy is U = 0. Then
Z t2 Z t2 µ ¶
∂T T
(δT + Wδ ) dt = δw + τ η dt (8.262)
t1 t1 ∂w

The variation δw satisfies

δw = η̇ + adw η (8.263)
η(t1 ) = η (t2 ) = 0 (8.264)
344 CHAPTER 8. ANALYTICAL MECHANICS

where adw is given by (8.152). This gives

Z t2 Z t2
∂T ¡ ¢
(δT + Wδ ) dt = η̇ + adw η + τ T η dt (8.265)
t1 t1 ∂w
Z t2 · µ ¶ ¸
d ∂T ∂T T
= − + adw +τ ηdt (8.266)
t1 dt ∂w ∂w

and, since η is arbitrary, Hamilton’s extended principle (8.245) gives

The equation of motion for rigid motion can be written in the form of a Euler-Poincaré
equation as
µ ¶T µ ¶T
d ∂T T ∂T
− adw =τ (8.267)
dt ∂w ∂w
where µ ¶
ω× v×
adw = (8.268)
0 ω×

T T
The Euler-Poincaré equation with τ = (F , LT ) gives the equations
µ ¶T µ ¶T
d ∂T ∂T
+ ω× = F (8.269)
dt ∂v ∂v
µ ¶T µ ¶T µ ¶T
d ∂T × ∂T × ∂T
+v +ω = L (8.270)
dt ∂ω ∂v ∂ω

which are known as Kirchhoff’s equations.

8.7 Lagrangian dynamics for PDE’s

8.7.1 Flexible beam dynamics
Lagrange’s equation of motion can also be used for systems described by partial differ-
ential equations. To illustrate this we will derive Lagrange’s equation of motion for an
Euler-Bernoulli beam (Meirovitch 1980). The beam is of length L, and the undeformed
beam is along the x axis. The elastic displacement in the z direction is denoted by
w(x, t). The kinetic energy is written
Z L
T (t) = T̂ [ẇ (x, t) , ẇ0 (x, t)] dx (8.271)
0

where T̂ dx is the kinetic energy of the the length element dx of the beam. The potential
energy is
Z L
U (t) = Û [w (x, t) , w0 (x, t) , w00 (x, t)] dx (8.272)
0

where Û dx is the potential energy of the length element dx. The Lagrangian can then
be defined as
Z L
L(t) = L̂ [w (x, t) , w0 (x, t) , w00 (x, t) , ẇ (x, t) , ẇ0 (x, t)] dx (8.273)
0
8.7. LAGRANGIAN DYNAMICS FOR PDE’S 345

where L̂ = T̂ − Û . The virtual work on dx due to nonconservative forces is

Ŵδ (x, t) = f (x, t)δw(x, t) (8.274)

and the virtual work from nonconservative forces on the beam is therefore
Z L
Wδ (t) = Ŵδ (x, t)dx (8.275)
0

The extended Hamilton principle for this system gives

Z t2 Z t2 Z L ³ ´
(δL + Wδ ) dt = δ L̂ + Ŵδ dxdt = 0 (8.276)
t1 t1 0

The variation of the Lagrangian density is

∂ L̂ ∂ L̂ ∂ L̂ ∂ L̂ ∂ L̂
δ L̂ = δw + 0
δw0 + 00
δw00 + δ ẇ + δ ẇ0 (8.277)
∂w ∂w ∂w ∂ ẇ ∂ ẇ0
and the extended Hamilton principle is therefore
Z t2 Z L Ã !
∂ L̂ ∂ L̂ 0 ∂ L̂ 00 ∂ L̂ ∂ L̂ 0
δw + δw + δw + δ ẇ + δ ẇ + f (x, t)δw(x, t) dxdt = 0
t1 0 ∂w ∂w0 ∂w00 ∂ ẇ ∂ ẇ0

Using partial integration and that δw = 0 and δw0 = 0 at t = t1 and t = t2 it is possible,

with some patience, to reach the following result:
Z t2 Z L " Ã ! Ã ! Ã !
∂ L̂ ∂ ∂ L̂ ∂2 ∂ L̂ ∂ ∂ L̂
− + 2 −
t1 0 ∂w ∂x ∂w0 ∂x ∂w00 ∂t ∂ ẇ
à ! #
∂2 ∂ L̂
+ + f (x, t) δwdxdt
∂x∂t ∂ ẇ0
 !# ¯L ¯L 
Z t2  " Ã ! Ã
¯ ¯ 
∂ L̂ ∂ ∂ L̂ ∂ ∂ L̂ ¯ ∂ L̂ 0¯
+ − + δw ¯ + δw ¯ dt = 0(8.278)
t1  ∂w ¯ ¯ 
0 ∂x ∂w00 ∂t ∂ ẇ0 ∂w00
0 0

As δw and δw0 are arbitrary for t1 < t < t2 this implies the Lagrangian equation of
motion in the form
à ! à ! à ! à !
∂ L̂ ∂ ∂ L̂ ∂2 ∂ L̂ ∂ ∂ L̂ ∂2 ∂ L̂
− + 2 − + + f (x, t) = 0 (8.279)
∂w ∂x ∂w0 ∂x ∂w00 ∂t ∂ ẇ ∂x∂t ∂ ẇ0

with boundary conditions

" Ã ! Ã !# ¯L
∂ L̂ ∂ ∂ L̂ ∂ ∂ L̂ ¯
¯
− + δw¯ = 0 (8.280)
∂w0 ∂x ∂w00 ∂t ∂ ẇ0 ¯
0
¯L
∂ L̂ ¯
¯
δw0 ¯ = 0 (8.281)
∂w00 ¯
0
346 CHAPTER 8. ANALYTICAL MECHANICS

Example 140 To reach the result (8.278) the following partial integrations are used.
Z ¯L Z Ã !
L
∂ L̂ ∂ L̂ ¯ L
∂ ∂ L̂
¯
δw0 dx = δw¯ − δwdx
0 ∂w0 ∂w0 ¯ 0 ∂x ∂w0
0
Z L ¯L Ã ! ¯L Z Ã !
∂ L̂ ∂ L̂ ¯ ∂ ∂ L̂ ¯ L
∂ 2
∂ L̂
¯ ¯
δw00 dx = δw0 ¯ − δw¯ + δwdx
0 ∂w
00 ∂w00 ¯ ∂x ∂w00 ¯ 0 ∂x
2 ∂w00
0 0
Z t2 Z t2 Ã !
∂ L̂ ∂ ∂ L̂
δ ẇdx = − δwdx
t1 ∂ ẇ t1 ∂t ∂ ẇ

 ¯L Z Ã ! 
Z t2 Z L Z t2 ¯ L
∂ L̂  ∂ L̂ δ ẇ¯¯ − ∂ ∂ L̂
δ ẇ0 dxdt = δ ẇdx dt
t1 0 ∂ ẇ0 t1 ∂ ẇ 0 ¯ 0 ∂x ∂ ẇ0
0
  ¯L  Z Ã ! 
Z t2 ¯ L 2
=  ∂  ∂ L̂ δw¯¯  − ∂ ∂ L̂
δwdx dt
t1 ∂t ∂ ẇ 0 ¯ 0 ∂x∂t ∂ ẇ0
0

8.7.2 Euler-Bernoulli beam

For an Euler Bernoulli beam the Lagrangian density is
1 1 2
L̂ = ρ(x) [ẇ(x, t)]2 − EI(x) [w00 (x)] (8.282)
2 2
The Lagrangian equation of motion (8.279) is in this case
à ! à !
∂2 ∂ L̂ ∂ ∂ L̂
− + f (x, t) = 0 (8.283)
∂x2 ∂w00 ∂t ∂ ẇ

which is evaluated to be
00
ρ(x)ẅ(x, t) + [EI(x)w00 (x)] = f (8.284)

8.7.3 Lateral vibrations in a string

The kinetic energy for lateral vibrations in a string is
Z
1 L 2
T = ρ(x) [ẇ(x, t)] dx (8.285)
2 0
while the potential energy is
Z L
1 2
U= P [w0 (x)] dx (8.286)
2 0

The displacement of the string is w. The string is displaced by a force P w0 . A change in

slope dw0 requires the work P w0 dw0 which integrates to
Z w0
1 2
Û = P w0 dw0 = P [w0 (x)] (8.287)
0 2
8.8. HAMILTON’S EQUATIONS OF MOTION 347

The Lagrangian density is

1 2 1 2
L̂ = ρ(x) [ẇ(x, t)] − P [w0 (x)] (8.288)
2 2
The Lagrangian equation of motion (8.279) is in this case
à ! à !
∂ ∂ L̂ ∂ ∂ L̂
− − + f (x, t) = 0 (8.289)
∂x ∂w0 ∂t ∂ ẇ

which gives the equation of motion

0
ρ(x)ẅ(x, t) = [P w0 (x)] = f (8.290)

8.8 Hamilton’s equations of motion

8.8.1 Introduction
Hamilton’s equations of motion are strongly related to Lagrange’s equation of motion, and
are based on energy expressions and generalized coordinates. In addition, the concept of a
generalized momentum vector is introduced. Hamilton’s equation of motion can be used
to establish physical properties that are important in controller design and in simulation.
In particular, this formulation is useful to establish energy functions that are invariant
with zero control input. This can be used to find Lyapunov function candidates, and for
checking the accuracy of numerical simulations. There are even specialized simulation
methods for Hamiltonian systems. In addition, the Hamiltonian formulation leads to the
Hamilton-Jacobi equation which is an important tool in optimal control theory. Basic
references for this section are (Lovelock and Rund 1989) and (Goldstein 1980).

8.8.2 Hamilton’s equation of motion

We consider a system with generalized coordinates q and Lagrangian

L(q, q̇, t) = T (q, q̇, t) − U (q) (8.291)

The momentum vector is defined by

∂L(q, q̇, t) T
p(q, q̇, t) = (8.292)
∂ q̇
We note that Lagrange’s equation of motion can be written
T
∂L(q, q̇, t)
ṗ(q, q̇, t) − =τ (8.293)
∂q
To define the Hamiltonian H from the Lagrangian L a change of variables from the
Lagrangian variables (q, q̇) to the Hamiltonian variables (q, p) is required. The velocity
vector q̇ is then regarded to be a function

q̇ = φ (q, p, t) (8.294)

of the Hamiltonian variables (q, p).

348 CHAPTER 8. ANALYTICAL MECHANICS

The Hamiltonian is defined by

H (q, p, t) = pT φ (q, p, t) −L(q, φ, t) (8.295)

Partial differentiation of (8.295) gives

µ ¶
∂H(q, p, t) ∂L(q, q̇, t) ∂φ (q, p, t)
= φT + pT −
∂p ∂ q̇ ∂p
µ ¶
∂H(q, p, t) ∂L(q, q̇, t) ∂φ (q, p, t) ∂L(q, φ, t)
= pT − −
∂q ∂ q̇ ∂q ∂q
µ ¶
∂H(q, p, t) ∂L(q, q̇, t) ∂φ (q, p, t) ∂L(q, φ, t)
= pT − −
∂t ∂ q̇ ∂t ∂t

where the definition (8.292) has been used. It follows from the definition (8.292) of the
momentum vector that
∂H(q, p, t)
= φT (8.296)
∂p
∂H(q, p, t) ∂L(q, φ, t)
= − (8.297)
∂q ∂q
∂H(q, p, t) ∂L(q, φ, t)
= − (8.298)
∂t ∂t
Insertion of Lagrange’s equation of motion and q̇ = φ (q, p, t) leads to the result:

Hamilton’s equations of motion are given by

∂H(q, p, t) T
q̇ = (8.299)
∂p
∂H(q, p, t) T
ṗ = − +τ (8.300)
∂q

The time derivative of the Hamiltonian is found from the chain rule to be
dH ∂H ∂H ∂H
= ṗ+ q̇+ (8.301)
dt ∂p ∂q ∂t
By inserting Hamilton’s equations of motion (8.299) and (8.300) we find that
à !
dH T ∂H T ∂H ∂H ∂H
= q̇ − +τ + q̇+ = q̇T τ + (8.302)
dt ∂q ∂q ∂t ∂t

which leads to:

The time derivative of the Hamiltonian is

dH(q, p, t) ∂H(q, p, t)
= q̇T τ + (8.303)
dt ∂t
8.8. HAMILTON’S EQUATIONS OF MOTION 349

The following result is useful:

If the Hamiltonian does not depend on time t, that is, if H = H(q, p), and if the system
is unactuated so that τ = 0, then

dH(q, p)
=0 (8.304)
dt

8.8.3 The energy function

Define the energy function h(q, q̇, t) by
∂L(q, q̇, t)
h (q, q̇, t) = q̇−L(q, q̇, t) (8.305)
∂ q̇
The time derivative of the energy function is found from the definition (8.305) to be
µ ¶
dh (q, q̇, t) d ∂L(q, q̇, t) dL
= q̇ − (q, q̇, t) (8.306)
dt dt ∂ q̇ dt
Insertion of (8.15) gives the result
dh (q, q̇, t) ∂L(q, q̇, t)
=− + τ T q̇ (8.307)
dt ∂t
From (8.292), (8.294), (8.295), and (8.305) it is possible to see that the Hamiltonian
H (q, p, t), has the same numerical value as the energy function h (q, q̇, t), that is,
H (q, p, t) = h (q, q̇, t) (8.308)
However, the two functions have different arguments, and should not be confused with
each other.
Suppose that the kinetic energy is quadratic in the velocity, that is,
1 T
T (q, q̇, t) = q̇ M (q, t) q̇ (8.309)
2
Then the the energy function becomes
h (q, q̇, t) = T (q, q̇, t) + U (q) (8.310)
which is the sum of the kinetic and potential energy. This explains the name energy
function.
The kinetic energy can in general be written in the form
1 T
T (q, q̇, t) = q̇ M (q, t) q̇ + α(q, t)T q̇ + T0 (q, t) (8.311)
2
= T2 (q, q̇, t) + T1 (q, q̇, t) + T0 (q, t) (8.312)
where T2 (q, q̇, t) is quadratic in the velocities, T1 (q, q̇, t) is linear in the velocities and
T0 (q, t) is independent of the velocity. We find that
· ¸
∂L
q̇ = q̇T Mq̇ + αT q̇ = 2T2 + T1 (8.313)
∂ q̇
and using (8.305) and (8.307) we may state the following result:
350 CHAPTER 8. ANALYTICAL MECHANICS

The energy function is given by

h(q, q̇, t) = T2 (q, q̇, t) + Ua (q, t) (8.314)

where Ua (q, t) = U (q, t) − T0 (q, t) may be considered to be an apparent potential. If

h = h(q, q̇) and the system is unactuated, then

dh(q, q̇)
=0 (8.315)
dt

8.8.4 Change of generalized coordinates

Consider a system with Lagrangian L(q, q̇, t) = T (q, q̇, t) − U (q), where
1 T
T (q, q̇, t) = q̇ M (q, t) q̇ (8.316)
2
The energy function is
h (q, q̇, t) = T (q, q̇, t) + U (q) (8.317)
while the momentum vector is
∂L(q, q̇, t) T
p(q, q̇, t) = = M (q, t) q̇ (8.318)
∂ q̇
and the Hamiltonian is
H (q, p, t) = pT φ−L(q, φ, t) (8.319)
where φ (q, p, t) = q̇.
A change in coordinates
q = q0 + Q (8.320)
from q to Q gives
1 T T 1
TQ (Q, Q̇, t) =Q̇ MQ (Q, t) Q̇ + q̇0 MQ (Q, t) Q̇+ q̇T0 MQ (Q, t) q̇0
2 2
= T (q, q̇, t) (8.321)
UQ (Q) = U (q) (8.322)

The Lagrangian in the new coordinates is

LQ (Q, Q̇, t) = TQ (Q, Q̇, t) − UQ (Q) = L(q, q̇, t) (8.323)

which means that the Lagrangian has the same numerical value after the change in
generalized coordinates. The momentum vector is

∂LQ (Q, Q̇, t)

T ³ ´
P(Q, Q̇, t) = = MQ (Q, t) Q̇ + q̇0 = p(q, q̇, t) (8.324)
∂ Q̇
The energy function becomes
1 T 1
hQ (Q, Q̇, t) = Q̇ MQ (Q, t) Q̇+UQ (Q) − q̇T0 MQ (Q, t) q̇T0
2 h 2 i
T
= h (q, q̇, t) − q̇T0 MQ (Q, t) Q̇ + q̇0 MQ (Q, t) q̇0 (8.325)
8.9. CONTROL ASPECTS 351

while the Hamiltonian is

HQ (Q, P, t) = PT Φ−LQ (Q, Φ, t) = H (q, p, t) − pT q̇0 (8.326)

where Φ = Q̇. We see that the Lagrangian and the canonical momentum vector has
the same numerical value after a change of coordinates, while the numerical value of the
energy function and the Hamiltonian changes when the coordinate is changed from q to
Q.

Example 141 We will demonstrate the effect of a change of generalized coordinates on

the Lagrangian and the Hamiltonian with an example. Consider a satellite modeled as a
mass point moving about the earth in a nominally circular orbit. Let the position of the
satellite be
r=R+q (8.327)
where R is the nominal circular motion with radius R = |R| and a constant angular
velocity ω c around the earth. The kinetic and potential energy of the satellite are
1 T m
Tr = mṙ ṙ, Ur = µ (8.328)
2 r
where r = |r|. The energy function
1 T m
hr (r, ṙ) = mṙ ṙ+µ (8.329)
2 r
is constant. Suppose that the velocity is given by

v = ω c Rc1 + q̇ (8.330)

where c1 is the unit vector along the tangent of the nominal orbit. The kinetic energy
may be written
1 1
Tq = mq̇T q̇+ω c RcT1 q̇+ mω 2c R2 (8.331)
2 2
With these variables the energy function is
1
hq (q, q̇) = mq̇T q̇+Ua (q) (8.332)
2
where
m 1
Ua (q) = µ − mω 2c R2 (8.333)
|R + q| 2
is the apparent potential.

8.9 Control aspects

8.9.1 Passivity of Hamilton’s equation of motion
Suppose that the Hamiltonian does not depend on time, so that H = H(q, p). Then from
(8.303) the time derivative of H along the solutions of Hamilton’s equation of motion
(8.299, 8.300) is
dH(q, p)
= q̇T τ (8.334)
dt
352 CHAPTER 8. ANALYTICAL MECHANICS

Next, suppose that the Hamiltonian is bounded from below, which means that there is a
constant H0 so that H ≥ H0 . Then it is possible to define a nonnegative storage function
V = H − H0 so that the time derivative of V along the solution of the system is

V̇ = q̇T τ (8.335)

From the results of Section 2.4.14 we see that this leads to the following conclusion:

If the Hamiltonian does not depend on time t, and if there is a constant H0 so that
H(q, p) ≥ H0 , then the system given by Hamilton’s equation of motion (8.299, 8.300) is
passive with input τ and output q̇, and V = H − H0 is a storage function.

Example 142 Suppose that the Lagrangian L does not depend on time t. Then it is seen
from (8.298) and the definition (8.305) of the energy function h that the Hamiltonian H
and the energy function h will not depend on t: This gives
dH(q, p) dh(q, q̇)
= = q̇T τ (8.336)
dt dt
If the actuator force is set to zero, then this implies that H(q, p) and h(q, q̇) are constants
for solutions of the system. In the terminology of Hamiltonian dynamics H and h are
said to be invariants of motion.

Example 143 Suppose that (8.336) holds. Velocity feedback in the form τ = −K q̇ will
then lead to
dH(q, p) dh(q, q̇)
= = −K q̇T q̇ ≤ 0 (8.337)
dt dt
which means that the energy h(q, q̇) = H(q, p) will be nonincreasing, and that the energy
will decrease whenever the velocity is nonzero.

8.9.2 Example: Manipulator dynamics

Consider a robotic manipulator with Lagrangian equation of motion

M (q) q̈ + C (q, q̇) q̇ + g (q) = τ (8.338)

where M is symmetric and positive definite, which implies that M is nonsingular, and
where Ṁ − 2C is skew symmetric. The Lagrangian is
1 T
L= q̇ M (q) q̇ − U (q) (8.339)
2
and the momentum vector is
T
∂L(q, q̇, t)
p= = M (q) q̇ (8.340)
∂ q̇
As M is nonsingular the velocity can be written

q̇ = M (q)−1 p (8.341)

The Hamiltonian is found from the definition (8.295) to be

1
H = pT M (q)−1 p− pT M (q)−1 M (q) M (q)−1 p + U (q) (8.342)
2
8.9. CONTROL ASPECTS 353

which simplifies to
1 T
H(q, p) = p M (q)−1 p + U (q) (8.343)
2
where the first term is the kinetic energy. Then, if U (q) is lower bounded, passivity of
the manipulator dynamics with input τ and output q̇ can be concluded from the general
result in Section 8.9.1. This agrees with the passivity analysis based on the Lagrangian
dynamics in Section 8.2.9. Note that it is easier to establish passivity from τ to q̇ in the
Hamiltonian formulation as it is only necessary to check that H does not depend on t
and that U is lower bounded.

8.9.3 Example: The restricted three-body problem

The restricted three-body problem (Szebehely 1967) is a classical problem in celestial
mechanics that has been adopted as a benchmark for numerical integrators (Hairer and
Wanner 1996), (Shampine, Allen and Preuss 1997). The reason for this is that there are
periodic solutions that are tabulated, and these solutions are very sensitive to changes in
initial conditions. Therefore, the accuracy of a numerical integrator can be investigated
by solving the system equations with initial conditions that correspond to a periodic
solution, and then check if the numerically computed solution is periodic. This is done
in Section 14.1.3.
The system includes three masses moving in a plane. The primary body is the earth,
the secondary body is the moon, and the third body is a satellite. The earth has mass
m1 , the moon has mass m2 , and the satellite has mass m3 , which is much smaller than
m1 and m2 . In the formulation of the problem it is assumed that the moon and the earth
interact in a gravitational field without being influenced by the satellite. A coordinate
frame b has the unit vector b1 along the axis from the earth to the moon, while the b3
vector is along the axis of rotation of the earth-moon system. According to the law of
gravitation the gravity force F1 on the earth from the moon, and the gravity force F2 on
the moon from the earth are given by
m1 m2
F1 = −F2 = k 2 b1 (8.344)
L2
where k is the Gaussian constant of gravitation and L is the distance between the two
bodies. The vector from the center of the earth to the center of the moon rotates with
an angular velocity ω = ω b3 . The earth has position R1 = −x1 b1 and the moon has
position R2 = x2 b1 , which implies that L = x1 + x2 . The accelerations are according to
(6.405)

a1 = ω × (ω × R1 ) = ω 2 x1 b1 (8.345)
2
a2 = ω × (ω × R2 ) = −ω x2 b1 (8.346)

Balance between the centrifugal forces and the gravitational forces imply that
m1 m2
k2 = m1 x1 ω 2 = m2 x2 ω 2 (8.347)
L2
From this we get Kepler’s third law:

k2 M
ω2 = (8.348)
L3
where M = m1 + m2 .
354 CHAPTER 8. ANALYTICAL MECHANICS

The satellite moves in the rotating gravitational field set up by the earth and the
moon. The position of the satellite is
r = xb1 + y b2 (8.349)
the velocity is
d b ³ ´
v= r + ω ib × r = ẋb1 + ẏ b2 + ω xb2 − y b1 (8.350)
dt
and from (6.405) the acceleration is
b 2 b
d d
a = 2
r + 2ω ib × r + αib × r + ω ib × (ω ib × r)
dt dt
³ ´ ³ ´
= ẍb1 + ÿ b2 + 2ω ẋb2 − ẏ b1 − ω 2 xb1 + y b2 (8.351)

The gravitational force on the satellite is the sum of the gravitational forces from the
earth and the moon, which gives
m1 m3 h i m2 m3 h i
F3 = −k 2 3 (x + x1 ) b1 + y b2 − k2 (x − x2 ) b 1 + y b 2 (8.352)
r1 r23
where q q
2 2
r1 = (x + x1 ) + y 2 , r2 = (x − x2 ) + y 2 (8.353)
Newton’s law in the x and y directions gives
· ¸
m1 m2
ẍ − 2ω ẏ − ω 2 x = −k2 (x + x 1 ) + (x − x2 ) (8.354)
r3 r23
µ 1 ¶
m1 m2
ÿ + 2ω ẋ − ω 2 y = −k2 + 3 (8.355)
r13 r2
This model is normalized by introducing
x y
ξ = , η = , τ = ωt (8.356)
L L
r1 r2
ρ1 = , ρ2 = (8.357)
L L
m1 x2 m2 x1
µ1 = = , µ2 = = (8.358)
M L M L
This gives the normalized model for the restricted three-body problem:
· ¸
d2 ξ dη µ1 (ξ + µ2 ) µ2 (ξ − µ1 )
−2 −ξ = − + (8.359)
dτ 2 dτ ρ31 ρ32
2
µ ¶
d η dξ µ1 η µ2 η
+2 −η = − + 3 (8.360)
dτ 2 dτ ρ31 ρ2
A constant energy function of the system is found from the kinetic energy
1
T = m3 v · v
2
1 h ³ ´i h ³ ´i
= m3 ẋb1 + ẏ b2 + ω xb2 − y b1 · ẋb1 + ẏ b2 + ω xb2 − y b1
2
1 ¡ ¢ 1 ¡ ¢
= m3 ẋ2 + ẏ 2 + m3 ω (−ẋy + ẏx) + m3 ω 2 x2 + y 2 (8.361)
|2 {z } | {z } |2 {z }
T1
T2 T0
8.9. CONTROL ASPECTS 355

and the potential energy µ ¶

2 m1 m2
U = −k m3 + (8.362)
r1 r2
Then (8.314) and (8.315) imply that an invariant energy function is given by
µ ¶
1 £ 2 2 2
¡ 2 2
¢¤ 2 m1 m2
h = m3 ẋ + ẏ − ω x + y − k m3 + (8.363)
2 r1 r2

In the normalized form the invariant energy function is

"µ ¶ µ ¶2 #
2
1 dξ dη 2 µ µ
κ= + −ξ −η − 1 − 2
2
(8.364)
2 dτ dτ ρ1 ρ2

8.9.4 Example: Attitude dynamics for a satellite

Consider a satellite that moves in a circular orbit about the earth with radius Rc . The
satellite has mass m and inertia dyadic Mc about the center of mass. The center of mass
has position Rc relative to the origin of frame i. Frame b is fixed in the satellite, and
frame i is a Newtonian frame with origin in the center of the earth and with axes pointing
at certain fixed stars. An orbital frame c is defined so that c1 is along the tangent of
the orbit in the positive velocity direction, c2 is perpendicular to the orbit, and c3 is
locally vertical and pointing downwards. Then Rc = −Rc c3 . The attitude dynamics are
assumed to be given by
³ ´
Mc · αib + ω ib × Mc · ω ib = τ c (8.365)

Here ω ib is the angular velocity of frame b relative to frame i, αib is the angular acceler-
ation, and τ c is the actuator torque.
A satellite moving in a circular orbit will have velocity v = ω c Rc c1 where
r
µ
ωc = (8.366)
Rc3

is the orbital frequency in the sense that T = 2π/ω c is the period of one orbit. In passing
we mention that this equation and the fact that the radius of the earth is 6,378 km
make it possible to compute the altitude of geostationary orbits as 35,863 km, because
this gives an orbital period of 24 hours. Also a low earth orbit with altitude 1,200 km
corresponds to a radius Rc = 7578 km and an orbital period of 109 min.
The kinetic energy of the satellite is
1 1
Ti = ω ib · Mc · ω ib + mω 2c Rc2 (8.367)
2 2
while the potential energy is
µm
U =− (8.368)
Rc
Then the energy function

hi = Ti2 − Ti0 + U
1 1 µm
= ω ib · Mc · ω ib − mω 2c Rc2 − (8.369)
2 2 Rc
356 CHAPTER 8. ANALYTICAL MECHANICS

is constant as long as the system is not actuated. The last two terms are constants, so
that the function
1
Vi = ω ib · Mc · ω ib (8.370)
2
will be a constant for the unactuated system.
The description based on the angular velocity ω ib gives the rotation of the satellite
relative to a star-fixed coordinate frame i. For the stabilization of the attitude in an orbit
it is better to study the dynamics in terms of the angular velocity ω cb of the satellite
relative to the orbit frame. The change of variables is done using
ω ib = ω cb + ω ic (8.371)
= ω cb + ω c c2 (8.372)
This gives the expression
1 1 1
Tc = ω cb · Mc · ω cb + ω c ω cb · Mc · c2 + ω 2c c2 · Mc · c2 + mω 2c Rc2 (8.373)
2 2 2
for the kinetic energy. Then the energy function, which is found from
hc = Tc2 − Tc0 + U (8.374)
1 1 1 µm
= ω cb · Mc · ω cb − ω 2c c2 · Mc · c2 − mω 2c Rc2 − (8.375)
2 2 2 Rc
is constant for the unactuated satellite. The last two terms on the right side are constants,
and this shows that
1 1
Vc = ω cb · Mc · ω cb − ω 2c c2 · Mc · c2 (8.376)
2 2
is a constant function for the unactuated satellite.

8.9.5 Example: Gravity gradient stabilization

In this section we introduce gravity gradient stabilization of the satellite in the previous
section. The material is adopted from (Hughes 1986). The gravity force acting on a mass
element is then
R
df = −µ 3 dm (8.377)
R
where R3 denotes |R|3 . Note that the gravity force has a gradient in the radial direction,
and this creates a torque about the center of mass. This torque is known as the gravity
gradient torque, and is given by
Z
r×R
gc = −µ 3
dm (8.378)
b R

where r = R − Rc is the vector from the center of mass to the mass element dm. The
gravitational potential is Z
dm
U = −µ (8.379)
b R
The expressions for the gravity gradient gc and the gravitational potential are difficult to
use in their present form, but can be approximated with more suitable expressions using
the binomial series à !
1 1 r · Rc
= 3 1 − 3 2 + ... (8.380)
R3 Rc Rc
8.9. CONTROL ASPECTS 357

and à !
1 1 r · Rc 1 r 2 3 (r · Rc )2
= 1− − + + ... (8.381)
R Rc Rc2 2 Rc2 2 Rc4

Then, using r × R = r × Rc the gravitational torque can be approximated by

Z Z
µ 3µ
gc = − 3 rdm × Rc − 5 Rc · rrdm × Rc
Rc b Rc b
Z
3µ
= − 3 c3 × rrdm · c3 (8.382)
Rc b
R
where b rdm = 0 is used. The inertia dyadic is
Z ³ ´
Mc = r2 I − rr dm (8.383)
b

It is found that
Z Z
c3 × rrdm · c3 = c3 × r2 Idm · c3 + c3 × Mc · c3 (8.384)
b b
Z
= c3 × c3 r2 dm + c3 × Mc · c3 = c3 × Mc · c3 (8.385)
b

which gives
gc = 3ω 2c c3 × Mc · c3 (8.386)
The equation of motion can then be written
³ ´
Mc · αib + ω ib × Mc · ω ib = 3ω 2c c3 × Mc · c3 + τ c (8.387)

From the binomial series the approximation

3
U = U0 + ω 2c c3 · Mc · c3 (8.388)
2
is found for the potential energy where U0 is a constant given by
µm 1 2
U0 = − − ω c TraceMc (8.389)
Rc 2
The kinetic energy is
1 1
Ti = ω ib · Mc · ω ib + mω 2c Rc2 (8.390)
2 2
where the velocity has been assumed to be v = ω c Rc c1 . Then the energy function

hi = Ti2 − Ti0 + U (8.391)

1 3 1
= ω ib · Mc · ω ib + ω 2c c3 · Mc · c3 + U0 − mω 2c Rc2 (8.392)
2 2 2
is constant as long as the system is not actuated, and as the last two terms of the energy
function are constants, the function
1 3
Vi = ω ib · Mc · ω ib + ω 2c c3 · Mc · c3 (8.393)
2 2
358 CHAPTER 8. ANALYTICAL MECHANICS

will be a constant of motion.

A change of variables using

ω cb = ω ib − ω ic (8.394)
= ω ib − ω c c2 (8.395)

will give the following expression for the kinetic energy

1 1 1
T = ω ib · Mc · ω ib − ω c ω ib · Mc · c2 + ω 2c c2 · Mc · c2 + mω 2c Rc2 (8.396)
2 | {z } 2 2
| {z } T
| {z }
1
T2 T0

Then the energy function is found to be

h = T2 − T0 + U
1 3 1 1
= ω cb · Mc · ω cb + ω 2c c3 · Mc · c3 − ω 2c c2 · Mc · c2 + U0 − mω 2c Rc2
2 2 2 2
Again, the last two terms of the energy function are constants, and it follows that the
function
1 3 1
Vc = ω cb · Mc · ω cb + ω 2c c3 · Mc · c3 − ω 2c c2 · Mc · c2 (8.397)
2 2 2
is a constant for the unactuated system.
The function Vc was used for Lyapunov analysis of gravity gradient stabilization of
a satellite in (Hughes 1986). What can be learned from this example is that nontrivial
energy functions for Lyapunov analysis can be derived using Hamilton theory, and in
particular, that this approach offer a systematic way of changing coordinates in the
description.

8.10 The Hamilton-Jacobi equation

In this section the Hamilton-Jacobi equation will be developed from Hamilton’s principle
(Lovelock and Rund 1989). The main idea is that the dynamics of a system with zero
input forces are found by minimization of the action integral A(C). Then, by introducing
a modified Lagrangian L∗ , it is possible to define the velocity of the system as a velocity
field that satisfies a partial differential equation known as the Hamilton-Jacobi equation.
This method is the underlying idea for the use of the Hamilton-Jacobi equation in optimal
control.
Let the Lagrangian L (q, q̇, t) and the associated momentum vector

∂L (q, q̇, t) T
p= (8.398)
∂ q̇
be given. Suppose that the velocity can be given by the function

q̇ = φ (q, p, t) (8.399)

The Hamiltonian is then

H (q, p, t) = pφ (q, p, t) − L (q, φ (q, p, t) , t) (8.400)

The action integral A(C) as defined in Section 8.6.4 satisfies δA(C) = 0 for a tra-
jectory C with fixed end points, and this implies Lagrange’s equation of motion. We
8.10. THE HAMILTON-JACOBI EQUATION 359

introduce a function S(q, t), which is yet to be specified, and define the alternative La-
grangian function
dS(q, t)
L∗ (q, q̇, t) = L (q, q̇, t) − (8.401)
dt
From the chain rule it follows that
dS ∂S ∂S
= + q̇ (8.402)
dt ∂t ∂q
Define the alternative action integral A∗ (C) by
Z t2 Z t2 µ ¶ Z t2 Z t2
∗ ∗ dS
A (C) = L dt = L− dt = Ldt − dS (8.403)
t1 t1 dt t1 t1

This implies that

A∗ (C) = A(C) − (S2 − S1 ) (8.404)
where S1 and S2 are the values of S at the end points. As the difference S2 − S1 is
independent of the curve C, it follows that minimum value for A∗ (C) is found for the
same trajectory as the minimum for A(C). Therefore the system can be analyzed in terms
of A∗ (C) and the function L∗ (q, q̇, t) instead of A(C) and L (q, q̇, t) .
Suppose that a velocity field ψ (q, t) defined according to
q̇ = φ (q, p, t) = ψ (q, t) (8.405)
so that for a suitably selected function S the alternative Lagrangian satisfies
L∗ (q, ψ (q, t) , t) = 0 (8.406)
L∗ (q, q̇, t) > 0 when q̇ 6= ψ (q, t) (8.407)
Then the field ψ (q, t) is called the geodesic field. Note that on the geodesic field L∗ = 0
and (8.401) implies that
dS
=L (8.408)
dt
We will now solve our problem under the assumption that a geodesic field exists.
The geodesic field q̇ = ψ (q, t) may be integrated with respect to the time t to give
a family of curves. We let C denote one of the curves of this family with initial time
t1 corresponding to a point P1 , and final time t2 corresponding to a point P2 . Then
q̇ = ψ (q, t) at each point along the curve. Moreover, along the curve we have A∗ (C) = 0,
whereas A∗ (K) > 0 for any other curve K between the points P1 and P2 . This means
that the curve C minimizes the alternative action integral A∗ , which implies that C is
the solution to the minimization of the action integral A. The curve C is therefore the
solution to Lagrange’s equation of motion with the Lagrangian L. This means that if we
are able to find the velocity field ψ (q, t), then we have the solution of the equation of
motion for any initial condition.
By introducing the function S and the geodesic field ψ (q, t) we have changed the
problem of minimizing the action integral A(C) with respect to C over the time interval
t1 ≤ t ≤ t2 , into the problem of minimizing the function L∗ (q, q̇, t) with respect to q̇
for each q and each t. This is simply a minimization problem for a function. In analogy
with the result from basic calculus that the minimum for f (x) is found for df /dx = 0,
the minimum of L∗ (q, q̇, t) for fixed q and t is found for

∂L∗ T
=0 (8.409)
∂ q̇
360 CHAPTER 8. ANALYTICAL MECHANICS

This gives
∂L∗ T ∂L T ∂S T ∂S T
0= = − =p− (8.410)
∂ q̇ ∂ q̇ ∂q ∂q
which implies that
∂S
p= (8.411)
∂q
on the geodesic field. Moreover, from (8.402) and (8.408) it follows that
dS ∂S
L [q, ψ (q, t) , t] = = + pT ψ (q, t) (8.412)
dt ∂t
This can be combined with
H(q, p, t) = pT ψ (q, t) −L (q, q̇, t) (8.413)
and the following result is found:

The momentum vector p can be found as a vector field

∂S(q, t) T
p(q, t) = (8.414)
∂q

where S(q, t) is the solution of the Hamilton-Jacobi equation

· ¸
∂S(q, t) ∂S(q, t)
+ H q, ,t = 0 (8.415)
∂t ∂q

which is a partial differential equation in S(q, t).

Example 144 The time derivative of p is found to be

T T
d ∂S (q, t) ∂ 2 S (q, t) ∂ 2 S (q, t)
ṗ = = + q̇ (8.416)
dt ∂q ∂t∂q ∂q∂q
Partial differentiation of the Hamilton-Jacobi equation gives
T
∂2S ∂H T ∂H ∂ 2 S (q, t)
+ + =0 (8.417)
∂t∂q ∂q ∂p ∂q∂q
T
Insertion of q̇ = (∂H/∂p) gives
T
∂2S ∂ 2 S (q, t) ∂H T
+ q̇ = − (8.418)
∂t∂q ∂q∂q ∂q
This shows that
∂H T
ṗ = − (8.419)
∂q
and it has been established that the solution of the Hamilton-Jacobi equation is consistent
with the Hamilton’s equations of motion
∂H T
q̇ = (8.420)
p
∂H T
ṗ = − (8.421)
∂q
Chapter 9

Mechanical vibrations

9.1 Introduction

Active damping of mechanical vibrations has been important for space structures where
there is little damping and where the structures are designed for low weight. As new
inexpensive sensors and actuators are becoming available, active vibration damping is
being more used also in civil engineering, crane systems, and transportation. In this
chapter vibration models will be developed for the string model and the Euler Bernoulli
beam model. These models represent important properties that are seen for models of
mechanical vibrations. Models are developed using assumed modes, mostly in the form of
orthogonal modes, and finite-element models. The Galerkin method is used to illustrate
similarities between the assumed mode method and the finite-element approach. Also
irrational transfer functions are derived, and examples with positive realness and non-
minimum phase dynamics are discussed. The chapter starts with systems with lumped
components, and progresses with distributed parameter models.

Elastic systems consisting of rigid bodies connected with springs and dampers are
described by ordinary differential equations, and are said to be lumped parameter sys-
tems. In contrast to this, systems with elastic bodies are described by partial differential
equations, and are said to be distributed parameter systems. In this chapter we will first
present results on elastic systems with lumped parameters and then proceed with results
on distributed parameter systems.

Energy considerations and passivity properties are important in the analysis and
controller design for elastic systems with distributed parameters. We will see that if
the applied force at a point is the input, and the velocity at the same point is the
output, then the input and output are said to be collocated, and the system is passive.
A system where force is input and velocity is output will not be passive if the input
and output are noncollocated , that is, if the velocity is not measured at the point where
the force is applied. Such systems may even be nonminimum phase, which may cause
severe restrictions on the performance of the closed loop system. In this section we will
study these phenomena closer. The material is adopted from (Meirovitch 1967), (Weaver,
Timoshenko and Young 1990) and (Rao 1990).

361
362 CHAPTER 9. MECHANICAL VIBRATIONS

9.2 Lumped elastic two-ports

9.2.1 Hybrid two-port

K
m1 F
x0
D x1

Figure 9.1: Mechanical two-port of the hybrid type with inputs ẋ0 and F .

The equation of motion for a mass m that is connected to a point x0 with a spring
with stiffness K and a damper with coefficient D is given by

mẍ + D(ẋ − ẋ0 ) + K(x − x0 ) = F (9.1)

Here x is the position of the mass, and F is a force acting on the mass. The system
is shown in Figure 9.1. In the following this simple system will be used as a building
block in some of the models in this section. In this connection it is useful to have
a two-port description of the system where one port has input ẋ0 and output F0 =
D(ẋ − ẋ0 ) + K(x − x0 ), and the other port has input F and output ẋ. This will be
termed a hybrid formulation as one port has the force as output, and the other has
velocity as output. The ports are therefore compatible for a serial interconnection.

9.2.2 Displacement two-port

K1
F1 m1 m2 F2
x1 D1 x2

Figure 9.2: Mechanical two-port of the displacement type with inpouts F1 and F2 .

Another possible building block for models of lumped elastic systems is two masses
connected by a spring and a damper. The system has two masses m1 and m2 connected
by a spring K1 and a damper D1 . An externally applied force F1 is acting on mass m1 ,
and a force F2 is acting on m2 . The position of mass i is denoted xi . The system is
shown in Figure 9.2, and the equations of motion for the system are

m1 ẍ1 + D1 (ẋ1 − ẋ2 ) + K1 (x1 − x2 ) = F1 (9.2)

m2 ẍ2 + D1 (ẋ2 − ẋ1 ) + K1 (x2 − x1 ) = F2 (9.3)

This can be written in vector form as

Mẍ + Dẋ + Kx = F (9.4)

9.2. LUMPED ELASTIC TWO-PORTS 363

where x = (x1 , x2 )T , F = (F1 , F2 )T , and the mass matrix M, the damping matrix D
and the stiffness matrix K are given by
µ ¶ µ ¶ µ ¶
m1 0 D1 −D1 K1 −K1
M= , D= , K= (9.5)
0 m2 −D1 D1 −K1 K1
In a network setting the system described by (9.2) and (9.3) is a two-port where port
1 has input F1 and output ẋ1 , and port 2 has input force F2 and output ẋ2 . This type
of two-port will be said to be in a displacement formulation as both ports have force as
input and velocity as output.
Example 145 The ports of a displacement two-port are compatible with connections to
springs and dampers, which have velocity as input and force as output. To demonstrate
this we consider the case where mass m1 is connected to a fixed point with a spring K0
and a damper D0 , and mass m2 is connected to a fixed point with a spring K2 and a
damper D2 . This means that the input port is connected to the one-port
F1 = − (K0 x1 + D0 ẋ1 ) (9.6)
and that the output port is connected to the one-port
F2 = − (K2 x2 + D2 ẋ2 ) (9.7)
Then the damping and stiffness matrices for the interconnected system is obtained by
inserting F1 and F2 into (9.2) and (9.3) which gives
µ ¶ µ ¶
D0 + D1 −D1 K0 + K1 −K1
D= , K= (9.8)
−D1 D1 + D2 −K1 K1 + K2

9.2.3 Three masses in the hybrid formulation

Consider the three hybrid two-ports
m1 ẍ1 = F1 (9.9)
m2 ẍ2 + D1 (ẋ2 − ẋ20 ) + K1 (x2 − x20 ) = F2 (9.10)
m3 ẍ3 + D2 (ẋ3 − ẋ30 ) + K2 (x3 − x30 ) = F3 (9.11)
The two-ports have compatible ports variables, and can be connected by assigning the
input forces to be
F1 = D1 (ẋ2 − ẋ1 ) + K1 (x2 − x1 ) + τ 1 (9.12)
F2 = D2 (ẋ3 − ẋ2 ) + K2 (x3 − x2 ) + τ 2 (9.13)
F3 = τ3 (9.14)
and the input displacements to be
x20 = x1 , x30 = x2 (9.15)
This gives the total system
m1 ẍ1 + D1 (ẋ1 − ẋ2 ) + K1 (x1 − x2 ) = τ 1 (9.16)
m2 ẍ2 + D1 (ẋ2 − ẋ1 ) + K1 (x2 − x1 ) (9.17)
+D2 (ẋ2 − ẋ3 ) + K2 (x2 − x3 ) = τ 2 (9.18)
m3 ẍ3 + D2 (ẋ3 − ẋ2 ) + K2 (x3 − x2 ) = τ 3 (9.19)
364 CHAPTER 9. MECHANICAL VIBRATIONS

In vector form this is written

Mẍ + Dẋ + Kx = τ (9.20)
where x = (x1 , x2 , x3 )T , τ = (τ 1 , τ 2 , τ 3 )T , and
 
m1 0 0
M =  0 m2 0  (9.21)
0 0 m2
 
D1 −D1 0
D =  −D1 D1 + D2 −D2  (9.22)
0 −D2 D2
 
K1 −K1 0
K =  −K1 K1 + K2 −K2  (9.23)
0 −K2 K2

9.2.4 Three masses in the displacement formulation

In this section we will derive the result of the previous section using displacement two-
ports. This procedure is of great interest as it resembles the method that is used to
interconnect elements in the displacement formulation of the finite element method. This
will be done by connecting two systems given as the displacement two-ports
m1 ẍ1 + D1 (ẋ1 − ẋ2 ) + K1 (x1 − x2 ) = τ 1 (9.24)
m2
ẍ2 + D1 (ẋ2 − ẋ1 ) + K1 (x2 − x1 ) = F12 (9.25)
2
and
m2
ẍ2 + D2 (ẋ2 − ẋ3 ) + K2 (x2 − x3 ) = F21 (9.26)
2
m3 ẍ3 + D2 (ẋ3 − ẋ2 ) + K2 (x3 − x2 ) = τ 3 (9.27)
The interconnection is done by requiring that the second mass of system 1 has the same
position as the first mass of system 2, that is, by requiring that ẋ2 is the same for the
two systems. In this case the port variables are not compatible, so the equations of
motion must be combined. To see how the equations must be combined it is noted that
to ensure that the masses are interconnected, there must be a constraint force F (c) so
that the forces acting on the two masses are given by
1 1
F12 = F (c) + τ 2 , F21 = −F (c) + τ 2 (9.28)
2 2
To derive the equations of motion of the interconnected system it is necessary to eliminate
the constraint force F (c) . This can be done by simply adding the equations (9.25) and
(9.26). This gives the equations of motion
m1 ẍ1 + D1 (ẋ1 − ẋ2 ) + K1 (x1 − x2 ) = τ 1 (9.29)
m2 ẍ2 + D1 (ẋ2 − ẋ1 ) + K1 (x2 − x1 ) (9.30)
+D2 (ẋ2 − ẋ3 ) + K2 (x2 − x3 ) = τ 2 (9.31)
m3 ẍ3 + D2 (ẋ3 − ẋ2 ) + K2 (x3 − x2 ) = τ 3 (9.32)
which is the same result as the one derived with the hybrid formulation in the previous
section.
9.3. VIBRATING STRING 365

9.2.5 Four masses

It is straightforward to connect one more mass to the system to have four interconnected
masses m1 , m2 , m3 and m4 . In addition, springs can be connected to terminate the ports
of mass 1 and 4. The equation of motion for the system is found to be

Mẍ + Dẋ + Kx = τ (9.33)

Here x = (x1 , x2 , x3 , x4 )T , τ = (τ 1 , τ 2 , τ 3 , τ 4 )T , M = diag(m1 , m2 , m3 , m4 ) and the

damping and stiffness matrices are
 
D0 + D1 −D1 0 0
 −D1 D1 + D2 −D2 0 
D= 
 (9.34)
0 −D2 D2 + D3 −D3 
0 0 −D3 D3 + D4
 
K0 + K1 −K1 0 0
 −K1 K1 + K2 −K2 0 
K= 
 (9.35)
0 −K2 K2 + K3 −K3 
0 0 −K3 K3 + K4
The total energy of the system is
4
X 2
X
1 1 2 1 1
V = mi ẋ2i + Ki (xi − xi+1 ) + K0 x20 + K4 x24 (9.36)
i=1
2 i=1
2 2 2

The
P4 time derivative of the energy for the solutions of the system will be the power
i=1 τ i vi supplied by the inputs τ i minus the power dissipated in the dampers. This is
written
X4 X3
1 1
V̇ = τ i vi − Di (ẋi − ẋi+1 ) − D0 ẋ20 − D4 ẋ24 (9.37)
i=1 i=1
2 2
Suppose that τ 2 = τ 3 = τ 4 = 0. Then the system with input τ 1 and output v1 will be
passive.

9.3 Vibrating string

9.3.1 Linearized model

wx, t
dx x
x 0 x L

Figure 9.3: Vibrating string of length L.

366 CHAPTER 9. MECHANICAL VIBRATIONS

f P  dP
  d
 dx
P

Figure 9.4: Differential string element.

A stretched string with tension P is fixed at x = 0 and x = L (Figure 9.3). The string
is supposed to have small transverse displacements w(x, t), and is excited by a transverse
force f (x, t). The equation of motion for a differential element of the string as shown in
Figure 9.4 is
∂2w
ρdx 2 = (P + dP ) sin(θ + dθ) − P sin θ + f dx (9.38)
∂t
where ρdx is the mass of the element, and θ is the slope of the string. Division with dx
leads to the partial differential equation
∂2w ∂
ρ = (P sin θ) + f (9.39)
∂t2 ∂x
For small angles we may approximate the sine function by
∂w
sin θ = (9.40)
∂x
which gives µ ¶
∂2w ∂ ∂w
ρ 2
= P +f (9.41)
∂t ∂x ∂x
If the tension is constant along the string, then P is a constant and the model is
∂2w ∂2w
ρ = P +f (9.42)
∂t2 ∂x2
The homogeneous form
∂2w ∂2w
2
= c2 2 (9.43)
∂t ∂x
of (9.42) is called the wave equation where
s
P
c= (9.44)
ρ

c is the propagation speed of the waves.

9.3.2 Orthogonal shape functions

The homogeneous problem as given by the wave equation (9.43) can be solved by separa-
tion of variables w(x, t) = q(t)φ(x). Then, according to basic textbooks on mathematics,
the following expression is found
q̈(t) c2 φ00 (x)
= = −ω 2 (9.45)
q(t) φ(x)
9.3. VIBRATING STRING 367

where ω is a constant to be determined. This leads to the differential equations

q̈(t) + ω 2 q(t) = 0 (9.46)
ω2
φ00 (x) + 2 φ(x) = 0 (9.47)
c
We investigate the solution of the equation of the vibrating string when the end-points
are fixed, that is, when w(0, t) = w(L, t) = 0, which implies that φ(0) = φ(L) = 0. Then
there are infinitely many solutions
r
2 kπ
φk (x) = sin x, k = 1, 2, . . . (9.48)
ρL L
and each solution φk (x) corresponds to the value
kπ
ωk = c (9.49)
L
and the solution qk of
q̈k + ω 2k qk = 0 (9.50)
The solution of the wave equation can then be written
∞
X
w(x, t) = qk (t)φk (x) (9.51)
k=1

Note that the shape functions are orthogonal in the sense that
Z L
ρφj (x)φk (x)dx = δ jk (9.52)
0

Moreover, the derivatives of the mode shapes are orthogonal, so that

Z L µ ¶2
kπ
φ0j (x)φ0k (x)dx = δ jk (9.53)
0 L

9.3.3 Galerkin’s method for orthogonal shape functions

The method of separation of variables works well for the wave equation when there
is no forcing term f . If there is a forcing term, then P∞ Galerkin’s method can be used
(Joshi 1989). In this method a solution w(x, t) = j=1 qj (t)φj (x) is assumed where
φj (x) belong to some set of shape functions. Orthogonal shape functions as found by
the separation of variables will be assumed. The equation of motion is obtained by
multiplying the wave equation by a shape function φi (s) and then integrating over the
interval. This gives
Z L ∞
X Z L
£ 2 00
¤
φi (x) ρ q̈j (t)φj (x) − c qj (t)φj (x) dx = φi (x)f (x, t)dx (9.54)
0 k=1 0

It is standard procedure to use partial integration for the term

Z L Z L
¯
φi (x)φ00j (x)dx = − φ0i (x)φ0j (x)dx + φi (x)φ0j (x)¯0
0 0
Z L
= − φ0i (x)φ0j (x)dx (9.55)
0
368 CHAPTER 9. MECHANICAL VIBRATIONS

x
h

-2 -1 0 1 2

Figure 9.5: Triangular shape functions ψ j (x).

Due to the orthogonality of the shape functions and the derivatives this simplifies to
Z L
q̈i (t) + ω 2i qi (t) = φi (x)f (x, t)dx (9.56)
0

Example 146 A point force F (t) at xF can be modelled as f (x) = F (t)δ(x − xF ). In

that case the equation of motion becomes

q̈i (t) + ω 2i qi (t) = φi (xF )F (t) (9.57)

We see that if φi (xF ) = 0 for some k, then the force F (t) will have no influence on qi (t).
In control terminology, this means that qi is not controllable when F is the control input.

9.3.4 Finite element shape functions

Instead of the orthogonal shape functions another set of shape functions will be intro-
duced in this section. This is done by using shape functions ψ j (x) leading to a finite-
element model. First we define N points along the string that are called nodes. The
position of node j is denoted xj . The distance between the nodes is set to h. The N
piece-wise linear shape functions
 x−xj−1
 h , xj−1 ≤ x ≤ xj
xj+1 −x
ψ j (x) = , xj ≤ x ≤ xj+1 , j = 1, 2, . . . , N (9.58)
 h
0, otherwise

that are shown in Figure 9.5 are used, and the solution is approximated by
N
X
w(x, t) = qj (t)ψ j (x) (9.59)
j=1

Note that
ψ j (xk ) = δ jk (9.60)
which implies that the displacement at node k is given by

w(xk , t) = qk (t) (9.61)

9.3. VIBRATING STRING 369

We insert the approximation into the wave equation (9.42) and get
N
X £ ¤
ρq̈j (t)ψ j (x) − P qj (t)ψ 00j (x) = f (x, t) (9.62)
k=1

We apply Galerkin’s method, which in this case involves the multiplication with ψ i (x)
and integration over the interval. This gives
Z L N
X Z L
£ 00
¤
ψ i (x) ρq̈j (t)ψ j (x) − P qj (t)ψ j (x) dx = ψ i (x)f (x, t)dx (9.63)
0 k=1 0

Partial integration gives

N Z
X L N
X Z L Z L
ρψ i (x)ψ j (x)dxq̈j (t) + P ψ 0i (x)ψ 0j (x)dxqj (t) = ψ i (x)f (x, t)dx (9.64)
k=1 0 k=1 0 0

This may be written in matrix form as

Mq̈ + Kq = F (9.65)

where q = (q1 , q2, . . . , qN )T , M = {mij }, K = {kij } and F =(f1 , f2, . . . , fN )T . The

components are given by
Z L
mij = ρ ψ i (x)ψ j (x)dx (9.66)
0
Z L
kij = P ψ 0i (x)ψ 0j (x)dx (9.67)
0
Z L
fi = ψ i (x)f (x, t)dx (9.68)
0

9.3.5 String element

The usual way of establishing a finite element model for this system is to define a string
element, and then to generate a model for the string by assembling string elements. The
string element is a model of a string of length h between two nodes. The elements of the
mass matrix Me of the element is then
Z h³
x ´2 ρh
m11 = ρ 1− dx = , m22 = m11 (9.69)
0 h 3
Z ³
h
x´ x ρh
m12 = m21 = ρ 1− dx = (9.70)
0 h h 6
while the stiffness matrix Ke of the element has elements
Z hµ ¶2
−1 P
k11 = P dx = , k22 = k11 (9.71)
0 h h
Z h
−1 P
k1,2 = k1,2 = P dx = − (9.72)
0 h2 h
370 CHAPTER 9. MECHANICAL VIBRATIONS

The model for the string element is found to be

Me q̈e + Ke qe = Fe (9.73)

where µ ¶ µ ¶
qe1 Fe1
qe = , Fe = (9.74)
qe2 Fe2
µ ¶ µ ¶
ρh 2 1 P 1 −1
Me = , Ke = (9.75)
6 1 2 h −1 1

9.3.6 Assembling string elements

Two string elements can be assembled by specifying that the second coordinate of the
first element equals the first coordinate of the second element. This is done by assigning
the coordinates to be q1 and q2 for the first element and q2 and q3 for the second element.
(c)
To keep the two elements together there must be a constraint force F2 . The models for
the two elements are
µ ¶ µ ¶ µ ¶
q̈1 q1 F1
Me + Ke = 1 (c) (9.76)
q̈2 q2 2 F2 + F2
µ ¶ µ ¶ µ (c)
¶
q̈2 q2 1
Me + Ke = 2 F2 − F2
(9.77)
q̈3 q3 F3
The models of the elements are assembled by cancelling the constraint force, which is
done by adding the last line of (9.76) with the first line of (9.77). This gives
     
q̈1 q1 F1
M  q̈2  + K  q2  =  F2  (9.78)
q̈3 q3 F3
   
2 1 0 1 −1 0
P
M =  1 4 1  , K =  −1 2 −1  (9.79)
h
0 1 2 0 −1 1
Note that the mass matrix is obtained by adding the mass matrices of the elements in
the sense that µ ¶ µ ¶
Me 0 0 0T
M= + (9.80)
0T 0 0 Me
In the same way the stiffness matrix is obtained from
µ ¶ µ ¶
Ke 0 0 0T
K= + (9.81)
0T 0 0 Ke

This procedure can be repeated to assemble more elements. For 5 nodes the mass
and stiffness matrix will be
   
2 1 0 0 0 1 −1 0 0 0
 1 4 1 0 0   −1 2 −1 0 0 
ρh  
 0 1 4 1 0 , K = P  0
 

M= −1 2 −1 0 (9.82)
6 
 0 0 1 4 1 
 h 
 0 0 −1 2 −1


0 0 0 1 2 0 0 0 −1 2
9.4. NONLINEAR STRING DYNAMICS 371

Example 147 The mass matrix as given by (9.66) is tridiagonal. It is called the con-
sistent mass matrix as it is derived using the shape functions. It is possible to have a
simpler model by using a lumped mass model where the mass is lumped at the nodes. This
leads to a diagonal mass matrix, which is called the nonconsistent mass matrix. In the
case of 5 nodes the nonconsistent mass matrix is
 1 
2 0 0 0 0
 0 1 0 0 0 
ρ  
M= 2  0 0 1 0 0   (9.83)
h 
0 0 0 1 0 
0 0 0 0 12

Then the model Mq̈ + Kq = F describes the dynamics of a mass spring damper arrange-
ment consisting of N masses that are interconnected by springs in a serial arrangement.
When the consistent mass matrix is used there will in general be spring connections be-
tween all masses.

9.4 Nonlinear string dynamics

9.4.1 Kirchhoff’s nonlinear string model
If the elastic deformation w is sufficiently large, then the tension will depend on the
deformation. Let the tension be constant along the string and given by P = P0 + EI∆x
where s
Z L µ ¶2 Z µ ¶2
∂w 1 L ∂w
∆x = 1+ dx − L ≈ dx (9.84)
0 ∂x 2 0 ∂x
is the stretching of the string due to w. Then the homogeneous string model (9.43)
becomes à ¶2 ! 2
Z Lµ
∂2w ∂w ∂ w
ρ 2 = P0 1 + EI dx (9.85)
∂t 0 ∂x ∂x2

This is Kirchhoff’s nonlinear string model (Shahruz 1999).

9.4.2 Marine cables

Tail x Towing vessel

buoy
z r
Streamers

t n
Depth controllers Lead-in
cable

Figure 9.6: Towing arrangement for a marine seismic cable.

372 CHAPTER 9. MECHANICAL VIBRATIONS

In this section the equation of motion for a towed marine cable will be presented. The
derivation relies results that will be presented in Section 10.2. In particular this involves
the definition of the material derivative D/Dt and the concept of material coordinates.
Towed marine cables are used in marine seismic operation to map oil and gas reservoirs
in an arrangement as indicated in Figure 9.6 Moreover, the model that will be presented
is also valid for the dynamics of anchor lines.
The position of a point on a cable is described by the spatial length p to the point
along the stretched cable, by the material length s to the point along the undeformed
cable, and the spatial position r(s, t) given in a spatial frame (x, y, z). The length p is
referred to as the length in spatial coordinates, while the length s is said to be the length
in material coordinates. The displacement of a point from its undeformed position is
denoted u = p − s. The material strain η is defined by
du dp dp
η= = −1⇒ =η+1 (9.86)
ds ds ds
We consider the cable element from s to s+ds of the unstretched cable. We call this a
material cable element as it contains the same material points as the cable is moved and
stretched. In the stretched case this cable element is from p to p + dp. The material cable
element is of constant length ds in material coordinates, and of length dp = (1 + η)ds in
spatial coordinates. The unit tangent vector of the stretched cable is
∂r ∂r ds 1 ∂r
t= = = (9.87)
dp ds dp 1 + η ∂s
As the tangent is a unit vector it can be written
¯ ¯−1
∂r ¯¯ ∂r ¯¯
t= (9.88)
ds ¯ ∂s ¯
This shows that the material strain η is given by
¯ ¯
¯ ∂r ¯
η = ¯¯ ¯¯ − 1 (9.89)
∂s
The volume of a material cable element is A0 ds = Adp where A0 is the cross section
of the unstretched cable, and A is the cross section of the stretched cable. The mass of
a material cable element is constant and is given by dm = ρ0 ds where ρ0 is the mass
per unit length of the unstretched cable. A detailed discussion on material and spatial
coordinates for this problem is found in (Lin and Segel 1974). The velocity v(s, t) and
the acceleration a(s, t) of a point on the cable are given by
Dr(s, t) ∂r(s, t) Dv(s, t) ∂ 2 r(s, t)
v(s, t) = = , a(s, t) = = (9.90)
Dt ∂t Dt ∂t2
The equation of motion for the material cable element ds is given by
D Dv ∂ 2 r(s, t)
(vρ0 ds) = ρ0 ds = ρ0 ds = (P +dP )t(t, s+ds)−P t(t, s)+f (t, s)dp (9.91)
Dt Dt ∂t2
where it is used that ρ0 ds is a constant. Here P is the tension, t is the tangent vector
along the cable, and f is the force per unit length of the stretched cable. Dividing by ds
we get
∂ 2 r(s, t) ∂ dp
ρ0 2
= (P t) + f (9.92)
∂t ∂s ds
9.5. EULER BERNOULLI BEAM 373

According to Hooke’s law the tension in the cable is P = EA0 η where E is the Young’s
modulus, and A0 is the cross section of the unstretched cable. The force due to the
tension P in the cable can be separated into a force along the tangent t and one force
orthogonal to the tangent according to
∂ 2 r(s, t) ∂P ∂t dp ∂η ∂t dp dp
ρ0 2
= t+P +f = EA0 t + P +f (9.93)
∂t ∂s ∂s ds ∂s ∂p ds ds
This gives
∂ 2 r(s, t) ∂2u ∂t
ρ0 = EA0 t + P (1 + η) + f (1 + η) (9.94)
∂t2 ∂s2 ∂p
Example 148 In (Aamo and Fossen 2000) the formulation
µ ¶
∂2r ∂ η ∂r
ρ0 2 = EA0 + f (1 + η) (9.95)
∂t ∂s 1 + η ∂s
due to (Triantafyllou 1990) was used to develop a finite-element model of anchor lines
for moored offshore vessels using Galerkin’s method with
N
X
r(t, s) = qk (t)ψ k (s) (9.96)
k=1

The shape functions ψ k (s) were selected as the hat functions shown in Figure 9.5.

9.5 Euler Bernoulli beam

9.5.1 Model
z

w
x
L

Figure 9.7: Euler Bernoulli beam.

An Euler Bernoulli beam is a mathematical model of lateral elastic deformations in

a slender beam. The length coordinate along the beam is denoted x, and the elastic
deformation in the z direction is denoted w(x, t) as shown in Figure 9.7. The model is
derived from the equations of motion for a volume element of length dx. The bending
moment is denoted M (x, t), and the shear force is denoted V (x, t) (Figure 9.8). The
momentum balance for the Euler Bernoulli beam is given by the two equations
∂2w
− (V + dV ) + f dx + V = ρdx (9.97)
∂t2
(M + dM ) − M − (V + dV ) dx = 0 (9.98)

where f is the external force per unit length of the beam. We see that the moment of
inertia is set to zero in the moment equation. This is due to the assumption that the
374 CHAPTER 9. MECHANICAL VIBRATIONS

f
M  dM
V


w V  dV
M
x
x  dx

Figure 9.8: Differential element of beam.

beam is slender. A more elaborate model for thick beams is the Timoshenko beam which
includes a nonzero moment of inertia. Division of the equations by dx leads to
∂V ∂2w
− +f = ρ (9.99)
∂x ∂t2
∂M
= V (9.100)
∂x
Combining these two equations we get
∂2M ∂2w
− + f = ρ (9.101)
∂x2 ∂t2
For the Euler Bernoulli beam the bending moment is given by the constitutive equation
∂2w
M (x, t) = EI(x) (9.102)
∂x2
where E is Young’s modulus, and I(x) is the moment of inertia about the y axis. It is
noted that (9.102) implies that the shear force is
∂3w
V = EI(x) (9.103)
∂x3
Combination of (9.101) and (9.102) gives the partial differential equation
· ¸
∂ 2 w(x, t) ∂2 ∂ 2 w(x, t)
ρ(x) + EI(x) = f (x, t) (9.104)
∂t2 ∂x2 ∂x2
For the case where the moment of inertia I(x) is a constant and the external force f (x, t)
is zero the following result is obtained:

A homogeneous Euler Bernoulli beam is described by the partial differential equation

∂2w ∂4w
2
+ c2 4 = 0 (9.105)
∂t ∂x
where
EI
c2 = (9.106)
ρ
9.5. EULER BERNOULLI BEAM 375

9.5.2 Boundary conditions

Clamped - free Clamped - pinned

Free - free Pinned - free

Clamped - clamped Pinned - pinned

Figure 9.9: Boundary conditions for an Euler Bernoulli beam.

Typical boundary conditions for the Euler Bernoulli beam as shown in Figure 9.9 are:

1. A clamped end is defined to have zero elastic deformation and zero elastic angle.
This means that
∂w
w = 0 and =0 (9.107)
∂x
2. A free end has zero bending moment and zero shear force. From (9.102) and (9.103)
this is seen to imply that
∂2w ∂3w
= 0 and =0 (9.108)
∂x2 ∂x3
3. A pinned end has zero elastic deformation and zero bending moment, so that
∂2w
w = 0 and =0 (9.109)
∂x2

4. An end with a point mass m will have zero bending moment, and a shear force
∂2w
V =m (9.110)
∂t2
This gives the boundary conditions
∂2w ∂3w ∂2w
= 0 and EI = m (9.111)
∂x2 ∂x3 ∂t2
5. An end point clamped to a rigid body with mass m and moment of inertia J will
have a shear force and bending moment given by
∂2w ∂ 2 ∂w
V =m 2
, M = −J 2 (9.112)
∂t ∂t ∂x
From (9.102) and (9.103) this is seen to imply that

∂2w ∂ 2 ∂w ∂3w ∂2w

EI = −J and EI = m (9.113)
∂x2 ∂t2 ∂x ∂x3 ∂t2
376 CHAPTER 9. MECHANICAL VIBRATIONS

9.5.3 Energy
For an Euler Bernoulli beam the kinetic energy is
Z
1 L
T = ρ(x)ẇ2 dx (9.114)
2 0
while the potential energy is
Z L
1
U= EI(x)(w00 )2 dx (9.115)
2 0

The total energy is then

Z L Z L
1 1 2 2
W = ρ(x)ẇ dx + EI (w00 ) dx (9.116)
2 0 2 0

and the time derivative of the energy along the solutions of the system is found to be
Z L
Ẇ = (ẇρẅ + w00 EI ẇ00 ) dx
0
Z L
= EI (−ẇw0000 + ẇf + w00 ẇ00 ) dx
0
Z L
L
= − ẇEIw000 |0 + EI (ẇf + ẇ0 w000 + w00 ẇ00 ) dx
0
Z L
L L
= − ẇEIw000 |0 + ẇ 0
EIw00 |0 + EI (ẇf − ẇ00 w00 + w00 ẇ00 ) dx
0
Z L
L
= − ẇV |L 0
0 + ẇ M |0 + EI ẇf dx (9.117)
0

The total energy of the Euler Bernoulli beam is

Z L Z L
1 1 2
W = ρ(x)ẇ2 dx + EI (w00 ) dx (9.118)
2 0 2 0

The time derivative along the solutions of the system is

Z L
L L
Ẇ = − ẇV |0 + ẇ0 M |0 + EI ẇf dx (9.119)
0

9.5.4 Orthogonal shape functions

For a homogeneous Euler Bernoulli beam with zero external force the method of sepa-
ration of variables leads to a very useful description based on orthogonal mode shapes.
This description will be developed in the following, and it will serve as a starting point
to explain transfer function models and finite element models that will be presented at
a later stage.
In the method of separation of variables the expression

w(x, t) = φ(x)q(t) (9.120)

9.5. EULER BERNOULLI BEAM 377

is inserted into the partial differential equation (9.105). This gives

φ(x)q̈(t) + c2 φ0000 (x)q(t) = 0 (9.121)
Following the standard procedure this equation is reformulated as
c2 φ0000 (x) q̈(t)
=− =C (9.122)
φ(x) q(t)
where it is seen that C must be a constant as it is a function of t alone and at the same
time a function of x alone. The only nontrivial solutions for φ are found for C > 0.
Therefore the constant ω is introduced so that C = ω 2 . This gives the two ordinary
differential equations
q̈(t) + ω 2 q(t) = 0 (9.123)
0000
φ (x) − β 4 φ(x) = 0 (9.124)
where the constant
ω2
β4 = (9.125)
c2
has been introduced.
The first equation (9.123) is recognized as a harmonic oscillator. The solution of the
second equation (9.124) is given by
φ(x) = C1 cos βx + C2 sin βx + C3 cosh βx + C4 sinh βx (9.126)
which has derivatives
φ0 (x) = β (−C1 sin βx + C2 cos βx + C3 sinh βx + C4 cosh βx) (9.127)
φ00 (x) = β 2 (−C1 cos βx − C2 sin βx + C3 cosh βx + C4 sinh βx) (9.128)
φ000 (x) = β 3 (C1 sin βx − C2 cos βx + C3 sinh βx + C4 cosh βx) (9.129)
Depending on which boundary conditions that apply for the beam, there will be condi-
tions on φ and its derivatives at x = 0 and x = L. The boundary conditions can then be
used to determine the constants Ci through the equations
    
φ(0) 1 0 1 0 C1
0
 β1 φ (0)   0 1 0 1   
 1 00 =   C2  (9.130)
 β 2 φ (0)   −1 0 1 0   C3 
1 000 0 −1 0 1 C4
β3
φ (0)
and
    
φ(L) cos βL sin βL cosh βL sinh βL C1
1 0
 β φ (L)
  − sin βL sin βL cosh βL sinh βL   C2 
 1 00
=   (9.131)
 β2
φ (L)   − cos βL − sin βL cosh βL sinh βL   C3 
1 000 sin βL − cos βL sinh βL cosh βL C4
β 3 φ (L)

For a given set of boundary conditions there will be a set of shape functions φi (x) with
associated constants β i and natural frequencies ω i = cβ 2i so that the solution is given by
∞
X
w(x, t) = φi (x)qi (t) (9.132)
i=1
378 CHAPTER 9. MECHANICAL VIBRATIONS

Type of end Boundary Condition 1 Boundary Condition 2

Clamped φ=0 φ0 = 0
Free φ00 = 0 φ000 = 0
Pinned φ=0 φ00 = 0
Mass m φ00 = 0 φ000 = − m 4
ρβ φ
Mass m and inertia J φ00 = − Jρ β 4 φ0 φ000 = − m 4
ρβ φ

Table 9.1: Boundary conditions for Euler Bernoulli beam.

Here φi (x) and qi (t) satisfy

q̈i (t) + ω 2i qi (t) = 0 (9.133)

0000
φi (x) − β 4i φi (x) = 0 (9.134)

The boundary conditions on the deflection w(x, t) imply the boundary conditions
given in Table 9.1 on the shape functions. Given the boundary conditions, the constants
Ci are found by formulating equations for the boundary conditions according to

Bc = 0 (9.135)

where c = (C1 , C2 , C3 , C4 )T . Then, to have nontrivial solutions for the constants in the
vector c, it is necessary that
det B = 0 (9.136)

9.5.5 Clamped-free Euler Bernoulli beam

We will derive the solution for an Euler Bernoulli beam which is clamped at the end at
x = 0, and free at x = L. The boundary conditions are

φ(0) = 0, φ0 (0) = 0, (9.137)

φ00 (L) = 0, φ000 (L) = 0 (9.138)

This can be written

    
φ(0) 1 0 1 0 C1
 β1 φ0 (0)   0 1 0 1  C2 
0 =  
 β12 φ00 (L)  =  − cos βL − sin βL cosh βL sinh βL


 (9.139)

C3
1 000 sin βL − cos βL sinh βL cosh βL C4
β3
φ (L)
| {z }| {z }
Bcf c

Nontrivial solutions are found when det Bcf = 0, which occurs when

1 + cos βL cosh βL = 0 (9.140)

This transcendental equation holds for infinitely many discrete values of β, and must
be solved numerically. The solutions are denoted β i , and for each β i there corresponds
resonance frequency ω i = β 2i c in agreement with (9.125), and a set of constants C1i ,
9.5. EULER BERNOULLI BEAM 379

Type of beam Characteristic equation

Clamped-free cos βL cosh βL + 1 = 0
Free-free and clamped-clamped cos βL cosh βL − 1 = 0
Clamped-pinned and pinned-free cos βL sinh βL − sin βL cosh βL = 0
Pinned-pinned sin βL = 0

Table 9.2: Characteristic equations for Euler Bernoulli beam.

C2i , C3i , C4i , and one shape function φi (x). We note from the first and second row of
(9.139) that the constants are related by C3i = −C1i and C4i = −C2i , and that the third
and fourth row of (9.139) implies that C2i can be expressed by C1i according to

cos β i L + cosh β i L
C2i = αi C1i where αi = (9.141)
sin β i L + sinh β i L

This shows that the shape functions are given by

φi (x) = C1i [(cos β i x − cosh β i x) + αi (sin β i x − sinh β i x)] (9.142)

Usually the constant C1i is normalized so that

Z
2
ρ [φi (x)] dx = 1 (9.143)
0

The solution is then

∞
X
w(x, t) = qi (t)φi (x) (9.144)
i=1

where the generalized coordinate qi satisfy the differential equation

q̈i (t) + ω 2i qi (t) = 0 (9.145)

Numerical values for the first modes are tabulated in textbooks like (Rao 1990), and
for the first three modes we have

β1 = 1.875104, α1 = 0.7341 (9.146)

β2 = 4.694091, α2 = 1.0185 (9.147)
β3 = 7.854757, α3 = 0.9992 (9.148)

In this example
ω2 β2
= 22 = 6.25
ω1 β1
Numerical values for β are tabulated in many textbooks on vibrations for simple cases
like pinned-pinned, free-free, fixed-fixed, fixed-free, fixed-pinned, and pinned-free.
The characteristic equation for different beams are given in Table 9.2.
380 CHAPTER 9. MECHANICAL VIBRATIONS

m2
m1
J1

x
0 L

Figure 9.10: Satellite antenna.

9.5.6 Beam fixed to an inertia and a mass

The antenna boom on a small satellite is modelled as an Euler Bernoulli beam. The
beam is modelled to be clamped at the satellite end at x = 0 to an inertial load with
mass m1 and moment of inertia J, and fixed to a mass m2 at the end of the boom at
x = L (Figure 9.10). The boundary conditions are then

J1 4 0 m1 4
φ00 (0) = − β φ (0), φ000 (0) = − β φ(0) (9.149)
ρ ρ
m2 4
φ00 (L) = 0, φ000 (L) = − β φ(L) (9.150)
ρ

which can be written

 1 00
  
β2
φ (0) + Jρ1 β 2 φ0 (0) C1
 1 000
φ (0) + mρ1 βφ(0)   C2 
 
 β3
1 00  = Ba  
 C3  = 0 (9.151)
 β2
φ (L) 
1 000
β3
φ (L) + mρ2 βφ(L) C4

Nontrivial solutions are found for

det Ba = 0 (9.152)

In the study of satellite dynamics the concept of constrained modes are often used.
These are the modes that occur when the satellite is assumed to have infinite inertia and
mass. Therefore the constrained modes for this antenna is found when m1 and J are
assumed to approach infinity. The condition for nontrivial solutions in this case is
m2
1 + cos β cosh β + (cos β − sin β sinh β ) = 0 (9.153)
ρ

In contrast to this the unconstrained modes are the modes are found for finite m1
and J. For a small satellite we assume that m1 =78 kg, J = 4 kg m2 , = 6 m, m2 = 4
kg, EI = 28.69 N·m2 (10,000 lbf ·inch2 ), ρ = 1 kg/6 m = 0.17 kg/m. Then the numerical
values were found for the constrained case and the unconstrained case using Maple. The
results are shown in Table 9.3. It was found that the spacecraft with gravity boom will
have resonances with periods 41.0 s, 1.14 s, 0.34 s,. . . The computations based on the
assumption of constrained modes predicts periods 20.2 s, 1.10 s, 0.34 s. It is seen that
the lowest natural frequency computed for the constrained modes is a factor of 2 greater
than the natural frequency associated with the unconstrained mode for this example.
9.5. EULER BERNOULLI BEAM 381

Constrained modes Unconstrained modes

ω 1 ω 1
Mode # β 2π (Hz) f (s) β 2π (Hz) f (s)
0 - - - 0 0 ∞
1 0.1529 0.0488 20.2 0.1082 0.0244 41.0
(9.154)
2 0.6592 0.9074 1.10 0.6487 0.8787 1.14
3 1.1810 2.9125 0.34 1.1782 2.8987 0.34
4 1.7037 6.0610 0.17 1.7025 6.0525 0.17
5 2.2268 10.354 0.10 2.2262 10.3487 0.10

Table 9.3: Natural frequencies for satellite antenna using constrained modes and uncon-
strained modes.

9.5.7 Orthogonality of the eigenfunctions

For specified boundary conditions there will be a set of solutions φi (x) of (9.124), where
each solution φi (x) corresponds to a value β i for β so that

φ0000 4
i (x) = β i φi (x) (9.155)

The solutions φi (x) are called the eigenfunctions (9.124), and the associated values λi =
β 4i are the eigenvalues of the system. It is assumed in the following that the eigenvalues
are distinct. Note that for each eigenfunction φi (x) there is one natural frequency

ω i = cβ 2i (9.156)

of the harmonic oscillator (9.123).

Consider the eigenfunction φi (x) with eigenvalue β 4i and the eigenfunction φj (x) with
eigenvalue β 4j . Then
Z Z
φi (x)φ0000
j (x)dx = β 4j φi (x)φj (x)dx (9.157)
0 0

The integral on the left side can also be evaluated by partial integration twice to be
Z Z
¯ ¯
φi (x)φ0000
j (x)dx = φi (x)φ000
j (x) ¯ − φ0
i (x)φ00
j (x) ¯ + φ00i (x)φ00j (x)dx (9.158)
0 0
0 0

Combination of (9.157) and (9.158) gives

Z Z
¯ ¯
β 4j φi (x)φj (x)dx = φi (x)φ000 ¯ 0 00 ¯
j (x) 0 − φi (x)φj (x) 0 + φ00i (x)φ00j (x)dx (9.159)
0 0

In the same way, by interchanging i and j in the expression the result

Z Z
¯ ¯
β 4i φi (x)φj (x)dx = φj (x)φ000
i (x) ¯ − φ0
j (x)φ00
i (x)¯ + φ00i (x)φ00j (x)dx (9.160)
0 0
0 0

appears. From (9.159) and (9.160) it is seen that if

¯ ¯ ¯ ¯
φi (x)φ000 ¯ 0 00 ¯ 000 ¯ 0 00 ¯
j (x) 0 − φi (x)φj (x) 0 = φj (x)φi (x) 0 − φj (x)φi (x) 0 (9.161)
382 CHAPTER 9. MECHANICAL VIBRATIONS

the eigenfunctions will satisfy

Z
(β 4j − β 4i ) φi (x)φj (x)dx = 0 (9.162)
0

Due to the assumption β i 6= β j , this implies that

Z
φi (x)φj (x)dx = 0, i 6= j (9.163)
0

Moreover, from (9.159) and (9.160) it follows that

Z
φ00i (x)φ00j (x)dx = 0, i 6= j (9.164)
0

Usually the eigenfunctions are normalized so that the following result is valid:

The eigenfunctions of (9.124) are orthogonal shape functions in the sense that they satisfy
Z
ρφi (x)φj (x)dx = δ ij (9.165)
0

In addition they satisfy

Z
ρφ00i (x)φ00j (x)dx = β 4i δ ij (9.166)
0
where δ ij = 1 when i = j and δ ij = 0 when i 6= j.

9.5.8 Galerkin’s method for orthogonal mode shapes

We will now introduce a control force u in our model of a Euler Bernoulli beam. The
control force u is assumed to be perpendicular to the beam at position xu . This can be
modelled with the Dirac delta δ(x) in the partial differential equation:
∂4w ∂2w
ρc2 (x, t) + ρ (x, t) = δ(x − xu )u (9.167)
∂x4 ∂t2
Due to the forcing term δ(x − xu )u the method of separation of variables cannot be used
directly.
The solution w(x, t) is assumed to be a linear combination of the eigenfunctions φj (x),
j ∈ {1, 2, . . .}, and we may write
∞
X
w(x, t) = qj (t)φj (x) (9.168)
j=1

The partial differential equations can then be written

X∞
∂ 4 φj (x) ∂ 2 qj (t)
[ρc2 qj (t) + ρ φj (x)] = δ(x − xu )u(t) (9.169)
j=1
∂x4 ∂t2

The partial differential equation is reformulated by insertion of (9.134) and ω 2j = c2 β 4j ,

which gives
X∞
ρφj (x)[ω 2j qj (t) + q̈j (t)] = δ(x − xu )u(t) (9.170)
i=1
9.5. EULER BERNOULLI BEAM 383

At this point Galerkin’s method is used. This involves the multiplication of equation
(9.170) with φi (x) and the integration of the result over the interval x ∈ [0, ]. This gives
Z ∞
X Z
φi (x) ρφj (x)dx[ω 2j qj (t) + q̈j (t)] = φj (x)δ(x − xu )dxu(t) (9.171)
0 j=1 0

R
For any function f (x) the Dirac delta gives 0 f (x)δ(x − xu )dx = f (xu ). This together
with the orthogonality of the eigenfunctions, see (9.165), lead to

ω 2i qi (t) + q̈i (t) = φi (xu )u(t) (9.172)

which has the Laplace transform

φi (xu )
qi (s) = u(s) (9.173)
ω 2i + s2

We assume that the measurement y(t) is the velocity ẇ(xy , t) of the elastic deflection
at position xy , that is,
y(t) = ẇ(xy , t) (9.174)
Then the measurement can be written
∞
X
y(t) = q̇i (t)φi (xy ) (9.175)
i=1

and the transfer function from u to y is seen to be

∞
y(s) X sφi (xy )φi (xu )
= (9.176)
u(s) i=1
ω 2i + s2

The following observations are important:

1. If input and output are collocated, which is the case whenever xu = xy = x0 , then
the transfer function from u to y is passive because
∞
Y (s) X sφ2i (x0 )
= (9.177)
U (s) ω 2 + s2
i=1 i

which is a parallel interconnection of passive systems. Alternatively passivity may

be established from energy considerations. This result agrees with an energy argu-
ment where V is the sum of kinetic and potential energy. Then

V̇ (t) = y(t)u(t) − d(t) (9.178)

RT
where 0 d(t)dt ≥ 0 is the dissipated energy in the system. It follows that the
system with input u and output y is passive.

2. If measurement and control are not collocated, then xu 6= xy , and it may be

that φi (xy ) and φi (xu ) have opposite signs for certain i. In this case the transfer
function y(s)/u(s) will not be positive real. This may cause difficulties in designing
a controller to damp out vibrations.
384 CHAPTER 9. MECHANICAL VIBRATIONS

3. If φi (xu ) = 0, the control u will have no influence on mode i. In the state-space

terminology this means that mode i is not controllable with the control u.

4. If φi (xy ) = 0, then mode i will not be noticeable in the measurement y. This means
that mode i is not observable from the measurement y.

Example 149 Consider a homogeneous beam of aluminium with a rectangular cross

section, length = 2 m, width b = 0.05 m, height h = 0.01 m, density ρ = b · h · 2700
kg/m3 = 1.35 kg/m, Young’s modulus E = 70 · 109 N/m2 and moment of inertia I =
bh3 /12 = 4.167 · 10−9 m4 . The beam is clamped at x = 0 and free at x = .
The shape functions can be found to be

φ1 (x) = −0.6086 · {cos(β 1 x) − cosh(β 1 x) − 0.7341 · [sin(β 1 x) − sinh(β 1 x)]} (9.179)

and

φ2 (x) = −0.6086 · {cos(β 2 x) − cosh(β 2 x) − 1.0185 · [sin(β 2 x) − sinh(β 2 x)} (9.180)

First collocation is tried with xu = xy = 2 m. Then

φ1 (xu ) = φ1 (xy ) = 1.22, φ2 (xu ) = φ2 (xy ) = −1.22

and the transfer function

y 1.5s 1.5s s(57.42 + s2 )

(s) = 2 2
+ 2 2
= 3.0 (9.181)
u 12.8 + s 80.1 + s (12.8 + s2 )(80.12 + s2 )
2

results, which is passive. Note that the complex conjugated zeros at s = ±j57.4 is located
between the poles in s = ±j12.8 and s = ±j80.1. A simple P controller

u = −Kp y (9.182)

will give stability, with a power dissipation of u(t)y(t) = −Kp y(t)2 . The gain Kp is only
limited by noise, quantization and discretization effects.
Next noncollocation is tried with xu = 0.5 m and xy = 2 m. Then

φ1 (xu ) = 0.12 φ1 (xy ) = 1.22

φ2 (xu ) = 0.51 φ2 (xy ) = −1.22

and the transfer function is

y 0.15s 0.62s s(47.62 − s2 )

(s) = 2 2
− 2 2
= −0.47 (9.183)
u 12.8 + s 80.1 + s (12.8 + s2 )(80.12 + s2 )
2

This transfer function has a zero in the right half plane at s = 47.6. This limits the
bandwidth of the system. Alternatively, we see that the transfer function is the sum of
two transfer functions that are not very different, except that they have opposite signs.
Thus if a P controller is used in a negative feedback, this will give stabilization and power
dissipation for mode 1, while it will give destabilization and added power to mode 2. In
fact the only possibility for stabilization is that the mode with positive feedback has gain
less that unity which ensures stability according to Bode-Nyquist stability theory.
9.6. FINITE ELEMENT MODEL OF EULER BERNOULLI BEAM 385

9.6 Finite element model of Euler Bernoulli beam

9.6.1 Introduction

An alternative technique for analyzing the Euler Bernoulli beam is to use the finite-
element method. The finite-element method can be seen as a model formulation based
on the Galerkin method, where special set of shape functions are used. The characteristic
feature of the finite-element method is that the shape functions are locally defined in the
sense that they are nonzero only in short intervals of the beam. This is in contrast to the
orthogonal shape functions, which are global function on the beam. An alternative way
of seeing the finite-element method is that the beam is divided into beam elements. The
equations of motion are then derived for the beam element using a cubic shape function,
and then the beam model is obtained by connecting the beam element models using
multiport techniques. The presentation that follows will focus on the formulation using
beam elements, but the Galerkin point of view will also be presented.

9.6.2 Beam element

In a finite-element model of an Euler Bernoulli beam the basic building block of the
model is a beam element of length h. The element is defined for the interval 0 ≤ x ≤ h.
At x = 0 the shear force is V1 and the bending moment is M1 , the elastic displacement
is w1 , and the elastic angle is w10 . This can be seen as one port with effort V1 and flow
ẇ1 , and one port with effort M1 and flow ẇ10 . At x = h the shear force is V2 , the bending
moment is M2 , the elastic deflection is w2 , and the elastic angle is this is w20 . This is
described as a port with effort V2 and flow ẇ2 , and one port with effort M2 and flow ẇ20 .
The usual finite-element model of the Euler Bernoulli beam is based on the displacement
formulation where the inputs to the model are the forces and torques, and the outputs
are the displacements and the displacement angles. In the multiport terminology this is
an admittance model where the efforts are input and the flows are outputs.
The displacement in the element is modeled as the cubic expression

w(x, t) = c0 (t) + c1 (t)x + c2 (t)x2 + c3 (t)x3 (9.184)

The motivation for using this expression is that in the stationary case the displacement
satisfies w0000 = 0, which has solution (9.184). The generalized coordinates ai (t) of the
beam element are defined as

a1 (t) = w1 (t) a2 (t) = w10 (t) (9.185)

a3 (t) = w2 (t), a4 (t) = w20 (t) (9.186)

Combination of (9.184), (9.185) and (9.186) leads to

4
X
w(x, t) = αi (x)ai (t) (9.187)
i=1
386 CHAPTER 9. MECHANICAL VIBRATIONS

where the shape functions αi (x) are given by

³ x ´2 ³ x ´3
α1 (x) = 1 − 3 +3 (9.188)
·³ ´h ³ ´ h ³ ´ ¸
x x 2 x 3
α2 (x) = h −2 + (9.189)
h h h
³ x ´2 ³ x ´3
α3 (x) = 3 −2 (9.190)
· h ³ ´ h³ ´ ¸
x 2 x 3
α4 (x) = h − + (9.191)
h h

These cubic shape functions are called the Hermitian shape functions.
Galerkin’s method for the beam element leads to

Me ä + Ke a = f (9.192)

where the mass matrix of the element is given by

 
Z h 156 22h 54 −13h
T ρh 
 22h 4h2 13h −3h2 

Me = ραα dx = (9.193)
0 420  54 13h 156 −22h 
−13h −3h2 −22h 4h2

and the stiffness matrix of the element is given by

 
Z h 6 3h −6 3h
00 T 2c2 ρ  3h 2h 2
−3h h2 
Ke = ρα (α ) dx = 3 
00


 (9.194)
0 h −6 −3h 6 −3h
3h h2 −3h 2h2

and f = (f1 , f2 , f3 , f4 )T where

Z h
fi = αi (x)f (x) (9.195)
0

Example 150 To simplify the model the mass is sometimes lumped to have a diagonal
mass matrix (Rao 1990). For translational degrees of freedom this is simply done by
lumping all mass at the nodes, while for rotational inertia the inertia can be computed
by having uniform mass distribution for half a beam on each side of the node. This gives
in this example
 
1 0 0 0
2
ρh  h
 0 12 0 0 

Me =  (9.196)
2 0 0 1 0 
2
0 0 0 h12
which is the lumped mass matrix.

9.6.3 Assembling a structure

To establish the model for a beam of length L where L = N h it is necessary to connect N
beam elements. Elements k and k + 1 can be connected by requiring that the end-point
variables satisfy ak,3 = ak+1,1 and ak,4 = ak+1,2 . Then, there must be forces and torques
9.6. FINITE ELEMENT MODEL OF EULER BERNOULLI BEAM 387

of constraints to hold the two element together, and the equations of motion for elements
k and k + 1 are given by
     
ak,1 ak,1 fk,1
d2  a   ak,2   fk,2 
Me 2  k,2  + Ke   =  (c)  (9.197)
dt  ak,3   ak,3   fk,3 + f3 
ak,4 ak,4 (c)
fk,4 + f4
     (c) 
ak+1,1 ak+1,1 fk+1,1 − f3
d2  a   ak+1,2  
 =  fk+1,2 − f4 
(c) 
Me 2  k+1,2  + Ke  (9.198)
dt  ak+1,3   ak+1,3   fk+1,3 
ak+1,4 ak+1,4 fk+1,4

These forces and torques of constraint are eliminated by adding rows 3 and 4 of element
k to rows 1 and 2 of element k + 1. This gives the model

Mq̈ + Kq = u (9.199)

where q = (ak,1 , ak,2 , ak+1,1 , ak+1,2 , ak+1,3 , ak+1,4 ). The mass matrix is obtained from
µ ¶ µ ¶
Me 04,2 02,2 02,4
M= + (9.200)
02,4 02,2 04,2 Me

In the same way the stiffness matrix is obtained from

µ ¶ µ ¶
Ke 04,2 02,2 02,4
K= + (9.201)
02,4 02,2 04,2 Ke

Alternatively, the model of the two elements can be written

d2
M̄ ā + K̄ā = f̄ (9.202)
dt2

ā = (a1 , . . . , ap )T (9.203)
M̄ = block diag(Me1 , . . . , Mep ), K̄ = block diag(Ke1 , . . . , Kep ) (9.204)
where the connection of the elements is obtained by requiring

ā = Cq, u = CT f̄ (9.205)

where  
1 0 0 0 0 0
 0 1 0 0 0 0 
 
 0 0 1 0 0 0 
 
 0 0 0 1 0 0 
C=


 (9.206)
 0 0 1 0 0 0 
 0 0 0 1 0 0 
 
 0 0 0 0 1 0 
0 0 0 0 0 1
when N = 2. Then the mass matrix and the stiffness matrix are found from

M = CT M̄C, K = CT K̄C (9.207)

388 CHAPTER 9. MECHANICAL VIBRATIONS

to be  
m11 m12 m13 m14 0 0
 m21 m22 m23 m24 0 0 
 
 m31 m32 m33 + m11 m34 + m12 m13 m14 
M=


 (9.208)
 m41 m42 m43 + m21 m44 + m22 m23 m24 
 0 0 m31 m32 m33 m34 
0 0 m41 m42 m43 m44
 
k11 k12 k13 k14 0 0
 k21 k22 k23 k24 0 0 
 
 k31 k32 k33 + k11 k34 + k12 k13 k14 
K=


 (9.209)
 k41 k42 k43 + k21 k44 + k22 k23 k24 
 0 0 k31 k32 k33 k34 
0 0 k41 k42 k43 k44
and the resulting model is
Mq̈ + Kq = u (9.210)

9.6.4 Finite element model and Galerkin’s method

A finite-element model for an Euler Bernoulli beam can alternatively be established
by applying Galerkin’s method with shape functions ψ i (x) based on the element shape
functions in (9.188—9.191). For the Euler Bernoulli beam, N nodes are defined at x1 <
x2 < . . . < xN , and the deflection is described by
N
X
w(x, t) = [αj,1 (x)aj,1 (t) + αj,2 (x)aj,2 (t)] (9.211)
j=1

which is expressed in the form

2N
X
w(x, t) = ψ j (x)qj (t) (9.212)
j=1

where the generalized coordinates are q = (a1,1 , a1,2 , . . . , aN,1 , aN,2 )T and the mode shape
vector is ψ =(ψ 1,1 , ψ 1,2 , . . . , ψ N,1 , ψ N,2 )T . The shape functions αj,1 (x) and αj,2 (x) for
the Euler Bernoulli beam are selected in agreement with (9.188—9.191) as the Hermitian
shape functions

 (x−xi )2 3

 1−3 2 + 2 (x−x3 i ) , xi ≤ x ≤ xi+1
i i
2 3
αi,1 (x) =
 3 (x−x2 i−1 ) − 2 (x−x3 i−1 ) , xi−1 ≤ x ≤ xi (9.213)

 i−1 i−1
0 otherwise
 2 3


(x−xi )
+ (x−x2 i ) , xi ≤ x ≤ xi+1
 x−2 i i
2
αi,2 (x) = (x−xi−1 ) (x−xi−1 )3 (9.214)
 − + 2 , xi−1 ≤ x ≤ xi


i−1 i−1
0, otherwise

These shape functions satisfy

ψ 2k−1 = αj,1 (xk ) = δ jk , ψ 02k−1 = α0j,1 (xk ) = 0 (9.215)

ψ 2k = αj,2 (xk ) = 0, ψ 02k = α0j,2 (xk ) = δ jk (9.216)
9.6. FINITE ELEMENT MODEL OF EULER BERNOULLI BEAM 389

1.2

1 ψ
1

0.8

0.6

0.4

0.2
ψ
2

-0.2
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5

Figure 9.11: Shape functions for Euler-Bernoulli beam.

This gives the following physical interpretation of the generalized coordinates q2k−1 =
ak,1 (t) and q2k = ak,2 (t):

q2k−1 = ak,1 (t) = w(xk , t) (9.217)

q2k = ak,2 (t) = w0 (xk , t) (9.218)

Insertion of (9.212) gives

2N
X £ ¤
ρq̈j (t)ψ j (x) + ρc2 qj (t)ψ 0000
j (x) = b(x)u(t) (9.219)
i=1

In the Galerkin method the equation of motion is premultiplied by ψ i (x) and integrated
over the interval x ∈ [0, ]. This gives the expression
Z 2N
X Z
ψ i (x) [ρc2 qj (t)ψ 0000
j (x) + ρq̈j (t)ψ j (x)]dx = ψ i (x)b(x)u(t)dx (9.220)
0 i=1 0

Partial integration twice gives

Z Z
¯ ¯
ψ i (x)ψ 0000
j (x)dx = ψ i (x)ψ 000 ¯
j (x) 0 − ψ 0i (x)ψ 00j (x)¯0 + ψ 00i (x)ψ 00j (x)dx (9.221)
0 0

and, if we assume that

¯ ¯
ψ i (x)ψ 000 ¯ 0 00 ¯
j (x) 0 − ψ i (x)ψ j (x) 0 = 0 (9.222)
390 CHAPTER 9. MECHANICAL VIBRATIONS

then
N
"Z Z # Z
X
ρψ i (x)ψ j (x)dxq̈i (t) + c2
ρψ 00i (x)ψ 00j (x)dxqj (t) = ψ i (x)b(x)dxu(t)
i=1 0 0 0

This can be written in vector form as

Mq̈ + Kq = bu (9.223)

Here q = (q1 , . . . , q2N ) T is the vector of generalized coordinates, M = {mij } is the mass
matrix and K = {kij} is the stiffness matrix, and b = (b1 , . . . , bN ) T with elements given
by
Z
mij = ρψ j (x)ψ i (x)dx (9.224)
0
Z
kij = c2 ρψ 00j (x)ψ 00i (x)dx (9.225)
0
Z
bi = ψ j (x)b(x)dx (9.226)
0

Note that

1. The mass matrix M and the stiffness matrix K are symmetric. Moreover, the M
and K matrices will typically be positive definite.

2. If the orthogonal mode shapes are used, then the mass matrix M and the stiffness
matrix K will be diagonal matrices.

9.7 Motor and Euler Bernoulli beam

9.7.1 Equations of motion

w x

m
Tm

Motor

Figure 9.12: Motor with Euler-Bernoulli beam

In this section we will study the dynamics of an Euler Bernoulli beam that is rotated
by a motor. This can be seen as a simplified model of a robotic joint with an elastic link.
The motor has moment of inertia Jm and joint angle θm , and the Euler Bernoulli beam
9.7. MOTOR AND EULER BERNOULLI BEAM 391

has length as shown in Figure 9.12. The position of a point x along the beam is given
by
η(x, t) = xθm (t) + w(x, t) (9.227)
where η(x, t) is the arc length from the reference position along a circle of radius x, and
w(x, t) is the elastic deflection.
For this system the dynamics are the same as for the usual Euler Bernoulli model,
except that the acceleration is ∂ 2 η/∂t2 in place of ∂ 2 w/∂t2 . This gives

∂4w ∂2η
c2 4
(x, t) + 2 (x, t) = 0 (9.228)
∂x ∂t
The equation of motion for the motor shaft can be found from

ḣ(t) = Tm (t) (9.229)

where Z
h(t) = Jm θ̇m + [xθ̇m + ẇ(x, t)]ρxdx (9.230)
0
is the angular momentum of the motor and beam. The equation of motion is found to
be Z
Jt θ̈m (t) + ρxẇ(x, t)dx = Tm (t) (9.231)
0
where Z
Jt = Jm + ρx2 dx (9.232)
0
is the total inertia seen from the motor.

9.7.2 Assumed mode shapes

An elastic beam is fixed to a moving base, and the elastic deflection of the beam is
described by
∞
X
w(x, t) = φj (x)qi (t) (9.233)
j=1

where φj (x) is a set of shape functions. Typically, the orthogonal shape functions may be
available from previous analysis, or even from textbooks, and a widely used method is to
approximate the solution by assuming that the orthogonal modes of the beam itself is a
sufficiently accurate approximation in (9.233). To demonstrate how this can be done we
will use the orthogonal shape functions φj (x) of a pinned Euler Bernoulli beam, which
is an accurate approximation if the inertia of the motor is large compared to the inertia
of the beam.
The angular momentum of the motor axis and beam is
Z X∞
h(t) = Jm θ̇m + [xθ̇m + φj (x)q̇j (t)]ρxdx (9.234)
0 j=1

which gives the equation of motion

∞
X
Jt θ̈m (t) + hj q̈j (t) = Tm (t) (9.235)
j=1
392 CHAPTER 9. MECHANICAL VIBRATIONS

where Z
hj = ρφj (x)xdx (9.236)
0
is the angular momentum coefficient for mode shape j.
Insertion of (9.227) and (9.233) into (9.228) gives
∞
X
[c2 φ0000
j (x)qj (t) + φj (x)q̈j (t)] + xθ̈ m (t) = 0 (9.237)
j=1

As the mode shapes φj are eigenfunctions that satisfy (9.124), this can be written
∞
X
φj (x)[ω 2j qj (t) + q̈j (t)] + xθ̈m (t) = 0 (9.238)
i=1

where the natural frequencies are given by ω 2j = c2 β 4j . Finally, we may find the differential
equations for the generalized coordinates qi by multiplying with φi (x) and integrating
over the interval x ∈ [0, ]. This gives
 
Z X∞ 
ρφi (x) φj (x)[ω 2j qj (t) + q̈j (t)] + xθ̈m (t) dx = 0 (9.239)
0  
j=1

Then, by accounting for the orthogonality of the shape functions as expressed in (9.165)
we arrive at the differential equations

ω 2i qi (t) + q̈i (t) + hi θ̈m (t) = 0 (9.240)

where i has been inserted for j, and the angular momentum coefficients hi are given by
Z
hi = ρφi (x)xdx (9.241)
0

The model can be written

       
Jt h1 . . . hi . . . θ̈m 0 0 ... 0 ... θm Tm
 h1 1 ... 0 ...   q̈1   0 ω 21 ... 0 ...   q1    0
       
 .. .. .. ..  ..   .. .. .. ..  ..    ..
 . . . . ...   . + . . . . ...   . =  .
       
 hi 0 ... 1 ...   q̈i   0 0 . . . ω 2i ... 
  
qi   0 
     
.. .. .. .. . . .. .. .. .. .. .. .. ..
. . . . . . . . . . . . .

This type of model formulation is used to simulate the dynamics of flexible antennas
mounted on satellites (Hughes 1974).

9.7.3 Finite elements

A model for a beam mounted on a motor axis can also be derived with the finite-element
method. Again, the position of a point x along the beam is given by η(x, t) = xθm (t) +
w(x, t), and the elastic deflection of the beam is described by
2N
X
w(x, t) = ψ j (x)qi (t) (9.242)
j=1
9.8. IRRATIONAL TRANSFER FUNCTIONS FOR BEAM DYNAMICS 393

where ψ j (x) are the Hermitian shape functions used in the finite-element method. The
partial differential equation (9.228) is then approximated by

c2 ψ 0000 (x)T q(t) + ψ(x)T q̈i (t) + xθ̈m (t) = 0 (9.243)

and Galerkin’s method leads to

Z h i
ρψ(x) c2 ψ 0000 (x)T q(t) + ψ(x)T q̈i (t) + xθ̈m (t) dx = 0 (9.244)
0

Insertion of (9.221) gives

Z h i
c2 ρψ 00 (x)T ψ 00 (x)qi (t) + ρψ(x)ψ(x)T q̈i (t) + ρψ(x)xθ̈m (t) dx = 0 (9.245)
0

The equation of motion for the motor is approximated by

2N
X
Jt θ̈m (t) + hj q̈j (t) = Tm (t) (9.246)
j=1

where Z
Jt = Jm + ρx2 dx (9.247)
0
is the total inertia seen from the motor, and
Z
hj = ρψ j (x)xdx (9.248)
0

are the influence coefficients. This leads to the model

µ ¶ µ ¶
Jt hT 0 0T
M = , K= (9.249)
h Mf e 0 Kf e
   
θm 1  
 q1   0  h1
     
q =  .  , b =  .  , h =  ...  (9.250)
 ..   .. 
hN
qN 0
where the mass matrix Mf e and the stiffness matrix Kf e of the finite-element method
are given by
Z Z
¡ ¢T
Mf e = ρψψ T dx, Kf e = ρψ 00 ψ 00 dx (9.251)
0 0

9.8 Irrational transfer functions for beam dynamics

9.8.1 Introduction
So far the dynamics of the Euler Bernoulli beam has been described using a series ex-
pansion of a solution based on shape functions. In this section it will be shown that
transfer functions for the beam dynamics can be derived directly from the partial differ-
ential equation. This is of great interest in itself, but it also gives useful information on
the dynamics of the system as there are no approximations involved. In particular, the
singularities and zeros of the dynamics can be found.
394 CHAPTER 9. MECHANICAL VIBRATIONS

w 2 ,V 2

Figure 9.13: Clamped beam with excitation force V2 and deflection w2 on the tip.

9.8.2 Clamped-free beam

Consider a clamped-free beam which is clamped at x = 0 and free at x = L as shown in
Figure 9.13. The beam is excited with a force V2 at the end, and we will in the following
derive the transfer function from the tip force V2 to the tip deflection w2 . Laplace
transformation of the partial differential equation (9.105) gives

∂4 s2
4
w(x, s) + 2 w(x, s) = 0 (9.252)
∂x c
Define the complex variable γ(s) by the relation

s2
γ4 = − (9.253)
c2
Then the Laplace transformation of the partial differential equation (9.266) gives the
ordinary differential equation
∂4
w(x, s) − γ 4 w(x, s) = 0 (9.254)
∂x4
The solution to the Laplace transformed model (9.254) is

w(x, s) = C1 cos γx + C2 sin γx + C3 cosh γx + C4 sinh γx (9.255)

while the derivatives are

w0 (x, s) = γ (−C1 sin γx + C2 cos γx + C3 sinh γx + C4 cosh γx) (9.256)

w00 (x, s) = γ 2 (−C1 cos γx − C2 sin γx + C3 cosh γx + C4 sinh γx) (9.257)
w000 (x, s) = γ 3 (C1 sin γx − C2 cos γx + C3 sinh γx + C4 cosh γx) (9.258)

For the clamped-free beam the boundary conditions imply that w1 = 0, w10 = θ1 = 0, and
that M2 and V2 are functions of time only. This means that w1 , θ1 , M2 and V2 should
be considered as input to the system, whereas w2 , θ2 , M1 and V1 can be considered as
outputs of the system dynamics. Note that it is critical in the method that is presented
here that the appropriate variables are selected as inputs and outputs. From (9.255—
9.258) we find that the input variables can be expressed in terms of the constant Ci
as
    
w1 (s) 1 0 1 0 C1
 1
θ (s)   0 1 0 1   C2 
 1γ 1    
 γ 2 EI M2 (s)  =  − cos γL − sin γL cosh γL sinh γL   C3  (9.259)
1 sin γL − cos γL sinh γL cosh γL C4
γ 3 EI V2 (s)
| {z } | {z } | {z }
u(s) G(s) c
9.8. IRRATIONAL TRANSFER FUNCTIONS FOR BEAM DYNAMICS 395

while the output variables can be expressed by

    
w2 (s) cos γL sin γL cosh γL sinh γL C1
 1
θ2 (s)   − sin γL cos γL sinh γL cosh γL  C2 
 1 γ    
 γ 2 EI M1 (s)  =  −1 0 1 0  C3  (9.260)
1 0 −1 0 1 C4
3
γ EI V 1 (s)
| {z } | {z }| {z }
y(s) K(s) c

Then, as G(s) is nonsingular, we can find the transfer function matrix H(s) from the
expression
y(s) = H(s)u(s) = K(s)G(s)−1 u(s) (9.261)
where u(s), y(s), G(s), and K(s) are defined in (9.259) and (9.260). Using a symbolic
program like MATLAB or MAPLE, we find that the transfer function from the tip force
V2 to the tip deflection w2 is given by
w2 (s) 1 cos γL sinh γL − sin γL cosh γL
= 3 (9.262)
V2 (s) γ EI 1 + cos γL cosh γL
The singularities of the transfer function are found by equating the denominator to zero.
This gives
1 + cos γL cosh γL = 0 (9.263)
which is the characteristic equation for the clamped-free beam. This means that the
natural frequencies of the irrational transfer functions are given by the natural frequencies
found from the analysis based on the orthogonal mode shapes. This makes sense, as both
methods are exact. The zeros of the transfer function are found from

cos γL sinh γL − sin γL cosh γL = 0 (9.264)

which is the characteristic equation for the clamped-pinned beam. This can be explained
as the dynamics associated with the zeros gives small amplification from the force V2 (s)
to the deflection w2 (s). The clamped-pinned beam has zero deflection w2 (s), and it
seems reasonable that this is reflected in the location of the zeros.

9.8.3 Motor and beam

In this section transfer functions for a motor and an Euler Bernoulli beam of length L
will be derived. The results are based on (Wie and Bryson 1987) and (Gevarter 1970).
The starting point for the derivation is the definition of the variable

η(x, t) = w(x, t) + xθ(t) (9.265)

which leads to the partial differential equation

∂4 1 ∂2
4
η(x, t) + 2 2 η(x, t) = 0 (9.266)
∂x c ∂t
which is the same as for the beam model. The position variable and the angles at the
end-points are given by

η 1 (s) = η(0, s), η 2 (s) = η(L, s) (9.267)

θ1 (s) = η 0 (0, s), θ2 (s) = η 0 (L, s) (9.268)
396 CHAPTER 9. MECHANICAL VIBRATIONS

while the shear forces and moments at the end-points are given by
1 1
V1 (s) = η 000 (0, s), V2 (s) = η 000 (L, s) (9.269)
EI EI
1 1
M1 (s) = η 00 (0, s), M2 (s) = η 00 (L, s) (9.270)
EI EI
The beam is pinned to the motor at x = 0 and is free at x = L, which leads to the
boundary conditions
η(0, t) = η 00 (L, t) = η 000 (L, t) = 0 (9.271)
1
η 00 (0, t) = M1 (t) = TL (t) (9.272)
EI
where TL (t) is the torque from the motor shaft on the beam.
We will now develop a transfer function model of the form y(s) = H(s)u(s) of the
motor and beam. First we have to select the input and output variables of the model.
The input and output variables are η1 , η 2 , θ1 , θ2 , V1 , V2 , M1 and M2 . The boundary
conditions (9.271) and (9.272) imply that η 1 , M2 and V2 are zero, and that M1 is a
function of time only. Therefore these variables must be in the input vector u of the
transfer function model. The remaining four variables η 2 , θ1 , θ2 , and V1 are in the
output vector y.
The partial differential equation (9.266) for η is the same as the partial differential
equation (9.105) for w. Therefore, the solution η(x, s) and its derivatives with respect to
x will be given by (9.255—9.258), and the input and output variables can be expressed by
 1    
γ 2 EI M1 (s) −1 0 1 0 C1
 η 1 (s)   1 0 1 0   C2 
 =   (9.273)
 γ 3 EI V2 (s)   sin γL
1
− cos γL sinh γL cosh γL   C3 
1 − cos γL − sin γL cosh γL sinh γL C4
γ 2 EI M2 (s)
| {z } | {z } | {z }
u(s) K(s) c

and
 1    
γ θ 1 (s) 0 1 0 1 C1
 η (s)   cos γL sin γL cosh γL sinh γL  C2 
 1 2    
 γ θ2 (s)  =  − sin γL cos γL sinh γL cosh γL  C3  (9.274)
1 0 −1 0 1 C4
γ 3 EI V1 (s)
| {z } | {z }| {z }
y(s) G(s) c

The matrix G(s) is nonsingular, we can find the transfer function matrix H(s) from the
expression
y(s) = H(s)c(s) = K(s)G(s)−1 u(s) (9.275)
where u(s), y(s), G(s), and K(s) are defined in (9.259) and (9.260). Using a symbolic
program like MATLAB or MAPLE, we find that the transfer function from the moment
to the angle at x = 0 is given by
θ1 1 1 + cos γL cosh γL
(s) = (9.276)
M1 EI γ (sin γL cosh γL − cos γL sinh γL)
The zeros and the singularities are easily found for this transfer function as the numerator
is equal to zero whenever the characteristic equation for the clamped-free beam is satis-
fied, whereas the denominator is zero when the characteristic equation of the pinned-free
9.8. IRRATIONAL TRANSFER FUNCTIONS FOR BEAM DYNAMICS 397

Zeros Singularities
s s
γL c/L2 γL c/L2
1.875104 ±j3.516 0 ±0 (9.277)
4.694091 ±j22.034 3.926602 ±j15.418
7.854757 ±61.697 7.068583 ±j49.964
10.995541 ±j120.9019 10.210176 ±j104.2477

Table 9.4: The first zeros and singularities for the transfer function from M1 (s) to θ1 (s).

beam is satisfied. These solutions are tabulated in standard textbooks like (Rao 1990).
Then the zeros and singularities are given by Table 9.4.
The singularities and zeros are along the imaginary axis as shown in Figure 9.14.
This agrees with the fact that the transfer function sθ1 /M1 is positive real, which can
be shown from energy arguments. The transfer function from the moment at x = 0 to

Figure 9.14: Singularities and zeros for the transfer function from the motor torque to
M1 the motor angle θ1 . The singularities are marked with crosses, and the zeros are
marked with circles.

the deflection at the other side of the beam is given by

η2 1 sin γL + sinh γL
(s) = (9.278)
M1 EI γ 2 (sin γL cosh γL − cos γL sinh γL)
The singularities are the same as for the transfer function (9.276), while it is possible
to verify that the numerator expression sin z + sinh z = 0 has solutions of the type
z = ρ(1 + j) where ρ is the solution of tan ρ + tanh ρ = 0. This gives the zeros and
singularities shown in Table 9.5.
Note that this transfer function has zeros in the right half plane as shown in Fig-
ure 9.15.
The beam is connected to the motor by requiring
Jm s2 θ1 = T − M1 (9.280)
which gives
1 + cos γL cosh γL
θ(s) = T (s) (9.281)
Jm s2 (1 + cos γL cosh γL) + EIγ (sin γL cosh γL − cos γL sinh γL)
398 CHAPTER 9. MECHANICAL VIBRATIONS

Zeros Singularities
s s
γL c/L2 γL c/L2
2.365 (1 + j) ±11.1865 0 ±0 (9.279)
5.498 (1 + j) ± 60.4560 3.926602 ±j15.418
8.639 (1 + j) ±149.2646 7.068583 ±j49.964

Table 9.5: The first zeros and singularities for the transfer function from the motor torque
M1 (s) to tip position η 2 (s) .

Figure 9.15: Singularities and zeros of the transfer function from the motor torque M1
to the end-point position η 2 . The singularities are marked with crosses, and the zeros
are marked with circles.
 Part IV

Balance equations

399
Chapter 10

Kinematics of Flow

10.1 Introduction
Balance equations are differential equations that are derived from conservation laws for
control volumes. The conservation laws include the conservation of mass, momentum,
angular momentum, and energy. Fluid flow phenomena are important in the derivation of
the balance laws of this chapter, and therefore the presentation starts with the kinematics
of fluid flow. Then the transport theorem is presented, and it is shown how the transport
theorem can be used to derive balance equations for typical control volumes. Balance
equations for mass, momentum, angular momentum and energy are then developed using
the mathematical tools presented in the first part of the chapter.

10.2 Kinematics
10.2.1 The material derivative
Let x1 , x2 , x3 be the coordinates of a Cartesian frame with with orthogonal unit vectors
a1 , a2 , a3 . A scalar function φ = φ(t, x1 , x2 , x3 ) of time and position is called a scalar
field. The time derivative of a scalar field φ is, according to the usual definition of the
derivative,
dφ ∂φ ∂φ dx1 ∂φ dx2 ∂φ dx3
= + + +
dt ∂t ∂x1 dt ∂x2 dt ∂x3 dt
T
We will also be dealing with vector fields u = u(t, x1 , x2 , x3 ) where u = (u1 , u2 , u3 ) . The
time derivative of a vector field u is
du ∂u ∂u dx1 ∂u dx2 ∂u dx3
= + + +
dt ∂t ∂x1 dt ∂x2 dt ∂x3 dt

The time derivative dφ dt of the scalar field φ clearly depends on the time derivatives
ẋi = dx
dt
i
of the position coordinates xi . This means that it must be specified for which
time function x(t) the derivative is taken. Two cases are common: One is the spatial
derivative which is the derivative at a specific point x = x0 where x0 is a constant vector.
Then the position coordinates xi are constants, and
¯
dφ ¯¯ ∂φ
¯ =
dt x=x0 ∂t

401
402 CHAPTER 10. KINEMATICS OF FLOW

x3 ux,t 

v

x, t
xt 

x2

x1

Figure 10.1: Coordinate frame with the motion x(t) of a particle.

In the same way the spatial derivative of a vector field is

¯
du ¯¯ ∂u
=
dt ¯x=x0 ∂t

The other usual case is the material derivative, which is the time derivative when follow-
T
ing a particular particle of the fluid. Then ẋ = v, where v = (v1 , v2 , v3 ) is the velocity
of the fluid. The material derivative is widely used, and this motivates the introduction
of the notation ¯ ¯
Dφ dφ ¯¯ Du du ¯¯
:= , :=
Dt dt ¯ẋ=v Vt dt ¯ẋ=v

The material derivative of a scalar field φ is defined by

Dφ ∂φ ∂φ ∂φ ∂φ
:= + v1 + v2 + v3 (10.1)
Dt ∂t ∂x1 ∂x2 ∂x3
while the material derivative of a vector field u is given by
Du ∂u ∂u ∂u ∂u
= + v1 + v2 + v3
Vt ∂t ∂x1 ∂x2 ∂x3

10.2.2 The nabla operator

The nabla vector operator ∇ is defined in the Cartesian coordinate system by
∂ ∂ ∂
∇ = a1 + a2 + a3 (10.2)
∂x1 ∂x2 ∂x3

When the vector operator ∇ is applied to a scalar field φ, we get the gradient vector
∂φ ∂φ ∂φ
∇φ = a1 + a2 + a3 (10.3)
∂x1 ∂x2 ∂x3
10.2. KINEMATICS 403

The nabla vector operator may be represented by a column vector as

 ∂ 
∂x1
∇ : = ∂
∂x2
 (10.4)
∂
∂x3

It follows that  
∂φ
∂x1
½ ¾
 ∂φ  T ∂uj
∇φ =  ∂x2  , ∇u = (10.5)
∂φ ∂xi
∂x3

It is then straightforward to show that

∂φ ∂φ ∂φ
v · ∇φ = vT ∇φ = v1 + v2 + v3 (10.6)
∂x1 ∂x2 ∂x3
With some care it is also found that
¡ ¢ ∂u ∂u ∂u
vT ∇u = vT ∇ u = v1 + v2 + v3 (10.7)
∂x1 ∂x2 ∂x3
In the vector notation with u = u1 a1 + u2 a2 +3 a3 , then
∂u ∂u ∂u
v · ∇u = (v · ∇)u = v1 + v2 + v3 (10.8)
∂x1 ∂x2 ∂x1

The material derivative of a scalar field φ and a vector field u can be written
Dφ ∂φ Du ∂u
:= + vT ∇φ, := + vT ∇u (10.9)
Dt ∂t Dt ∂t
or, alternatively, in vector form as

Dφ ∂φ Du ∂u
= + v · ∇φ, = + v · ∇u (10.10)
Dt ∂t Dt ∂t

10.2.3 Divergence
The divergence of a vector field u = u1 a1 + u2 a2 + u3 a3 is the scalar
∂u1 ∂u2 ∂u3
∇ · u= ∇T u = + + (10.11)
∂x1 ∂x2 ∂x3
The divergence appears in many results. In particular, its usefulness is due to the diver-
gence theorem:

The Divergence Theorem: Consider a volume V with a closed surface ∂V and an outwards
pointing surface normal n, where n is a unit vector. Let dV be a volume element and
dA a surface element. Then, for any vector field u = u (x) we have
ZZ ZZZ
u · ndA = ∇ · udV (10.12)
∂V (t) V (t)
404 CHAPTER 10. KINEMATICS OF FLOW

x3
u

n

x2
x1

Figure 10.2: Volume V with outwards pointing surface normal n and the vector u.

Example 151 A result related to the divergence theorem is

ZZ ZZZ
φndA = ∇φdV (10.13)
∂V (t) V (t)

The first component of this vector equation follows from the divergence theorem by letting
u=φa1 , and the second and third element is found in a similar way.

Example 152 The divergence of the vector φu is

∂φu1 ∂φu2 ∂φu3

∇ · (φu) = + +
∂x1 ∂x2 ∂x3
µ ¶
∂φ ∂φ ∂φ ∂u1 ∂u2 ∂u3
= u1 + u2 + u3 + φ + + (10.14)
∂x1 ∂x2 ∂x3 ∂x1 ∂x2 ∂x3

and we see that

∇ · (φu) = (∇φ)·u + φ(∇ · u) (10.15)

10.2.4 Curl
The curl of a vector u = u (x) is the vector
¯ ¯
¯ a1 a2 a3 ¯ X3 X 3
¯ ∂ ¯ ∂uk
¯ ∂
∇ × u = ¯ ∂x1 ∂x2 ∂x ∂ ¯= εijk ai (10.16)
3 ¯ ∂xj
¯ u1 u2 u3 ¯ i=1 j=1

The coordinate form is seen to be

 ∂ ∂    ∂u3 ∂u2

0 − ∂x ∂x u1 ∂x2 − ∂x3
  u2  =  
3 2
∇× u =  ∂x ∂
3
∂
0 − ∂x 1

∂u1
∂x3 − ∂u3
∂x1  (10.17)
∂ ∂ u ∂u2 ∂u1
− ∂x2 ∂x1 0 3 ∂x1 − ∂x2

where ∇× is the skew symmetric form of ∇. The curl of a vector is used in Stokes’
Theorem:
10.2. KINEMATICS 405

x3 u n

ds

S
x2
x1

Figure 10.3: Surface S with surface normal n and tangential differential vector ds along
the closed boundary ∂S.

Stokes’ Theorem: Consider a surface S with a boundary ∂S, which is a closed curve.
Let ds be the differential position increment which is tangent to the curve ∂S. Let dA
be an area element on the surface, and let n be a surface normal so that the direction
of ds corresponds to a counter-clockwise rotation around n. Then for any vector field
u = u(x) we have I ZZ ³ ´
u·ds= ∇ × u · ndA
∂S S

We see from Stokes’ Theorem that if the surface S is taken to be dA, then
I ³ ´
1
u·ds= ∇ × u · n (10.18)
dA ∂S

The condition ∇ × u=0 implies that

I
u·ds = 0 (10.19)
∂S

This is equivalent to the existence of a scalar function ψ(x) called the potential of u (x),
so that u(x) =∇ψ(x). This is shown in basic textbooks on vector analysis. Here we just
comment that this is a consequence of the result
Z x2
ψ(x2 )−ψ(x1 ) = ∇ψ(x)·ds (10.20)
x1

Example 153 The skew symmetric form of the column vector form ∇× u is found to be
½ ¾
¡ × ¢× ∂ui ∂uj
∇ u = − (10.21)
∂xj ∂xi
406 CHAPTER 10. KINEMATICS OF FLOW

3
x3 t
2

t0 1
particle
motion

x2
x1

Figure 10.4: The spatial coordinate frame (x1 , x2 , x3 ) and the material coordinate frame
(ξ 1 , ξ 2 , ξ 3 ). At time t0 the two frames coincide, then the matrial frame moves with the
particles of the fluid so that each particle has the same coordinates ξ 1 , ξ 2 , ξ 3 for all t.

Example 154 The following results can be verified on the component level:
³ ´
∇· ∇×u = 0 (10.22)
³ ´
∇ × ∇φ = 0 (10.23)
³ ´
∇ × (φu) = ∇φ × u + φ∇ × u (10.24)
1 ³ ´ ³ ´
(u · u) = u·∇ u+u× ∇×u (10.25)
2 ³ ´ ³ ´
2
∇ u = ∇ ∇·u −∇× ∇×u (10.26)

10.2.5 Material coordinates

For fluids and deformable bodies the concept of spatial coordinates and material coordi-
T
nates is useful. The spatial coordinates x = (x1 , x2 , x3 ) define a spatial grid which is
T
constant. In contrast to this, the material coordinates ξ = (ξ 1 , ξ 2 , ξ 3 ) define a material
grid where each particle of the fluid has a given position in the grid. Then, as the fluid
deforms, the material grid deforms with the fluid so that each particle maintains its po-
sition in the grid. At initial time t0 the spatial coordinates and the material coordinates
coincide, that is, ξ(t0 ) = x(t0 ). For t ≥ t0 the material coordinates will be a function of
the spatial coordinates and vice versa, so that

ξ = ξ [x(t), t] and x = x [ξ(t), t] (10.27)

10.2.6 The dilation

The dilation or expansion of a fluid (Aris 1989) is closely related to the divergence of
the velocity, which will be shown in this section. A control volume V (t) is considered.
The control volume is assumed to contain the same fluid particles as it moves with the
flow as indicated in Figure 10.5. This means that the mass contained in the volume is
10.2. KINEMATICS 407

x3

Vt 

Vt 0 

x2
x1

Figure 10.5: Material control volume containing the same set of particles for all t. The
material volume V moves along with the particles of the fluid, and may be deformed and
stretched.

x3 n
dA
v
V

x2
x1

Figure 10.6: Material control volume V where the change in volume due to the velocity
v of a surface element dA is indicated.

constant, while the volume and the surface may change with the flow. Such a control
volume will be called a material control volume. An area element dA(t) of the surface
∂V (t) of V (t) moves with the velocity v. Therefore, the motion of the area element dA
results in a change of volume with a rate v · ndA where n is the outwards surface normal
at dA. Integrating over the whole surface ∂V (t) of the control volume results in
ZZ
DV
= v · ndA (10.28)
Dt ∂V (t)

Then the divergence theorem leads to

ZZZ
DV
= ∇ · vdV (10.29)
Dt V (t)

It is seen that if the divergence ∇ · v is negative for all of V (t), then the time derivative
dV (t)
dt will be negative and the control volume will contract and become smaller in size.
408 CHAPTER 10. KINEMATICS OF FLOW

Likewise, if the divergence is positive for all V (t), the control volume will expand. An
interesting result appears if the control volume is taken to be a infinitesimal material
volume element dV (t) which is an infinitesimal volume element containing the same
particles as the fluid flows. Then the divergence ∇ · v can be taken to be a constant over
the control volume, and the time derivative of dV is the material derivative. Then the
divergence theorem gives the following result:

The time derivative of a material volume element dV is given by

D (dV ) ³ ´
= ∇ · v dV (10.30)
Dt
This shows that the material volume element will diverge if the divergence of velocity is
positive, and that it will contract is the divergence is negative.

Example 155 Consider the specific volume

dV 1
V̂ := = (10.31)
dm ρ

which is widely used in thermodynamics. Here dV is the volume of the mass element dm,
and ρ is the density. The material derivative of V̂ becomes
³ ´
µ ¶ ∇ · v dV ³ ´
DV̂ D dV 1 D (dV )
= = = = ∇ · v V̂ (10.32)
Dt Dt dm dm Dt dm

where v is the velocity vector.

10.3 Orthogonal curvilinear coordinates

10.3.1 General results
So far we have been working with a Cartesian coordinate system (x1 , x2 , x3 ) with orthog-
onal unit vectors i1 , i2 , i3 along the coordinate axes. Other useful coordinate systems are
cylindrical coordinates and spherical coordinates, which are examples of orthogonal curvi-
linear coordinate systems. The material in this section is based on (Milne-Thomson 1996,
p. 62). The section can be skipped at a first reading. An orthogonal curvilinear coordi-
nate system (y1 , y2 , y3 ) is a coordinate system where the surfaces defined by y1 = y1P ,
y2 = y2P , y3 = y3P intersect orthogonally when y1P , y2P and y3P are constants. The
point of intersection between the planes is denoted by P . If we draw the surfaces cor-
responding to y1 = y1P , y2 = y2P , y3 = y3P and y1 = y1P + ∆y1 , y2 = y2P + ∆y2 ,
y3 = y3P + ∆y3 we get a figure which to the first order approximation is a parallel piped
where one corner is the point P . The edges of the parallel piped are of length h1 ∆y1 ,
h2 ∆y2 and h3 ∆y3 where h1 , h2 and h3 are functions of the coordinates. To proceed we
introduce a Cartesian coordinate system (z1 , z2 , z3 ) with origin in the point P and with
orthogonal unit vectors j1 , j2 , j3 along the coordinate axes (y1 , y2 , y3 ), which coincide
with the edges of the parallel piped. It is assumed that the unit vectors j1 , j2 , j3 form a
right handed system. In the (z1 , z2 , z3 ) system the length of the edges are denoted ∆z1 ,
∆z2 and ∆z3 . This means that
∆zi = hi ∆yi (10.33)
10.3. ORTHOGONAL CURVILINEAR COORDINATES 409

and that the volume of the parallel piped is

∆V = ∆z1 ∆z2 ∆z3 = h1 h2 h3 ∆y1 ∆y2 ∆y3 (10.34)
The nabla operator at the point P described in the Cartesian coordinate system
(z1 , z2 , z3 ) is given by
∂ ∂ ∂
∇= j1 + j2 + j3 (10.35)
∂z1 ∂z2 ∂z3
and, if φ is any scalar field, the gradient of φ at the point P is
∂φ ∂φ ∂φ
∇φ= j1 + j2 + j3 (10.36)
∂z1 ∂z2 ∂z3
The gradient can be expressed in the orthogonal curvilinear coordinates (y1 , y2 , y3 ) using
∂φ ∂y1 ∂φ ∂y2 ∂φ ∂y3
∇φ = j1 + j2 + j3
∂y1 ∂z1 ∂y2 ∂z2 ∂y3 ∂z3
1 ∂φ 1 ∂φ 1 ∂φ
= j1 + j2 + j3 (10.37)
h1 ∂y1 h2 ∂y2 h3 ∂y3
and we may conclude that the nabla vector can be expressed in orthogonal curvilinear
coordinates as
j1 ∂ j2 ∂ j3 ∂
∇= + + (10.38)
h1 ∂y1 h2 ∂y2 h3 ∂y3
The divergence at the point P of a vector
u = u1 j1 + u2 j2 + u3 j3 (10.39)
is
3
X 3 h³
X ´ ³ ´i
∇·u=∇· ui ji = ∇ui · ji + ui ∇ · ji (10.40)
i=1 i=1
while the curl is
3
X 3 h³
X ´ ³ ´i
∇×u=∇× ui ji = ∇ui × ji + ui ∇ × ji (10.41)
i=1 i=1

Here we have used the identities

³ ´ ³ ´
∇ · (φa) = ∇φ · a + φ ∇ · a (10.42)
³ ´ ³ ´
∇ × (φa) = ∇φ × a + φ ∇ × a (10.43)

To compute the expressions for the divergence and the curl we need some intermediate
results, namely expressions for the divergence and curl of the unit vectors ji . We do this
by introducing a potential given by φ = y1 with gradient
j1
∇y1 = (10.44)
h1
³ ´ ³ ´
Then from the identity ∇× ∇φ = ∇ × ∇ φ = 0 we get
à ! µ ¶
³ ´ j1 1 1 ³ ´
0 = ∇ × ∇y1 = ∇ × =∇ × j1 + ∇ × j1 (10.45)
h1 h1 h1
410 CHAPTER 10. KINEMATICS OF FLOW

which implies that the curl of the unit vector j1 is

µ ¶
1
∇ × j1 = h1 ∇ × j1
h1
µ ¶µ ¶
1 1 ∂h1 1 ∂h1 1 ∂h1
= h1 − 2 j1 + j2 + j3 × j1
h1 h1 ∂y1 h2 ∂y2 h3 ∂y3
j2 ∂h1 j3 ∂h1
= − (10.46)
h1 h3 ∂y3 h1 h2 ∂y2

The curl of j2 and j3 is found in the same way, and following expressions for the curl of
the unit vectors result:

j2 ∂h1 j3 ∂h1
∇ × j1 = − (10.47)
h1 h3 ∂y3 h1 h2 ∂y2
j3 ∂h2 j1 ∂h2
∇ × j2 = − (10.48)
h2 h1 ∂y1 h2 h3 ∂y3
j1 ∂h3 j2 ∂h3
∇ × j3 = − (10.49)
h3 h2 ∂y2 h3 h1 ∂y1

The divergence of the unit vector j1 is found from

³ ´
∇ · j1 = ∇ · j2 × j3
³ ´ ³ ´
= ∇ × j2 · j3 − j2 · ∇ × j3
1 ∂h2 1 ∂h3
= −
h2 h1 ∂y1 h3 h1 ∂y1
1 ∂ (h2 h3 )
= (10.50)
h1 h2 h3 ∂y1

The divergence of j2 and j3 is found in the same way, and we can conclude that

1 ∂ (h2 h3 )
∇ · j1 = (10.51)
h1 h2 h3 ∂y1
1 ∂ (h3 h1 )
∇ · j2 = (10.52)
h1 h2 h3 ∂y2
1 ∂ (h1 h2 )
∇ · j3 = (10.53)
h1 h2 h3 ∂y3

With these results it is a straightforward, although it is a time-consuming exercise,

to verify that the divergence of a vector in orthogonal curvilinear coordinates is
· ¸
1 ∂ ∂ ∂
∇·u= (u1 h2 h3 ) + (u2 h3 h1 ) + (u3 h1 h2 ) (10.54)
h1 h2 h3 ∂y1 ∂y2 ∂y3

while the curl is ¯ ¯

¯ h1 j1 h2 j2 h3 j3 ¯¯
1 ¯ ∂
∇×u= ¯ ∂ ∂ ¯ (10.55)
h1 h2 h3 ¯ ∂y1 ∂y2 ∂y3 ¯
¯ h u h2 u2 h u ¯
1 1 3 3
10.3. ORTHOGONAL CURVILINEAR COORDINATES 411

2
We also present the result for the Laplacian ∇ φ of a scalar field φ, which is found by
2
letting u = ∇φ. Then ∇ φ = ∇ · u, and it is seen that
· µ ¶ µ ¶ µ ¶¸
2 1 ∂ ∂φ h2 h3 ∂ ∂φ h3 h1 ∂ ∂φ h1 h2
∇ φ= + + (10.56)
h1 h2 h3 ∂y1 ∂y1 h1 ∂y2 ∂y2 h2 ∂y3 ∂y3 h3

Example 156 For spherical coordinates (r, φ, θ) the orthogonal unit vectors jr , jφ , jθ
form a right handed system, and we have hr = 1, hφ = r and hθ = r sin φ. The vector u
is written
u = ur jr + uφ jφ + uθ jθ (10.57)
while the gradient of a scalar field ψ is

∂ψ 1 ∂ψ jθ ∂ψ
∇ψ = jr + jφ + (10.58)
∂r r ∂φ r sin φ ∂θ

We find that the divergence of u is

¡ ¢
1 ∂ ur r2 1 ∂ (uφ sin φ) 1 ∂uθ
∇·u= 2 + + (10.59)
r ∂r r sin φ ∂φ r sin φ ∂θ
while the curl is ¯ ¯
¯ jr rjφ r sin φjθ ¯¯
1 ¯ ∂
∇×u= 2 ¯ ∂ ∂ ¯ (10.60)
r sin φ ¯¯ ∂r ∂φ ∂θ ¯
ur ruφ r sin φu ¯
θ

10.3.2 Cylindrical coordinates

For cylindrical coordinates (r, θ, z) the orthogonal unit vectors jr , jθ , jz form a right
handed system, and we have hr = 1, hθ = r and hz = 1. The vector u is written

u = ur jr + uθ jθ + uz jz (10.61)

the expression for ∇ is

∂ jθ ∂ ∂
∇ = jr + + jz (10.62)
∂r r ∂θ ∂z
The divergence of the vector u is found to be

1 ∂ (ur r) 1 ∂uθ ∂uz

∇·u= + + (10.63)
r ∂r r ∂θ ∂z
while the curl is ¯ ¯
¯ j rjθ jz ¯¯
1 ¯ ∂r
∇ × u = ¯¯ ∂r ∂
∂θ
∂ ¯
∂z ¯
(10.64)
r¯
ur ruθ uz ¯
The Laplacian of the scalar field ψ is
· µ ¶ µ ¶ µ ¶¸
2 1 ∂ ∂ψ ∂ ∂ψ 1 ∂ ∂ψ
∇ ψ = r + + r (10.65)
r ∂r ∂r ∂θ ∂θ r ∂z ∂z
∂2ψ 1 ∂ 2 ψ ∂ 2 ψ 1 ∂ψ
= + + + (10.66)
∂r2 r2 ∂θ2 ∂z 2 r ∂r
412 CHAPTER 10. KINEMATICS OF FLOW

The velocity v may also be written

v = v1 i1 + v2 i2 + v3 i3 (10.67)
= vr jr + vθ jθ + vz jz (10.68)
where the unit vectors i1 , i2 , i3 of the Cartesian system (x1 , x2 , x3 ) are constant vectors,
while the unit vectors jr and jθ of the cylindrical coordinate system changes as the
particle moves. The material derivative is
Dv Dv1 Dv2 Dv3
= i1 + i2 + i3 (10.69)
Dt Dt Dt Dt
Dvr Dvθ Dvz Djr Djθ
= jr + jθ + jz + vr + vθ (10.70)
Dt Dt Dt Dt Dt
where the time derivatives of the unit vectors are
D jr ∂ jr vθ ∂ jr vθ
= + = jθ (10.71)
Dt ∂t r ∂θ r
Djθ ∂ jθ vθ ∂ jθ vθ
= + = − jr (10.72)
Dt ∂t r ∂θ r
∂ jr ∂ jθ
= jθ , = −jr (10.73)
∂θ ∂θ
This gives the following expressions for the material derivative of the vector v:
µ ¶ µ ¶
Dv Dvr vθ2 Dvθ vr vθ Dvz
= − jr + + jθ + jz (10.74)
Dt Dt r Dt r Dt
and for the scalar field ψ:
Dψ ∂ψ ∂ψ ∂ψ vθ ∂ψ ∂ψ
= + v · ∇v = + vr + + vz (10.75)
Dt ∂t ∂t ∂r r ∂θ ∂z
The Laplacian of the velocity is found from
2
³ 2 ´ ³ 2 ´ ³ 2 ´
∇ v = ∇ v1 i1 + ∇ v2 i2 + ∇ v2 i3 (10.76)
2
³ ´ 2
³ ´ 2
³ ´
= ∇ vr jr + ∇ vθ jθ + ∇ vz jz (10.77)

To proceed we need the Laplacian of the components in the cylindrical coordinate system,
which are given by
à !
2
³ ´ ∂ 2 vr 1 ∂ 2 vr ∂vr ∂ jr ∂ 2 jr ∂ 2 vr 1 ∂vr
∇ vr jr = 2
jr + 2 2 jr + 2 + vr 2 + jr + jr
∂r r ∂θ ∂θ ∂θ ∂θ ∂z 2 r ∂r
µ 2 ¶
∂ vr 1 ∂ 2 vr ∂ 2 vr 1 ∂vr vr 2 ∂vr
= + + + − jr + 2 jθ (10.78)
∂r2 r2 ∂θ2 ∂z 2 r ∂r r2 r ∂θ
à !
2
³ ´ ∂ 2 vθ 1 ∂ 2 vθ ∂vθ ∂ jθ ∂ 2 jθ
∇ vθ jθ = jθ + 2 jθ + 2 + vθ 2 jθ
∂r2 r ∂θ2 ∂θ ∂θ ∂θ
µ 2 2 2
¶
∂ vθ 1 ∂ vθ ∂ vθ 1 ∂vθ vθ 2 ∂vθ
= 2
+ 2 2 + 2
+ − 2 jθ − 2 jr (10.79)
∂r r ∂θ ∂z r ∂r r r ∂θ
2
³ ´ ∂ 2 vz 1 ∂ 2 vz ∂ 2 vz 1 ∂vz
∇ vz jz = 2
j z + 2 2 jz + 2
jz + jz
∂r r ∂θ ∂z r ∂r
10.4. REYNOLDS’ TRANSPORT THEOREM 413

This leads to the following expression for the Laplacian of the velocity in cylindrical
coordinates:
µ 2 ¶
2 ∂ vr 1 ∂ 2 vr ∂ 2 vr 1 ∂vr 2 ∂vθ vr
∇ v = + 2 + + − 2 − 2 jr
∂r2 r ∂θ2 ∂z 2 r ∂r r ∂θ r
µ 2 2 2
¶
∂ vθ 1 ∂ vθ ∂ vθ 1 ∂vθ 2 ∂vr vθ
+ + 2 + + + 2 − 2 jθ
∂r2 r ∂θ2 ∂z 2 r ∂r r ∂θ r
µ 2 2 2
¶
∂ vz 1 ∂ vz ∂ vz 1 ∂vz
+ + 2 + + jz (10.80)
∂r2 r ∂θ2 ∂z 2 r ∂r

10.4 Reynolds’ transport theorem

10.4.1 Introduction
In the derivation of balance equations we will typically define a control volume that
will depend on the geometry of the specific problem. The modeling procedure will then
typically involve the calculation of the rate of change of mass, momentum or energy in
the control volume. In connection with this calculation Reynolds’ transport theorem is of
great use. In the following we will present the theorem and show how it can be adapted
to the case where the control volume is a material volume, a spatially fixed volume, or a
general control volume with a moving boundary.

10.4.2 Basic transport theorem

The concept of a control volume Vc is used in the derivation of models based on conserva-
tion laws. In this section we will present an important kinematic result called Reynolds’
transport theorem (Aris 1989), (White 1999). Reynolds’ transport theorem shows the
relation between the time derivative of the volume integral
ZZZ
φ(x, t)dV (10.81)
Vc (t)

and the time derivative of φ(x, t). The boundary of Vc (t) is denoted ∂Vc (t), and the
velocity of a point on the boundary ∂Vc (t) is denoted vc . We recall the following standard
result from calculus:
Z Z b(t)
d b(t) ∂f (x, t) db da
f (x, t)dx = dx + f (b, t) − f (a, t) . (10.82)
dt a(t) a(t) ∂t dt dt
In analogy with this, the time derivative of the integral in (10.81) is equal to the vol-
ume integral of the time derivative of the integrand, and one term due to the changing
boundary of the volume Vc (t).

For a general time-varying control volume Vc the transport theorem is given by

ZZZ ZZZ ZZ
d ∂φ(x, t)
φ(x, t)dV = dV + φvc · ndA (10.83)
dt Vc (t) Vc ∂t ∂Vc

The last term can be explained as follows: The volume element dA on the surface
∂Vc (t) has velocity vc , and the rate of change of the integral due to this is φvc · ndA
where n is the outwards unit normal of the surface. Integration over the whole surface
gives the total rate of change due to the change in the volume Vc (t).
414 CHAPTER 10. KINEMATICS OF FLOW

10.4.3 The transport theorem for a material volume

Of particular interest is Reynolds’ transport theorem for a material volume Vm (t). The
reason for this is that balance laws will typically be formulated for material volumes. By
material volume it is meant a volume containing a specific set of particles. It is assumed
that initially, say at t = t0 , these particles filled the volume Vm (t0 ) = V0 , while at time
t the same particles fill the volume Vm (t) = V . If we apply Reynolds transport theorem
with Vc = Vm , then vc (t) is equal to the particle velocity v(t) and Reynolds transport
theorem gives
ZZZ ZZZ ZZ
d ∂φ(x, t)
φ(x, t)dV = dV + φv · ndA (10.84)
dt Vm (t) Vm (t) ∂t ∂Vm (t)

To avoid the need to explain whether the volume is material or not, we define the notation
ZZZ ZZZ ZZ
D ∂φ(x, t)
φ(x, t)dV := dV + φv · ndA (10.85)
Dt V V ∂t ∂V

Note that in this notation, the volume V need not be a material volume, it is merely
assumed that some material volume Vm (t) coincides with V at time t.

The result can be further developed by applying the divergence theorem to the last
term on the right side of (10.85), and by using (10.10) and (10.15):

Reynolds’ transport theorem for a material volume coinciding with V at time t is given
in material form as
ZZZ ZZZ · ³ ´¸
D Dφ(x, t)
φ(x, t)dV = + φ ∇ · v dV (10.86)
Dt V V Dt

and in divergence form as

ZZZ ZZZ · ¸
D ∂φ(x, t)
φ(x, t)dV = + ∇(φv) dV (10.87)
Dt V V ∂t

10.4.4 The transport theorem and balance laws

As we will see in the following there are important physical laws that can be formulated
in terms of the material derivative given by (10.85). In particular, this is the case for the
mass balance, the momentum balance, the angular momentum balance, and the energy
balance. In the derivation of a model, however, we will often use a control volume that
is not a material volume, but rather a volume that is determined from the geometry of
the problem. From (10.83) and (10.85) we have the following result

For a general control volume Vc (t) where a point on the surface has velocity vc the
transport theorem gives
ZZZ ZZZ ZZ
d D
φ(x, t)dV = φ(x, t)dV − φ (v − vc ) · ndA (10.88)
dt Vc Dt Vc ∂Vc
10.4. REYNOLDS’ TRANSPORT THEOREM 415

If the volume Vf is fixed in spatial coordinates, we get

ZZZ ZZZ
d ∂φ(x, t)
φ(x, t)dV = dV, Vf is fixed. (10.89)
dt Vf Vf ∂t

as there is no term due to a changing boundary. Combining this with (10.85) we find:

For a fixed volume Vf the transport theorem gives

ZZZ ZZZ ZZ
∂φ(x, t) D
dV = φ(x, t)dV − φv · ndA (10.90)
Vf ∂t Dt Vf ∂Vf

.
416 CHAPTER 10. KINEMATICS OF FLOW
Chapter 11

Mass, momentum and energy

balances

11.1 The mass balance

11.1.1 Differential form
We will now derive the continuity equation using the principle of mass conservation,
which states that the mass of a material volume must be a constant. The mass of a
material volume Vm (t) is ZZZ
m= ρdV (11.1)
Vm (t)

where ρ is the fluid density. This means that principle of mass conservation can be
expressed in the form ZZZ
D
ρdV = 0 (11.2)
Dt V

Then Reynolds’ transport theorem with φ = ρ leads to

ZZZ · ¸
Dρ
+ ρ(∇ · v) dV = 0 (11.3)
V Dt

in material form and ZZZ · ¸

∂ρ
+ ∇ · (ρv) dV = 0 (11.4)
V ∂t
in divergence form.
As the volume V is arbitrary, the integrand in both integral forms (11.3) and (11.4)
of the continuity equation must be identically zero, and this leads to the differential
formulation of the continuity in material form

Dρ ³ ´
+ ρ ∇·v =0 (11.5)
Dt
|{z} | {z }
rate of change of rate of change of density
density in material due to divergence of
volume element material volume element

417
418 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

and in divergence form,

∂ρ
+ ∇ · (ρv) =0 (11.6)
∂t
|{z} | {z }
rate of change of density
rate of change of
due to convection out of
density in spatial
spatial volume element
volume element

Example 157 It is possible to derive the divergence form of the continuity equation from
the material form using ³ ´
∇ · (ρv) = ∇ρ · v + ρ(∇ · v) (11.7)
and the definition of the material derivative.

11.1.2 Integral form

For a fixed volume Vf we find from (10.90) that
ZZZ ZZ
d
ρdV = − ρv · ndA , Vf is constant (11.8)
dt Vf ∂Vc
| {z } | {z }
rate of change net increase of
of mass in mass by
fixed volume convection

where n is a unit normal pointing out of the volume Vf .

From (10.88) we have the following equation for a control volume Vc where vc is the
velocity of the surface ∂Vc of the volume:
ZZZ ZZZ ZZ
d D
ρdV = ρdV − ρ (v − vc ) · ndA (11.9)
dt Vc Dt Vc ∂Vc

Using the principle of mass conservation as expressed in (11.2) we arrive at the result
ZZZ ZZ
d
ρdV = − ρ (v − vc ) · ndA (11.10)
dt Vc ∂Vc

11.1.3 Control volume with compressible fluid

Consider a control volume Vc which may be time varying, and which is filled with a
compressible fluid. Moreover, assume that the density ρ is the same all over the control
volume. The fluid is assumed to be compressible with bulk modulus β so that
dρ dp
= (11.11)
ρ β
Then from (10.88) we have the mass balance in the form
ZZZ ZZZ ZZ
d D
ρdV = ρdV − ρ (v − vc ) · ndA (11.12)
dt Vc Dt Vc ∂Vc
| {z } | {z } | {z }
rate of change This term equals net mass
of mass in zero in view of flow into
control volume (11.2) control volume
11.1. THE MASS BALANCE 419

This equation states that the time derivative of the mass in Vc is equal to the net mass
flow into the control volume, which makes sense. We denote the mass flow into the
volume by w1 = ρq1 , and the mass flow out of the volume by w2 = ρq2 , where q1 and q2
are the corresponding volumetric flows. Then the mass balance can be written
d
(ρVc ) = w1 − w2 = ρ(q1 − q2 ) (11.13)
dt
Moreover, assume that the density ρ is the same all over the control volume. The fluid
is assumed to be compressible with bulk modulus β so that
dρ dp ρ
= ⇒ ρ̇ = ṗ (11.14)
ρ β β

Then the mass balance of a control volume Vc with compressible fluid with bulk modulus
β can be written
Vc
ṗ + V̇c = q1 − q2 (11.15)
β

11.1.4 Mass flow through a pipe

A fluid of constant density is flowing through a pipe of cross section A with velocity v
along the direction of the pipe, which is the x direction with unit vector i. It is assumed
that the velocity is constant over the cross section, and given by v = v i. If the flow is
into the volume, then the outwards-pointing surface normal is n = −i, and the mass flow
is ZZ ZZ
w=− ρv · ndA = − ρv · ndA = ρvA (11.16)
A A

If the velocity varies over the cross section of the pipe, then the mass flow is
ZZ Z
w=− ρv · ndA = ρ vdA = ρv̄A (11.17)
∂Vc

where v̄ is the average velocity.

Let the control volume be Vc = AL, which is the fixed volume from x1 = 0 to x2 = L
of the pipe. Then the boundary ∂Vc of the volume Vc is the wall of the pipe plus the
cross sections at x1 and at x2 . The outwards-pointing normal vector is n = −i at x1 and
n = i at x2 . Then the mass balance is
d
mc = ρ1 v̄1 A − ρ2 v̄2 A = w1 − w2 (11.18)
dt
where ZZZ
mc = ρdV (11.19)
Vc

is the mass inside the control volume.

Example 158 We consider gas of density ρ in a fixed volume V shown in Figure 11.1
with inlet through a pipe of cross section A1 and outlet through a pipe of cross section
A2 . We suppose that ρ is constant over the fixed volume V , while the density is ρ1 at the
inlet. We assume that the velocity in the inlet pipe is in the x direction and of magnitude
420 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

A1 V,  A2
w1 w2

Figure 11.1: Volume V with mass flow w1 into the wolume and w2 out of the volume.

v1 . In the same way, the velocity in the outlet pipe is assumed to be in the x direction
and of magnitude v2 . The velocity is assumed to be constant over the cross section of the
pipe. Then the balance equation of mass is
dρ
V = A1 ρ1 v̄1 − A2 ρv̄2 (11.20)
dt
which may also be written
dm
= w1 − w2 . (11.21)
dt
Here m = ρV is the mass contained in the volume, w1 = A1 ρ1 v1 is the mass flow into
the volume, and w2 = A2 ρ2 v2 is the mass flow out of the volume.
Example 159 Water of constant density ρ is flowing into a tank of cross section A with
mass flow w1 and flows out with mass flow w2 . The water level is h. The mass balance
is
d
(ρAh) = w1 − w2 (11.22)
dt
which can be written
1
ḣ = (w1 − w2 ) (11.23)
ρA

11.1.5 Continuity equation and Reynolds’ transport theorem

It turns out that by combining the continuity equation with Reynold’s transport theorem
we can derive alternative forms of Reynold’s transport theorem. This is useful in the
development of the momentum balance in Section 11.2.1.
First it is noted that the divergence form (10.87) of Reynolds’ transport theorem for
the function ρφ gives
ZZZ ZZZ · ¸
D ∂ (ρφ)
ρφdV = + ∇ · (ρφv) dV (11.24)
Dt V V ∂t
Then it is used that the material form (10.86) of Reynolds’ transport theorem gives
ZZZ ZZZ · ¸
D D (ρφ)
ρφdV = + ρφ(∇ · v) dV
Dt V V Dt
ZZZ ½ · ¸¾
Dφ Dρ
= ρ +φ + ρ(∇ · v) dV (11.25)
V Dt Dt
The last two terms of the integrand cancel, which can be seen from the continuity equation
(11.5). This gives the following important result for a volume V .
11.1. THE MASS BALANCE 421

The continuity equation in combination with the transport theorem gives the result
ZZZ ZZZ
D Dφ
ρφdV = ρ dV (11.26)
Dt V V Dt

By comparing this with (11.24) and accounting for the fact that the volume V is arbitrary,
it is found that
Dφ ∂ (ρφ)
ρ = + ∇ · (ρφv) (11.27)
Dt ∂t

Note that the last term on the right hand side of (11.27) is a divergence term. The
importance of this is made clear by integrating the equation over a volume V and using
the divergence theorem. This gives
ZZZ ZZZ ZZ
Dφ ∂ (ρφ)
ρ dV = dV + ρφ (v · n) dA (11.28)
V Dt V ∂t ∂V

where V can be any volume, and n is the outwards pointing surface normal. We see that
the first term on the right side is the rate of change of the quantity of ρφ in the volume,
while the second term on the right side is the flow of ρφ into the volume over the volume
boundary.
We note that φ may be the component of a vector u, that is, φ = ui which leads to
the following vector equations

The continuity equation and the transport theorem for a vector u gives the results
ZZZ ZZZ
D Du
ρudV = ρ dV (11.29)
Dt V V Dt

and
Du ∂ (ρu)
ρ = + ∇ · (ρvu) (11.30)
Dt ∂t

The last term in (11.30) is verified in a Cartesian coordinate system with orthogonal
unit vectors ai with the following computation:
 
X ∂ X X X ∂
∇ · (ρvu) = ak · ρ vj aj ui ai  = (ρvj ui ) ai (11.31)
∂xk j i i
∂xj
k

The integral form of (11.30) is found to be

ZZZ ZZZ ZZ
Du ∂ (ρu)
ρ dV = dV + ρu (v · n) dA (11.32)
V Dt V ∂t ∂V

This result has a nice structure, and the terms on the right side has the same physical
interpretation as in the scalar case.
Finally we note that from the expressions in (10.10) of the material derivative, the
following alternative expressions are obtained
422 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

µ ¶
∂ (ρφ) Dφ ∂φ
+ ∇ · (ρφv) = ρ =ρ + v · ∇φ (11.33)
∂t Dt ∂t
µ ¶
∂ (ρu) Du ∂u
+ ∇ · (ρvu) = ρ =ρ + v · ∇u (11.34)
∂t Dt ∂t

11.1.6 Multi-component systems

To describe systems with chemical reactions we may need the continuity equation for a
volume with several components, and where different mass components are generated or
used in the chemical reactions. The presentation is adopted from the introductory part
of (de Groot and Mazur 1984). We consider a fluid with n components where there may
be r chemical reactions between the components. The mass mk of component k in a
material volume V satisfies
Xr ZZZ
d
mk = ν kj Jj dV (11.35)
dt j=1 V

where ν kj Jj is the rate of production of component k per unit volume in reaction j. Using
the transport theorem, this gives
ZZZ · ¸ Xr ZZZ
∂ρk
+ ∇ · (ρk vk ) dV = ν kj Jj dV (11.36)
V ∂t j=i V

where ρk is the density of component k, and vk is the velocity of component k. As the

volume V is arbitrary, it follows that the continuity equation of component k is

X r
∂ρk
+ ∇ · (ρk vk ) = ν kj Jj (11.37)
∂t j=i

Since mass is conserved in each of the chemical reactions it follows that

n
X
ν kj Jj = 0 (11.38)
k=i

Then, by adding the continuity equations of all components the continuity equation

∂ρ
+ ∇ · (ρv) = 0 (11.39)
∂t
where ρ is the total density
n
X
ρ= ρk (11.40)
k=1

and v is the barycentric velocity, which is defined as the velocity of the center of mass
n
X ρk vk
v:= (11.41)
ρ
k=1
11.2. THE MOMENTUM BALANCE 423

Define the barycentric material derivative by

D ∂
= +v·∇ (11.42)
Dt ∂t
where v is the barycentric velocity. Insertion into the continuity equation (11.37) for
component k gives
Xr
Dρk
− v · ∇ρk + ∇ · (ρk vk ) = ν kj Jj (11.43)
Dt j=i

The last term on the left side is expanded to give

X r
Dρk
− v · ∇ρk + ∇ · (ρk v) + ∇ · [ρk (vk − v)] = ν kj Jj (11.44)
Dt j=i

By defining the diffusion flow of component k as

jk = ρk (vk − v) (11.45)
and, accounting for (10.15), we find the following result:

The continuity equation for component k is found to be

Dρk ³ ´ X r
= −ρk ∇ · v − ∇ · jk + ν kj Jj (11.46)
Dt j=i

while from (11.39) and (11.42) the continuity equation for the total density is

Dρ ³ ´
+ρ ∇·v =0 (11.47)
Dt

Example 160 In terms of mass fractions

ρk
ck = (11.48)
ρ
the continuity equation for component k becomes
X r
Dck
ρ = −∇ · jk + ν kj Jj (11.49)
Dt j=i

which is found by inserting ρk = ρck in the continuity equation (11.46) for component k
and then use (11.47).

11.2 The momentum balance

11.2.1 Euler’s equation of motion
We consider a material volume element dV of a fluid with density ρ and velocity v. The
momentum of the volume element is ρvdV . Newton’s law for this set of particles is
D
(ρvdV ) = dF . (11.50)
Dt
424 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

x3

n

tn 

x2
x1

Figure 11.2: The stress vector in an inviscid fluid is parallel to the surface normal.

where dF is the force acting on the differential volume dV . Note that the material
derivative is used, as Newton’s law applies to a material volume element. The mass ρdV
of the particles in the material volume element is constant, and it follows that

D D Dv Dv
(ρvdV ) = (ρdV ) v + ρ dV = ρ dV (11.51)
Dt Dt Dt Dt
We may therefore write Newton’s law in the form

Dv
ρ dV = dF (11.52)
Dt

The force dF denotes the total force on the volume element, which is the mass force plus
the surface force. When this is integrated over the material volume V we get
ZZZ ZZZ
Dv
ρ dV = dF = F (r) (11.53)
V Dt V

where F (r) is the resultant force acting on the volume V . The surface forces cancel out
inside the volume due to Newton’s third law of action and reaction. This is referred to
as the principle of local equilibrium of the stresses. Because of this the total force F (r)
is given by the sum of surface forces acting on ∂V plus the mass force on the volume.
Assume that the fluid is inviscid in which case the only surface forces are the pressure
forces. This gives
ZZ ZZZ
F (r) = − pndA + ρf dV (11.54)
∂V V

where ρf is the mass force, and −pndA is the surface force in the form of pressure forces.
The divergence theorem and (10.13) then gives
ZZZ ZZZ ³ ´
Dv
ρ dV = −∇p + ρf dV (11.55)
V Dt V

The volume V is arbitrary, and this leads to Euler’s equation of motion

11.2. THE MOMENTUM BALANCE 425

Euler’s equation of motion for an inviscid fluid is given by

Dv
ρ = −∇p + ρf (11.56)
Dt
Alternative formulations of Euler’s equation found from (11.34) are the divergence form

∂ (ρv)
+ ∇ · (ρvv) = −∇p + ρf (11.57)
∂t
and the formulation ³ ´
∂v
ρ + ρ v · ∇ v = −∇p + ρf (11.58)
∂t

Example 161 We consider the one-dimensional case where the velocity is v in the x
direction. Then, if the pressure gradient is zero, the mass forces are zero, and ρ is a
constant, Euler’s equation as given by (11.58) gives
∂v ∂v
+v =0 (11.59)
∂t ∂x
which is known as Burger’s equation (Evans 1998). This simple equation is interesting
as it may have analytical solutions that can be used to check the accuracy of numerical
solution techniques, and it may exhibit shocks where the velocity gradient approaches
infinity.

11.2.2 The momentum equation for a control volume

From (11.29) we have the following expression
ZZZ ZZZ
D Dv
ρvdV = ρ dV (11.60)
Dt V V Dt
From (10.88) we have the following equation for a control volume Vc
ZZZ ZZZ ZZ
d D
ρvdV = ρvdV − ρv (v − vc ) · ndA (11.61)
dt Vc Dt Vc ∂Vc

where v is the velocity of the fluid and vc is the velocity of the surface ∂Vc of the control
volume. Combining these two equations with (11.53) we get
ZZZ ZZ
d (r)
ρvdV = F
|{z} − ρv (v − vc ) · ndA (11.62)
dt V ∂Vc
| {zc } resultant force
| {z }
rate of change on fluid in net increase of
of momentum control momentum
in control volume by convection
volume

Example 162 For the system in Example 158 the momentum conservation in the x
direction gives
ZZZ
d
vρdV = F + p1 A1 − p2 A2 + v1 w1 − v2 w2 (11.63)
dt V
426 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

where F is the force in the x direction acting on the gas from the tank, p1 A1 is the force
due to pressure on the inlet, and p2 A2 is the force due to pressure at the outlet. It is
assumed that the velocity is constant over the cross section.

11.2.3 Example: Waterjet

A1
n A2 
n
Pump
w w

Figure 11.3: Schematic diagram of a waterjet.

We consider a waterjet (Figure 11.3) where water enter through the intake which is
a pipe with cross section A1 , and flows out through an outlet pipe of cross section A2 .
A pump is used to force the water through the waterjet. The water flows axially in the
pipes with velocity v =v1 i = −v1 n at the inlet and v =v2 i = v2 n at the outlet where i
is the unit vector in the x direction. Stationary conditions are assumed. Moreover, the
water is assumed to be incompressible, so that the mass flow in is equal to the mass flow
out. Then the continuity equation gives

A1 ρv1 = A2 ρv2 = w (11.64)

where w is the mass flow. We assume that the pressure forces over the cross sections A1
and A2 of the pipes can be left out. Then the momentum equation in the x direction
gives
F + v1 A1 ρv1 − v2 ρA2 v2 = 0. (11.65)
We define the thrust T of the waterjet as the force from the fluid on the casing. The
thrust is given by T = −F , and we get the result
µ ¶
A2
T =− 1− wv2 ≈ −wv2 (11.66)
A1
where it is assumed that A2 ¿ A1 . We see that if the outlet area is much smaller than
the inlet area, then the thrust is equal to mass flow times outlet velocity, and that the
thrust is directed in the opposite direction of the flow through the waterjet. Suppose that
the outlet cross section is reduced. Then if the pump is sufficiently powerful so that the
mass flow w is unchanged, then v2 = w/ (A2 ρ) will increase, and the thrust T ≈ −wv2
will increase in magnitude.

11.2.4 Example: Sand dispenser and conveyor

Sand is dispensed from a container with mass flow w down on a conveyor belt as shown in
Figure 11.4. The conveyor belt is driven by a motor torque T acting on a shaft of radius
r with angular velocity ω m . The velocity of the conveyor belt is therefore v = ω m r.
11.2. THE MOMENTUM BALANCE 427

w
 m ,T

Figure 11.4: Sand of mass flow w falling down on a conveyor belt.

Here the mass m and the momentum p = mv of the sand are conserved quantities. The
balance equation for the mass is
d
m = w − we , (11.67)
dt
while the balance equation for the momentum is
d
(mv) = −vwe + F (11.68)
dt
Here F is the force from the conveyor belt on the sand. The equation of motion for the
belt is
J ω̇ m = T − F r (11.69)
where J is the inertia experienced by the motor. The equation of motion can be expressed
in terms of the velocity to give
J 1
v̇ = T − F. (11.70)
r2 r
The momentum equation gives

ṁv + mv̇ = −vwe + F (11.71)

and insertion of the mass balance and the equation of motion gives
µ ¶
J 1
m + 2 v̇ = T − vw (11.72)
r r
The results seem reasonable as the belt is slowed down when sand with zero horizontal
velocity falls down on the belt.

11.2.5 Irrotational Bernoulli equation

The convective term ρ(v · ∇)v in (11.58) can be written
³ ´ µ ¶ ³ ´
1 2
v · ∇ v=∇ v −v× ∇×v (11.73)
2
which can be verified by evaluation the components on both sides. It follows that for
irrotational flow, which occurs for ∇ × v = 0, the Euler equation can be written
µ ¶
∂v 1 2 1
+∇ v − f + ∇p = 0 (11.74)
∂t 2 ρ
428 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

Suppose that the fluid is incompressible so that ρ is a constant. Moreover, assume that
the mass force is the gradient f = −∇(gz) of the gravitational potential gz, where z is
the coordinate in the vertical upwards direction. As ∇ × v = 0 there will be a velocity
potential φ so that v = ∇φ. Then Euler’s equation can be written as the gradient
equation · ¸
∂φ 1 2 p
∇ + v + gz + =0 (11.75)
∂t 2 ρ
where it is used that ρ is a constant for incompressible fluids. This implies that
∂φ 1 2 p
+ v + + gz = constant (11.76)
∂t 2 ρ
which is the irrotational Bernoulli equation. In the stationary case we then have
1¡ 2 ¢ (p2 − p1 )
v2 − v12 + + (z2 − z1 ) g = 0 (11.77)
2 ρ
for irrotational flow of an inviscid and incompressible fluid.
Example 163 The velocity term can be expressed using the gradient of the velocity po-
tential, which gives
∂φ 1 ³ ´ p
+ ∇φ · ∇φ + + gz = constant (11.78)
∂t 2 ρ

11.2.6 Bernoulli’s equation along a streamline

It is seen from (11.58) and (11.73) that the Euler equation can be written
µ ¶ ³ ´
∂v 1 2 1
+∇· v − v × ∇ × v − f + ∇p = 0 (11.79)
∂t 2 ρ

To proceed we need to eliminate the term v × (∇ × v). There are two ways to do this
that give interesting results (White 1999). The first approach, which was discussed in
the previous section, is to require that ∇ × v=0, which is the case for irrotational flow.
The second approach, which will be investigated here, is to integrate the expression on
the left hand side of (11.79) along a streamline.
Consider the following integral form of the Euler equation (11.79):
Z · µ ¶ ³ ´ ¸
∂v 1 2 1
+∇· v − v × ∇ × v − f + ∇p · dx=0 (11.80)
∂t 2 ρ
where the differential dx is parallel to the velocity and satisfies dx/dt = v. Then
³ ´
v × ∇ × v · dx= 0 (11.81)

and the integral expression becomes

Z · µ ¶ ¸
∂v 1 2 1
+∇· v − f + ∇p · dx=0 (11.82)
∂t 2 ρ

We assume that f = −ga3 , and denote the vertical coordinate z = x3 , and write |dx| =
ds. This gives Z 2 Z 2 µ ¶ Z 2 Z 2
∂ |v| 1 2 dp
ds + d v + gdz + =0 (11.83)
1 ∂t 1 2 1 1 ρ
11.2. THE MOMENTUM BALANCE 429

x3  z

 v
g 
dx


xt 
1 2
streamline

x2

x1

Figure 11.5: Streamline x(t) with two points 1 and 2 on the streamline.

where 1 and 2 denotes two points on the same streamline. Two of the integrals are exact,
and we find that
Z 2 Z 2
∂ |v| 1¡ 2 2
¢ dp
ds + v2 − v1 + + g (z2 − z1 ) =0 (11.84)
1 ∂t 2 1 ρ

which is Bernoulli’s equation for frictionless flow along a streamline. Under stationary
conditions ∂ |v| /∂t = 0, and
Z 2
1¡ 2 ¢ dp
v2 − v12 + + (z2 − z1 ) g = 0 (11.85)
2 1 ρ

For incompressible flow ρ is a constant and

1¡ 2 ¢ (p2 − p1 )
v − v12 + + (z2 − z1 ) g = 0, 1 and 2 on a streamline (11.86)
2 2 ρ
which is the Bernoulli equation for stationary frictionless flow along a streamline for an
incompressible fluid. We see that if z1 = z2 , then the pressure along a streamline will
decrease when the velocity increases.
The additional assumption that was made for the irrotational Bernoulli’s equation
was that the flow is irrotational. The equation (11.77) is valid for arbitrary points 1 and
2 in the fluid, whereas Bernoulli’s equation (11.86) along a streamline is only valid if the
points 1 and 2 are on a streamline.

11.2.7 Example: Transmission line

A hydraulic transmission line is a pipe of cross section A and length L with a compressible
fluid. The dynamic model for a hydraulic transmission line is developed from the mass
balance and momentum balance of a differential control volume Adx where A is the cross
sectional area of the pipe and x is the length coordinate along the pipe. It is assumed
that the density of the fluid is not varying over the cross section, so that ρ = ρ(x, t). The
mass flow is Z
w(x, t) = ρvdA = ρv̄A (11.87)
A
430 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

A
q q  dq
, p

x x  dx

Figure 11.6: Volume element for hydraulic transmission line.

where v̄ is the average velocity. The mass balance is taken for the fixed differential control
volume Adx from x to x + dx. The mass flow into the volume is w at x, while the mass
flow out of the volume is w + dw at x + dx. The mass balance is then found from (11.18)
to be
∂ρ
Adx = w − (w + dw) = −dw
∂t
Dividing by Adx we get
∂ρ 1 ∂w
=− (11.88)
∂t A ∂x
A change of variables from density ρ to pressure p is achieved in the mass balance using
the constitutive equation dp = (β/ρ)dρ where β is the bulk modulus of the fluid. This
gives
∂p β ∂w
=−
∂t ρA ∂x
The momentum equation is found from (11.63) to be
Z Z
∂ 2
£ ¡ ¢¤
(ρv̄) Adx = Ap − A(p + dp) + ρv dA − ρv 2 + d ρv 2 dA − F dx (11.89)
∂t A A

where F dx is the friction force. This gives

Z
∂w ∂p ∂
= −A −A ρv 2 dA − F (11.90)
∂t ∂x ∂x A

We will assume that the average velocity v̄ is close to zero, so that the second term on
the right side can be set to zero. The model becomes

∂p β ∂w
= − (11.91)
∂t ρA ∂x
∂w ∂p
= −A −F (11.92)
∂t ∂x

These equations are usually formulated in terms of the pressure p and the volumetric
flow q by treating the density as a constant ρ0 so that w = ρ0 q.
11.2. THE MOMENTUM BALANCE 431

v1

v2

p 1 ,A 1 p 2 ,A 2

Figure 11.7: Incompressible fluid flowing through a pipe of cross section A1 with a
restriction with cross section A2 .

The transmission line model linearized around q = 0 and ρ = ρ0 is given by

∂p β ∂q
= − (11.93)
∂t A ∂x
∂q A ∂p F
= − − (11.94)
∂t ρ0 ∂x ρ0

11.2.8 Liquid mass flow through a restriction

We consider a liquid, that is an incompressible fluid, which flows through a pipe with cross
sectional area A1 with a restriction with cross sectional area A as shown in Figure 11.7.
The continuity equation implies that the mass flow w1 = ρq2 at the inlet is the same as
the mass flow w2 = ρq2 at the outlet. As the fluid is incompressible, this implies that
also the volumetric flow is the same at the inlet and the outlet, so that the volumetric
flow q is given by
q = v1 A1 = v2 A2 (11.95)

Bernoulli’s equation (11.86) gives

1 2 p1 1 p2
v1 + = v22 + (11.96)
2 ρ 2 ρ

which gives
" µ ¶2 #
ρ¡ 2 2
¢ ρ A2
p1 − p2 = v − v1 = 1− v22
2 2 2 A1
" µ ¶2 # 2
A2 ρq2
= 1− (11.97)
A1 2A2

This gives the following result:

432 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

 t ,T t
h q
pp +

Ap A

Figure 11.8:

Frictionless and incompressible flow through a restriction A2 in a pipe with cross section
A1 is given by v
u 2 (p − p )
q = A2 u
1 2
t ³ ´2 (11.98)
ρ
1− A A1
2

If the flow is from a volume, then A1 → ∞, and the expression becomes

r
2
q = A2 (p1 − p2 ) (11.99)
ρ

This expression (11.98) is adjusted with the discharge coefficient Cd to account for the
effect that the cross section of the flow will be somewhat smaller than the cross section
A2 of the restriction. This gives
v
u 2 (p − p )
q = Cd A2 u
1 2
t ³ ´2 (11.100)
ρ A2
1 − A1

At very low flow rates the friction will be the dominating physical phenomenon. Then
Bernoulli’s equation is no longer valid, and the flow becomes linear in the pressure dif-
ference. This is discussed in Section 4.2.2.

11.2.9 Example: Water turbine

Model
In this section we will study the dynamics of a hydroelectric power system consisting of
a pipe that transports water from a reservoir with water level h to an impulse turbine
with a Pelton wheel (White 1999) at water level 0. Between the outlet of the pipe and
the turbine there is a control device that sets the cross section A of the water flowing
into the turbine. The cross section A is the input control variable of the system, while
the turbine torque Tt is the output. The turbine torque is of interest as the equation of
11.2. THE MOMENTUM BALANCE 433

motion for the turbine shaft is

Jt ω̇ t = Tt − TL (11.101)
where Jt is the moment of inertia of the turbine shaft, ω t is the turbine shaft speed and
TL is the load torque which will typically be the driving torque for an electrical generator.
The model will be developed by deriving the model for the pipe, the model for the control
device, and the model for the turbine. Then the complete model is obtained by connecting
the three component models. This approach makes it easy at a later stage to change the
pipe model from an incompressible flow model to a compressible flow model. Also this
approach will hopefully give some structure to the presentation so that the reader will
not get lost in the many equations.
The pipe is of length L, and has inlet at the elevation h where the inlet pressure is
zero. The outlet of the pipe has pressure pp and volumetric flow q. We treat the pipe
as a two-port with pressure and volumetric flow as port variables. The pressure at the
line ends are the inputs to the model of the pipe. The inlet pressure is supposed to be
the constant ambient pressure pa = 0. Therefore, the flow q at the outlet of the pipe
will depend on the outlet pressure pp . To get a result that will be valid for different pipe
models, we will at this stage assume that the linearized dynamics of the pipe are given
by the transfer function
−∆pp
Hpq (s) := (s) (11.102)
∆q
where ∆q = q − q0 and ∆pp = pp − pp0 are deviations from a constant solution (q0 , pp0 ).
Note that the negative pressure change −∆pp is used in the definition of the transfer
function to ensure that the Hpq (s) has positive gain.
The inlet of the control device has a constant cross section Ap , and the inlet pressure
is pp . At the outlet of the control device the cross section is controlled to A, the pressure
is p, and water velocity is v = q/A. It is assumed that the outlet pressure p is small
and constant so that p = 0 can be used. It is assumed that the mass of the water in
the control device is small so that Bernoulli’s equation applies to describe the relation
between the inlet pressure and velocity and the outlet pressure and velocity. According
to (11.96) this gives µ ¶
ρ q2 q2
pp = − 2 (11.103)
2 A2 Ap
Linearization of the control device equation (11.103) around the nominal area A0 and a
corresponding nominal flow q0 gives
ραq0 ρq 2
∆pp = 2 ∆q − 30 ∆A
A0 A0

where α = 1 − A20 /A2p . Dividing by ∆q and rearranging we find that

ρq02 ∆A ραq0 −∆pp

3 (s) = + (s) (11.104)
A0 ∆q A20 ∆q
The transfer function from the control input A to the flow q is then found to be given
by
∆q q0 1
HqA (s) := (s) = (11.105)
∆A αA0 1 + A20 Hpq (s)
αρq0

where we have used (11.102). We note that the control device can be connected to a
particular pipe by inserting the transfer function Hpq (s) of the pipe.
434 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

The shaft torque Tt for an impulse turbine with a Pelton wheel is given by (White
1999) µ 2 ¶
q
Tt = 2rt ρq(v − rt ω t ) = 2rt ρ − qrt ω t (11.106)
A
where rt is the radius of the wheel. We will treat the shaft speed ω t as a constant in the
linearization of the shaft torque. Linearization of the turbine torque equation (11.106)
will then give µ ¶
q02 A0
∆T = 2rt ρ 2 β ∆q − ∆A (11.107)
A0 q0
where
rt ω t0 A0
β =2−
q0
is a constant of linearization. The transfer function from the control input A to the
turbine torque is then found to be
µ ¶
∆Tt q2 A0
(s) = 2rt ρ 02 β HqA (s) − 1 (11.108)
∆A A0 q0
Insertion of HqA (s) from (11.105) gives
 
∆Tt q02  β 1
(s) = 2rt ρ 2 A20
− 1
∆A A0 α 1 +
αρq0 Hpq (s)

and some algebra leads to the transfer function in the form (Hutarew 1969), (Ervik 1971)
A2
∆Tt 2rt ρ q02 1 − γ αρq00 Hpq
(s) = (11.109)
∆A γ A20 1 + A20 Hpq
αρq0

where the constant γ is given by

1 1
γ= β
≈ rpt ω t0 A0
α −1 1− q0

The power on the turbine shaft is P = T ω t , and if we assume α = 1, then it is a

straightforward exercise to show that β = 1.5 and γ = α/(1.5 − α) ≈ 2 at full load where
the power is maximized.

Water turbine with incompressible water supply

The pipe is of length L and cross section Ap , and the reservoir has water level h. The
water is assumed to be incompressible with density ρ. The volumetric flow is q, and the
velocity of the water is vp = q/Ap . The equation of motion for water in the pipe is

Lρq̇ = mgh + Ap (p0 − pp ) (11.110)

where mgh is the constant gravity force in the flow direction that acts on the water in
the pipe, p0 is the constant ambient pressure, and pp is the pressure at the end of the
pipe. Laplace transformation leads to the pipe transfer function
−∆pp ρLs
Hpq (s) := (s) = (11.111)
∆q Ap
11.2. THE MOMENTUM BALANCE 435

Note that the negative pressure is used in the definition of the transfer function Hpq (s)
to achieve a transfer function with a positive gain. The transfer function HqA (s) can
then be found from (11.105) to be
q0 1
HqA (s) =
αA0 1 + µ T2r s

where we have defined the time constant Tr and the nondimensional flow constant µ by

LA20 qmax q0
Tr = 2 , µ= (11.112)
αq02 Ap qmax

The transfer functions for the complete system is found from (11.109) to be
¡ ¢
∆Tt 2 rt ρq02 1 − γµ T2r s
(s) = ¡ ¢ (11.113)
∆A γ A20 1 + µ T2r s

At full load with µ = 1 and γ = 2, the transfer function is

∆Tt rt ρq02 (1 − Tr s)
(s) = ¡ ¢ (11.114)
∆A A20 1 + T2r s

Example 164 Francis or Kaplan type turbines are reaction turbines that are driven by
power transfer from the water flow. The shaft torque is
P
Tf t = (11.115)
ωt
where µ ¶
1 2 ρ q3
P = qρ v = (11.116)
2 2 A2
is the power supplied to the turbine. Linearization of the power expression gives
µ ¶
ρ q02 q0
∆P = 3∆q − 2 ∆A (11.117)
2 A20 A0
Then the transfer function from A to P can be found from
µ ¶
∆P ρ q02 q0
(s) = 3HqA (s) − 2
∆A 2 A20 A0

by inserting (11.105). This gives

µ ¶ µ ¶
A20 ∆P ρq03 3 A20
1+ Hpq (s) (s) = 3 −2−2 Hpq (s)
ραq0 ∆A A0 α ραq0
and, using the reasonable approximation α = 1, we arrive at the well-known power trans-
fer function (Hutarew 1969)

∆P ρq 3 (1 − µTr s)
(s) = 30 ¡ ¢ (11.118)
∆A A0 1 + µ T2r s

where Tr and µ are given in (11.112).

436 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

Water turbine with compressible water supply

We now include compressibility effects in the supplying pipe. The inlet of the pipe is
open, so the transfer function from the volumetric flow w = ρq at the lower end of the
pipe to the pressure p at the same place is given by (4.180) and (4.195) as
−∆p ρc L
Hpq = (s) = tanh s (11.119)
∆q Ap c
Note that Hpq (s) tends to the incompressible solution ρLs/Ap when c → ∞, which corre-
sponds to the incompressible case where β → ∞. We find that when the compressibility
effects of the water in the pipe is included the transfer functions to torque and power
becomes ³ ´
A2 c L
2 1 − γ αq0 A tanh c s
∆Tt 2 rt ρq0 0 p
(s) = ³ ´ (11.120)
∆A γ A20 1+
A20 c
tanh L s
αq0 Ap c

11.2.10 Example: Waterhammer

L, A
v P

Figure 11.9: The waterhammer effect occurs when the pipe is suddenly closed at P .

The waterhammer effect (Merritt 1967), (Holmboe and Rouleau 1967) occurs when
fluid is flowing through a pipe, and the pipe is suddenly closed for example by a valve
(Figure 11.9). A fluid with velocity v and density ρ flowing in a pipe of length L and
cross section A will have a kinetic energy
1
K= ρV v 2 (11.121)
2
where V = LA is the volume of the fluid. We note that for a material volume V of a set
of particle with mass m and density ρ = m/V the volume differential is
µ ¶
m m mρ V
dV = d = − 2 dρ = − 2 dp = − dp (11.122)
ρ ρ ρ β β
Then it follows that if the fluid is instantaneously stopped the kinetic energy K will give
an increase ∆P due to compression, which is given by
Z 2 Z 2
V 1V ¡ 2 ¢
∆P = − pdV = p dp = p2 − p21 (11.123)
1 1 β 2β
where p1 is the pressure just before the pipe is closed, and p2 is the pressure just after
the pipe is closed. From K = ∆P the pressure increase is seen to be
q
p22 − p21 = ρcv (11.124)
11.3. ANGULAR MOMENTUM BALANCE 437
p
where c = β/ρ is the sonic speed. In the case that the initial pressure p1 is small, this
is approximated by
p2 = ρcv (11.125)
Pa
Example 165 For water c = 1500 m/s and ρ = 103 kg/m3 , and p2 = 1.5 · 106 m/s · v, or
atm
p2 = 15 m/s ·v, so that 5 m/s gives a pressure rise of 75 bar. For hydraulic fluids c = 1250
atm
m/s and ρ = 800 kg/m3 which gives p2 = 10 m/s · v so that 5 m/s gives a pressure rise of
50 bar.

11.3 Angular momentum balance

11.3.1 General expression
The angular momentum equation is important in the modeling of compressors and tur-
bines. Whereas the momentum equation is derived from Newton’s law for an infinitesimal
material volume, the angular momentum equation is derived from Euler’s law of angular
momentum
D
ρ (r × v) dV = r × dF (11.126)
Dt
for a material volume element dV . Here r is the position vector of the volume element
from a specified point o. The force dF denotes the resultant force on the volume element,
which is the mass force plus the surface force. When this is integrated over the material
volume V we get
ZZZ ZZZ
D D
ρ (r × v) dV = r × ρvdV = No (11.127)
V Dt Dt V

where ZZZ
No = r × dF (11.128)
V

is the moment about the point o.

The angular momentum equation is given by

ZZZ
D
r × ρvdV = No (11.129)
Dt V

For a general control volume Vc the angular momentum equation is written

ZZZ ZZ
d
r × ρvdV + (r × ρv) (v − vc ) · ndA = No (11.130)
dt Vc ∂Vc

where vc is the velocity of a point on the surface of Vc .

11.3.2 Centrifugal pump with radial blades

A pump is a device where power is supplied from the pump axis to a fluid to make the
fluid flow with a mass flow w. The pump axis may be driven by an electrical motor, an
engine or a turbine. We will first consider a centrifugal pump with radial blades acting
on an incompressible fluid (Figure 11.10). The pump has angular shaft velocity ω. The
438 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

er e r 2

r 1

r1
r2 
+

Figure 11.10: Centrifugal pump with radial blades.

fluid enters in the center, and flows through an arrangement of radial blades with inner
blade tips at a radius r1 and outer blade tips at a radius r2 . We define a frame with
orthogonal unit vectors er , eθ , ez where er is in the radial direction, eθ in the tangential
direction, and ez is along the pump axis. We note that the inner tip speed of the blades
is U1 = r1 ωeθ while the outer tip speed of the blades is U2 = r2 ωeθ . We will assume that
the fluid flow is constant and with a mass flow

w = 2πr1 bv1r = 2πr2 bv2r (11.131)

where b is the width of the pump, v1r is the radial fluid velocity at the blade inlet, and
v2r is the radial fluid velocity at the blade outlet. We will consider the moment about
the pump axis, which means that the point o is in the center of the pump, so that we
have r1 = r1 er at the blade inlet and r2 = r2 er at the outlet. The fluid velocity at the
blade inlet is denoted v1 and the velocity at the blade outlet is denoted v2 where
w w
v1 = er + r1 ωeθ , v2 = er + r2 ωeθ (11.132)
2πr1 b 2πr2 b
This gives
r1 × v1 = r12 ωez , r2 × v2 = r22 ωez (11.133)
The control volume Vc is taken to be the volume between the blade inlet and the blade
outlet. This is a volume that is fixed in space, so that vc = 0. The outwards pointing
surface normal is n = −eθ at the inlet and n = eθ at the outlet. The angular momentum
balance (11.130) gives
ZZZ
∂ ¡ 2 ¢ ¡ ¢
ρr ωez dV + w r22 ω − r12 ω ez = Tp (11.134)
Vc ∂t
11.3. ANGULAR MOMENTUM BALANCE 439

where Tp = Tp ez is the load torque on the shaft. This gives

¡ ¢
Jf ω̇ + wω r22 − r12 = Tp (11.135)

where
πbρ ¡ 4 ¢ mf ¡ 2 ¢
Jf = r2 − r14 = r1 + r22 (11.136)
2 2
¡ ¢
is the moment of inertia due to the fluid, and mf = πbρ r22 − r12 is the mass of the fluid
in Vc . We see that the stationary shaft torque needed to pump a mass flow of w is
¡ ¢
Tp = wω r22 − r12 (11.137)

The shaft power is ¡ ¢

Pp = Tp ω = wω 2 r22 − r12 (11.138)

11.3.3 Euler’s turbomachinery equation

In a more well-designed centrifugal pump the blades will be curved, and the blades will
have an inlet angle β 1 and outlet angle β 2 . The velocity v1 at the inlet and the velocity
v2 at the blade outlet is written

v1 = v1r er + v1t eθ , v2 = v2r er + v2t eθ (11.139)

We get
r1 × v1 = r1 v1t ez , r2 × v2 = r2 v2t ez (11.140)
Then, proceeding as in the previous section, we get the shaft power

Tp = w (r2 v2t − r1 v1t ) (11.141)

The shaft power is found to be

Pp = T ω = wω (r2 v2t − r1 v1t ) (11.142)

A turbine is a device where a fluid delivers power to the turbine shaft by changing
the momentum of the fluid. This means that a turbine converts kinetic energy in a fluid
to mechanical energy in the form of rotational energy of the shaft. We note that for the
centrifugal pump the shaft torque T is zero when the shaft speed ω is zero. This shows
that the centrifugal pump with radial blades cannot be used as a turbine.

11.3.4 Pump instability

The direction of the velocity vectors v1 and v2 are described by the flow angles
v1t v2t
tan α1 = , tan α2 = (11.143)
v1r v2r

We define W1 and W2 by

v1 = U1 + W1 , v2 = U2 + W2 (11.144)

W1 = W1r er + W1t eθ , W2 = W1r er + W1t eθ (11.145)

At the blade outlet the fluid flow will be along the blade, so that the velocity W2 will
have direction given by the blade outlet angle β 2 . At design speed a design rule is to
440 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

select the inlet blade angle β 1 so that the inlet flow will be along the blade at the inlet,
so that W1 will have direction given by β 1 . Then

W1t = −v1r cotanβ 1 , W2t = −v2r cotanβ 2 (11.146)

will be the tangential fluid velocities relative to the blades. We will consider the situation
when there is no pre-whirl, which means that the tangential speed at the blade inlet is
zero. Then
v1r
v1t = 0 ⇒ tan β 1 = (11.147)
U1
and the torque is found to be

T = wr2 (U2 − v2r cotanβ 2 )

µ ¶
w
= wr2 ωr2 − cotanβ 2 (11.148)
2πr2 bρ

Suppose that the pump is delivering an incompressible fluid to a pipe of cross section
A and length L. The velocity at the inlet of the pipe is denoted v, and it is assumed that
the mass flow is
w = ρAv (11.149)
The equation of motion for the fluid is

ρALv̇ = F − Fout (11.150)

where Fout is the force acting at the pipe outlet. We assume that the shaft power T ω
is converted to kinetic power F v for the fluid in the pump so that T ω = F v. Then the
force F at the inlet of the pipe is found to be
µ ¶
ω ω Av
F = T = ρA T = ρAωr2 ωr2 − cotanβ 2 (11.151)
v w 2πr2 b

and the equation of motion becomes

ρA2 ω
ρALv̇ = ρAω 2 r22 − v cotanβ 2 − Fout (11.152)
2πb
The force consist of a term that is proportional to ω 2 which can be considered as the
forcing term. In addition there is the second term on the right side of (11.152) which is
proportional to the outlet fluid velocity v. If β 2 > 90◦ , which is the case if the blade have
a backsweep at the outlet, then the velocity term will have the same effect as viscous
friction, and has a stabilizing effect. However, if the blades are swept forward, then
β 2 > 90◦ , and the second term on the right side of (11.152) will give the same effect as
a positive velocity feedback, which may cause the system to be unstable.

Example 166 The pump delivers an incompressible fluid through a pipe of cross section
A to a basin. The fluid level in the basin is√denoted h. Water flows out of the basin
through at throttle with mass flow wt (h) = C h. The model for the system is

A g r2 ω 2
v̇ = − ωcotanβ 2 v − h + 2 (11.153)
2πbL L L
A 1
ḣ = v− wt (h) (11.154)
Ab Ab ρ
11.4. THE ENERGY BALANCE 441

where the pump velocity ω 2 is considered to be the control input. This can be achieved by
velocity control of the motor driving the pump. Linearization gives
v̇ = a11 v + a12 h + bω 2 (11.155)
ḣ = a21 v + a22 h (11.156)
and the characteristic equation of the linearized system is found to be
λ2 − (a11 + a22 ) λ − a12 a21 = 0 (11.157)
Stability results whenever
A 1 dwt
a11 + a22 = − ωcotanβ 2 − <0 (11.158)
2πbL Ab ρ dh
which is the case if
2πbLω dwt
cotanβ 2 < − (11.159)
ρAAb dh
This means that if the blade outlets are backswept so that cotanβ 2 ≥ 0, then the system
will be stable. Forward swept blade outlets may cause instability depending on the system
parameters.
Example 167 Under stationary conditions is may be assumed that the mechanical power
T ω from the shaft is converted to power F v supplied to the fluid, so that
ω
F = T (11.160)
v
In transients there will be energy loss until the stationary flow pattern is established. It is
reasonable to assume that these transient flow will last for at least the time time it takes
a fluid particle to flow through the pump, and in some cases up to 5 times of this time.
Then a reasonable model for the transients in the shaft torque is
1 ³ω ´
Ḟ = T −F (11.161)
αTtrans v
where Ttrans can be taken to be the transport time of a fluid particle through the pump,
and α is in the range from 1 to 5.

11.4 The energy balance

11.4.1 Material volume
A material volume has a fixed set of particles. Therefore the total energy of a material
volume is conserved. This means that the rate of change of the total energy of a material
volume is equal to the net rate of energy supplied to the volume. We assume here that
the total energy in a volume element dV is ρedV where
1
e = u + v2 + φ (11.162)
2
is the specific energy, u is the specific internal energy, (1/2)v 2 is the specific kinetic energy,
and φ is the specific potential energy. We assume that the body forces are derived from
the potential φ in the sense that
Dφ ³ ´
∇φ = −f ⇒ = ∇φ · v= −f · v (11.163)
Dt
442 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

The material time derivative of the total energy in a volume V is equal to the net rate of
energy supplied to the volume. Suppose that the net supplied energy is the sum of the
heat flow into the volume due to the heat flux density jQ plus the power added from the
pressure force −pn acting on the surface. This is written
ZZZ ZZ ZZ
D
ρedV = − pv · ndA − jQ · ndA (11.164)
Dt V ∂V ∂V

The volume V is arbitrary, and it follows from the divergence theorem that
µ ¶
D 1 2
ρ u+ v +φ =− ∇ · (pv) − ∇ · jQ (11.165)
Dt 2 | {z } | {z }
| {z }
pressure work heat
rate of change
on the surface of conduction
in internal, kinetic
the volume element
and potential energy
for material
volume element

The divergence form is found by changing the left hand side as follows:
De ∂
ρ = (ρe) + ∇ · (ρev) (11.166)
Dt ∂t
If we leave out the potential energy, then (11.163) can be used to express the energy
equation in the form
µ ¶
D 1 2
ρ v +u = − ∇ · (pv) − ∇ · jQ
Dt 2 | {z } | {z }
| {z }
pressure work heat
rate of change
on the surface of conduction
in internal and
the volume element
kinetic energy
for material
volume element

+ ρv · f (11.167)
| {z }
work of body
forces on volume
element

Example 168 If the pressure is constant over the volume, then the pressure work can
be written ZZ ZZ
DV
pv · ndA = p v · ndA = p (11.168)
∂V ∂V Dt
which means that the pressure work is equal to the pressure times the time derivative of
the material volume.

11.4.2 Fixed volume

If the volume V is fixed then the energy balance can be written in the material form
using ZZZ ZZZ ZZ
D d
ρedV = ρedV + ρev · ndA (11.169)
Dt V dt V ∂V
11.4. THE ENERGY BALANCE 443

Insertion of (11.164) gives the result

ZZZ ZZ µ ¶ ZZ
d p
ρedV = − ρ e+ v · ndA − jQ · ndA (11.170)
dt V ∂V ρ ∂V

where the first term on the right side is the convected energy plus the pressure work on
the volume. At this stage it is useful to introduce the specific enthalpy h which is defined
by
p
h=u+ (11.171)
ρ
Then the energy balance can be written
ZZZ µ ¶ ZZ µ ¶
d 1 1 2
ρ u + v 2 + φ dV = − ρ h+ v + φ v · ndA
dt V 2 2
| {z } | ∂V {z }
rate of change convected enthalpy,
of energy kinetic energy and
in fixed volume potential energy
ZZ
− jQ · ndA (11.172)
∂V
| {z }
heat
conduction

Note that in the convection term the enthalpy h enters in place of the internal energy u
as the pressure work is included in the convection term.

Example 169 Suppose that the specific energy of the system in Example 158 is simply
e = u, which means that the kinetic and potential energy can be neglected. Moreover,
suppose that there is no heat flow into the volume, that is, jQ = 0, that there is no mass
or energy generation in the volume, and that ρ and u are constants over the volume.
Then the energy balance is

d
V (uρ) = u1 A1 ρ1 v1 − uA2 ρv2 + p1 A1 v1 − pA2 v2 (11.173)
dt
which gives
d
E = h1 w1 − hw2 (11.174)
dt
where we used E = mu where m = ρV , and where we have used the enthalpy h = u+p/ρ.
We may obtain an equation for the specific internal energy u by expanding the left side.
This gives
mu̇ + ṁu = h1 w1 − hw2 . (11.175)
Combining this with the mass balance

ṁ = w1 − w2 (11.176)

we get a differential equation for the specific internal energy in the form

mu̇ = h1 w1 − hw2 − u (w1 − w2 ) (11.177)

444 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

Steam in
Fluid
inlet
Instantaneous
liquid level

Condensate out

Figure 11.11: Water tank heated by a coil filled with steam.

which gives
w1 w2 p
u̇ = (h1 − u) − (11.178)
m mρ
We will later see that this leads to a differential equation for the temperature by using
h = cp T and u = cv T .

Example 170 This example and the next example are adopted from (Bird, Stewart and
Lightfoot 1960, p. 473). A cylindrical tank with cross section A is filled with a liquid with
a mass flow w (Figure 11.11). The volume of the liquid in the tank is V = Az where z
is the height of the liquid surface. The liquid in the tank is heated with a coil filled with
steam of temperature Ts . The heat transfer coefficient per length unit of the coil from the
coil to the liquid is G. The tank is stirred so that the temperature of the liquid in the tank
is uniform. The energy of the liquid is supposed to be u = cp T . The mass and energy
balances are
d
(ρV ) = w (11.179)
dt
d
(ρuV ) = wu1 + Gz (Ts − T ) . (11.180)
dt
The first term on the right side of the energy balance is the convected internal energy,
while the second term is a heat conduction term as in the general expression (11.172).
The energy balance can be written out as
µ ¶
d dT
ρV cp T + ρV cp = wcp T1 + Gz (Ts − T ) . (11.181)
dt dt

Insertion of the mass balance in the energy balance gives

dT
ρV cp = wcp T1 − wcp T + Gz (Ts − T ) (11.182)
dt
11.4. THE ENERGY BALANCE 445

w2
P T, ,V

T 1 ,w 1

Figure 11.12: Heated tank.

and some straightforward manipulations lead to the model

w
ż = (11.183)
ρA
w G
Ṫ = (T1 − T ) + (Ts − T ) . (11.184)
ρAz ρAcp
Example 171 A liquid is heated by pumping it through a tank with an electrical heating
element supplying the power P as shown in Figure 11.12. The temperature of the liquid
in the tank is T, the density ρ is constant, and the volume of the tank is V . The inlet
has mass flow w1 and temperature T1 , while the outlet has mass flow w2 . The liquid
flowing through the outlet has the temperature T of the liquid in the tank. The energy of
the liquid in the tank is mcp T where m = ρV is the mass of the liquid in the tank. The
mass balance implies that
w1 = w2 = w (11.185)
The energy balance is then
d
(mcp T ) = cp w (T1 − T ) + P. (11.186)
dt
From the mass balance we have ṁ = 0, and we get
w P
Ṫ = (T1 − T ) + . (11.187)
m cp m

11.4.3 General control volume

For a general control volume Vc Reynolds’ transport theorem (10.88) gives
ZZZ ZZZ ZZ
d D
ρedV = ρedV − ρe (v − vc ) · ndA (11.188)
dt Vc Dt Vc ∂Vc

From equation (11.164) we have the following expression for the material derivative of
the energy: ZZZ ZZ ZZ
D
ρedV = − pv · ndA − jQ · ndA (11.189)
Dt Vc ∂Vc ∂Vc
446 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

Combining the two equations we find that

ZZZ ZZ µ ¶
d p
ρedV = − ρ e+ (v − vc ) · ndA
dt Vc ∂Vc ρ
ZZ ZZ
− pvc · ndA − jQ · ndA (11.190)
∂Vc ∂Vc

The first term on the right side is the convected energy plus the pressure work on the
volume. The specific enthalpy h = u + p/ρ is inserted. Then the energy balance can be
written
ZZZ µ ¶ ZZ µ ¶
d 1 2 1 2
ρ u + v + φ dV = − ρ h + v + φ (v − vc ) · ndA
dt Vc 2 ∂Vc 2
| {z } | {z }
rate of change convected enthalpy,
of energy kinetic energy and
in control volume potential energy
ZZ ZZ
− pvc · ndA − jQ · ndA (11.191)
∂Vc ∂Vc
| {z } | {z }
pressure work heat
due to change in conduction
control volume

Note that the velocity in the convection term is v − vc which is the particle velocity
relative to the surface of the control volume Vc .

Example 172 If the pressure is constant over the volume then the pressure work on the
surface of the control volume can be written
ZZ ZZ
pvc · ndA = p vc · ndA = pV̇c (11.192)
∂Vc ∂Vc

11.4.4 The heat equation

Heat conduction in a solid is described by the heat equation. The energy appears in the
form of internal energy u = cp T , and energy flow is due to heat conduction according to
the constitutive equation in the form of Fourier’s law

jQ = −α∇(ρcp T ) (11.193)

where the α is the thermal diffusivity in m2 /s. The energy balance is simply

∂u
ρ = −∇ · jQ (11.194)
∂t
which in combination with Fourier’s law with constant α and ρ gives
∂T
− α∇2 T = 0 (11.195)
∂t

where ∇2 = ∇ · ∇ is the Laplacian operator.

11.4. THE ENERGY BALANCE 447

11.4.5 Transfer function for the heat equation

The heat equation in one dimension for the temperature T (x, t) in a bar from x = 0 to
x = L is given by
∂T (x, t) ∂ 2 T (x, t)
−α =0 (11.196)
∂t ∂x2
Suppose that the bar is insulated at x = 0, and that the heat flux jQ is controlled at
x = L according to jQ = −αρcp u where u is the control variable. Then the boundary
conditions are
∂T (0, t) ∂T (L, t)
= 0, =u (11.197)
∂x ∂x
Laplace transformation of (11.196) gives

∂ 2 T (x, s) s
2
− T (x, s) = 0 (11.198)
∂x α
which has the solution
µr ¶ µr ¶
s s
T (x, s) = A cosh x + B sinh x (11.199)
α α

with derivative
r µr ¶ r µr ¶
∂T (x, s) s s s s
=A sinh x +B cosh x (11.200)
∂x α α α α

The boundary condition at x = 0 gives B = 0, and the boundary condition at x = L

gives r µr ¶
s s
A sinh L =u (11.201)
α α
so that the temperature is given by
¡p s ¢
cosh αx
T (x, s) = p s ¡p s
¢ u(s) (11.202)
α sinh αL

The transfer function from the heat flux to the temperature at x = L is found to be
¡p s ¢
T (L, s) cosh αL
= ps ¡p s
¢ (11.203)
u(s) α sinh αL

The zeros of the transfer function are found from

r ³
s π´ s 1 ³ π ´2
L = j kπ + ⇒ = − 2 kπ + (11.204)
α 2 α L 2

while the singularities are given by

r
s s 1 2
L = jkπ ⇒ = − 2 (kπ) (11.205)
α α L

Numerical values are given in Table 11.1.

448 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

Zeros Singularities
p ps
L αs L2 αs L α L2 αs
1. 570 8 −2. 467 4 0 0
(11.206)
4. 712 4 −22. 207 3. 141 6 −9. 869 6
7. 854 0 −61.685 6. 283 2 −39. 478
10.995541 −120.9019 9. 424 8 −88. 826

Table 11.1: Singularites for the one-dimensional heat equation when the beam is insulated
at x = 0, and the heat flux is controlled at x = L.

Example 173 The heat equation is studied for the bar of the previous example, but the
boundary condition at x = L is changed so that the bar is in contact with a reservoir with
temperature u, which is the control input. The heat-transfer coefficient is β. Then the
boundary conditions are changed to
∂T (0, t) ∂T (L, t)
= 0, = β[u − T (L, t)] (11.207)
∂x ∂x
The boundary condition at x = 0 gives B = 0, while the boundary condition at x = L in
combination with (11.199) gives
r µr ¶ · µr ¶¸
s s s
A sinh L = β u(s) − A cosh L
α α α
This implies that
β
A = ps ¡p s ¢ ¡p s ¢ u(s) (11.208)
α sinh α L + β cosh αL
and insertion in (11.199) gives the transfer function
¡p s ¢
T (L, s) cosh L
= 1ps ¡p s ¢ α ¡p s ¢ (11.209)
u(s) β α sinh α L + cosh αL

11.5 Viscous flow

11.5.1 Introduction
So far the balance equations for momentum and energy have been developed for inviscid
fluids, that is, for fluids without viscosity. In some problems, viscous effects may be
important, and in the following balance equations for the viscous case will be developed.
The mathematical level is somewhat more advanced than for the inviscid case. The
main reason for this is the appearance of the viscous stress tensor which necessitates the
introduction of tensor notation.

11.5.2 Tensor notation

The derivation of certain important results in fluid mechanics are best done in tensor
notation (Aris 1989), (Lovelock and Rund 1989). This involves a systematic notation
11.5. VISCOUS FLOW 449

for doing vector operations at the component level. We will see in the following that
tensor notation is of particular use in connection with the computation of gradients and
divergence of complicated vector expressions. All tensors in the following are Cartesian.
Let a be a Cartesian frame with orthogonal unit vectors a1 , a2 , a3 and let u be a vector
and D be a dyadic given by
3
X 3 X
X 3
u= ui ai , D = dij ai aj (11.210)
i=1 i=1 j=1

where ui = u · ai is component i of u and dij = ai · D · aj is component i, j of D. The

vector u and the dyadic D are uniquely defined by their components. This means that
we may represent u by its generic component ui and we may represent D by its generic
component dij .

The generic component ui = u · ai of a vector u is a first order Cartesian tensor, while

the generic component dij = ai · D · aj of a dyadic D is a second order Cartesian tensor.

We introduce the summation convention where the summation symbol may be left
out when it is summed over a repeated index in a product as in the scalar product
3
X
ui vi := ui vi = uT v (11.211)
i=1

The summation convention also applies to the vector expression

3
X
u= ui ai = ui ai (11.212)
i=1

and the dyadic expression

3 X
X 3
D= dij ai aj = dij ai aj (11.213)
i=1 j=1

In addition we will use a notation where subscript , i denotes partial differentiation

with respect to xi , so that

∂φ ∂vi
φ,i := and vi,j := (11.214)
∂xi ∂xj

In addition, subscript , ij denotes partial differentiation with respect to xi and xj , that

is,
¡ ¢ ∂2φ ∂ 2 vi
φ,ij := φ,i ,j = and vi,jk := (vi,j ),k = (11.215)
∂xi ∂xj ∂xj ∂xk
We let the summation convention apply to differentiation expressions so that

∂ ³ T ´
3
X 3
X
vj,ji := vj,ji = ∇ v and vi,jj := vi,jj = ∇2 vi (11.216)
j=1
∂xi j=1
450 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

Example 174 The tensor form of the material derivative of a scalar φ is

Dφ ∂φ
= + vi φ,i (11.217)
Dt ∂t
while the material derivative of a vector u with elements ui is
Dui ∂ui
= + vj ui,j (11.218)
Dt ∂t

Example 175 The divergence of velocity ∇ · v can written

3
X
∇ · v= vi,i = vi,i (11.219)
i=1

while the scalar product between the v and the gradient of a scalar ψ can be written
3
X
v · ∇ψ = vi ψ ,i = vi ψ ,i (11.220)
i=1

Theorem 1 The divergence theorem (10.12) can be written in tensor notation as

ZZ ZZZ
ui ni dA = ui,i dV (11.221)
∂V (t) V (t)

Moreover, the related result (10.13) is written

ZZ ZZZ
φni dA = ∇φ,i dV (11.222)
∂V (t) V (t)

Example 176 The Laplacian of a scalar φ is

3
X
∇2 φ = φ,ii = φ,ii (11.223)
i=1

where ∇2 = ∇ · ∇= ∂x
∂ ∂
i ∂xi
is the Laplacian operator.

Example 177 The Hessian matrix of a scalar φ is

 ∂ ∂ ∂ ∂ ∂ ∂   
∂x1 ∂x1 ∂x1 ∂x2 ∂x1 ∂x3 φ,11 φ,12 φ,13
∇∇T φ =  ∂
∂x2
∂
∂x1
∂
∂x2
∂
∂x2
∂ ∂
∂x2 ∂x3
 φ =  φ,21 φ,22 φ,23  (11.224)
∂ ∂ ∂ ∂ ∂ ∂ φ,31 φ,32 φ,33
∂x3 ∂x1 ∂x3 ∂x2 ∂x3 ∂x3
n o
where ∇∇T = ∂ ∂
∂xi ∂xj is the Hessian operator in matrix form.

Example 178 The scalar ∇·(φu), which is the divergence of the vector φu can be written
in tensor notation as (φui ),i . From the usual rule for the differentiation of products it
follows that
(φui ),i = φ,i ui + φui,i (11.225)
which is the tensor form of (10.15).
11.5. VISCOUS FLOW 451

Example 179 Matrix multiplication is conveniently expressed in tensor notation. Let

x = {xi }, y = {yi }, A = {aij }, B = {bij } C = {cij }, and D = {dij }. Then it is
straightforward to verify that
y = Ax ⇔ yi = aij xj (11.226)
C = AB ⇔ cij = aik bkj (11.227)
D = AT B ⇔ dij = aki bkj (11.228)
E = ABT ⇔ eij = aik bjk (11.229)

11.5.3 The velocity gradient tensor

Viscous forces appear in fluids because of velocity gradients. To describe velocity gradi-
ents it is convenient to introduce the velocity gradient tensor defined by
∂vi
vi,j = (11.230)
∂xj
This is the tensor form of the velocity gradient dyadic
∇v = vi,j ai aj (11.231)
while the corresponding matrix form is (∇vT )T .

The velocity gradient tensor vi,j is written

vi,j = eij + Ωij (11.232)

where
1
eij := (vi,j + vj,i ) (11.233)
2
is the symmetric rate of strain tensor , which is also called the deformation tensor, and
1
Ωij := (vi,j − vj,i ) (11.234)
2
is the skew-symmetric part of the velocity gradient tensor.

The matrix form of the rate of strain tensor is written

½ ¾
1
E := {eij } = (vi,j + vj,i ) (11.235)
2
while ½ ¾
1
Ω := {Ωij } = (vi,j − vj,i ) (11.236)
2
Example 180 Define ds by
ds2 = dxi dxi (11.237)
T
which means that ds is the length of the differential vector dx = (dx1 , dx2 , dx3 ) . Then
the material derivative of ds2 is found to be
1 D 1 D D
(ds)2 = (dxi dxi ) = dxi dxi
2 Dt 2 Dt Dt
D ∂xi ∂vi
= dxi dξ j = dxi dξ (11.238)
Dt ∂ξ j ∂ξ j j
452 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

x3 v0
  x

p vp
x
o 


x2
x1

Figure 11.13: Rigid body with velocity vo of point o and angular velocity ω.

where ξ j are the material coordinates. Then, because

∂vi ∂vi
dξ = dxj (11.239)
∂ξ j j ∂xj
it follows that
1 D 2
(ds) = dxi eij dxj (11.240)
2 Dt
We see that the deformation tensor eij is related to the stretching of the differential dx.

11.5.4 Example: The velocity gradient for a rigid body

Consider a rigid body with angular velocity ω(t) and velocity vo and position ro (t) of
some specified point o, which is fixed in the rigid body. We consider a fixed point p in
the rigid body with position r(t, x1 , x2 , x3 ). The velocity of the point p is
v (t, x1 , x2 , x3 ) = vo (t) + ω (t) × r(t, x1 , x2 , x3 ) (11.241)

where r = x − ro . The velocity gradient ∇v which describes the velocity variations over
the rigid body is
∇v = ∇vo + ∇ (ω × r) = ω × ∇r = ω × ∇x = ω × · ∇x = ω × · I
= ω× (11.242)
Here we have used the fact that vo , ro and ω are functions of time only, and that
∇r = ∇x = I. We see that for rigid-body motion the velocity gradient tensor is skew
symmetric and given by
∇v = ω × (11.243)
This means that for rigid body motion we have
Ω = ω× , eij = 0 (11.244)
The opposite is also true: If the rate of strain tensor eij is zero, then the motion is a
rigid body motion. From this we conclude that a nonzero rate of strain tensor eij is a
measure of how much the velocity field differs from rigid body motion, and in this sense
eij is related to the rate of deformation of the fluid.
11.5. VISCOUS FLOW 453

n
tn    T
x3
n

x2
x1

Figure 11.14: Volume V with surface normal n and stress vector t(n) .

Example 181 We recognize 2Ω in (11.236) as the skew symmetric form of the curl
vector ∇× v as given in (10.21). This means that
¡ × ¢×
∇ v =2Ω (11.245)
From this and (11.244) we conclude that for rigid body motion we have
∇ × v = 2ω (11.246)
and ³ ´×
∇×v = 2∇v (11.247)

11.5.5 The stress tensor

RRR
The forces acting on a material volume V (t) are divided into mass forces ρf dV and
RR V
surface forces ∂V t(n) dA. Here t(n) is the stress vector and t(n) dA is the contact force
acting on the area element dA. The subscript (n) indicates that the stress vector t(n) is
acting on a surface with outwards surface normal n as shown in Figure 11.14. The stress
vector t(n) can be expressed in the form

t(n) = n · T

where T is the dyadic form of the the stress tensor Tij . We note that the corresponding
matrix representation is T = {Tij }. By convention, the element Tij denotes the contact
force in the j direction exerted on a plane with surface normal in the i direction.
It is not straightforward to handle divergence terms involving the stress tensor in
vector form. Because of this we use the more powerful tensor notation in component
form, where the summation convention is used. Then, component i of the stress vector
t(n) is written
t(n)i = Tji nj (11.248)
where the summation over j is implied by the summation convention. The divergence
theorem (11.221) leads to the equation
ZZ ZZZ
Tji nj dA = Tji,j dV (11.249)
∂V V
454 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

where the notation

3
∂Tji X ∂Tji
Tji,j = = (11.250)
∂xj j=1
∂xj

is used. In vector form this is written

ZZ ZZ ZZZ ³ ´
t(n) dA = n · T dA = ∇ · T dV (11.251)
∂V ∂V V

The stress tensor can generally be written

T = −pI + τ (11.252)

where p is the pressure and τ is the viscous stress tensor . We note that the equivalent
tensor form is
Tij = −pδ ij + τ ij (11.253)
where δ ij is the Kronecker delta which is given by δ ii = 1 and δ ij = 0 for i 6= j. From

Tji,j = − (pδ ij ),j + τ ji,j = −p,i + τ ji,j (11.254)

we find that
∇ · T = −∇p + ∇ · τ (11.255)
and we may use (11.251) to establish the following useful relation:
ZZ ZZZ ³ ´
t(n) dA = −∇p + ∇ · τ dV (11.256)
∂V V

which has the component form

ZZ ZZZ
t(n)i dA = (−p,i + τ ji,j ) dV (11.257)
∂V V

Example 182 We note that the matrix form of the main equations of this section is
given by
T = −pI + τ (11.258)
³ ´T ³ ´T
∇T T = −∇p + ∇T τ (11.259)

and ZZ ZZZ · ³ ´T ¸
T
t(n) dA = −∇p + ∇ τ dV (11.260)
∂V V

Remark 4 Some authors use T = −pI − τ , which means that they use the opposite sign
for the viscous stress tensor.

11.5.6 Cauchy’s equation of motion

The resultant force acting on a fluid of volume V is the sum of the surface forces and the
body forces, which is written
ZZ ZZZ
(r)
F =− t(n) dA + ρf dV (11.261)
∂V V
11.5. VISCOUS FLOW 455

The divergence theorem gives

ZZZ ³ ´
F (r) = ∇ · T + ρf dV (11.262)
V

and by combining this with the momentum balance (11.53) the result is
ZZZ ZZZ ³ ´
Dv
ρ dV = ∇ · T + ρf dV (11.263)
V Dt V

As the volume V is arbitrary, this leads to Cauchy’s equation of motion

Dvi
ρ = Tji,j + ρfi (11.264)
Dt

Dv
ρ = ∇ · T + ρf (11.265)
Dt
which is stated both in component and vector form.
We express the stress tensor in terms of the pressure p and the viscous stress tensor
τ ij as given in (11.253).

Cauchy’s equation of motion for a fluid with viscosity is given in tensor form and vector
form as
Dvi
ρ = −p,i + τ ji,j + ρfi (11.266)
Dt
Dv
ρ = −∇p + ∇ · τ + ρf (11.267)
Dt
Cauchy’s equation of motion can be written in divergence form by inserting (11.30). This
gives
∂ (ρv)
+ ∇ · (ρvv) = − ∇p + ∇ · τ + ρf (11.268)
∂t

Example 183 The component form of Cauchy’s equation of motion (11.266) is

Dv1 ∂p ∂τ 11 ∂τ 21 ∂τ 31
ρ = − + + + + ρf1 (11.269)
Dt ∂x1 ∂x1 ∂x2 ∂x3
Dv2 ∂p ∂τ 12 ∂τ 22 ∂τ 32
ρ = − + + + + ρf2 (11.270)
Dt ∂x2 ∂x1 ∂x2 ∂x3
Dv3 ∂p ∂τ 13 ∂τ 23 ∂τ 33
ρ = − + + + + ρf3 (11.271)
Dt ∂x3 ∂x1 ∂x2 ∂x3

Example 184 The momentum equation for a multi-component fluid is

X n
Dv
ρ =∇·T + ρk fk (11.272)
Dt
k=1

We see that the only difference from a single-component fluid is in the term related to
the mass forces fk .
456 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

11.5.7 Newtonian fluids

The results that have been derived in the previous sections for the stress tensor are
kinematic and are valid for any fluid. The constitutive equations give expressions for
the stress tensor of a particular fluid. Fluids can be arranged in different classes of
fluids depending on which constitutive equations they satisfy. An example of this are
the Newtonian fluids, which include water and oil. It is shown that the insertion of
the viscous stress tensor of an incompressible Newtonian fluid in Cauchy’s equation of
motion leads to the Navier-Stokes equation.
Some notable fluids are not Newtonian. One example is blood, and another is the
mud used in oil drilling. Non-Newtonian fluids will not be further discussed, but it is
remarked that the equations of motion for such fluids can be obtained by inserting the
viscous stress tensor of the specific fluid into Cauchy’s equation of motion (11.266).

Newtonian fluids are fluids where the stress tensor is given by

³ ´
τ = λ ∇ · v I+2µE (11.273)

where λ and µ are the Lamé coefficients, and E=eij ai aj is the rate of strain tensor
defined by (11.233).

The component form of the constitutive equation (11.273) is

τ ij = λvk,k δ ij + 2µeij (11.274)
We note that for a Newtonian fluid the stress tensor is symmetric, so that τ ji = τ ij . To
evaluate the term τ ji,j = τ ij,j in Cauchy’s equation of motion (11.266) we first calculate
1 1 1
eij,j = (vi,j + vj,i ),j = (vi,jj + vj,ij ) = (vi,jj + vj,ji ) (11.275)
2 2 2
This gives
τ ij,j = (λvk,k δ ij ),j + 2µeij,j
= λvk,ki + µvj,ji + µvi,jj (11.276)
We may change indices in vj,ji by noting that j is a dummy index so that vj,ji = vk,ki .
This gives
τ ij,j = (λ + µ) vk,ki + µvi,jj (11.277)
In vector form this is written
³ ´ 2
∇ · τ = (λ + µ) ∇ ∇ · v + µ∇ v (11.278)

Example 185 Let the velocity field be given by v = (cx22 , 0, 0)T , or equivalently, by
v = cx22 a1 (11.279)
where a1 , a2 , a3 are orthogonal unit vectors along the x1 , x2 , x3 axes of a Cartesian co-
ordinate frame. The position is given by x = (x1 , x2 , x3 )T . Then the velocity gradient
tensor is
µ ¶
∂ ∂ ∂
∇v = a1 + a2 + a3 cx22 a1
∂x1 ∂x2 ∂x3
= 2cx2 a2 a1 (11.280)
11.5. VISCOUS FLOW 457

The deformation tensor is the symmetric part of the velocity gradient tensor and is given
by
E = cx2 a2 a1 + cx2 a1 a2 (11.281)
The divergence is found to be
∇·v =0 (11.282)
while the Laplacian is
µ ¶
2 ∂2 ∂2 ∂2
∇ v= 2 + 2+ 2 cx22 a1 = 2ca1 (11.283)
∂x1 ∂x2 ∂x3
Then the viscous stress tensor is
τ = 2µcx2 a2 a1 + 2µcx2 a1 a2 (11.284)
We see that the term ∇ · τ in the equation of motion is given by
2
∇ · τ = µ∇ v = 2µca1 (11.285)
Consider the surface with the surface normal n = a2 . Then the stress vector on that
surface is
t(n) = n · τ = 2µcx2 a1 (11.286)
The stress vector is seen to be in the direction of the flow.
Example 186 It is common practice to use an assumption due to Stokes, which involves
setting the relation between the Lamé coefficients so that
2
λ+ µ=0 (11.287)
3
We will explain the motivation for this relation. Denote the mean of the diagonal terms
of the stress tensor Tij by −p̄. By noting that the divergence is related to the diagonal
terms of the rate of strain tensor according to
∇ · v=vk,k = ekk (11.288)
we find that
1 2
−p̄ :=
Tii = −p + (λ + µ)ekk (11.289)
3 3
The reasoning that leads to (11.287) starts with the observation that for incompressible
fluids ekk = 0, which implies that p̄ = p. If it is assumed that (11.287) is valid, then
p̄ = p also for compressible fluids. More details on this is found in (Aris 1989).
Example 187 The constitutive equation for a Newtonian fluid originates from three
assumptions (Aris 1989): First it is assumed that the stress tensor is linear in the velocity
gradient tensor, which means that the viscous stress tensor can be written
τ ij = Nijkl vk,l (11.290)
for some four-dimensional tensor Nijkl . Second, it is assumed that Nijkl is isentropic.
This implies that Nijkl can be expressed in terms of the Lamé parameters λ, µ and κ as
Nijkl = λδ ij δ kl + µ(δ ik δ jl + δ il δ jk ) + κ(δ ik δ jl − δ il δ jk ) (11.291)
Third, symmetry is assumed in the sense that τ ij = τ ji , which implies that κ = 0. This
gives the constitutive equation for a Newtonian fluid as
λδ ij δ kl + µ(δ ik δ jl + δ il δ jk )vk,l = λδ ij vk,k + µ(vi,j + vj,i ) (11.292)
458 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

11.5.8 The Navier-Stokes equation

If it is assumed that a fluid is Newtonian, then the momentum equation in the form of
Cauchy’s equation of motion is found by inserting (11.278) into (11.266)

Cauchy’s equation of motion for a Newtonian fluid is given by

Dv ³ ´ ³ 2 ´
ρ = −∇p + (λ + µ)∇ ∇ · v + µ ∇ v + ρf (11.293)
Dt

We note that the tensor form is

Dvi
ρ = −p,i + (λ + µ) vk,ki + µvi,jj + ρfi (11.294)
Dt
We may write this out in each of the three orthogonal directions as

Dv1 ∂p ∂ ³ ´ ³ 2 ´
ρ = − + (λ + µ) ∇ · v + µ ∇ v + ρf1 (11.295)
Dt ∂x1 ∂x1
Dv2 ∂p ∂ ³ ´ ³ 2 ´
ρ = − + (λ + µ) ∇ · v + µ ∇ v + ρf2 (11.296)
Dt ∂x2 ∂x2
Dv3 ∂p ∂ ³ ´ ³ 2 ´
ρ = − + (λ + µ) ∇ · v + µ ∇ v + ρf3 (11.297)
Dt ∂x3 ∂x3
The divergence form of momentum equations for a Newtonian fluid is found to be

∂ (ρv) ³ ´ 2
+ ∇ · (ρvv) = −∇p + (λ + µ)∇ ∇ · v + µ∇ v + ρf (11.298)
∂t
or, in tensor form,

∂ (ρvi )
+ (ρvi vj ),j = −p,i + (λ + µ)vk,ki + µvi,jj + ρfi (11.299)
∂t
In the incompressible case the divergence of the velocity is zero, and the equation of
motion is found by inserting ∇ · v = 0 into (11.293). This leads to the Navier-Stokes
equation.

Cauchy’s equation of motion for an incompressible Newtonian fluid is given by

Dv 2
ρ = −∇p + µ∇ v + ρf (11.300)
Dt
This is the classical Navier-Stokes equation

The tensor form of the classical Navier-Stokes equation is

Dvi
ρ = −p,i + µvi,jj + ρfi (11.301)
Dt
The Euler equation for inviscid flow appears by setting the viscous forces to zero by
inserting µ = 0. Control of the Navier-Stokes equation is treated in great detail in
(Aamo and Krstić 2003).
11.5. VISCOUS FLOW 459

Example 188 Insertion of the identity

2
³ ´ ³ ´
∇ v =∇ ∇·v −∇× ∇×v (11.302)

gives the alternative form

Dv ³ ´ ³ ´
ρ = −∇p + (λ + 2µ)∇ ∇ · v − µ∇ × ∇ × v + ρf (11.303)
Dt
of the momentum equations for a Newtonian fluid. It is interesting to note how the
divergence ∇ · v and the curl ∇ × v of the velocity appear in the two terms due to the
stress tensor. If the fluid is incompressible, then the divergence term is zero, and if there
is potential flow, then the curl term is zero.

Example 189 In cylindrical coordinates the Navier-Stokes equation is found by applying

the results of Section 10.3.2 to the coordinate-free form

Dv 2
ρ = −∇p + µ∇ v + ρf (11.304)
Dt
of (11.300) where
v = vr jr + vθ jθ + vz jz (11.305)
and
∂ jθ ∂ ∂
∇ = jr + + jz (11.306)
∂r r ∂θ ∂z
This results in
∂vr ∂vr vθ ∂vr ∂vr v2 1 ∂p
+ vr + + vz − θ =−
∂t ∂r r ∂θ ∂z r ρ ∂r
µ ¶
µ ∂ 2 vr 1 ∂ 2 vr ∂ 2 vr 1 ∂vr 2 ∂vθ vr
+ + + + − − (11.307)
ρ ∂r2 r2 ∂θ2 ∂z 2 r ∂r r2 ∂θ r2
∂vθ ∂vθ vθ ∂vθ ∂vθ vr vθ 1 ∂p
+ vr + + vz + =−
∂t ∂r r ∂θ ∂z r ρr ∂θ
µ ¶
µ ∂ 2 vθ 1 ∂ 2 vθ ∂ 2 vθ 1 ∂vθ 2 ∂vr vθ
+ + 2 + + + 2 − 2 (11.308)
ρ ∂r2 r ∂θ2 ∂z 2 r ∂r r ∂θ r
∂vz ∂vz vθ ∂vz ∂vz 1 ∂p
+ vr + + vz =−
∂t ∂r r ∂θ ∂z ρ ∂z
µ 2 2 2
¶
µ ∂ vz 1 ∂ vz ∂ vz 1 ∂vz
+ + 2 + + (11.309)
ρ ∂r2 r ∂θ2 ∂z 2 r ∂r

Further details on the systematic derivation of these equations are found in (Aris 1989)
where also results for spherical coordinates are presented.

11.5.9 The Reynolds number

The Navier-Stokes equation can be made dimensionless by using the dimensionless vari-
ables
v x tU p fL
v∗ = , x∗ = , t∗ = , p∗ = , ∇∗ = ∇L, f∗ = (11.310)
U L L ρU 2 U2
460 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

where U is a characteristic velocity and L is a characteristic length. This gives

Dv∗ U 2 ∗ ∗ ρU
2
2 U U
ρ ∗
= −∇ p + µ (∇∗ ) v∗ 2 + ρf ∗ 2 (11.311)
Dt L L L L
which simplifies to
Dv∗ 1 2
∗
= −∇∗ p∗ + f ∗ + (∇∗ ) v∗ (11.312)
Dt Re
where Re is the Reynolds number given by
UL
Re = (11.313)
ν
and
µ
ν= (11.314)
ρ
is the kinematic viscosity. It is seen from (11.312) that the Reynolds number indicates
the relative importance of viscosity compared to inertial forces.

11.5.10 The equation of kinetic energy

The kinetic energy 12 ρv2 is not a conserved quantity. However, an equation for the kinetic
energy can be found from the momentum equation by observing that
µ ¶ µ ¶
D 1 2 Dv
ρ v =v· ρ (11.315)
Dt 2 Dt
This means that the equation for the kinetic energy can be found by premultiplying
Cauchy’s equation of motion (11.267) by v, which gives
µ ¶ ³ ´
D 1 2
ρ v = ρv · f − v · ∇p + v · ∇ · τ (11.316)
Dt 2
= ρvi fi − vi p,i + τ ij,i vj (11.317)

In the last term on the right hand side we have interchanged the i and j index, which is
possible as the term is a scalar, so that vi τ ji,j = τ ij,i vj .
From the product rule of differentiation of products it is straightforward to verify the
expression
(τ ij vj ),i = τ ij,i vj + τ ij vj,i (11.318)
The vector form of (11.318) is not quite as simple to derive, but we state the result,
which is ³ ´ ³ ´
∇ · (τ · v) = v · ∇ · τ + τ : ∇v (11.319)
where : is the double dot product defined for four vectors u, v, w, z by

uv : wz = (u · w) (v · z) (11.320)

For two dyadics D = dij ai aj and E = eij ai aj this gives

D : E = dij eij (11.321)

For matrix representations D = {dij } and E = {eij } the double dot product is defined
in the same way as the scalar D : E =dij eij .
11.5. VISCOUS FLOW 461

Equation (11.319) is verified in a Cartesian frame a by the dyadic computations

∂ ∂
∇ · (τ · v) = ak · τ ij vj ai = (τ ij vj ) = (τ ij vj ),i (11.322)
∂xk ∂xi
³ ´ µ ¶
∂ ∂τ ij
v· ∇·τ = vl al · ak · τ ij ai aj = vj = τ ij,i vj (11.323)
∂xk ∂xi
³ ´ ∂vl ∂vj
τ : ∇v = τ dij ai aj : ak al = τ ij = τ ij vj,i (11.324)
∂xk ∂xi
To complete the discussion we mention that the matrix formulation of the result can be
shown on the component level to be
³ ´ ³ ´
∇T (τ v) = ∇T τ v + τ : ∇vT (11.325)

where at least the second term on the right side is not obvious in a derivation based on
the use of the matrix notation.
Combination of (11.319) and (11.316) gives the following result:

The equation for the kinetic energy can be written

µ ¶ ³ ´
D 1 2
ρ v = ρv · f − ∇ · (pv) + p ∇·v
Dt 2 | {z } | {z } | {z }
| {z }
work of body pressure work
rate of change reversible
forces on volume on volume
in kinetic energy conversion
element element surface
for material to internal
volume element energy
³ ´
+ ∇ · (τ · v) − τ : ∇v (11.326)
| {z } | {z }
viscous work irreversible viscous
on surface of conversion to
volume element internal energy

Using the divergence theorem we find that the integral for of the equation for kinetic
energy is
ZZZ ZZZ h ³ ´ ³ ´i ZZ
D ρ 2
v dV = ρv · f + p ∇ · v − τ : ∇v dV + v·(−pn + n · τ ) dA
Dt Vc 2 Vc ∂Vc

Example 190 Still another expression for the left hand side of (11.326) is found from
(11.24): µ ¶ · µ ¶ µ ¶ ¸
D 1 2 ∂ 1 2 1 2
ρ v =ρ v +∇· v v (11.327)
Dt 2 ∂t 2 2
´ 191 For a Newtonian fluid the stress tensor is symmetric, and the term τ :
Example
³
T
∇v = τ : E is found to be positive from the calculation
³ ´ ³ ´
τ : ∇v = λ ∇ · v I : E + 2µE : E
= λ (eii )2 + 2µeij eij (11.328)
³ ´
which clearly shows that τ : ∇v ≥ 0.
462 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

Example 192 Let the velocity field be given by v = (cx22 , 0, 0)T , or equivalently, by

v = cx22 a1 (11.329)

where a1 , a2 , a3 are orthogonal unit vectors along the x1 , x2 , x3 axes of a Cartesian coor-
dinate frame. The position is given by x = (x1 , x2 , x3 )T . Then the deformation tensor
is
E = cx2 a2 a1 + cx2 a1 a2 (11.330)
irreversible conversion to internal energy is found from (11.328) to be
³ ´
τ : ∇v = λ (eii )2 + 2µeij eij = 4µc2 x22 (11.331)

which clearly is nonnegative. The viscous work on the surface corresponds to the diver-
gence term
µ ¶
∂ ∂ ∂
∇ · (τ · v) = a1 + a2 + a3 · (2µcx2 a2 a1 + 2µcx2 a1 a2 ) · cx22 a1
∂x1 ∂x2 ∂x3
µ ¶
∂ ∂ ∂
= a1 + a2 + a3 · 2µc2 x32 a2
∂x1 ∂x2 ∂x3
= 6µc2 x22 (11.332)

Example 193 If the stress tensor Tij is symmetric, then

1
Tji vi,j = (Tji vi,j + Tij vj,i ) = Tij eij (11.333)
2
where eij is the rate of strain tensor. This gives
³ ´
T : ∇v = T : E, T symmetric (11.334)

11.5.11 The energy balance for a viscous fluid

The material time derivative of the total energy in a volume V is equal to the heat flow
into the volume due to the heat flux density jQ plus the power added from the contact
stress
t(n) = −pn + n · τ (11.335)
acting on the surface. This is written
ZZZ ZZ ZZ ZZ
D
ρedV = − pv · ndA + v · (n · τ ) dA − jQ · ndA (11.336)
Dt V ∂V ∂V ∂V

The volume V is arbitrary, and it follows from the divergence theorem that
µ ¶
De 1
ρ u + v2 + φ = − ∇ · (pv) + ∇ · (τ · v) − ∇ · jQ
Dt 2 | {z } | {z } | {z }
| {z }
pressure work viscous work heat
rate of change
on the surface of on the surface of conduction
in internal, kinetic
the volume element the volume element
and potential energy
for material
volume element
11.5. VISCOUS FLOW 463

Example 194 The equation for the internal energy is found by subtracting (11.326)
from this equation, which gives
Du ³ ´ ³ ´
ρ =− p ∇·v + τ : ∇v − ∇ · jQ (11.337)
Dt}
| {z | {z } | {z } | {z }
rate of change reversible irreversible viscous heat
in internal energy conversion conversion to conduction
for material to internal internal energy
volume element energy

11.5.12 Fixed volume

If the volume Vf is fixed then Reynolds’ transport theorem (10.90) gives
ZZZ ZZZ ZZ
d D
ρedV = ρedV − ρev · ndA (11.338)
dt Vf Dt Vf ∂Vf

Insertion of (11.336) gives the result

ZZZ ZZ µ ¶ ZZ ZZ
d p
ρedV = − ρ e+ v · ndA + v · (n · τ ) dA − jQ · ndA
dt Vf ∂Vf ρ ∂Vf ∂Vf
(11.339)
where the first term on the right side is the convected energy plus the pressure work on
the volume. Then the energy balance can be written
ZZZ µ ¶ ZZ µ ¶
d 1 1
ρ u + v 2 + φ dV = − ρ h + v 2 + φ v · ndA
dt Vf 2 ∂Vf 2
| {z } | {z }
rate of change convected enthalpy,
of energy kinetic energy and
in fixed volume potential energy
ZZ ZZ
+ v · (τ · n) dA − jQ · ndA (11.340)
∂Vf ∂Vf
| {z } | {z }
viscous work heat
on volume surface conduction

Note that in the convection term the enthalpy h enters in place of the internal energy u
as the pressure work is included in the convection term.

11.5.13 General control volume

For a general control volume V Reynolds’ transport theorem (10.88) gives
ZZZ ZZZ ZZ
d D
ρedV = ρedV − ρe (v − vc ) · ndA (11.341)
dt Vc Dt Vc ∂Vc

From equation (11.336) we then find that

ZZZ ZZ µ ¶ ZZ
d p
ρedV = − ρ e+ (v − vc ) · ndA − pvc · ndA
dt Vc ∂Vc ρ ∂Vc
ZZ ZZ
+ v · (n · τ ) dA − jQ · ndA (11.342)
∂Vc ∂Vc
464 CHAPTER 11. MASS, MOMENTUM AND ENERGY BALANCES

where the first term on the right side is the convected energy plus the pressure work on
the volume. Insertion of h = u + p/ρ gives
ZZZ ZZ µ ¶
d 1
ρ (e) dV =− ρ h + v 2 + φ (v − vc ) · ndA
dt Vc ∂Vc 2
| {z } | {z }
rate of change convected enthalpy,
of energy kinetic energy and
in control volume potential energy
ZZ ZZ ZZ
− pvc · ndA + v · (τ · n) dA − jQ · ndA (11.343)
∂Vc ∂Vc ∂Vc
| {z } | {z } | {z }
pressure work viscous work heat
due to change in on volume surface conduction
control volume

Note that the velocity in the convection term is v − vc which is the particle velocity
relative to the surface of the control volume Vc .
Chapter 12

Gas dynamics

12.1 Introduction
The balance equations must be combined with thermodynamic results in the modeling
of systems where gas flow plays an important part. This is the case in the modeling of
engines, gas turbines and compressors, and for pipeline dynamics in the production and
transport of oil and gas. Engines, turbines and compressors are designed so that losses
due to viscosity and heat conduction are kept at a minimum. It will be shown in the
following that such losses will be associated with the generation of entropy in the system.
Because of this a well designed and optimized system will typically have gas dynamics that
are close to isentropic. On background of this observation the concept of isentropic gas
dynamics will be developed in the beginning of this chapter, and the results on isentropic
gas dynamics will be used to formulate balance equations for important systems with gas
dynamics.

12.2 Energy, enthalpy and entropy

12.2.1 Energy
We consider a gas of mass m in a control volume V . The energy of the gas in the mass
element dm is called the specific energy e, which is supposed to be given by
1
e = u + v2 + φ (12.1)
2
where u is the specific internal energy, 12 v 2 is the specific kinetic energy, and φ is the
specific potential energy. The energy E of the total volume is given by
ZZZ ZZZ ZZZ ZZZ
1 2
E = eρdV = uρdV + v ρdV + φρdV (12.2)
V V V 2 V
= U +K +Φ (12.3)
where U is the internal energy, K is the kinetic energy and Φ is the potential energy.

12.2.2 Enthalpy
Enthalpy appears in convection terms in the energy balance to account for the convection
of internal energy plus pressure work. This is seen in Example 169 and in equation

465
466 CHAPTER 12. GAS DYNAMICS

(11.343). The specific enthalpy is of great importance in gas flow problems where both
internal energy and pressure work appear. This is the case in the modeling of compressors
and turbines used in turbochargers, jet engines and power plants, and in the description
of pipelines for gas transport. Note that enthalpy is not a conserved quantity.
The inverse of density ρ appears in many equations, and it is customary to define the
specific volume V̂ as volume per mass unit, which is the inverse of density. In accordance
with the notation used for other specific quantities, we would have liked to use v for
specific volume, but we already used v for velocity. Therefore we use the notation
1
V̂ = (12.4)
ρ
The specific enthalpy is defined by

h = u + pV̂ (12.5)

where V̂ = 1/ρ is the specific volume. Enthalpy is not a conserved quantity.

From the definition of specific enthalpy (12.5) we see that

dh = du + pdV̂ + V̂ dp (12.6)

12.2.3 Specific heats

The internal energy for an ideal gas is a function of the temperature, and we define the
specific heat cv (T ) by
du = cv (T )dT (12.7)
In many cases cv (T ) will be a constant, and in that case a change in the specific internal
energy is given by
∆u = cv ∆T (12.8)
For an ideal gas we have pV̂ = RT , where R is the universal gas constant, and it
follows that
h = u + RT, ideal gas (12.9)
This implies that also the specific enthalpy will be a function of temperature for an ideal
gas. We define the specific heat cp (T ) by
dh = cp (T )dT (12.10)
From (12.9) is follows that
dh = du + RdT, ideal gas (12.11)
and, consequently, for any gas that satisfy the ideal gas law we have the relation
cp (T ) = cv (T ) + R, ideal gas (12.12)
If cv (T ) is a constant, then a change in enthalpy is given by
∆h = cp ∆T (12.13)
We define κ to be the ratio between cp and cv :
cp
κ := (12.14)
cv
The numerical value for air is κ = 1.4.
12.2. ENERGY, ENTHALPY AND ENTROPY 467

Example 195 For an ideal gas we have the following useful expressions:
cp
R cp − cv cv −1
= = = κ − 1, ideal gas (12.15)
cv cv 1
and cp
R cp − cv cv −1 κ−1
= = cp = , ideal gas (12.16)
cp cp cv
κ

12.2.4 Entropy
The concept of entropy provides us with a number of useful modeling tools for thermody-
namic systems. The reason for this is that in engines, turbines and compressors, entropy
can be seen as a book-keeping tool for certain energy-loss phenomena. In particular, if a
component is designed so that no entropy is being generated in connection with gas flow,
then this means that there is no energy loss due to viscosity or heat conduction. This
will be highly desirable for most system components as this will typically imply that the
thermal efficiency is high. If a process is designed so that no entropy is being generated,
then the process is said to be isentropic. The design of systems that are isentropic or
close to isentropic will therefore be the goal of the mechanical design. Then, as a suc-
cessful design will be isentropic or close to isentropic, it can often be assumed in the
development of the dynamic modeling that the gas flow is isentropic. This is very useful
in the development of the balance equations. It is interesting to note that in this context,
the important question for the control engineer is whether entropy is being generated or
not, while the absolute value of the entropy is not an issue. Therefore, we will describe
entropy in terms of its differential, and we will not be concerned about its absolute value.
The entropy of a mass element dm is called the specific entropy and is denoted s.

The specific entropy is defined in terms of its differential ds by

T ds = du + pdV̂ (12.17)

Combination of this definition and (12.6) gives the following alternative expression:

T ds = dh − V̂ dp (12.18)

12.2.5 The entropy equation

The entropy equation which will be derived in this section is quite interesting, as it
gives a mathematical formulation of the second law of thermodynamics. This material
is included to make it easier for the reader to understand the physical interpretation of
entropy in the setting of the present chapter. The entropy equation is derived using the
mass balance and the equation for internal energy. In addition, the derivation relies on
the fact that certain forms of energy transfer can only run in one direction. In particular,
heat flow is always in the direction of decreasing temperature, and internal work due to
viscosity in combination with velocity gradients will always lead to a reduction of kinetic
energy and an increase of internal energy.
From the definition (12.17) of entropy we have

Ds Du DV̂
T = +p (12.19)
Dt Dt Dt
468 CHAPTER 12. GAS DYNAMICS

The continuity equation (11.5) gives

µ ¶
DV̂ D 1 1 Dρ ³ ´
ρ =ρ =− = ∇·v (12.20)
Dt Dt ρ ρ Dt
Insertion of this equation and (11.337) into (12.19) gives the following form of the entropy
equation:
Ds 1 1 ³ ´
ρ =− ∇ · j̇Q + τ : ∇v (12.21)
Dt}
| {z |T {z } |T {z }
rate of change change in entropy increase in entropy
in entropy in due to heat from irreversible
material volume conduction viscous dissipation
element
This can be further developed into a formulation which offers a more detailed inter-
pretation by applying the product rule for differentiation to the first term on the right
side.

This entropy equation is given by

à !
Ds jQ 1 1 ³ ´
ρ =− ∇· − j · ∇T
2 Q
+ τ : ∇v
Dt
| {z } T T
| {z } |T {z }
| {z }
rate of change increase in entropy increase in entropy
change in entropy
in entropy in due to heat from irreversible
due to heat flow
material volume conduction within viscous dissipation
element volume element within volume element
(12.22)
The physical interpretation offered in the equation is supported by the integral form
ZZZ ZZ Ã ! ZZZ µ ¶
D jQ 1 1 ³ ´
ρsdV = − · ndA + − 2 jQ · ∇T + τ : ∇v dV
Dt V ∂V T V T T
(12.23)

An illustration of the different phenomena is given in Figure 12.1. The heat generation
τ : (∇v) due to viscous dissipation of mechanical energy will always be positive. Moreover,
it has been established that the heat flow will always have a negative component along
the temperature gradient. These two results are summed up as
³ ´
τ : ∇v ≥0, jQ · ∇T ≤ 0 (12.24)

Using the inequalities of (12.24) we then get the Clausius-Duhem inequality (Lin and
Segel 1974) which we state both in differential and integral form:
à !
Ds jQ
ρ ≥ −∇ · (12.25)
Dt T
ZZZ ZZ Ã !
D jQ
ρsdV ≥ − · ndA (12.26)
Dt V ∂V T
The Clausius-Duhem inequality is often referred to as the second law of thermodynamics.
Note, however, that a more precise formulation of the second law of thermodynamics is
either of the equations (12.22) or (12.23) together with (12.24).
12.2. ENERGY, ENTHALPY AND ENTROPY 469

j Q V, ,s
T

j Q
1
 : 
v  
T
T T2

Figure 12.1: The rate of change of entropy in the material control volume V depends on
the entropy flow jQ /T into the material volume due to heat flow over the boundary, and
on the internal entropy production terms −jTQ ∇T /T 2 and τ : ∇vT /T , where the first
term is related to temperature gradients and heat flow, and the second term is due to
velocity gradients in combination with viscosity.

V, ,s

 T
j Q
 v
: 
T T2

Figure 12.2: Isolated tank of volume V . The internal entropy production terms are
indicated.

Example 196 Consider a constant control volume V (Figure 12.2) filled with gas and
with no exchange of gas with the outside. The walls of the volume are assumed to be
isolated so that there is no heat conduction over the boundary ∂V . Then (12.26) give
ZZZ
D
ρsdV ≥ 0 (12.27)
Dt V

which shows that the total entropy is constant or increasing. This result can be made
more precise by using equation (12.23) which gives
ZZZ ZZZ ZZZ
D 1 1 ³ ´
ρsdV = − 2
j Q · ∇T dV + τ : ∇v dV (12.28)
Dt V V T V T

We see from this equation and (12.24) that the entropy of the gas in an isolated tank
with no gas exchange with the outside cannot decrease, and that the entropy will increase
if there are either temperature gradients ∇T , or velocity gradients ∇v.
470 CHAPTER 12. GAS DYNAMICS

wi wo

V
, V
m, T
jQ

Figure 12.3: Cylinder with variable volume V due to a moving piston. The inlet mass
flow in wi , the outlet massflow is wo , and the heat flow into the volume is jQ .

Example 197 In Fourier’s law of heat conduction

jQ = −k∇T (12.29)

the heat flow is proportional to the negative temperature gradient. This gives
³ ´2
jQ · ∇T = −k ∇T (12.30)

which clearly is negative.

12.2.6 Internal energy equation in terms of temperature

We consider a cylinder of variable volume V = Ax filled with an ideal gas. The tem-
perature T and the pressure p are assumed to be constant over the volume. Gas with
temperature Ti and pressure pi is flowing into the volume with mass flow wi = ρi qi ,
where qi is the volumetric flow and ρi is the density. Gas is flowing out of the volume
with mass flow wo = ρqo where qo is the volumetric flow out of the volume. The mass
balance is
d
m = wi − wo (12.31)
dt
We assume that the kinetic and potential energy is zero so that E = U = mu. Then the
energy balance is
d
(mu) = hi wi − hwo − pAẋ + jQ (12.32)
dt
where jQ is the heat flow. We assume that cv and cp are constants, so that u = cv T and
h = cp T . This gives

dm
cv T + mcv Ṫ = cp Ti wi − cp T wo − pAẋ + jQ (12.33)
dt
12.2. ENERGY, ENTHALPY AND ENTROPY 471

Insertion of the mass balance (12.31) gives

cv T (wi − wo ) + mcv Ṫ = cp Ti wi − cp T wo − pAẋ + jQ (12.34)

which leads to

mcv Ṫ = cp wi Ti − [cv wi + (cp − cv ) wo ] T − pAẋ + jQ (12.35)

Finally, it is noted that for an ideal gas, pV = mRT , and therefore pA = mRT /x.
Insertion of this and the use of (12.15) leads to the energy balance in the form

κwi wi + (κ − 1) wo ẋ 1
Ṫ = Ti − T − (κ − 1) T + jQ (12.36)
m m x mcv

12.2.7 Energy balance in terms of pressure

For an ideal gas with constant cv the internal energy can be written
cv 1
U = mcv T = pV = pV. (12.37)
R κ−1
The energy balance is assumed to be

dU
= hi wi − hwo − pAẋ + jQ (12.38)
dt
and we get
1 ³ ´
ṗV + pV̇ = hi wi − hwo − pAẋ + jQ (12.39)
κ−1
which is simplified to

ṗV = (κ − 1) (hi wi − hwo + jQ ) − κpAẋ. (12.40)

The specific enthalpy can be written

cp κ
h = cp T = pV̂ = pV̂ (12.41)
R κ−1
which gives ³ ´
ṗV = κ pi V̂i wi − pV̂ wo + (κ − 1) jQ − κpAẋ (12.42)

We introduce the input volumetric flow qi = V̂i wi and the output volumetric flow qo =
V̂ wo and get the expression

qo qi ẋ jQ
ṗ = −κ p + κ pi − κ p + (κ − 1) . (12.43)
V V x V
Example 198 An alternative way of writing this is

V pi (κ − 1)
ṗ = −qo + qi − Aẋ + jQ (12.44)
κp p κp

which is reminiscent of the mass balance for a hydraulic control volume.

472 CHAPTER 12. GAS DYNAMICS

12.2.8 Piston motion

The equation of motion for the piston is

mp ẍp = pA − FL (12.45)

where FL is the load force. Suppose that the piston is connected to a crankshaft with
a rod, and that the kinematic relation between the piston velocity ẋ and the crankshaft
velocity ω cs is given by
ẋ = r(θcs )ω cs (12.46)
where θcs is the crankshaft angle. Then the equation of motion for the crankshaft is

J ω̇ cs = F r(θ) − TL (12.47)

where TL is the load torque. The crankshaft is a mechanical two-port where one port
has input F and output ω cs , and one port with input TL and output ω cs .
Flexibility in the crankshaft can easily be included. This is done by inserting an
elastic transmission and an inertia which is modelled as a mechanical two-port given by

J1 ω̇ 1 = TL − T1 (12.48)
TL = D1 (ω m − ω 1 ) + K1 (θm − θ1 ) . (12.49)

where ω 1 is the shaft speed of the load shaft. This means that the transmission is
modelled as a torsional spring with spring constant K1 in parallel with a torsional damper
with damping coefficient D1 . The input port has effort TL and flow ω m , while the output
port has effort T1 and flow ω 1 .

12.3 Isentropic conditions

12.3.1 Isentropic processes
Isentropic processes are important in a wide range of applications. Isentropic processes
are processes where there is no entropy production in the sense that ds = 0. In view
of (12.21) isentropic processes can occur if there is no heat conduction and no internal
viscous work. From (12.17) and (12.18) we see that ds = 0 implies

du = −pdV̂ (12.50)
dh = V̂ dp (12.51)

For gases that satisfy the ideal gas law pV̂ = RT , this can be written
RT
cv (T )dT = − dV̂ (12.52)
V̂
RT
cp (T )dT = dp (12.53)
p
which leads to
dT R dV̂
= − (12.54)
T cv (T ) V̂
dT R dp
= (12.55)
T cp (T ) p
12.3. ISENTROPIC CONDITIONS 473

In view of (12.15) and (12.16) this gives the differential isentropic relations

dT dV̂
= − [κ(T ) − 1] (12.56)
T V̂
dT κ(T ) − 1 dp
= (12.57)
T κ(T ) p

and, by combination of the two equations, the differential isentropic relation

dp dV̂
= −κ(T ) (12.58)
p V̂

Consider the differential equation

dx dy ¡ ¢a
= a ⇒ ln x = a ln y + C1 ⇒ eln x = C eln y (12.59)
x y
⇒ x = Cy a (12.60)

Thus, if x1 = Cy1a is one solution, then the constant c is found from C = x1 /y1a , and any
solution x2 = Cy2a must satisfy
µ ¶a
x2 y2
= (12.61)
x1 y1
A reasonable assumption is that κ is a constant. Then we see from this derivation
that (12.56, 12.57, 12.58) leads to the isentropic relations
à !κ−1
T1 V̂2
= (12.62)
T2 V̂1
µ ¶ κ−1
T1 p1 κ
= (12.63)
T2 p2
à !κ
p1 V̂2
= (12.64)
p2 V̂1

Example 199 The identity ρV̂ = 1 implies

dρ dV̂
=− (12.65)
ρ V̂

For isentropic processes (12.58) gives

p
dp = κ(T ) dρ (12.66)
ρ

For an ideal gas this leads to the result

dp = c2 dρ (12.67)
√
where c = κRT , which we will see is the speed of sound.
474 CHAPTER 12. GAS DYNAMICS

Example 200 The mass balance of a gas flowing through a fixed volume V with mass
flow win into the volume and mass flow wout out of the volume is
V ρ̇ = win − wout (12.68)
where ρ is the density of the gas. If the gas is ideal and the conditions are isentropic,
then (12.67) applies, and the mass balance becomes
V
ṗ = win − wout (12.69)
c2

12.3.2 Stagnation state

In the modeling of flow the concept of a stagnation state is useful. The stagnation state
is characterized by a stagnation velocity v0 = 0, the specific stagnation enthalpy h0 , the
stagnation temperature T0 , and the stagnation pressure p0 .
The specific stagnation enthalpy
v2
h0 = h + (12.70)
2
is defined as the sum of the specific enthalpy h and the specific kinetic energy v 2 /2. The
stagnation temperature T0 is the temperature corresponding to the specific stagnation
enthalpy h0 in the sense that
h0 − h
T0 − T = (12.71)
cp
when cp is constant. The stagnation pressure p0 is defined by
µ ¶ κ−1 κ
p0 T0
= (12.72)
p T
From the definitions (12.71) and (12.72) it it seen that the stagnation temperature and
the stagnation pressure can be found from
µ ¶ κ−1
κ
v2 v2
T0 = T + , p0 = p 1 + (12.73)
2cp 2cp T
If a fluid with enthalpy h, temperature T , pressure p and velocity v is slowed down
to zero velocity in an isentropic process, then the gas will be in the stagnation state
with stagnation velocity v0 = 0, stagnation enthalpy h0 , stagnation temperature T0 ,
and stagnation pressure p0 . To make a clear distinction between the pressure p and the
stagnation pressure p0 it is customary to refer to p as the static pressure.

12.3.3 Energy balance for isentropic processes

In this section we investigate which condition that must be fulfilled for the energy balance
to describe an isentropic process. We start with the relation
p
dT = − dV̂ (12.74)
cv
which follows from T ds = cv dT + pdV̂ = 0. Division with dt gives
µ ¶
p d V p pV
Ṫ = − =− V̇ + 2 ṁ (12.75)
cv dt m mcv m cv
12.3. ISENTROPIC CONDITIONS 475

Then the ideal gas law pV = mRT gives

à !
ṁ V̇
Ṫ = (κ − 1) − T (12.76)
m V

The energy balance (12.36) satisfies this equation if the heat flux jQ is zero, and the tem-
perature of the mass flow into the volume has the same temperature as the temperature
in the volume.

12.3.4 The speed of sound

The derivation of the speed of sound is based on the following mass and momentum
balances for a differential control volume:
∂ρ
= −∇ · (ρv) (12.77)
∂t
Dv
ρ = −∇p (12.78)
Dt
The change of variable in the first equation from ρ to p requires an expression for dp as
a function of dρ. This is obtained by assuming isentropic conditions. For an ideal gas
the ideal gas law p = ρRT under isentropic conditions leads to

dp = κRT dρ (12.79)

This gives
∂p
= −κRT ∇ · (ρv) (12.80)
∂t
³ ´
Linearization around ρ0 , T0 , v0 = 0 gives

∂p
= −c2 ρ0 ∇ · v (12.81)
∂t
∂v
ρ0 = −∇p (12.82)
∂t
where c2 = κRT0 . Then, time differentiation of the linearized pressure equation gives

∂2p ∂ ³ ´
= −c2 ρ0 ∇·v
∂t2 ∂t
∂v
= −c2 ∇ · ρ0
∂t
Insertion of the linearized velocity equation gives the wave equation

∂2p 2
2
= c2 ∇ p (12.83)
∂t
√
where c = κRT0 .
d’Alembert’s solution of the wave equation (12.83) (Kreyszig 1979) is in the form

p(x, t) = f (x + ct) + g (x − ct) (12.84)

476 CHAPTER 12. GAS DYNAMICS

where f (x + ct) + g (x − ct) are required to satisfy initial conditions and boundary con-
ditions. To verify that this is a solution to the wave equation we calculate

∂2
[f (x + ct) + g (x − ct)] = c2 [f 00 (x + ct) + g 00 (x − ct)] (12.85)
∂t2
and
∂2
[f (x + ct) + g (x − ct)] = f 00 (x + ct) + g 00 (x − ct) (12.86)
∂x2
The function f (x + ct) can be interpreted as a wave that goes in the negative x direction
with velocity −c, while g(x − ct) can be interpreted as a wave moving in the positive x
direction
√ with velocity c. From this we can conclude that the speed of sound is equal
to c = κRT . For air at 15 degrees Centigrade this gives 340 m/s, which agrees with
experimental data.

Example 201 In the works of Sir Isaac Newton the speed of sound was derived under
conditions that correspond to isothermal conditions. For ideal gases this gives dp = RT dρ
and the wave equation becomes
∂2p 2
2
= c̄2 ∇ p (12.87)
∂t
√
The speed of sound according to this result is c̄ = RT , which for air at 15 degrees
Centigrade gives 287 m/s. This was is not in agreement with what was observed in
experiments.

12.3.5 Helmholtz resonator

L, A
V
w
p  RT

Figure 12.4: Helmholtz resonator

A Helmholtz resonator is a pipe of length L and cross section A which is open at one
end, and which is connected to a volume V , called the plenum, at the other end. This
is shown in Figure 12.4. It is assumed that the compressibility effects in the pipe are
negligible, which means that it is assumed that the pressure p and the mass flow w do
not vary along the pipe. Further, it is assumed that the velocity in the plenum can be
approximated by zero. This means
√ that the conditions in the plenum are isentropic. This
implies dp = c2 dρ where c = κRT is the sonic speed. The dynamics of the Helmholtz
12.4. ACOUSTIC RESONANCES IN PIPES 477

resonator is derived from the mass balance of the plenum and the momentum balance of
the pipe. The model is found to be
V dp
= w (12.88)
c2 dt
dw
L = −pA (12.89)
dt
The model can be further developed by differentiating the mass balance with respect
to time, and by inserting the momentum balance. This gives
d2 p
+ ω 2H p = 0 (12.90)
dt2
Here ω H is the Helmholtz frequency defined by
r
A
ωH = c (12.91)
VL
Example 202 Consider a Helmholtz resonator with a pressurizing device on the input
that supplies air at a mass flow w with a pressure p1 = p1 (w). In addition, air is flowing
out of the plenum with mass flow w2 = w2 (p). The model is then
dp c2
= [w − w2 (p)] (12.92)
dt V
dw A
= [p1 (w) − p] (12.93)
dt L
Linearization gives
d∆p c2 c2 dw2
= ∆w − ∆p (12.94)
dt V V dp
d∆w A A dp1
= − ∆p + ∆w (12.95)
dt L L dw
For simplicity it is assumed that dw2 /dp ≈ 0. Then the characteristic equation is
A dp1
λ2 − λ + ω 2H = 0 (12.96)
L dw
It is clear that the system will be stable if dp1 /dw ≤ 0. Therefore, if p1 (w) has a positive
slope, then the system becomes unstable (Figure 12.5).

12.4 Acoustic resonances in pipes

12.4.1 Dynamic model
Acoustic waves for an ideal gas are described by the wave equation
∂2p
= c2 ∇2 p (12.97)
∂t2
√
where c = κRT under the assumption of isentropic conditions. This is derived as for
fluids. The wave variables defined in Section 4.5.7 will satisfy the transfer functions
equations
a2 (s) = exp(−T s)a1 (s), b1 (s) = exp(−T s)b2 (s) (12.98)
478 CHAPTER 12. GAS DYNAMICS

p1

w
Unstable Stable
region region

Figure 12.5: Pressure p1 as a function of massflow w. The unstable and stable regions
are indicated.

and the boundary conditions

a1 (s) = g1 (s)b1 (s), b2 (s) = g2 (s)a2 (s) (12.99)

Combining these equations we get

[1 − exp (−2T s) g1 (s)g2 (s)] a1 (s) = 0 (12.100)

12.4.2 Pipe closed at both ends

We consider acoustic waves in a pipe of length , and describe the dynamics of the gas
by the pressure p and the volumetric flow q through the pipe. The wave variables a and
b are defined as for fluid flow through pipes. The boundary conditions are set to

q(0, s) = q( , s) = 0 (12.101)

which means that the pipe is closed at both ends. The impedances at the two ends are
defined by p(0, s) = z1 (s)q(0, s) and p( , s) = z2 (s)q( , s). The boundary conditions then
correspond to impedances z1 = ∞ and z2 = ∞. This leads to g1 = g2 = 1 and b1 = a1 ,
a2 = b2 , which gives
a1 (s) = exp (−2T s) a1 (s) (12.102)
or
[1 − exp (−2T s)] a1 (s) = 0 (12.103)
It is straightforward to check that the solution is the same if the pipe is open at both
ends. This system has singularities whenever

exp (−2T s) = 1 (12.104)

which means that the system has infinitely many singularities in

kπ
s=j , k = 0, ±1, ±2, . . . (12.105)
T
12.4. ACOUSTIC RESONANCES IN PIPES 479

For each singularity there is an undamped resonance. The first resonance is at

π
s=j (12.106)
T
which corresponds to a frequency of
1 π 1 c
f= = = (12.107)
2π T 2T 2
For air at 15 degrees Centigrade (288 K) the speed of sound is c = 340 m/s. A pipe of
length 10 m then gives a resonance frequency at f = 340/20 = 17 Hz. To get a frequency
of 440 Hz the pipe must be = 340/(2 · 440) = 0.386 m. To get 256 Hz the pipe must be
0.66 m.

12.4.3 Pipe closed at one end

The boundary condition is changed to
p(0, s) = 0 (12.108)
q( , s) = 0 (12.109)
which gives g1 = −1 and g2 = 1. Then the system equation becomes
[1 + exp (−2T s)] a(s) = 0 (12.110)
and the singularities are found at
π + 2kπ
s=j , k = 0, ±1, ±2, . . . (12.111)
2T
The first resonance is now at
1 c
f= = (12.112)
4T 4
To get 440 Hz at 288 K the pipe must then be 0.193 m, while 55 Hz results from a pipe
of length 3.145 m.
Consider a pipe of length L and cross section A which is open at the inlet end, denoted
by subscript 1, and which ends in a volume V at the outlet end denoted by subscript
2. It is assumed that the pressure is constant over the volume V, and that this pressure
is equal to the pressure p2 = p( ) at the outlet of the pipe. In this case the boundary
conditions are
p(0, s) = 0 (12.113)
p( , s) = z2 (s)q( , s) (12.114)
The impedance z2 (s) at the outlet is found from the mass balance of the volume V. The
mass balance is
d (ρ2 V )
= ρ2 q2 (12.115)
dt
which gives
dp2 κp2
V = ρq2 (12.116)
dt ρ2
which is simplified to
dp2 κp2
= q2 (12.117)
dt V
and it is seen that for small q2 the impedance is
κp2 1
z2 (s) = (12.118)
V s
480 CHAPTER 12. GAS DYNAMICS

p 0 ,T 0 v
p, T,A Outlet

v0  0
Throat

Figure 12.6: Isentropic gas flow from stagnation state p0 , T0 and velocity v0 = 0 through
a restriction with throat cross section A. The throat states are denoted p and T, and the
velocity at the throat is denoted v.

12.4.4 Pressure measurement in diesel engine cylinder

Due to the high temperatures in a diesel cylinder it is desirable to measure the pressure
in the cylinder by placing the pressure sensor in a bored pipe of small diameter. Suppose
that the length of the bore is 3 cm, and that the pressure sensor is in the closed end of
the pipe, while the open end is on the cylinder face. Then the resonance frequency of
the pipe is
r
340 1000
f= = 5.28 kHz (12.119)
4 × 0.03 288
In a motor running at 6000 rpm, which corresponds to a shaft speed of 36000 degrees
per second, the shaft will rotate 36000/5280 = 6.8 degrees in one resonance period. This
means that accurate pressure measurements can not have a resolution better than about
6.8 degrees of crank-shaft angle under these conditions.

12.5 Gas flow

12.5.1 Gas flow through a restriction
Gas flow through a valve or a restriction is usually modelled as nozzle flow , which is
isentropic flow that is computed using the isentropic relations between the stagnation
state in the inlet volume and the state in the throat of the restriction. It is convenient
to use the Mach number
v
M= (12.120)
c
in the derivation of the mass flow. √
This is a dimensionless velocity, where the velocity v
is scaled by the sonic velocity c = κRT .
The starting point for the derivation is the relation between the mass flow w and
Mach number M = v/c which is found from the following calculation:
√ Ap
w = ρAv = ρAM c = ρAM κRT = √ Mκ (12.121)
κRT
Here the ideal gas law p = ρRT is used, and A is the cross section area of the nozzle.
12.5. GAS FLOW 481

This can then be written √

w κRT
= κM (12.122)
Ap
where the left hand side is a dimensionless form of the mass flow. To be able to reference
the mass flow to the stagnation state 0, we use the expression
√ √ r r
w κRT0 w κRT T0 p T0 p
= = κM (12.123)
Ap0 Ap T p0 T p0

where p0 , T0 are stagnation states at the inlet. The isentropic relation

µ ¶(κ−1)/κ
T0 p0
= (12.124)
T p

can then be used to find the following expression for isentropic flow of an ideal gas:
√ µ ¶(κ+1)/2(κ−1) µ ¶(κ+1)/2κ
w κRT0 T p
= κM = κM (12.125)
Ap0 T0 p0

Next, the steady-state energy equation for nozzle flow is written

v2 M 2 c2 M 2 κRT
cp T0 = cp T + = cp T + = cp T + (12.126)
2 2 2
where T0 is the stagnation temperature at the inlet, T is the temperature at the throat,
while v is the velocity at the throat. Using R/cp = (κ − 1)/κ, the energy equation can
be written in the form
T0 κ−1 2
=1+ M (12.127)
T 2
or, equivalently,

· µ ¶¸1/2 ( "µ ¶
(κ−1)/κ
#)1/2
2 T0 2 p0
M= −1 = −1 (12.128)
κ−1 T κ−1 p

Insertion of this expression for M in equation (12.125) for mass flow then leads to

√ µ ¶(κ+1)/2κ ( "µ ¶
(κ−1)/κ
#)1/2
w κRT0 p 2 p0
=κ −1
Ap0 p0 κ−1 p

By straightforward operations this can be rearranged into the well-known expressions for
isentropic nozzle flow of an ideal gas.

Isentropic flow through a restriction will have mass flow given by

√ µ ¶1/κ ( " µ ¶(κ−1)/κ #)1/2
w κRT0 p 2 p
=κ 1− (12.129)
Ap0 p0 κ−1 p0
482 CHAPTER 12. GAS DYNAMICS

We note that this expressions are valid for M ≤ 1. Sonic flow occurs when M = 1.
It is seen from (12.127) that if κ = 1.4 this corresponds to
¯ µ ¶ κ−1
κ
p ¯¯ 2
= = 0.52828 (12.130)
p0 ¯sonic κ+1

The resulting mass flow at sonic conditions is seen from (12.125) to be

√ ¯ µ ¶ 2(κ−1)
κ+1
w κRT0 ¯¯ 2
=κ (12.131)
Ap0 ¯sonic κ+1

12.5.2 Example: Discharge of gas from tank

Insulation

Nozzle
V
A
T, p,

Figure 12.7: Insulated tank with gas discharge

The example of this section is adopted from (Bird et al. 1960, p. 480). A fluid
is contained in an insulated tank of volume V as shown in Figure 12.7. The gas is
discharged with mass flow w through a small nozzle with opening area A. The nozzle
flow is assumed to be isentropic from the stagnation state T, p, ρ in the tank to the throat
state Ta , pa , ρa at the outlet of the nozzle. The throat state is set equal to the ambient
state of the surrounding air, and the throat area is set to be the area A of the nozzle
(Heywood 1988).
The mass balance is
dm
= −w (12.132)
dt
The energy of the gas is assumed to be equal to the internal energy U = ρuV where u is
the specific internal energy. The energy balance is then

d p
(mu) = −wu − w = −wh (12.133)
dt ρ

where h = u + p/ρ is the specific enthalpy. Using the product rule for differentiation and
mass balance we get
pw
u̇ = − (12.134)
ρm
Finally, we assume that u = cv T and that the ideal gas law p = ρRT is valid. Then the
model can be written

ṁ = −w (12.135)
w
Ṫ = − (κ − 1) T (12.136)
m
12.5. GAS FLOW 483

where κ > 1 and the mass flow is found from the nozzle flow equation to be
µ ¶1/κ ( " µ ¶(κ−1)/κ #) 12
Ap pa 2 pa
w= √ κ 1− (12.137)
κRT p κ−1 p

and the throat pressure pa is the ambient pressure, and the stagnation pressure p is found
from the ideal gas law to be
mRT
p= . (12.138)
V
From the model it is clear that the time derivatives of mass m and temperature T
will be less than or equal to zero.

12.5.3 The Euler equation around sonic speed

We write the continuity equation (11.6) in the form
∂ρ ³ ´ ³ ´
= − ∇ρ · v − ρ ∇ · v (12.139)
∂t
and the Euler equation in the form (11.58) with zero mass force:
· ³ ´¸
∂v
ρ + v · ∇v = −∇p (12.140)
∂t
We consider an ideal gas under isentropic conditions, which according to (12.67) implies
that
∇p = c2 ∇ρ (12.141)
The Euler equation then becomes
∂v ³ ´
ρ = −ρv · ∇v − c2 ∇ρ (12.142)
∂t
We take the scalar product between the Euler equation and the velocity and get the
following set of equations:
∂ρ ³ ´ ³ ´
= − ∇ρ · v − ρ ∇ · v (12.143)
∂t
∂v ³ ´ ³ ´
ρ · v = −ρv ∇v · v − c2 ∇ρ · v (12.144)
∂t
We consider a point, and align our coordinate frame so that the velocity v is along the
x1 axis, that is,
v = v1 a1 (12.145)
Then, at this point our set of equations become
∂ρ
= −ρ,1 v1 − ρv1,1 −ρv2,2 − ρv3,3 (12.146)
µ ¶∂t
∂v1
ρ v1 = −ρv12 v1,1 − c2 ρ,1 v1 (12.147)
∂t
By adding the equations we get
µ ¶
1 ∂ρ 1 ∂v1 ¡ ¢
− 2 v1 = − 1 − M 2 v1,1 − (v2,2 + v3,3 ) (12.148)
ρ ∂t c ∂t
484 CHAPTER 12. GAS DYNAMICS

Under stationary conditions we have

¡ ¢
v2,2 + v3,3 = − 1 − M 2 v1,1 (12.149)

This result has interesting implications if we consider speeds close to the sonic speed.
Then the Mach number is close to zero, and in particular, at sonic speed M = 1, and
1 − M 2 = 0. This gives
v2,2 + v3,3 = 0 · v1,1 (12.150)
This is satisfied in a flow field where v2 = v3 = 0, which is the case if all particles flow
along the x1 axis. However, if M → 1 when v2,2 + v3,3 6= 0, then v1,1 → ∞. This
means that the model consisting of the Euler equation and the continuity equation has
a singularity at M = 1. This agrees with experimental evidence which shows that if the
Mach number M approaches unity when there are disturbances in the flow patterns so
that v2,2 +v3,3 6= 0, then a large gradients in density and velocity occur, and this is called
a shock . In a shock the assumption of isentropic conditions are no longer valid as there
will be significant viscous energy dissipation, and (12.67) will no longer be valid.
Chapter 13

Compressor dynamics

13.1 Introduction
13.1.1 Compressors
Compressors are used in a wide variety of applications. These includes turbojet engines
used in aerospace propulsion, power generation using industrial gas turbines, turbocharg-
ing of internal combustion engines, pressurization of gas and fluids in the process indus-
try, transport of fluids in pipelines and so on. Compressors can be divided into four
general types: reciprocating, rotary, centrifugal and axial. Some authors use the term
radial compressor when referring to a centrifugal compressor. Reciprocating and rotary
compressors work by the principle of reducing the volume of the gas, and will not be
considered further in this work. Centrifugal and axial compressors, also known as tur-
bocompressors or continuous flow compressors, work by the principle of accelerating the
fluid to a high velocity and then converting this kinetic energy into potential energy by
decelerating the gas in diverging channels. In axial compressors the deceleration takes
place in the stator blade passages, and in centrifugal compressors it takes place in the
diffuser. The increase in potential energy of the fluid is manifested by a pressure rise.
This conversion can be explained from Bernoulli’s equation (11.86), which is written:

p1 C2 p2 C22
+ 1 + gz1 = + + gz2 (13.1)
ρ 2 ρ 2

Here p is the pressure, ρ is the density of the fluid, C is the velocity of the fluid and gz
the potential energy per unit mass. Subscripts 1 and 2 denotes properties before and
after deceleration, respectively. The equation is a special case of the law of conservation
of energy developed for flowing fluids. Bernoulli’s equation states that the sum of kinetic
2
energy C2 , potential energy gz and pressure head pρ at one set of conditions is equal
to their sum at another set of conditions. Hence, decrease in kinetic energy implies an
increase in potential energy and pressure. One obvious difference between the two types
of turbocompressors is that, in axial compressors, the flow leaves the compressor in the
axial direction, whereas in centrifugal compressors, the flows leaves the compressor in a
direction perpendicular to the axis of the rotating shaft. Axial compressors can accept
higher mass flow rates than centrifugal compressors for a given frontal area. This is
one reason for axial compressors dominance in jet engines, where frontal area is of great

485
486 CHAPTER 13. COMPRESSOR DYNAMICS

importance. Another reason for this is that for gas turbines, or jet engines, specific
fuel consumption decreases with increasing pressure ratio. The axial compressor, using
multiple stages, can achieve higher pressure ratio and efficiency than the centrifugal.

13.1.2 Surge and rotating stall

The useful range of operation of turbocompressors, both axial and centrifugal, is limited
by choking at high mass flows when sonic velocity is reached in some component. The
mass flow at which choking occurs depends on the rotational speed of the compressor,
and the compressor is said to have reached the stone wall area in its operating domain
when choking occurs. The mass flow at sonic velocity, that is M = v/c = 1, can be
calculated by using (12.131).
At low mass flows the operation is limited by the onset of two instabilities known as
surge and rotating stall. Surge is an axisymmtrical oscillation of the flow through the
compressor, and is characterized by a limit cycle in the compressor characteristic. Surge
oscillations are in most applications unwanted, and can in extreme cases even damage
the compressor. Rotating stall is an instability where the circumferential flow pattern is
disturbed. This is manifested through one or more stall cells of reduced, or stalled, flow
that propagate around the compressor annulus at a fraction of the rotor speed. There is
two different methods of dealing with the surge/stall-problem:

1. Traditionally, surge and rotating stall have been avoided by using control systems
that prevent the operating point of the compressions system to enter the unstable
regime to the left of the surge line, that is the stability boundary.

2. A fundamentally different approach, known as active surge/stall control, is to use

feedback to stabilize this unstable regime. This approach requires a model of the
compression system in order to design a stabilizing controller. Active control can
allow for both operation in the peak efficiency and pressure rise regions located in
the neighborhood of the surge line, as well as an extension of the operating range
of the compressor.

13.2 Centrifugal Compressors

13.2.1 Introduction
The centrifugal compressor consists essentially of a stationary casing containing a rotating
impeller which imparts a high velocity to the fluid, and a number of fixed diverging
passages in which the air is decelerated with a consequent rise in static pressure. The
latter process is one of diffusion, and consequently, the part of the compressor containing
the diverging passages is known as the diffuser. The impeller is mounted on a shaft which
is either a direct extension of the drive shaft or a separate shaft supported by bearings
and driven through a coupling. The shaft and impeller assembly, called the rotor, are
seated in the casing.
The fluid flows into the inducer (also referred to as the impeller eye) and flow through
the blade passage of the rotating impeller. Because of the rotation the tangential velocity
of the fluid increases when the fluid flows outwards in the impeller, and the associated
increase in the centrifugal force makes the static pressure increase through the impeller.
A further pressure increase is obtained in the diffuser, where the pressure increase is due
to the reduction of the velocity. It is common practice to design a compressor so that
13.2. CENTRIFUGAL COMPRESSORS 487

about half the pressure rise occurs in the impeller and half in the diffuser. As no work
is done on the fluid in the diffuser, all the energy is supplied to the fluid in the impeller,
thus the energy transfer from the shaft power to the fluid energy will be determined by
the conditions at the inlet and outlet of the impeller.
It is more difficult to obtain efficient deceleration of flow than it is to obtain efficient
acceleration. If divergence in the diffuser is too rapid, the fluid will tend to break away
from the walls of the diverging passage, reverse its direction and flow back in the direction
of the pressure gradient. This may lead to the formation of eddies with consequent
transfer of some kinetic energy into internal energy and a reduction of useful pressure
rise. On the other hand, a small angle of divergence will lead to a long diffuser and high
losses due to friction. In order to carry out the diffusion in as short a length as possible,
the air leaving the impeller may be divided into a number of separate diverging passages
separated by fixed diffuser vanes, resulting in a vaned diffuser. However, in industrial
applications where size may be of secondary importance a vaneless diffuser may have the
economic advantage as it is much cheaper to manufacture than the vaned diffuser. A
vaneless diffuser is a simple annular channel, and is therefore also known as an annular
diffuser, in which the radial velocity component is reduced by area increase and the
tangential velocity component by the requirement of constant fluid angular momentum.
If the disadvantage of the annular diffuser is its bulk, the advantage is its wide range of
operation. A vaned diffuser may have a higher peak efficiency than an annular diffuser,
but its mass flow range is considerable less because of early stall of the diffuser vanes.
We consider a compressor driven by a drive torque τ d . The compressor is modeled as
a momentum source in a duct of length L and cross section A. The duct is connected to
a plenum, which is a volume V . The gas flows from the duct into the plenum, and then
out of the plenum through a throttle. This is illustrated in Figure 13.4. The momentum
delivered to the gas is the ideal momentum from the rotor minus momentum loss due to
incidence loss and fluid friction loss. The usual Greitzer surge model (Greitzer 1976) for
this system is derived from the mass balance of the plenum and the momentum equation
of the duct.

13.2.2 Shaft dynamics

We assume that the compressor is driven by an electrical motor as illustrated in Figure
13.1. The shaft dynamics are given by

J ω̇ = τ d − τ c (13.2)

where J is the inertia of the compressor shaft and compressor wheel, ω is the angular
velocity of the shaft, τ d is the drive torque of the motor, and τ c is the compressor torque,
which is the torque acting on the compressor shaft from the rotor blades. The compressor
torque is equal to the rate of change of angular momentum
τ c = w(r2 Cθ2 − r1 Cθ1 ) (13.3)
where w is the mass flow, Cθ1 is the tangential velocity of the fluid at the rotor inlet
and Cθ2 is the tangential fluid velocity at the rotor outlet. The radius at the inlet is r1
and the rotor radius at the outlet is denoted r2 , as shown in Figure 13.2. Assuming no
pre-whirl, an assumption that is met in e.g. most turbocharges, we have that Cθ1 = 0,
and the torque becomes
τ c = wr2 Cθ2 (13.4)
488 CHAPTER 13. COMPRESSOR DYNAMICS

compressor
Motor
d  c

Figure 13.1: Centrifugal compressor with motor.

For backswept blades, we see from the velocity triangle in Figure 13.3 that

Cθ2 = U2 − Cr2 cot β 2b

= (1 − φ cot β 2b ) U2
= µ (φ) U2 (13.5)

where Cr2 is the radial flow velocity,

U2 = r2 ω (13.6)

is the tangential speed of the impeller tip,

Cr2 w r2
φ= = (13.7)
U2 ρ1 A1 ω

is the flow coefficient, and β 2b is the blade angle at the impeller tip. The function

µ (φ) = 1 − φ cot β 2b (13.8)

is the energy transfer coefficient. The torque is seen to be given by

τ c = wµ (φ) r22 ω (13.9)

In the case of radial vanes β 2b = 90◦ we have that cot β 2b = 0 and µ = 1. Backwards
swept blades have β 2b < 90◦ in which case µ decreases with increasing flow coefficient
φ. In practice the energy transfer coefficient is slightly less than the ideal value, and we
may write
µ (φ) = σ (1 − φ cot β 2b ) (13.10)
where σ is the slip factor, which is slightly less than unity. A usual approximation is the
Stanitz slip factor:
2
σ ≈1− (13.11)
N
where N is the number of blades.
13.2. CENTRIFUGAL COMPRESSORS 489

C2
C 2 C r2

r1 
Vanes
r2

Impeller

Figure 13.2: The rotating impeller viewed face-on. Note the radial vanes.

13.2.3 Compressor system

We will derive the dynamic equations for a compressor system that includes a compressor,
a volume Vp called the plenum, and a duct of length L running from the compressor to
the plenum. The model is derived from the mass balance of the plenum, the momentum
equation for the duct and the pressure rise for the compressor. In the derivation the duct
and the plenum is treated as in the derivation of the Helmholtz resonator. This means
that the fluid in the duct is assumed to be incompressible and with mass flow

w = ρAC (13.12)

where ρ is the density in the duct, A is the constant cross section of the duct, and C is
the velocity of the fluid in the duct, which is assumed to be constant along the duct.

13.2.4 Mass balance

The mass balance for the plenum is

Vp ρ̇p = w − wt (pp ) (13.13)

where ρp is the density in the plenum, w is the mass flow from the duct into the plenum,
and wt (pp ) throttle mass flow going out of the plenum. Suppose that the gas in the
plenum is ideal and isentropic. Following the derivation leading to equation (12.67), we
then have
dpp = c2p dρp
where the sonic velocity in the plenum is given by
p
cp = κRTp

and the mass balance becomes

490 CHAPTER 13. COMPRESSOR DYNAMICS

U 2  r2 

W2 C r2
C 2 = W2 + U 2
 2b
C 2

r2


Figure 13.3: Velocity triangle at the impeller exit. Note the backswept vanes.

Compressor Plenum
Vp
Duct
p 01 pp
p 02 A w w t (p p )
Tp
L Throttle

Figure 13.4: Compression system consisting of compressor, duct, plenum volume and
throttle.

c2p
ṗp = [w − wt (pp )] (13.14)
Vp

In Greitzer’s original derivation of this model, it is assumed that cp ≈ c01 , where c01 is
the sonic velocity at ambient conditions. This is equivalent to assume Tp ≈ T01 .

13.2.5 Momentum equation

We will now employ the momentum balance in order to find a differential equation for
the duct mass flow. The mass flow in the duct is given by (13.12). The pressure at
the inlet of the duct is the pressure at the outlet of the diffusor. The fluid enters the
diffusor from the impeller with the stagnation pressure p02 (w, ω). In the diffuser the fluid
is slowed down through an isentropic process where the stagnation enthalpy and hence
the stagnation pressure is kept constant. Due to the reduction in the fluid velocity C the
static pressure p = p0 − 12 C 2 is increased, and if the velocity is reduced to a value close
13.3. COMPRESSOR CHARACTERISTIC 491

to zero, the static pressure at the outlet of the diffuser will be approximately equal to
the stagnation pressure at the outlet of the impeller. Therefore, the pressure at the inlet
of the duct is set to the stagnation pressure at the outlet of the impeller.
The momentum equation of the duct is then
d
(md C) = Ap02 (w, ω) − App (13.15)
dt
where
md = LAρ (13.16)
is the mass in the duct, L is the duct length, p02 (w, ω) is the stagnation pressure at the
outlet of the compressor rotor which is assumed to be equal to the static pressure at
the inlet of the duct, and pp is the plenum pressure. The velocity C of the fluid can be
written
q w
C= = (13.17)
A ρA
where q is the duct volume flow. By combining (13.15), (13.16) and (13.17), we get the
momentum equation for the duct written as a differential equation in mass flow.

Lẇ = A [p02 (w, ω) − pp ] (13.18)

13.3 Compressor characteristic

13.3.1 Derivation
In (13.18) we need an expression for the pressure p02 (w, ω) at the outlet of the compressor
The pressure rise in the compressor is from the stagnation pressure p01 at the inlet, to
the stagnation pressure p02 at the rotor outlet. Generally, this rise in stagnation pressure
can be obtained in an isentropic process 1 → 2s where p02s = p02 . This isentropic
process involves an increase in the stagnation enthalpy ∆h02s = h02s − h01 . In practice,
the pressure increase in the compressor is not isentropic, and there is an increase in
entropy due to friction ∆hf and incidence losses ∆hi . The incidence losses arise from
misalignment of the flow with respect to vane angles at off-design conditions, and the
friction losses are due to friction between the fluid and the various solid surfaces in the
compressor passages. To account for these losses, the compression from p01 to p02 is
modelled as an isentropic process in series with an isobaric process, where the isobaric
process involves an increase in entropy. We recall that for the isentropic process ds = 0
and the enthalpy differential is dh = vdp, while for the isobaric process dp = 0 and
dh = T ds. This results in the following description: First, there is an isentropic process
1 → 2s which ends in a state with pressure p02s = p02 and stagnation enthalpy h02s .
Then, there is an isobaric process 2s → 2, which ends in a state with pressure p02 and
stagnation enthalpy h02 . The total process 1 → 2 is illustrated in Figure 13.5. This gives
Z 2
∆h02 = ∆h02s + T ds
2s
= ∆h02s + ∆hi + ∆hf (13.19)
Here ∆hi is the incidence loss, ∆hf is the friction work, which both will be defined below,
and
∆h02s := ∆h02 − ∆hi − ∆hf (13.20)
492 CHAPTER 13. COMPRESSOR DYNAMICS

p 02
h

2
h 02
p 01

2s
h 02s

h 01
1

Figure 13.5: The isentropic and isobaric process in the compressor.

so that ∆h02s is the change in specific stagnation enthalpy that contributes to the accel-
eration of the gas in the duct.
The mechanical power transferred from the shaft to the gas is

P = τ cω

As there is no energy storage in the gas in the impeller, the power transferred from the
rotor to the gas is equal to the rate of stagnation enthalpy increase

P = w∆h02

of the fluid, where ∆h02 is the increase in specific stagnation enthalpy. By using (13.3),
this gives
w∆h02 = P = τ c ω = w(U2 Cθ2 − U1 Cθ1 ) (13.21)
which is known as Euler’s pump equation. Alternatively, we can use (13.9) and it follows
that
w∆h02 = P = τ c ω = wµ (φ) r22 ω 2 (13.22)
and
τ cω
∆h02 (ω, C) = = µ (φ) r22 ω 2 (13.23)
w
In off-design operation of the compressor there will be a mismatch between the fixed
blade angle β 1b and the direction of the gas stream β 1 . The loss associated with this is
termed the incidence loss. The incidence loss ∆hi , expressed as a change of enthalpy, is
determined according to the so-called NASA model of (Futral and Wasserbauer 1965).
It is assumed that ∆hi is equal to a reduction in kinetic energy corresponding to an
instantaneous change in fluid velocity at the blade inlet such that the axial velocity is
unchanged, while the tangential velocity is changed to accommodate the fixed blade angle
β 1b . The difference between the flow angle and the blade angle,

β i = β 1b − β 1
13.3. COMPRESSOR CHARACTERISTIC 493

Inducer blade

C 1
U1
 1b
1
C a1
i C1
W1 W 1b

1
W 1

Figure 13.6: Velocity triangles at the inducer. Section through inducer at radius r1 .

is called the incidence angle. The velocity of the incoming gas relative to the inducer is
termed W1 . Thus, with Cθ1 = 0, we get from the triangles of Figure 13.6
1 2
∆hi = W
2 θ1
1
= (U1 − cot β 1b Ca1 )2
2
µ ¶2
r12 A
= ω − cot β 1b v
2 A1 r1
r12 2
= (ω − αw) (13.24)
2
where U1 = r1 ω is the rotor blade speed, Ca1 = C (A/A1 ) is the axial velocity at the
blade inlet, ρ1 is the density at the blade inlet which is considered to be a constant, and
A
α = cot β 1b (13.25)
A1 r1
In order to calculate the friction loss, we treat the flow passages in the impeller as
pipes with circular cross section. The friction loss is then modelled as
2
l W1b
∆hf = Ch
D 2
where Ch is the surface friction loss coefficient, l is the mean channel length and D is the
mean hydraulic diameter. The hydraulic diameter for a pipe of non-circular cross section
can be calculated as
4A
D=
a
where A is the cross sectional area and a is the length of the wetted perimeter. Using
Figure 13.6, the friction loss can now be shown to be given by

∆hf = kf w2 (13.26)

where
Ch l
kf =
2Dρ21 A21 sin2 β 1b
494 CHAPTER 13. COMPRESSOR DYNAMICS

is a constant. Combining (13.23), (13.20), (13.24) and (13.26) results in the enthalpy
transfer
r2 2
∆h02s (ω, w) = σr22 ω 2 − 1 (ω − αw) − kf w2 (13.27)
2
The stagnation pressure ratio is given by the stagnation temperature ratio of the
isentropic process 1 → 2s according to
µ ¶ κ−1
κ
p02 p02s T02s
= = (13.28)
p01 p01 T01

An expression for T02s /T01 is found as

T02s h02s h01 + ∆h02s ∆h02s ∆h02s

= = =1+ =1+ (13.29)
T01 h01 h01 h01 cp T01

From (12.63) and (13.27) we find the expression for the pressure ratio
µ ¶ κ
p02 ∆h02s κ−1
Ψc (w, ω) = = 1+ (13.30)
p01 cp T01
à r2 2
! κ−1
κ

σr22 ω 2 − 21 (ω − αv) − kf w2
Ψc (w, ω) = 1+ (13.31)
cp T01

such that
p02 (w, ω) = Ψc (w, ω)p01 (13.32)
Inserting (13.32) into (13.18) gives the momentum balance
õ ¶ κ !
A ∆h02s κ−1
ẇ = 1+ p01 − pp (13.33)
L cp T01

or
A
ẇ = (Ψc (w, ω)p01 − pp )
L
From (13.2), (13.14) and (13.33), the resulting compressor model is found to be

c2p
ṗp = (w − wt ) (13.34)
Vp
à r12 2
! κ−1
κ

L µr2 ω 2 − (ω − αw) − kf w2
ẇ = 1+ 2 2
p01 − pp (13.35)
A cp T01
J ω̇ = τ d − wr22 µω (13.36)

13.3.2 The compressor characteristic at zero mass flow

Let C denote the fluid velocity and h the specific enthalpy. The stagnation enthalpy
1
h0 = h + C 2 (13.37)
2
13.3. COMPRESSOR CHARACTERISTIC 495

is increased over the impeller by the the energy

τ c ω = (r2 Cθ2 − r1 Cθ1 ) ωw = (U2 Cθ2 − U1 Cθ1 ) w (13.38)

that is transferred from the compressor blades to the fluid. Thus, if we let 1 denote the
impeller inlet and 2 denote the impeller outlet we have

(h20 − h10 ) w = (U2 Cθ2 − U1 Cθ1 ) w (13.39)

When the mass flow tends to zero it seems reasonable to assume that the increase in static
enthalpy is continuous at w = 0. Moreover, at zero mass flow through the compressor
we may assume that Cθ1 = U1 and Cθ2 = U2 . This gives the following increase in static
enthalpy for zero mass flow:
¡ ¢
∆h0s = h20 − h10 = U22 − U12 = ω 2 r22 − r12 (13.40)

where r2 is the radius of the impeller outlet and r1 is the radius of the impeller inlet.
The associated pressure rise for the compressor at zero mass flow is
µ ¶ κ Ã ¡ ¢ ! κ−1
κ

p02 h02 − h01 κ−1 ω 2 r22 − r12

= 1+ = 1+
p01 cp T01 cp T01
à ¡ !
¢ κ−1
κ

ρ01 ω 2 r22 − r12

= 1+ κ (13.41)
p01 κ−1

where the ideal gas law has been used together with cp = Rκ/ (κ − 1).

Example 203 By combining (13.37) and (13.39), the increase in enthalpy as opposed
to stagnation enthalpy is found from
µ ¶ µ ¶
1 2 1 2
h1 + C1 − U1 Cθ1 w = h2 + C − U2 Cθ2 w. (13.42)
2 2 2

We introduce the relative velocity Wi = Ci − Ui . In terms of the tangential components

(·)θi and the radial components (·)ri we have

Cθi = Ui + Wθi , Cri = Wri (13.43)

and it follows that

Ci2 = Cθi
2 2
+ Cri = Ui2 + 2Ui Wθi + Wθi
2 2
+ Wri = Ui2 + 2Ui Wθi + Wi2 (13.44)

We insert this expression in the enthalpy equation, and get

µ ¶ µ ¶
1 2 1 2 1 2 1 2
h1 + W1 − U1 w = h2 + W2 − U2 w (13.45)
2 2 2 2

which in the terminology of (Cumpsty 1989) says that the rothalpy

1 1
Ii = hi + Wi2 − Ui2 (13.46)
2 2
496 CHAPTER 13. COMPRESSOR DYNAMICS

is unchanged over the impeller. We may then compute the change in enthalpy at zero
mass flow by assuming continuity when w → 0. At zero mass flow we have Ci = Ui at
the inlet and at the outlet, and therefore Wi = 0. It follows that the increase in enthalpy
over the impeller at zero mass flow is
1¡ 2 ¢ ω2 ¡ 2 ¢
h2 − h1 = U2 − U12 = r − r12 (13.47)
2 2 2
where r2 is the radius of the impeller outlet and r1 is the radius of the impeller inlet. As
1¡ 2 ¢
h20 − h10 = h2 − h1 + C2 − C12 (13.48)
2
this is consistent with the result for the stagnation enthalpy with C1 = U1 and C2 = U2 .
Assuming isentropic pressure rise,
µ ¶ κ
p02 ∆h κ−1
Ψc (w, ω) = = 1+ , w > 0. (13.49)
p01 cp T01
holds, for details see (Gravdahl, Egeland and Vatland 2001), and by combining (13.46)
and (13.49), we get at zero mass flow
µ ¶ κ
π 2 N 2 (D22 − D12 ) κ−1
Ψc (0, N ) = Ψo = 1 + , (13.50)
2cp T01
where N = 2πω is the rotational speed in revolutions per second.

Remark 5 Close to zero mass flow there may be high incidence losses in the diffuser so
that the fluid is not slowed down isentropically, and the kinetic energy is not recovered as
a pressure rise. Therefore, the pressure at the inlet of the duct may be set to the static
pressure at the outlet of the impeller, in which case the pressure rise is
µ ¶ κ Ã ¡ ¢ ! κ−1
κ

p2 h02 − h01 κ−1 ρ01 ω 2 r22 − r12

= 1+ = 1+ κ (13.51)
p01 cp T01 2p01 κ−1

Example 204 The centrifugal force on a material volume element is

ρω 2 rdV (13.52)

Integrating this gives the centrifugal loading at r = r2 for zero mass flow:
Z r2
1 ¡ ¢
ρω 2 rdr = ρ1 ω 2 r22 − r12 (13.53)
r1 2
were the density has been approximated by ρ = ρ1 . Then the increase in static pressure
at zero mass flow is
1 ¡ ¢
p2 ≈ p1 + ρ1 ω 2 r22 − r12 (13.54)
2
and the pressure rise in terms of static pressure is
¡ ¢
p2 ρ1 ω 2 r22 − r12
=1+ (13.55)
p1 2p1
This expression for the pressure rise is approximately equal to the expression (13.51) for
a small pressure rise.
13.4. COMPRESSOR SURGE 497

13.4 Compressor surge

13.4.1 The Greitzer surge model
Compressor surge is a serious problem in the control of compressors, and the main ob-
jective of compressor control systems is to avoid surge, as it reduces performance, and
may lead to the destruction of the compressor. The compressor may surge if the pressure
rise increases with increasing mass flow, which is the case in the unstable region of the
compressor map. The mass flow and the plenum pressure will form an oscillating system,
and even flow reversal may occur.
In the Greitzer model for the description of compressor surge dynamics the shaft
dynamics are left out, and the compressor model (13.34—13.36) is written
c2p
ṗp = [w − wt (pp )] (13.56)
Vp
LρĊ = (Ψc (w, ω) − 1) p01 + p01 − pp (13.57)
The addition of ±p01 in (13.57) has been done to facilitate the use of pressure differences
below. This model is known as Greitzer’s compressor model in dimensional form, and
was first introduced in (Greitzer 1976). We introduce the nondimensional variables
t
ξ= 1 (13.58)
ωH

pp − p01 (Ψc (w, ω) − 1) p01

ψ= 1 2
, ψ c (φ) = 1 2
(13.59)
2 ρU 2 ρU
C w wt (pp )
φ= = , φt (ψ) = (13.60)
U ρAU ρAU
where ξ is the dimensionless time variable, ψ is the pressure coefficient, ψ c (φ) is the
compressor characteristic, φ is the flow coefficient, φt (ψ) is the throttle flow coefficient,
and s
A
ω H = cp (13.61)
Vp L
is the Helmholtz frequency of the duct-plenum system.
Example 205 If the temperature increase over the compressor is small, then the com-
pressor characteristic can be approximated by
"µ ¶ κ #
(Ψc (w, ω) − 1) p01 ∆h02s (ω, C) κ−1 p01
ψ c (φ) = 1 2
= 1+ −1 1 2
ρ
2 01 U c T
p 01 2 01 U
ρ
κ ∆h02s (ω, C) p01
≈ 1 2
κ−1 cp T01 2 ρ01 U
∆h02s (ω, C) p01
= 1 2
2U
RT01 ρ01
∆h02s (ω, C)
= 1 2
2U
µ ¶2
r1 2
= 2µ − (1 − r2 αφ) − 2kf (ρ01 A) φ2 (13.62)
r2
498 CHAPTER 13. COMPRESSOR DYNAMICS

c
Radial vanes 2  2
Backswep
t vanes
2

Figure 13.7: The compressor characteristic from (13.62).

where the expression for ∆h02s (ω, w) from (13.27) has been used. Equation (13.62)
clearly shows that the compressor characteristic is a function of the flow coefficient φ
when ∆T02s is small. The characteristic is plotted in Figure 13.7.

Using the normalized variables the model becomes

1 2 dψ c2p
ρU ω H = ρAU [φ − φt (ψ)] (13.63)
2 dξ Vp
dφ 1 2 1 1
U ωH L = U ψ c (φ) − ρU 2 ψ (13.64)
dξ 2 2 ρ

This gives the following result:

The Greitzer surge model is given by

dψ 1
= [φ − φt (ψ)] (13.65)
dξ B
dφ
= B [ψ c (φ) − ψ] (13.66)
dξ
where the nondimensional parameter
r
U Vp U
B= = (13.67)
2cp AL 2Lω H

is Greitzer’s B-parameter.

A large B corresponds to a small ω H . In axial compressors, the numerical value of

B gives information about the type of instability the compressor will enter if operating
beyond the surge line. A large B indicates that the compressor will enter surge, and a
small B indicates that rotating stall will occur.
13.4. COMPRESSOR SURGE 499

Example 206 Remark 6 The pressure coefficient

∆p
ψ= 1 2
(13.68)
2 ρU

has also been characterized as a temperature coefficient

∆h0s
ψ Cohen = 1 2 (13.69)
2U

in (Cohen, Rogers and Saravanamuttoo 1996).

13.4.2 Linearization
Linearization of the Greitzer model (13.65)-(13.66) gives
µ ¶ µ 1 1 ¶µ ¶
d ψ − Bg B ψ
= t (13.70)
dξ φ −B Bgc φ
where µ ¶−1
∂ψ c ∂φt
gc = , gt = (13.71)
∂φ ∂ψ
are the slopes of the compressor characteristic and the throttle characteristic in a φ, ψ
diagram. The characteristic equation is
µ ¶ µ ¶
1 gc
λ2 + − Bgc λ + 1 − =0 (13.72)
Bgt gt
and the system is seen to be stable if
1
gc < (13.73)
B 2 gt
and
gc < gt (13.74)
¡ 2 ¢−1
The case gc > B gt is usually referred to as a dynamic instability, while the case
gc < gt is called a static instability. We consider a situation where the slope gc of the
compressor characteristic increases from some negative value. We see that if B becomes
small, then a larger gc is tolerated before the dynamic instability is encountered. If B
becomes sufficiently small, that is if B < gt−1 , then the static instability will occur before
the dynamic instability.
Example 207 At the top of the compressor characteristic we have gc = 0, and the
characteristic equation becomes
1
λ2 + λ+1=0 (13.75)
Bgt
This corresponds to a normalized undamped natural frequency ω ∗0 = 1. We recall that
the normalized time variable ξ = tω H is scaled with the Helmholtz frequency, and the
corresponding undamped natural frequency is found to be
ξ
ω 0 = ω ∗0= ωH (13.76)
t
We see that the undamped natural frequency for an equilibrium point at the top of the
compressor characteristic is the Helmholtz frequency ω H .
500 CHAPTER 13. COMPRESSOR DYNAMICS

Example 208 We now assume that there is a valve at the compressor outlet. The valve
characteristic is v
u 2ρ (p − p )
ẇ = CD A2 u
1 2
t ³ ´2 (13.77)
1− A 2
A1

The plenum volume in this case is small, and we may assume that B < gt−1 . Then the
stability requirement is
gc < gt (13.78)
that is, the static instability must be avoided.

Example 209 If the compressor delivers gas to a pipeline the plenum volume is Vp → ∞
and it follows that B → ∞. In this case ψ will be a constant, and the system is described
by the first order model
dφ
= B [ψ c (φ) − ψ] (13.79)
dξ
This model is stable whenever
gc < 0 (13.80)
This means that a compressor connected to a pipeline is stable as long as the compressor
characteristic has a negative slope.

13.4.3 Passivity of the Greitzer surge model

We will now investigate the passivity properties of the Greitzer model. The model is
given as
dψ 1
= [φ − φt (ψ)] (13.81)
dξ B
dφ
= B [ψ c (φ) − ψ] (13.82)
dξ

where (13.81) is the pressure dynamics, and (13.82) is the mass flow dynamics. The
throttle flow φt (ψ) is given as p
φt (ψ) = γ t ψ (13.83)
Consider the positive function
µ ¶
1 1
V = V1 + V2 = Bψ + φ2 2
(13.84)
2 B

The time derivative of V1 along the solutions of (13.81) is

dV1
= ψφ − ψφt (ψ) (13.85)
dξ

Integrating (13.85) we see that

Z T Z T
ψφdξ = V1 (T ) − V1 (0) + ψφt (ψ) dξ
0 0
Z T
> −V1 (0) + ψφt (ψ) dξ
0
13.4. COMPRESSOR SURGE 501

which implies that the system with φ as input and ψ as output has certain passivity
RT
properties depending on the value of the term 0 ψφt (ψ) dξ. The throttle in a compres-
sion system is a passive component, and from (13.83) it can be seen that a coefficient κ1
can always be chosen sufficiently small such that the throttle characteristic satisfies the
condition
ψφt (ψ) ≥ κ1 ψ 2 (13.86)
We now find that Z Z
T T
ψφdξ > −V1 (0) + κ1 ψ 2 dξ (13.87)
0 0

and it follows that the pressure dynamics are (strictly) passive.

The time derivative of V2 along the solutions of (13.82) is

dV2
= φψ c (φ) − ψφ (13.88)
dξ
Following the same procedure as above, we find that
Z T Z T
−ψφdξ = V2 (T ) − V2 (0) + φψ c (φ) dξ
0 0
Z T
> −V2 (0) + φψ c (φ) dξ (13.89)
0

and the passivity properties of the system with input −ψ and output φ depends on the
RT
value of the term 0 φψ c (φ) dtξ. For operating points where the slope of the compressor
characteristic is negative, the sector condition

φψ c (φ) ≥ κ2 φ2

will hold, and

Z T Z T
−ψφdξ > −V2 (0) + κ2 φ2 dξ (13.90)
0 0

and it follows that the mass flow dynamics are (strictly) passive. This leads us to the
known result that the feedback interconnection of the two systems, that is the Greitzer
model, is a passive system when the operating point is located on a part of the compressor
characteristic with negative slope.
In the case where the operating point is not located in an area of negative slope,
that is the system is in an unstable (surge) condition, we still have that the pressure
dynamics are passive according to (13.87). The flow dynamics are now not passive
according to (13.90), but if it is possible to manipulate, through some external actuator,
the compressor characteristic ψ c (φ), (13.89) can be used for controller design. This
approach was taken in (Gravdahl and Egeland 1998) using a close coupled valve for
active surge control, and in (Gravdahl and Egeland 2002) using the drive itself.

13.4.4 Curvefitting of compressor characteristic

When modelling a real compression system it might be difficult to obtain models of
the compressor characteristic like (13.30) or (13.62). A compressor is usually equipped
with a measured compressor map. This might be measured when the compressor is
manufactured or when it is installed. It is then an alternative to use an approximation
502 CHAPTER 13. COMPRESSOR DYNAMICS

of this measured characteristic in the model of the system. In this example it is shown
how to use a polynomial approximation. The solid circles in Figure 13.8 are points read
from an actual compressor map for a centrifugal compressor used in pipeline natural gas
transport in the North Sea in Norway. It is to be noted that points near the stonewall
area of the measured characteristic were not used. In that area choking is the dominant
effect, and this is not taken into account in the modeling. Also, in order to ensure that
the approximated constant speed lines do not cross each other in the negative flow area,
one point for negative flow was chosen for each speed line. The zero flow pressure rise for
each speed as calculated by (13.50), and the dotted lines are the third order polynomial
approximations of the speed lines at five different speeds.

5.5
500/s

4.5
460/s

3.5
400/s
3

2.5 340/s

2
300/s

1.5

1
-2 0 2 4 6 8 10

Figure 13.8: The measured speed lines (solid lines) and the polynomial approximations
(dashed lines).

The approximations are calculated using the Matlab function polyfit, and for the
five chosen speed lines, results are:
Ψc (w, 300) = 1.6024 − 0.0625w + 0.1668w2 − 0.0441w3 ,
Ψc (w, 340) = 1.8291 − 0.0966w + 0.1825w2 − 0.0304w3 ,
Ψc (w, 400) = 2.2511 − 0.1443w + 0.1908w2 − 0.0240w3 ,
Ψc (w, 460) = 2.8092 − 0.1658w + 0.1783w2 − 0.0177w3 ,
Ψc (w, 500) = 3.2699 − 0.2051w + 0.1744w2 − 0.0147w3 .
A compressor map is also continuous in the rotational speed, as can be seen from (13.30),
so in order to simulate the system, there is a need for making the approximated map also
continuous in rotational speed. For this reason, the coefficients of the third order poly-
nomials in are chosen to be functions of rotational speed. The polynomial approximation
for each speed line can be written as
Ψc (w, N ) = c0 (N ) + c1 (N ) w + c2 (N ) w2 + c3 (N ) w3 ,
where the functions
ci (N ) = ci0 + ci1 N + ci2 N 2 + ci3 N 3
13.4. COMPRESSOR SURGE 503

are calculated by using polynomial approximation yet again. In Figure 13.9, the polyno-
mial coefficients of the five polynomials are plotted as a function of rotational speed. As
can be seen, a fairly good fit can be made with third order.

-0.01 0.195

0.19
-0.02
0.185

c 2 (N)
c3 (N)

-0.03 0.18

0.175
-0.04
0.17

-0.05 0.165
300 350 400 450 500 300 350 400 450 500
N [1/s] N [1/s]

0.1 3.5

0 3
c 1 (N)

-0.1 c 0 (N)2.5

-0.2 2

-0.3 1.5
300 350 400 450 500 300 350 400 450 500
N [1/s] N [1/s]

Figure 13.9: The coefficients ci as functions of speed N (solid lines), and their polynomial
approximations (dashed lines).

Remark 7 For the presentation of compressor characteristics which describe the sta-
tionary performance of compressors another normalization is often used. This is based
on using the Mach number to represent flow and blade velocity. We introduce as nondi-
mensional variables the Mach numbers
C U
MC = , MU = (13.91)
c0 c0
where p
c0 = κRT0 (13.92)
2
c
and T0 = T + 2 is the stagnation temperature. Then, using w = ρAC, we find that
√
C w wRT0 ρ0 ρ0 w RT0
MC = = √ = √ = √ (13.93)
c0 ρA κRT0 Ap0 κRT0 ρ ρ κ Ap0
and that
U ωr 1 ωr
MU = =√ =√ √ (13.94)
c0 κRT0 κ RT0
For plotting of compressor characteristics the following dimensionless variables are used:
The corrected mass flow √
√ w RT
κMC = (13.95)
Ap
504 CHAPTER 13. COMPRESSOR DYNAMICS

the corrected speed,

√ ωr
κMU = √ (13.96)
RT
and the pressure ratio p02 /p01 .

13.4.5 Compression systems with recycle

As mentioned in Section 13.1.2, the surge problem in centrifugal compressor can be
handled by using a recycle line around the compressor and thereby ensuring that the
flow is maintained above a certain minimum value. Such a recycle concept is illustrated
in Figure 13.10.

wr
Recycle line
w1
wf V 1 ,p 1 ,T 1 V 2 ,p 2 ,T 2 wt

Q1 , d Q2
Compressor

Drive

c
w1

Figure 13.10: A centrifugal compressor with drive unit, upstream volume, downstream
volume and recycle line.

A model of this system can be found by calculating the mass balances for the two
volumes, calculating the momentum balance of the downstream compressor duct, and
calculating the torque balance of the shaft:

c201
ṗ1 = (wf − w1 + wr ) (13.97a)
V1
c201
ṗ2 = (w1 − wr − wt ) (13.97b)
V2
A
ẇ1 = (Ψc (w1 , ω)p1 − p2 ) (13.97c)
Lc
1
ω̇ = (τ d − τ c ) , (13.97d)
J
where, p1 is the pressure in the upstream volume V1 , p2 is the pressure in the downstream
volume V2 , wf is the feed mass flow, w1 is the mass flow through the compressor, wr is
the recycle mass flow, wt is the throttle flow, Ψc (w1 , ω) is the compressor characteristic,
ω is the rotational speed of the compressor, τ d is the drive torque, τ c is the compressor
torque, J is the inertia of all rotating equipment, c01 is the sonic velocity at ambient
conditions.
The mass flow through the recycle valve is given as
p
wr = kr ∆pr , (13.98)
13.4. COMPRESSOR SURGE 505

where ∆pr is the pressure drop across the valve.

In order to ensure high efficiency, compression systems are often equipped with cool-
ers. In order to take this into account, we need to study the energy flow in the system.
By combining the mass balance

ṁ = win − wout

for a volume Vj with the energy balance

d X
U= wi hi − Q,
dt i

and assuming ideal gas such that

pV = mRT,
one gets
d X
(mu) = wcp Ti − Q
dt i

(win − wout ) u + mcv Ṫ = win cp Tin − wut cp T + Q

RT
Ṫ = (win cp Tin − (wut R + win cv ) T + Q) (13.99)
pV cv
which is the energy balance for the volume in terms of temperature. Here U is internal
energy, u = cv T is specific internal energy, h = cp T is specific enthalpy, cp and cv are the
specific heats at constant pressure and volume, R = cp − cv is the specific gas constant
and Q is heat flow. One energy balance (13.99) is used per volume.
506 CHAPTER 13. COMPRESSOR DYNAMICS
 Part V

Simulation

507
Chapter 14

Simulation

14.1 Introduction
14.1.1 The use of simulation in automatic control
Simulation of dynamic processes involves the numerical solution of differential equations
which are normally in the form of initial value problems. The numerical schemes that
are used for this are called numerical integrators. There is a large literature on numerical
integrators (Hairer, Nørsett and Wanner 1993), (Hairer and Wanner 1996), (Lambert
1991), (Shampine 1994), and a wide range of methods are available. These methods have
different properties and the selection of which method to use depends on the properties of
the system to be simulated. In this chapter a range of numerical integrators are presented
and analyzed, and it is attempted to give some advice on how a suitable method can be
selected for important dynamic systems.
Simulation plays an important part in the design, maintenance and upgrading of
control systems. Dynamic systems to be controlled are usually described by differential
equations or transfer functions, and simulation is used to check the qualitative behavior
of the system for typical parameter values and for expected modes of operation. When a
controller is designed for a system it is usual practice to test the controller in simulations
before implementing it. This allows for rapid changes and correction of errors before the
system is designed. Also it is important that procedures for handling of discrete events
and errors can be tested. For systems where a controller has already been developed,
quantitative aspects of simulation is important for the fine tuning of controller parameters
and the redesign of the system to be controlled.
An example where this is useful is in the development of industrial robots. In appli-
cations like spot welding in car production lines there are ever-increasing demands on the
robot to finish spot welding tasks faster while maintaining the weld quality specifications.
Then, simulation must be used to improve controller parameters, to try out friction com-
pensation, and to improve the mechanical construction so that elastic deformations can
be reduced. The alternative to using simulation would be iterative mechanical redesign
which is costly and time consuming.
In car engines new regulations of emissions are enforced, and there is a demand for
engines that are lighter, that use less fuel and that pollute less. The introduction of new
electronic control systems is necessary to achieve this. Car manufacturers use simulation
systems to reduce mechanical vibrations, to shape the combustion chamber for efficient
combustion of the fuel, to reduce the formation of pollutants, to optimize electronic

509
510 CHAPTER 14. SIMULATION

Figure 14.1: The model implemented in Simulink.

controls for components like fuel injectors, turbochargers, and valves for exhaust gas
recycling. Also in the design and testing of control systems for ABS brakes simulation is
an indispensable tool.
For ship control systems there are large costs involved in commissioning of control
systems, which involves installation and calibration. Moreover, a typical situation is that
there is very little time available for the control engineer to commission the controllers
before the ship is to be set into commercial operation. By use of simulation the time for
commissioning can be reduced significantly, and this may be a decisive factor to make
control systems commercially attractive for the marine industry.
The last few years new and powerful tools have been made available for simulation
which makes it much easier to run simulations than what have been the case. Also
simulation tools and control systems development tool have been integrated, and the
role of simulation in automatic control is becoming even more important than it used to
be. Still, it is important to know the properties of the numerical schemes that are used
so that the results can be interpreted in the right way.
In the following, three examples are presented where the dynamics of systems with-
out controllers are presented. Simulation of the dynamics of these systems reveals the
qualitative properties of the systems, and this is useful a starting point for designing
controllers. MATLAB code is included for two of the examples to make it easy for the
reader to simulate the systems with MATLAB or SIMULINK.

14.1.2 The Moore Greitzer model

A jet engine consists basically of a compressor, a combustion chamber, a turbine and
connecting ducts. The compressor delivered compressed air to the combustion chamber
14.1. INTRODUCTION 511

0.8

0.7

0.6

0.5

0.4
Φ, Ψ, J

0.3

0.2

0.1

-0.1

-0.2
0 10 20 30 40 50 60 70 80 90 100
Time

Figure 14.2: Simulation of rotating stall. Φ, Ψ and J are plotted with dashed solid and
dash-dotted lines respectively.

1.2

0.8
Φ, Ψ

0.6

0.4

0.2
0 5 10 15 20 25 30 35
Time
4

2
J

0
0 5 10 15 20 25 30 35
Time

Figure 14.3: Simulation of surge. Φ, Ψ and J are plotted with dashed solid and dash-
dotted lines respectively.
512 CHAPTER 14. SIMULATION

where fuel is added. The gas expands and drives the turbine which delivers the required
power to the compressor over a shaft. The thrust force comes from the mass flow through
the restriction behind the turbine, and in modern jet from a fan that is driven by the
turbine If the jet-stream of another aeroplane meets the intake, or if the aspect angle of
the aeroplane becomes too large there will be a severe disturbance in the fluid flow at
the compressor inlet. This may cause the mass flow through the compressor to become
smaller than a critical value given by the surge line, and the engine will enter one of two
unstable operating modes known as surge and rotating stall. Surge is an axisymmetric
pulsation of the flow through the compressor, while rotating stall is an instability where
the circumferential flow pattern is disturbed. If the engine enters rotating stall it will
be necessary to shut down the engine, which may lead to the plane falling out of the
sky. These physical phenomena are described by the Moore-Greitzer model (Moore
and Greitzer 1986) that describes the transients in an axial compression system like an
aircraft jet engine. Based on this model a number of control systems have been designed
to stabilize surge and stall. This is discussed in (Gravdahl and Egeland 1999). In the
model the turbine is modeled as a throttle and the combustion chamber is called the
plenum. The model consists of three nonlinear differential equations, and the states are
plenum pressure, mass flow and rotating stall amplitude. The rotating stall amplitude
is a measure of the unstable non-axisymmetric flow disturbance. All states have been
made normalized. The Moore-Greitzer model is written
1 ³ √ ´
Ψ̇ = Φ − γ T Ψ (14.1)
4lc B 2
à µ ¶µ ¶ µ ¶3 !
H Ψ − ψ c0 3 Φ J 1 Φ
Φ̇ = − +1+ −1 1− − −1 (14.2)
lc H 2 W 2 2 W
à µ ¶2 !
˙ Φ J
J =J 1− −1 − σ (14.3)
W 4

where Ψ is the nondimensional plenum pressure (pressure divided by density and the
square of compressor rotational speed), Φ is the average mass flow coefficient (axial flow
velocity divided by compressor rotational speed), and J is the squared amplitude of
rotating stall amplitude. The constant lc is the total length of the compressor and duct,
Ac is the cross sectional flow area, γ T is a parameter proportional to the throttle opening,
and H, W , ψ c0 and σ are constants describing the compressor. Finally,
r
U Vp
B= (14.4)
2c Ac lc
is Greitzer’s B-parameter, where U is the tangential speed of the compressor, c is the
speed of sound, Vp is the plenum volume, and Ac is the cross sectional flow area.
Numerical values for a laboratory compression system in unstable operation is H =
0.18, W = 0.25, ψ c0 = 0.30, σ = 0.38, γ T = 0.5 and lc = 2. Initial conditions that
correspond to a stable operating point are given by Ψ(0) = Φ(0) = 0.6, J(0) = 0.1. By
setting the B-parameter at B = 0.1, the engine will go into rotating stall, and by setting
the B-parameter at B = 1, the engine will start to surge.
The model (14.1)-(14.3) can be simulated in SIMULINK under MATLAB by imple-
menting the following SIMULINK s-function:
function [sys,x0] = MooreGreitzer(t,x,u,flag)
H=0.18;
14.1. INTRODUCTION 513

W=0.25;
l_c=2;
psi_co=0.30;
s=0.38;

if flag == 1,
%return state derivatives
gamma_T=u(1);
B=u(2);
sys(1)=1/(4*l_c*B^2)*(x(2)-gamma_T*sqrt(x(1)));
sys(2)=H/l_c*(-(x(1)-psi_co)/H+1+1.5*(x(2)/W-1)*(1-0.5*x(3))
-0.5*(x(2)/W-1)^3);
sys(3)=x(3)*(1-(x(2)/W-1)^2-x(3)/4)*s;
elseif flag == 0,
% return initial conditions
sys=[3;0;3;2;0;0];
x0=[0.6;0.6;0.1];
elseif flag == 3,
% return outputs
sys=[x(1) x(2) x(3)];
else
sys = [];
end

The model may now be simulated in SIMULINK by making a block diagram as shown
in Figure 14.1. The simulation result for both rotating stall and surge is shown in Figures
14.2 and 14.3.

14.1.3 The restricted three-body problem

The restricted three-body problem describes the motion of a satellite moving in the
combined gravitational field of the moon and the earth. There are three bodies in the
problem, the satellite, the moon and the earth. The mass of the spacecraft is assumed
to be so small that it does not influence the motion of the moon or the earth. The
normalized model is derived in Section 8.9.3, and is given by

ẋ = vx (14.5)
ẏ = vy (14.6)
m1 (x + m2 ) m2 (x − m1 )
v̇x = 2vy + x − − (14.7)
r13 r23
m1 y m2 y
v̇y = −2vx + y − 3 − 3 (14.8)
r1 r2

where
q
r1 = (x + m2 )2 + y 2 (14.9)
q
r2 = (x − m1 )2 + y 2 (14.10)
m1 + m2 = 1 (14.11)
514 CHAPTER 14. SIMULATION

Orbit Numerical values

x0 0.994
y0 0
vx0 0
1
vy0 −2.00158510637908252240537862224
T 17.0652165601579625588917209
m2 0.012277471
x0 0.994
y0 0
vx0 0
2
vy0 −2.0317326295573368357302057924
T 11.124340337266085134999734047
m2 0.012277471
x0 1.2
y0 0
vx0 0
3
vy0 −1.04935750983031990726
T 6.192169333131963970674
−1
m2 (82.45)

Table 14.1: Initial conditions and periode T for three periodic orbit of the resticted
three-body problem.

Here x and y are the position coordinates of the satellite, vx is the velocity in the x
direction and vy is the velocity in the y direction. The mass of the earth is m1 and the
mass of the moon is m2 . The acceleration terms are due to the gravitational field, and
Coriolis and centrifugal effects due to the rotation of the earth-moon system. The energy
function of the system is given by
1¡ 2 ¢ m1 m2
h= v + vy2 − x2 − y 2 − − (14.12)
2 x r1 r2
and the conservation of energy implies that h is a constant during the motion of the
system. This can be used to check the accuracy of computed solutions.
We would like to compute the solution of the differential equation using a numerical
scheme. Several periodic orbits have been found for this system that can be used to check
the accuracy of numerical integrators (Hairer and Wanner 1996), (Shampine et al. 1997).
It turns out that the solution is very sensitive close to the moon at (x, y) = (1, 0), and
close to the earth at (x, y) = (0, 0). As a consequence of this, a standard fixed step
integrator will be useless for the integration of this system. The widely used Euler’s
method give large errors even with 24000 time steps per orbit, and even the fourth order
Runge-Kutta method RK4 gives significant errors with 6000 time steps. The parameters
describing three periodic orbits are given in Table 14.1.
The solutions were computed with the ode45 function in MATLAB with a relative
tolerance of 10−6 . The computation of the first orbit took 697 steps for Orbit 1, 621
steps for Orbit 2, and 601 steps for orbit 3. The results for the computation of two orbits
are shown in Figures 14.4—14.6. Note that the integration is sufficiently accurate for the
two orbits to coincide.The code for generating the plots is the MATLAB script

tf1=17.0652165601579625588917206249; %Periode
14.1. INTRODUCTION 515

1.5

0.5

-0.5

-1

-1.5
-1.5 -1 -0.5 0 0.5 1 1.5

Figure 14.4: Periodic orbit 1 of the restricted three-body problem

1.5

0.5

-0.5

-1

-1.5
-1.5 -1 -0.5 0 0.5 1 1.5

Figure 14.5: Periodic orbit 2 of the restricted three-body problem

1.5

0.5

-0.5

-1

-1.5
-1.5 -1 -0.5 0 0.5 1 1.5

Figure 14.6: Periodic orbit 3 of the restricted three-body problem

516 CHAPTER 14. SIMULATION

tf2=11.124340337266085134999734047;
tf3=6.192169333131963970674;
x0=0.994; x03=1.2; y0=0; vx0=0.0; %Initial conditions
vy01=-2.00158510637908252240537862224;
vy02=-2.0317326295573368357302057924;
vy03=-1.04935750983031990726;
N=1; %Number of orbits
options = odeset(’RelTol’,1e-6);
[t,y] = ode45(’OrbitODEEq’,[0 N*tf1],[x0 0 0 vy01],options);
plot(y(:,1),y(:,2),0,0,’.’,1,0,’.’);
axis([-1.5 1.5 -1.5 1.5]); grid; size(t) %number of steps
options = odeset(’RelTol’,1e-6);
[t,y] = ode45(’OrbitODEEq’,[0 N*tf2],[x0 0 0 vy02],options);
figure; plot(y(:,1),y(:,2),0,0,’.’,1,0,’.’);
axis([-1.5 1.5 -1.5 1.5]);grid; size(t) %number of steps
options = odeset(’RelTol’,1e-6);
[t,y] = ode45(’OrbitODEEq2’,[0 N*tf3],[x03; 0; 0; vy03],options);
figure; plot(y(:,1),y(:,2),0,0,’.’,1,0,’.’);
axis([-1.5 1.5 -1.5 1.5]); grid; size(t) %number of steps

and the function

function dydt = OrbitODEEq(t,y)

m2 = 0.012277471;
m1 = 1 - m2;
r13 = (((y(1) + m2)^2 + y(2)^2) ^1.5);
r23 = (((y(1) - m1)^2 + y(2)^2) ^1.5);
dydt = [ y(3)
y(4)
(2*y(4) + y(1) - m1*((y(1)+m2)/r13)...
- m2*((y(1)-m1)/r23))
(-2*y(3) + y(2) - m1*(y(2)/r13)...
- m2*(y(2)/r23)) ];

The function OrbitODEEq2 is identical to OrbitODEEq except for the numerical value
of m2 .

14.1.4 Mass balance of chemical reactor

A chemical reaction

A → B (slow)
B+B → B+C (very fast)
B+C → A+C (fast)

in a closed tank has the mass balance equations

ẏ1 = −0.04y1 + 104 y2 y3 y1 (0) = 1

ẏ2 = 0.04y1 − 104 y2 y3 −3 · 107 y22 y2 (0) = 0 (14.13)
ẏ3 = 3 · 107 y22 y3 (0) = 0
14.2. PRELIMINARIES 517

The solution of these equations can be computed numerically. This is a difficult system,
however, to integrate, as it has both very fast and slow dynamics. Because of this,
the process has been used as a benchmark for testing the performance of numerical
integrators.

14.2 Preliminaries
14.2.1 Notation
We will investigate the problem of computing a numerical solution to the initial value
problem
ẏ = f (y,t), y(t0 ) = y0 (14.14)
The system has the exact solution y(t), and we would like to compute a numeric solution
which approximates the exact solution with satisfactory accuracy. This will be done
with a time step h so that the solution is computed for (t0 , t1 , . . . , tn , . . . , tN ) where
tn+1 − tn = h. The numerical solution at time tn is denoted yn , while the exact solution
at time tn is denoted y(tn ).

14.2.2 Computation error

To analyze the accuracy of a computed solution it is useful to have a measure of how
much the error increases in one time-step. To do this we introduce the concept of a local
solution yL (tn ; t), which is the exact solution of (14.14) with initial condition yn at tn ,
that is,
ẏL (tn ; t)= f [yL (tn ; t)], yL (tn ; tn ) = yn (14.15)
In particular we will be concerned with the local solution at the next time-step, which
is yL (tn ; tn+1 ). The deviation of the computed solution yn+1 from the local solution
yL (tn ; tn+1 ) will then be the error introduced by the numerical scheme from time tn to
time tn+1 .

The local error en+1 is the is the difference of the computed solution yn+1 from the local
solution yL (tn ; tn+1 ) at time tn+1 :

en+1 = yn+1 − yL (tn ; tn+1 ) (14.16)

The global error En+1 is the error in the computed solution yn+1 relative to the exact
solution y(tn+1 ) at time tn+1 :

En+1 = yn+1 − y(tn+1 ) (14.17)

The local error en+1 is the error in the solution resulting from the computation from
time tn to tn+1 .The global error En+1 is the error in the solution resulting from the
computation from initial time t0 to tn+1 .

14.2.3 The order of a one-step method

A one-step method is a numerical scheme which computes yn+1 as a function of yn
according to
yn+1 = yn + hφ(yn , tn ) (14.18)
518 CHAPTER 14. SIMULATION

where φ(·) is given by the particular numerical method that is used. We would like our
method to give a small error in some sense when the time step is small. One way of
characterizing different methods is the concept of the order of the method. We say that
the method is of order p if p is the smallest integer so that

en+1 = O(hp+1 ) (14.19)

Here we have used the order notation O (·) (Lin and Segel 1974). The function φ(x)
satisfies
φ(x) = O[ψ(x)] (14.20)
if there exists a constant C > 0 so that

|φ(x)| ≤ C |ψ(x)| (14.21)

when x is close to zero.

Example 210 The expression φ(x) = O (xm ) implies that there exists a C > 0 so
that |φ(x)| ≤ C |xm |. Moreover, if C > 0, then Chm = O(hm ), which is implied by
|Chm | ≤ C |hm |.

To investigate the order of a method it is useful to develop the Taylor series expansion
of yL (tn ; tn+1 ) around yn . The Taylor series is given by

hp dp−1 f (yn , tn ) hp+1 dp f [yL (τ ), τ ]

yL (tn ; tn+1 ) = yn + hf (yn , tn ) + . . . + + (14.22)
p! dtp−1 (p + 1)! dtp

where tn ≤ τ ≤ tn+1 . As the local error is en+1 = yn+1 − yL (tn ; tn+1 ), and we arrive at
the following result

A one-step method is of order p if p is the smallest integer so that en+1 = O(hp+1 ). If

the numerical solution yn+1 satisfies the equation

hp dp−1 f (yn , tn )
yn+1 = yn + hf (yn , t) + . . . + + O(hp+1 ) (14.23)
p! dtp−1

then en+1 = O(hp+1 ), and it follows that the method is of order p.

Analysis of the global error is somewhat more complicated. However, we state without
further analysis that for one-step methods the global error is En+1 = O(hp ).

14.2.4 Linearization
The stability and performance of a one-step method for the numerical integration of
(14.14) can be investigated in terms of the linearized system equations, and in this
section we will see how this can be done. The basic idea is to apply a one-step method
to the linearized system. From basic systems theory it is known that the dynamics of a
linearized system is to a large extent determined by the location of the eigenvalues of the
Jacobian matrix. In the same way, the performance of a one-step method applied to a
linear system can be described by the eigenvalues of the Jacobian in terms of the stability
function of the method. We will first establish the necessary mathematical background
for this.
14.2. PRELIMINARIES 519

Suppose that y∗ (t) is a solution of the differential equation

ẏ∗ = f (y∗ ,t), y∗ (t0 ) = y0∗ (14.24)
Linearization of the differential equation around the solution y∗ (t) is based on writing
y = y∗ + ∆y and using the Taylor series
¯
∂f (y,t) ¯¯
ẏ∗ + ∆ẏ = f (y∗ ,t) + ∆y (14.25)
∂y ¯y=y∗

We define the Jacobian J of the system to be

¯ ( ¯ )
∂f (y,t) ¯¯ ∂fi (y,t) ¯¯
J= = (14.26)
∂y ¯y=y∗ ∂yj ¯y=y∗

and obtain the linearization of (14.14) which is

∆ẏ = J∆y (14.27)
The solution ∆y(t) of (14.27) can be expressed as a linear combination
d
X
∆y = qi (t)mi (14.28)
i=1

of solutions qi (t) of the scalar differential equations

q̇i = λi qi , i = 1, . . . , d (14.29)
where q = (q1 , . . . , qn )T , λi are the eigenvalues of J, and mi are the eigenvectors of J.
This means that we can study the dynamics of the linearized system (14.27) by finding
the eigenvalues of J. In particular, if we apply a one-step method to (14.27), then the
solution ∆yn+1 will be the same as if we apply the method to (14.29) and compute
d
X
∆yn+1 = (qi )n+1 mi (14.30)
i=1

Suppose that there is a function R(s), which will be called the stability function, so
that the one-step method gives the numerical solution
(qi )n+1 = R(hλi )(qi )n (14.31)
when applied to (14.29). Then
d
X
∆yn = Rn (hλi )(qi )0 mi (14.32)
i=1

and the following conclusion may be drawn:

The numerical solution ∆yn of the linearized system (14.27) is stable if the magnitude
of the stability function is less than or equal to unity for all the eigenvalues, that is, if

|R(hλi )| ≤ 1, i = 1, . . . , d (14.33)

where h is the step-length and λi is an eigenvalue of J.

520 CHAPTER 14. SIMULATION

Example 211 Consider the system

ẏ1 = y2 (14.34)
ẏ2 = −γy13 − cy2 (14.35)

The linearization around y1 = 0, y2 = 0 is

µ ¶ µ ¶µ ¶
ẏ1 0 1 y1
= (14.36)
ẏ2 0 −c y2

which has eigenvalues λ1 = 0 and λ2 = −c. The linearization around a solution y1∗ (t),
y2∗ (t) is
µ ¶ µ ¶µ ¶
∆ẏ1 0 1 ∆y1
= 2 (14.37)
∆ẏ2 −3γ (y1∗ ) −c ∆y2

The eigenvalues are given by

2
λ2 + cλ + 3γ (y1∗ ) = 0 (14.38)

which gives
r³ ´
c c 2
λ=− ± − 3γ (y1∗ )2 (14.39)
2 2
¡ ¢2
which implies that Re[λ] ≤ 0. We see that for large |y1∗ |, that is when 3γ (y1∗ )2 À 2c ,
then the system becomes oscillatory, while for small |y1∗ | the system is overdamped.

14.2.5 The linear test function

Important insight on the properties of a numerical integration scheme is gained by analyz-
ing the performance of the method for the linearization of the system. From the previous
section it is clear that the performance of a numerical integrator for linear systems can
be investigated by applying the method to the linear test system

ẏ = λy (14.40)

The numerical solution for this system is

yn+1 = R(hλ)yn (14.41)

where R(hλ) is the stability function for the method. Stability of the numerical scheme
is ensured if the difference equation satisfies

|yn+1 | ≤ |yn | (14.42)

and we see that this is ensured if

|R(hλ)| ≤ 1 (14.43)

This gives conditions on the time-step h and the location of the eigenvalue λ for the
numerical solution to be stable.
14.3. EULER METHODS 521

14.3 Euler methods

14.3.1 Euler’s method
A simple but important numerical integration scheme is Euler’s method, where the nu-
merical solution is computed from

yn+1 = yn + hf (yn , tn ) (14.44)

Comparison with (14.23) shows that the method is of order 1.

y n+1
k1 yL (t)

yn

t
tn t n+1

Figure 14.7: Euler’s method

The linear stability of Euler’s method can be investigated with the scalar test system

ẏ = λy (14.45)

Euler’s method gives

yn+1 = yn + hλyn = (1 + hλ)yn (14.46)
which shows that the stability function is

R(hλ) = 1 + hλ (14.47)

Stability is ensured whenever

|R(hλ)| = |1 + hλ| ≤ 1 (14.48)

This is the case if hλ is inside the circle of radius one around −1. For real eigenvalues λ
stability is ensured when
2
− ≤λ≤0 (14.49)
h
or, equivalently,
2
h≤− (14.50)
λ
The region of stability is shown in Figure 14.14.
522 CHAPTER 14. SIMULATION

Example 212 The system

ẏ = −y, y(0) = 1 (14.51)
was integrated for 0 ≤ t ≤ 8 with Euler’s method. The stability limit for the time step
is h = 2, as λ = −1 for this system. First a solution was calculated with h = 0.5, then
with h = 1.5, then with the stability limit h = 2.0, and finally with the unstable value
h = 2.2. The results are shown in Figure 14.8. It is clear from the results that the time
step should be less than h = 0.5 to achieve a reasonably accurate solution.

1.5

h=2.2
0.5 exp(-t)

h=1.5
h= 0.5
0

-0.5 h=2.0

-1

-1.5
0 1 2 3 4 5 6 7 8

Figure 14.8: Calculated solutions for the system ẏ = −y with Euler’s method for four
different time steps h. The exact solution exp(−t) is shown for comparison.

Example 213 The first order system

ẏ = −βy 3 (14.52)

in Euler’s method gives the algorithm

yn+1 = yn − hβyn3 (14.53)

The linearization of the differential equation around zero gives

ẏ = 0 (14.54)

that is, the test system ẏ = λy with λ = 0. This is stable for all time steps h.

Example 214 The system

ẏ = −αy − βy 3 (14.55)
in Euler’s method gives ¡ ¢
yn+1 = yn − h αyn + βyn3 (14.56)
The linearization of the differential equation around zero gives

ẏ = −αy (14.57)
14.3. EULER METHODS 523

while the linearization around y ∗ gives

h i
2
∆ẏ = − α + 3β (y ∗ ) ∆y (14.58)

A largeh |y ∗ | requiresia small h for the stability condition to hold. Here, the eigenvalue is
λ = − α + 3β (y ∗ )2 , and the stability condition for the linearized system is

2
h≤ 2 (14.59)
α + 3β (y ∗ )

Example 215 Consider the second order system

ẍ = F (x, ẋ) (14.60)

To apply Euler’s method, the system must first be brought into the form (14.14). This
can be done by defining y1 = x and y2 = ẋ. This gives

ẏ1 = y2 (14.61)
ẏ2 = F (y1 , y2 ) (14.62)

and Euler’s method gives the integration algorithm

y1,n+1 = y1,n + hy2,n (14.63)

y2,n+1 = y2,n + hF (y1 , y2 ) (14.64)

14.3.2 The improved Euler method

The improved Euler method includes an evaluation y bn+1 = yn + hf (yn , tn ) according to
Euler’s method. Then an approximation of f (b yn+1 , tn+1 ) at the time tn+1 is computed
bn+1 . This value is used to improve the accuracy of the numerical solution yn+1 .
using y
The method is given by

k1 = f (yn , tn ) (14.65)
k2 = f (yn + hk1 , tn + h) (14.66)
h
yn+1 = yn + (k1 + k2 ) (14.67)
2
To find the order of this method a Taylor series expansion around (yn , tn ) is used. The
Taylor series of k2 is

df h2 d2 f
k2 = f (yn , tn ) + h (yn , tn ) + (yn , tn ) + O(h3 ) (14.68)
dt 2 dt2
This gives the Taylor series

h2 df h3 d2 f
yn+1 = yn + hf (yn , tn ) + (yn , tn ) + (yn , tn ) + O(h4 ) (14.69)
2 dt 4 dt2
The first two terms coincide with the Taylor series expansion of the local solution
yL (tn ; tn+1 ), and the remaining terms are O(h3 ). Comparison with (14.23) leads to
the conclusion that the improved Euler’s method is of order p = 2.
524 CHAPTER 14. SIMULATION

To investigate stability of the method, we apply the method to the test equation

ẏ = λy

This results in

k1 = λyn
k2 = λ (1 + hλ) yn
à !
2
(hλ)
yn+1 = 1 + hλ + yn
2

which is stable whenever ¯ ¯

¯ (hλ)2 ¯¯
¯
¯1 + hλ + ¯≤1
¯ 2 ¯

On the real axis this corresponds to −2/h ≤ λ ≤ 0. The region of stability is shown in
Figure 14.14

Example 216 The first order system

ẏ = −αy 3 (14.70)

in the improved Euler method gives the algorithm

k1 = −αyn3 (14.71)
3
k2 = −α (yn + hk1 ) (14.72)
h
yn+1 = yn + (k1 + k2 ) (14.73)
2
Example 217 Consider the second order system

ẍ + cẋ + γ(x)x = 0 (14.74)

which is set in standard form by using y1 = x and y2 = ẋ. Then the differential equation
is written µ ¶ µ ¶
ẏ1 y2
= (14.75)
ẏ2 −γ(y1 )y1 − cy2
The improved Euler’s method gives the integration algorithm
µ ¶ µ ¶
k1,1 y2,n
= (14.76)
k1,2 −γ(y1,n )y1,n − cy2,n
µ ¶ µ ¶
k2,1 y2,n + hk1,2
= (14.77)
k2,2 −γ(y1,n + hk1,1 )(y1,n + hk1,1 ) − c (y2,n + hk1,2 )

h
y1,n+1 = y1,n + (k1,1 + k2,1 ) (14.78)
2
h
y2,n+1 = y2,n + (k1,2 + k2,2 ) (14.79)
2
14.3. EULER METHODS 525

14.3.3 The modified Euler method

The modified Euler method , also called the explicit midpoint rule, is derived in a similar
way ¡as the improved¢ Euler method. In the modified Euler method an approximation of
f at y(t + h2 ), t + h2 is used to find the solution. This approximation is computed using
Euler’s method to find an estimate of y(t + h2 ). The method is illustrated in Figure 14.9
and is given by
k1 = f (yn , tn ) (14.80)
h h
k2 = f (yn + k1 , tn + ) (14.81)
2 2
yn+1 = yn + hk2 (14.82)
A Taylor series expansion of k2 gives

k1 k2

y L (t n)
y n+1

k2

yn

tn t n+1
t

Figure 14.9: The modified Euler method

¡ h ¢2
h df d2 f
k2 = f (yn , tn ) + (yn , tn ) + 2
(yn , tn ) + O(h3 ) (14.83)
2 dt 2 dt2
which gives
h2 df h3 d2 f
yn+1 = yn + hf (yn , tn ) + (yn , tn ) + (yn , tn ) + O(h4 ) (14.84)
2 dt 8 dt2
The method is seen to be of order p = 2.
Application of the method to the test systems ẏ = λy gives
µ ¶
h
k2 = λ 1 + λ yn
2
à !
(hλ)2
yn+1 = 1 + hλ + yn
2

which leads to the same stability conditions as for the improved Euler method.
526 CHAPTER 14. SIMULATION

14.4 Explicit Runge-Kutta methods

14.4.1 Introduction
It was demonstrated above that Euler’s method, which is of order p = 1, can be modified
to a method of order p = 2 by computing yn+1 as a linear combination of f (yn , tn ) and
an approximation of f [y (tn + ch) , tn + ch] where 0 < c ≤ 1. This result can be extended
to higher order methods by computing more approximations of f over the interval, and
then compute yn+1 as a linear combination of these approximations. This is done in
the explicit Runge-Kutta methods. A Runge-Kutta method is said to have σ stages if σ
approximations, or stages, of the function derivative f is used.

14.4.2 Numerical scheme

An explicit Runge-Kutta method with σ stages for the system

ẏ = f (y, t) (14.85)

is given by
i−1
X
ki = f (yn + h aij kj , tn + ci h), i = 1, . . . , σ
j=1
σ
X
yn+1 = yn + h bj kj
j=1

The explicit Runge-Kutta method can be written out as

k1 = f (yn , tn ) (14.86)
k2 = f (yn + ha21 k1 , tn + c2 h) (14.87)
k3 = f (yn + h (a31 k1 + a32 k2 ) , tn + c3 h) (14.88)
..
. (14.89)
kσ = f (yn + h (aσ1 k1 + . . . + aσ,σ−1 kσ−1 ) , tn + cσ h) (14.90)
yn+1 = yn + h(b1 k1 + . . . + bσ kσ ) (14.91)

The equations for k1 , . . . , kσ are called the stage computations. The interpolation pa-
rameters ci , i ∈ {2, . . . , σ} are in the range 0 ≤ ci ≤ 1 and form an increasing sequence,
that is, 0 ≤ c1 ≤ . . . ≤ cσ ≤ 1. The weighting parameters at stage i are denoted
aij , i ∈ {2, . . . , σ}, j ∈ {1, . . . , i − 1}, and satisfy the normalization condition

i−1
X
aij = ci ≤ 1 (14.92)
j=1

P parameters bi of the solution yn+1 are required to satisfy the normalization

The weighting
condition σi=1 bi = 1. Each explicit Runge-Kutta method is described by its parameters
14.4. EXPLICIT RUNGE-KUTTA METHODS 527

aij , bi and ci , which can be arranged in a Butcher array of the form

0
c2 a21
c3 a31 a32
.. .. .. .. (14.93)
. . . .
cσ aσ1 aσ2 ... aσ,σ−1
b1 b2 ... bσ−1 bσ
Alternatively, the parameters can be expressed by the matrix A and the vectors b and
c defined by
     
0 0 ... 0 0 b1 0
 a21 0 ... 0 0   b2   c2 
     
 a31 a32 . . . 0 0   b3   
A= , b =   , c =  c3 
 .. .. . .. .
.. . 
..   .   . 
 . .  ..   .. 
aσ1 aσ2 . . . aσ,σ−1 0 bσ cσ
The Butcher array is then written
c A
bT
We note that the matrix A is singular for explicit Runge-Kutta methods.

14.4.3 Order conditions

The parameters of an explicit Runge-Kutta method of σ stages must satisfy certain
conditions to be of order p. Here, a derivation of the conditions for a method with σ = 2
stages to be of order p = 2 will be done. The Butcher array is
0
c2 a21
b1 b2
A Taylor series expansion of k2 gives
df
k2 = f (yn , tn ) + a21 h (yn , tn ) + O(h2 )
dt
The condition a21 = c2 from (14.92) then gives
df
yn+1 = yn + (b1 + b2 ) hf (yn , tn ) + b2 c2 h2 (yn , tn ) + O(h3 ) (14.94)
dt
and it is seen that the right hand side is equal to the Taylor series expansion of yn+1 for
terms up to h2 if the parameters satisfy
b1 + b2 = 1 (14.95)
1
b2 c2 = (14.96)
2
Example 218 The improved Euler method has b1 = b2 = 12 and c2 = 1, which satisfies
the conditions in (14.95) and (14.96). The modified Euler method has b1 = 0, b2 = 1 and
c2 = 12 , which also agrees with the conditions (14.95) and (14.96). The is in agreement
with the result that both of these methods have σ = 2 stages, and are of order p = 2.
528 CHAPTER 14. SIMULATION

In the same way 4 conditions can be found for σ = p = 3, while 8 conditions can be
found for σ = p = 4.
For higher order methods there are certain lower bounds for how many stages that
are needed (Hairer et al. 1993). For order 5 ≤ p ≤ 6, an explicit Runge-Kutta method
must have σ ≥ p + 1 stages. For order p = 7, an explicit Runge-Kutta method must have
σ ≥ p + 2 stages, while to achieve order p ≥ 8, a method with at least σ ≥ p + 3 stages.

14.4.4 Some explicit Runge-Kutta methods

The following explicit Runge-Kutta methods are of order p = σ. Euler’s method, which
is of order 1, has the Butcher array
0
1
The improved Euler method is an explicit Runge-Kutta method with array

0
1 1
1 1
2 2

The modified Euler method has the array

0
1 1
2 2
0 1

Heun’s method has the following array

0
1 1
3 3
2 2
3 0 3
1 3
4 0 4

The region of stability is shown in Figure 14.14.

The famous fourth order Runge-Kutta method RK4 is of order 4 and has the array

0
1 1
2 2
1 1
2 0 2
1 0 0 1
1 2 2 1
6 6 6 6

The region of stability is shown in Figure 14.14.

14.4.5 Case study: Pneumatic spring

Consider the pneumatic spring system in Figure 14.10. The cylinder has cross section
A = 0.01 m2 and a vertical center axis pointing upwards with coordinate x. The cylinder
14.4. EXPLICIT RUNGE-KUTTA METHODS 529

m g

p=k ρκ

Figure 14.10: Pneumatic spring with gravity acting on the position.

is filled with air and has a piston of mass m = 200 kg that compresses the air. The
density of the air inside the cylinder is
ma ma
ρ= = (14.97)
Va Ax
where ma is the mass and Va = Ax is the volume of the air. The air is assumed to be
isentropic which implies that the pressure inside the cylinder is
µ ¶κ ³ x ´κ
ρ 0
p = p0 = p0 (14.98)
ρ0 x
where κ = 1.4 and p0 = 2 · 105 N/m2 is the pressure corresponding to a piston position
x0 = 1 m, and the density ρ0 := ma / (Ax0 ) . The total force acting on the piston is
gravity and pressure forces:
³ x ´κ
0
F = −mg + Ap = −mg + Ap0 (14.99)
x
where g = 10 m/s2 . Inserting the numerical values we see that mg = Ap0 , which implies
that when x = x0 the force is F = 0 and the system is at an equilibrium at x = x0 . The
equation of motion can then be written
£ ¤
ẍ + g 1 − x−κ = 0 (14.100)

The standard form is ẏ = f (y) is obtained by setting

µ ¶ µ ¶
x v
y= , f= (14.101)
v −g [1 − x−κ ]
where v = ẋ is the velocity of the piston. We see that the system has an equilibrium at
x = 1, v = 0, where ẍ = 0.
Linearization around x∗ = 1 gives

∆ẏ = J∆y (14.102)

530 CHAPTER 14. SIMULATION

Euler's method

2.5

2
Position

1.5

RK4
0.5

0
0 1 2 3 4 5 6 7 8 9 10
Time

Figure 14.11: Position of piston integrated with Euler’s method with h = 0.005 s and
with RK4 with h = 0.1 s.

where µ ¶ µ ¶
x−1 0 1
∆y = , J= ∗ −(κ+1) (14.103)
v −gκ (x ) 0
The eigenvalues of the linearization are found to be
q
λ1,2 = ±jω 0 , ω 0 = gκ (x∗ )−(κ+1) (14.104)

Numerical values are

x∗ 0.5 1 2
(14.105)
ω0 4.3 3.7 3.3
The total energy E is the sum of the internal energy U = pV /(κ − 1), the gravity
potential mgx, and the kinetic energy 12 mv 2 :

1 1
E= p0 Ax−(κ−1) + mgx + mv 2 (14.106)
κ−1 2
The total energy has its minimum value at the equilibrium state were the energy is
1
Emin = p0 A + mg = 7000 J (14.107)
κ−1
The system was simulated with Euler’s method with time step h = 0.005, and with the
fourth order RK4 method with time step h = 0.1. The result is shown in Figure 14.11.The
solution computed with Euler’s method was unstable even with the very short time step
of 0.005, while the solution with RK4 was stable with a time step that was 20 times larger
than for the Euler solution. To check the accuracy of the solutions the total energy was
computed for the numerical solutions. For the exact solution the total energy will be
constant as there is no energy loss terms in the equation of motion. The solution from
14.4. EXPLICIT RUNGE-KUTTA METHODS 531

9000

8800

8600

8400
Total energy

Euler's method
8200

8000

7800

RK4

7600
0 1 2 3 4 5 6 7 8 9 10
Time

Figure 14.12: Total energy of solutions computed with Euler’s method and RK4. It is
seen that Euler’s method increases the energy in the system, while RK4 gives a slight
decrease in energy.

Euler’s method gave a steady increase in energy which is not in agreement with the
physics of the system is it has no energy source. The RK4 solution gave a slight decrease
in energy, which means that the RK4 introduced some damping in the system. The
results are shown in Figure 14.12.
The system has eigenvalues ±jω 0 on the imaginary axis, and the stability limit for
RK4 is then h = 2.83/ω 0 , which can be seen from Figure 14.14. As the largest eigenvalue
occurs for ω 0 = 4.3 this indicates that the stability limit would be hmin = 0.65 s. In
simulations it turned out that a slightly smaller value, h = 0.52 s was the stability limit
for this trajectory. This is demonstrated in Figure 14.13. The difference between the
theoretical value and the value found in simulations should be due to the system being
nonlinear.

14.4.6 Stability function

A general formula for the stability function of an explicit Runge-Kutta method in terms
of c, A and b is found as follows: Application of a general explicit Runge-Kutta method
to the linear time-invariant test system

ẏ = λy

gives

k1 = λyn
..
.
kσ = λ [yn + h (aσ1 k1 + . . . + aσ,σ−1 kσ−1 )]
yn+1 = yn + h (b1 k1 + . . . + bσ kσ )
532 CHAPTER 14. SIMULATION

2.5
h=0.521

2
Position

1.5

0.5 h=0.5
h=0.1

0
0 1 2 3 4 5 6 7 8 9 10
Time

Figure 14.13: Simulation of peumatic spring usning RK4 for three different step lengths.

T T
In vector notation with κ = (k1 , k2, . . . , kσ ) and 1 = (1, 1, . . . , 1) this can be written

κ = λ (1yn + hAκ) (14.108)

yn+1 = yn + hbT κ (14.109)

Here κ can be solved from (14.108) and inserted into (14.109), which gives

R(hλ) = 1 + λhbT (I − hλA)−1 1 (14.110)

Alternatively, the system (14.108, 14.109) can be written

µ ¶µ ¶ µ ¶
I − hλA 0 κ λ1
= yn
−hbT 1 yn+1 1
From Cramer’s rule it is seen that the stability function can be written
h ³ ´i
det I − λh A − 1bT
R(hλ) = (14.111)
det (I − λhA)
This formula has the advantage that it clearly shows how the numerator and denominator
depend on hλ, A and b. For an explicit Runge-Kutta method the A matrix have nonzero
elements only below the diagonal, and it follows that det (I − λhA) = 1. Using (14.111)
we find:

For an explicit Runge-Kutta method the stability function can be written

h ³ ´i
RE (hλ) = det I − λh A − 1bT (14.112)

This expression shows that for explicit Runge-Kutta methods

14.4. EXPLICIT RUNGE-KUTTA METHODS 533

1. |RE (hλ)| will tend to infinity when |λ| goes to infinity

2. RE (hλ) is a polynomial in hλ of order less than or equal to σ.

Example 219 Consider the improved Euler method where

µ ¶ µ ¶
0 0 1 1
A= , b=
1 0 2 1

Then µ ¶
1 + λh2
λh
2
(λh)2
RE (hλ) = det = 1 + λh + (14.113)
− λh
2 1 + λh
2 2
We see that RE (hλ) is a polynomial in hλ of order 2 which is equal to the number of
stages.

Next we will comment on explicit Runge-Kutta methods where the number of stages
equals the order of the method. The local solution yL (tn ; tn+1 ) starting from yL (tn ; tn ) =
yn is given by
yL (tn ; tn+1 ) = eλh yn
A Taylor series expansion of the local solution is therefore
" #
2 3
(hλ) (hλ)
yL (tn ; tn+1 ) = 1 + hλ + + + . . . yn (14.114)
2! 3!

Therefore, if an explicit Runge-Kutta method of order p is used, then the numerical

solution yn+1 for a linear test system with will have the Taylor series expansion
" #
(hλ)2 (hλ)p p+1
yn+1 = 1 + hλ + + ... + + O(h ) yn (14.115)
2! p!

It follows that the stability function for a explicit method of order p satisfies
2 p
(hλ) (hλ)
RE (λh) = 1 + hλ + + ... + + O(hp+1 ) (14.116)
2! p!
The stability function of an explicit Runge-Kutta method with σ stages is a polynomial
in λh of degree less than or equal to the number of stages σ. If the method is of σ = p ≤ 4
stages the stability function must have exactly p terms, and this is only possible if
2 p
(hλ) (hλ)
RE (λh) = 1 + hλ + + ... + when p = σ
2! p!

Example 220 The improved Euler method has stability function

(λh)2
R(λh) = 1 + λh +
2
which coincides with the Taylor series expansion with two terms. This agrees with the
fact that the method has 2 stages and is of order 2.
534 CHAPTER 14. SIMULATION

Figure 14.14: Regions of stability in s = hλ for the test system ẏ = λy for the explicit
Runge-Kutta methods. ERK(1): Euler’s method, ERK(2): The modified and the im-
proved Euler method, ERK(3): Heun’s third order method, and ERK(4): The fourth
order Rung-Kutta method RK4.

14.4.7 FSAL methods

We will here take a closer look at explicit Runge-Kutta methods of the FSAL type.

An explicit Runge-Kutta method is said to be an FSAL method if

kσ = f (yn+1 , tn+1 ) (14.117)

From the definition we see that in an FSAL method gives some savings in computa-
tions as
knσ = kn+1
1 (14.118)
where knσ denotes the last stage in the calculation of yn+1 , and kn+11 denotes the first
stage in the computation of yn+2 . This is the reason for calling such methods First Same
As Last, which is abbreviated to FSAL. In an FSAL method the weighting vector b is
equal to the last row in the stage matrix A.

14.5 Implicit Runge-Kutta methods

14.5.1 Stiff systems
When a system ẏ = f (y, t) is integrated with an explicit Runge-Kutta method the time
step h cannot be selected so that h |λmax | is significantly larger than unity, where λmax
14.5. IMPLICIT RUNGE-KUTTA METHODS 535

is the largest eigenvalue of the Jacobian J = ∂f (y,t)/∂y. As an example of this, h |λmax |

must be less than 2 for Euler’s method, and it is seen from Figure 14.14 that approxi-
mately the same hold for e.g. RK4. Some systems have a large spread in eigenvalues, and
as the time-step of an explicit method must be selected to ensure stability, it follows that
very many time steps are required to compute the dynamics corresponding to the small
eigenvalues. This gives problems with simulation time and accuracy. Systems that have
a large spread in eigenvalues of the Jacobian are referred to as stiff systems. Stiff systems
are difficult to solve with explicit methods. This has lead to a recent and more pragmatic
definition of stiff systems as systems that are difficult to solve with explicit methods. Ex-
amples of stiff systems are the restricted three-body problem in Section 14.1.3, and the
mass balances in Section 14.1.4. We will see that stiff problems can be solved efficiently
by implicit Runge-Kutta method, that are presented in the following.

14.5.2 Implicit Runge-Kutta methods

An implicit Runge-Kutta method with σ stages for the system

ẏ = f (y, t) (14.119)

is given by

k1 = f (yn + h (a11 k1 + . . . + a1σ kσ ) , tn + c1 h) (14.120)

..
. (14.121)
kσ = f (yn + h (aσ1 k1 + . . . + aσσ kσ ) , tn + cσ h) (14.122)
yn+1 = yn + h(b1 k1 + . . . + bσ kσ ) (14.123)

As for explicit Runge-Kutta methods, the interpolation parameters ci , i ∈ {1, . . . , σ}

Pσ in the range 0 ≤ ci ≤ 1. The weighting factors satisfy the normalization
are Pσ equation
b
i=1 i = 1, and usually the weighting factors at each stage satisfy a
j=1 ij = ci .

14.5.3 Implicit Euler method

The implicit Euler method is an implicit Runge-Kutta method with one stage described
by the following array:
1 1
1
This gives

k1 = f (yn + hk1 , tn+1 )

yn+1 = yn + hk1

This method is said to be a Radau IIA method.

The stability function is found by applying the method to the linear test system
ẏ = λy. Then k1 = λyn + λhk1 can be solved for k1 , and inserting this into the equation
for yn+1 we get
hλ 1
yn+1 = yn + yn = yn (14.124)
1 − hλ 1 − hλ
536 CHAPTER 14. SIMULATION

The stability function is seen to be

1
R (hλ) = (14.125)
1 − hλ
The region in the complex plane where the method is stable is given by |hλ − 1| ≥ 1 and
shown as the shaded region in Figure 14.15.

Figure 14.15: The shaded area shows where the implicit Euler method is stable as a
function of the complex variable s = λh.

14.5.4 Trapezoidal rule

Consider the implicit Runge-Kutta method

k1 = f (yn , tn ) (14.126)
· ¸
h
k2 = f yn + (k1 + k2 ) , tn + h (14.127)
2
h
yn+1 = yn + (k1 + k2 ) (14.128)
2
which is a Lobatto IIIA method of order 2. The Butcher array is

0 0 0
1 1
1 2 2
1 1
2 2

A closer look reveals that

k2 = f (yn+1 , tn+1 ) (14.129)
T
as the last row in A is equal to b . This implies that the expression for yn+1 can be
rewritten in the form
h
yn+1 = yn + [f (yn , tn ) + f (yn+1 , tn+1 )] (14.130)
2
14.5. IMPLICIT RUNGE-KUTTA METHODS 537

which is known as the trapezoidal rule.

The stability function is found from (14.130) which for the test equation gives
hλ
yn+1 = yn + (yn + yn+1 ) (14.131)
2
and it follows that
hλ
1+ 2
R(λh) = hλ
(14.132)
1− 2
We see that ¡ £ ¤¢2 ¡ £ hλ ¤¢2
2 1 + Re hλ2 + Im 2
|R(λh)| = ¡ £ hλ ¤¢2 ¡ £ hλ ¤¢2 (14.133)
1 − Re 2 + Im 2
and it follows that |R(λh)| ≤ 1 and the method is stable for all λ that have negative real
part. The area in the complex plane where the trapezoidal rule is stable is therefore the
left half plane as shown in Figure 14.16.

Figure 14.16: The shaded area shows where the trapezoidal rule is stable as a function
of the complex variable s = λh.

14.5.5 Implicit midpoint rule

We consider the implicit Runge-Kutta method
µ ¶
h h
k1 = f yn + k1 , tn +
2 2
yn+1 = yn + hk1

with Butcher array

1 1
2 2
1
This method is a Gauss method of order 2. This implicit Runge-Kutta method can
be reformulated as a scheme known as the implicit mid-point rule. To do this we first
538 CHAPTER 14. SIMULATION

note that the equation for yn+1 gives hk1 = yn+1 − yn . Inserting this into the stage
computation gives
· ¸
1 h
yn+1 − yn = hf yn + (yn+1 − yn ) , tn +
2 2
which is simplifies to the following scheme
µ ¶
yn + yn+1 h
yn+1 = yn + hf , tn + (14.134)
2 2
which is called the implicit mid-point rule.
From (14.134) we find that the stability function for the implicit mid-point rule is
hλ
1+ 2
R(λh) = hλ
(14.135)
1− 2

which is identical to the stability function for the trapezoidal rule. Therefore the stability
properties of the two methods are the same for linear time-invariant systems. However it
turns out that for nonlinear systems the implicit mid-point rule has much better stability
properties, as will be seen in the following sections.

14.5.6 The theta method

Consider the implicit Runge-Kutta method
k1 = f (yn , tn ) (14.136)
k2 = f [yn + h [θk1 + (1 − θ) k2 ] , tn + h] (14.137)
yn+1 = yn + h [θk1 + (1 − θ) k2 ] (14.138)

where θ ∈ [0, 1] is a parameter. The Butcher array is

0 0 0
1 θ 1−θ
θ 1−θ

As for the trapezoidal rule, the second stage can be written k2 = f (yn+1 , tn+1 ) . Then
the expression for yn+1 can be rewritten in the form

yn+1 = yn + h [θf (yn , tn ) + (1 − θ) f (yn+1 , tn+1 )]

which is known as the theta method . We see that for θ = 1 the method is Euler’s method,
for θ = 12 the method is the trapezoidal rule, and for θ = 0 it is the implicit Euler method.
The stability function is
1 + hλθ
R(hλ) = (14.139)
1 − hλ (1 − θ)

14.5.7 Stability function

The application of an implicit Runge-Kutta method to a linear test system gives
κ = λ (1yn + hAκ) (14.140)
yn+1 = yn + hbT κ (14.141)
14.5. IMPLICIT RUNGE-KUTTA METHODS 539

as for explicit methods, where the notation is defined in connection with equations
(14.108) and (14.109). From (14.110) and (14.111) we may conclude as follows:

The stability function for an implicit Runge-Kutta method is given by the two alternative
expressions h i
−1
R(hλ) = 1 + λhbT (I − hλA) 1 (14.142)
h ³ ´i
det I − λh A − 1bT
R(hλ) = (14.143)
det (I − λhA)

From (14.143) it is seen that the stability function for an implicit Runge-Kutta method
is a rational expression in s = λh. We will see in the following that certain properties of
the implicit methods will depend on the degree of the numerator and denominator poly-
nomials in the stability function. In particular it will be shown that the most important
implicit methods have stability functions R(s) given by Padé approximations of es , and
that interesting conclusions can be drawn from this fact. However, first we will present
some implicit methods and a case study.

14.5.8 Some implicit Runge-Kutta methods

Gauss order 2, which is the implicit mid-point rule

1 1
2 2
1

Gauss order 4, which is the Hammer and Hollingsworth method of order 4

√ √
1 3 1 1 3
2 − −
√6 4√ 4 6
1 3 1 3 1
2 + 6 4 + 6 4
1 1
2 2

Radau IA, order 3

1
0 4 − 14
2 1 5
3 4 12
1 3
4 4

Radau IIA, order 3

1 5 1
3 12 − 12
3 1
1 4 4
3 1
4 4

Lobatto IIIA, order 2, which is the trapezoidal rule

0 0 0
1 1
1 2 2
1 1
2 2
540 CHAPTER 14. SIMULATION

Lobatto IIIB order 2

1
0 2 0
1
1 2 0
1 1
2 2

Lobatto IIIC order 2

1
0 2 − 12
1 1
1 2 2
1 1
2 2

14.5.9 Case study: Pneumatic spring revisited

The pneumatic spring from Section 14.4.5 was simulated with implicit Runge-Kutta
methods with a time step h = 0.5 s, which was found to be the stability limit for this
system when the explicit RK4 was used (Figure 14.13). The methods that were used was
the Gauss method of order 2 (the implicit mid-point rule), Radau IIA of order 3, Lobatto
IIIC of order 2 and the implicit Euler method. The results are shown in Figure 14.17.
The Gauss method gave no damping, while the Radau method gave some damping, the
Lobatto method gave more damping than the Radau method, and the implicit Euler
method gave the most damping. This is clearly seen in Figure 14.18 where the total
energy corresponding to the numerical solutions is plotted. It is seen that the energy of
the solution from the Gauss method fluctuates around the correct value, while the other
methods introduce what can be termed numerical dissipation of energy. In particular
it is seen that the implicit Euler method gave a solution where the total energy quickly
converged to the energy of the equilibrium state.

2.5

Radau IIA, 3 Gauss, 2

2
Position

1.5

0.5

Implicit Euler Lobatto IIIC, 2

0
0 1 2 3 4 5 6 7 8 9 10
Time

Figure 14.17: Position of piston computed with the implicit Runge-Kutta methods Gauss
of order 2, Radau IIA of order 3, Lobatto IIIC of order 2 and the implicit Euler method.

To study how Runge-Kutta methods work for stiff oscillatory systems the pneumatic
spring system was modified to include a mechanical resonance in the mass as shown
14.5. IMPLICIT RUNGE-KUTTA METHODS 541

7900

Gauss, 2
7800

7700
Radau IIA, 3

7600

Lobatto IIIC, 2
7500
Total energy

7400

7300 Implicit Euler

7200

7100

7000
Equilibrium energy level

6900
0 1 2 3 4 5 6 7 8 9 10
Time

Figure 14.18: Total energy for the pneumatic system when the solution is computed with
the implicit Runge-Kutta methods. The energy oscillates around the correct value for
the Gauss of order 2, while the energy is numerically dissipated with the methods Radau
IIA of order 3, Lobatto IIIC of order 2 and the implicit Euler method.

in Figure 14.19. This was done by splitting the mass m = 200 kg into two masses
m1 = m2 = 100 kg, which are connected by a spring with stiffness K = m1 ω 22 /2 with
axis along the vertical axis. The position of m1 is denoted x1 and the position of m2 is
denoted x2 . The coordinate x2 is given an offset so that x1 = x2 when the spring force
is zero. The equilibrium energy of this system with two degrees of freedom is the same
as for the one degree of freedom system studied above. The equations of motion are
µ ¶
m −κ ω2
ẍ1 + g 1 − x1 + 2 (x1 − x2 ) = 0 (14.144)
m1 2
ω2
ẍ2 + g + 2 (x2 − x1 ) = 0 (14.145)
2

where ω 2 = 1000 rad/s, m1 = m2 = 100 kg and m = m1 + m2 = 200 kg. The standard

form is ẏ = f (y) is obtained by setting

   
x1 v1
 v1   m −κ
−g(1 − m x ) −
ω 22 
y=  
f = 2 (x1 − x2 ) 
 x2  ,
1  (14.146)
 v2 
v2 ω2
−g − 22 (x2 − x1 )

where vi = ẋi is the velocity of the piston. We see that the system has an equilibrium
at x∗1 = 1, v1∗ = 0, x∗2 = 1 − m ∗ ∗
K where ẍi = 0. Linearization around x1 , x2 gives
2

∆ẏ = J∆y (14.147)

542 CHAPTER 14. SIMULATION

x
m2

K
g
m1

p=k ρκ

Figure 14.19: Pneumatic spring with mechanical resonance in load. The x coordinate of
mass m2 is given an offset so that x1 = x2 when the spring is unloaded.

where
   
x1 − x∗1 0 1 0 0
   m
−g m κ (x∗ −(κ−1)
) −
ω 22
0
ω 22
0 
v1  
∆y =  
 x2 − x∗2  , J = 1 2 2
 (14.148)
 0 0 0 1 
v2 ω22 ω2
2 0 − 22 0

The eigenvalues of the linearization around the equilibrium are found to be

λ1,2 = ±jω 1 , ω 1 = 3.7 rad/s (14.149)

λ3,4 = ±jω 2 , ω 2 = 1000 rad/s (14.150)

This means that the system has the eigenvalues at ±j3.7 as the pneumatic spring, and
in addition a new set of eigenvalues have been introduced at ±j1000.
The system was simulated with RK4 with a time step h = 0.0005 s, where the step
size was selected so that hω 2 = 0.5. The simulation result as shown in Figure 14.20
is fairly accurate, but it is seen from Figure 14.21 that the high frequency motion is
damped out even though there is no damping in the system. The phenomenon is clearly
seen from the plot of the total energy in Figure 14.22 where it is seen that the energy
converges to the energy level of the slow dynamics corresponding to λ1,2 = ±j3.7, which
is the energy that is obtained if x1 = x2 .
The system was then simulated with the implicit methods Gauss of order 2 and
Lobatto IIIC of order 2 with a time step h = 0.05. This gives Nyquist frequency ω N =
π/0.05 = 62.8 rad/s, so that the resonance at 1000 rad/s is well above the Nyquist
frequency. The solution of the Gauss method, which is shown in Figure 14.23 gave no
damping, but the aliasing effect moved the energy of the fast dynamics associated with
the eigenvalues ±j1000 to oscillations with frequency lower than the Nyquist frequency.
This gave a beat phenomenon which is clearly seen in Figure 14.24, while it is seen from
Figure 14.25 that the total energy is constant for the Gauss solution. The Lobatto IIIC
solution gave quick damping of the fast dynamics, and a slight damping of the slow
dynamics. It is seen from Figure 14.25 that the energy associated with the fast dynamics
is dissipated in one step, and the total energy remains on the level of the energy associated
with the slow dynamics.
14.5. IMPLICIT RUNGE-KUTTA METHODS 543

1.25

1.2

1.15

Position, RK4 solution 1.1

1.05

0.95

0.9

0.85

0.8
0 1 2 3 4 5 6 7 8 9 10
Time

Figure 14.20: Position of the two masses computed with RK4 with time step h = 0.0005
s.

0.01

0.008

0.006

0.004
Position difference, RK4 solution

0.002

-0.002

-0.004

-0.006

-0.008

-0.01
0 1 2 3 4 5 6 7 8 9 10
Time

Figure 14.21: Offset equilibrium for the spring between masses one and two computed
with RK4 with h = 0.0005 s. The oscillation is seen to be lightly damped by the
integration method.
544 CHAPTER 14. SIMULATION

10000

Total energy of exact soultion

9500

9000

8500
Total energy of RK4 solution

Total energy
8000

7500
Total energy of slow dynamics

7000

Equilibrium energy
6500

6000
0 1 2 3 4 5 6 7 8 9 10
Time

Figure 14.22: Total energy corresponding to the numerical solution computed with RK4
with h = 0.0005 s. It is seen that the energy related to the fast dynamics is slowly
damped out.

14.6 Stability of Runge-Kutta methods

14.6.1 Aliasing
We consider the test equation ẏ = λy, and write the eigenvalue in the form

λ = σ + jω (14.151)

where σ is the real part and jω is the imaginary part. It is assumed that ω < π/h, that
is, ω is assumed to be less that the Nyquist frequency π/h. The local solution of the test
system is
yL (tn ; tn+1 ) = eλh yn
Consider a system ẏ = µy, which has the local solution

yL (tn ; tn+1 ) = eµh yn (14.152)

The two systems will give the same local solutions at tn+1 whenever eλh = eµh which is
implied by ³
π π´
µ = λ + j2k = σ + j ω + 2k , k = 0, ±1, ±2, . . . (14.153)
h h
If
π
µ = λ + 2kj , k = ±1, ±2, . . . (14.154)
h
then the system ẏ = µy where Im (λ) > π/h will have the same solution as the system
ẏ = λy where Im (λ) < π/h. This phenomenon is called aliasing.

14.6.2 A-stability, L-stability

The test system ẏ = λy is stable when Re λ ≤ 0. We consider an integration method
which gives yn+1 = R(λh)yn when applied to the test system. It would seem to be a
14.6. STABILITY OF RUNGE-KUTTA METHODS 545

1.25
Lobatto IIIC, 2 Gauss, 2

1.2

1.15

1.1

1.05
Position

0.95

0.9

0.85

0.8
0 1 2 3 4 5 6 7 8 9 10
Time

Figure 14.23: Position of the two masses computed with a Gauss method of order 2 and
a Lobatto IIIC method of order 3. The time step was h = 0.05 with both methods.

0.015

Gauss, 2

0.01

0.005
Positon difference

-0.005

-0.01
Lobatto IIIC, 2

-0.015
0 1 2 3 4 5 6 7 8 9 10
Time

Figure 14.24: Offset in position between the two masses calculated with Gauss order 2
and Lobatto IIIC order 2 with h = 0.05 s. The Gauss method gave no damping, but
the energy of the fast dynamics was shifted to frequencies below the Nyquist frequency
ω N = 62.8 rad/s. The Lobatto method damped out the fast dynamics in one step.
546 CHAPTER 14. SIMULATION

10000

Gauss, 2
9500

9000

8500

Total energy
8000 Lobatto IIIC, 2

7500
Total energy of slow dynamics

7000
Equilibrium energy

6500

6000
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time

Figure 14.25: Total energy corresponding to the numerical solution of Gauss order 2 and
Lobatto IIIC order 2. The Gauss method gave a constant total energy in agreement with
the exact solution, while the Lobatto method damped out the energy associated witht
the fast dynamics.

useful property for an integration method if the method was stable for all stable test
systems. This property is called A-stability.

An method is A-stable if |R(λh)| ≤ 1 for all Re λ ≤ 0.

Integration methods that are A-stable will be stable also for systems with very fast
dynamics which in this context are systems that have dynamics which is significantly
faster that the time step h of the integration method. In particular, aliasing can be
problematic for A-stable methods, as high frequency oscillations will appear in the com-
puted solution as an oscillation with frequency below the Nyquist frequency π/h. As the
integration cannot give an accurate computation of such fast dynamics, it may be useful
that the method damp out the fast dynamics. This is the case for L-stable integration
methods.

A method is L-stable if it is A-stable and, in addition, if |R(jωh)| → 0 when ω → ∞ for

all systems ẏ = λy where λ = jω.

Example 221 We note that explicit Runge-Kutta methods have stability functions
h ³ ´i
RE (λh) = det I − λh A − 1bT (14.155)

It is clear that |RE (λh)| → ∞ whenever |λ| → ∞, and it follows that an explicit Runge-
Kutta method cannot be A-stable.

14.6.3 Stiffly accurate methods

We will here take a closer look at implicit Runge-Kutta methods that are stiffly accurate.
The stability function of an implicit Runge-Kutta method is found in the same way as
14.6. STABILITY OF RUNGE-KUTTA METHODS 547

for explicit Runge-Kutta methods. Therefore, the expression

h i
R(hλ) = 1 + λhbT (I − hλA)−1 1

(14.110) can be used also for implicit Runge-Kutta methods. Suppose that A is non-
singular, and consider the case where λh tends to infinity. Then if the limit R(∞) :=
lims→∞ R(s) exists, it is given by
h i
−1
R(∞) = lim 1 + sbT (I − sA) 1
s→∞
h i
−1
= lim 1 − sbT (sA) 1
s→∞
= 1 − bT A−1 1 (14.156)

Moreover, it is noted that for an implicit Runge-Kutta method where

kσ = f (yn+1 , tn+1 ) (14.157)

then the weighting vector b is equal to the last row in the stage matrix A. This gives

b = AT eσ (14.158)

where eσ = (0, 0, . . . , 1)T is a σ-dimensional unit vector. Insertion of (14.158) into

(14.156) gives
−1
R(∞) = 1 − λheTσ A(λhA) 1 =1 − eTσ 1 =0 (14.159)

An implicit Runge-Kutta method is said to be stiffly accurate if the stage matrix A is

nonsingular and in addition b = AT eσ .

From (14.159) we find that

An A-stable Runge-Kutta method that is stiffly accurate will be L-stable.

Moreover, from (14.159) we may conclude that a stiffly accurate method will damp
out dynamics corresponding to eigenvalues λi that are large in the sense that |λi h| are
much larger that unity. Consider the case where a stiffly accurate method is applied to a
stiff system, and the time step h is selected in the dynamic range of the slow dynamics.
Then the fast dynamics will have eigenvalues so that |λi h| À 1. The fast dynamics will
therefore be damped out, and the solution will mainly correspond to the slow dynamics.
In particular, if there is an eigenvalue λj so that |λj h| → ∞, then the dynamics associated
with this eigenvalue will be damped to zero.

Example 222 It is clear that a Gauss method cannot be stiffly accurate as for these
methods |R(jω)| = 1 for all ω.

Example 223 The Trapezoidal rule is a Lobatto IIIA method with

µ ¶ µ 1 ¶
0 0 2
A= 1 1 , b = 1
2 2 2
548 CHAPTER 14. SIMULATION

The last row in A is equal to bT , but the matrix A is singular. Thus the method is not
stiffly accurate. This agrees with the fact that the stability function is
R(s) = 1 + sbT (I − sA)−1 1
" Ã !µ ¶#
s ¡ ¢ 1 0 1
= 1+ 1 1 s
1
2 2
1− 2s 1− 2s 1
s
1+ 2
= s
1− 2

which means that

|R(jω)| = 1
for all ω.

14.6.4 Padé approximations

The stability function R(s) of a Runge-Kutta method, which may be implicit or explicit,
is given by a rational expression in s, which is seen from (14.143) with s = λh. Properties
like A-stability and L-stability depend on the stability function. Also the region of
stability for a method is found from the stability function. Instead of checking such
properties for each method, it is possible to have more general results. This is done by
introducing a classification of Runge-Kutta methods based on a special characterization of
the stability functions. This will be done in the following using the Padé approximations
of the exponential function es .
First it is noted that the local solution of the test equation ẏ = λy over the time step
from tn to tn+1 is
yL (tn ; tn+1 ) = eλh yn (14.160)
while the numerical solution is
yn+1 = R(λh)yn (14.161)
From these two equations it is seen that the accuracy of the numerical solution yn will
depend on to what extent the stability function approximates the exponential function,
that is, the accuracy of the numerical solution yn depends on the difference
es − R(s) (14.162)
between the exponential function in the exact solution (14.160), and stability the numer-
ical solution (14.161). Here we have used s = λh to simplify the notation. An explicit
Runge-Kutta method with σ stages approximates the exponential function es by the
polynomial approximation
R(s) = 1 + β 1 s + . . . + β σ sσ
In the special case that the method is of order p = σ ≤ 4 the stability function is given
by the Taylor series expansion of es given by
sp
R(s) = 1 + s + . . . +
p!
An implicit Runge-Kutta method of σ stages approximates es by the rational approxi-
mation
1 + β 1 s + . . . + β k sk
R(s) =
1 + γ 1 s + . . . + γ m sm
14.6. STABILITY OF RUNGE-KUTTA METHODS 549

k
0 1 2 3
s2 s2 s3
0 1 1+s 1+s+ 2! 1+s+ 2! + 3!
1 1+ 12 s 1+ 23 s+ 16 s2 1+ 34 s+ 14 s2 + 24
1 3
s
m 1 1−s 1− 12 s 1− 13 s 1− 14 s
1 2 3 2 1 3
1 1+ 13 s 1+ 12 s+ 12 s 1+ 35 s+ 20 s + 60 s
2 2
1− 23 s+ 16 s2 1− 12 s+ 121 2
s 1− 25 s+ 201 2
s
1−s+ s2!
1 2 2 3
1 1+ 14 s 1+ 25 s+ 20 s 1+ s2 + s10 + 120
s
3 2 3
1− 34 s+ 14 s2 − 24
1 3
s 1− 35 s+ 20
3 2
s − 601 3
s 2 s3
1−s+ s2! − s3! 1− s2 + s10 − 120

k
Table 14.2: The Padé approximations Pm (s) for m, n = 0, 1, 2, 3

k
0 1 2 3
0 Euler’s method Mod. Euler Heun’s, 3
m 1 Radau, 1 Gauss, 2, Trapez.
2 Lobatto IIIC, 2 Radau, 3 Gauss, 4
3 Lobatto IIIC, 4 Radau, 5 Gauss, 6

k
Table 14.3: Methods that have Pade approximations Pm (s) as stability functions.

where m ≤ σ and k ≤ σ.
One particular rational approximation of the exponential function is the Padé ap-
proximation (Golub and van Loan 1989).

k
The Padé approximation Pm (s) of the exponential function es is a rational function of s
with a numerator of degree k and a denominator of degree m. The Padé approximation
k
Pm (s) of es is the rational approximation of es which has the highest order in s when
the numerator is of order k and the denominator is of order m.

k
The Padé approximation Pm (s) is given by

k Qmk (s)
Pm (s) = (14.163)
Qkm (−s)

where
Xk
k! (m + k − i)! si
Qmk (s) = 1 + (14.164)
i=1
(k − i)! (m + k)! i!

The error in the approximation is given by

k
(−1) m!k! sk+m+1
es − Pm
k
(s) = + O(sk+m+2 )
(k + m)! (k + m + 1)!
k
which shows that the approximation is of order k + m. The Padé approximations Pm (s)
for m, k = 0, 1, 2, 3 are shown in Table 14.2.
550 CHAPTER 14. SIMULATION

Example 224 Long division of P11 (s) gives

1 + 12 s s2 s3
= 1 + s + + + O(s4 )
1 − 12 s 2 4

s3
where the error is 6 + O(s4 ).

Example 225 We note that an explicit Runge-Kutta method with p = σ ≤ 4 have

stability functions RE (s) = P0p (s)

14.6.5 Stability for Padé approximations

An important result related to A-stability of methods is the following:
k
Pm (s) ≤ 1, when Re [s] ≤ 0 for k ≤ m ≤ k + 2 (14.165)

This is derived using order stars in (Hairer and Wanner 1996). Moreover, in relation to
k
L-stability it is interesting to study Pm (jω). From Table 14.2 it is seen that the Padé
approximations where the degree of the numerator polynomial equals the denominator
polynomials satisfy
m
|Pm (jω)| = 1, for all ω (14.166)
whereas for Padé approximations where the degree of the numerator polynomial is less
than the degree of the denominator polynomials we have
k
|Pm (jω)| → 0, when ω → ∞ for m > k (14.167)

Combining these results we arrive at the following result:

A one-step method with stability function

m
R(s) = Pm (s) (14.168)

is A stable. A one-step method with stability function

k
R(s) = Pm (s) where m = k + 1 or m = k + 2 (14.169)

is L-stable.

Example 226 The Gauss methods, including the implicit mid-point rule, the Lobatto
IIIA, including the trapezoidal rule, and the Lobatto IIIB have stability functions
m
R(s) = Pm (s) (14.170)

which implies that the methods are A-stable.

Example 227 Radau methods and Lobatto IIIC methods have stability functions
k
R(s) = Pm (s), k <m≤k+2 (14.171)

and this implies that the methods are L-stable.

14.6. STABILITY OF RUNGE-KUTTA METHODS 551

Deflection of free end, Gauss2

5

-1

-2

-3

-4

-5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time

Figure 14.26: Simulation of vibrations in an Euler-Bernoulli beam with a Gauss method.

There is an initial excitation that is switched off after 1 s. After this the vibrational
energy of the system is a constant. The simulation reflects this.

14.6.6 Example: Mechanical vibrations

An Euler-Bernoulli beam was simulated using finite elements spatial discretization. The
beam was modelled as having no damping, and this means that the vibrational energy
will be constant when the beam is not excited from external forces. The vibration of
the beam will occur at resonance frequencies ω i , i = 1, 2, . . .. The number of resonant
frequencies in the model used for simulation depends on the way the model is imple-
mented. A simulation was done with a discretization using 10 finite elements leading to
10 resonance frequencies. The resulting system is stiff as the fastest resonances have very
large eigenvalues λi = jω i h, so that |λi | = ω i h À 1. The system was simulated with a
Gauss method and a Lobatto IIIC method (Kristiansen 2000). The results are shown in
Figures 14.26 and 14.27.

14.6.7 Frequency response

The Runge-Kutta methods have stability functions

R(s) = 1 + sbT (I − sA)−1 1

where R(s) appear in yn+1 = R(λh)yn when the method is applied to the test equation
ẏ = λy. To study the performance of Runge-Kutta methods it is of interest to plot |R(s)|
as a function of the complex variable s. One approach to this is to plot the order stars of
a method (Hairer and Wanner 1996), which are contour plots of |R(s)/es in the complex
plane. We will follow a different approach in the following where we plot the magnitude of
R(s) for the imaginary axis s = jω, and for s = σ for −∞ < σ ≤ 0 which is the negative
part of the real axis. We note that for s = λh the Nyquist frequency is found at s = ±jπ.
The absolute value |R(jω)| of the stability function evaluated on the imaginary axis is
shown for explicit Runge-Kutta methods in Figure 14.28, and for implicit Runge-Kutta
552 CHAPTER 14. SIMULATION

Deflection of free end, LobattoIIIC

5

-1

-2

-3

-4

-5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time

Figure 14.27: Simulation of vibrations in an Euler-Bernoulli beam with a Lobatto IIIC

method. There is an initial excitation that is switched off after 1 s. After this the
vibrational energy of the system is a constant. The simulation method is seen to introduce
damping of the response, and the high frequency components are seen to be more damped
than the low frequency components.

methods in Figure 14.29. The absolute value of the stability function evaluated on the
negative part of the real axis is shown in Figures 14.30 and 14.31.
A special case occurs when R(λh) has a zero, that is, when there is a λz (h) so that
R(λz h) = 0. This results in a dead-beat response

yn+1 = 0 when R(λh) = 0 (14.172)

In this section we will take a closer look at the stability functions for the Runge-Kutta
methods that have the Padé approximations as their stability functions. To simplify
notation we use s = λh. The explicit methods of order p ≤ 4 with σ = p stages have
stability functions


 1+s when p = 1

 s2
1+s+ 2 when p = 2
R(s) = Pp0 (s) = s2 s3

 1 + s + + when p = 3

 2
2 3 6 4
1 + s + s2 + s6 + 24 s
when p = 4

We see that R(s) = 0 occurs for

s1 = −1 when p = 1
s1,2 = −1 ± j when p = 2
s1,2 = −0.7020 ± j1.8073, s3 = −1.5961 when p = 3
s1,2 = −0.2706 ± j2.5048, s3,4 = −1.7294 ± j0.8890 when p = 4
To study the performance of the method when the test equation has a pole λ = jω on
the imaginary axis we insert s = jω in the stability function and get R(jω). It is seen
that for all the explicit Runge-Kutta methods, |R(jω)| → ∞ when ω → ∞.
14.6. STABILITY OF RUNGE-KUTTA METHODS 553

2
10

RK4
Improved
Euler

1
10 Heun Euler's method
Gain of stability function R(s)

0
10
π

-1
10
-1 0 1 2
10 10 10 10
Frequency

Figure 14.28: Stability function R(jω) of explicit Runge-Kutta methods evaluated for
λh = jω. The Nyquist frequency ω N is plotted at ω N h = π.

1
10

0 π
10
Gain of stabiltiy function R(s)

−1
10 R7

L2 L4 R3 R5

−2
10
−1 0 1 2
10 10 10 10
Frequency

Figure 14.29: Stability function R(jω) of implicit Runge-Kutta methods evaluated for
λh = jω. The Nyquist frequency ω N is plotted at ω N h = π. The methods are seen to
damp out frequency components over the Nyquist frequency. The Radau methods have
a roll-off of -1, and the Lobatto IIIC methods have a roll-off of -2.
554 CHAPTER 14. SIMULATION

2
10

1
10
Gain of stability function R(s)

0 π
10
Heun
RK4

-1
10 Euler's
method

exp(-s)
Impr.
Euler
-2
10
-2 -1 0 1 2
10 10 10 10 10
-s on real axis

Figure 14.30: Absolute value of stability function |R(−s)| of explicit Runge-Kutta meth-
ods evaluated for λh = −s. The exact value exp(−s) is plotted for comparison.

1
10

0 π
10
Gain of stability function R(s)

Gauss

-1
10

Radau

-2
10

Lobatto
exp(-s)

-3
10
-1 0 1 2 3
10 10 10 10 10
-s on real axis

Figure 14.31: Absolute value of stability function |R(−s)| of implicit Runge-Kutta meth-
ods evaluated for λh = −s. The exact value exp(−s) is plotted for comparison. The
Radau and Lobatto IIIC have a roll-off for −s large, while |R(−s)| → 1 when −s → ∞
for the Gauss methods.
14.6. STABILITY OF RUNGE-KUTTA METHODS 555

A Gauss method with σ stages, which is an implicit Runge-Kutta method of order

2σ, has stability function

 1+ 12 s

 1 when σ = 1
 1
1− 2s
1 2
1+ 2 s+ 12 s
R(s) = Pσσ (s) = 1− 12 s+ 12
1 2 when σ = 2


s

 1+ 1
s+ 1 2
s + 1
s3
2 10 120
1− 1 s+ 1 s2 − 1 s3
when σ = 3
2 10 120

The stability function is zero for

s1 = −2 when σ = 1
s1,2 = −3 ± j1.7321 when σ = 2
s1,2 = −3.6778 ± j3.5088, s3 = −4.6444 when σ = 3

It is quite interesting to study the stability function of Gauss methods for s = jω. Then,
for p = 1 we see that ¯ ¯
¯ 1 + 12 jω ¯
|R(jω)| = ¯¯ ¯=1
1 − 1 jω ¯ 2

The Radau IIA methods of order p = 2σ − 1 have stability functions

 1

 1−s when σ = 1
 1+ 13 s
σ−1
R(s) = Pσ (s) = 1− 23 s+ 16 s2
when σ = 2

 2 1 2
 1+ 5 s+ 20 s
when σ = 3
1− 3 s+ 3 s2 − 1 s3
5 20 60

There is no zero in R(s) for σ = 1, while there is a zero in s = −3 for σ = 2. For Radau
IIA with σ = 1 ½
1 when ω ¿ ω R1
|R(jω)| = 1
ω when ω À ω R2
The Lobatto IIIC methods of order p = 2σ − 2 have stability functions
( 1
2 when σ = 2
σ−2 1−s+ s2!
R(s) = Pσ (s) = 1
1−s+ 1 s2 − 1 s3
when σ = 3
2 6

For σ = 2 we have ½
1 when ω ¿ ω L1
|R(jω)| = 1
ω2 when ω À ω L2

14.6.8 AN-stability
Before we turn our attention to the nonlinear stability analysis of Runge-Kutta methods
we will present an intermediate result on linear time-varying systems. In this connection
the linear time-varying test system

ẏ = λ(t)y

will be used. The exact solution for linear time-varying test system satisfies
·Z tn+1 ¸
y(tn+1 ) = y(tn )exp λ(t)dt
tn
556 CHAPTER 14. SIMULATION

It is clear that the system is stable in the sense that |y(tn+1 )| ≤ |y(tn )| if Re [λ(t)] ≤ 0
for all t ∈ [tn , tn+1 ].
An implicit Runge-Kutta method for this system can be written
κ = Λ (1yn + hAκ) (14.173)
yn+1 = yn + hbT κ (14.174)
where κ = (k1 . . . kσ )T and 1 = (1 . . . 1)T and
Λ =diag (λ1 , . . . , λσ ) , λi = λ(tn + ci h) (14.175)
−1
Equation (14.173) gives κ = (1 − hΛA) Λ1yn , and insertion into (14.174) gives
yn+1 = RAN (hΛ)yn
where we have defined the stability function
−1
RAN (hΛ) = 1 + bT (I − hΛA) hΛ1 (14.176)

An implicit Runge-Kutta method is said to be AN-stable if Re[λi ] ≤ 0 implies

that |RAN (hΛ)| ≤ 1 and that (1 − hΛA) is nonsingular.
From this definition it is clear that

AN-stability ⇒ A-stability (14.177)

Example 228 The trapezoidal rule has

µ ¶ µ 1
¶
0 0 2
A= 1 1 , b = 1
2 2 2

and after some algebra it is found that

hλ1
1+ 2
RAN (hΛ) = hλ2
1− 2

It is seen that if λ2 = 0 and λ1 is large, then |RAN (hΛ)| > 1, and the method is not AN
stable.
Example 229 Cramer’s rule can be used to find the function RAN (hΛ) defined in
(14.176) in the same way as the stability function R(hλ). This gives
h i
det I − (A − 1bT )hΛ
RAN (hΛ) = (14.178)
det (I − AhΛ)
In Lobatto IIIA the first row of A and the last row of A − 1bT have only zeros. This
means that the numerator of RAN (hΛ) is not a function of λσ , and the denominator of
RAN (hΛ) is not a function of λ1 . This means that if λ2 = . . . = λσ = 0, then |RAN (hΛ)|
can be made arbitrarily large by selecting a large |λ1 | . This means that Lobatto IIIA is
not AN-stable.
Example 230 In Lobatto IIIB the last column of A and the first column of A − 1bT
have only zeros. This means that the numerator of RAN (hΛ) is not a function of λ1 ,
and the denominator of RAN (hΛ) is not a function of λσ . This means that if λ1 = . . . =
λσ−1 = 0, then |RAN (hΛ)| can be made arbitrarily large by selecting a large |λσ | . This
means that Lobatto IIIB is not AN-stable.
14.6. STABILITY OF RUNGE-KUTTA METHODS 557

14.6.9 B-stability
Using the concept of B-stability it is possible to analyze the stability of Runge-Kutta
methods for contracting nonlinear systems. For such systems we will see that the stability
of Runge-Kutta methods can be studied in terms of a simple algebraic condition on the
Runge-Kutta parameters A and b.
Consider the nonlinear systems

ẏ = f (y, t) (14.179)

and the scalar nonnegative function

1
V = e )T P (y − y
(y − y e) (14.180)
2
d
where ye is a solution of dt e = f (e
y y, t) and P is a positive definite symmetric matrix.
We note that the eigenvalues of P must be real and positive, and we denote the largest
eigenvalue by λmax (P) and the smallest eigenvalue by λmin (P). The time derivative of
V along solutions of the systems is
T
e ) P [f (y, t) − f (e
V̇ = (y − y y, t)]

Suppose that the system is contracting, which means that there exists a positive definite
symmetric matrix P and a constant γ ≥ 0 so that

e )T P [f (y, t) − f (e
(y − y e)T P (y − y
y, t)] ≤ −γ (y − y e) , e
∀y, y

This implies that

V̇ ≤ −2γV
and it follows that
V (t) ≤ V (t0 )e−2γ(t−t0 )
and that
µ ¶ 12
λmax (P)
e (t)k ≤
ky(t) − y e (t0 )k e−γ(t−t0 )
ky(t0 ) − y
λmin (P)
e (t) of a contracting system will converge
This means that the two solutions y(t) and y
exponentially to each other.

Suppose that the system (14.179) is contracting with P = I so that

e )T [f (y, t) − f (e
(y − y e )T (y − y
y, t)] ≤ −γ (y − y e)

and that a numerical solution is computed using a Runge-Kutta method. Then, if the
computed solutions yn+1 starting from yn and yen+1 starting from y
en satisfies the con-
dition
en+1 k ≤ kyn − y
kyn+1 − y en k
then the Runge-Kutta method is said to be B-stable.

Consider the linear time-varying test system

ẏ = λ(t)y (14.181)
558 CHAPTER 14. SIMULATION

Then, for two solutions y(t) and ỹ(t) we have

1
V = (y − ỹ)2 ⇒ V̇ = (y − ỹ)λ(t)(y − ỹ) (14.182)
2
It follows that if Re λ(t) ≤ 0 for all t, then the linear time invariant test equation (14.181)
is contracting. Therefore, if a B-stable method is used the numerical solutions will satisfy
|yn+1 − ỹn+1 | ≤ |yn − ỹn |. As ỹ(t) = 0 is a solution it follows that a B-stable method will
also be AN-stable. We may then conclude that

B-stability ⇒ AN-stability ⇒ A-stability (14.183)

14.6.10 Algebraic stability

The property of B-stability is important as it applies to nonlinear contracting systems,
and as B-stability implies A-stability. It is problematic, however, to check if a method is
B-stable by working with the nonnegative function V defined in (14.180). Therefore it is
better to work with algebraic stability which can be established by algebraic manipula-
tions of the elements of the Butcher array. In this section algebraic stability is defined,
and it will be shown that algebraic stability implies B-stability.

An implicit Runge-Kutta method with Butcher array

c A
bT

is said to be algebraically stable if bi ≥ 0 for i = 1, . . . , σ and

T
M =diag (b) A + AT diag(b) − bb ≥ 0

We note that the elements of M are given by

mij = bi aij + bj aji − bi bj (14.184)

An algebraically stable Runge-Kutta method is B-stable, that is,

Algebraic stability ⇒ B-stability (14.185)

This is shown as follows (Hairer et al. 1993): First we make a change of variables in
the Runge-Kutta methods and write
σ
X
Yi = yn + h aij f (Yj , tn + cj h) , i = 1, . . . , σ (14.186)
j=1
Xσ
yn+1 = yn + h bi f (Yi , tn + ci h) (14.187)
i=1
14.6. STABILITY OF RUNGE-KUTTA METHODS 559

en by
Then we denote the differences between the two solutions yn and y
∆yn en , ∆yn+1 = yn+1 − y
= yn − y ei
en+1 , ∆Yi = Yi − Y
h ³ ´i
∆fi = h f (Yi , tn + ci h) − f Ye i , tn + ci h

e i is a vector corresponding to
where Yi is a vector corresponding to the vector yn , and Y
en . Subtraction of the Runge-Kutta equations for the solution y
the vector y en from the
equations of yn gives
σ
X
∆Yi = ∆yn + aij ∆fj
j=1
Xσ
∆yn+1 = ∆yn + bi ∆fi
i=1

Then, we have
σ
X
T T
(∆yn+1 ) ∆yn+1 = (∆yn ) ∆yn + 2 bi (∆fi )T ∆yn
i=1
σ X
X σ
+ bi bj (∆fi )T ∆fj
i=1 j=1
 
σ
X σ
X
T T
= (∆yn ) ∆yn + 2 bi (∆fi ) ∆Yi − aij ∆fj 
i=1 j=1
σ X
X σ
+ bi bj (∆fi )T ∆fj
i=1 j=1
σ
X
T T
= (∆yn ) ∆yn + 2 bi (∆fi ) ∆Yi
i=1
σ X
X σ
T
− mij (∆fi ) ∆fj (14.188)
i=1 j=1

where mij is element (i, j) of matrix M. As

h ³ ´iT ³ ´
(∆fi )T ∆Yi = h f (Yi , tn + ci h) − f Ye i , tn + ci h ei ≤ 0
Yi − Y
Pσ Pσ T
by assumption, and as i=1 j=1 mij (∆fi ) ∆fj ≥ 0 for positive semidefinite M, it
follows that
k∆yn+1 k ≤ k∆yn k
which shows that an algebraically stable method is B-stable.
Example 231 If there exists a positive definite symmetric matrix P so that
y, t)]T P (y − y
[f (y, t) − f (e e) = 0 (14.189)
then it follows from the derivation above that a Runge-Kutta method that satisfies
M =diag (b) A + AT diag(b) − bbT = 0
will give
T T
(∆yn+1 ) P∆yn+1 = (∆yn ) P∆yn (14.190)
560 CHAPTER 14. SIMULATION

Method Order Stability Linear Nonlinear Stiffly

function stability stability Accurate
Explicit, p = σ σ P0σ |hλ| small - No
Gauss 2σ Pσσ A Algebraic No
Radau IA 2σ − 1 Pσσ−1 L Algebraic No
Radau IIA 2σ − 1 Pσσ−1 L Algebraic Yes
σ−1
Lobatto IIIA 2σ − 2 Pσ−1 A not AN No
σ−1
Lobatto IIIB 2σ − 2 Pσ−1 A not AN No
Lobatto IIIC 2σ − 2 Pσσ−2 L Algebraic Yes

Table 14.4: Order and stability properties for some important Runge-Kutta methods.

14.6.11 Properties of Runge-Kutta methods

The properties of some important Runge-Kutta methods are summarized in Table 14.4.

14.7 Automatic adjustment of step size

14.7.1 Estimation of the local error for Runge-Kutta methods
The selection of the step size h is critical for the performance of a Runge-Kutta method.
The main issues in this connection is accuracy and stability. In general the accuracy
of the computed solution depends on the step size. We will see in this section that it
is possible to specify the desired accuracy of the computed solution, and then to have
automatic selection of the step size that ensures the required accuracy. This feature is
used in the standard integrators of MATLAB.
In some applications it may be desirable for simplicity to compute the solution with
a constant step size. For explicit methods the step size must then be selected so that the
computations are stable. For non-stiff systems that are approximately linear in the sense
that the eigenvalues of the Jacobian J do not vary much, it will normally be possible to
select a reasonable step size that ensures stability and a certain accuracy. For systems
with strong nonlinearities so that the eigenvalues of J exhibit large variations, the step
size of an explicit Runge-Kutta method may have to be very small to account for worst-
case situations. For such systems the use of a constant step size is not recommended.
The automatic selection of the step size h is based on finding an estimate of the local
error, and then adjusting the time step so that the local error is less than some specified
tolerance. This is done by computing the numerical solution with two explicit Runge-
Kutta method with different order. Assume that the solution yn+1 is computed with a
method
c A
bT
bn+1 of the same system is computed with a method
of order p, and that the solution y

b
c b
A
b
bT

bn . Then the
of order pb = p + 1. The computation of the value at tn+1 starts with yn = y
14.7. AUTOMATIC ADJUSTMENT OF STEP SIZE 561

local solution yL (tn ; tn+1 ) satisfies

bn+1 + b
yL (tn ; tn+1 ) = yn+1 + en+1 = y en+1

where en+1 = O(hp+1 ) is the local error in the computation of yn+1 , and b
en+1 = O(hp+2 )
is the local error in the computation of ybn+1 . Because b
en+1 is of higher order in h than
en+1 , we can find an estimate of en+1 from

bn+1 − yn+1 = en+1 − b

y en+1 ≈ en+1

The step size can then be adjusted to achieve a specified accuracy in the local error en+1 .
The estimated local error en+1 is an estimate of the local error of the lower order
solution yn+1 . However, the solution ybn+1 is more accurate, so it makes more sense to
use ybn+1 as a starting point for the next time step. The use of y
bn+1 instead of yn+1 is
called local extrapolation, and is normally used. When local extrapolation is used then
yn+1 will be used to denote the high order solution, while ybn+1 denotes the embedded
low order solution.
To make the computations efficient, the two methods are usually designed so that
c =b
c and A =A.b Then the stage computations will be the same in both methods, and
b is said to be an embedded solution in this case.
need only be done once. The solution y
The algorithm is

k1 = f (yn , tn ) (14.191)
..
. (14.192)
kσ = f (yn + h (aσ1 k1 + . . . + aσ,σ−1 kσ−1 ) , tn + cσ h) (14.193)
yn+1 = yn + h(b1 k1 + . . . + bσ kσ ) (14.194)
bn+1
y = yn + h(bb1 k1 + . . . + bbσ kσ ) (14.195)
en+1 bn+1 − yn+1
= y (14.196)

bn+1 is described by an array as follows:

The computation of yn+1 and y

c A
y bT
b
y bT
b
e ET

b − b.
where E = b
Runge-Kutta-Fehlberg 4(5) is a method where y is computed with order p = 4 using
b is of order p = 5 and is computed using six stages.
five stages. The embedded solution y
The method is optimized for accuracy in the fourth order solution yn+1 . The method is
562 CHAPTER 14. SIMULATION

given by the following array.

0
1 1
4 4
3 3 9
8 32 32
12 1932
13 2197 − 7200
2197
7296
2197
439 3680 845
1 216 −8 513 − 4104
1 8
2 − 27 2 − 3544
2565
1859
4104 − 11
40
25 1408 2197
y 216 0 2565 4104 − 15
16 6656 28561 9 2
yb 135 0 12825 56430 − 50 55
1 128 2197 1 2
∆e 360 0 − 4275 − 75240 50 55

Dormand-Prince 5(4) is a method where yn+1 is computed with order p = 5. This

requires six stages. The embedded solution y b is of order p = 4 and is computed using
seven stages. The seventh stage is k7 = yn+1 to reduce the number of computations.
This is recognized as a FSAL method. The method is optimized for accuracy in the fifth
order solution yn+1 . This is the standard MATLAB method for integrating initial value
problems (Shampine and Reichelt 1997). The method is given by the following array.

0
1 1
5 5
3 3 9
10 40 40
4 44
5 45 − 56
15
32
9
8 19372
9 6561 − 25360
2187
64448
6561 - 212
729
9017
1 3168 − 355
33
46732
5247
49
176
5103
− 18656
35 500 125
1 384 0 1113 192 − 2187
6784
11
84
35 500 125
y 384 0 1113 192 − 2187
6784
11
84
5179 7571 393 92097 187 1
yb 57600 0 16695 640 − 339200 2100 40
71 71 71 17253 22 1
∆e 57600 0 − 16695 1920 − 339200 525 − 40

Another variable-step explicit Runge-Kutta method used in MATLAB is the BS23

method of Bogacki and Shampine (Shampine and Reichelt 1997). This is a method where
yn+1 is computed with a third order method, and the error estimate is found by com-
paring the result with an embedded second order method. Also here local extrapolation
is used. The Butcher array is

0
1 1
2 2
3 3
4 0 4
2 1 4
1 9 3 9
21 1 3 1
y 72 4 9 8
2 1 4
yb 9 3 9
5 1 1
∆e − 72 12 9 − 18
14.8. IMPLEMENTATION ASPECTS 563

14.7.2 Adjustment algorithm

Suppose that the specified accuracy is specified in terms of a tolerance etol on the local
error, and that the size of the local error is described by εn+1 = |ei,n+1 | where ei,n+1
is the element of highest absolute value of the vector en+1 . Then, because the method
is of order p we will have εn+1 ≤ Chp+1 for some constant C. Let hnew be defined by
etol = Chp+1
new . Then, if εn+1 > etol the local error is larger than the specified tolerance.
We may then expect that the tolerance can be obtained by using the new and smaller
time step
µ ¶ 1
etol p+1
hnew = h (14.197)
εn+1
In practice a somewhat smaller value may be used by adjusting with a factor of about
0.8. If the tolerance is met, then the time step can be carefully increased.
An alternative adjustment algorithm that has given good results is based on a PI
control method. The derivation of this algorithm is based on the resulting equation
when the logarithm of the adjustment algorithm (14.197) is taken:

1
ln hnew = ln h − (ln εn+1 − ln etol ) (14.198)
p+1
This can be compared with an incremental form of a PI controller

h
un+1 = un − Kp (en − en−1 ) − Kp en (14.199)
Ti
One may compare the adjustment formula with an I controller. Proportional action is
included using

h
ln hnew = ln h − Kp (ln εn+1 − ln εn ) − Kp (ln εn+1 − ln etol ) (14.200)
Ti
which gives the adjustment formula
µ ¶Kp Th µ ¶Kp
etol i εn
hnew = h (14.201)
εn+1 εn+1

which is simplified to
³ ´
µ ¶Kp 1+ Th µ ¶Kp
etol i εn
hnew = h (14.202)
εn+1 etol

The following parameters have been suggested.

Kp = 0.4/(p + 1), Ti = 1.3h (14.203)

14.8 Implementation aspects

14.8.1 Solution of implicit equations
The implicit Runge-Kutta methods involves the solution of a set of implicit nonlinear
equations. To solve these equations it is useful to make a change of variables and write
564 CHAPTER 14. SIMULATION

the stage computations in the form

z1 = h [a11 f (yn + z1 , tn + c1 h) + . . . + a1σ f (yn + zσ , tn + cσ h)]

..
.
zσ = h [aσ1 f (yn + z1 , tn + c1 h) + . . . + aσσ f (yn + zσ , tn + cσ h)]

and to compute the solution yn+1

yn+1 = yn + h [b1 f (yn + z1 , tn + c1 h) + . . . + bσ f (yn + zσ , tn + cσ h)]

Example 232 If A is nonsingular, the update can be found from

yn+1 = yn + d1 z1 + . . . + dσ zσ (14.204)

where
(d1 , . . . , dσ ) = (b1 , . . . , bσ ) A−1 (14.205)
In particular, if aσi = bi then
yn+1 = yn + zσ (14.206)

To solve for z1 , . . . , zσ , a Newton search method is used. The equation is written in

vector form as
Z = h (A ⊗ Iσ ) F(Z)
where    
z1 f (yn + z1 , tn + c1 h)
   
Z =  ...  , F(Z) =  ..
. 
zσ f (yn + zσ , tn + cσ h)
are vectors of dimension dσ, Iσ is the σ × σ identity matrix and
 
a11 Iσ . . . aσ1 Iσ
 .. .. .. 
A ⊗ Iσ =  . . . 
a1σ Iσ . . . aσσ Iσ

is the Kronecker tensor product of A and Iσ .

The solution is found by minimizing the function
T
L = [Z − h (A ⊗ Iσ ) F(Z)] [Z − h (A ⊗ Iσ ) F(Z)] (14.207)

with respect to Z using a Newton search, which is done by the iteration

¡ ¢
H Zi+1 − Zi = −Zi + h (A ⊗ Iσ ) F(Zi ) (14.208)

which is solved for Zi+1 . Here Zi is iteration i of Z, and

 
1 − ha11 J . . . −haσ1 J
 .. .. .. 
H = I−h (A ⊗ J) =  . . . 
−ha1σ J . . . 1 − haσσ J

is an approximation to the Hessian matrix of dimension nσ × nσ, where

∂f
J= (yn , tn )
∂y
14.8. IMPLEMENTATION ASPECTS 565

is the Jacobian evaluated at (yn , tn ). The initial value for the iterations is Z0 = 0.
To solve for Zi+1 , a Gaussian elimination is used where the LU decomposition of H is
needed. The reason for using the approximation of a constant J is that this makes it
possible to use only one LU decomposition at each time step. Details on how to do the
Gaussian elimination is given in (Golub and van Loan 1989) where also algorithms are
included.

14.8.2 Dense outputs

A Runge-Kutta method computes the numerical solution . . . yn−1 , yn , yn+1 . . . at discrete

time instants . . . tn−1 , tn , tn+1 . . . for the system

ẏ = f (y,t), y(t0 ) = y0 (14.209)

In some situations it is not sufficient to have the function values only at the time-steps.
The reason for this is that for some systems it is very important to detect the exact
time of certain events. In particular this is important for systems with discontinuities
in f (y,t). In addition it may be desirable to have function values between the timesteps
for plotting. The solution to this problem is to use an interpolation scheme where an
interpolation yn (α) is computed so that

yn (α), α ∈ [0, 1], yn (0) = yn , yn (1) = yn+1 (14.210)

This can be done by using the original stage computations of the Runge-Kutta method,
possibly with some additional stages, and then interpolating the solution by interpolat-
ing the weighting factors bj . The resulting scheme is called a continuous Runge-Kutta
method .

A continuous Runge-Kutta method is a Runge-Kutta method where interpolation is used

to compute dense outputs yn (α), α ∈ [0, 1] from the scheme
∗
σ
X
ki = f (yn + h aij kj , tn + ci h), i = 1, . . . , σ ∗ (14.211)
j=1
∗
σ
X
yn (α) = yn + h bj (α)kj (14.212)
j=1

∗
The dense outputs are of order p∗ if yn (α)−yL (tn ; tn +αh) = O(hp +1
), where yL (tn ; tn +
αh) is the local solution starting at yL (tn ; tn ) = yn .

For the Dormand-Prince 5(4) method, which is the numerical integration method of
the ode45 in MATLAB, a dense output with order 4 can be computed with the original
stage computations using Hermite interpolation (Dormand and Prince 1986), (Hairer
566 CHAPTER 14. SIMULATION

et al. 1993). The weighting factors of the method are given by the Hermite polynomials.

b1 (α) = α2 (3 − 2α)b1 + α(α − 1)2

5(2, 558, 722, 523 − 31, 403, 016α)
−α2 (α − 1)2 (14.213)
11, 282, 082, 432
b2 (α) = 0 (14.214)
100(882, 725, 551 − 15, 701, 508α)
b3 (α) = α2 (3 − 2α)b3 + α2 (α − 1)2 (14.215)
32, 700, 410, 799
25(443332067 − 31, 403, 016α)
b4 (α) = α2 (3 − 2α)b4 − α2 (α − 1)2 (14.216)
1, 880, 347, 072
32805(23, 143, 187 − 3, 489, 224α)
b5 (α) = α2 (3 − 2α)b5 + α2 (α − 1)2 (14.217)
199, 316, 789, 632
55(29, 972, 135 − 7, 076, 736α)
b6 (α) = α2 (3 − 2α)b6 − α2 (α − 1)2 (14.218)
822, 651, 844
10(7, 414, 447 − 829, 305α)
b7 (α) = α2 (α − 1) + α2 (α − 1)2 (14.219)
29, 380, 432

Note that bj (0) = 0, and that bj (1) = bj , where bj are the coefficients of the fifth order
solution in the Dormand-Prince 5(4) method. It is therefore clear that yn (0) = yn
and yn (1) = yn+1 . In addition, it can be shown that the time derivatives of the dense
solutions satisfy ẏn (0) = hf (yn , tn ) and ẏn (1) = f (yn+1 , tn+1 ). This means that the
dense outputs and their derivatives are continuous at the time-steps tn .

14.8.3 Event detection

Event detection can be formulated as a zero crossing problem by defining a function g
so that the event is given by the condition

g(y,t) = 0 (14.220)

The event can then be detected by computing the numerical solution yn and for each
step check if there is a change of sign from g(yn ,tn ) to g(yn+1 ,tn+1 ). If there is a change
in sign, then the dense output yn (α) is used to find the time of event by solving

g[yn (α),t + αh] = 0 (14.221)

numerically for α. Then the time of the event is given by tn + αh.

This type of event detection can be used for systems with signum terms in f (y,t), as
for problems with dry friction. Then the event that the velocity becomes zero, or leaves
zero, may be detected with this method.

14.8.4 Systems with inertia matrix

There are important applications where the differential equation may be in the form

Mu̇ = φ(u) (14.222)

14.9. INVARIANTS 567

where M is a nonsingular matrix. Runge-Kutta methods can the be implemented with

the stage computations
i−1
X
ki = φ(un + aij kj , tn + ci h), i = 1, . . . , σ
j=1
 
σ
X
un+1 = un + M−1 h bj kj 
j=1

Example 233 One example of this is in robotics where the equation of motion is of the
form
M(q)q̈ + C (q, q̇) q̇ + g(q) = τ
where q is the vector of generalized coordinates and τ is the vector of input general-
ized forces. The matrix M, which is called the inertia matrix is positive definite and
symmetric. The system can be written
µ ¶
I 0
u̇ = φ(u)
0 M(q)

where µ ¶ µ ¶
q q̇
u= , φ(u) =
q̇ −C (q, q̇) q̇ − g(q) + τ
The system could have been written in the form that has been used so far, that is,
µ ¶
q̇
ẏ = f (y) = −1
M (q) [−C (q, q̇) q̇ − g(q) + τ ]

Then the evaluation of f (y) would involve a computationally expensive Gauss elimination.
Therefore it is advantageous to leave the system in the form (14.222) and do a slight
modification to the Runge-Kutta algorithm.

14.9 Invariants
14.9.1 Introduction
The material presented on linear and quadratic invariants in this section is based on
(Hairer 1999), while the section of Hamiltonian systems is based on (Sanz-Serna and
Calvo 1994) and (Hairer 1999).

14.9.2 Linear invariants

Suppose that there is a function
L (y) = wT y (14.223)
T
where w = (w1 . . . wd ) is a vector of constants, so that for all y the time derivative
along solutions of the system ẏ = f (y, t) is zero, that is

L̇ (y) := wT ẏ = wT f (y, t) = 0 for all y (14.224)

568 CHAPTER 14. SIMULATION

Such a function is called a linear invariant. Then it follows from

i−1
X
ki = f (yn + aij kj , tn + ci h) (14.225)
j=1

that
wT ki = 0 (14.226)
and
σ
X
wT yn+1 = wT yn + h bj wT kj = wT yn (14.227)
j=1

We see that linear invariants will be conserved when the solution is computed with
any Runge-Kutta method.

Example 234 Consider a chemical reaction A + B → C in a closed tank. The total

mass m of the chemical components will be constant due to the principle of conservation
of mass. This is written
m = mA + mB + mC = const. (14.228)
where mA , mB and mC are the masses of each of the components A, B and C. This
means that the total mass m is a linear invariant of the system. The mass balance is
assumed to be
d
mA = −ν CA (mA , mB , mC ) (14.229)
dt
d
mB = −ν CB (mA , mB , mC ) (14.230)
dt
d
mC = ν CA (mA , mB , mC ) + ν CB (mA , mB , mC ) (14.231)
dt
where ν CA is the rate of mass transfer from A to C, and ν CB is the rate of mass transfer
from B to C. Then, if the numerical solution to the mass balances (14.229—14.231) is
computed with a Runge-Kutta method, the mass m will be conserved in the numerical
solution.

14.9.3 Quadratic functions

Consider a system
ẏ = f (y, t) (14.232)
and a quadratic function
1 T
V (y) = y Py (14.233)
2
The time derivative of V along solutions of the system is given by

∂V (y)
V̇ (y) := ẏ = yT Pf (y, t) (14.234)
∂y

We then have the following result (Hairer 1999):

14.9. INVARIANTS 569

If the system (14.232) is integrated with a Runge-Kutta method satisfying

mij = bi aij + bj aji − bi bj = 0 (14.235)

then for the quadratic function V (y) = (1/2)yT Py with time derivative V̇ (y) along the
solutions of the system, the following result apply:
σ
X
V (yn+1 ) = V (yn ) + 2 bi V̇ (Yi ) (14.236)
i=1

where Yi is the function value for y corresponding to stage i as defined in (14.186). This
means that if bi ≥ 0, then

V̇ (y) > 0, ∀y =⇒ V (yn+1 ) > V (yn ) (14.237)

V̇ (y) = 0, ∀y =⇒ V (yn+1 ) = V (yn ) (14.238)
V̇ (y) < 0, ∀y =⇒ V (yn+1 ) < V (yn ) (14.239)

This result follows from the calculation

à !T  
σ
X σ
X
T
yn+1 Pyn+1 = yn + h bi f (Yi ) P yn + h bj f (Yj )
i=1 j=1
σ
X
= ynT Pyn + 2h bi YiT Pf (Yi ) (14.240)
i=1
σ X
X σ
2
−h mij f T (Yi ) Pf (Yj ) (14.241)
i=1 j=1

where equations (14.186) and (14.187) are used, and the calculation is done along the
lines of (14.188). The computed solution yn+1 will then result in an increasing V (yn+1 )
if V [y (t)] is increasing for the exact solution y (t), it will result in an invariant V (yn+1 )
if V [y (t)] is an invariant, and it will give a decreasing V (yn+1 ) if V [y (t)] is a decreasing
function for the exact solution.

14.9.4 Quadratic invariants

In this section we will look closer at the case where the quadratic function V defined in
(14.233) is invariant, which is the case if V̇ (y) = 0 for all y. A numerical solution yn is
found from a Runge-Kutta method that will be characterized by the matrix M = {mij }
where mij is defined in (14.235). The topic that is addressed in this section is to find
conditions for the numerically computed invariant V (yn ) to increase, decrease or stay
invariant.
Suppose that the matrix P is positive definite and symmetric. Then it follows that
there is a matrix Q so that P = QT Q. Define gi = Qf i (Yi ). Then, with M = {mij }
we have
Xσ X σ X d
β= mij gjT gi = vkT Mvk (14.242)
i=1 j=1 k=1
570 CHAPTER 14. SIMULATION

Method eigenvalues of M
Implicit Euler 1
Gauss order 2 (Implicit midpoint rule) 0
Gauss methods of any order 0, . . . , 0
Radau IA order 3 0, 0.1250
Radau IIA order 3 0, 0.5
Lobatto IIIA order 2 (Trapezoidal rule) −0.25, 0.25
Lobatto IIIB order 2 −0.25, 0.25
Lobatto IIIC order 2 0, 0.5

Table 14.5: Eigenvalues of M for some implicit Runge-Kutta methods

where the vector vk is defined by

 
(g1 )k
 .. 
vk :=  .  (14.243)
(gσ )k

and (gi )k is element k of the vector gi . As it is assumed that V̇ (y) = 0, ∀y, it follows
from (14.241) that
Xd
T
yn+1 Pyn+1 = ynT Pyn − h2 vkT Mvk
k=1

We recall that a quadratic form can be expressed in term of the eigenvalues of the matrix
according to
X σ
vkT Mvk = λi (M) α2k (14.244)
i=1

where the scalars α2kdepend on M and vk . Therefore the quadratic form is positive if all
the eigenvalues λi (M) of M are positive, the quadratic form is zero if all the eigenvalues
are zero, and the quadratic form is negative if all the eigenvalues are negative. We
conclude that

Let V be the quadratic function V = (1/2)yT Py with symmetric and positive P, and
let λi (M) be an eigenvalue of M = {mij }, where mij is defined in (14.235). Then

V̇ (y) = 0, ∀y and λi (M) < 0,∀i =⇒ V (yn+1 ) ≥ V (yn ) (14.245)

V̇ (y) = 0, ∀y and λi (M) = 0,∀i =⇒ V (yn+1 ) = V (yn ) (14.246)
V̇ (y) = 0, ∀y and λi (M) > 0,∀i =⇒ V (yn+1 ) ≤ V (yn ) (14.247)

Table 14.5 shows the eigenvalues of M for some implicit Runge-Kutta methods, while
Table 14.6. shows the eigenvalues for some explicit methods. The eigenvalues were
computed using the Symbolic Math Toolbox in MATLAB using a script like the one
shown below which computes the eigenvalues for the Lobatto IIIC method.
syms A b M;
A=[1 -1; 1 1]/2;
b=[1 1]’/2;
14.9. INVARIANTS 571

Method eigenvalues of M
Euler’s method −1
Improved Euler −0.5, 0
Modified Euler −1.2, 0.2
Heun’s method of order 3 −0.89, −0.06, 0.3
RK4 −0.27, −0.1, 0, 0.1

Table 14.6: Eigenvalues of M for some explicit Runge-Kutta methods.

M=diag(b)*A+A’*diag(b)-b*b’
eig(M)

Example 235 In the description of rigid body motion the rotation can be described using
Euler parameters µ ¶
η
y= (14.248)
²

where η = cos θ2 , ² = ksin θ2 , and k is a unit vector has the quadratic invariant

yT y = 1 (14.249)

These parameters are used e.g. in strap-down inertial navigation systems. The Euler
parameters satisfy the differential equation
µ ¶ µ ¶
η̇ 1 ωT ²
= (14.250)
²̇ 2 ηω + S (²) ω

where S (²) is the skew-symmetric form of ² and ω is the angular velocity vector of the
rigid body. The invariant yT y = 1 will hold when the system is integrated with a Runge-
Kutta method with mij = bi aij + bj aji − bi bj = 0. In applications it is usual to integrate
with an explicit method and a projection
y
y := (14.251)
|y|

14.9.5 Symplectic Runge-Kutta methods

We consider a system with the Hamiltonian H = H (p, q, t) where
   
p1 q1
   
p =  ...  , q =  ...  (14.252)
pd qd

are in the phase space Ω ⊂ R2d , that is (p, q) ∈ Ω ⊂ R2d . The Hamiltonian system of
differential equations with Hamiltonian H is given by

dpi ∂H dqi ∂H
=− , = , i = 1, . . . , d
dt ∂qi dt ∂pi

We refer to this system as ΣH .

572 CHAPTER 14. SIMULATION

We note that H = H (p, q) , which means that the Hamiltonian is not a function of
time, then the Hamiltonian is an invariant of the Hamiltonian system, which is seen from

∂H dpi ∂H dqi
Ḣ = + =0 (14.253)
∂pi dt ∂qi dt

We define µ ¶
p
y=
q
Then the Hamiltonian system can be written

d
y = J−1 ∇H
dt
where µ ¶ µ ¶
0 I 0 −I
J= , J−1 = JT =
−I 0 I 0
is a skew-symmetric matrix and à !
∂
∂p
∇= ∂
∂q

is the gradient operator.

We note that the Hamiltonian system has zero divergence in the sense that

divẏ = ∇T J−1 ∇H = 0

which is a consequence of J being skew-symmetric. Alternatively this is shown by

Xd µ ¶ X d µ ¶
∂ dpi ∂ dqi ∂2H ∂2H
divẏ = + = − + =0
i=1
∂pi dt ∂qi dt i=1
∂qi ∂pi ∂pi ∂qi

Let y = y(t) be the state of a Hamiltonian system at time t, and let y∗ = y(t∗ ) be
the state at time t∗ . Then a fundamental property of Hamiltonian systems is that
µ ¶T µ ¶
∂y∗ ∂y∗
J =J (14.254)
∂y ∂y

where µ ¶ Ã !
∂p∗ ∂p∗
∂y∗ ∂p ∂q
= ∂q∗ ∂q∗ (14.255)
∂y ∂p ∂q

A system that satisfies (14.254) is said to be symplectic. A system is Hamiltonian if and

only if it is symplectic.

Example 236 Consider the system

q̇ = p, ṗ = −ω 20 q (14.256)

which has the Hamiltonian

1 2 ω 20 2
H= p + q (14.257)
2 2
14.9. INVARIANTS 573

We see that
Ḣ = pṗ + ω 20 q q̇ = −ω 20 qp + ω 20 qp = 0 (14.258)
Then, as H is a quadratic invariant for this system, it follows that a Runge-Kutta method
satisfying
T
M =diag (b) A + AT diag(b) − bb = 0
will result in
Hn+1 = Hn (14.259)
Consider the sensitivity function
∂y(t)
Ψ(t) := (14.260)
∂y0
which gives the sensitivity of the true solution y (t) with respect to the initial condition
y (0) = y0 . We see that Ψ ∈ R2d×2d . The time derivative of the sensitivity when the
system evolves is found from the computation
µ ¶
d d ∂y(t) ∂ dy(t) ∂f (t) ∂f (t) ∂y(t)
Ψ(t) = = = = (14.261)
dt dt ∂y0 ∂y0 dt ∂y0 ∂y(t) ∂y0
to be
∂f
Ψ̇ =
Ψ (14.262)
∂y
If a Runge-Kutta method is applied to this equation, the computational scheme is given
by
∂f
K1 = (Ψn + h (a11 K1 + . . . + a1σ Kσ ) , tn + c1 h) (14.263)
∂y
..
. (14.264)
∂f
Kσ = (Ψn + h (aσ1 K1 + . . . + aσσ Kσ ) , tn + cσ h) (14.265)
∂y0
Ψn+1 = Ψn + h(b1 K1 + . . . + bσ Kσ ) (14.266)
where Ki , i = 1, . . . , σ are matrices of the same dimension as Ψ. Suppose that the same
Runge-Kutta method is applied to the system ẏ = f (y, t) , and that the solution at time
tn+1 is computed to be yn+1 . Then the sensitivity of the computed solution is found to
be given by

µ ¶
∂k1 ∂f ∂yn ∂k1 ∂kσ
= ( + h a11 + . . . + a1σ , tn + c1 h) (14.267)
∂y0 ∂y ∂y0 ∂y0 ∂y0
..
. (14.268)
µ ¶
∂kσ ∂f ∂yn ∂k1 ∂kσ
= ( + h aσ1 + . . . + aσσ , tn + cσ h) (14.269)
∂y0 ∂y ∂y0 ∂y0 ∂y0
µ ¶
∂yn+1 ∂yn ∂k1 ∂kσ
= + h b1 + . . . + bσ (14.270)
∂y0 ∂y0 ∂y0 ∂y0
Comparison with (14.263—14.266) shows that
∂yn+1
= Ψn+1 (14.271)
∂y0
574 CHAPTER 14. SIMULATION

The condition for the solution to be symplectic is that

µ ¶T
∂yn+1 ∂yn+1
J =J (14.272)
∂y0 ∂y0
is therefore equivalent to
ΨTn+1 JΨn+1 = J (14.273)
This expression will hold for all Runge-Kutta methods for which

mij = bi aij + bj aji − bi bj = 0 (14.274)

as
V (Ψ) = ΨT JΨ = J (14.275)
holds for the exact solution, which shows that V (Ψ) is a quadratic invariant.
This means that for a Runge-Kutta method with mij = 0, which is the case for the
Gauss methods, the numerically computed solution will be symplectic, and hence it must
be the solution of a Hamiltonian system with a Hamiltonian that we denote H ∗ . For small
time-steps and a smooth Hamiltonian H, the Hamiltonian system described by H ∗ will
have solutions that are close to the solutions of the Hamiltonian system described by H.
From Tables 14.5 and 14.6 we may conclude that the quadratic invariant V (Ψ) will
decrease for methods like Radau IA, Radau IIA, and Lobatto IIIC, while it will increase
for methods like Euler’s method, Modified Euler, and improved Euler. For methods like
Lobatto IIIA, Lobatto IIIB, Heun’s method of order 3, and RK4 the invariant V (Ψ) may
decrease or increase.

14.10 Rosenbrock methods

A Rosenbrock method with σ stages for the system

ẏ = f (y, t) (14.276)

is given by (Hairer and Wanner 1996)

i−1
X
ki = f (yn + h aij kj , tn + ci h)
j=1
i
X
+hJ ρij kj + ρi hḟ (yn , tn ), i = 1, . . . , σ
j=1
σ
X
yn+1 = yn + h bj kj
j=1

where J is the Jacobian given by J =∂f (yn , tn )/∂y, the interpolation constants satisfy
P
ci = i−1
j=1 aij as for the Runge-Kutta methods, and

i
X
ρi = ρij
j=1

The first term on the right side of the stage computations has the same form as the
stage in an explicit Runge-Kutta method. A linearized term is added to the stage, which
14.10. ROSENBROCK METHODS 575

makes the method implicit. However, while a Newton search is required at each time step
to compute the stages in an implicit Runge-Kutta method, the stage computations in a
Rosenbrock method can be done without iterations according to the formula
i−1
X
Vi ki = f (yn + h aij kj , tn + ci h)
j=1
i−1
X
+hJ ρij kj + ρi hḟ (yn , tn ) (14.277)
j=1

where
Vi = I−hρii J
is a nonsingular matrix for a sufficiently small time step h.
For the test equation
ẏ = λy
a Rosenbrock method gives
κ = λ (1yn + hλ(A + R)κ)
yn+1 = yn + hbT κ
where κ = (k1 . . . kσ )T , 1 = (1 . . . 1)T and R = {ρij }.

The stability function of a Rosenbrock method is given by

h ³ ´i
det I − λh A + R − 1bT
R(hλ) = (14.278)
det[I − λh(A + R)]

It is seen that Rosenbrock methods can have the same type of stability function as an
implicit Runge-Kutta method of the diagonally implicit type, which are implicit method
with aij = 0 for i > j. The main advantage with Rosenbrock methods is that they can
be used for stiff systems without Newton iterations in the stage computations.
A second order method with L-stability developed by Wolfbrandt is given by
Vk1 = f (yn )
2 4
Vk2 = f (yn + hk1 ) − √ hJk1
3 3(2 + 2)
h
yn+1 = yn + (k1 + 3k2 )
4
where
1
√ hJ
V = I−
2+ 2
A modified second order Rosenbrock method with step size control is given by
Vk1 = f (yn , tn ) + hρḟ (yn , tn )
µ ¶
h h
Vk2 = f (yn , tn ) + f yn + k1 , tn + − k1 + hρḟ (yn , tn )
2 2
yn+1 = yn + hk2
1
V = I − hρJ, ρ = √
2+ 2
576 CHAPTER 14. SIMULATION

with step size control using a FSAL computation

³ µ ¶
√ ´ h h
Vk3 = 2f (yn , tn ) + 6 + 2 f yn + k1 , tn + + f (yn+1 , tn+1 )
2 2
³ √ ´
−2k1 − 6 + 2 k2 + hρḟ (yn , tn )
h
error = (k1 − 2k2 + k3 )
6
This method is similar to a Rosenbrock method, but the computation of the second stage
has a term of the type −k1 instead of hJρ21 k1 . This method is used in the MATLAB
function ode23s (Shampine and Reichelt 1997).

14.11 Multistep methods

14.11.1 Explicit Adams methods
The explicit Adams methods, also called Adams-Bashforth methods, has the equation
Z tn+1
y(tn+1 ) = y(tn ) + f (y(t), t)dt
tn

as a starting point. The idea is to calculate a numerical solution from the approximation
Z tn+1
yn+1 = yn + P(t)dt
tn

where P(t) is a polynomial approximation of f of order q so that

P(tn+1−i ) = f (yn+1−i , tn+1−i ) =: fn+1−i , i = 1, 2, . . . , q. (14.279)

This is done with the polynomial

q
X
P(t) = fn+1−i Li (t)
i=1

where Li (t), i = 1, . . . , q are the fundamental Lagrange polynomials (Shampine et al.

1997)
Yq µ ¶
t − tn+1−j
Li (t) = , i = 1, . . . , q
tn+1−i − tn+1−j
j=1, j6=i

These polynomials have the property that

½
1 if i = j
Li (tn+1−j ) =
0 if i 6= j

as shown in Figure 14.32. The polynomial P (t) is shown in Figure 14.33

It is convenient to describe the methods in terms of backward differences. To do this
we define the backward difference operator ∇ by

∇yn = yn − yn−1
14.11. MULTISTEP METHODS 577

Fundamental Lagrange polynomials, order 5

2
L5(t) L1(t)
1.5
L4(t) L2(t)
L3(t)
1

0.5

-0.5

-1

-1.5
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5

Figure 14.32: The Lagrange ploynomials Li (t6−j ) for i = 1, . . . , 5 when h = 1.

P(t) f n-2
f n+1-q f n-1 fn
fn+2-q

yn+1-yn

tn+1-q tn+2-q tn-2 tn-1 tn t n+1

Figure 14.33: The explicit Adams method

Repeated use of the backward difference operator gives

∇m+1 yn = ∇ (∇m yn ) = ∇m yn − ∇m yn−1
for m = 0, 1, 2, . . . where
∇0 yn = yn
A constant h is assumed. Then in the interval tn ≤ t ≤ tn+1 the polynomial P(t) can
be written using a Newton interpolation formula
q−1
X α(α + 1) . . . (α + m − 1) m
P(tn + αh) = ∇ fn
m=0
m!

This leads to the explicit Adams method of order q :

q−1
X
yn+1 = yn + h γ m ∇m f (yn , tn )
m=0
578 CHAPTER 14. SIMULATION

where Z 1
α(α + 1) . . . (α + m − 1)
γm = dα
0 m!
It can be shown that γ m can be found recursively from the recurrence equation
1 1
γ m + γ m−1 + . . . + γ =1
2 m+1 0
The numerical values for γ m are found from the recurrence equation to be

m 0 1 2 3 4
1 5 3 251
γm 1 2 12 8 720

The numerical algorithms are found by inserting the expression for the backwards
difference operator. The algorithms are

yn+1 = yn + hfn
µ ¶
3 1
yn+1 = yn + h fn − fn−1
2 2
µ ¶
23 4 5
yn+1 = yn + h fn − fn−1 + fn−2
12 3 12
µ ¶
55 59 37 9
yn+1 = yn + h fn − fn−1 + fn−2 − fn−3
24 24 24 24
We see that the first order explicit Adams method is Euler’s method.

14.11.2 Implicit Adams methods

In implicit Adams methods, which are also called Adams-Moulton methods, the approx-
imating polynomial P(t) is required to satisfy

P(tn+1−i ) = f (yn+1−i , tn+1−i ) , i = 0, 1, . . . , q − 1 (14.280)

as shown in Figure 14.34.

This is achieved with
Xq
∗ (α − 1)α(α + 1) . . . (α + m − 2) m
P (tn + αh) = ∇ fn+1
m=0
m!

This gives the implicit Adams method of order q + 1:

q
X
yn+1 = yn + h γ ∗m ∇m f (yn+1 , tn+1 )
m=0

where Z 1
(α − 1)α(α + 1) . . . (α + m − 2)
γ ∗m = dα
0 m!
Numerical values are
m 0 1 2 3 4
γ ∗m 1 − 12 1
− 12 1
− 24 19
− 720
14.11. MULTISTEP METHODS 579

P(t) f n-2 f n+1

fn+2-q f n-1 fn
fn+3-q

yn+1 -y n

tn+2-q tn+3-q tn-2 tn-1 tn t n+1

Figure 14.34: The implicit Adams method

which can also be found from the recurrence equation

1 1
γ ∗m + γ ∗m−1 + . . . + γ ∗ = 0, m ≥ 0, γ ∗0 = 1
2 m+1 0
The resulting algorithms are
yn+1 = yn + hfn+1
µ ¶
1 1
yn+1 = yn + h fn+1 + fn
2 2
µ ¶
5 8 1
yn+1 = yn + h fn+1 + fn − fn−1
12 12 12
µ ¶
9 19 5 1
yn+1 = yn + h fn+1 + fn − fn−1 + fn−2
24 24 24 24
It is seen that the first order implicit Adams method is the implicit Euler method, and
that the second order implicit Adams method is the trapezoidal rule.

14.11.3 Predictor-Corrector implementation

An approximate implementation of the implicit Adams method is based on computing a
predictor
q−1
X
bn+1 = yn + h
y γ m ∇m f (tn , xn )
m=0

bn+1 ) in the place of fn+1

with the explicit Adams method, and then use f̂n+1 := f (tn+1 , y
in the implicit Adams method. This gives
yn+1 = yn + hf̂n+1
µ ¶
1 1
yn+1 = yn + h f̂n+1 + fn
2 2
µ ¶
5 8 1
yn+1 = yn + h f̂n+1 + fn − fn−1
12 12 12
µ ¶
9 19 5 1
yn+1 = yn + h f̂n+1 + fn − fn−1 + fn−2
24 24 24 24
580 CHAPTER 14. SIMULATION

This is called a Predictor-Corrector method, which is abbreviated to PECE.

14.11.4 Backwards differentiation methods

In the Backwards Differentiation Formula (BDF) the vector P(t) of polynomials of order
q is required to satisfy the q + 1 constraints

P(tn−q+1 ) = yn−q+1 , . . . , P(tn ) = yn , P(tn+1 ) = yn+1 (14.281)

In this method, the numerical solution at tn+1 is generated by requiring the polynomial
P(t) to satisfy
Ṗ(tn+1 ) = f (yn+1 , tn+1 )

as shown in Figure 14.35. This is done with the Newton interpolating polynomial

P(t) y n-2 y n+1

y n-1 fn+1
yn+1-q yn
yn+2-q

t tn+2-q t n-2 tn-1 tn t n+1

n+1-q

Figure 14.35: The BDF method

à q
!
X (α − 1)α(α + 1) . . . (α + m − 2) m
P(tn + αh) = 1+ ∇ yn+1
m=1
m!

From
¯ Xq ¯
d ¯ d (α − 1)α(α + 1) . . . (α + m − 2) ¯¯
P(tn + αh)¯¯ = ¯ ∇m yn+1
dα α=1 m=1
dα m! α=1

Xq
1 m
= ∇ yn+1
m=1
m

the BDF method of order q is found to be

Xq
1 m
∇ yn+1 = hf (yn+1 , tn+1 )
m=1
m
14.11. MULTISTEP METHODS 581

This gives the following algorithms for q = 1, . . . , 6:

yn+1 − yn = hfn+1
3 1
yn+1 − 2yn + yn−1 = hfn+1
2 2
11 3 1
yn+1 − 3yn + yn−1 − yn−2 = hfn+1
6 2 3
25 4 1
yn+1 − 4yn + 3yn−1 − yn−2 + yn−3 = hfn+1
12 3 4
137 10 5 1
yn+1 − 5yn + 5yn−1 − yn−2 + yn−3 − yn−4 = hfn+1
60 3 4 5
147 15 20 15 6 1
yn+1 − 6yn + yn−1 − yn−2 + yn−3 − yn−4 + yn−5 = hfn+1
60 2 3 4 5 6
In this case the first order method is the implicit Euler method.
A variant of the BDF method is the NDF method (Numerical Differentiation Formu-
las) (Shampine and Reichelt 1997) where an additional term is introduced as follows
q q
à q
!
X 1 m X 1 X
m
∇ yn+1 = hf (tn+1 , yn+1 ) + κ yn+1 − ∇ yn
m=1
m m=1
m m=0

14.11.5 Linear stability analysis

Multistep methods are of the form
¡ ¢
αq yn+1 + αq−1 yn + . . . + α0 yn+1−q = h β q fn+1 + β q−1 fn + . . . + β 0 fn+1−q
which is written
N (z)yn+1 = hD(z)fn+1 (14.282)
where

N (z) = αq z q + αq−1 z q−1 + . . . + α0

D(z) = β q z q + β q−1 z q−1 + . . . + β 0
and z is viewed as the time-shift operator defined by z −1 yn+1 = yn .
Consider the linear test system
ẏ = λy
Then the multistep method gives
N (z)yn+1 = hλD(z)yn+1

Introduction of the z transform gives

N (z)y(z) = hλD(z)y(z)

where z is the complex z transform variable, and y(z) is the z transform of the numerical
solution yn+1 . Equivalently, this is written
[N (z) − hλD(z)] y(z) = 0 (14.283)
The stability of the method can then be investigated by studying the roots of the char-
acteristic equation
N (z) − hλD(z) = 0
582 CHAPTER 14. SIMULATION

which implies stability of the multistep method if the roots are inside the unit circle.
Also the location of the continuous time poles λ can be found as a function of z from
N (z)
hλ =
D(z)
This equation makes it possible to find the poles λ that correspond to the limit of
stability for the multistep method. The stability limit occurs when |z| = 1, which can
be parameterized by z = ejω , −π ≤ ω ≤ π. Then the limit of stability in the s plane is
found by plotting
N (ejθ )
hλ = , −π ≤ θ ≤ π (14.284)
D(ejθ )

14.11.6 Stability of Adams methods

In terms of the z transformation the backwards differences operator ∇ is replaced by
1 − z −1 . This is seen from the z transform of
∇yn = yn − yn−1
which gives
Z {∇yn } = (1 − z −1 )y(z)
For explicit Adams methods the z transform gives
q−1
X
zy(z) = y(z) + hλ γ m (1 − z −1 )m y(z)
m=0

This gives
z−1
hλ = Pq−1
γ m (1 − z −1 )m
m=0
The regions of stability for methods of order 1 to 4 were computed as in (14.284), and
are shown in Figure 14.36. For implicit Adams methods the z transform gives
q
X
zy(z) = y(z) + hλ γ ∗m (1 − z −1 )m zy(z)
m=0

which gives
1 − z −1
hλ = Pq ∗ −1 )m
m=0 γ m (1 − z
The stability regions can then be plotted as in equation (14.284). This gives the stability
regions shown in Figure 14.37. In the PECE Adams method the solution ybn+1 of the
explicit Adams method is inserted for yn+1 on the right hand side of the implicit method.
© £ ¤ ª
zy(z) = y(z) + hλ γ ∗0 zb
y (z) + γ ∗1 [zb
y (z) − y(z)] + γ ∗2 zby (z) − 2y(z) + z −1 y(z) + . . .
After some calculation it can be established that hλ satisfies the second order equation
A (hλ)2 + Bhλ + C = 0
à q
! " q−1 #
X X
A = γ ∗m γ m (1 − z −1 m
)
m=0 m=0
q
X q
X
B = (1 − z) γ ∗m +z γ ∗m (1 − z −1 )m
m=0 m=0
C = 1−z
14.11. MULTISTEP METHODS 583

Figure 14.36: The stability regions of the explicit Adams (Adams-Bashforth) methods
of order 1 to 4. We recognize the stability area of AB(1) as that of Euler’s method.

14.11.7 Stability of BDF methods

For BDF, z transformation gives

Xq
1
(1 − z −1 )m y(z) = hλy(z)
m=1
m

and it follows that

Xq
1
hλ = (1 − z −1 )m
m=1
m

The stability areas are found by plotting hλ for z = ejθ , −π ≤ θ ≤ π, and are shown in
Figure 14.38. It is seen that both the first order and the second order BDF are stable
for ẏ = λy whenever Re(λ) ≤ 0.

14.11.8 Frequency response

From
[N (z) − hλD(z)] y(z) = 0 (14.285)

the dynamics of the numerical solution of ẏ = λy can be analyzed in the z plane as a

function of hλ. We recall that if there is a zp so that

N (zp ) − hλD(zp ) = 0 (14.286)

then the dynamics of y(z) have a pole in the z plane at z = zp . If zp = 0, then this gives
a one-step reset response where yn+1 = 0. If zp = 1, then we have the dynamics of an
integrator where yn+1 = yn .
584 CHAPTER 14. SIMULATION

Figure 14.37: Stability areas of implicit Adams (Adams-Moulton) methods of order 1 to

4. The methods are denoted by AM(q) where q + 1 is the order of the method. Note
that AM(1) is the implicit Euler method, and AM(2) is the trapezoidal rule.

Figure 14.38: Stability areas for BDF methods of order 1 to 6. We note that BDF1 is
the implicit Euler method.
14.12. DIFFERENTIAL-ALGEBRAIC EQUATIONS 585

14.11.9 Adams methods

Explicit Adams methods have the dynamics
" q
#
X
z−1−s γ m (1 − z −1 )m y(z) = 0
m=1

while implicit Adams methods have the dynamics

" q
#
X
∗ −1 m
z−1−s γ m (1 − z ) z y(z)
m=0

It is seen that for s = 0 both methods have one pole which is at z = 1. Moreover, when
|s| → ∞, the explicit method has poles defined by
q
X
γ m (1 − z −1 )m = 0
m=1

Clearly, at least one of the poles for |s| → ∞ is at z = 1, which is also the case for the
implicit methods. This means that high frequency modes are not damped out in the
Adams methods.

14.11.10 BDF methods

When a BDF method is applied to the test equation ẏ = λy we have the expression
" q #
X 1
−1 m
(1 − z ) − λh y(z) = 0
m=1
m

which shows that when λh = 0, there is a pole at z = 1.

The expression
αq yn+1 + αq−1 yn + . . . + α0 = λhyn+1 (14.287)
leads to
(αq − λh) yn+1 + αq−1 yn + . . . + α0 = 0 (14.288)
We see that when λh → ∞, then yn+1 → 0. In the z transform the result is found from
£ ¤
(αq − λh) z q + αq−1 z q−1 + . . . + α0 y(z) = 0 (14.289)

where it is seen that when λh → ∞ the dynamics tend to z q y(z) = 0 which is q poles
at the origin of the z plane. This means that dynamics corresponding to λh À 1 are
damped out, and because of this the BDF methods are well suited for stiff systems. The
standard MATLAB integrator for stiff systems is ode15s which is a variable order BDF
solver (Shampine and Reichelt 1997).

14.12 Differential-algebraic equations

Consider the system
Mu̇ = φ(u)
586 CHAPTER 14. SIMULATION

where u ∈Rd and M is a square matrix of dimension d × d. To begin with we assume

that M has the simple form
µ ¶
Id1 0
M= (14.290)
0 Id2

where d1 + d2 = d and is a constant. It is seen that M is singular whenever = 0.

Let
µ ¶ µ ¶
y f
= u, =φ
z g

where y, f ∈Rd1 and z, g ∈Rd2 . Then the system can be written in the form

ẏ = f (y, z) (14.291)
ż = g(y, z) (14.292)

It is seen that if 6= 0 then the system is of order d, and is described by the differential
equations above, while for = 0 the system is of order d1 and is described by the
differential-algebraic equation

ẏ = f (y, z) (14.293)
0 = g(y, z) (14.294)

If
½ ¾
∂g(y, z) ∂gi
=
∂z ∂zj

is nonsingular the differential algebraic equation is said to be of index 1. It is then

possible to solve z from 0 = g(y, z) giving

z = z(y)

and the dynamics of the system can be written

ẏ = f [y, z(y)] (14.295)

The system (14.295) can be solved with any numerical integration scheme, and the alge-
braic condition is automatically satisfied.
However, in some cases it is desirable to leave the system in the original form and
let tend to zero. in particular, this is done if there is no explicit solution z = z(y)
available, or that the system is in the form Mu̇ = φ(u) where M is possibly nonsingular.
The system
Mu̇ = φ(u)

is said to be a differential algebraic equation of index 1 if it can be transformed in to a

index 1 system as defined above by a change of variables.
14.12. DIFFERENTIAL-ALGEBRAIC EQUATIONS 587

14.12.1 Implicit Runge-Kutta methods for index 1 problems

An implicit Runge-Kutta method for the system (14.291, 14.292) is given by

σ
X
Yi = yn + h aij f (Yj , Zj )
j=1
Xσ
Zi = zn + h aij g(Yj , Zj )
j=1
σ
X
yn+1 = yn + h bi f (Yi , Zi )
i=1
Xσ
zn+1 = zn + h bi g(Yi , Zi )
i=1

We will now show how this scheme can be reformulated so that the equation for zn+1
does not include . This is done by solving g(Yj , Zj ) from the equation for Zi , which
gives

σ
X
hg(Yj , Zj ) = ω ij (Zi − zn )
j=1

where A−1 = Ω = {ω ij }. This expression is inserted into the equation for zn+1 , and
the result is

σ
X σ
X
zn+1 = zn + bi ω ij (Zi − zn )
i=1 j=1
 
σ X
X σ σ X
X σ
= 1 − bi ω ij  zn + bi ω ij Zi
i=1 j=1 i=1 j=1

We note that may be cancelled from this equation, and recall from (14.156) that

σ X
X σ
T −1
R(∞) = 1 − b A 1 =1 − bi ω ij (14.296)
i=1 j=1

This leads to the expression

σ X
X σ
zn+1 = R(∞)zn + bi ω ij Zi
i=1 j=1
588 CHAPTER 14. SIMULATION

which can be used to compute zn+1 , and we get a reformulation of the Runge-Kutta
method in the form
σ
X
Yi = yn + h aij f (Yj , Zj )
j=1
Xσ
Zi = zn + h aij g(Yj , Zj )
j=1
σ
X
yn+1 = yn + h bi f (Yi , Zi )
i=1
σ X
X σ
zn+1 = R(∞)zn + bi ω ij Zi
i=1 j=1

Note that only appears in the equation for Zi . If we let go to zero, the Runge-Kutta
method becomes
σ
X
Yi = yn + h aij f (Yj , Zj )
j=1
σ
X
0 = aij g(Yj , Zj )
j=1
σ
X
yn+1 = yn + h bi f (Yi , Zi )
i=1
σ X
X σ
zn+1 = R(∞)zn + bi ω ij Zi
i=1 j=1

The following observations for the case = 0 are important. At each stage the algebraic
equation g(Yj , Zj ) = 0 is satisfied because A is nonsingular. The algebraic condition is
not necessarily satisfied for zn+1 . If R(∞) = 0, we get

σ X
X σ
zn+1 = bi ω ij Zi
i=1 j=1

where zn+1 is a linear combination of stages Zi . Still the algebraic equations

g(yn+1 , zn+1 ) = 0

are not necessarily satisfied. However, if the method is stiffly accurate, that is, if it has a
nonsingular A matrix and the last row of A equals bT , then yn+1 = Yσ and zn+1 = Zσ ,
and as g(Yσ , Zσ ) = 0 it follows that g(yn+1 , zn+1 ) = 0.
To conclude: Suppose that a stiffly accurate Runge-Kutta method is used to solve
(14.291,14.292) for = 0. Then the computed solution will be the same as if the Runge-
Kutta method was applied to the system ẏ = f [y, z(y)]. The same method can be used
for an arbitrarily small .
14.12. DIFFERENTIAL-ALGEBRAIC EQUATIONS 589

For a general possibly singular M the method is written

σ
X
M (Ui − un ) = h aij φ(Uj )
j=1
σ X
X σ
un+1 = R(∞)un + bi ω ij Ui
i=1 j=1

Also in this case the algebraic condition is satisfied for the stages, and also for yn+1 if a
stiffly accurate method is used.

14.12.2 Multistep methods for index 1 problems

The BDF and NDF methods are of the form
q
X
αq yn+m−q = hf (tn+1 , yn+1 )
m=1

when applied to systems of the form

ẏ = f (t, y)

For index 1 systems in the form

Mu̇ = φ(u)
the BDF and NDF method are given by
q
X
αq (Mu)n+m−q = hφ(un+1 )
m=1

This method works also for singular M. In the case

µ ¶
I 0
M=
0 0

the system can be written

ẏ = f (y, z)
0 = g(y, z)

where µ ¶ µ ¶
y f
= u, =φ
z g
Then, BDF and NDF gives
q
X
αq yn+m−q = hf (yn+1 , zn+1 )
m=1
0 = hg(yn+1 , zn+1 )

It is seen that the algebraic condition is satisfied at each time step.

590 CHAPTER 14. SIMULATION
Chapter 15

Computational fluid dynamics

15.1 Introduction
As shown in Chapter 10, models from fluid mechanics and thermodynamics often involve
partial differential equations. In particular this applies to the transport equations and
the Navier-Stokes equations. These equations cannot be solved analytically except in
special cases and there is a need for finding approximate solutions. Computational fluid
dynamics (CFD) is the collection of tools and methods used for solving this through sim-
ulation. CFD covers both mathematical modelling of the application at hand, methods
of discretization, numerical grid generation and methods of solving the sets of nonlinear
algebraic equations arising from the discretization. CFD methods can be divided into
at least three groups by the method of discretization. These three are finite difference,
finite volume and finite element methods. Due to its simplicity and the fact that the
approximation terms can be given a physical interpretation, we will in the following use
the finite volume method to compute approximate numerical solutions to some important
types of fluid dynamic problems.
The finite volume method uses the integral form of the balance equations as its
starting point. The solution domain is divided into a finite number of control volumes
(CV), and the conservation equations are applied too each CV. At the centroid of each
CV lies a computational node at which the variables are to be calculated. Interpolation
of the nodal vales of neighboring nodes are used at the CV surfaces. This finally results in
an algebraic equation for each node, in which a number of neighbor node values appear,
and a system of equations can now be found. The size of this system depends on the
size of the domain and the grid spacing, but a system of one million equations is not
uncommon. The finite volume method always yields systems with equations with a large
number of zero entries. The structure of the equations depend on the differencing scheme
used, but in most cases a diagonal structure is ensured. This can be exploited by the
solution technique, and iterative methods are often used. Further details on CFD are
found in (Patankar 1980), (Anderson 1995), (Ferziger and Perić 1999) and (Versteeg and
Malalasekera 1995)

15.2 Governing equations

The governing equations for fluid properties such as mass, velocity and energy have many
similarities, and it will be advantageous to introduce a general variable φ to denote a

591
592 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

conserved quantity like mass, momentum or energy. The balance equation for a conserved
quantity φ is written

∂ (ρφ)
+ ∇T (ρφv) = ∇T (Γ∇φ) + Sφ (15.1)
| ∂t
{z } | {z } | {z } |{z}
Net rate of flow rate of change rate of change
rate of change
of φ out of spatial of φ due to of φ due to
of φ of spatial
fluid element diffusion sources
fluid element

in divergence form. Here v is velocity, ρ is density and Γ is a diffusion coefficient.

Equation (15.1) gives the continuity equation (11.6) by substituting φ = 1, the momen-
tum (Navier-Stokes) equations (11.299) by substituting φ = vi , and the energy equation
(11.165) by setting φ = e.
The key step in the finite volume method is the integration of equation (15.1) over a
constant control volume V to give
ZZZ ZZZ ZZZ ZZZ
∂ (ρφ) T T
dV + ∇ (ρφv) dV = ∇ (Γ∇φ) dV + Sφ dV
V ∂t V V V

By use of the divergence theorem as given by equation (10.12), we find

µZ Z Z ¶ ZZ ZZ ZZZ
∂ T T
ρφdV + n (ρφv) dA = n (Γ∇φ) dA + Sφ dV
∂t
| {zV } | ∂V
{z } | ∂V
{z } | V
{z }
rate of change Net rate of change Net rate of change Net rate of
of φ of φ due to convection of φ due to diffusion creation of φ
across the boundaries across the boundaries

which describes the conservation of a fluid property for a finite size control volume V .
In steady state problems, the rate of change term is zero, which leads to the integrated
form of the steady state transport equation:
ZZ ZZ ZZZ
nT (ρφv) dA = nT (Γ∇φ) dA + Sφ dV
∂V ∂V V

In transient time-dependent problems it is also necessary to integrate with respect to

time t:

Z µZ Z Z ¶ Z ZZ
∂
ρφdV dt + nT (ρφv) dAdt
∂t V ∂V
Z ZZ Z ZZZ
T
= n (Γ∇φ) dAdt + Sφ dV dt
∂V V

15.3 Classification
Classification of partial differential equations is an important concept for solving such
equations. Different methods can be developed for different types of equations such that
15.3. CLASSIFICATION 593

distinct properties of the equations can be utilized. The governing partial differential
equations of fluid dynamics as derived in Chapter 10 are quasi linear. This means that
the highest order derivatives occur linearly, they appear by themselves, multiplied with
coefficients which are functions of the dependent variables. It is useful to examine the
mathematical properties of such equations, as any numerical solution of the equations
should exhibit the property of obeying the general mathematical properties of the gov-
erning equations. We will establish a classification of three types of differential equations:
elliptic, hyperbolic and parabolic — all three of which are encountered in fluid dynamics.

P

bon
dar
y co
ndi
tio n
s

initial conditions

Figure 15.1: General propagation problem.

Consider the system of quasi-linear equations given below

∂u ∂u ∂v ∂v
a1 + b1 + c1 + d1 = f1 (15.2)
∂x ∂y ∂x ∂y
∂u ∂u ∂v ∂v
a2 + b2 + c2 + d2 = f2 (15.3)
∂x ∂y ∂x ∂y
where u and v are the dependent variables, and the coefficients ai , bi , ci , di and fi may be
functions of x, y, u and v. Note that these are not the flow equations, but they are similar
in some respects. In Figure 15.1, we have sketched a general problem where the solution
for u and v is known below and on the curve Σ. The solution propagates from the known
initial conditions and with known boundary conditions. In a point P on this curve we
know the values of u and v and their directional derivatives in directions ”downwards”
from Σ. We are now interested in whether it is possible to decide if the solution above P
is given by the information below and on the curve Σ. Or, equivalently: Are these data
sufficient to decide the directional derivatives in P in directions ”upwards” from Σ? The
directional derivatives can be found from the total differentials of u and v, which are
given by
∂u ∂u
du = dx + dy (15.4)
∂x ∂y
∂v ∂v
dv = dx + dy (15.5)
∂x ∂y
Thus, the total differentials du and dv can be found if ∂u ∂u ∂v ∂v
∂x , ∂y , ∂x and ∂y are known,
and we will now investigate under which conditions the partial derivatives in P can be
594 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

calculated from the values of u and v on Σ. Equations (15.2) to (15.5) can be written as
a linear system in the four unknowns ∂u ∂u ∂v ∂v
∂x , ∂y , ∂x and ∂y :

  ∂u  
a1 b1 c1 d1 ∂x f1
 a2 b2 c2 d2   ∂u   f2 
  ∂y =  (15.6)
 dx dy 0 0  ∂v   du 
∂x
0 0 dx dy ∂v dv
∂y
| {z }
A

If u and v are known in P , then ai , bi , ci , di and fi are also known, and if Σ is known
then dx and dy are known. Moreover, if u and v are known on Σ, then du and dv are
known. A unique solution for the four partial derivatives will then exist if and only if
det A 6= 0, .and the directional derivatives will have the same values above and below
Σ. If on the other hand det A = 0, (15.6) will have multiple solutions, and the partial
derivative might be discontinuous over Σ. The characteristic equation of the system is
found by setting det A = 0 :
¯ ¯
¯ a1 b1 c1 d1 ¯
¯ ¯
¯ a2 b2 c2 d2 ¯
¯ ¯=0
¯ dx dy 0 0 ¯¯
¯
¯ 0 0 dx dy ¯

⇓
(a1 c2 − a2 c1 ) (dy)2 − (a1 d2 − a2 d1 + b1 c2 − b2 c1 ) dxdy + (b1 d2 − b2 d1 ) (dx)2 = 0
⇓
µ ¶2
dy dy
(a1 c2 − a2 c1 ) − (a1 d2 − a2 d1 + b1 c2 − b2 c1 ) + (b1 d2 − b2 d1 ) = 0 (15.7)
dx dx
The directions given by (15.7) are called characteristic directions, and a curve, plane or
hyperplane of points where (15.7) is satisfied, is called a characteristic. The characteristic
directions can be real and distinct, real and coinciding or imaginary dependent on the
discriminant

D = (a1 d2 − a2 d1 + b1 c2 − b2 c1 )2 − 4 (a1 c2 − a2 c1 ) (b1 d2 − b2 d1 ) = B 2 − 4AC

being positive, zero or negative. If D > 0, two distinct characteristic lines exist through
each point in the xy-plane, and the system (15.2)-(15.3) is called hyperbolic, if D = 0 one
characteristic line exist through each point in the xy-plane, and the system (15.2)-(15.3)
is called parabolic, and if D < 0 the characteristic lines are imaginary, and the system is
called elliptic.
Alternatively, consider the second order equation

∂2u ∂2u ∂2u ∂u ∂u

a 2
+b +c 2 +e +g + hu = f (15.8)
∂x ∂x∂y ∂y ∂x ∂y

where a, b, c, d, e, f and g may be functions of x, y, u, ∂u

∂x and
∂u
∂y . As before, we will find
2 2 2
∂ u ∂ u ∂ u
conditions so that ∂x2 , ∂x∂y and ∂y2 are uniquely defined on Σ based on known values
15.3. CLASSIFICATION 595

of u, ∂u
∂x and
∂u
∂y on Σ. Using the same argument leading to (15.6) we have
µ ¶
∂u ∂2u ∂2u
d = dx + dy (15.9)
∂x ∂x2 ∂x∂y
µ ¶
∂u ∂2u ∂2u
d = dx + 2 dy (15.10)
∂y ∂y∂x ∂y

Equations (15.8), (15.9) and (15.10) can now be written

   ∂ 2 u   f − e ∂u − g ∂u − hu 
a b c ∂x2 ∂x ¡ ¢∂y
 dx dy 0   ∂2u 
 ∂x∂y  = 
 d³ ∂u
∂x ´


2 ∂u
0 dx dy ∂ u
2
d ∂y
∂y

with characteristic equation

a (dy)2 − bdxdy + c (dx)2 = 0

⇓
µ ¶2
dy dy
a −b +c=0 (15.11)
dx dx
Based on (15.11), equation (15.8) will be hyperbolic if b2 −4ac > 0, parabolic if b2 −4ac =
0 and elliptic if b2 − 4ac < 0.

Remark 8 Notice that the classification of the PDE (15.8), is dependent only on the
coefficients of the second order derivatives.

We will now review some characteristic properties for each type of equation.

15.3.1 Hyperbolic equations

The characteristics of a hyperbolic equation in two variables are plotted in Figure 15.2.
Information at point P influences only the region between the characteristics, that is the
effect of a small disturbance at point P is felt only in the region of influence. Assume that
boundary conditions have been specified on the y-axis. Then the solution can be found by
”marching forward” along the x-axis, starting from the given boundary. The solution at
P will depend only on the part of the boundary conditions that are given between points
a and b on the y-axis. The region to the left of P is called the domain of dependence,
that is properties at P depends only on what is happening in this region. The solution
of hyperbolic equations can be set up as ”marching” solutions, that is starting with the
initial conditions, i.e. at the y-axis, and sequentially calculating the flow field step by step
marching in the x direction. This technique is known as space-marching. Steady inviscid
supersonic flow is an example of flow that is governed by a hyperbolic equation. Another
example is unsteady, inviscid flow. But in this case the governing equation is hyperbolic
with respect to time, as depicted in Figure 15.3. In such equations, the marching variable
is always time, and the technique is known as time-marching.

Example 237 The second order wave equation is given as

∂2u ∂2u
2
− c2s 2 = 0 (15.12)
∂t ∂x
596 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

a P
Domain of
dependence
Region of
influence

Figure 15.2: Characteristics for a hyperbolic equation in two dimensions.

From (15.8) we recognize a = 1, b = 0 and c = −c2s . The discriminant is given by

D = b2 − 4ac = 4c2s > 0

and the equation is hyperbolic. To find the characteristics, we now select dx/dt so that
µ ¶2
dx dx
a −b +c=0 (15.13)
dt dt
which is the characteristic equation, see (15.11). The characteristic directions are given
by p
dx 0 ± 0 + 4 · 1 · c2s
= = ±cs
dt 2·1
We conclude that the second order wave equation has two real characteristics.

15.3.2 Parabolic equations

The characteristic of a parabolic equation is shown as the dotted vertical line in Fig-
ure 15.4. Assume that initial conditions are given along the line ab, and boundary condi-
tions are known along cd and ab. Information at P influences the region to the right of the
characteristic, the region of influence. Parabolic equations are also solved by marching
techniques. Steady boundary-layer flows and unsteady thermal conduction are examples
of problems governed by parabolic equations. The latter being parabolic with respect to
time.

Example 238 The diffusion equation

∂φ ∂2φ
=α 2 (15.14)
∂t ∂x
is a typical example of a parabolic equation. This can be established by recognizing from
(15.8) that a = α, b = c = 0. We also have that g = 1, but this does not influence the
15.3. CLASSIFICATION 597

Region of
influence

Domain of
dependence

x
a b

Figure 15.3: Characteristics for a equation hyperbolic with respect to time.

classification. The discriminant is

D = b2 − 4ac = 0

and it follows that (15.14) is parabolic. The characteristic is found by

µ ¶2
dt dt
a −b +c = 0 (15.15)
dx dx
µ ¶2
dt
α = 0 (15.16)
dx
dt
= 0 (15.17)
dx
dt
The parabolic equation has one characteristic given by dx = 0.

15.3.3 Elliptic equations

For elliptic equations there are no limited regions of influence or domains of dependence.
Information is propagated everywhere in all directions. Consider point P in Figure 15.5.
A disturbance at P will be felt everywhere throughout the region, and the solution at P
is influenced by the entire boundary abcd. Therefore the solution at P must be carried
out simultaneously with the solution of all the other points in the domain and boundary
conditions must be applied for the entire boundary. For this reason, marching solutions
can not be used for elliptic equations. Boundary conditions can be a specification of
the dependent variables u and v in which case the conditions are known as a Dirichlet
condition, or they can be a specification of the derivatives such as ∂u
∂x in which case they
are called Neumann conditions. The boundary conditions can also be of mixed form,
that is involving both Dirichlet and Neumann conditions. Examples of flow that are
598 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

Bo
a un
dar
y con
Domain of dit
dependence ion
sk
P now
n
b
Bou Region of
nd ary influence
con
diti
on sk
now d
n

c
x

Figure 15.4: Characteristic of a parabolic equation

b c

Region of
influence
P

a d
x

Figure 15.5: Domain and boundaries for an elliptic equation in two dimensions.

governed by elliptic equations include steady, subsonic, inviscid flow and incompressible
inviscid flow.

Example 239 The Laplace equation which e.g. describes steady state conductive heat
transfer is a typical example of an elliptic equation. In two dimensions we have

∇2 φ = 0 (15.18)
∂2φ ∂2φ
+ 2 = 0
∂x2 ∂y

By recognizing that a = 1, b = 0 and c = 1 and calculating the discriminant

D = b2 − 4ac = −4,
15.4. DIFFUSION 599

it can be concluded that (15.18) is elliptic. Further, the characteristic is found by

µ ¶2
dx dx
a −b +c = 0 (15.19)
dy dy
µ ¶2
dx
= −1 (15.20)
dy
dx
= ±i (15.21)
dy
The Laplace equation has no real characteristic, a fact that applies to all elliptic equations.

We will now present and apply the finite volume method of CFD to a number of well
known problems of increasing complexity.

15.4 Diffusion
15.4.1 Introduction
We start by considering the simplest possible transport process: pure diffusion in steady
state, which is governed by the equation

∇T (Γ∇φ) + Sφ =0 (15.22)
| {z } |{z}
rate of change rate of change
of φ due to of φ due to
diffusion sources

and can be found by deleting the two terms on the left hand side of (15.1).

15.4.2 Finite volume method for stationary diffusion

1D stationary diffusion
In one dimension, (15.22) is given by
µ ¶
d dφ
Γ +S =0 (15.23)
dx dx

where Γ is called the diffusion coefficient, and S is the source term. Boundary conditions
φA = cA and φB = cB where cA and cB are constants are provided. This process will
be used to illustrate the three basic steps grid generation, discretization, solution of
equations, in the finite volume method.

Grid generation In the finite volume method we want to divide the domain into dis-
crete control volumes. A number of nodal points are placed between the boundaries A
and B. Each node is placed in the center of its control volume, such that the control vol-
ume surfaces are positioned mid-way between the nodes. This arrangement is illustrated
in Figure 15.6, and it is further detailed in Figure 15.7. The general node is named P ,
600 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

and its adjacent nodes W (west) and E (east), respectively. Lower case letters w and e
are used for the boundaries. The distance between P and E is termed δxP E , with other
distances labeled in a similar manner, see Figure 15.7. Notice that the control volume
width is given by ∆x = δxwe . In two and three dimensions, the notation is extended
with the points N (north), S (south), T (top) and B (bottom) and corresponding lower
case letters for the boundaries.
Nodes

P
W E
A B

Control volume
boundaries
Control volume

Figure 15.6: Control volume and boundaries.

δxWP δxPE
δxwP δxPe

W P e E

∆x=δxwe

Figure 15.7: Labeling of nodes and distances in the finitie volume method.

Discretization Integration of (15.23) over the control volume ∆V gives

Z µ ¶ Z
d dφ
Γ dV + SdV = 0
∆V dx dx ∆V

and by using the divergence theorem

Z µ ¶ Z
dφ
nT Γ dA + SdV = 0
A dx ∆V

and evaluate at the east and west surfaces of the control volume we get
µ ¶ µ ¶
dφ dφ
ΓA − ΓA + S∆V = 0 (15.24)
dx e dx w

where A is the cross-sectional area of the control volume face, ∆V is the volume, and S
is average value of S over the control volume. Notice that this discretized equation has
15.4. DIFFUSION 601

a clear physical interpretation. It states that the diffusive flux of φ leaving the e surface
minus diffusive flux of φ entering through the w surface is equal to the generation of
φ inside the control volume. This is in fact one of the most attractive features of this
method. The resulting solution would imply that the integral conservation of quantities
such as mass, temperature and energy is exactly satisfied over any group of control
volumes. This holds for any number of grid points, and even coarse grid solutions exhibit
exact integral balances. We will now use the central difference to evaluate the value of
the diffusion coefficient Γ and the gradient dφ dx at the interfaces of the volume:

ΓW + ΓP
Γw = (15.25)
2
ΓP + ΓE
Γe = (15.26)
2
such that the diffusive terms in (15.24) can be evaluated as
µ ¶
dφ φ − φW
ΓA = Γw Aw P (15.27)
dx w δxW P
µ ¶
dφ φE − φP
ΓA = Γe Ae (15.28)
dx e δxP E
The source term may be a function of φ, and in such cases it may be linearized as

S∆V = Su + Sp φP . (15.29)

Substitution of (15.27), (15.28) and (15.29) into (15.24) gives

µ ¶ µ ¶ µ ¶
Γe Γe Γw Γe
Ae + Aw − Sp φP = Aw φW + Ae φE + Su
δxP E δxW P δxW P δxP E
| {z } | {z } | {z }
aP aW aE

or

aP φP = aW φW + aE φE + Su (15.30)

which represent the discretized version of (15.22)

Finite volume method for 2D steady state diffusion

In two dimensions, (15.22) is given by
µ ¶ µ ¶
∂ ∂φ ∂ ∂φ
Γ + Γ +S =0 (15.31)
∂x ∂x ∂y ∂y

Integration over the control volume gives

Z µ ¶ Z µ ¶ Z
∂ ∂φ ∂ ∂φ
Γ dxdy + Γ dxdy + SdV = 0
∆V ∂x ∂x ∆V ∂y ∂y ∆V
602 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

and by following the same technique as in the 1D case with Ae = Aw = ∆y and An =

As = ∆x, we get
· µ ¶ µ ¶ ¸ · µ ¶ µ ¶ ¸
∂φ ∂φ ∂φ ∂φ
Γe Ae − Γw Aw + Γn An − Γs As +S∆V = 0 (15.32)
∂x e ∂x w ∂y n ∂y s

By substituting
µ ¶
∂φ φE − φP
Γe Ae = Γe Ae
∂x δxP E
µ ¶e
∂φ φ − φW
Γw Aw = Γw Aw P
∂x δxW P
µ ¶w
∂φ φ − φP
Γn An = Γn An N
∂y δyP N
µ ¶n
∂φ φ − φS
Γs As = Γs As P
∂y s δySP

into (15.32) and rearranging, we get

µ ¶
Γe Ae Γe Aw Γs As Γn An
+ + + − Sp φP =
δxP E δxW P δySP δyP N
| {z }
aP
µ ¶ ¶ µ µ ¶ µ ¶
Γw Aw Γe Ae Γs As Γn An
φW + φE + φS + φN + Su
δxW P δxP E δySP δyP N
| {z } | {z } | {z } | {z }
aW aE aS aN

or

aP φP = aW φW + aE φE + aS φS + aN φN + Su (15.33)

which is the discretized version of (15.31).

Finite volume method for 3D steady state diffusion

In three dimensions, (15.22) is given by
µ ¶ µ ¶ µ ¶
∂ ∂φ ∂ ∂φ ∂ ∂φ
Γ + Γ + Γ +S =0
∂x ∂x ∂y ∂y ∂z ∂z

which using the same technique as in the 2D case can be discretized as

aP φP = aW φW + aE φE + aS φS + aN φN + aB φB + aT φT + Su

where
aW aE aS aN aB aT aP
Γw Aw Γe Ae Γs As Γn An Γb Ab Γt At
δxW P δxP E δySP δyP N δzBP δzP T aW + aE + aS + aN + aB + aT − Sp
15.5. SOLUTION OF EQUATIONS 603

15.5 Solution of equations

In order to solve the problem, discretized equations like (15.30) have to be set up at each
of the nodal points, and at the boundaries, the boundary conditions is incorporated.
This will result in a system of ordinary linear algebraic equations which can be solved
by a number of algorithms.

15.5.1 Worked example on stationary diffusion

This section is based on Example 4.2 in (Versteeg and Malalasekera 1995). Consider a
large (so large that temperature gradients are only significant in the x direction) plate
of thickness L = 2cm with thermal conductivity of k = 0.5 K/(Wm) and uniform heat
generation of q = 1000 kW/m3 . The boundary conditions are TA = 100◦ C and TB =
200◦ C. The temperature distribution is governed by
µ ¶
d dT
k +q =0
dx dx
We divide the domain into five control volumes, that is δx = 0.004 m, and consider a
unit area in the yz plane. Integration over a control volume (using S∆V = q∆V ) gives
Z µ ¶ Z
d dT
k dV + qdV = 0
dx dx
µ ∆V ¶ µ ¶ ∆V
dT dT
kA − kA + q∆V = 0
dx e dx w
TE − TP TP − TW
ke A − kw A + qAδx = 0
δx δx
which is rearranged to
µ ¶
ke A kw A ke A kw A
+ TP = TW + TE + qAδx (15.34)
δx δx δx δx
which, when comparing to (15.30), has the following coefficients:

aW aE aP Sp Su
kA kA
δx δx aW + aE − Sp 0 qAδx

Equation (15.34) have to be somewhat modified at the boundaries in order to include

the boundary conditions. A linear approximation is used for temperatures between the
boundary point and the nodal point. So for node 1, we have
TE − TP TP − TA
ke A − kA A + qAδx = 0
δx δx/2
which have, using kA = k, the following table of coefficients:

aW aE aP Sp Su
kA
0 δx aW + aE − Sp - 2kA
δx qAδx + 2kA
δx TA

Similar for node 5, we get

aW aE aP Sp Su
kA
δx 0 aW + aE − Sp - 2kA
δx qAδx + 2kA
δx TB
604 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

Using numerical values, we get the following coefficients for all the nodes:

Node aW aE Su Sp aP
1 0 125 4000 + 250TA −250 375
2 125 125 4000 0 250
3 125 125 4000 0 250
4 125 125 4000 0 250
5 125 0 4000 + 250TB −250 375

This set of equations can be written in matrix form:

    
375 −125 0 0 0 T1 29000
 −125 250 −125 0 0   T2   4000 
    
 0 −125 250 −125 0  T3 = 4000 
    
 0 0 −125 250 −125   T4   4000 
0 0 0 −125 375 T5 54000

The matrix of coefficients is seen to be tridiagonal, which is generally the case in problems
like this. The system can be found to have the solution
   
T1 150
 T2   218 
   
 T3  =  254 
   
 T4   258 
T5 230

which can be found by using an appropriate technique for solving systems of ordinary
algebraic equations. Particularly well suited is the TDMA (tridiagonal matrix algorithm)
or Thomas algorithm, which can be applied iteratively.

15.6 Stability issues

Before we proceed to study more complicated flow phenomena, we need some concepts
regarding stability of the numerical solutions. Convergence is the property of a numer-
ical method to produce a solution which approaches the exact solution of the original
problem as the grid spacing tends to zero. Consistent numerical schemes produce sys-
tems of algebraic equations which can be demonstrated to be equivalent to the original
governing equation as the grid spacing tends to zero. Stability means that all solutions
of the numerical scheme are uniformly bounded functions of the initial conditions for
sufficiently small grid spacing. Convergence is usually very difficult to establish theoret-
ically. However, for linear problems, convergence can be established using the following
theorem.

Theorem 2 (Lax’s equivalence theorem) If a differential approximation to a initial and

boundary value problem is consistent, stability is a necessary and sufficient condition for
the scheme to be convergent.

For CFD problems this is of limited value as the governing equations often are nonlin-
ear, and in that case stability and consistency are necessary, but not sufficient conditions
for convergence. (Patankar 1980) presented rules which yields robust finite volume cal-
culations schemes. Robust schemes has three properties: conservativeness, boundedness
15.7. FINITE VOLUME METHOD FOR DIFFUSION DYNAMICS 605

and transportiveness. As we shall see, these concepts are designed into finite volume
schemes, and they are according to (Versteeg and Malalasekera 1995) commonly ac-
cepted as alternatives for the more mathematically rigorous concepts of convergence,
consistency and stability. Stability will be studied in some more detail in section 15.10.

• Conservativeness: Any discretization scheme for diffusion convection problems

must ensure that the flux through a common face is represented in a consistent
manner in adjacent control volumes. Ideally, when calculating the overall flux
balance, the fluxes between the control volumes should cancel out, leaving only the
fluxes at the boundaries,
• Boundedness: The discretized equations at each node give rise to a set of alge-
braic equations that need to be solved. This set might be quite large, and iterative
solution methods are often used. The Scarborough criterion (Scarborough 1966)
states that a sufficient condition for a convergent iterative method can be expressed
in terms of the values of the coefficients of the discretized equations. The criterion
is given by P ½
|anb | ≤ 1 at all nodes
0 (15.35)
|aP | < 1 at one node at last

where a0P is the net coefficient for the central node P , that is

a0P = aP − SP
P
and |anb | is the net coefficient for all the neighboring nodes. Another criterion
for boundedness is that all coefficients of the discretized equation have the same
sign. Physically, this implies that the increase of φ in one node should result in an
increase of φ in the neighboring nodes.
• Transportiveness: Define the Peclet number as
F ρu
Pe = =
D Γ/δx

The Peclet number P e is a measure of the relative strength of diffusion and con-
vection, and the terms F and D will be further treated in Section 15.8. For pure
diffusion we have P e = 0, and for pure convection we have P e = ∞.

15.7 Finite volume method for diffusion dynamics

We now include a rate of change term in the governing equation in order to study time
dependent problems. Unsteady diffusion will be studied by looking at heat conduction,
which in one dimension, is described by
µ ¶
∂T ∂ ∂T
ρc = k +S (15.36)
∂t ∂x ∂x

where ρ is the density (assumed constant), c is the specific heat of the material and k
is the thermal conductivity. As seen in Section 15.3.2 this problem is parabolic in time,
and we will solve it by marching in time from an initial distribution of temperature T .
606 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

Given the nodal values of T at time t (denoted TP0 , TW0

, TE0 and so on), we will find the
1 1 1
values of T at time t + ∆t (denoted TP , TW , TE and so on). Integrating over the control
volume using the same techniques as before, we get
Z Z µ ¶ Z
∂T ∂ ∂T
ρc dV = k dV + SdV (15.37)
∆V ∂t ∆V ∂x ∂x ∆V
Z e µ ¶ µ ¶
∂T ∂T ∂T
ρc dV = kA − kA + S∆V (15.38)
w ∂t ∂x e ∂x w
Assuming that the temperature is uniform within each control volume, we have
µ ¶ µ ¶
∂T ∂T ∂T
ρc ∆V = kA − kA + S∆V
∂t ∂x e ∂x w
Using the theta method described in Section 14.5.6, we get.
£ ¡ ¢¤
ρc∆V TPn+1 = ρc∆V TPn + ∆t θf (TPn , tn , x) + (1 − θ) f TPn+1 , tn+1 , x (15.39)
where
TEn − TPn T n − TWn
f (TPn , tn , x) = ke − kw P + S∆V
δxP E δxW P
and
¡ ¢ T n+1 − TPn+1 T n+1 − TW n+1
f TPn+1 , tn+1 , x = ke E − kw P + S∆V
δxP E δxW P
As before, central differencing has been applied to the diffusion terms.
In (15.39), θ = 1 gives the Euler method, θ = 12 gives the trapezoidal rule (usually
referred to as the Crank-Nicolson method in the CFD literature) and θ = 0 gives the
implicit Euler method. Using ∆V = A∆x, equation (15.39) can also be written as
µ µ ¶¶
∆x ke kw ke ¡ ¢
ρc + (1 − θ) + TPn+1 = (1 − θ) TEn+1 + θTEn
∆t δxP E δxW P δxP E
kw ¡ n+1 n
¢
+ (1 − θ) TW + θTW
δxW P
µ µ ¶¶
∆x ke kw
+ ρc −θ + TPn
∆t δxP E δxW P
+S∆x
or

¡ ¢ ¡ ¢
aP TPn+1 n+1
= aW (1 − θ) TW n
+ θTW + aE (1 − θ) TEn+1 + θTEn (15.40)
¡ ¢
+ a0P − θaW − θaE TPn + S∆x

where
aP = (1 − θ) (aW + aE ) + a0P
∆x
a0P = ρc
∆t
kw
aW =
δxW P
ke
aE =
δxP E
15.7. FINITE VOLUME METHOD FOR DIFFUSION DYNAMICS 607

We have now arrived at a discretized version of (15.36), and we will review three different
schemes given by the theta method.

Euler (θ = 1)
In the Euler method, or explicit scheme, θ = 1, and the source term is linearized as

S∆x = Su + SP TPn

The discretized heat transfer equation is then given by

¡ ¢
aP TPn+1 = aW TW
n
+ aE TEn + a0P − aW − aE + SP TPn + Su (15.41)

where

aP = a0P (15.42)
∆x
a0P = ρc
∆t
kw
aW =
δxW P
ke
aE =
δxP E

For this method to be stable, all coefficients in the discretized equation (15.41) need to
be positive. This is satisfied if a0P − aW − aE > 0. When assuming uniform grid spacing
δxP E = δxW P = ∆x and constant k = ke = kw , this can be written

2
(∆x)
∆t < ρc (15.43)
2k

This sets a maximum limit on the time step size ∆t, and reduction of ∆x to improve
spatial accuracy forces us to chose a much smaller time step in order to ensure stability.

Crank-Nicolson (θ = 12 )
1
In the Crank-Nicolson method, θ = 2 and the source term is linearized as

TPn + TPn+1
S∆x = Su + SP
2
The discretized heat transfer equation is

n+1 µ ¶
n
TW + TW TEn + TEn+1 aW aE SP
aP TPn+1 = aW + aE 0
+ aP − − + TPn + Su
2 2 2 2 2
608 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

where
1 SP
aP = (aW + aE ) + a0P −
2 2
∆x
a0P = ρc
∆t
kw
aW =
δxW P
ke
aE =
δxP E
Again, all coefficients in the discretized equation must be positive, which is satisfied if

(∆x)2
∆t < ρc
k

where δxP E = δxW P = ∆x and constant k = ke = kw . Again, we notice the restriction

on the time step ∆t, which is only slightly less restrictive than for the Euler scheme.

Implicit Euler (θ = 0)
In the implicit Euler scheme, θ = 0 and the source term is linearized as

S∆x = Su + SP TPn+1

and the discretized heat transfer equation is

aP TPn+1 = aW TW
n+1
+ aE TEn+1 + a0P TPn + Su

where

aP = aW + aE + a0P − SP
∆x
a0P = ρc
∆t
kw
aW =
δxW P
ke
aE =
δxP E
As all coefficients in the discretized equation are positive, the method is unconditionally
stable. CFD computations for two- and three-dimensional transient diffusion can be done
using the same methods as in this section.

Example 240 Simulation of temperature control using the Finite Volume CFD
method
Consider an insulated rod equipped with a heating element at the east end as depicted in
Figure 15.8. The one dimensional transient heat conduction equation is
µ ¶
∂T ∂ ∂T
ρc = k +S
∂t ∂x ∂x
15.7. FINITE VOLUME METHOD FOR DIFFUSION DYNAMICS 609

x0 insulation xL

T meas
T
x
t, 0  0
S  k p T ref  T meas 

Figure 15.8: Temperature control of a rod. A heating element at the east end is controlled
with feedback from the temperature at the middle of the rod. The rest of the rod is
insulated.

where ρc = 107 JK/m3 , k = 10 WK/m, and length L = 0.02 m. Initially T (x, 0) = 300 K.
The insulated west end implies the boundary condition ∂T ∂x (t, 0) = 0. We will use a P
controller to control the power S of the heating element which is located at the east end
(x = L). Temperature will be measured at the middle of the rod so that the source term
will be given by
S = kp (Tref − Tmeas )
where kp > 0 is the controller gain Tref is the temperature set point and Tmeas is the
temperature measurement. We use the Euler method for time discretization and divide
the spatial domain into five control volumes such that ∆x = 0.004 m. The computational
grid will be similar to the one used in Section 15.5.1. By using equation (15.40), we find
for node 1 µ n+1 ¶
TP − TPn k
ρc ∆x = (T n − TPn )
∆t ∆x E
The discretized equations for node 2,3 and 4 are given by (15.41) and (15.42), and for
node 5 we find
µ n+1 ¶
TP − TPn k
ρc ∆x = kp (Tref − T3n ) − (T n − TW
n
)
∆t ∆x P

where Tmeas = T3 has been used as the measurement is made at the middle. The time
step ∆t must satisfy (15.43), which leads to ∆t < 8.We chose ∆t = 2. Using numerical
values, it can be shown that the equation for each node is given by

Node 1: 200T1n+1 = 25T2n + 175T1n

Node 2-4: 200TPn+1 = 25TWn
+ 25TEn + 150TPn (15.44)
n+1
Node 5: 200T5 = 25T4 + 175T5n + kp0 (Tref − T3 )
n

Notice that due to measurement and control not being collocated, the introduction of T3 in
the equation for node 5 breaks the tridiagonal structure of the equations. The equation set
(15.44) was solved numerically using MATLAB. The result is plotted in Figure 15.9. As
can be seen, the desired temperature is reached throughout the rod, which is not surprising
due to the insulation.

The equation set in the example was evaluated with the following MATLAB script.
kp=50; % Controller gain
T_ref=400; % Desired temperature
610 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

for i=1:5, % Initialization

T(1,i)=300;
end
for t=2:150; % 150 timesteps
T(t,1)=1/200*(25*T(t-1,2)+175*T(t-1,1)); %node 1
T(t,2)=1/200*(25*T(t-1,1)+150*T(t-1,2)+25*T(t-1,3)); %node 2
T(t,3)=1/200*(25*T(t-1,2)+150*T(t-1,3)+25*T(t-1,4)); %node 3
T(t,4)=1/200*(25*T(t-1,3)+150*T(t-1,4)+25*T(t-1,5)); %node 4
T(t,5)=1/200*(25*T(t-1,4)+175*T(t-1,5)+kp*(T_ref-T(t-1,3))); %node 5
end
figure(1)
surf(T’)

550

500
Temperature [K]

450

400

350

300
5

4 150

3 100

2 50

1 0
Node #
time [s]

Figure 15.9: Simulation of the equation set (15.44).

15.8 Finite volumes for Convection-Diffusion

15.8.1 Introduction
Convection is caused by fluid flow. In this section we will study methods to obtain a
solution for φ for a given flow field. The transport equation for convection-diffusion of a
general property φ is given by (15.1), and repeated here for convenience:

∂ (ρφ)
+ ∇T (ρuφ) = ∇T (Γ∇φ) + Sφ (15.45)
| ∂t
{z } | {z } | {z } |{z}
Net rate of flow rate of change rate of change
rate of change
of φ out of of φ due to of φ due to
of φ of
fluid element diffusion sources
fluid element

where ρ is the density, u is the flow velocity vector and Γ is the diffusion coefficient.
15.8. FINITE VOLUMES FOR CONVECTION-DIFFUSION 611

δxWP δxPE
δxwP δxPe
uw ue

W P e E

∆x=δxwe

Figure 15.10: Control volume for 1D convection and diffusion. ue is the flow velocity
entering the CV, and uw is the flow velocity leaving the CV.

15.8.2 Finite volume method for 1D diffusion and convection dy-

namics
In one dimension, (15.45) is given by
µ ¶
∂ ∂ ∂ ∂φ
(ρφ) + (ρuφ) = Γ + S, (15.46)
∂t ∂x ∂x ∂x

where u is the x component of the flow velocity vector.Using the same methodology as
in Section 15.7, integration of (15.46) over the control volume in Figure 15.10 gives
Z µ ¶ µ ¶
∂ dφ dφ
(ρφ) dV + (ρAuφ)e − (ρAuφ)w = ΓA − ΓA (15.47)
∆V ∂t dx e dx w
µ ¶ µ ¶
∂φ dφ dφ
ρ ∆V + (ρAuφ)e − (ρAuφ)w = ΓA − ΓA (15.48)
∂t dx e dx w
We now define the convection and diffusion terms
F = ρu
Γ
D =
δx
so that
Fw = (ρu)w , Fe = (ρu)e
Γw Γe
Dw = ,D=
δxW P δxP E
We choose here to use the implicit Euler scheme of Section 15.7 for integration, but other
schemes might just as well be used. Linearizing the source term as
S∆V = Su + SP φn+1
P

the integrated version of (15.47) can be written

£ ¡ ¢
ρ∆V φn+1
P = ρ∆V φnP + ∆t −Fe φn+1 e + Fw φn+1
w + De φn+1
E − φn+1
P (15.49)
¡ n+1 n+1
¢ n+1
¤
−Dw φP − φW + Su + SP φP
In (15.49), the diffusion terms are approximated by using the central difference. Calcula-
tion of the convection terms at the e and w surfaces can be done by a number of different
schemes, and some of them will now be reviewed.
612 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

Central difference
One alternative is to use the central difference to approximate the convection terms. By
substituting

φP + φE
φe =
2
φW + φP
φw =
2
into (15.49) and rearranging, we get
µ µ ¶ µ ¶ ¶
ρ∆V Fw Fe
+ Dw + + De − + (Fe − Fw ) − SP φn+1
P
∆t 2 2
µ ¶ µ ¶
ρ∆V n Fw Fe
= φ + Dw + φn+1
W + De − φn+1 + Su
∆t P 2 2 E

or
aP φn+1
P = a0P φnP + aW φn+1 n+1
W + aE φE + Su
and we have the same type of discretized equation that we derived for the pure diffusion
problems. Regarding conservativeness, it can easily be shown that fluxes of the type
(φ −φi )
Γ i+1∆x will cancel out in pairs, leaving only the fluxes at the boundaries, and thus
the central differencing scheme is conservative. For a flow that is governed by (15.49)
and simultaneously satisfies continuity, that is

Fe − Fw = 0

it follows that, a0P = aW + aE + a0P . The Scarborough criterion can then be calculated as
P ¯ ¯
|anb | |aW | + |aE | + ¯a0P ¯
=
|a0P | |aW + aE + a0P |

and it can be seen that the conditions of (15.35) are satisfied. As aE = De − F2e ,
this coefficient can be negative if convection is sufficiently strong. Provided all other
coefficients are positive, this will violate the requirement for boundedness. Moreover

Fe Fe
aE > 0 =⇒ De − > 0 =⇒ = P ee < 2
2 De

This means that the central differencing scheme in this case will produce bounded solu-
tions only if the Peclet number satisfies P ee < 2. Regarding transportiveness, the scheme
does not recognize the direction of the flow, and would not be useful for high P e.The
central difference scheme is accurate to second-order.

The upwind scheme

To remedy this, the upwind scheme takes flow direction into account. If the flow direction
is positive (from west in Figure 15.10), that is uw > 0 and ue > 0, we chose

φw = φW and φe = φP .
15.8. FINITE VOLUMES FOR CONVECTION-DIFFUSION 613

When inserted into (15.49) this gives the discretized equation

µ ¶
ρ∆V
+ (Dw + Fw ) + De + (Fe − Fw ) − SP φn+1 P
∆t
ρ∆V n
= φ + (Dw − Fw ) φn+1 W + De φE
n+1
+ Su . (15.50a)
∆t P
If the flow direction is negative (from east in Figure 15.10), that is uw < 0 and ue < 0,
we chose
φw = φP and φe = φE .
and consequently
µ ¶
ρ∆V
+ Dw + (De − Fe ) + (Fe − Fw ) − SP φn+1
P
∆t
ρ∆V n
= φ + Dw φn+1
W + (De − Fe ) φE
n+1
+ Su (15.51)
∆t P
By combining (15.50a) and (15.51), the upwind scheme can be written

aP φn+1
P = a0P φnP + aW φn+1 n+1
W + aE φE

where
aP = a0P + aW + aE + (Fe − Fw )
ρ∆V
a0P =
∆t
aW = Dw + max(Fw , 0)
aE = De + max(0, −Fe )

The upwind scheme is conservative, the fact that aP = aW + aE implies that the Scar-
borough criterion is met. Also, all the coefficient are positive, and transportiveness is
built into the formulation. The upwind scheme is accurate to first order.

The hybrid scheme

The hybrid differencing is based on a combination of central and upwind differencing
schemes. For small Peclet numbers (P e < 2), central differencing is used while the
upwind scheme is used for large Peclet numbers. The reason for this is to make use of
the second order accuracy of the central differencing scheme. It can be shown that the
hybrid scheme can be written

aP φn+1
P = a0P φnP + aW φn+1 n+1
W + aE φE

with

aP = a0P + aW + aE + (Fe − Fw )
ρ∆V
a0P =
∆t µ µ ¶ ¶
Fw
aw = max Fw , Dw + ,0
2
µ µ ¶ ¶
Fe
aw = max −Fe , De − ,0
2
614 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

Quadratic upwind scheme (QUICK)

The quadratic upstream interpolation for convective kinetics (QUICK) scheme (Leonard
1979) uses a three-point upstream-weighted quadratic interpolation for cell face values.
The value of φ at the cell face between two bracketing nodes i and i − 1 and an upstream
node n. − 2 is given by
6 3 1
φf ace = φi−1 + φi − φi−2
8 8 8
When uw > 0 and ue > 0, a quadratic fit through W W , W , and P is used to evaluate φw
and a quadratic fit through W , P , and E is used to evaluate φe . If uw < 0 and ue < 0,
values at W , P , and E is used to evaluate φw and P , E, and EE is used to evaluate
φe . The diffusion terms are evaluated using central differencing. For one-dimensional
convection diffusion, the QUICK-scheme can be summarized as

aP φn+1
P = a0P φnP + aW φn+1 n+1
W + aE φE + aW W φn+1 n+1
W W + aEE φEE

where

aP = a0P + aW + aE + (Fe − Fw ) + aW W + aEE

6 1 3
aW = Dw + αw Fw + αe Fe + (1 − αw ) Fw
8 8 8
1
aW W = − αw Fw
8
3 6 1
aE = De − αe Fe − (1 − αe ) Fe − (1 − αw ) Fw
8 8 8
1
aEE = (1 − αe ) Fe
8
where αw = 1 for Fw > 0, αw = 0 for Fw < 0, αe = 1 for Fe > 0 and αe = 1 for Fw < 0.
The QUICK algorithm is accurate to third order. Notice that the QUICK algorithm will
yield a penta-diagonal system of equations.

15.9 Pressure-velocity coupling

15.9.1 Introduction
Viscous incompressible flow is governed by the incompressible Navier-Stokes equations.
These equations exhibit a mixed elliptic-parabolic behavioral, and therefore, as discussed
in Section 15.3, the problem can not be solved by time marching techniques. The velocity
components are governed by the momentum equations, which are particular cases of the
general differential equation for φ. The momentum equations, e.g. in 2D, are given by

µ ¶ µ ¶
∂ (ρu) ∂ ¡ 2¢ ∂ ∂ ∂u ∂ ∂u ∂p
+ ρu + (ρuv) = µ + µ − + Su (15.52)
∂t ∂x ∂y ∂x ∂x ∂y ∂y ∂x
µ ¶ µ ¶
∂ (ρv) ∂ ∂ ¡ 2¢ ∂ ∂v ∂ ∂v ∂p
+ (ρuv) + ρv = µ + µ − + Sv (15.53)
∂t ∂x ∂y ∂x ∂x ∂y ∂y ∂y

These equations can be shown to be parabolic. The velocity also satisfy the continuity
equation
15.9. PRESSURE-VELOCITY COUPLING 615

∂ ∂
(ρu) + (ρv) = 0 (15.54)
∂x ∂y

which is elliptic. The incompressible Navier-Stokes equations can be obtained from the
compressible form simply by setting constant density ρ, which would lead us to believe
that the equations can be solved numerically by using one of the previous mentioned
techniques. This is unfortunately not the case. (Anderson 1995) cites a result where an
approximate stability criterion for an explicit Navier-Stokes solution is shown to be
1
∆t ≤ q
2 2
|u| /∆x + |v| /∆y + a 1/ (∆x) + 1/ (∆y)

where α is the sonic velocity. For a compressible flow a is finite, and it is possible to
find a finite ∆t to guarantee stability of the solution. However, for an incompressible
flow, the sonic velocity is theoretically infinite, and we would get ∆t = 0. Consequently,
another method has to be found. In addition to the equations (15.52), (15.53) and (15.54)
being nonlinear, another difficulty in solving this problem is the presence of the pressure
∂p ∂p
gradient terms ∂x and ∂y . As can be seen we have no differential equation for ∂p∂t . We
also recognize that a consequence of the system (15.52)-(15.54) being a parabolic-elliptic,
we have no transient term in (15.54). It is interesting to compare this to a system of
differential-algebraic equations (DAE).

Remark 9 If the flow is compressible, the continuity equation

∂ ∂ ∂
(ρ) + (ρu) + (ρv) = 0 (15.55)
∂t ∂x ∂y
may be used as a transport equation for density, and the energy equation is a transport
equation for temperature. The pressure may then be obtained using the equation of state
p = p(T, ρ), and the the problem can be solved by using techniques described previously
in this Chapter.

For incompressible flow, there is by definition no connection between pressure and

density. In this case, coupling between pressure and density introduces a constraint on the
solution of the flow field: if the correct pressure field is applied in the momentum equa-
tions (15.52) and (15.53), the resulting velocity field should satisfy continuity (15.54).For
these reasons, solution techniques for incompressible Navier-Stokes equations are differ-
ent than for compressible Navier-Stokes equations. The pressure correction model, which
will be presented next, are one such technique.

15.9.2 The staggered grid

When including the pressure gradient in the calculations, another problem arises. This
is usually illustrated by an example.

Example 241 The, checkerboard pressure field of Figure 15.11 is highly irregular. Let us
examine how this is represented in the discretized momentum equations. If the pressure
drop across the CV is obtained by linear interpolation, we have in the x direction:
pW + pP pP + pE pW − pE
pw − pe = − =
2 2 2
616 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

100 50 100 50 100 50 100 50

50 100 50 100 n 50 100 50 100

w e y
100 50 100 50 100 50 100 50
s
50 100 50 100 x 50 100 50 100

100 50 100 50 100 50 100 50

Figure 15.11: Discrete checkerboard pressure field

which implies
∂p pw − pe pW − pE 100 − 100
= = = =0 (15.56)
∂x δx 2δx 2δx
and similarly in the y direction
∂p pN − pS 100 − 100
= = =0 (15.57)
∂y 2δy 2δy
We see that the value at the node P is cancelled out and not used in either of the cal-
culations (15.56) or (15.57). This means that the pressure differences are calculated for
alternate nodes, and not for adjacent ones. This might reduce the accuracy of the solu-
tion. However this is not the main problem: far more serious is the fact that a pressure
field like the one in Figure 15.11 results in all the discretized gradients being zero at the
nodal point. As a result, this pressure field would give the same momentum source as a
uniform field.

Remark 10 The problem described in Example 241 does not occur if the flow is com-
pressible. The term ∂ρ
∂t in the continuity equation (15.55) would damp out the checker-
board pattern.

It is clear from the Example 241 that if velocities are defined at the scalar grid
nodes, the pressure gradients are not properly represented in the discretized momentum
equations. A remedy for this is the staggered grid. The idea is to evaluate velocities
on another grid than all the other variables. Pressure, temperature etc. are calculated
at the ordinary nodal points, but velocities are calculated at the CV faces between the
nodal points. A 2D staggered grid is shown in Figure 15.12. The x direction velocity u is
calculated at the faces that are normal to the x direction, and the y direction velocity v is
calculated at the faces that are normal to the y direction. The staggered grid arrangement
solves the problem of the checkerboard pressure field. Alternatively, (Anderson 1995),
suggest to use another approximation when calculating the pressure differences in (15.56)
and (15.57). The upwind scheme might be one such scheme.

Remark 11 Although example 241 illustrates that the staggered grid solves the problem
of the checkerboard pressure field, it is rather unrealistic because it contains a nonphysical
15.9. PRESSURE-VELOCITY COUPLING 617

pressure field. The main advantage of the staggered grid is in fact that it allows for the
pressure difference between two adjacent grid points to be the driving force for the velocity
component located between these grid points.

n
P
w e
s

Figure 15.12: A staggered two dimensional grid. The u’s are stored at the −→, the v’s
at the ↑ and other variables φ at the •.

15.9.3 The momentum equations

P e E

Figure 15.13: Control volume for u.

A staggered control volume for the momentum equation (15.52) is shown in Figure
15.13. As can be seen, the pressure difference pP −pE can be used to calculate the pressure
force acting on the control volume for the velocity u. Using similar techniques as in the
previous sections, the discretized transient momentum equation in the x direction on the
staggered grid can be written in the form
X
ae un+1
e = anb unnb + a0e une + b + (pnP − pnE ) Ae (15.58)

where
∆Vu
Au =
∆x
and b is a collection of other source terms than the pressure gradient. In (15.58), inte-
gration with respect to time has been carried out by an explicit scheme such as the Euler
method. The values of the coefficients a0e , ae and anb may be calculated by any of the
618 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

previous presented methods suitable for convection-diffusion problems (hybrid, upwind,

QUICK).
Similarly, the discretized momentum equation in the y direction on the staggered grid
can be written in the form
X
an vnn+1 = n
anb vnb + a0n vnn + b + (pnP − pnN ) An (15.59)

where
∆Vv
An =
∆y
The extension to the z direction in a 3D problem is trivial.

Remark 12 Notice that the term (pP − pE ) Au in (15.58) is the pressure force acting
on the u control volume. This is physically correct and would not be possible without the
staggered grid.

15.9.4 The transient SIMPLE algorithm

SIMPLE (Patankar and Spalding 1972) stands for Semi-Implicit Method for Pressure-
Linked Equations. The complete algorithm will be described below. Solving the mo-
mentum equations requires knowledge of the pressure field, and using an incorrect field
results in a velocity field not satisfying continuity. We will make an initial guess p∗ of the
pressure field, and an imperfect velocity field based on this guessed pressure field will be
denoted by u∗ , v ∗ , w∗ . Using (15.58), the u∗ component will result from the solution of
X
ae (u∗ )n+1
e = anb (u∗ )nnb + a0e (u∗ )ne + b + ((p∗ )nP − (p∗ )nE ) Ae (15.60)

n+1
where the notation (u∗ )e is used for expressing the velocity u∗ at position e and
time n + 1. Similar equations is solved for v ∗ and w∗ . Now we define the correction p0 as

p = p∗ + p0

and similarly for the velocities

u = u∗ + u0 , v = v ∗ + v 0 , w = w∗ + w0

Subtracting (15.60) from (15.58) gives

n+1
X n n ¡ n n¢
ae (u0 )e = anb (u0 )nb + a0e (u0 )e + (p0 )P − (p0 )E Ae

At this time we set

X n
anb (u0 )nb = 0 (15.61)
n
(u0 )e = 0

This is the main approximation of the SIMPLE algorithm and it results in the velocity-
correction formula
n+1 ¡ n n¢
(u0 )e = de (p0 )P − (p0 )E
¡ n n¢
un+1
e = (u∗ )n+1
e + de (p0 )P − (p0 )E (15.62)
15.9. PRESSURE-VELOCITY COUPLING 619

where
Ae
de = ,
ae
and ae = f (∆t; ∆x) depending on the numerical scheme used. The reason behind the
approximation (15.61) is that we want to derive a equation for pressure correction p0 .
We construct a formula for p0 that ensures convergence of the velocity field to a solution
that satisfies continuity. As p0 is a numerical artifice, and there is no reason to expect
that the formula for predicting p0 from one step to the next is physical correct. The
approximation is analyzed in (Patankar 1980) and (Anderson 1995).
The above procedure can also be carried out in the y direction, which would lead to
the velocity-correction formula
n+1 ¡ n n¢
vnn+1 = (v ∗ )n + dn (p0 )P − (p0 )N (15.63)
We will now use the continuity equation (15.54) as a pressure correction equation. In-
tegrating the continuity equation over the control volume centered at P and using the
central difference for discretizing in x and y directions, we get
un+1
e − un+1
w un+1 − un+1
ρ +ρ n s
=0 (15.64)
∆x ∆y
when evaluating the time invariant equation at time t + ∆t. Inserting (15.62) and (15.63)
into (15.64) it follows that
n n n n n
aP (p0 )P = aE (p0 )E + aW (p0 )W + aN (p0 )N + aS (p0 )S + b0 (15.65)
where the term
(u∗ )n+1 − (u∗ )n+1 (u∗ )n+1 − (u∗ )n+1
b0 = e w
+ n s
∆x ∆y
is the continuity imbalance arising from the incorrect velocity field, and
dw + de dn + ds dw de ds dn
aP = + , aW = − , aE = − , aS = − , aN = −
∆x ∆y ∆x ∆x ∆y ∆y
The complete SIMPLE algorithm will now be stated for all three dimension.
1. Guess the pressure field p∗
2. Solve the momentum equations for u∗ , v ∗ , w∗
3. Solve the pressure correction equation for p0
4. Calculate p and u, v, w from p∗ and u∗ , v ∗ , w∗
5. Solve the discretized equations for all other φ.(If a φ does not influence the flow
field it is better to calculate it after the flow field calculations have converged)
6. If the solutions has converged: end, otherwise set p∗ = p, and start over from step
2.
Remark 13 The term semi-implicit in the name of the algorithm stems from the approx-
imation (15.61). Without the approximation, the complete pressure correction field would
have been coupled in one equation, giving a fully implicit equation. The approximation
allows the pressure correction equation (15.65) to include terms only from the neighbor
nodes, and it was termed only as semi-implicit by (Patankar and Spalding 1972).
Refinements to the SIMPLE algorithm in terms of computations effort and stability
have produced algorithms such as SIMPLER, SIMPLEC and PISO. For an overview
consult (Versteeg and Malalasekera 1995).
620 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

15.10 Von Neuman stability method

Although convergence might be difficult to establish, there exist a quite simple method
to establish the stability of partial difference equations. Named after its inventor, the
von Neuman method will be presented here. First the general idea will be presented, and
then the detail will be made clear through a couple of examples.
Let D be the exact solution of the difference equation in question, and N be the
corresponding numerical solution. The round off error, or hereafter the error, is then
defined as
ε=N −D ⇒N =D+ε (15.66)
By showing that ε also satisfy the difference equation, it follows that the numerical
solution N is stable if εi shrinks as the solution progresses from step n to step n + 1.
That is, the solution will be stable if
¯ n+1 ¯
¯ εP ¯
¯ ¯
¯ εn ¯ ≤ 1 (15.67)
P

and unstable otherwise. How to represent the error ε is shown in the example below.

Example 242 Consider the one dimensional heat conduction equation

∂T ∂2T
=α 2
∂t ∂x
∂T ∂2T
By representing ∂t with a forward difference, and ∂x2 with a central difference we get

TPn+1 − TPn n
α (TW − 2TPn + TEn )
= (15.68)
∆t (∆x)2
Let D be the exact solution of the difference equation (15.68), and N be the numerical
solution. The error, is defined as

ε = N − D ⇒ N = D + ε.

The numerical solution N must satisfy the difference equation:

DPn+1 + εn+1 − DPn − εnP Dn + εnW − 2DPn − 2εnP + DE

n
+ εnE
P
= W
α∆t (∆x)2
and as D also satisfy (15.68), so must the error ε :

εn+1 − εnP εn − 2εnP + εnE

P
= W (15.69)
α∆t (∆x)2
Equation (15.69) is stable if ¯ n+1 ¯
¯ εP ¯
¯ ¯
¯ εn ¯ ≤ 1 (15.70)
P
The error ε can be written as a Fourier series:
X X
ε(x, t) = Am (t)eikm x = Am (t) (cos km x + i sin km x) (15.71)
m m
15.10. VON NEUMAN STABILITY METHOD 621

√
where i = −1,
2π
km =
λ
is the wave number and m is to be determined. It is assumed that the length of the
domain on which the equation is solved is L. By using N + 1 grid points, we have that

L
∆x =
N
The smallest allowable wavelength in the Fourier series (15.71) is

2L
λmin =
N
which is the wavelength of a sine (or cosine) function having all three zeros in adjacent
gridpoints. Thus the highest wave number in the series is

2π 2π N
km,max = =
λmin L 2

which gives us the summation limits for (15.71):

N/2 N/2
X X
ikm x
ε(x, t) = Am (t)e = Am (t) (cos km x + i sin km x)
m=1 m=1

where µ ¶
2π
km = m
L
It is further assumed that the amplitude Am varies with time as Am (t) = eat ,where a is
a constant. This implies
N/2
X
ε(x, t) = eat eikm x
m=1

By substituting one term

εm (x, t) = eat eikm x (15.72)
into (15.69), we get

ea(t+∆t) eikm x − eat eikm x eat eikm (x+∆x) − 2eat eikm x + eat eikm (x−∆x)
= 2
α∆t (∆x)

which is simplified to

α∆t ¡ ikm ∆x ¢
ea∆t = 1+ 2 e + e−ikm ∆x − 2
(∆x)
2α∆t
= 1+ (cos (km ∆x) − 1)
(∆x)2
µ ¶
4α∆t 2 km ∆x
= 1− 2 sin (15.73)
(∆x) 2
622 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

By combining (15.70), (15.72) and (15.73) we get

¯ n+1 ¯ ¯ a(t+∆t) ik x ¯
¯ εP ¯ ¯ e m ¯¯ ¯¯ a∆t ¯¯
¯ ¯ = ¯e
¯ εn ¯ ¯ eat eikm x ¯ = e
P
¯ µ ¶¯
¯ 4α∆t km ∆x ¯¯
¯ 2
= ¯1 − sin ¯≤1
¯ (∆x)2 2 ¯

which is the stability criterion. The factor

¯ µ ¶¯
¯ 4α∆t km ∆x ¯¯
¯
G , ¯1 − 2 sin
2
¯
¯ (∆x) 2 ¯

is known as the amplification factor. The condition G ≤ 1 has two solutions, where the
first is trivial, and the other leads to
µ ¶
4α∆t 2 km ∆x
1− 2 sin ≥ −1
(∆x) 2

which is simplified to
α∆t 1
2 ≤
(∆x) 2
which is the stability requirement for the difference equation (15.68) to be stable.

The exact form of the stability criterion G ≤ 1 depends on the form of the difference
equation, that is both the original PDE and the discretization methods used. This is
illustrated in the next example.

Example 243 Given the partial differential equation for one dimensional convection (or
the one dimensional wave equation)

∂φ ∂φ
+c =0
∂t ∂x
∂φ ∂φ
Using a forward difference for ∂t and Upwind difference for ∂x :

φn+1 − φnP (φn − φnE )

P
+c P =0
∆t ∆x
which can be simplified to
φn+1
P = (1 − C) φnP + CφnE (15.74)
where
∆t
C=c
∆x
is called the Courant number. The von Neuman stability analysis applied to (15.74),
using
εm (x, t) = eat eikm x
gives
ea(t+∆t) eikm x = (1 − C) eat eikm x + Ceat eikm (x−∆x)
15.10. VON NEUMAN STABILITY METHOD 623

and the amplification factor is given by

εn+1
G = P
= eat
εnP
= (1 − C) + Ceat eikm ∆x
= 1 − C (1 − cos km ∆x) − iC sin km ∆x

We now demand that

q
|G| = (1 − C (1 − cos km ∆x))2 + (C sin km ∆x)2 ≤ 1

which after some elementary trigonometric manipulations gives

q
C 2 + (C. − 1)2 ≤ 1 ⇒ C ≤ 1

The above is an example of a more general stability result known as the Courant-
Friedrichs-Lewy (CFL) condition. This condition, that is

∆t
C=c ≤1 (15.75)
∆x

applies generally to explicit schemes for hyperbolic equations. Physically, the CFL condi-
tion indicates that for stability, a particle of fluid should not travel more than one spatial
step-size ∆x in one time step ∆t.
The CFL condition is also the stability condition for the second order wave equation

∂2u 2
2∂ u
− c =0 (15.76)
∂t2 ∂x2
studied in Example 237. There is a connection between the characteristic lines of a
hyperbolic equation and the CFL condition. The following presentation is based on
(Anderson 1995). The characteristic lines
½
ct (right-running)
x=
−ct (left-running)

for (15.76) are plotted in Figure 15.14. In both figures 15.14 a) and b) point b is the
intersection of the right-running characteristic through grid-point W and the left-running
characteristic through grid point E. This point also has a connection to the CFL condi-
tion. Let ∆tC=1 denote the value of ∆t given by (15.75) when C = 1, that is
∆x
∆tC=1 =
c
In Figure 15.14 a) and b) ∆tC=1 is exactly the distance between point P and point b given
by the intersection of the characteristics. Now, study Figure 15.14 a) and assume C < 1.
Then, ∆tC<1 < ∆tC=1 . Let point d correspond to the grid point directly above point P
existing at time t+∆tC<1 . Since properties at point d are calculated numerically from the
difference equation using information at grid points E and W , the numerical domain for
point d is the triangle adc. The numerical domain is denoted Dn . The analytical domain
for d is the shaded area defined by the characteristics through d. The analytical domain
624 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS

is denoted Da . Note that in this case, the numerical domain contains the analytical
domain.
Now, consider Figure 15.14 b). Then C > 1 and ∆tC>1 > ∆tC=1 . Let point d
correspond to the grid point directly above point P existing at time t+∆tC>1 . As before,
the numerical domain in this case is the triangle adc, and the analytical domain is defined
by the characteristics through d. Note that in this case, the numerical domain does not
contain all of the analytical domain. The case in 15.14 b) considered an unstable solution
as C > 1, and we can now state the following interpretation of the CFL condition: For
stability, the numerical domain must include all of the analytical domain, that is

Da ⊂ Dn

Figure 15.14 a) can also illustrate accuracy. The analytical domain for point d is the
shaded triangle. Note however, that the numerical grid points E and W are outside the
domain of dependence of d and should therefore theoretically not influence the properties
at d. On the other hand, the numerical calculation of the properties at d takes information
from E and W into account. This leads to inaccurate results as there is a mismatch
between the domain of dependence and the actual numerical data used. In view of
accuracy it is therefore desirable to have the Courant number as close as possible to
unity.
15.10. VON NEUMAN STABILITY METHOD 625

d Da

t C1

x
ct



t C1

c
Dn
x

t
a c
t
W P E

x
x
(a)
t

Dn
b
x

t C1

ct

c


t
x

t C1 Da

a c
t
e W P E f

x

x
(b)

Figure 15.14: (a) Stable (b) Unstable

626 CHAPTER 15. COMPUTATIONAL FLUID DYNAMICS
Bibliography

Aamo, O. and Fossen, T.: 2000, Finite element modelling of mooring lines, Mathematics
and Computers in Simulation 53, 415 — 422.

Aamo, O. and Krstić, M.: 2003, Flow Control by Feedback. Stabilization and Mixing,
Springer-Verlag, London.

Anderson, B. and Moore, J.: 1989, Optimal Control. Linear Quadratic Methods, Prentice-
Hall, Englewood Cliffs.

Anderson, B. and Vongpanitlerd, S.: 1973, Network Analysis and Synthesis, Prentice-
Hall, Englewood Cliffs, New Jersey.

Anderson, J.-D.: 1995, Computational fluid dynamics, McGraw-Hill, New York.

Angeles, J.: 1988, Rational Kinematics, Springer-Verlag, New York.

Antsaklis, P. and Michel, A.: 1997, Linear Systems, McGraw-Hill, New York.

Arimoto, S.: 1996, Control Theory of Nonlinear Mechanical Systems: A Passivity-Based

and Circuit-Theoretic Approach, Oxford University Press, Oxford.

Aris, R.: 1989, Vectors, tensors, and the basic equations of fluid mechanics, Dover Pub-
lications, New York.

Armstrong-Hélouvry, B.: 1990, Stick-slip arising from stribeck friction, Proceedings of the
1990 International conference on robotics and automation, Cincinatti, OH, pp. 1377—
1382.

Armstrong-Hélouvry, B., Dupont, P. and Canudas de Wit, C.: 1994, A survey of models,
analysis tools and compensation methods for the control of machines with friction,
Automatica 30(7), 1083—1138.

Arnold, V.: 1989, Mathematical Methods of Classical Mechanics, 2nd edn, Springer-
Verlag, New York.

Aström, K. and Furuta, K.: 2000, Swinging up a pendulum by energy control, Automatica
36(2), 287—295.

Barabanov, N. and Ortega, R.: 2000, Necessary and sufficient conditions for passivity of
the LuGre friction model, IEEE Transactions on Automatic Control 45(4), 830—832.

Bathe, K.-J.: 1996, Finite Element Procedures, Prentice-Hall, Englewood Cliffs, New
Jersey.

627
628 BIBLIOGRAPHY

Bird, R., Stewart, W. and Lightfoot, E.: 1960, Transport Phenomena, John Wiley, New
York.
Blanke, M.: 1981, Ship Propulsion Losses Related to Automatic Steering and Prime
Mover Control, PhD thesis, Danmarks tekniske højskole.
Bremer, H.: 1988, über eine zentralgleichung in der dynamik, Z. angew. Math. Mech.
68, 307—311.
Canudas de Wit, C., Olsson, H., Åström, K. and Lischinsky, P.: 1995, A new model
for control of systems with friction, IEEE Transactions on Automatic Control
40(3), 419—424.
Chen, C.: 1999, Linear System Theory and Design, Oxford University Press, Oxford.
Cohen, H., Rogers, G. and Saravanamuttoo, H.: 1996, Gas turbine theory, 4th edn,
Longman, Essex.
Crandall, S., Karnopp, D., E.F. Kurtz, J. and Pridmore-Brown, D.: 1968, Dynamics of
Mechanical and Electromechanical Systems, McGraw-Hill, New York.
Cumpsty, N.: 1989, Compressor Aerodynamics, Longman.
Dahl, P.: 1968, A solid friction model, Technical Report TOR-0158(3107-18)-1, The
Aerospace corporation, El Segundo, Calif. 90245.
Dahl, P.: 1976, Solid friction damping of mechanical vibrations, AIAA Journal
14(12), 1675—1682.
de Groot, S. and Mazur, P.: 1984, Non-Equilibrium Thermodynamics, Dover Publica-
tions, New York.
der Waerden, B. V.: 1976, Hamilton’s discovery of the quaternions, Mathematics Maga-
zine 5, 227 — 234.
Dorf, R. and Bishop, R.: 2000, Modern Control Systems, 9th edn, Addison-Wesley,
Reading, Massachusetts.
Dormand, J. and Prince, P.: 1986, Runge-kutta triplets, Comp. and Maths. with Applc.
12A, 1007—1017.
Dupont, P., Hayward, V., Armstrong, B. and Altpeter, F.: 2002, Single state elastoplastic
friction models, IEEE Transactions Automatic Control 47(5), 787 —792.
Egeland, O. and Godhavn, J.-M.: 1994, Passivity-based attitude control of a rigid space-
craft, IEEE Transactions on Automatic Control 39, 842—846.
Ellman, A. and Piché, R.: 1999, A two regime orifice flow formula for numerical simula-
tion, Journal of Dynamic Systems, Measurement, and Control 121(4), 721 — 724.
Ervik, M.: 1971, Regulatorer for Vannkraft Makiner, Universitetsforlaget, Oslo.
Evans, L.: 1998, Partial Differential Equations, Graduate Studies in Mathematics, Amer-
ican Mathematical Society, Providence, Rhode Island.
Farell, J. and Barth, M.: 1999, The Global Positioning System and Inertial Navigation,
McGraw-Hill, New York.
BIBLIOGRAPHY 629

Ferziger, J. and Perić, M.: 1999, Computational Methods for Fluid Dynamics, second
edn, Springer Verlag, Heidelberg.

Fitzgerald, A., Kingsley, C. and Umans, S.: 1983, Electric Machinery, 4th edn, McGraw-
Hill, New York.

Fossen, T.: 1994, Guidance and Control of Ocean Vehicles, John Wiley & Sons, Chich-
ester.

Fossen, T.: 2002, Marine Control Systems: Guidance, Navigation and Control of Ships,
Rigs and Underwater Vehicles, Marine Cybernetics, Trondheim, Norway.

Fuller, C., Elliott, S. and Nilsen, P.: 1996, Active Control of Vibration, Academic Press,
London.

Futral, S. and Wasserbauer, C.: 1965, Off design performance prediction with experi-
mental verification for a radial-inflow turbine, Technical Report TN D-2621, NASA.

Gavronski, W., Beech-Brandt, J., Ahlstrom, H. and Manieri, E.: 2000, Torque-bias
profile for improved tracking of deep space network antennas, IEEE Antennas and
Propagation magazine 42, 35 — 45.

Gevarter, W.: 1970, Basic relations for control of flexible vehicles, AIAA Journal
8(4), 666 — 672.

Goldstein, H.: 1980, Classical Mechanics, Addison Wesley, Reading, Massachusetts.

Golub, G. and van Loan, C.: 1989, Matrix computations, 2nd edn, The John Hopkins
University Press, Baltimore, Maryland.

Goodson, R. and Leonard, R.: 1972, A survey of modeling techniques for fluid line
transients, Trans. of the ASME. Journal of Basic Engineering. pp. 474 —482.

Gravdahl, J. and Egeland, O.: 1998, Two results on compressor surge control with
disturbance rejection, Proceedings of the 37th IEEE Conference on Decision and
Control. To appear.

Gravdahl, J. and Egeland, O.: 1999, Compressor surge and rotating stall: modeling and
control, Advances in Industrial Control, Springer-Verlag, London.

Gravdahl, J. and Egeland, O.: 2002, Drive torque actuation in active surge control of
centrifugal compressors, Automatica 38(11), 1881—1893.

Gravdahl, J., Egeland, O. and Vatland, S.: 2001, Active surge control of centrifugal
compressors using drive torque, Proceedings of 40th IEEE Conference on Decision
and Control, pp. 1286 —1291.

Greitzer, E.: 1976, Surge and Rotating stall in axial flow compressors, Part I: Theoretical
compression system model, Journal of Engineering for Power 98, 190—198.

Haessig, Jr., D. and Friedland, B.: 1991, On the modeling and simulation of friction,
ASME J. of Dynamic Systems, Measurement and Control 113, 354—362.

Hairer, E.: 1999, Numerical geometric integration, Report found on

http://www.unige.ch/math/folks/hairer/polycop.html.
630 BIBLIOGRAPHY

Hairer, E., Nørsett, S. and Wanner, G.: 1993, Solving ordinary differential equations I:
Nonstiff problems, 2nd edn, Springer-Verlag Berlin.

Hairer, E. and Wanner, G.: 1996, Solving ordinary differential equations II: Stiff and
differential-algebraic problems, 2nd edn, Springer-Verlag Berlin.

Hartman, P.: 1982, Ordinary Differential Equations, Elsevier, New York.

Hess, D. and Soom, A.: 1990, Friction at a lubricated line contact operating at oscillating
sliding velocities, ASME Journal of Tribology 112(1), 147—152.

Heywood, J.: 1988, Internal combustion engine fundamentals, McGraw-Hill.

Holmboe, E. and Rouleau, W.: 1967, The effect of viscous shear on transients in liquid
lines, Journal of Basic Engineering, Trans. ASME, Series D 89, 174 — 180.

Hughes, P.: 1974, Dynamics of flexible space vehicles with active attitude control, Ce-
lestial Mechanics 9, 21—39.

Hughes, P.: 1986, Spacecraft Attitude Dynamics, John Wiley, New York.

Hutarew, G.: 1969, Regelungstechnik : Kurze Einführung Am Beispiel der

Drehzahlregelung Von Wasserturbinen, Springer, Berlin.

IEEE: 1987, IEEE standard on piezoelectricity, Technical Report Standard 176-1987,

ANSI/IEEE.

Isidori, A.: 1989, Nonlinear Control Systems, 2nd. edn, Springer-Verlag, Berlin.

Joshi, S.: 1989, Control of Large Flexible Space Structures, Vol. 131 of Lecture Notes in
Control and Information Science, Springer-Verlag, Berlin.

Kane, T. and Levinson, D.: 1983, The use of Kane’s dynamical equations in robotics,
Int. J. Robotics Research 2, 3—21.

Kane, T. and Levinson, D.: 1985, Dynamics: Theory and Applications, McGraw-Hill,
New York.

Kane, T., Likins, P. and Levinson, D.: 1983, Spacecraft Dynamics, McGraw-Hill, New
York.

Karnopp, D.: 1985, Computer simulation of stick-slip friction in mechanical synamic

systems, ASME J. of Dynamic Systems, Measurement and Control 107, 100—104.

Karnopp, D., Margolis, D. and Rosenberg, R.: 2000, System Dynamics. Modeling and
Simulation of Mechatronic Systems, 3rd. edn, Wiley-Interscience, New York.

Kelkar, A. and Joshi, S.: 1996, Control of Nonlinear Multibody Flexible Space Structures,
Springer-Verlag, London.

Kelly, R., Carelli, R. and Ortega, R.: 1989, Adaptive motion control design of
robot manipulators: An input-output approach, International Journal of Control
50(6), 2563—2581.

Khalil, H.: 1996, Nonlinear systems, 2nd edn, Prentice-Hall, Inc.

BIBLIOGRAPHY 631

Kiencke, U. and Nielsen, L.: 2000, Automotive Control Systems for Engine, Driveline,
and Vehicle, Springer, Berlin.

Kreyszig, E.: 1979, Advanced Engineering Mathematics, 4th edn, Wiley, New York.

Kristiansen, E.: 2000, Energy-based observer design for manipulators with flexible links,
Master’s thesis, Norwegian University of Science and Technology.

Krstić, M., Kanellakopoulos, I. and Kokotović, P.: 1995, Nonlinear and Adaptive Control
Design, Wiley, New York.

Kuo, B.: 1995, Automatic Control Systems, 7th edn, Prentice-Hall, Englewood Cliffs,
New Jersey.

Lamb, H.: 1945, Hydrodynamics, Dover Publications, New York.

Lambert, J.: 1991, Numerical methods for ordinary differential equations. The initial
value problem, John Wiley, Chichester.

Lanczos, C.: 1986, The Variational Principles of Mechanics, Dover Publications, New
York.

Leonard, B.: 1979, A stable and accurate convective modelling procedure based on
quadratic upstream interpolation, Computer Methods in Applied Mechanics and En-
gineering 19(1), 59—98.

Leonard, N.: 1997, Stability of bottom-heavy underwater vehicle, Automatica 33(3), 331—
346.

Leonhard, W.: 1996, Control of Electrical Drives, 2nd edn, Springer, Berlin.

Lewis, D. and Simo, J.: 1994, Conserving algorithms for the dynamics of hamiltonian
systems on lie groups, J. Nonlinear Science 4, 253 — 299.

Lin, C. and Segel, L.: 1974, Mathematics applied to deterministic problems in the natural
sciences, Macmillan, New York.

Lohmiller, W. and Slotine, J.: 1998, On contraction analysis for nonlinear systems,
Automatica 34, pp. 683—696.

Lovelock, D. and Rund, H.: 1989, Tensors, Differential Forms, and Variational Princi-
ples, Dover Publications, New York.

Lozano, R., Brogliato, B., Egeland, O. and Maschke, B.: 2000, Dissipative Systems
Analysis and Control. Theory and Applications, Springer-Verlag, London.

Luh, J., Walker, M. and Paul, R.: 1980, On-line computational scheme for mechanical
manipulators, ASME J. Dynamic Syst., Meas., Contr. pp. 69—76.

Mäkinen, J., Piché, R. and Ellman, A.: 2000, Fluid transmission line modeling using
a variational method, J. Dynamic Systems, Measurement, and Control 122, 153 —
162.

Marsden, J. and Ratiu, T.: 1994, Introduction to Mechanics and Symmetry, Springer-
Verlag, New York.
632 BIBLIOGRAPHY

McCarthy, J.: 2000, Geometric Design of Linkages, Springer-Verlag, New York.

Meirovitch, L.: 1967, Analytical Methods in Vibrations, Mamillan, New York.

Meirovitch, L.: 1980, Computational Methods in Structural Dynamics, Sijthoff & Noord-
hoff, Alphen aan den Rijn, The Netherlands.

Meisel, J.: 1966, Principles of Electromechanical Energy Conversion, McGraw-Hill, New

York.

Merritt, H.: 1967, Hydraulic Control Systems, John Wiley, New York.

Milne-Thomson, L.: 1996, Theoretical Hydrodynamics, Dover Publications, New York.

Mohan, N., Undeland, T. and Robbins, W.: 1989, Power Electronics: Converters, Ap-
plications and Design, John Wiley, New York.

Moore, F. and Greitzer, E.: 1986, A theory of post-stall transients in a axial compressor
systems: Part I—Development of equations, Journal of Engineering for Gas Turbines
and Power 108, 68—76.

Murray, R., Li, Z. and Sastry, S.: 1994, A Mathematical Introduction to Robotic Manip-
ulation, CRC Press, Boca Raton.

Nicklasson, P.: 1996, Passivity-based control of electric machines, PhD thesis, Norwegian
Institute of Technology, Dept of Engineering Cybernetics.

Niemeyer, G. and Slotine, J.: 1991, Stable adaptive teleoperation, IEEE J. Oceanic
Engineering 16(1).

Nijmeijer, H. and der Schaft, A. V.: 1990, Nonlinear Dynamical Control Systems,
Springer-Verlag, Berlin.

Nilsson, J.: 1983, Electric Circuits, Addison-Wesley, Reading, Massachusetts.

Ortega, R., Loria, A., Nicklasson, P. and Sira-Ramirez, H.: 1998, Passivity-Based Control
of Euler-Lagrange Systems: Mechanical, Electrical and Electromechanical Applica-
tions, Springer-Verlag.

Patankar, S.: 1980, Numerical Heat Transfer and Fluid Flow, Hemisphere Publishing
Corporation.

Patankar, S. and Spalding, D.: 1972, A calculation procedure for heat, mass and mo-
mentum transfer in three-dimensional parabolic flows, Int. J. Heat Mass Transfer
15, 1787—1806.

Pazy, A.: 1983, Semigroups of Linear Operators and Applications to Partial Differential
Equations, Vol. 44 of Applied Mathematical Sciences, Springer-Verlag, New York.

Piché, R. and Ellman, A.: 1996, A Fluid Transmission Line Model for Use with ODE
Simulators, Research Studies Press, Somerset, England, pp. 221 — 236.

Rao, S.: 1990, Mechanical Vibrations, Addison-Wesley, Reading, Massachusetts.

Robertson, R. and Schwertassek, R.: 1988, Dynamics of Multibody Systems, Springer-

Verlag, Berlin.
BIBLIOGRAPHY 633

Rugh, W.: 1996, Linear System Theory, 2nd edn, Prentice-Hall, Englewood Cliffs.
Sagatun, S. and Fossen, T. I.: 1991, Lagrangian formulation of underwater vehicles’
dynamics, Proc. IEEE Int. Conf. Systems, Man and Cybernetics, pp. 1029—1034.
Samson, C., Borgne, M. L. and Espiau, B.: 1991, Robot Control: The Task Function
Approach, Vol. 22 of Oxford Engineering Science Series, Oxford University Press,
Oxford.
Sanz-Serna, J. and Calvo, M.: 1994, Numerical Hamiltonian problems, Chapman and
Hall, London.
Scarborough, J.: 1966, Numerical Mathematical Analysis, sixth edn, John Hopkins Press,
Baltimore.
Sciavicco, L. and Siciliano, B.: 2000, Modeling and Control of Robot Manipulators,
Springer-Verlag, London berlin Heidelberg.
Sepulchre, R., Janković, M. and Kokotović, P.: 1997, Contructive Nonlinear Control,
Springer-Verlag, London.
Shahruz, S.: 1999, Boundary control of kirchhoff’s non-linear string, Int. Journal of
Control 72(6), 560 — 563.
Shampine, L.: 1994, Numerical Solution of Ordinary Differential Equations, Chapman
and Hall, New York.
Shampine, L., Allen, R. and Preuss, S.: 1997, Fundamentals of numerical computing,
John Wiley, New York.
Shampine, L. and Reichelt, M. W.: 1997, The matlab ode suite, SIAM Journal on
Scientific Computing 18(1). also in MATLAB Helpdesk, Online Manuals (in PDF).
Shepperd, S.: 1978, Quaternion from rotation matrix, J. Guidance and Control 1, 223—
224.
Skogestad, S. and Postlethwaite, I.: 1996, Multivariable Feedback Control. Analysis and
Design, John Wiley, Chichester.
Slotine, J.: 1991, Applied Nonlinear Control, Prentice-Hall, Englewood Cliffs, New Jersey.
Slotine, J. and Li, W.: 1988, Adaptive manipulator control: A case study, IEEE Trans-
actions on Automatic Control 33, 995—1003.
Spong, M. and Vidyasagar, M.: 1989, Robot Dynamics and Control, John Wiley, New
York.
Stecki, J. and Davis, D.: 1986a, Fluid transmission lines - distributed parameter models.
part 1: A review of the state of the art, Proc. Instn. Mech. Engrs. 200, 215 — 228.
Stecki, J. and Davis, D.: 1986b, Fluid transmission lines - distributed parameter models.
part 2: Comparison of models, Proc. Instn. Mech. Eng. 200, 229 — 236.
Strang, G.: 1988, Linear Algebra and its Applications, Saunders.
Stribeck, R.: 1902, Die wesentlichen Eigenschaften der Gleit- und Rollenlager, Zeitschrift
des Vereines Deutcher Ingenieure 46(39), 1432—37.
634 BIBLIOGRAPHY

Szebehely, V.: 1967, Theory of orbits, Academic press, New York.

Takegaki, M. and Arimoto, S.: 1981, A new feedback method for dynamic control of
manipulators, ASME J. Dyn. Syst. Meas. Control 102, 119—125.
Titterton, D. and Weston, J.: 1997, Strapdown Inertial Navigation Technology, Peter
Peregrinus on behalf of IEEE, London.
Triantafyllou, M.: 1990, Cable mechanics with marine applications, Lecture notes, MIT.
Vas, P.: 1990, Vector Control of AC Machines, Oxford University Press, Oxford.
Versteeg, H. and Malalasekera, W.: 1995, An Introduction to Computational Fluid Dy-
namics: The Finite Volume Method, Longman, Essex.
Walker, M. and Orin, D.: 1982, Efficient dynamic computer simulation of robotic mech-
anisms, Journal of Dynamic Systems, Measurement and Control 104, 205 — 211.
Watton, J.: 1989, Fluid Power Systems. Modeling, Simulation, Analog and Microcom-
puter Control, Prentice-Hall, New York.
Weaver, W., Timoshenko, S. and Young, D.: 1990, Vibration Problems in Engineering,
5th Ed., John Wiley & Sons, New York.
Wen, J.-Y. and Kreutz-Delgado, K.: 1991, The attitude control problem, IEEE Trans-
actions on Automatic Control 36, 1148—1162.
White, F.: 1999, Fluid mechanics, 4th edn, McGraw-Hill, New York.
Wie, B. and Bryson, A.: 1987, Pole-zero modeling of flexible space structures, Journal
of Guidance, Control and Dynamics 11(6), 554 — 561.
Woods, R.: 1983, A first-order square-root approximation for fluid transmission lines,
in M. Franke and T. Drzewiecki (eds), Fluid Transmission Line Dynamics 1983,
ASME, pp. 37—49.
Yang, W. and Tobler, W.: 1991, Dissipative model approximation of fluid transmission
lines using linear friction model, ASME Journal of Dynamic Systems, Measurement
and Control 113, 152—162.
Index
A-stability, 546, 550 Coenergy, 103, 104
Adams methods Collocation, 361
Explicit, 576 Complemetary sensitivity function, 15
Implicit, 578 Componets, 210
Adams-Bashforth methods, 576 Composite rotation, 220
Adams-Moulton methods, 578 Compressor characteristic
Adjoint transformation Nondimesional form, 497
Rigid motion, 333 Compressor dynamics, 485
Affine in the control, 4 Computational causality, 26
Algebraic stability, 558 Configuration, 289
Aliasing, 544 Configuration space, 289
Amplification factor, 622 Conservativness, 605
AN-stability, 555 Consistent mass matrix, 371
Analytic, 18 Continuity equation, 417
Angle-axis parameters, 227 Multi-component systems, 422
Angular velocity, 240 Coordinate transformation matrix, 219
Coordinate vector, 210
B-stability, 557 Coordinate vectors
Barycentric material derivative, 423 Differentiation, 242
Barycentric velocity, 422 Coordinates, 210
BDF methods, 580 Corrected mass flow, 503
Bernoulli’s equation Corrected speed, 504
Along a streamline, 429 Coulomb friction, 192
Irrotational, 428 Couple, 265
Bond graphs, 26 Courant number, 622
Bound vector, 263 Courant-Friedrichs-Lewy condition, 623
Boundedness, 605 Curvilinear coordinates, 408
Bulk modulus, 151
Burger’s equation, 425 d’Alembert’s principle, 291
Butcher array, 527 Dahl effect, 191
deformation tensor, 451
Cauchy’s equation of motion, 455 Degrees of freedom, 290
Causality, 41 Denavit-Hartenberg convention, 224
Cayley transformation, 239 Dense outputs, 565
Cayley’s formula, 239 Differential-algebraic equations, 585, 586
Center of gravity, 266 Index 1, 586
Center of mass, 261 Multistep methods, 589
Characteristic impedance, 34 Runge-Kutta method, 587
Christoffel symbols, 320 Diffusion flow, 423
Clausius-Duhem inequality, 468 Dilation, 406
Closed loop transfer function, 14 Direction cosines, 219

635
636 INDEX

Dirichlet condition, 597 FSAL method, 534, 562

Discharge coefficient, 432
Divergence, 406 Gain margin, 44
Divergence thorem, 403 Gauss method, 537, 539, 547, 551, 560,
Dormand-Prince 5(4), 562 570
Driving-point impedance, 26 Gear, 80
Dyadic, 215 Gear ratio, 80
Identity, 216 Generalized coordinates, 289, 314, 340
Inertia, 214 Generalized force, 315
Global error, 517
Elasticity Greitzer surge model, 487, 498
Distributed parameter, 361 Linearization, 499
Lumped parameter, 361 Normalized, 497
Electrical time constant, 88 Greitzer’s B-parameter, 498
Elliptic equations, 594
Helmholtz frequency, 499
Embedded solution, 561
Helmholtz resonator, 476
Energy balance
Hessian matrix, 450
Isentropic processes, 474
Heun’s method, 528
Pressure form, 471
Homogeneous transformation matrix, 223
Temperature form, 470
Hydraulic gear, 185
Energy function, 349
Hydraulic motor
Enthalpy, 443
Pump controlled, 185
Stagnation, 474
Valve controlled, 157
Entropy
Hydrodynamic motor, 141
Material derivative, 468
Hydrostatic motor, 141
Specific, 467
Hyperbolic equations, 594
Entropy equation, 468
Hyrostatic gear, 185
Euler angles, 225
Classical, 226 Ideal gas, 466
Roll-Pitch-Yaw, 225 Identity dyadic, 216
Euler Bernolli beam, 373 Implicit Adams methods, 578
Euler parameters, 231, 571 Implicit Euler method, 535
From rotation matrix, 236 Implicit methods
Euler rotation vector, 237 Numerical solution, 563
Euler’s equation of motion, 424 Implicit mid-point rule, 538, 539, 550
Euler’s method, 521, 528, 570, 574, 578 Improved Euler method, 523, 528, 533
Euler-angle singularity, 247 Impulse response, 12
Euler-Bernoulli beam, 390, 551 Inertia dyadic, 214, 270
Euler-Rodrigues parameters, 238 Inertia matrix, 274
Event detection, 566 Inertia tensor, 274
Explicit Adams methods, 576 Inertial frame, 259
Explicit midpoint rule, 525 Infinite dimensional systems, 18
Internal energy, 465
Fehlberg 4(5), 561 Inverteed pendulum, 281
Field weakening, 119 Inviscid fluid, 424
Field-oriented control, 130 Irrational transfer function, 17
Flow coefficient, 497 Isentropic process, 472
Flux linkage, 102 Isentropic relations, 473
Forces of constraints, 290
Fourier’s law, 446, 470 Jacobi identity, 213
INDEX 637

Jacobian, 519, 565 Neumann condition, 597

Newton search, 564
Kane’s equation of motion, 297, 340 Newtonian fluid, 456, 458
Kinematic differential equations Newtonian frame, 259
Attitude deviation, 245 Noncollocation, 361
Euler angles, 247 Nonconsistent mass matrix, 371
Euler parameters, 248 Normal plane, 254
Euler rotation, 250 Nozzle flow, 480
Euler-Rodrigues parameters, 250 Numerical dissipation, 540
Rotation matrix, 240
Kronecker tensor product, 564 O(.) notation, 518
One-step method, 517
L-stability, 546, 547, 550 Order, 518
Lame coefficients, 456 Osculating plane, 254
Laplacian, 450
Lie bracket Pade approximation, 18
Rotations, 332 Pade approximations, 548
Line of action, 263 Parabolic equations, 594
Linear invariant, 568 Parallel axes theorem, 275
Linear test system, 520 Passive electrical one-port, 67
Linearization, 5, 519 Passivity, 361
Lobatta IIIC, 551 Energy formulation, 63
Lobatto IIIA, 536, 539, 547, 550, 556, PID controller, 65
560, 570, 574 PD controller
Lobatto IIIB, 540, 550, 556, 560, 570, Mechanical analog, 31
574 PECE, 580
Lobatto IIIC, 540, 550, 555, 560, 570, Peclet number, 605
574 Permutation symbol, 211
Local error, 517 Phase margin, 44
Local extrapolation, 561 Piezoelectric actuator, 114
Local solution, 517 Piezoelectricity, 114
Loop transfer function, 14 Pole at infinity, 11
Positive real, 56
Mach number, 480 Pre-whirl, 440
Magnetomotive force (mmf), 101, 122 Predictor-Corrector method, 580
Material control volume, 407 Pressure
Material coordinates, 406 Stagnation, 474
Material derivative, 402 Static, 474
Material volume, 414 Pressure coefficient, 497
Matrix exponential, 230 Principle of virtual work, 290
Mechanical time constant, 88 Proper transfer function, 11
mmf, 101, 122 Pump controlled motor, 185
Modified Euler method, 525, 528
Momentum balance, 423 Quaternion, 232
Momentum vector, 347 Quaternions, 231
Multiport, 21 Identity quaternion, 234
Inverse quaternion, 234
Nabla operator, 402 Quaternion product, 232
Navier-Stokes equation, 459
NDF methods, 581 Radau IA, 539, 560, 570, 574
638 INDEX

Radau IIA, 539, 560, 570, 574 Space marching, 595

Radau methods, 550 Spatial coordinates, 406
Rate of production, 422 Specific energy, 465
Rate of strain tensor, 451, 456 Specific entropy, 467
Rational transfer function, 11 Specific force, 258
Rectifying plane, 254 Specific internal energy, 465
Reduction gear, 80 Specific kinetic energy, 465
Restriction Specific volume, 408, 466
Gas flow, 480 Speed of sound, 475
Liquid flow, 431 Stability
Sonic gas flow, 482 A-stability, 546
Reynolds number, 459 Algebraic stability, 558
Reynolds’ transport theorem, 413 AN-stability, 555
Rigid body, 259 B-stability, 557
RK4, 528, 535, 570, 574 BDF methods, 583
Rosenbrock method, 574 Explicit Adams methods, 582
Stability function, 575 Implicit Adams methods, 582
Rotation dyadic, 229 L-stability, 546
Rotation matrix, 219 Multistep methods, 581
Basic, 221 Nonlinear analysis, 557
Classical Euler angles, 226 Pade approximations, 550
Composite rotations, 221 Stability function, 518, 519, 531
Euler angles, 225 Stage computations, 526
Euler parameters, 231 Stages, 526
Euler rotation vector, 237 Staggered grid, 616
Euler-Rodrigues parameters, 239 Stagnation state, 474
Simple rotation, 221 State space model, 4
Rothalpy, 495 Static pressure, 474
Runge-Kutta method Step size selection, 560
Continuous, 565 Stick-slip friction, 192
Differential-Algebraic, 587 Stiff systems, 535
Dormand-Prince 5(4), 562 Stiffly accurate, 546, 588
Explicit, 526, 560 Stoke’s assumption, 457
Fehlberg 4(5), 561 Stokes’ Theorem, 405
Implicit, 535, 563 Storage function, 63
Inertia matrix, 566 Stress tensor, 453
Stability function, 531 Stribeck curve, 192
Step size selection, 560 Stribeck effect, 195
Symplectic, 571 Strictly proper transfer function, 11
Runge-Kutta methods Summation convention, 449
Property table, 560
Telemanipulation, 70
Sensitivity function, 14 Theta method, 538
Serret-Frenet frame, 253 Time marching, 595
Shock, 484 Torque, 265
Singularity of transfer function, 18 Torsion, 254
Skew-symmetric form, 211 Transfer function
SO(3), 220 Proper, 11
Sonic flow Strictly proper, 11
Restriction, 482 Transmission line, 18
INDEX 639

Electric, 33
Hydraulic, 73
Lossless, 38
Transportiveness, 605
Trapezoidal rule, 537, 539, 547, 550, 556
Twist vector, 246

Undamped natural frequency, 42

Unit quaternion, 233
Unit step function, 12

Valve controlled motor, 157

Variable displacement pump, 185
Variations, 325
Vector, 209
Vectors
Differentiation, 242
Virtual change, 325
Virtual displacement, 290
Viscous stress tensor, 454

Wave equation, 475

d’Alembert’s solution, 475
Wave variables, 37

Zero at infinity, 11
Zero crossing, 566