Université de Liège

Faculté des Sciences Appliquées

Département d'Aérospatiale, Mécanique et Matériaux
Vibrations et Identication des Structures

Integrated Simulation

and Reduced-Order Modeling

of Controlled Flexible Multibody Systems

Thèse de doctorat
présentée en vue
de l'obtention du grade de
Docteur en Sciences Appliquées

par

Olivier BRÜLS
Ingénieur Civil Electro-Mécanicien (Mécatronique - Productique)
Aspirant F.N.R.S.

Février 2005
 Abstract
A mechatronic system is an assembly of technological components, such as a mech-
anism, sensors, actuators, and a control unit. Recently, a number of researchers and
industrial manufacturers have highlighted the potential advantages of lightweight paral-
lel mechanisms with respect to the accuracy, dynamic performances, construction cost,
and transportability issues. The design of a mechatronic system with such a mechanism
requires a multidisciplinary approach, where the mechanical deformations have to be
considered. This thesis proposes two original contributions in this framework.
(i) First, a modular and systematic method is developed for the integrated sim-
ulation of mechatronic systems, which accounts for the strongly coupled dynamics of

the mechanical and non-mechanical components. The equations of motion are formu-
lated using the nonlinear Finite Element approach for the mechanism, and the block
diagram language for the control system. The time integration algorithm relies on
the generalized-α method, known in structural dynamics. Hence, well-dened concepts
from mechanics and from system dynamics are combined in a unied formulation, with
guaranteed convergence and stability properties. Several applications are treated in the
elds of robotics and vehicle dynamics.
(ii) Usual methods in exible multibody dynamics lead to complex nonlinear mod-
els, not really suitable for control design. Therefore, a systematic nonlinear model reduc-
tion technique is presented, which transforms an initial high-order Finite Element model

into a low-order and explicit model. The order reduction is obtained using the original
concept of Global Modal Parameterization: the motion of the assembled mechanism is
described in terms of rigid and exible modes, which have a global physical interpreta-
tion in the conguration space. The reduction procedure involves the component-mode
technique and an approximation strategy in the conguration space. Two examples are
presented: a four-bar mechanism, and a parallel kinematic machine-tool.
Finally, both simulation and modeling tools are exploited for the dynamic analysis
and the control design of an experimental lightweight manipulator with hydraulic ac-
tuators. A Finite Element model is rst constructed and validated with experimental
data. A reduced model is derived, and an active vibration controller is designed on
this basis. The simulation of the closed-loop mechatronic system predicts remarkable
performances. The model-based controller is also implemented on the test-bed, and the
experimental results agree with the simulation results. The performances and the other
advantages of the control strategy demonstrate the relevance of our developments in
mechatronics.
 Acknowledgements
This dissertation is the result of several years of research at the University of
Liège, and I feel deeply indebted to a number of people who directly or indirectly
inspired its realization.
First, I would like to thank my advisor Professor Jean-Claude Golinval for his
precious help and support during those years. My gratitude also goes to Professor
Pierre Duysinx for his technical guidance, and for his enthusiastic encourage-
ments. Every discussion with Professor Michel Géradin had a constructive inu-
ence on the orientation of this research, I am pleased to acknowledge him. Those
three Professors highly contributed to the improvement of the draft manuscript, I
am grateful to them for their advice during its elaboration.
I wish to express my thanks to Professor Wayne J. Book for his cheerful
welcome at the Georgia Institute of Technology in summer 2004, and for making
my stay as protable as it could be from all points of view. During the stay,
I especially appreciated the collaboration with Ryan Krauss, who initiated me
with the Ralf test-bed. Our numerous discussions were highly valuable for the
elaboration of chapter 6. I also acknowledge the technical support of J.D. Huggins
during this experimental work.
The IAP/AMS framework oered great opportunities to cooperate with bel-
gian specialists in mechatronics. The research on the semi-active suspension has
been initiated by the teams of KULeuven-PMA and UCL-PRM, which are grate-
fully acknowledged. In particular, I am indebted to Professor Paul Fisette (PRM)
who took the time to provide the data and the complementary results associated
with this problem. I am grateful to Denis Joassin, whose master thesis contributed
to the modeling of the suspension.
I wish to thank Professor Rodolphe Sepulchre for interesting discussions about
the application of control theories to exible multibody systems.
Professor Philippe Wenger (IRCCyN, France) is also acknowledged for mak-
ing the Orthoglide data available.
I would like to thank the Oofelie community, in particular the team of Open
Engineering, for its invaluable support in the software implementation of the
methods presented in this dissertation. Special thanks go to Professor Alberto
Cardona and Elisabet Lens (CIMEC-INTEC, Argentina), whose active collabo-
vi

ration made possible the development of capabilities for mechanism analysis in

Oofelie.
The University of Liège has been a stimulating scientic environment for
me. For their friendship, I am grateful to my colleagues from the Vibration and
Structural Identication group and from the ASMA department.
Many thanks also go to Professors Peter Eberhard, Jean-Claude Samin, Hen-
drik Van Brussel, and Doctor Gaetan Kerschen, who accepted to participate in
the jury of this doctoral thesis.
Finally, I would like to express my deepest gratitude to my family, for its sup-
port, and in particular to my parents for their repeated encouragements throughout
the preparation of this dissertation. Many friends also supported me during this
period, I wish to thank them all individually.
 Contents

1 Introduction 1
1.1 Integrated design in mechatronics . . . . . . . . . . . . . . . . . . 2
1.2 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 State of the Art 9

2.1 Dynamic models for multibody systems . . . . . . . . . . . . . . . 9
2.1.1 Rigid mechanisms . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.2 Flexible mechanisms . . . . . . . . . . . . . . . . . . . . . 11
2.2 Simulation of multibody systems . . . . . . . . . . . . . . . . . . 13
2.2.1 Time-integration in linear structural dynamics . . . . . . . 13
2.2.2 Time-integration in exible multibody dynamics . . . . . . 14
2.3 Integrated simulation of mechatronic systems . . . . . . . . . . . . 16
2.4 Model reduction techniques for control design . . . . . . . . . . . 24
2.4.1 Reduced-order modeling of a linear system . . . . . . . . . 26
2.4.2 Approximated representation of a nonlinear model . . . . . 28
2.4.3 Nonlinear reduction in exible multibody dynamics . . . . 29
2.5 Control of exible mechanisms . . . . . . . . . . . . . . . . . . . . 30
2.5.1 Methods based on Linear Time Invariant dynamics . . . . 31
2.5.2 Quasi-linear control methods . . . . . . . . . . . . . . . . . 32
2.5.3 Nonlinear control techniques . . . . . . . . . . . . . . . . . 34
2.5.4 Control strategy selected in this work . . . . . . . . . . . . 37

3 The Finite Element Approach in Multibody Dynamics 39

3.1 Motion parameterization . . . . . . . . . . . . . . . . . . . . . . . 39
3.1.1 Kinematic description of a planar mechanism . . . . . . . 40
3.1.2 Description of spatial rotations . . . . . . . . . . . . . . . 42
viii CONTENTS

3.2 Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.2.1 The constraint elimination method . . . . . . . . . . . . . 44
3.2.2 The augmented Lagrangian method . . . . . . . . . . . . . 44
3.3 Time-integration . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3.1 Newmark algorithm . . . . . . . . . . . . . . . . . . . . . . 46
3.3.2 Generalized-α method . . . . . . . . . . . . . . . . . . . . 47
3.3.3 Linear stability analysis . . . . . . . . . . . . . . . . . . . 50
3.3.4 Convergence analysis . . . . . . . . . . . . . . . . . . . . . 54
3.4 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

4 Integrated Simulation of Mechatronic Systems 61

4.1 Simulation of control systems . . . . . . . . . . . . . . . . . . . . 62
4.1.1 State equations . . . . . . . . . . . . . . . . . . . . . . . . 63
4.1.2 Semi-explicit index-1 assumption . . . . . . . . . . . . . . 65
4.1.3 Time-integration algorithm . . . . . . . . . . . . . . . . . . 67
4.1.4 Linear stability analysis . . . . . . . . . . . . . . . . . . . 70
4.2 Integrated simulation of mechatronic systems . . . . . . . . . . . . 72
4.2.1 Coupled equations of motion . . . . . . . . . . . . . . . . . 72
4.2.2 Time-integration algorithm . . . . . . . . . . . . . . . . . . 73
4.2.3 Linear stability analysis: the unconstrained case . . . . . . 78
4.2.4 Linear stability analysis: the constrained case . . . . . . . 81
4.2.5 Convergence analysis . . . . . . . . . . . . . . . . . . . . . 83
4.2.6 Optimal choice for the algorithmic parameters . . . . . . . 84
4.3 Systems with discontinuous dynamics . . . . . . . . . . . . . . . . 86
4.3.1 Impulse phenomena . . . . . . . . . . . . . . . . . . . . . . 88
4.4 Implementation issues . . . . . . . . . . . . . . . . . . . . . . . . 89
4.5 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.5.1 Rigid four-bar mechanism . . . . . . . . . . . . . . . . . . 91
4.5.2 Scara robot . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.5.3 Vehicle semi-active suspension . . . . . . . . . . . . . . . . 100
4.6 Summary and concluding remarks . . . . . . . . . . . . . . . . . . 110

5 Nonlinear Model Reduction in Flexible Multibody Dynamics 111

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.2 Linear component-mode synthesis . . . . . . . . . . . . . . . . . . 115
CONTENTS ix

5.2.1 Principle of a linear reduction method . . . . . . . . . . . 116

5.2.2 Component-mode selection . . . . . . . . . . . . . . . . . . 117
5.3 Parameterization of the rigid kinematics . . . . . . . . . . . . . . 120
5.3.1 Rigid mechanisms . . . . . . . . . . . . . . . . . . . . . . . 121
5.3.2 Flexible mechanisms . . . . . . . . . . . . . . . . . . . . . 126
5.4 The Global Modal Parameterization (GMP) . . . . . . . . . . . . 129
5.4.1 Partitioned coordinate transformation . . . . . . . . . . . 130
5.4.2 Denition of the local modes . . . . . . . . . . . . . . . . . 131
5.4.3 From local to global modes . . . . . . . . . . . . . . . . . . 132
5.5 Reduced-order model . . . . . . . . . . . . . . . . . . . . . . . . . 137
5.5.1 Elastic forces . . . . . . . . . . . . . . . . . . . . . . . . . 137
5.5.2 Inertia forces . . . . . . . . . . . . . . . . . . . . . . . . . 139
5.5.3 Reduced equations of motion . . . . . . . . . . . . . . . . 146
5.5.4 Algorithm for local model reduction . . . . . . . . . . . . . 146
5.6 Approximation in the conguration space . . . . . . . . . . . . . . 148
5.6.1 Piecewise strategy . . . . . . . . . . . . . . . . . . . . . . 149
5.6.2 Local polynomial approximation . . . . . . . . . . . . . . . 151
5.6.3 Automatic conguration space decomposition . . . . . . . 157
5.6.4 Additional parameters of the model . . . . . . . . . . . . . 161
5.7 Summary of the reduction procedure . . . . . . . . . . . . . . . . 162
5.8 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
5.8.1 Four-bar mechanism . . . . . . . . . . . . . . . . . . . . . 162
5.8.2 The Orthoglide . . . . . . . . . . . . . . . . . . . . . . . . 170
5.9 Summary and concluding remarks . . . . . . . . . . . . . . . . . . 172

6 Modeling, Simulation and Control of an Experimental Manipu-

lator 179
6.1 Description of the test-bed . . . . . . . . . . . . . . . . . . . . . . 180
6.1.1 Workspace description . . . . . . . . . . . . . . . . . . . . 182
6.1.2 Transfer functions acquisition . . . . . . . . . . . . . . . . 183
6.2 Model of the manipulator . . . . . . . . . . . . . . . . . . . . . . 184
6.2.1 Previous models of Ralf . . . . . . . . . . . . . . . . . . . 184
6.2.2 Structural model . . . . . . . . . . . . . . . . . . . . . . . 185
6.2.3 Validation and actuators models . . . . . . . . . . . . . . . 186
x CONTENTS

6.2.4 Nonlinear model reduction . . . . . . . . . . . . . . . . . . 189

6.3 Control design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
6.3.1 Slow controller . . . . . . . . . . . . . . . . . . . . . . . . 198
6.3.2 Fast controller . . . . . . . . . . . . . . . . . . . . . . . . . 200
6.3.3 Summary and rst results . . . . . . . . . . . . . . . . . . 204
6.4 Closed-loop system: simulations and experiments . . . . . . . . . 207
6.4.1 Model of the mechatronic system . . . . . . . . . . . . . . 209
6.4.2 Dynamic behavior at home conguration . . . . . . . . . . 210
6.4.3 Trajectories in the conguration space . . . . . . . . . . . 216
6.5 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . 223

7 Conclusion 229

Bibliography 235
 I

Introduction

Engineers are encouraged by the market demands to the creation of tech-

nological products with increasing complexity. Therefore, they generally exploit
a modular and hierarchical system approach, which suggests that a product is
a system, that can be obtained by appropriate combination of several simpler
subsystems.
This thesis concerns mechatronic systems, whose purpose is to generate a
controlled motion. For example, robots, machine-tools or vehicles are mecha-
tronic systems. At a highly abstracted level, they are composed of hardware and
software technological components, such as mechanisms, actuators, sensors and
control units, as illustrated in Figure 1.1. The interactions between the subsys-
tems are functional: it is assumed that they occur at localized points, namely the
ports of the subsystems.
The design problem consists in rening each subsystem description, leading
to an optimal system for the task under consideration. Numerous design methods
and tools are available to assist the engineer in this challenge; they usually rely
on models, which are simplied and abstracted representations of the reality.

Reference -
Controller - Ampliers - Actuators - Transmission - Mechanism -
trajectory
6

Sensors 

Figure 1.1: A mechatronic system.

2 CHAPTER 1. INTRODUCTION

Most methods are automated in computer software, able to achieve eciently

and systematically expensive and complex computational treatments.
Traditional design tools are discipline-oriented, and can only be used in a se-
quential approach : each subsystem is designed independently, assuming roughly
idealized interactions with the other subsystems. Hence, the initial optimization
problem naturally dened at the system level, should be translated at the com-
ponent level, involving the judgment of the engineer. As pointed out by Van
Brussel [VB96], this pragmatic approach obviously leads to a sub-optimal design,
and great improvements are expected from a concurrent approach, where all the
interacting components are considered in a unied framework.
In particular, over the last twenty years, dedicated methods have been de-
veloped for the dynamic analysis of exible mechanisms. In a mechatronic con-
text, the mechanism interacts with the control system, and the objective of this
dissertation is to propose modeling tools useful for the integrated design of the
mechatronic system. This idea is further developed in the following section.

1.1 Integrated design in mechatronics

The interactions between a mechanism and its control system are highly
simplied under the assumptions of (i) a rigid-body behavior, (ii) a simple serial
mechanical topology. Therefore, those assumptions are often considered as arti-
cial design constraints, and we shall briey highlight the potential benets of
their relaxation.
The rigid-body constraint implies that the natural frequencies of the mecha-
nism are far beyond the bandwidth of the actuators. Removing this constraint,
a lighter mechanical design is possible, with the advantages of a reduced moving
mass, a better arm weight to payload weight ratio, a lower power demand, smaller
actuators, and a better transportability. A exible mechanism oers interesting
properties to achieve compliant force control. The actuator bandwidth can also
be extended, yielding faster motions. However, a subsequent design challenge is
to control the exible dynamics of the mechanism, in order to guarantee tracking
performances, stability and disturbance rejection.
Compared to serial topologies, parallel mechanisms are appreciated for their
higher static and dynamic stiness. Moreover, the actuators may be located on
1.1. INTEGRATED DESIGN IN MECHATRONICS 3

the base, reducing the moving mass and simplifying the transmission mechanisms.
A diculty comes from the complexity in the geometrical workspace and in the
nonlinear kinematic relation between the motion of the actuated joints and the
eector location.
From this early analysis, the advantages of exible and/or parallel designs
appear especially attractive for high-speed applications (e.g. high-speed robots
and machine-tools), and problems where the mass is a critical issue (e.g. large
manipulators, space robots and foldable structures). In order to design a con-
troller for those complex dynamic systems, an appropriate mechanical model is
highly desirable at this level.
For a wide range of practical problems such as car suspensions and tradi-
tional machine-tools, the exibility of the mechanical system is not dominant,
and the complex kinematic behavior can be approximately represented by lin-
ear equations. After a design based on simplied linear models, the engineer
may worry about the validity of the predicted performances. Then a multidisci-
plinary simulation tool may be extremely valuable at a pre-prototyping stage, in
order to achieve inexpensive checks and performance predictions before physical
implementation and testing.
Even though the rigid body and serial topology constraints are discarded,
the sequential design is still a pragmatic approach, which allows a relative inde-
pendence in the design of the mechanism and of the controller. This sequential
approach suers from inherent limitations which can be overcome by the devel-
opment of a concurrent optimization of the whole system. At this level, a mul-
tidisciplinary simulation tool is valuable for the estimation of the optimization
criterion and its sensitivities.
Those modeling, simulation, and optimization concepts are relevant at spe-
cic stages of the design procedure. In this work, we assume that a rst design
of the system is available, which includes the description of the mechanism and
of the actuators. Hence, the objective is to dene the control strategy, and to
optimize the remaining design variables. A procedure to solve this problem may
involve the following steps:

1. the elaboration of a mechanical model suitable for control design,

2. the design of a control system, relying on the model dened in step 1,

4 CHAPTER 1. INTRODUCTION

3. the integrated modeling and simulation of the mechatronic system (pre-

prototyping),

4. the concurrent optimization of the remaining design variables, based on

step 3.

In order to obtain the most appropriate information at each step of the design
procedure, various models should be considered. Two important distinctions
among modeling concepts are introduced in the next two paragraphs.
Mathematically, the dynamic model of a mechanism is naturally expressed
by the equations of motion, which dene an instantaneous relationship between
the external actions (actuator forces, disturbances), the mechanism generalized
coordinates and their time-derivatives (velocities and accelerations). The equa-
tions of motion, supplemented with a solver algorithm, constitute a simulation
model of the system, relating time-domain responses (displacements, strains, etc)
with time-domain actions (applied forces, commands, etc). In system theory, the
same distinction can be done between the state equations, and their formulation
with a simulation algorithm.
For complex multibody systems, a modeling software is helpful to formulate
automatically the equations of motion from a high-level description. Among the
computer modeling methods, symbolic methods allow to build the equations of
motion in symbolic format, whereas numerical methods produce the equations
of motion as complex numerical procedures. The symbolic format has the ad-
vantages of portability and eciency, and it provides interesting insights in the
analytical structure of the equations. However, numerical methods are able to
deal with a more general class of problems, and they are especially suitable to
model the dynamics of a exible mechanism with complex topology in a system-
atic way.
After this clarication, let us further characterize the modeling requirements
in the design procedure, which are directly associated with the objectives of this
research.
In step 2, the control design usually exploits the mathematical structure of
the mechanical equations of motion formulated in step 1. If a linear model is
able to capture the essential dynamics, linear control theory can be eciently ap-
plied. Using linearizations at dierent operating points, it is sometimes possible
1.2. OUTLINE 5

to generalize this paradigm for nonlinear systems. For the critical applications
considered here, a nonlinear control technique is necessary, which advantageously
exploits the structure of a nonlinear model. Since the order of the controller is
often proportional to the order of the model, a low-order model in explicit and
analytical format is desirable at this stage. However, existing modeling meth-
ods for complex exible mechanisms are not able to construct such a nonlinear
model; they lead to high-order and computationally demanding models, which
are denitely not appropriate. An important contribution of this thesis is a sys-
tematic reduced-order modeling method, leading to nonlinear mechanical models,
compatible with the requirements of a control design procedure.
Steps 3 and 4 involve a multidisciplinary simulation model, for which less
stringent simplicity requirements are imposed: the computational load and the
complexity of the model should be balanced with the generality, reliability, modu-
larity and systematic implementation issues. In this thesis, we propose an original
integrated simulation tool for mechatronic systems based on a unifying Finite El-
ement formulation. This reliable method allows a modular model denition for
both the exible multibody dynamics and the dynamics of the control system.
In the presentation, a specic eort is delivered to dene natural connections
between the elds of modeling and control. Standard formulations from both
areas are combined for the developments of our original mechatronic concepts.
Hence, we hope that our point of view can be understood, exploited and developed
by specialists from both elds.

1.2 Outline
For the sake of consistency, the dissertation follows a progression from analy-
sis to design concepts, in contradiction with the sequence of the design procedure.
Thus, the topics of integrated simulation, model reduction and control design are
successively addressed in chapters 4, 5 and 6.
Prior to the presentation of our personal contribution, an extended state of
the art is developed in chapter 2. Preliminary modeling and simulation concepts
in multibody system dynamics are reviewed before specialized discussions on the
simulation of mechatronic systems, the reduced-order modeling and the control
of exible mechanisms. The Finite Element approach proposed by Géradin and
6 CHAPTER 1. INTRODUCTION

Cardona for exible multibody systems [GC01, Car89, CG89, CG88] is at the
core of our developments, and an overview is presented in chapter 3.
The simulation of mechatronic systems is addressed in chapter 4. The block
diagram language is selected for the description of the control system dynamics,
and the generalized-α integration scheme is extended to deal with the dynamics
of the state variables. After a formal presentation of the simulation method, theo-
retical convergence and stability results are established. Three examples illustrate
the generality and the eciency of the formulation: a four-bar mechanism, an
industrial robot, and a car equipped with a semi-active suspension.
Chapter 5 is devoted to the reduced-order modeling of exible mechanisms.
The method is an extension of the linear component-mode synthesis, which ac-
counts for the nonlinear kinematics of the system. Therefore, the original concept
of Global Modal Parameterization is dened, and the reduced equations of mo-
tion are formulated. The variations of the model in the conguration space are
approximated by piecewise polynomial functions. Critical implementation issues
are discussed, in connection with the control design requirements. The method
is illustrated with two examples: a rigid parallel machine-tool, and a exible
four-bar mechanism.
In chapter 6, the theoretical tools presented and validated in chapters 4 and 5
are exploited for the modeling and control design of a large and exible manipu-
lator, which is represented in Figure 1.2. A reduced-order model is validated with
experimental results, and the new insights in the dynamics of the manipulator
lead to the development of an original control law. Numerical simulations of the
closed-loop mechatronic system predict that a compromise can be obtained be-
tween the motion bandwidth and the stabilization of the mechanical vibrations,
which is conrmed by the experimental results. The various advantages of the
control strategy demonstrate the relevance of a global approach in mechatronics.
Our contribution is thus described from the general to the particular: simula-
tion of mechatronic systems, modeling of exible mechanisms, and application to
an experimental manipulator. Finally, chapter 7 draws conclusions and discusses
possible directions for future research.
1.2. OUTLINE 7

Figure 1.2: Large and exible manipulator.

8 CHAPTER 1. INTRODUCTION
 II

State of the Art

This research is developed in a multidisciplinary mechatronic framework,

and it is thus hopeless to attempt a comprehensive survey of the literature. In
order to situate our work, selected topics of interest are reviewed in this chapter:
dynamic modeling and simulation of multibody systems, integrated simulation
of mechatronic systems, model reduction techniques, and control strategies for
exible mechanisms. On this basis, our strategic choices are motivated for the
dierent contributions of the dissertation.

2.1 Dynamic models for multibody systems

This section discusses existing formulations for the equations of motion in
rigid and exible multibody dynamics, pertaining to the simulation and the con-
trol design problems. Simulation is a direct problem (or forward problem), i.e.
the equations of motion are solved in the time-domain from given information
about the applied forces and driven degrees-of-freedom (dofs). Open-loop and
feedforward control are inverse problems, i.e. the applied forces and driven dofs
are determined in order to yield a specied motion. The design of a closed-loop
controller is a non-standard problem, that may exploit the mathematical struc-
ture of the equations of motion.
The equations of motion of a mechanism rely on a kinematic parameteriza-
tion, using a set of generalized coordinates. Sometimes, a unique set of inde-
pendent coordinates is able to describe arbitrary motion of the system, and the
equations of motion are formulated as a set of Ordinary Dierential Equations
10 CHAPTER 2. STATE OF THE ART

(ODEs). But in general, the kinematic description involves redundant coordi-

nates, interrelated by algebraic and dierential constraints. Therefore, the equa-
tions of motion are formulated as a set of Dierential and Algebraic Equations
(DAEs), which is less convenient both for control design and simulation.

2.1.1 Rigid mechanisms

A mechanism made of rigid bodies connected by kinematic joints can be de-
scribed in terms of relative coordinates, i.e. the angles of the revolute joints and
the displacements of the prismatic joints. Usually those coordinates are physically
associated with the actuators, which make them relevant for control design, espe-
cially for the inverse dynamics problem [AS86]. For tree-topologies, the relative
coordinates are independent, which is an additional advantage. Alternatively,
Cartesian coordinates have been proposed for the ecient and systematic simu-
lation of complex mechanisms [DJB94].
Two approaches can be used in order to obtain the equations of motion: the
Newton-Euler formulation, and the Lagrangian formulation.
The Newton-Euler equations come from the application of the motion laws
to each rigid-body, involving coupling reaction forces and torques. According to
d'Alembert's principle, the reaction forces can be eliminated by projection onto
the conguration space, leading to a compact set of independent equations. For
mechanisms with tree-topology, recursive implementations of the Newton-Euler
equations lead to a direct formulation of the independent equations in terms
of the relative coordinates, with a high eciency [LWP80, SF03]. For parallel
mechanisms, recursive methods can still be exploited using the cut-joint method
to break the loops, and adding closure kinematic constraints between the relative
coordinates.
In the Lagrangian formulation [DJB94], the system dynamics is described in
terms of work and energy using generalized coordinates, e.g. relative or Carte-
sian coordinates. If the generalized coordinates are independent, the reaction
forces and torques are automatically eliminated, leading to compact equations
of motion. Redundancy in the generalized coordinates requires the introduction
of Lagrange multipliers. Optimized recursive Lagrangian formulations have been
presented in the literature [Hol80]. Both Newton-Euler and Lagrangian formula-
2.1. DYNAMIC MODELS FOR MULTIBODY SYSTEMS 11

tions can be implemented symbolically or numerically.

A detailed review of the literature about the dynamics of rigid mechanisms
is presented by Schiehlen [Sch97].

2.1.2 Flexible mechanisms

The kinematic description of a deformed body requires an innite set of dofs.
In order to develop a systematic and general modeling method, an approximation
can be obtained by spatial discretization. The motion is thereby parameterized
by a nite number of coordinates, which are the amplitudes associated with shape
functions. Dierent discretization techniques are possible in exible multibody
dynamics, leading to dierent formulations of the equations of motion; a few of
them are considered here.

The nonlinear Finite Element method

The most general and versatile approach in exible multibody dynamics is

the nonlinear Finite Element method. Each exible body is divided into elements,
and a piecewise approximation strategy is adopted for the displacement eld. The
amplitudes of the shape functions are nodal parameters, namely translations and
rotations with respect to an inertial frame. The consistent description of the large
displacements and rotations requires a nonlinear geometric formulation [Sim85,
SVQ86]. A systematic methodology for mechanisms analysis using the Finite
Element method is proposed by Géradin and Cardona [GC01], who select an
updated Lagrangian point of view to handle the large rotations. In order to
circumvent the dicult treatment of rotations, Shabana [Sha98] proposed the
Absolute Nodal Coordinate Formulation, where the nodal dofs are translations
and slopes.
The nodal parameterization of the motion is highly redundant, so that alge-
braic constraints arise in the equations of motion. The resulting large but sparse
system of DAEs is well-suited for simulation, but not for control design. For
this reason, alternative approaches leading to simpler models are discussed in the
following.
12 CHAPTER 2. STATE OF THE ART

The oating frame of reference approach

According to the oating frame of reference approach [DV76, SW83, MK91,

CG91, LHKB91, SM00, Nik03, SF03], the motion of a exible body is subdivided
into a reference motion and a deformation. The former is formally identied
with the motion of a rigid oating frame attached to the body, which can be
described according to rigid multibody formalisms. The analysis of the deforma-
tion within the oating frame is usually simplied assuming small displacements
and rotations, so that assumed-mode techniques can be exploited. Three major
diculties are associated with the selection of the reference frame, the denition
of the modal shape functions, and the formulation of the coupled kinetic energy.
The oating frame approach was stimulated by simulation problems. How-
ever, Book [Boo84] developed a similar method for control design. Relying on a
recursive Lagrangian formulation, this method can be implemented numerically
or symbolically.
The motion of the reference frame and the deformations are highly cou-
pled by the kinetic energy, requiring a strongly coupled formulation. However,
the inuence of the deformations on the motion of the reference frame becomes
negligible for low-speed applications, so that a weakly coupled scheme can be
developed in a co-simulation conguration [VA99, Arn04]: the dynamics of the
oating frame is simulated in a software dedicated to rigid multibody systems,
whereas the deformations are analyzed in a general purpose Finite Element code.

Lumped approaches

According to a lumped approach, a exible body is described as a lumped

model composed of rigid bodies interconnected by springs and dampers. Several
methods have been proposed for simulation by Huston [Hus91] and Wittbrodt and
Wojciech [WW95], and for control design by Seto and co-workers, e.g. [SH04].
The dynamic analysis can follow any standard methodology for rigid multi-
body systems. However, the typical number of elements required to model a
deformable structure is high, so that the eciency of the method becomes ques-
tionable for a complex three-dimensional mechanism.
2.2. SIMULATION OF MULTIBODY SYSTEMS 13

2.2 Simulation of multibody systems

The numerical simulation of a distributed system requires a space-time dis-
cretization of the partial dierential equations.
In the previous section, the nonlinear Finite Element method was recom-
mended for the systematic spatial semi-discretization of a exible multibody sys-
tem. The dynamics is then represented by a large but sparse set of index-3
DAEs1 .
The time-discretization and the denition of integration formulae are the
next step in the development of a simulation method. Since linear structural
dynamics is a special case of exible multibody dynamics, it is instructive to
review the numerical integration methods developed in this eld.

2.2.1 Time-integration in linear structural dynamics

A Finite Element model of a dynamic structure is typically a large set of
ODEs. These equations accurately represent the low-frequency modes of the
physical system, but their high-frequency content is dominated by numerical phe-
nomena resulting from the Finite Element discretization. Due to those numerical
modes, the equations of motion are sti 2 .
At this level, it is necessary to distinguish explicit and implicit integration
schemes, as explained by Géradin and Rixen [GR97]. The stability of an explicit
integration scheme is only guaranteed if the time step is small enough with respect
to the natural frequencies of the system. On the other side, the commonly used
implicit algorithms are unconditionally stable (or A-stable), which means that
the numerical solution is stable whatever the frequency content of the mechanical
system. This is a desirable property for the reliable simulation of sti equations of
motion, but the price to pay for an implicit method is an increased computational
complexity.
According to Hughes [Hug87], a time-integration scheme in structural dy-
1 The index characterizes the algebraic structure of a set of DAEs. It is the minimum number
of times that all or part of the equations must be dierentiated with respect to time in order
to obtain a set of ODEs [BCP96].
2 The concept of sti dierential equations is dened and discussed in details by Hairer and

Wanner [HW91].
14 CHAPTER 2. STATE OF THE ART

namics should ideally combine the following properties: unconditional stability

for linear systems, no more than one set of implicit equations to be solved at
each step, second-order accuracy, controllable numerical dissipation in the higher
modes, and self-starting procedure.
One-step methods are attractive since they are self-starting and lead to simple
code architecture. Among them, the Newmark family of implicit solvers have
been extensively applied in structural dynamics. From the original formulae
proposed by Newmark [New59], improved algorithms have been constructed, such
as the Hilber-Hughes-Taylor algorithm [HHT77], and the generalized-α method
by Chung and Hulbert [CH93], in order to combine high-frequency dissipation
with second-order accuracy. Since unconditional stability can be demonstrated,
those methods satisfy the specications proposed by Hughes.
The Newmark schemes are well adapted for the second-order dierential
equations of a structural model. Alternatively, general purpose ODE-solvers have
been developed by mathematicians, such as the one-step, multistage Runge-Kutta
methods [Gea71, HNW87]. These solvers can reach a high-order of accuracy, with
guaranteed stability properties. Analysts from structural dynamics are sometimes
reluctant with respect to this approach, fearing the computational burden of a
multi-stage algorithm. However, Owren and Simonsen [OS95] have shown that a
class of Runge-Kutta methods could be implemented with an acceptable compu-
tational cost. They concluded that those algorithms are also good candidates for
the simulation of dynamic structures.

2.2.2 Time-integration in exible multibody dynamics

An additional diculty for the simulation of exible mechanisms comes from
the DAE nature of the equations of motion. Brenan et al. [BCP96] analyze
the conditions under which standard ODEs solvers lead to acceptable solution
when applied to DAEs. They present the implementation of multistep backward
dierence methods in DASSL, a general purpose code designed for index-0 and
index-1 DAEs. For higher index DAEs, several implementation diculties are
reported, and the authors recommend a reformulation of the equations.
Following this advice, the index-3 equations of motion should be modied
according to a constraint elimination or an index reduction. The constraint elim-
2.2. SIMULATION OF MULTIBODY SYSTEMS 15

ination [WH82, DJB94] is a projection of the initial DAEs on an underlying

set of ODEs, whose solution computed with ODEs solvers naturally satises the
constraints. The index reduction involves the dierentiation of the constraint
equations, in order to express them at the velocity level (index-2 formulation),
or at the acceleration level (index-1 formulation). Unfortunately, the numerical
solution of the reduced-index system does not fulll the original constraints at the
displacement level and, due to error propagation, a drift-o from these constraints
occurs. To avoid this phenomenon, a stabilization technique [Bau72] should be
implemented, which usually brings articial high-frequencies in the response.
Géradin and Cardona [Car89, CG89, GC01] analyzed the behavior of the
Newmark, Hilber-Hughes-Taylor and generalized-α algorithms when directly ap-
plied to the original equations of motion. The iterative nature of these implicit
algorithms allows to satisfy the algebraic constraints with a level of accuracy
specied by the user. In the linear regime, they demonstrated that a weak insta-
bility is expected in the Lagrange multipliers, which can be eliminated by adding
a small numerical dissipation over the high-frequencies. Those methods have
been implemented in the Samcef-Mecano commercial software [SAM99]. Even
though no stability proof is available in the nonlinear regime, this methodology
turned out to be very ecient and powerful for the wide range of applications
encountered by the users of Samcef-Mecano.
Recently, several authors developed a class of algorithms that naturally
preserve invariants of the dynamic system, such as energy and canonical mo-
menta [STW92, BBT03, LCG04]. If energy preservation does not necessary leads
to the accuracy of the solution, those algorithms usually perform remarkably well.
In order to solve eciently sti problems, Bauchau et al. [BBT03] introduced
high-frequency dissipation, leading to a so-called energy-decaying scheme, with
guaranteed unconditional stability in the nonlinear regime. The construction of
the integration formulae relies on the Lagrangian structure of the mechanical
system. In a mechatronic framework, it is usually not possible to describe the
control system with a Lagrangian function, and unfortunately, those methods
become irrelevant.
As a conclusion, among time-integration methods, improved Newmark algo-
rithms (Hilber-Hughes-Taylor and generalized-α methods) turn out to be good
candidates for the numerical simulation of mechatronic systems with exible
16 CHAPTER 2. STATE OF THE ART

mechanisms.

2.3 Integrated simulation of mechatronic systems

The functional decomposition of a mechatronic system into simpler interact-
ing subsystems is of practical interest for design purpose. Modularity is also a
key issue for the integrated simulation of the mechatronic system. This section
addresses the comparison of existing modular simulation concepts based on three
fundamental questions:

1. how are the subsystems described?

2. how are the couplings between the subsystems dened?

3. how does the simulation algorithm deal with the coupled problem?

Before this comparison, let us analyze general properties of simulation methods.

Basic simulation concepts

In this work, we are especially interested in computational modeling tech-
niques, which have followed the exponential development of computer capabilities
over the last decades. Any high-level operation of the user is supported by nu-
merous modeling concepts and dedicated treatments implemented at the lower
levels. However, the high-level description should be suciently intuitive, so that
the underlying levels can be "forgotten" in the modeling process. For example,
modern multibody dynamics software are equipped with powerful graphical user
interface, so that the user manipulates visually the idealized bodies and joints of
the model, intuitively associated with the actual system.
Hence, a modular simulation software for mechatronic systems requires a
user-friendly interface, associated with a consistent low-level implementation.
Most existing software are characterized by the standard architecture sketched in
Figure 2.1. This overview is certainly simplied, and a more detailed description
would include other libraries, and highlight the deep connections between them.
The lower level is the programming language. The compiler, the operating sys-
tem and the computer hardware could have been mentioned below, but they will
2.3. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS 17

Figure 2.1: General structure of a modular simulation software.

not be considered here. At the kernel level, the concepts are built using the syn-
tax of the programming language, and are exploited to establish the element and
analysis libraries. The element library aims at describing every subsystem and its
contribution to the global equations of motion, whereas the analysis module de-
nes the treatments to solve those equations. At the top, the user interface gives
access to the element and analysis libraries in order to construct the simulation
model.
In the following, several modular simulation approaches are successively con-
sidered: the Finite Element method, physical modeling techniques (the bond
graph and the linear graph), variational principles, mathematical approaches
(the block diagram). We shall discuss their ability to represent the dynamics
of a exible mechanism within the mechatronic system. A general comparison is
summarized in Tables 2.1 and 2.2: the answers to the three fundamental questions
(elementary description, coupling strategy and time-integration) are decomposed
according to the software architecture given in Figure 2.1. We encourage the
reader to refer to them during the presentation of the dierent methods.

The Finite Element method

The relevance of the Finite Element method for the mechanical modeling
of complex exible multibody systems has been demonstrated earlier. Here the
method is presented in a more general framework as a modular modeling approach
for distributed systems.
Modularity comes from the geometric decomposition of the distributed sys-
tem into simple elements coupled by nodal dofs. The numerical assembly of
elementary contributions leads to the formulation of the coupled equations of
motion. Several time-integration algorithms have been proposed in the literature
18 CHAPTER 2. STATE OF THE ART

for the simulation of Finite Element models, see section 2.2.

The user interface species a high-level syntax for the geometric and physical
description of the elements, and for the denition of the nodes. The user is also
able to select the time-integration algorithm and its numerical parameters. At a
lower level, each element is dened by its dofs and the algorithms to compute its
numerical contribution to the dynamic equilibrium. The integration schemes are
implemented at the analysis level, whereas the general concepts of element, dof
and node are dened at the kernel level.
Traditionally, the Finite Element method is applied for the simulation of
mechanical systems, but it can be naturally extended to represent non-mechanical
systems. For example, in a previous work [BG03, BDG03], the dynamics of a
controller has been considered in a Finite Element model in order to analyze the
active vibration control of a complex mechanical structure. The design of the
control system could thus be guided by the simulation results.

Physical modeling: Bond Graph / Linear Graph

In physical modeling, one assumes that each physical component exchanges
energy with its environment through localized ports. Hence, the interaction be-
tween subsystems is characterized by port variables, which are physical quantities
satisfying conservation laws at the system level (e.g. power, voltage, current,
ow, eorts...). For instance, the conservation equations of an electrical network
are the Kirchho equations, directly connected with the topology of the graph
describing the network. Those equations should be complemented by the consti-
tutive equations of each component (e.g. Ohm's law of a resistor) to characterize
the dynamics of the system. By exploiting a graph analysis algorithm, it is pos-
sible to organize the equations in a compact form, thereby eliminating redundant
variables. This operation can be achieved symbolically, leading to an explicit
analytical expression of the minimal set of equations.
Additional information may be given by the user in order to help the graph
analysis algorithm. For bond graphs, this information consists in a causality as-
signment of the interactions, whereas for linear graphs, the user should specify
a topological tree. Since the conservation equations are implemented at the ker-
nel level, the development of a new element only requires the denition of its
constitutive equations, which is fairly simple.
2.3. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS 19

Physical modeling methods are especially well-suited when the port variables
are scalar quantities. Among the numerous references in the literature, we may
cite [OE97, KMR00] for bond graphs and [KTKH67] for linear graphs. Recently,
a signicant eort has been spent for the standardization of multidisciplinary sim-
ulation techniques, leading to the development of the unifying modeling language
Modelica [EMO98, Til01], which is based on bond graph theory.
Several authors have extended the concept of bond graph and linear graph
for multibody systems, see [Kar97, Fav97] and [McP96, McP98, McP03, SM03a],
respectively. According to the comparison study by Sass [Sas04], the linear graph
is more appropriate in that case, and it is noticeable that the topological analysis
automatically solves the kinematic problem. The extension of the linear graph
method for exible multibody systems was also presented in [SM00, SMH01],
relying on an a priori spatial discretization with assumed-modes. McPhee et
al. [SM03a, SM03b] demonstrated the relevance of the linear graph method in
electromechanics.

Methods based on variational principles

Several authors [HP95, MEFK97] applied the virtual work principle for the
modeling of electromechanical systems. Its scalar formulation is a denite advan-
tage to mix various contributions from electrical and mechanical subsystems. The
method relies on the selection of a minimum set of independent generalized coor-
dinates, able to describe the electromechanical conguration of the system. For
serial rigid mechanisms, this principle leads to the standard Newton-Euler equa-
tions in terms of independent joint coordinates. Hadwitch and Pfeier [HP95]
applied the virtual work principle to rigid multibody systems interacting with
electrical networks. From electrical and mechanical generalized coordinates, all
other redundant mechanical and electrical variables can be deduced using the
kinematic equations and the Kirchho laws, respectively.

Mathematical modeling: Block Diagram

In the aforementioned methods, the dynamic equations are formulated from
physical principles. However, the input/output behavior of a controller follows an
algorithm designed in an articial way. This is probably the reason why control
20 CHAPTER 2. STATE OF THE ART

engineers prefer the more general block diagram formalism, where the equations
of the subsystems are directly manipulated. The block diagram is a mathemat-
ical language for the modular description of dynamic systems in terms of input,
output, and internal variables; it should not be associated with a particular sim-
ulation method. Depending on the coupling strategy, three dierent approaches
are considered below.
The weakly coupled strategy, available in several commercial software (Simulink,
Vissim, etc.) is often regarded as the standard approach for the simulation of
a block diagram model. Each individual subsystem is usually represented by
ODEs, but algebraic equations come from input/output interconnections. The
time-integration is conducted separately for each block in an asynchronous way,
and the coupling results from the exchange of input/output numerical values. The
sequence followed for the successive treatments of the blocks is obtained from a
causality analysis, itself relying on the input/output causality property of every
block. However, this causality analysis breaks down in case of algebraic loops3 , so
that reliable algebraic loop detection and solver algorithms are essential for con-
sistent simulation results. As represented in Table 2.2, the ODE solver directly
deals with the individual blocks, so that the separation between the element and
analysis library is virtual. For this reason, the weak coupling strategy is not
suitable for globally implicit integration schemes, and it is not recommended for
sti problems.
Partitioned simulation methods can also be formulated mathematically in
terms of block diagrams. Each subsystem can hide a highly complex dynamics
(e.g. the mechanical or the electrical part of a mechatronic system), and a parti-
tioned treatment intends to improve the computational eciency and the mod-
ularity of the simulation. Among those methods, let us mention the multi-rate
integration strategies for ODEs suggested by Gear and Wells [GW84], their gener-
alization by Schiehlen and co-workers to DAEs with algebraic loops [RS98, KS00],
and the methods based on Gauss-Seidel iteration for systems coupled by alge-
braic constraints developed by Arnold et al. [Arn01, HAV03]. Co-simulation
techniques are also partitioned methods, where the subsystems are implemented
in dierent specialized software [VBSV99, VGVDS+ 99]. In this case, a master
3 An algebraic loop is a topological loop where all the blocks are direct-feedthrough, which
means that their outputs are instantaneously aected by their inputs.
2.3. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS 21

linker module manages the task scheduling, and the calling procedures of the
other software.
Finally, the block diagram language leads to a strongly coupled simulation if
the coupling is imposed prior to the integration scheme: an assembly procedure
is necessary to build the coupled equations, which are treated by a monolithic
solver. Sass [Sas04] recently adopted this point of view to establish symbolically
the coupled dynamic equations of an electromechanical system composed of two
subsystems: a rigid mechanism and an electrical network. The strongly coupled
approach is also selected in this thesis, but the assembly is performed numerically,
as in the Finite Element method.
A few methodologies escape to the strict classication between weak and
strong coupling, such as the fastest-rst approach evoked in [GW84], or the sub-
cycling technique proposed by Cardona [Car89] for the dynamic analysis of mech-
anisms with hydraulic actuators. In both cases, smaller time-steps are applied for
the fast subsystem according to a weakly coupled approach, but a strong coupling
is considered at every global time-step.

Approach selected in this work (Table 2.3)

Chapter 4 will be devoted to the simulation of mechatronic systems composed
of a complex exible mechanism and a control system.
The nonlinear Finite Element method is appropriate to deal with the me-
chanical problem, whereas the block diagram approach seems especially suitable
for the modeling of the control system. Thus, we propose to integrate the block
diagram language into the Finite Element framework, so that the strongly coupled
equations are obtained using the standard Finite Element assembly procedure.
This choice oers the following advantages:

- the detailed, accurate and reliable dynamic representation of the exible

mechanism,

- the generality of the block diagram formulation, which can be extended to

discontinuous and sampled systems,

- the modularity of the block diagram library, leading to an intuitive modeling

process,
22 CHAPTER 2. STATE OF THE ART

Element
Coupling strategy Time-integration
description

Element type,
User
Finite Element method

geometry, physical Nodes Solver selection

interface properties
Element Numerical
Element Dofs
library contributions
Analysis Integration scheme
library
Dof + node concepts,
Kernel Element concept
assembly procedure

Ports, causality
User Element type, physical
assignment/tree Solver selection
interface properties
Physical modeling

selection
Element Constitutive equations Port variables
library
Analysis ODE/DAE solver
library
Port concept,
Kernel Element concept topological equations,
graph analysis

User Element type, physical Generalized

Variational principles

Solver selection
interface properties coordinates denition
Element Energetic Generalized/redundant
library contributions coordinates
Analysis ODE/DAE solver
library
Relation generalized/
Kernel Element concept
redundant coordinates

Table 2.1: Comparison between modular simulation concepts.

2.3. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS 23

Block description Coupling strategy Time-integration

User Mathematical model I/O connections Solver selection

interface
Element
Weak coupling

ODEs I/O variables Feedthrough nature

library
Analysis ODE solver
library
Block order selection,
Kernel Block concept I/O variable concept algebraic loop
detection and solver

User Solvers/scheduler
Partitioned methods

Subsystem model I/O connections

interface selection
Element Dedicated methods I/O variables
ODE/DAE solver,
library extrapolation method
Analysis Scheduler
library
Kernel Block concept I/O variable concept

User Subsystem model I/O connections Solver selection

interface
Strong coupling

Element Dedicated methods I/O variables

library
Analysis DAE solver
library
I/O variable concept,
Kernel Block concept
assembly procedure

Table 2.2: Comparison between modular simulation concepts: block diagram

approaches. "I/O" means "Input/Output".
24 CHAPTER 2. STATE OF THE ART

Element Coupling Simulation

Mechanism System
Type, geometry, I/O dofs,
User
physical mathematical Nodes, I/O dofs Solver selection
interface properties model
Element Numerical
ODEs Element Dofs
library contributions
Analysis Coupled DAE
library solver

Dof/node
Block diagram concepts,
Kernel Element concept
concept numerical
assembly

Table 2.3: Approach selected in this work.

- the formal equivalence with widespread block diagram software, such as

Simulink, so that the modeling concepts are very natural for the control
engineer, and the exchange of models is facilitated,

- the reliability of the method, since the strongly coupled approach prevents
any diculty due to algebraic loops.

Two important contributions of this thesis are the formulation of the block di-
agram concepts in a Finite Element framework, and the extension of the
generalized-α method for the time-integration of the coupled equations.

2.4 Model reduction techniques for control design

Modern formalisms in multibody dynamics allow a detailed and reliable rep-
resentation of complex mechanical systems. However, high levels of accuracy
and generality can only be reached at the price of more sophisticated mod-
els, which require increased computational resources. Therefore, Eberhard and
Schiehlen [ES98] advocated a hierarchical modeling approach, which suggests
that dierent models should be developed to meet the specic requirements at
the dierent stages of a design procedure.
2.4. MODEL REDUCTION TECHNIQUES FOR CONTROL DESIGN 25

This work addresses the control design of exible mechanisms. At this design
stage, it is commonly accepted that an ideal mechanical model should combine
the following properties:

1. accurate input/output representation in the domain of interest, which is

dened from the actuator bandwidth viewpoint and from the workspace
viewpoint,

2. low-order, since the order of the controller is often related with the order
of the model,

3. represented by ODEs (independent coordinates, no constraint equations),

4. available in closed-form - at least in a structured format, since the structure

of the model opens the range of control strategies that can be considered,

5. computationally ecient, since it might be implemented in real-time, or

involved in an optimization procedure,

6. obtained using a general and systematic formulation, which is especially

valuable for complex systems,

7. portable, in order to be available in software environments dedicated to

control design.

The objective of this section is to compare existing modeling techniques with

respect to those criteria.
For rigid mechanisms with serial topology, an appropriate inverse dynamic
model can be obtained using recursive methods. Due to the presence of algebraic
constraints, the ecient inverse dynamics of rigid mechanisms with complex par-
allel topology is still an open problem [BDG06].
In exible multibody dynamics, the nonlinear Finite Element method and
the oating frame approach were pointed as relevant modeling techniques in sec-
tion 2.1.2. For complex mechanisms with linear elastic deformations and parallel
topologies, those approaches are compared in Table 2.4, according to the seven
aforementioned criteria.
Both methods lead to a suciently accurate model. However, the Finite Ele-
ment parameterization involves a large number of redundant coordinates, leading
26 CHAPTER 2. STATE OF THE ART

Finite Element Assumed-mode

Accuracy yes yes
Order no ∼
ODEs no ∼
Structure ∼ yes
Computational eciency no ∼
Generality yes yes
Systematic formulation yes no
Portability no yes
Table 2.4: The assumed-mode technique vs. the Finite Element method for
exible mechanisms with complex topologies and linear elasticity.

to a high-order and computationally demanding model. As a numerical method,

it suers from a poor portability, and its strongest advantage in this context fol-
lows from its systematic formulation. The oating frame of reference approach
relies on a more compact parameterization, with a separation between rigid and
exible coordinates. The resulting model is more structured, which is an ad-
vantage for the design of the control algorithm. Symbolic implementations of
assumed-mode methods lead to high portability and ecient computations. But
in case of complex parallel mechanisms, due to the large number of modes (several
modes are necessary for every exible body), and due to the presence of nonlin-
ear algebraic constraints, the resulting model is still too complicated for control
design. Moreover, the denition of the mode shapes cannot be realized in a truly
systematic way.
A possible breakthrough may come from reduction or approximation tech-
niques, whose developments in linear and nonlinear frameworks are discussed in
the following.

2.4.1 Reduced-order modeling of a linear system

A linear reduction technique transforms an original model into a lower-order
model, with a minimized loss in accuracy. The control system is expected to be
robust with respect to the omitted dynamics. Several methods were proposed
2.4. MODEL REDUCTION TECHNIQUES FOR CONTROL DESIGN 27

in both elds of linear system theory and linear structural dynamics. We shall
review them before considering the reduction of a nonlinear exible multibody
system.
Several state-space reduction methods have been proposed for linear time-
invariant systems. The reduction relies on two low-dimensional subspaces: the
former denes a coordinate transformation, and the second, a projection operator
for the state equations. The objective of a reduction technique is to select those
subspaces in order to minimize the accuracy loss. Some methods, well-suited
for large-scale problems, are based on Krylov subspaces [GVVD04], and others
rely on a truncated balanced realization, as presented by Gawronski [Gaw98]. In
order to preserve physical properties of the system, such as dissipativity, congru-
ence transformations4 may be advantageous, especially for passive RLC electrical
circuits and passive mechanical systems [PLS03].
In linear structural dynamics, dedicated reduction techniques exploit the
second-order and lightly damped nature of the equations of motion. Initially,
those methods were not developed for control applications, but for substructured
analysis of complex mechanical systems: the reduced contributions of each sub-
structure are established independently in a rst step, in order to simplify the
analysis of the assembled structure in a second step. A congruence transforma-
tion is dened in the space of generalized coordinates as a mode shape matrix.
Hurty [Hur65], Craig and Bampton [CB68], Herting [Her79] and Craig [Cra87]
proposed various methods to select the truncated modal basis. They are usu-
ally denoted component-mode techniques, and a more detailed presentation will
be given in chapter 5. Interesting combinations of reduction methods in system
theory and in structural dynamics have been investigated by Gawronski [Gaw98]
and De Fonseca [DF00].
The component-mode synthesis leads to an optimal modal representation of
a exible body, which can be exploited in a oating frame formulation, leading
to the concept of superelement [SW83, CG91, GC01]. But again, for a complex
exible multibody system, the resulting model does not meet the requirements
of a control design procedure.

4 In a congruence transformation, the subspaces for the coordinate transformation and for
the projection are identical.
28 CHAPTER 2. STATE OF THE ART

2.4.2 Approximated representation of a nonlinear model

A linear dynamic system can be represented using a few matrices (e.g. state-
space matrices or mass, damping and stiness matrices), whereas a nonlinear
dynamic system may have a very complex structure, which can hardly be casted
in a generic format. Sometimes, a nonlinear system is represented explicitly
by analytical ODEs, which constitute an appropriate basis for control design.
However, the model of a complex exible mechanism is formulated implicitly in
terms of a costly and non-portable numerical procedure. Therefore, it is highly
relevant to seek for an approximated explicit formulation, i.e. to build a simplied
metamodel of the initial model. In the eld of optimization, this concept has been
applied for response surface techniques [MM95], in order to reduce the number of
runs of a full model. In robotics, a nonlinear dynamic model can be pre-computed
oine and stored in a look-up table, as discussed by Raibert [Rai77]; this look-up
table can then be implemented online for computed-torque control. More general
approximation methods have been developed in the eld of control, and a quick
overview is presented in the following.
Articial neural networks are theoretically able to represent any kind of non-
linearity, and they oer an interesting solution for the approximation of dynamic
systems. During a training stage, the neural network learns an input/output be-
havior from a set of data collected on the initial model. The diculties of this
general method are associated with the choice of the structure of the network, the
implementation of the training algorithm, and the black-box nature of the result-
ing model. The development of neural networks for the control of mechanisms
has been considered by Gutiérres et al. [GLL98].
Besides neural networks, other simpler approximation functions are possi-
ble, such as low-order polynomials, rational functions or kriging functions. In
every cases, the data necessary to build the model should be selected with great
care, and theories have been specically developed for this design of experiment
problem [Mon97]. In particular, adaptive techniques, where the set of data is con-
structed iteratively in order to minimize the approximation error, have a clear
superiority over regular gridding techniques.
In many practical situations, a dynamic system can be characterized by its
linearized behavior around specic operating points. For instance, around a con-
2.4. MODEL REDUCTION TECHNIQUES FOR CONTROL DESIGN 29

guration of a mechanism, the linearized equations of motion contain relevant

information about its dynamic behavior. Hence, a global model might be ob-
tained by appropriate combination of local models, according to the local model
network approach, also denoted the multiple model approach [TS85, Nel99, MS94,
Ang01, Sch01]. Takagi and Sugeno [TS85] proposed to build a smooth global
model using concepts from fuzzy sets theory, whereas other authors adopted in-
terpolation methods. Actually, a local model network can be interpreted as a
Linear Parameter Varying system, which belongs to the more general class of
Linear Time-Varying systems. The parametric dependence can be expressed in
several ways, leading for instance to Polytopic Linear Models analyzed by An-
gelis [Ang01], or models based on Linear Fractional Transformation considered
by Scherer [Sch01]. Control theories have been proposed exploiting these specic
structures; Caswara and Unbehauen [CU02] and Symens [Sym04] applied some of
them for the control of exible manipulators. However, the combination of local
models established for dierent operating points may lead to a non-consistent
global model, especially if the dynamic coordinates (e.g. the state variables) do
not have a global physical interpretation.

2.4.3 Nonlinear reduction in exible multibody dynamics

The model associated with an assembled exible mechanism is nonlinear, but
around a given conguration, it can be locally characterized by a linear model,
complemented by a quadratic gyroscopic term.
An original nonlinear reduction technique is presented in chapter 5, which
extracts a low-order and simplied model from an initial Finite Element model.
The reduction relies on the assumption of a linear elastic behavior, and it proceeds
in two steps:

- The order reduction results from a Global Modal Parameterization, which is

inspired by the component-mode technique, with the conceptual dierences
that the mode shapes are associated with the whole mechanism, and depend
on the mechanical conguration. The clear physical interpretation of the
modal coordinates guarantees the consistency of the reduced model in the
conguration space. The reduction procedure should be repeated for every
conguration changes, and as such, it can hardly be exploited for control
30 CHAPTER 2. STATE OF THE ART

design.

- The nonlinear variations of the reduced-order model in the conguration

space are approximated using a piecewise polynomial interpolation. This
strategy is based on an adaptive conguration space inspection algorithm,
which minimizes the computational resources to satisfy a specication on
the approximation error.

This systematic and general method leads to a highly compact dynamic rep-
resentation of complex exible mechanisms. The resulting model is appropriate
for control design, but it can also be used to build a more structured model, e.g. a
polytopic linear model or a linear fractional model. As a special case, the method
is applicable to parallel rigid mechanisms [BDG06].

2.5 Control of exible mechanisms

This section explores active techniques to reduce vibration problems of mech-
anisms. These methods are opposed to passive damping methods, which rely either
on the use of high damping materials, such as composites, or on the implementa-
tion of lumped passive dampers, tuned for specic operating conditions. In many
situations, the use of damping materials is excluded, and it may be dicult to
add lumped passive dampers to the system. Active damping is then an interesting
alternative, and the design of the control algorithm may lead to higher eciency
for a broad range of operating conditions.
According to the classical design approach, a rigid mechanism is driven by
collocated joint actuators and sensors. A general assumption in this dissertation is
that the number of actuators equals the number of rigid modes; in other words,
the mechanism is assumed to be fully actuated. In this case, classical control
strategies, such as the computed torque technique [AS86], can be applied if an
inverse dynamic model is available in real-time. For mechanisms with complex
topologies, our original model reduction method can be advantageously exploited
for this purpose [BDG06].
When the rigidity assumption needs to be reconsidered, collocated pairs of
sensors and actuators may be added in order to achieve a direct vibration con-
trol [Mei90, Pre97, Gaw98]. This method is widely exploited for active damping
2.5. CONTROL OF FLEXIBLE MECHANISMS 31

of large space structures. Another solution is to avoid extra actuators, and use
the existing actuators for the simultaneous control of the large amplitude motion
and of the vibrations. This is the indirect vibration control strategy, which is
especially suitable for the control of robots and machine-tools. This approach is
cheaper and simpler from a mechanical point of view. However, if one is interested
in controlling the tip-position, the non-collocated conguration of the sensors and
the actuators complicates the design of the control law. This research focuses on
the indirect vibration control problem, which has received a major attention in
the literature.
The control methods reported here rely on a model resulting from spatial
discretization with a nite number of coordinates. Since a exible system has, in
principle, an innite number of dofs, the higher order dynamics is neglected, but
its interaction with the control system may cause spillover instability. Special
care is necessary to prevent the system from such undesirable phenomena.
Most control techniques rely on the structure of a mathematical model, hence,
we successively consider methods based on linear approaches, on quasi-linear ap-
proaches, and nally on nonlinear concepts; of course, each method is conditioned
by the availability of an appropriate model.

2.5.1 Methods based on Linear Time Invariant dynamics

The early experimental work realized by Cannon and Schmitz [CS84] focuses
on the end-point control of a single exible beam rotating around its origin. The
dynamic equations of this mechanical system can be reasonably approximated by
a linear model, experimentally identied. The authors report that the transfer
function from the actuators to the tip position is non-minimum phase, which can
be explained by the non-collocated conguration of the joint actuator and the
tip position sensor. In the following, we shall insist on the consequences of this
property for the design of an open-loop or of a closed-loop control law.
A rst control design problem is the command generation problem. For rigid-
body mechanisms, it is often accepted that bang-bang acceleration proles (i.e.
piecewise constant acceleration proles) lead to optimal trajectories. When deal-
ing with exible mechanisms, the high-frequency content of such proles may
excite the vibrations of the system. Bayo and Paden [BP87] recommended the
32 CHAPTER 2. STATE OF THE ART

use of smoother Gaussian velocity proles. Further improvements can be obtained

using input shaping methods proposed by Singer and Seering [SS90], which con-
sist in pre-ltering the command signal in order to reduce the detrimental eect of
mechanical exibility on the dynamic response. Designed for linear systems, the
eciency of input shaping has been demonstrated in several cases for nonlinear
mechanisms [RV02], especially when combined with a nonlinearity compensa-
tion feedback [KJT95]. The command generation problem can also be reformu-
lated as an inverse dynamics problem: nd the torque function to be applied
so that the eector follows a specied trajectory in the time-domain. Solutions
obtained by feedforward pole/zero cancellation are not applicable because of the
non-minimum phase dynamics. Actually, the solution is necessarily non-causal,
which means that the torque should start before the tip motion, and continue
after reaching the nal position. Bayo [Bay87] proposed an o-line frequency
domain method, later extended to multilink mechanisms [DB94], whereas Kwon
and Book [KB94] advocated a time-domain approach.
Input shaping and inverse dynamics are open-loop methods, which suer
from model sensitivity, and poor disturbance rejection. Therefore, they should be
combined with a feedback control strategy. Usually, one is interested in controlling
the eector of the manipulator, which calls for a noncollocated end-point feed-
back design. Cannon and Schmitz [CS84] have successfully implemented an LQG
state-feedback controller, optimizing the tip response. However, they pointed a
high performance sensitivity with respect to variations of the dynamics, which is
attributed to the non-collocated conguration. Wang and Vidyasagar [WV91b]
argued that the regulation of the tip position can be improved using a reected
tip position feedback, since the corresponding transfer function benets from the
minimum-phase property.
The essential diculties encountered for the control of linear slewing beams
also appear in the multi-link case. But more general approaches are necessary to
deal with the nonlinear dynamics caused by the conguration changes.

2.5.2 Quasi-linear control methods

Even though the dynamics of a multilink mechanism is nonlinear, the validity
of a linearized model about a particular operating point is sometimes acceptable,
2.5. CONTROL OF FLEXIBLE MECHANISMS 33

especially in case of small deformations and reasonable velocities. A pragmatic

approach consists in a linear control design, which is expected to be insensitive to
the nonlinearities. Oakley and Cannon [OC89] proposed an LQR controller for
a two-link mechanism, whose parameters are optimized for the nominal congu-
ration. For some applications, a satisfactory solution can be obtained combining
a gross-motion controller based on the rigidity assumption, with a linear vibra-
tion controller active close to the nal position. When many nal congurations
are possible, one vibration controller should be implemented for each linearized
local model, which may become cumbersome. Alternatively, De Fonseca [DF00]
has presented a linear H∞ robust controller for a 3-axis machine tool, where the
dependence of the dynamics with respect to the conguration of the machine
is treated as a model uncertainty. Stability is therefore guaranteed in the whole
workspace, but the author reports that this strategy leads to a rather conservative
design when compared to non-robust linear designs.
A better dynamic representation can be obtained using a linear time-varying
model (or even a nonlinear time-varying model), instead of a linear time-invariant
model. Assuming slow variations of the model parameters, an adaptive control
strategy [Lan79] can be developed: the basic idea is to construct a stable adap-
tation procedure for the controller, based on real-time measurements. The varia-
tions of the model are supposed to be unpredictable, and the adaptation is typi-
cally slow. Hence, adaptive control is best suited to deal with uncertainty in the
model, in the payload or in the external forces. It has been applied for the control
of exible manipulators by several authors [YBS89, GLL98, CU02, CYC02].
However, if the dependence of the linear model with respect to a few param-
eters is characterized a priori, and if the parameters are directly available from
measurements, it is more relevant to apply a gain scheduling control strategy.
For instance, the linear quadratic regulator theory can be extended for varying
systems, and this approach has been exploited by Meirovitch and Chen [MC95]
for the vibration control of exible space robots. Scott [Sco95] developed a gain-
scheduling controller for recongurable structures: (i) a set of mechanical cong-
urations is selected, (ii) for each of them a linear controller is designed from a lin-
earized model, (iii) the global controller follows from a smooth interpolation of the
local controllers in the conguration space. Of course, the global stability of this
strategy cannot be assessed theoretically. Symens [Sym04] investigated the active
34 CHAPTER 2. STATE OF THE ART

vibration control of a exible beam with varying length, in particular, he devel-

oped a low-authority controller in order to bring local damping near the resonance
frequencies. Following the hierarchical HAC/LAC design [Pre97, VBS02], this
low-authority vibration controller (LAC) is complemented with a high-authority
motion controller (HAC), responsible for the attenuation of wide-band distur-
bances on the rigid motion. Among various scheduling techniques reviewed in
his work, Symens points out the necessary trade-o between theoretical stabil-
ity guarantees and practical performances, as well as the dicult elaboration of
a low-order model able to capture the conguration-dependent dynamics. We
believe that the systematic reduced-order modeling technique presented in this
report gives a positive answer to this last problem.

2.5.3 Nonlinear control techniques

There is no general method for the control of the nonlinear dynamics of a
exible multilink mechanism. However, some tools have been proposed in the
literature in order to assist the control engineer in the design procedure.
Optimal control theory is a general framework for the construction of a con-
trol policy. The statement of a typical optimal control problem can be expressed
as follows: nd a feasible control such that the system starting from the given
initial conditions transfers its state to the target, and minimizes a performance in-
dex. Pontryagin's Maximum Principle and Bellman's Dynamic Programming are
two powerful tools for this class of variational problems. For linear systems, stan-
dard solutions are available for quadratic performance indexes (i.e. the linear
quadratic regulator), and for indexes expressed by the H∞ norm of the trans-
fer function of the system. For a class of rigid mechanisms5 , the minimum-time
motion is obtained with bang-bang commands. For more complex cases, the opti-
mal control problem can be solved numerically involving expensive computations.
Some general methodologies have been recently proposed for systems represented
by DAEs [SP01, SLP03], and for multibody systems [Str98, BCGF04, BC04].
The availability of a computationally ecient dynamic model is always a strong
advantage when applying such numerical methods.
5 Fully actuated mechanisms, free from singularities, without frictional eects and without
disturbances.
2.5. CONTROL OF FLEXIBLE MECHANISMS 35

There exists a special class of systems, called dierentially at systems, for
which there is a one-to-one correspondence between trajectories of a set of "at
outputs" and full state and input trajectories [FLOR95, VNM98, FLMR99]. This
special invertibility property leads to an ecient trajectory planning tool: a tra-
jectory specied in output space can be lifted to the state and input space through
an algebraic mapping. Ollivier and Sedoglavic [OS01] exploited this tool for the
trajectory planning of a very exible rod, and Thummel et al. [TOB01] for
exible joint robots. Since this formalism requires the analytical structure of the
dynamic equations, its systematic implementation for complex exible mecha-
nisms seems to be a dicult problem.
Feedback linearization is a geometric method to transform a nonlinear system
into a virtual linear system by a state transformation and a feedback transforma-
tion [SL91, Isi95]. Then, linear control theory can be applied for the virtual linear
system. For a rigid robot, this method is equivalent to the classical computed
torque technique. Sometimes, feedback linearization is classied among inverse
dynamics techniques, since it realizes a cancellation of the nonlinear part of the
dynamics. Feedback linearization typically requires a full state estimation, which
is not possible if part of the dynamics is not observable at the outputs. This unob-
servable dynamics is called the zero dynamics and it is interpreted as the system
dynamics when the outputs are forced to zero. By denition, a nonlinear system
is minimum-phase if its zero dynamics is asymptotically stable. Several authors,
such as Wang and Vidyasagar [WV91a], applied those advanced concepts to the
indirect vibration control of mechanisms with exible links, and they concluded
that the full state dynamics cannot be linearized. The choice of the outputs has a
decisive inuence on the stability of the zero dynamics: the collocated joint angles
lead to an oscillatory zero dynamics, whereas the noncollocated tip position leads
to an unstable behavior [WV91a, DLS93]. The output redenition concept can
also be generalized from the linear case [MPK00]. Feedback linearization relies
on an analytical expression of the dynamic equations, so that this tool may be
hardly applied for complex parallel mechanisms. Moreover, some authors [SJK97]
described the pitfalls of the feedback linearization technique, in particular its lack
of robustness, since it may involve useless and expensive cancellation of stabilizing
nonlinearities, using their dangerous destabilizing counterpart.
Singular perturbation theory has also been investigated for the control of ex-
36 CHAPTER 2. STATE OF THE ART

ible multilink manipulators [SB88, MC95, GLL98, MPK00, GS00, SV01, CYC02].
This theory assumes the partitioning of the state variables into slow and fast vari-
ables, so that the dynamics can be represented by two reduced subsystems: the
slow subsystem obtained by neglecting the fast oscillatory motion, and the fast
subsystem obtained by freezing the slow variables. A composite control can then
be applied, combining control laws established separately for each subsytems. In
the context of exible mechanisms, the elastic coordinates are naturally selected
as fast variables. The slow subsystem corresponds to a corrected rigid system
(e.g. including the static contribution of the exible variables), whereas the fast
subsystem has a linear dynamics, parametrically depending on the conguration.
The slow controller can be designed on the basis of well-established joint tracking
schemes for rigid manipulators, and the fast subsystems can be stabilized using
standard methods for linear systems [SB88, GLL98, GS00, SV01] or for linear
parameter-varying systems [MC95, MPK00, CYC02]. This theory is extremely
appealing, but, as pointed out by Book [Boo93], the performances are limited by
the necessary time-scale separation between the rigid and the exible controllers.
The Lyapunov method is a powerful tool for the stability analysis of non-
linear systems. However, for closed-loop systems, the construction of the con-
trol algorithm does not completely follow from this theory and involves engineer
judgement and intuition.
If the transfer function between the input and a chosen output is passive 6 ,
Lyapunov-type control schemes can be more systematically developed with guar-
anteed stability. Usually, this method leads to simple and robust feedback strate-
gies. For a exible mechanism, the passivity property of collocated transfer func-
tions can be easily assessed, leading to a classical joint controller. Dening a
passive output that includes deformation eects is a more dicult problem.
Considering plane mechanisms that are not inuenced by gravity, Ge et al.
[GLZ96] proposed an energy-based feedback which includes deformation eects
and leads to a stable closed-loop system in the sense of Lyapunov. An attractive
property of this strategy comes from the independence of the control design with
respect to the system dynamics, so that no model is necessary. However, a draw-
back comes from its conservatism (i.e. its low performances), and no guideline is
available to optimize the parameters of the controller.
6 An exact denition of passivity in system theory is given by Sepulchre et al. [SJK97].
2.5. CONTROL OF FLEXIBLE MECHANISMS 37

Specic nonlinear control design methods have been developed for systems
with a cascade structure, e.g. the states of one subsystem are the control variables
for the following. A step-by-step recursive design procedure can then be exploited,
such as backstepping or forwarding [SJK97]. Backstepping has been explicitly
applied for exible joint robots [OL99]. Recursivity is also at the basis of the
inertial damping concept developed for the control of macro/micromanipulators.
In the macro/micro conguration, a small rigid robot is mounted at the tip of a
large and exible manipulator. Several approaches have been proposed for the
control of such system, and an extended review is presented by George [Geo02].
Among them, the inertial damping method consists in controlling the motion of
the rigid robot in such a way that the inertia forces transmitted to the supporting
exible manipulator have a damping eect on the vibration modes. This is a true
cascade system: the control of the micromanipulator motion allows the active
damping of the macromanipulator.

2.5.4 Control strategy selected in this work

After this review, the compromise between stability, performances and ro-
bustness appears more than ever critical when designing a control law for a exible
mechanism. In chapter 6, the control of an experimental manipulator will be in-
vestigated on the basis of the two-time-scale approach: a slow joint-tracking con-
troller is complemented with a fast vibration controller. Related with HAC/LAC
control techniques and with approaches developed from singular perturbation
theory, this method has several advantages:

- a simple hardware and software implementation,

- an adjustable compromise between the joint tracking performances and the

vibration suppression,

- the absence of inversion or cancellation of the nonminimum phase dynamics,

so that the performances are less sensitive to modeling errors.

The fast vibration controller relies on an extension of the inertial damping

concept, which exploits the particular structure of the inertia forces in a low-order
mechanical model. Hence, the nonlinear model reduction technique developed in
38 CHAPTER 2. STATE OF THE ART

chapter 5 will be extremely valuable for the design and the implementation of the
control law.
 III

The Finite Element Approach in

Multibody Dynamics

According to chapter 2, the Finite Element method is appropriate for the

modeling and the simulation of complex parallel mechanisms with signicant
exibility. This chapter briey presents the approach developed by Géradin and
Cardona [CG88, Car89, CG89, GC01], which is at the center of our developments.
The motion parameterization, the dynamics and the time-integration strategy are
successively discussed.

3.1 Motion parameterization

In the literature, several coordinate choices have been proposed for the kine-
matic description of a mechanism, such as:

- the minimal coordinates,

- the relative coordinates, also denoted Lagrangian coordinates or joint coor-

dinates,

- the Cartesian coordinates, or reference point coordinates,

- the Finite Element coordinates, also denoted natural coordinates, or abso-

lute coordinates,

- the mixed coordinates.

40 THE FE APPROACH IN MULTIBODY DYNAMICS

(a) (b)

(c) (d)

Figure 3.1: (a) Minimal coordinates, (b) Relative coordinates, (c) Cartesian co-
ordinates, (d) Finite Element coordinates or natural coordinates.

We propose to illustrate these concepts for a simple planar mechanism; after-

wards, the representation of spatial rotations will be specically addressed.

3.1.1 Kinematic description of a planar mechanism

The rigid four-bar mechanism represented in Figure 3.1 has one kinematic
dof, and its description in terms of minimal coordinates requires a single parame-
ter: the angle θ. However, in Figure 3.1(a), two congurations can be associated
with a single value of θ, so that this parameterization is only valid for a restricted
part of the conguration space. The strong advantage of the minimal coordinates
comes from their independence, i.e. the absence of algebraic constraint.
The relative coordinates (Figure 3.1(b)) allow a recursive description of the
mechanical conguration. They are especially suitable for mechanisms with a
tree-topology, since their independence is then guaranteed. For a parallel mecha-
nism, such as the four-bar example, the relative coordinates are dependent, and
3.1. MOTION PARAMETERIZATION 41

connected by loop-closure constraints.

The Cartesian coordinates (Figure 3.1(c)) are dened as the position and
orientation of the center of mass of each body. The connections between bodies
are dened a posteriori, using appropriate kinematic constraints. This allows
a simple and systematic formulation, which leads to a rather large but sparse
system of equations.
Those three types of coordinates are competing solutions for the kinematic
description of rigid multibody systems. However, in exible multibody dynamics,
the deformations should be described with additional (modal) coordinates, lead-
ing to a hybrid set of coordinates. The coupling between the exible coordinates
and the other coordinates generally leads to a dicult formulation of the kinetic
energy. In this context, the Finite Element coordinates are more appropriate:
each body is characterized by a set of nodes, and each node is characterized by
its own set of coordinates, see Figure 3.1(d). Let us dene l1 , l2 , l3 and l4 as the
respective lengths of the four bars between the nodes 1-2, 3-4, 5-6 and 1-6. The
Finite Element coordinates are redundant, and three types of constraints are to
be considered:

- boundary nodal constraints:

x1 = y1 = y6 = 0 and x6 = l4 (3.1)

- assembly nodal constraints, imposed by boolean identication:

x2 = x3 , y2 = y3 , x4 = x5 , y4 = y5 (3.2)

- rigidity constraints, which may be expressed as a constant length require-

ment:
li2 = (x2i − x2i−1 )2 + (y2i − y2i−1 )2 i = 1, 2, 3 (3.3)

A mixed formulation may involve several types of coordinates. For instance,

despite our choice of absolute coordinates, it may be interesting to consider the
relative angle θ as an additional dof. Since the parameterization with absolute
coordinates is complete, θ is redundant, and connected to the absolute variables
by an algebraic constraint:

y2 cos θ − x2 sin θ = 0 (3.4)

42 THE FE APPROACH IN MULTIBODY DYNAMICS

Mixed coordinates will be extremely useful to develop our reduction technique,

which will be discussed later.

3.1.2 Description of spatial rotations

The denition of the Finite Element coordinates is straightforward for a
simple planar mechanism, as seen in the previous section, but for a spatial mech-
anism, the parameterization of the absolute nodal rotation is a challenging prob-
lem. Generally, a nite rotation can be represented using a proper unitary or
orthonormal 3 × 3 rotation matrix. Such a matrix can be described by three
independent parameters, and several choices have been proposed in the litera-
ture, such as the Euler angles, the Bryant angles, the Cartesian rotation vector,
the Rodrigues parameters, the conformal rotation vector or the linear parameters.
However, all those parameterizations suer from singularities. The four Euler pa-
rameters should also be mentioned, which lead to a non-singular but redundant
parameterization.
Due to the various advantages and drawbacks of each parameterization, the
optimal parameterization is problem-dependent. Following the conclusion of Car-
dona [Car89, page 65], the most relevant choices for a Finite Element formula-
tion in exible multibody dynamics are the Cartesian rotation vector and the
conformal rotation vector. The former was at the basis of the Samcef-Mecano
software [SAM99]; both of them are combined for the current implementation
in the Oofelie software [CKG94], for reasons that are beyond the scope of this
dissertation.

3.2 Dynamics
The Finite Element coordinates are absolute coordinates, and the total mo-
tion (rigid-body motion and elastic deformation) is directly referred to an inertial
frame. Due to the large displacements and rotations of the mechanical elements
with respect to this frame, the linear theory of elasticity is not applicable, and a
three-dimensional nonlinear theory is necessary.
An updated Lagrangian point of view for the rotation parameters has been
recommended by Géradin and Cardona [GC01]. This means that the rotations
3.2. DYNAMICS 43

are described as increments with respect to a previous conguration, so that the

singularities of the rotation parameterization are avoided, full symmetry of the
operators is obtained for the conservative loading case, and the time-integrator
is quite naturally extended to treat the rotational dofs.
The kinetic and potential energies K and V may be formulated using the
n × 1 vector of Finite Element coordinates q, and the equations of motion can
be derived from the Hamilton's principle:
Z t2
δ (L(q, q̇) + W) dt = 0 subject to Φ(q) = 0 (3.5)
t1

where the Lagrangian L is dened by

L = K − V, (3.6)

W denotes the virtual work of external non-conservative loads Q:

δW = δqT Q (3.7)

and Φ are m scleronomic constraints. The formulation may also include non-
holonomic constraints, but we shall not insist on this point here.
The problem (3.5) can be replaced by the equivalent formulation:
Z t2
δ (L(q, q̇) + W − λT Φ) dt = 0 (3.8)
t1

where λ is the m×1 vector of Lagrange multipliers associated with the constraints.
Performing the variation and the integration by part, one obtains the Lagrange
equations:   
d ∂L ∂L
 − + ΦTq λ = Q
dt ∂ q̇ ∂q
(3.9)
 Φ(q) = 0
where Φq denotes the constraint gradient. Assuming that K can be expressed as
a quadratic form of generalized velocities with a symmetric mass matrix M(q):
1 T
K= q̇ M(q) q̇ (3.10)
2
the equations of motion can be put in matrix form:
(
M q̈ + ΦTq λ = g(q, q̇)
(3.11)
Φ(q) = 0
44 THE FE APPROACH IN MULTIBODY DYNAMICS

where the vector of apparent forces g(q, q̇) collects external forces, internal forces
and complementary inertia forces:
 
∂V X X ∂Mki 1 ∂Mij
gk = Qk − − − q̇i q̇j (3.12)
∂qk i j
∂qj 2 ∂qk

This system of equations, and its linearized counterpart, can be directly used
for numerical analysis, e.g. in the time domain. Several enhanced formulations
have been proposed for more ecient numerical treatments. Among them, we
shall review the constraint elimination technique and the augmented Lagrangian
method.

3.2.1 The constraint elimination method

The constraint elimination method aims at replacing the initial problem by an
equivalent unconstrained problem. It relies on a partitioning of the n generalized
coordinates q into two subsets: the n − m independent coordinates θ, and the m
dependent coordinates q∗ . Then, we seek for an explicit elimination formula:

q∗ = Φ∗ (θ) (3.13)

Since the m algebraic constraints Φ are nonlinear and implicit, the expression of
Φ∗ cannot be constructed analytically. However, the Jacobian of the transforma-
tion can be computed by implicit dierentiation:
" #
∗ I
∂q ∂q
= −Φ−1
q∗ Φθ ⇒ = qθ = (3.14)
∂θ ∂θ −Φ−1
q∗ Φθ

Numerical treatments usually rely on Newton-Raphson iterations of the linearized

dynamic equations. At this level, the Jacobian qθ can be exploited to transform
the initial linearized problem in terms of the redundant variables q into a reduced
linearized problem in terms of the independent variables θ. This technique is
explained with more details by De Jalón and Bayo [DJB94].

3.2.2 The augmented Lagrangian method

The objective of the augmented Lagrangian method is to improve the nu-
merical conditioning of the Lagrange multiplier method by adding a penalty term
3.2. DYNAMICS 45

of moderate amplitude. At the same time, appropriate constraint scaling is in-

troduced in order to generate system matrices of the same order of magnitudes.
The augmented functional of Hamilton's principle (3.8) is
Z t2
δ (L(q, q̇) + W − k λT Φ − p ΦT Φ) dt = 0 (3.15)
t1

where k and p are the scaling and penalty factors. The dynamic equations follow:
(
M q̈ + ΦTq (p Φ + k λ) = g
(3.16)
k Φ(q) = 0

Since the penalty term vanishes at the solution point, it is easily observed that
the method provides the exact solution of the initial problem.
Since the time-integration procedure involves a Newton-Raphson procedure,
it is important to formulate the linearized equations for the corrections of the
displacements, velocities and accelerations:
" # " # " # " # " # " #
M 0 ∆q̈ Ct 0 ∆q̇ Kt k ΦTq ∆q
+ +
0 0 ∆λ̈ 0 0 ∆λ̇ k Φq 0 ∆λ
" #
−resq
= Φ
+ O(∆2 )
−res
(3.17)
where

- resq and resΦ denote the residual vectors of equilibrium and constraints,
with:
resq = −g + M q̈ + ΦTq (p Φ + k λ) (3.18)

- Ct is a tangent damping matrix resulting from the variation of the gyro-

scopic forces with respect to the velocities:
∂g
Ct = − (3.19)
∂ q̇

- Kt is a tangent stiness matrix resulting from variations with respect to

the displacements:


∂g ∂(M q̈) ∂ ΦTq (p Φ + k λ)
Kt = − + + (3.20)
∂q ∂q ∂q
46 THE FE APPROACH IN MULTIBODY DYNAMICS

Usually, a good approximation is given by:

∂g
Kt ' − + p ΦTq Φq (3.21)
∂q
The penalty term adds a positive denite contribution to the tangent sti-
ness, which improves the numerical conditioning and the convergence of the
Newton-Raphson iterations.

3.3 Time-integration
In chapter 2, the relevance of the Newmark family of implicit algorithms for
the simulation of exible mechanisms was highlighted. This section presents the
original Newmark scheme [New59] and its generalized-α extension by Chung and
Hulbert [CH93]. Theoretical results will also be detailed, related with the issues
of stability and convergence.

3.3.1 Newmark algorithm

The original Newmark integration formula are obtained from a Taylor series
expansion of the displacements and velocities, with respect to the time-step size h:

qn+1 = qn + h q̇n + h2 ( 12 − β) q̈n + h2 β q̈n+1

(3.22)
q̇n+1 = q̇n + h (1 − γ) q̈n + h γ q̈n+1

where the constants β and γ are numerical parameters. It may be shown that the
optimal choice of the parameters corresponds to the average constant acceleration
formula, obtained with the particular values:
1 1
γ= and β= (3.23)
2 4
They give maximal second-order accuracy, and unconditional linear stability, as
will be demonstrated later. Numerical damping can be introduced in the New-
mark formula according to:
 2
1 1 1
γ = +α and β= γ+ α>0 (3.24)
2 4 2
where α is the numerical damping parameter. This choice allows to increase the
numerical damping in the system while remaining on the stability boundary of the
3.3. TIME-INTEGRATION 47

integration scheme. A more systematic presentation of the Newmark algorithm

in structural dynamics is described by Géradin and Rixen [GR97].
The numerical solution of the dynamic equations (3.16) follows an iterative
predictor-corrector scheme, exploiting the linearized equations (3.17). From given
values of qn , q̇n , q̈n at time tn , an initial prediction at time tn+1 is obtained with
the zero acceleration assumption:
q̈0n+1 = 0
q̇0n+1 = q̇n + h (1 − γ) q̈n (3.25)
q0n+1 = qn + h q̇n + h2 ( 21 − β) q̈n

and to satisfy the Newmark formulae (3.22), the iterative corrections should sat-
isfy:
1
4q̈n+1 = 4qn+1
β h2
γ
(3.26)
4q̇n+1 = βh
4qn+1
so that the linearized system (3.17) becomes:
" # " # " #
Sqq k ΦTq ∆q −resqk
t
= (3.27)
k Φq 0 ∆λ −resΦ
k

with the augmented tangent matrix:

γ 1
Sqq
t = Kt + Ct + M (3.28)
βh β h2
The solution to the system (3.27) is the correction for the Newton-Raphson pro-
cedure. The complete integration algorithm is summarized in Figure 3.2.

3.3.2 Generalized-α method

Due to the Finite Element discretization and to the presence of algebraic
constraints, the dynamic equations are numerically sti, and high-frequency nu-
merical damping is desirable. However, the stabilized Newmark scheme is only
rst-order accurate. Several alternatives have been proposed in order to introduce
numerical damping at high-frequency without degrading the order of accuracy.
Among them, the generalized-α method described by Chung and Hulbert [CH93]
includes as particular cases the most important integration schemes in structural
dynamics, such as the Hilber-Hughes-Taylor algorithm [HHT77], and it is there-
fore a very general framework for theoretical investigations. It consists of keeping
48 THE FE APPROACH IN MULTIBODY DYNAMICS

Time incrementation
-
t=t+h

Initial prediction
?

q0 , q̇0 , q̈0 (eqn (3.25))

Evaluation of residues
?


resq (q, q̇, q̈, λ), resΦ (q)

Check for convergence

?
yes
kresq k < , kresΦ k < η
no ?
Evaluation of corrections
4q, 4λ (eqn (3.27))

Incrementation
?

q, q̇, q̈, λ

Figure 3.2: Newmark algorithm

3.3. TIME-INTEGRATION 49

the Newmark formulae (3.22), whereas the residual equations are modied by
averaging the dierent contributions between both time instants:
(
∗
(1 − αm ) (M q̈)n+1 + αm (M q̈)n + (1 − αf ) gn+1 + αf gn∗ = 0
(3.29)
(1 − αf ) k Φn+1 + αf k Φn = 0
where g∗ is a notation for ΦTq (k λ + p Φ) − g, while αm and αf are numer-
ical parameters. The Hilber-Hughes-Taylor algorithm is obtained for αm = 0
and αf ∈ [0, 1/3]. The optimal parameters of the generalized-α method can be
computed from the desired spectral radius at innity ρq∞ :
2 ρq∞ − 1 ρq
αm = q and αf = q ∞ (3.30)
ρ∞ + 1 ρ∞ + 1
Dening αf m = αf − αm , the Newmark parameters are now given by:
 2
1 1 1
γ = + αf m and β= γ+ (3.31)
2 4 2
It will be demonstrated later that this scheme is second-order accurate even for
αf m > 0.
The numerical solution is obtained with a predictor-corrector strategy as for
the Newmark algorithm. At the correction step, the augmented tangent matrix
becomes:
γ 1
Sqq
t = (1 − αf ) Kt + (1 − αf ) Ct + (1 − αm ) M (3.32)
βh β h2
and the linearized equations:
" # " # " #
Sqq k (1 − αf ) ΦTq ∆q −resqk
t
= (3.33)
k (1 − αf ) Φq 0 ∆λ −resΦ
k

where the residues are computed according to (3.29).

Remark 3.1 As a consequence of the modication of the residual equation in the
generalized-α method, the acceleration q̈n+1 is a poor rst-order approximation for
the true acceleration q̈(tn+1 ) [EBB02]. Assuming that an order 2 approximation
is available for qn+1 , q̇n+1 and λn+1 , an order 2 approximation an+1 would satisfy
the original residual equation at time tn+1 :
(M a)n+1 + g∗ (qn+1 , q̇n+1 , λn+1 ) = 0 (3.34)
Using the rst equation of (3.29) to eliminate g∗ , an equivalent and useful con-
dition is obtained if the matrix M is constant and non-singular:
(1 − αf ) an+1 + αf an = (1 − αm ) q̈n+1 + αm q̈n (3.35)
50 THE FE APPROACH IN MULTIBODY DYNAMICS

3.3.3 Linear stability analysis

When dealing with sti systems of equations, the stability of the integration
scheme is critical for a successful simulation. In exible multibody dynamics, the
presence of algebraic constraints strongly aects the stability of the numerical
solution. For unconstrained problems, the stability of the generalized-α method
is analyzed by Chung and Hulbert [CH93] in the linear regime, and by Erlicher et
al. [EBB02] in the nonlinear regime. For constrained problems, linear stability
results are presented by Géradin and Cardona [GC01].
In this thesis, all stability analyses are realized for linear systems; as Géradin
and Cardona [GC01], we believe that these results are relevant indications of the
stability properties for nonlinear systems. This section presents a demonstration
of the unconditional stability of the generalized-α method inspired by the work
of Géradin and Cardona [GC01]; the unconstrained and constrained cases are
successively addressed. In chapter 4, those developments will be extended for the
stability analysis of our integrated simulation method.

Unconstrained system
The stability analysis of an integration scheme applied to a linear system of
equations:
M q̈ + K q = 0 (3.36)

is hereby conducted according to the developments of Chung and Hulbert [CH93].

The problem is simplied by invoking the diagonalization of the equations with
a linear transformation. It is straightforward that the numerical behavior will be
the same for the initial and the transformed systems. Since each modal equation
can be analyzed independently, it is sucient to consider the stability of the
integrator applied to the scalar test equation

q̈ + ω 2 q = 0 (3.37)

According to the generalized-α methodology, the time-discretized expression is

(1 − αm ) q̈n+1 + αm q̈n + (1 − αf ) ω 2 qn+1 + αf ω 2 qn = 0 (3.38)

3.3. TIME-INTEGRATION 51

Complementing this equation with the Newmark formulae (3.22), we get the
matrix relation
     
(1 − αf )ω 2 0 (1 − αm ) q αf ω 2 0 αm q
     

 1 0 −βh2   q̇ 
  +
 −1 −h −( 21 − β)h2 
  q̇  = 0
 
0 1 −γh q̈ 0 −1 −(1 − γ)h q̈
n+1 n
(3.39)
and we seek its eigensolutions, characterized by a synchronous behavior:
   
q q0
with (3.40)
   
 q̇  = ϕn  q̇0  ϕn+1 = ζ ϕn
   
q̈ q̈0
n

ϕ is the amplitude, [q0 q̇0 q̈0 ]T is the eigenvector, and ζ is the eigenvalue. If we
dene the polynomials,
P (ζ) = (1 − ) ζ +   = {αm , αf }
Pγ (ζ) = γ ζ + (1 − γ)
(3.41)
Pβ (ζ) = β ζ + ( 12 − β)
P1 (ζ) = ζ − 1
the characteristic equation is
 
Pαf (ζ)ω 2 0 Pαm (ζ)
q
(ζ) = det  (3.42)
 
Pωh  P1 (ζ) −h −Pβ (ζ)h2 
=0
0 P1 (ζ) −Pγ (ζ)h
which is equivalent to:
q
Pωh (ζ) = Pαf (Pγ + P1 Pβ ) (ωh)2 + P12 Pαm = 0 (3.43)
The algorithm is stable if all the roots of Pωh
q
(ζ) = 0 are inside the unit circle.
Unconditional stability requires this property to be satised whatever the value
of ωh. Therefore, an important property of the algorithm is the spectral radius
associated with the roots ζiq of Pωhq
for ωh → ∞:
ρq∞ = max |ζiq | (3.44)
i

which should be less than one for unconditional stability. This spectral radius is
computed from the characteristic equation:
q
P∞ = Pαf (Pγ + P1 Pβ ) = 0 (3.45)
52 THE FE APPROACH IN MULTIBODY DYNAMICS

If we consider the optimal parameters β and γ given by (3.31), this equation has
one simple root and one double root:

−αf 1 − αf m
ζ1q = q
ζ2,3 =− (3.46)
1 − αf 1 + αf m

ζ1q is the so-called spurious root, and ζ2,3

q
are the principal roots. A usual design
requirement is |ζ1q | ≤ |ζ2,3
q
|. The condition |ζ1q | = |ζ2,3
q
| = ρq∞ leads to the optimal
formulae (3.30) proposed by Chung and Hulbert [CH93], and the linear stability
is then guaranteed by the design constraint ρq∞ < 1. It is easy to verify that the
Hilber-Hughes-Taylor algorithm has a spectral radius less than one.

Constrained system
The linearized equations of an undamped system are:

M q̈ + K q + BT λ = 0
(3.47)
Bq = 0

As pointed out by Géradin and Cardona [GC01], this system is not diagonalizable
due to the presence of the constraints. However, these authors demonstrated that
it could be transformed into a canonical form by a linear transformation:

Theorem 3.1 If K is symmetric positive semi-denite, M is symmetric positive

denite, and B has full rank, there exists a transformation matrix Tm such that
 
" # Ω2 0 0
K BT
(3.48)
 
TTm Tm = 
 0 0 Icc 
B 0 
0 Icc 0
 
" # Irr 0 0
M 0
(3.49)
 
TTm Tm = 
 0 I
cc
0 
0 0 
0 0 0

where Ω2 is a (n − m) × (n − m) diagonal matrix, Icc is the identity matrix of

dimension m × m, Irr is the identity matrix of dimension (n − m) × (n − m), and
0 are null matrices with appropriate dimensions.
3.3. TIME-INTEGRATION 53

This transformation matrix Tm leads to the denition of a new set of me-

chanical variables qr , qc , and qλ :
   
" # qr " # qr
q Tqr Tqc 0
(3.50)
   
= Tm  c  =  qc 
q
λ 
λ
 Tλr Tλc Tλλ 
λ

q q

qc are m constrained dofs, qλ are m multipliers-like dofs, and qr are n − m

independent dofs. Actually, those names do not reect the true complexity of the
transformation, since the Lagrange multipliers λ are also connected with qr and
qc .
After transformation, the set of dynamic equations becomes:

q̈r + Ω2 qr = 0
q̈c + qλ = 0 (3.51)
qc = 0

Following the procedure detailed for the unconstrained case, the characteristic
equation is the determinant of the following matrix:
 
Pαf Ω2 0 0 0 0 0 Pα m I 0 0
 
 0 0 Pα f I 0 0 0 0  Pα m I 0
 

 0 Pα f I 0 0 0 0 0 0 0 

 

 P1 I 0 0 −h I 0 0 −Pβ h2 I 0 0 

0 P1 I 0 0 −h I 0 0 −Pβ h2 I 0
 
 
 

 0 0 P1 I 0 0 −h I 0 0 −Pβ h2 I 

 

 0 0 0 P1 I 0 0 −Pγ h I 0 0 

0 0 0 0 P1 I 0 0 −Pγ h I 0
 
 
0 0 0 0 0 P1 I 0 0 −Pγ h I
(3.52)
After lengthy but straightforward algebraic manipulations, and referring to the
characteristic polynomial Pωhq
associated with the unconstrained problem, one
nds that the characteristic equation is equivalent to:
 
Pαf (ζ) Ω2 0 Pαm (ζ) I
q
(P∞ (ζ))2m det −h I −Pβ (ζ)h2 I  = 0 (3.53)
 
P1 (ζ)I
0 P1 (ζ)I −Pγ (ζ)h I
54 THE FE APPROACH IN MULTIBODY DYNAMICS

The second factor can be interpreted as the characteristic equation when the
algorithm is applied to an unconstrained diagonal system, so that equation (3.53)
can be restated:
(3.54)
Y q
q
(P∞ (ζ))2m Pωi h (ζ) = 0
i

where ωi are the diagonal terms of Ω. Any root of this equation is either a
root of the polynomial Pωqi h obtained in the unconstrained case, or a root of the
same polynomial at innite frequency. We conclude that if the generalized-α
algorithm leads to a stable integration of the unconstrained problem, and if the
spectral radius at innity ρq∞ is smaller than 1, the global stability is guaranteed.
If ρq∞ = 1, it can be demonstrated that the numerical integration is weakly
unstable [GC01].

3.3.4 Convergence analysis

This section investigates the second-order accuracy of the generalized-α al-
gorithm, and as for the linear stability, the analysis starts with the unconstrained
case.

Local truncation error: the unconstrained case

If the mass matrix is constant and non-singular, it is equivalent to study the
behavior of the algorithm applied to the system:

q̈ = f (q, q̇, t) (3.55)

According to the generalized-α algorithm, the discrete-time equation is

(1 − αm ) q̈n+1 + αm q̈n = (1 − αf ) fn+1 + αf fn (3.56)

In this work, qn+1 , q̇n+1 , and q̈n+1 conventionally denote the numerical solutions,
whereas q(tn+1 ), q̇(tn+1 ), and q̈(tn+1 ) refer to the exact solutions. The numerical
solutions qn+1 , q̇n+1 satisfy the Newmark formulae (3.22), and the dierences
with the exact solutions q(tn+1 ), q̇(tn+1 ) are the local truncation errors:

en+1 = q(tn+1 ) − qn+1 (3.57)

e0n+1 = q̇(tn+1 ) − q̇n+1 (3.58)
3.3. TIME-INTEGRATION 55

The order 2 condition is:

en+1 = O(h3 ) (3.59)
e0n+1 = O(h3 ) (3.60)
In order to exploit the theory of linear multistep algorithms, it is useful to
note that the one-step generalized-α method hides a two-step linear formula for
the velocities, and a three-step linear formula for the displacements [EBB02].
Indeed, from the integration formulae (3.22) and (3.56) developed at time tn+1
and tn+2 , the elimination of the accelerations q̈n , q̈n+1 and q̈n+2 leads to the
two-step formulae:
2 1 2
(3.61)
X X X
αk qn+k + h µk q̇n+k = h2 βk fn+k
k=0 k=0 k=0
2 2
(3.62)
X X
αk q̇n+k = h γk fn+k
k=0 k=0

with
α0 = −αm , α1 = −1 + 2αm , α2 = 1 − αm ,
µ0 = −αm , µ1 = −1 + αm ,
(3.63)
β0 = (1/2 − β) αf , β1 = 1/2 − β − αf /2 + 2 β αf , β2 = β (1 − αf ),
γ0 = (1 − γ) αf , γ1 = 1 − γ − αf + 2 γ αf , γ2 = γ (1 − αf ).
and a further elimination of the velocities in the rst equation would lead to a
three-step formula for the displacements. Equation (3.62) is a good basis for the
analysis of the local truncation error at the velocity level. Since it is a standard
multistep formula, the order 2 condition can be found in classical textbooks such
as Hairer et al. [HNW87]:
α0 + α1 + α2 = 0 (3.64)
α1 + 2 α2 = γ0 + γ1 + γ2 (3.65)
α1 + 4 α2 = 2 γ1 + 4 γ2 (3.66)
These relations come from the substitution of q̇ and f by their Taylor series
expansion in the expression of the local error. The rst two equations are auto-
matically satised by the parameters given in (3.63), whereas the last equation
yields
1
γ = + αf − αm (3.67)
2
56 THE FE APPROACH IN MULTIBODY DYNAMICS

which is the order 2 condition for the generalized-α method (3.31).

A the displacement level, the order condition is obtained by a similar analysis
of the three-step formulation. For conciseness, the demonstration is not repro-
duced here; it has been developed by Erlicher et al. [EBB02].
Those results rely on the initial assumption of a constant mass matrix. In
exible multibody dynamics, the mass matrix is actually varying between tn and
tn+1 , so that the discretized equation becomes

(1 − αm ) Mn+1 q̈n+1 + αm Mn q̈n = (1 − αf ) gn+1 + αf gn (3.68)

which is not equivalent to equation (3.56). However, a deeper analysis of the

updated Lagrangian formulation shows that the variation of the mass matrix
over one time-step is usually of order 2, leading to O(h2 ), O(h3 ) and O(h4 ) local
truncation error in q̈n+1 , q̇n+1 and qn+1 , respectively. Therefore, the order 2
condition is usually observed for the displacements and velocities.

Global convergence: the unconstrained case

The local truncation error analysis gives some information about the error
over one time-step. However, when a nite time-interval is considered, the prop-
agation of the errors may jeopardize the convergence of the numerical solution.
According to Hairer et al. [HNW87], a linear multistep algorithm is con-
vergent of order p if it is stable and if its local truncation error is of order p. In
particular, let us consider the nx dimensional rst-order ODE:

ẋ = f (x, t) (3.69)

solved using a k-step stable method, whose local truncation error satises the
order p condition
kx(tn+1 ) − xn+1 k ≤ M hp+1 (3.70)

We also dene the (k nx ) × 1 vectors collecting the coordinate vectors of the last
k steps:

XTi = (xTi+k−1 , ..., xTi ) (3.71)

X(ti )T = (x(ti+k−1 )T , ..., x(ti )T ) (3.72)
3.3. TIME-INTEGRATION 57

The following bound on the global error after n time-steps is given by Hairer et
al. [HNW87]
M hp nhL∗
(3.73)
∗
kX(tn ) − Xn k ≤ kX(t0 ) − X0 k enhL +


e − 1
L∗
where L∗ is a Lipschitz constant. An initial error in the initial conditions is at
most amplied by a coecient depending exponentially on the length of the time
interval nh. Therefore, it is especially important to limit this phenomenon by the
denition of consistent initial conditions.
Since the generalized-α method hides a multistep algorithm, the error prop-
agation might be characterized by a similar property, and the consistency of the
initial conditions is of practical interest to guarantee the global convergence. For
a multistep method, the initial conditions are given as the vector X0 of the solu-
tion at the k rst steps, which is necessary to initiate the integration algorithm.
However, in its pseudo-one-step formulation, the generalized-α algorithm starts
on the basis of the values q0 , q̇0 and q̈0 . For full consistency, q̈0 should satisfy:

(1 − αm ) q̈0 + αm q̈−1 = (1 − αf ) f0 + αf f−1 (3.74)

In practice, this equation is helpless for the determination of q̈0 , and it is usually
replaced by the simplied consistency condition:

q̈0 = f0 (3.75)

This expression satises (3.74) with an O(h) error. From a dimensional analysis,
this initial error at the acceleration level will contaminate the displacements and
velocities during the rst steps with O(h3 ) and O(h2 ) errors, respectively. This
error in the initial steps will be coupled, propagated and amplied by the integra-
tion procedure, leading to a maximal global error of O(h2 ), so that second-order
convergence is still guaranteed. However, if the simplied consistency condi-
tion (3.75) is not satised, the error in the initial values may drop one order,
with disastrous consequences on the global convergence of the results.

The constrained case

The previous convergence results can be extended for constrained systems.
In Figure 3.2, the time-integration algorithm developed for exible multibody
58 THE FE APPROACH IN MULTIBODY DYNAMICS

systems guarantees that the algebraic constraints:

Φ(q) = 0 (3.76)

are satised at each time-step, if they were satised by the initial conditions.
Hidden constraints should also be satised by the dynamic variables at all dier-
entiated levels, and in particular, at the velocity and acceleration levels:

Φ̇ = Φq q̇ = 0 (3.77)
Φ̈ = Φ̇q q̇ + Φq q̈ = 0 (3.78)

Those equations are, however, not considered in the algorithm, which might lead
to signicant errors.
In the special case of linear constraints, Φq is constant, and the constraints
at the acceleration level become:

Φq q̈ = 0 (3.79)

After inspection of the predictor and corrector steps of the numerical algorithm,
the constraints at displacement, velocity and acceleration levels are satised at
tn+1 if they were initially satised at tn . Therefore, the local truncation error is
of order 2, as in the unconstrained case.
Since the constraints are satised at each time grid point, the constraint vi-
olation of the numerical solution is propagated at a very high frequency which
depends on the step size. In agreement with the stability analysis, the ampli-
cation factor of this phenomenon is associated with the spectral radius of the
integration algorithm at high frequencies. Therefore, the numerical damping is
responsible for the decay of this error.
In case of nonlinear but smooth constraints, if we do not consider nonlinear
amplication eects, the violation of the hidden constraints resulting from the
nonlinearity may be seen as a disturbance, which is ltered by the algorithm in
the same manner. In this sense, those errors do not accumulate throughout the
integration process, and the convergence results demonstrated for unconstrained
systems are still relevant.
From a practical perspective, the computation of consistent initial veloci-
ties and accelerations may rely on equations (3.77) and (3.78) (or its linearized
counterpart (3.79)).
3.4. IMPLEMENTATION 59

3.4 Implementation
The concepts presented in this chapter have been implemented in the Mecano
program of the Samcef software [SAM99]. In 2005, this program has reached
industrial maturity for several years, it is well documented and it contains a huge
library of nite elements and types of analysis, which is very convenient for the
user. Mecano has been extensively exploited for problems in aerospace, robotics,
automotive engineering and machine-tools.
However, for the developments realized in this research, the Oofelie multi-
physics platform [CKG94] has been selected for several reasons. Oofelie is an
acronym for Object Oriented Finite Element Led by Interactive Executor. It is
written in C++, and modern programming concepts are favourable for an open
architecture, well-suited for new developments. Its kernel has been developed
to deal with strongly coupled multiphysics problems, with great care about the
modularity and eciency issues. Since the developments related with exible
multibody dynamics only started a couple of years ago, a restricted library of the
most signicant elements and algorithms is currently available in Oofelie. Nev-
ertheless, those capabilities are sucient to demonstrate the relevance and the
eciency of the innovative concepts developed in this research.
60 THE FE APPROACH IN MULTIBODY DYNAMICS
 IV

Integrated Simulation of Mechatronic

Systems

In the introduction, the motivation for an integrated simulation tool able to

predict the dynamic behavior of a mechatronic system was clearly demonstrated.
The analysis should account for all the technological components: the mechanism,
the control system, the actuators, the sensors, etc. We have also seen that the
Finite Element method is a general and reliable modeling approach able to deal
eciently with complex exible mechanisms.
This chapter presents some extensions of the Finite Element method for the
simulation of mechatronic systems. The description of non-mechanical elements
within a Finite Element context is not an innovative idea, and for instance, user-
elements are available in most industrial software for multibody systems. How-
ever, the implementation of a user-element is an intricate, time-consuming and
sometimes unreliable process, which motivates a more systematic and theoreti-
cally founded approach. Hence, we propose to integrate the block diagram lan-
guage within the Finite Element framework, as we have suggested in [BDG04b].
Relying on a nonlinear state-space description, this approach presents the advan-
tages of modularity, generality, with a language very familiar to control engineers.
The methodology developed here is illustrated in Figure 4.1. The mechanical
system is modeled using the Finite Element formalism for exible multibody
systems presented in chapter 3, whereas the control system is described using the
block diagram language. The strongly coupled mechanical and state equations
are obtained by numerical assembly, and their time-domain simulation is based
62 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

Mechanism Control system

(Multibody formalism) (Block diagram language)
@
R
@
Strongly coupled DAEs

?
Generalized-α solver

Figure 4.1: Strongly coupled and modular simulation.

on an extension of the generalized-α method.

In a mechatronic system, discontinuous phenomena may occur, e.g. due
to the sampling eect in a digital electronic system or to the saturation of a
component. From a numerical point of view, a discontinuous transition requires
a computation restart procedure, which is also considered here.
The organization of the chapter is as follows. First, a state-space formalism
for the simulation of control systems is presented in section 4.1, with convergence
and stability results. Section 4.2 describes the integrated simulation method and
analyzes its stability properties. A specic treatment for systems with discon-
tinuous phenomena is developed in section 4.3. After the presentation of a few
implementation issues in section 4.4, three examples with increasing complex-
ity illustrate the method: a four-bar mechanism, a Scara robot, and a vehicle
semi-active suspension.

4.1 Simulation of control systems

Figure 4.2 illustrates the interactions between a mechanism and a continu-
ous control system. The control system includes all the non-mechanical elements,
such as the controller, the actuators and the sensors. Mathematically, the inter-
actions are represented by two vectors: the nm × 1 vector wm contains the sensor
measurements exploited by the control system and any other mechanical variables
that may inuence its dynamic behavior, whereas the na × 1 vector ga represents
the generalized forces exerted by the actuators on the mechanism.
This section focuses on the description of the control system using the block
4.1. SIMULATION OF CONTROL SYSTEMS 63

ga- wm
Mechanism

Control system 

Figure 4.2: Interactions between a mechanism and a control system.

diagram language, and on the simulation of the resulting state-space model. The
simulation of the coupled system will be considered later, in section 4.2.

4.1.1 State equations

The most general formalism for the description of the control system is cer-
tainly the descriptor state-space format:

f s∗ (wm , x, ẋ, t) = 0
(4.1)
f o∗ (wm , x, ga , t) = 0

where the nx × 1 vector x represents the state variables, f s∗ are the nx dierential
equations of the dynamic states, whereas f o∗ are the na algebraic output equa-
tions. Control engineers are more familiar with the explicit state-space format,
which is a special case of the descriptor state-space format:

ẋ = f s (wm , x, t)
(4.2)
ga = f o (wm , x, t)

In general, the transformation from (4.1) to (4.2), is not always possible nor
trivial.
Such a global input/output black box description is compact and ecient.
However, a functional decomposition into subsystems may lead to an advanta-
geous modular approach, as illustrated in Figure 4.3. At the subsystem level,
the explicit state-space equations can be formulated more easily, and we consider
that each element e is characterized by explicit state equations, with respect to
64 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

ga- wm
Mechanism

y(2) (2) y(1) (1)

System 2 u System 1
u


u(3)- System 3 y(3)

Figure 4.3: Modular approach for the description of the control system.

its own inputs u(e) and outputs y(e) :

ẋ(e) = f s(e) (u(e) , x(e) , t)

(4.3)
y(e) = f o(e) (u(e) , x(e) , t)

Hence, one denes the global input and output vectors:

   
u(1) y(1)
.. .. (4.4)
. .
   
u=

,
 y=



u(ne ) y(ne )

with the connectivity relations:

u = Lim wm + Lio y and ga = Lao y (4.5)

where Lim , Lio and Lao are boolean localization matrices.

The assembled equations of the block diagram model are summarized by:

ẋ = f s (u, x, t) (4.6)
y = f o (u, x, t) (4.7)
u = Lim wm + Lio y (4.8)

with the actuator outputs:

ga = Lao y (4.9)
4.1. SIMULATION OF CONTROL SYSTEMS 65

u and ga are explicitly dened by boolean equations; in the numerical code,

they are not implemented as independent variables. However, they appear in the
formal presentation for notational convenience.
Those equations are equivalent to the descriptor state-space equations (4.1),
and we observe that the transformation to the explicit form (4.2) would require
the elimination of u and y, using an algebraic transformation. This means that
even though we restrict the subsystems description to the ODEs (4.3), the as-
sembled equations are DAEs, where the algebraic equations come from the in-
put/output connections between subsystems. The theoretical results of the time-
integration scheme will rely on an additional technical assumption on the alge-
braic structure of those DAEs, which is presented hereafter.

4.1.2 Semi-explicit index-1 assumption

In the general theory of DAEs [BCP96], a semi-explicit index-1 system is
written as nx state equations and nz constraint equations:
ẋ = f (x, z, t) (4.10)
0 = g(x, z, t) (4.11)
with the property that the nz × nz Jacobian gz exists and is invertible in the
neighborhood of the exact solution. From the implicit function theorem, g can
thus be solved for the algebraic variables z in terms of x and t (i.e., the function
z = ge(x, t) can be dened locally). Hence, a very natural implementation of
a time-integration scheme is to require the constraint (4.11) to be satised by
the algebraic variables z at each time-step, while applying a standard multistep
formula to (4.10) for the time-integration of the dynamic variables x. Indeed, the
algorithm behaves as if it were applied to the equivalent ODE:
ẋ = f (x, g
e(x, t), t) (4.12)
A multistep integrator applied to semi-explicit index-1 DAEs in this manner is
thus stable and convergent to the same order of accuracy as if it were applied to
standard sti ODEs.
In our case, the input and output variables are the algebraic variables, and
the block diagram model (4.6), (4.7) and (4.8) is a semi-explicit index-1 system
if the Jacobian satises the following regularity condition.
66 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

u- y
System

Figure 4.4: A trivial system.

Assumption 4.1 " #

I −fuo
det 6= 0 (4.13)
−Lio I

In this case, the implicit function theorem can be invoked to solve (4.7) and (4.8)
for the inputs and outputs:

y = feo (wm , x, t) (4.14)

u = feu (wm , x, t) (4.15)

and the numerical algorithm behaves as if it were applied to the equivalent ODE:

ẋ = fes (wm , x, t) (4.16)

with fes (wm , x, t) = f s (feu (wm , x, t), x, t).

Assumption 4.1 is not satised if the block diagram model hides algebraic
constraints for the states x and/or for the mechanical inputs wm . This patho-
logical situation can be illustrated with the trivial system described by the scalar
linear equations (Figure 4.1.2):

ẋ = a x + b u
y = cx + u (4.17)
u = y

The input/output equations hide the constraint x = 0, so that the output variable
y = u plays the role of a Lagrange multiplier in the dynamic equation. The block
diagram formalism presented here is not really appropriate for such a higher-
index DAE. In the following, Assumption 4.1 is supposed to be satised, so that
an ecient strategy can be developed with guaranteed reliability.
4.1. SIMULATION OF CONTROL SYSTEMS 67

4.1.3 Time-integration algorithm

This section presents a methodology for the simulation of a block diagram
model isolated from the mechanical system. Therefore, we consider that the mea-
surements are known functions of time wm = wm (t).
A simple and interesting integrator for rst-order ODEs is the trapezoidal
rule:
h
xn+1 = xn + (ẋn + ẋn+1 ) (4.18)
2
This formula is second-order accurate and A-stable, but it is not appropriate
for sti problems, due to the absence of numerical ltering at high frequen-
cies [HW91]. An alternative, stabilized scheme is given by

xn+1 = xn + h (1 − θ) ẋn + h θ ẋn+1 θ ∈ [1/2, 1] (4.19)

with only rst-order accuracy for θ 6= 12 .

In order to introduce high-frequency numerical damping without tremendous
accuracy loss, improved integration formula are necessary, such as multistep or
Runge-Kutta methods [HW91]. The default of the one-step formula (4.19) is
similar to the default of the original Newmark formulae. Introducing the modied
residual equation, Hilber, Hughes and Taylor [HHT77], followed by Chung and
Hulbert [CH93], proposed an elegant generalization of the Newmark scheme as
a three-step method, with an implementation very close to a one-step method.
The resulting generalized-α method was presented in section 3.3.2. Jansen et
al. [JWH00] proposed an extension of this method for rst-order dierential
equations in uid dynamics.
We propose to apply a similar strategy here, which suggests the modied
residual equation

s
(1 − δm ) ẋn+1 + δm ẋn − (1 − δf ) fn+1 − δf fns = 0 (4.20)
o
(1 − δf ) (yn+1 − fn+1 ) + δf (yn − fno ) = 0 (4.21)

where δm and δf are numerical parameters of the method. The modication of

the output equation does not alter the local error on the output variables. A
perfect analogy can be observed between the treatment of the state variables
in this method and the treatment of the velocities in the generalized-α method.
68 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

Dening δf m = δf − δm , we immediately conclude that second-order accuracy is

obtained for
1
θ= + δf m (4.22)
2
According to section 3.3.4, consistent initial conditions should satisfy:

ẋ0 = f0s (4.23)

y0 = f0o (4.24)

leading to the propagation of an O(h2 ) initial error on the states.

The integration scheme follows a predictor-corrector procedure. The predic-
tion formula is obtained with the zero state rates assumption:

ẋ0n+1 = 0
x0n+1 = xn + (1 − θ) h ẋn (4.25)
0
yn+1 = yn

and the correction of the state rates follows:

1
4ẋn+1 = 4xn+1 (4.26)
θh

The linearized form of the discretized state equations (4.20), combined with
the linearized input/output relations are:

(1 − δm ) ∆ẋ − (1 − δf ) fxs ∆x − (1 − δf ) fus ∆u = −ressk (4.27)

(1 − δf ) (∆y − fxo ∆x − fuo ∆u) = −resok (4.28)
∆u = Lio ∆y (4.29)

which can be put in matrix form using (4.26), after elimination of ∆u and ∆ẋ:
" # " # !" # " #
1 I 0 −fxs −fus Lio ∆x −ressk
(1 − δm ) + (1 − δf ) =
θh 0 0 −fxo I − fuo Lio ∆y −resok
(4.30)
In chapter 3, the linear stability analysis of the generalized-α method was in-
vestigated for second-order undamped equations. We may reproduce this analysis
for rst-order equations.
4.1. SIMULATION OF CONTROL SYSTEMS 69

Time incrementation
-
t=t+h

Initial prediction
?

x0 , ẋ0 , y0 (eqn (4.25))

Evaluation of residues
?


ress (x, ẋ, y), reso (x, y)

Check for convergence

?
yes
kress k < µ, kreso k < ν
no ?
Evaluation of corrections
4x, 4y (eqn (4.30))

Incrementation
?

x, ẋ, y

Figure 4.5: Integration scheme for a control system.

70 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

4.1.4 Linear stability analysis

In order to assess the stability properties of the algorithm in the linear regime,
let us consider the linearized equations, obtained as in the previous paragraph:
" # " # " # " #
I 0 ẋ −fxs −fus Lio x
+ =0 (4.31)
0 0 ẏ −fxo I − fuo Lio y
Using Assumption 4.1, this DAE system is equivalent to an ODE system supple-
mented with an algebraic output equation. Indeed, since I−fuo Lio is not singular,
it is possible to extract explicitly the algebraic outputs:

−1
y+Cx=0 with C = − I − fuo Lio fxo (4.32)
and introducing this result in the rst equation, we obtain the ODE system:

−1
ẋ + A x = 0 with A = −fxs − fus Lio I − fuo Lio fxo (4.33)
Independent modal equations can be derived by a further linear transfor-
mation. Assuming that the state matrix A is diagonalized by the matrix of
eigenvectors Tx , we dene the modal coordinates x∗ and y∗ :
" # " # " #
x Tx 0 x∗
= (4.34)
y −C Tx I y∗
so that (4.31) is equivalent to:
" # " # " # " #
I 0 ẋ∗ Σ 0 x∗
+ =0 (4.35)
0 0 ẏ∗ 0 I y∗
where Σ is the diagonal matrix of the eigenvalues of A. It is sucient to study
the stability of the algorithm for each independent scalar equation.
We begin with the equation of a dynamic state :
ẋ∗ + σ x∗ = 0 (4.36)
which is discretized according to the modied residual equation (4.20)
(1 − δm ) ẋ∗n+1 + δm ẋ∗n + (1 − δf ) σ x∗n+1 + δf σ x∗n = 0 (4.37)
Combining this equation with the integration formula (4.19), we get the matrix
relation
" #" # " # " #
(1 − δf ) σ (1 − δm ) x∗ δf σ δm x∗
+ = 0 (4.38)
1 −θh ẋ∗ −1 −(1 − θ)h ẋ∗
n+1 n
4.1. SIMULATION OF CONTROL SYSTEMS 71

and we seek its eigensolutions, characterized by a synchronous behavior:

" # " #
x∗ x∗0
= ϕn with ϕn+1 = ζ ϕn (4.39)
ẋ∗ ẋ∗0
n

ϕ is the amplitude, [x∗0 ẋ∗0 ]T is the eigenvector, and ζ is the eigenvalue. The
characteristic equation is
" #
Pδf (ζ)σ Pδm (ζ)
x
Pσh = det =0 (4.40)
P1 (ζ) −Pθ (ζ)h

with the polynomials:

P (ζ) = (1 − ) ζ +   = {δm , δf }
Pθ (ζ) = θ ζ + (1 − θ) (4.41)
P1 (ζ) = ζ − 1

Developing the determinant, we obtain:

x
Pσh = Pδf Pθ σh + P1 Pδm = 0 (4.42)

The absolute stability is characterized by the spectral radius when σh → ∞:

x
P∞ = Pδf Pθ = 0 (4.43)

This equation has two simple roots:

−δf 1−θ
ζ1x = ζ2x = − (4.44)
1 − δf θ

where ζ1x is the spurious root (equivalent to the spurious root ζ1q associated with
the simulation of a mechanical system), and ζ2x is the principal root. With the
optimal parameters for second-order accuracy θ = 1/2 + δf m , we get
1 − 2 δf m
ζ2x = − (4.45)
1 + 2 δf m

Finally, in the system (4.35), the algebraic equation associated with an output

y∗ = 0 (4.46)

is discretized according to equation (4.21)

∗
(1 − δf ) yn+1 + δf yn∗ = 0 (4.47)
72 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

The single eigenvalue of the amplication matrix is the spurious root ζ1x .
As a conclusion, |ζ1x | < 1 requires δf < 1/2, and |ζ2x | < 1 requires δf m > 0,
so that the stability is guaranteed for:
1
δm < δf < (4.48)
2
Equivalent results were obtained by Jansen et al. [JWH00]. It is easily demon-
strated that the optimal condition |ζ1x | = |ζ2x | = ρx∞ leads to
3 ρx∞ − 1 ρx∞
 
1
δm = and δf = (4.49)
2 ρx∞ + 1 ρx∞ + 1

These formulae are dierent from the optimal formulae (3.30) obtained for the
simulation of a purely mechanical system.

4.2 Integrated simulation of mechatronic systems

The formalisms developed for the analysis of exible mechanisms (chapter 3)
and for the analysis of control systems (previous section) can be combined for the
analysis of mechatronic systems.

4.2.1 Coupled equations of motion

In a mechatronic system, the input/output variables of the control system
are directly connected with the mechanical dofs and their derivatives. The sensor
measurements wm are associated with the displacements, velocities and acceler-
ations:
wm = Lmq q + Lmq̇ q̇ + Lmq̈ q̈ (4.50)
where Lmq , Lmq̇ , Lmq̈ are sensor localization matrices. The expression of the
input variables follow from (4.8)

u = Liq q + Liq̇ q̇ + Liq̈ q̈ + Lio y (4.51)

with Liq = Lim Lmq , Liq̇ = Lim Lmq̇ , and Liq̈ = Lim Lmq̈ . The generalized forces
of the actuators ga produce the mechanical virtual work

δW a = (Lqa ga )T δq = (Lqa Lao y)T δq = (Lqo y)T δq (4.52)

4.2. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS 73

where Lqa and Lqo = Lqa Lao are the actuator and output localization matrices.
The whole set of coupled equations is then:

M q̈ + ΦTq (k λ + p Φ) − g(q, q̇, t) − Lqo y = 0 (4.53)

k Φ(q) = 0 (4.54)
ẋ − f s (u, x, t) = 0 (4.55)
y − f o (u, x, t) = 0 (4.56)
u − Liq q − Liq̇ q̇ − Liq̈ q̈ − Lio y = 0 (4.57)

Equation (4.53) represents the dynamics of the mechanical system, equation (4.54),
the kinematic constraints, equation (4.55), the state dynamics, equations (4.56)
the algebraic output equations, and (4.57) the input localizations.

4.2.2 Time-integration algorithm

A usual method to simulate such a coupled system of equations is to re-
formulate the second-order DAEs of the mechanism as rst-order DAEs, by the
introduction of the mechanical state vector xmech = qT q̇T . Any DAE inte-
 T

gration scheme can then be exploited, such as multistep or Runge-Kutta meth-

ods [BCP96], provided that the algebraic content of the coupled equations leads
to a "reasonable" index. This algebraic content comes from the kinematic con-
straints (4.54), and from the input/output equations (4.56),(4.57). Considering
an unconstrained mechanical system, under a few technical assumptions, the cou-
pled system is equivalent to a semi-explicit index-1 system, so that the simulation
can be done with the same convergence and stability properties as if the integrator
were applied to sti ODEs. In the presence of kinematic constraints formulated
at the position level, the system has an index-3 structure, and standard integra-
tion methods may lead to unreliable results [BCP96]. Therefore, the index of
the system should be reduced, either using a constraint elimination technique,
or a constraint dierentiation technique. In the literature, numerous methods
have been proposed for this dicult problem, which is still the object of intensive
research.
We have seen that the generalized-α method, known in structural dynamics,
can be considered for the simulation of constrained exible multibody systems.
Theoretical convergence and stability results are available, and the method ap-
74 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

pears to be quite ecient in practice. Moreover, the generalized-α method is also

applicable for state-space models, as shown in section 4.1.
Hence, we shall demonstrate that the generalized-α method leads to a unied
framework for the simulation of mechatronic systems. Many convergence and
stability properties of the original scheme developed for mechanical systems are
inherited thanks to this approach, as well as the systematic treatment of the
kinematic constraints. The numerical algorithm is presented and analyzed in the
following.
Combining the discretized mechanical equations (section 3.3) and the dis-
cretized state equations (section 4.1), we obtain the following coupled set of
equations:

∗

 (1 − αm ) (M q̈)n+1 + αm (M q̈)n + (1 − αf ) gn+1 + αf gn∗ =0


 (1 − αf ) k Φn+1 + αf k Φn = 0


 qn+1 = qn + h q̇n + h2 ( 21 − β) q̈n + h2 β q̈n+1

q̇n+1 = q̇n + h (1 − γ) q̈n + h γ q̈n+1



s s
 (1 − δm ) ẋn+1 + δm ẋn − (1 − δf ) fn+1 − δf fn =0


o
(1 − δf ) (yn+1 − fn+1 ) + δf (yn − fno ) = 0

n+1 = xn + h (1 − θ) ẋn + h θ ẋn+1

 x
(4.58)
with the discretized inputs:
un+1 = Liq qn+1 + Liq̇ q̇n+1 + Liq̈ q̈n+1 + Lio yn+1 (4.59)
where g∗ is a notation for ΦTq (k λ + p Φ) − g(q, q̇) − Lqo y. The rst set of
equations contains the discretized dynamic equations associated with the me-
chanical dofs and the Newmark formulae. The second contains the discretized
state equations and their time-integration formula. At this level, two remarks
may be formulated, which suggest a reformulation of those equations.
Remark 4.1 In order to uniformize the treatment of the state and displacement
variables, the dummy dynamic variables z are introduced:
Z t
z(t) = x(τ ) dτ (4.60)
0

The value of z is only meaningful at the velocity level (ż = x) and at the accelera-
tion level (z̈ = ẋ); nevertheless, an articial displacement-like Newmark formula
4.2. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS 75

is added for z. Moreover, we impose a constraint on the choice of the algorithmic

parameters
δm = αm , δf = αf , θ=γ (4.61)

Remark 4.2 According to Remark 3.1, q̈n+1 is a poor approximation for q̈(tn+1 ).
Therefore, we exclude the acceleration term from the input equation,
u = Liq q + Liq̇ q̇ + Lio y (4.62)

and we dene an observer equation for each acceleration measurement q̈ i :

y a = q̈ i (4.63)

This equation is discretized according to

a
(1 − αf ) yn+1 + αf yna = (1 − αm ) q̈n+1
i
+ αm q̈ni (4.64)

so that yn+1
a
is an order 2 approximation for q̈ i (tn+1 ), which can be safely con-
nected to any input port. The global output equation becomes
y = f o (u, x, t) + Loq̈ q̈ (4.65)

The discretized equations of motion are then restated,


∗


 (1 − αm ) (M q̈)n+1 + αm (M q̈)n + (1 − αf ) gn+1 + αf gn∗ =0

(1 − αf ) k Φn+1 + αf k Φn =0





s
(1 − αm ) z̈n+1 + αm z̈n − (1 − αf ) fn+1 − αf fns =0





−(1 − αm ) Loq̈ q̈n+1 − αm Loq̈ q̈n





o
) + αf (yn − fno )

 + (1 − αf ) (yn+1 − fn+1 =0




qn+1 = qn + h q̇n + h2 ( 21 − β) q̈n + h2 β q̈n+1





zn+1 = zn + h żn + h2 ( 12 − β) z̈n + h2 β z̈n+1









 q̇n+1 = q̇n + h (1 − γ) q̈n + h γ q̈n+1

 ż = ż + h (1 − γ) z̈ + h γ z̈
n+1 n n n+1
(4.66)
The advantage of this equivalent expression comes from the observation that q
and z receive exactly the same treatment. Hence, the formulation, the implemen-
tation and the theoretical analysis of the algorithm are highly simplied.
76 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

The prediction for qn+1 , zn+1 , q̇n+1 , żn+1 follows from the standard Newmark
formulae with the zero acceleration assumption q̈0n+1 = 0, z̈0n+1 = 0. Using the
linearized form of the input equation (4.62):
 
γ iq̇
∆u = iq
L + L 4q + Lio ∆y (4.67)
βh

the correction equation follows:

   
4q −resqk
 4λ   −resΦ
   
k
(4.68)

St 
 4z  =  −ress
  

   k 
4y −resok

where
γ 1
St = (1 − αf ) Kt + (1 − αf ) Ct + (1 − αm ) Mt (4.69)
βh βh2
Kt , Ct , Mt are respectively given by

     
Kt k ΦTq 0 −Lqo Ct 0 0 0 M 0 0 0
     
 k Φq 0 0 0 0 0 0 0  ,  0 0 0 0
   
 ,  
 −fus Liq 0 0 −fus Lio s iq̇
 −fu L s
0 −fx 0   0 0 I 0
     
 
−fuo Liq 0 0 I − fuo Lio o iq̇
−fu L o
0 −fx 0 −Loq̈
0 0 0
(4.70)
The correction equation involves non-symmetric matrices. Thanks to the
penalization term, Kt is positive denite and it is possible to apply a direct
solver without pivoting strategy. Good performances were observed with a non-
symmetric direct solver optimized for sparse matrices. This choice is reasonable
since a mechatronic system usually involves no more than a few hundreds dofs.
For systems with much more dofs, iterative solvers could be advantageously ap-
plied.
The time-integration algorithm is described in Figure 4.6. In order to analyze
its stability and convergence properties, it is equivalent to analyze the dynamic
system after elimination of the input and output variables from the state equa-
tions. Under Assumption 4.1, the input and output variables dened by (4.62)
and (4.65) can be formulated in explicit format (4.14), (4.15), and replaced in the
4.2. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS 77

Time incrementation
-
t=t+h

Initial prediction
?

q0 , q̇0 , q̈0 , x0 , ẋ0 (eqn (3.25), (4.25))

Evaluation of residues
?


resq , resΦ , ress , reso

Check for convergence

?
yes
kresa k < a , a = {q, Φ, s, o}
no ?
Evaluation of corrections
4q, 4λ, 4x, 4y (eqn (3.27), (4.30))

Incrementation
?

q, q̇, q̈, λ, x, ẋ, y

Figure 4.6: Time-integration algorithm for mechatronic systems

78 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

dynamic equations (4.53), (4.55):

M q̈ − g(q, q̇, t) + ΦTq (k λ + p Φ) − Lqo feo (x, q, q̇, q̈, t) = 0 (4.71)

k Φ(q) = 0 (4.72)
ẋ − fes (x, q, q̇, q̈, t) = 0 (4.73)

where fes (x, q, q̇, q̈, t) = f s (feu (x, q, q̇, q̈, t), x, t).

4.2.3 Linear stability analysis: the unconstrained case

For a purely mechanical system, the linear stability of the generalized-α
method has been analyzed under the assumption of a negligible damping matrix.
In case of a mechatronic system, the dynamics of the state variables is described
by rst-order dierential equations, and this assumption is no more relevant. For
this reason, an extended linear stability analysis is necessary, which is realized
here for the unconstrained case, and in the next section for the constrained case.
In order to investigate the stability of the algorithm when the mechanical
system is unconstrained, the linearized form of equations (4.71) and (4.73) is
considered:
M
f q̈ + C q̇ + K q + G x = 0
(4.74)
ẋ + A x + P q + D q̇ + F q̈ = 0

The matrices M f , C and K include the possible contributions of the direct feed-
back control actions feq̈o , feq̇o and feqo , respectively (e.g. M
f 6= M). Hence, Mf, C
and K are not necessarily symmetric positive denite, which is an important
dierence compared to the mass, damping, and stiness matrices of a passive
mechanical system.
The velocity term C q̇ may come from the internal damping of the material,
or from a direct feedback action. It is expected that this latter contribution has a
stabilizing eect, otherwise, the whole dynamic system may get unstable as well
as the numerical simulation. Therefore, the case C = 0 can be seen as the worst
case situation, where stability in the simulation results is desirable; the damping
matrix C is thus omitted in this analysis.  
The diculty may be reduced by diagonalization of matrices M f −1 K and
A:
Ω2 = T−1 q M
f −1 K Tq and Σ = T−1
x A Tx (4.75)
4.2. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS 79

where Ω and Σ are diagonal matrices, Tq and Tx are transformation matrices.

Dening the modal coordinates q∗ and x∗ :

q = Tq q∗ and x = Tx x∗ (4.76)

and the coupling matrices:

Γ = T−1 −1 −1 −1
q G Tx , Π = Tx P T q , ∆ = Tx D Tq , Υ = Tx F Tq (4.77)

we obtain the equivalent system:

q̈∗ + Ω2 q∗ + Γ x∗ = 0
(4.78)
ẋ∗ + Σ x∗ + Π q∗ + ∆ q̇∗ + Υ q̈∗ = 0

where the coupling matrices Γ, Π, ∆ and Υ have no specic structure, that could
be exploited for a component-wise analysis.
Those equations are discretized according to (4.58):
   
(1 − αf ) Ω2 0 (1 − αm ) I (1 − αf ) Γ 0 q∗
   ∗ 

 I 0 −βh2 I 0 0 

 q̇ 
 
 ∗ 
0 I −γh I 0 0  q̈ 
 
 
   
 (1 − α ) Π (1 − α ) ∆ (1 − α ) Υ (1 − α ) Σ (1 − α ) I   x∗ 
 f f m f m   
∗
0 0 0 I −γh I ẋ
n+1

   
αf Ω2 0 αm I αf Γ 0 q∗
   ∗ 
 −I
 −h I −( 21 − β)h2 I 0 0 

 q̇ 
 
 ∗ 
+ 0 −I −(1 − γ)h I 0 0  q̈  = 0
 

   
 α Π α ∆ αm Υ αf Σ αm I   x∗ 
 f f   
∗
0 0 0 I −(1 − γ)h I ẋ
n
(4.79)
where I and 0 still denote identity and null matrices of appropriate dimensions.
We seek the eigensolutions of this dierence equation, characterized by a
synchronous behavior:
   
q∗ q∗0
 ∗   
 q̇ 
 

 q̇∗0 

 ∗ 
 q̈  = ϕn 
 ∗ 
q̈0  with ϕn+1 = ζ ϕn (4.80)
   
 x∗   x∗0 
   
∗ ∗
ẋ ẋ0
n
80 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

so that the characteristic equation is

 
Pαf (ζ) Ω2 0 Pαm (ζ) I Pαf (ζ) Γ 0
 
 P1 (ζ)I
 −h I −Pβ (ζ)h2 I 0 0 

det 

0 P1 (ζ)I −Pγ (ζ)h I 0 0 =0

(4.81)
 
 P (ζ) Π P (ζ) ∆ P (ζ) Υ P (ζ) Σ P (ζ) I 
 αf αf αm αf αm 
0 0 0 P1 (ζ)I −Pγ (ζ)h I

with the polynomials Pαm , Pαf , Pβ , Pγ , P1 dened in equation (3.41).

If Γ = 0 or Π = ∆ = Υ = 0, a block triangular structure appears, and the
characteristic equation may be splitted into one equation for each block of the
diagonal, indicating an uncoupling in the time-integration of both subsystems. In
this case, the stability of the global integration scheme follows from the stability
of the scheme applied to the independent subsystems, which has been analyzed
before.
In any other situation, the analysis cannot be pursued without further as-
sumption. As proposed by Gear and Wells in a similar context [GW84], let us
consider the scalar case (one mechanical dof q and one state variable x). Then,
the characteristic equation becomes:
 
Pαf (ζ) ω 2 0 Pαm (ζ) Pαf (ζ) g 0
 
 P1 (ζ)
 −h −Pβ (ζ)h2 0 0 

det 
 0 P1 (ζ) −Pγ (ζ)h 0 0 =0

(4.82)
 
 P (ζ) p Pαf (ζ) d Pαm (ζ) f Pαf (ζ) σ Pαm (ζ) 
 αf 
0 0 0 P1 (ζ) −Pγ (ζ)h

In order to analyze the stability in the presence of sti dynamics in the mechanical
systems and/or in the control system, the characteristic equation is developed in
three subcases:

1. ωh → ∞:
 
−h −Pβ (ζ)h2 0 0
 
 P1 (ζ) −Pγ (ζ)h 0 0
Pαf (ζ) ω det  (4.83)
2

=0
 Pαf (ζ) d Pαm (ζ) f Pαf (ζ) σ Pαm (ζ) 
 

0 0 P1 (ζ) −Pγ (ζ)h

which is equivalent to P∞
q x
(ζ) Pσh (ζ) = 0.
4.2. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS 81

2. σh → ∞:
 
Pαf (ζ) ω 2 0 Pαm (ζ) 0
−Pβ (ζ)h2
 
P1 (ζ) −h 0
Pαf (ζ) σ det  (4.84)
 
=0

 0 P1 (ζ) −Pγ (ζ)h 0 

0 0 0 −Pγ (ζ)h

which is equivalent to Pωh

q x
(ζ) P∞ (ζ) = 0.

3. ωh → ∞ and σh → ∞:
 
−h −Pβ (ζ)h2 0 0
 
 P1 (ζ) −Pγ (ζ)h 0 0
Pαf (ζ) ω det 
2

 
 Pαf (ζ) d Pαm (ζ) f Pαf (ζ) σ Pαm (ζ) 


0 0 P1 (ζ) −Pγ (ζ)h

  (4.85)
P1 (ζ) −h 0 0
 
0 P1 (ζ) 0 0
+Pαm (ζ) det 
 
 =0
 Pαf (ζ) p Pαf (ζ) d Pαf (ζ) σ Pαm (ζ) 


0 0 P1 (ζ) −Pγ (ζ)h

which is equivalent to Pωh

q x
(ζ) Pσh (ζ) = 0.

As a conclusion, the linear stability of the integrator applied to the indepen-

dent subsystems is a sucient condition for the linear stability of the integrator
applied to the unconstrained mechatronic system. We assume that this result,
established for a scalar system (1 mechanical dof and 1 state), is a relevant indi-
cation for the stability of the algorithm in a more general context.

4.2.4 Linear stability analysis: the constrained case

The linearized equations of a constrained system are:
f q̈ + K q + BT λ + G x = 0
M
Bq = 0 (4.86)
ẋ + A x + P q + D q̇ + F q̈ = 0

As in the previous section, the damping term is neglected.

Inspired by section 3.3.3, it would be convenient to apply Theorem 3.1
page 52 in order to obtain a canonical form of the mechanical equations. This
82 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

would be possible if the stiness matrix K were real symmetric positive semi-
denite, and the mass matrix M f , real symmetric positive denite. Therefore, the
developments of this section assume that K and M f satisfy those conditions, as in
the passive case. This is always the case if no direct acceleration or displacement
feedback is present.
Using the notations of sections 3.3.3 and 4.2.3, we dene
Γr = TTqr G Tx , Πr = T−1 r −1 r −1
x P Tqr , ∆ = Tx D Tqr Υ = Tx F Tqr
Γc = TTqc G Tx , Πc = T−1 c −1 c −1
x P Tqc , ∆ = Tx D Tqc Υ = Tx F Tqc
(4.87)
We obtain the following set of equations:
q̈r + Ω2 qr + Γr x∗ = 0
q̈c + qλ + Γc x∗ = 0
(4.88)
qc = 0
ẋ∗ + Σ x∗ + Πr qr + Πc qc + ∆r q̇r + ∆c q̇c + Υr q̈r + Υc q̈c = 0

The characteristic equation is the determinant of the following matrix:

 
Pαf Ω2 0 0 0 0 0 Pαm I 0 0 Pαf Γr 0
 0 0 Pαf I 0 0 0 0 Pαm I 0 Pαf Γc 0 
0 Pαf I 0 0 0 0 0 0 0 0 0
 
 
 P1 I 0 0 −h I 0 0 −Pβ h2 I 0 0 0 0 
 
 0 P1 I 0 0 −h I 0 0 −Pβ h2 I 0 0 0 
−Pβ h2 I
 0 0 P1 I 0 0 −h I 0 0 0 0

 
0 0 0 P1 I 0 0 −Pγ h I 0 0 0 0
 
 
 0 0 0 0 P1 I 0 0 −Pγ h I 0 0 0 
 
 0 0 0 0 0 P1 I 0 0 −Pγ h I 0 0 
Pαf Πr Pαf Πc Pαf ∆r Pαf ∆c Pαm Υr Pαm Υc
 0 0 0 Pαf Σ Pαm I

0 0 0 0 0 0 0 0 0 P1 I −Pγ h I
(4.89)
and after lengthy but straightforward algebraic manipulations, one nds the
equivalent form:
 
Pαf (ζ) Ω2 0 Pαm (ζ) I Pαf (ζ) Γr 0
−h I −Pβ (ζ)h2 I
 
 P1 (ζ)I 0 0 
det
q
 
(P∞ (ζ))2m  0 P1 (ζ)I −Pγ (ζ)h I 0 0 =0

 Pαf (ζ) Πr Pαf (ζ) ∆r Pαm (ζ) Υr Pαf (ζ) Σ Pαm (ζ) I 
 

0 0 0 P1 (ζ)I −Pγ (ζ)h I

(4.90)
The determinant in (4.90) is strictly equivalent to the characteristic equation
obtained in the unconstrained case (4.81). In this sense, the presence of algebraic
4.2. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS 83

constraints has a similar impact on the stability of the mechatronic integration

scheme than in the purely mechanical case.

4.2.5 Convergence analysis

After elimination of the input and output variables, the equivalent dynamic
equations can be formulated with the vector of dummy variables z (ż = x):

M q̈ − g(q, q̇, t) + ΦTq (k λ + p Φ) − Lqo feo (ż, q, q̇, q̈, t) = 0 (4.91)

k Φ(q) = 0 (4.92)
z̈ − fes (ż, q, q̇, q̈, t) = 0 (4.93)

For a purely mechanical system, the convergence of the generalized-α method

has been assessed from the assumption of a constant mass matrix (section 3.3.4).
Considering a mechatronic system, due to possible acceleration measurements,
the functions fes and feo depend on q̈. If this dependence is strongly nonlinear, we
may not be able to guarantee a good convergence for the generalized-α method.
As we shall see, the generalized-α method behaves in a reliable way if the depen-
dence of fes and feo with respect to the acceleration is linear, which motivates the
following technical assumption:

Assumption 4.2 The functions fes (x, q, q̇, q̈, t) and feo (x, q, q̇, q̈, t) are linear in q̈:
∗
fes = fes (ż, q, q̇, t) + feq̈s q̈ (4.94)
∗
feo = feo (ż, q, q̇, t) + feq̈o q̈ (4.95)

where feq̈o and feq̈o are constant. Moreover, the matrix M

f = M − Lqo feo is not
q̈
singular.
The dynamic equations can then be restated:
" # " # " ∗ # " #
M
f 0 q̈ ΦTq (k λ + p Φ) − g − Lqo feo 0
+ ∗ = (4.96)
−feq̈s I z̈ −fes 0
Φ(q) = 0 (4.97)

These equations have the same structure than the equations of a purely me-
chanical system. Hence, the same convergence properties are observed, and we
immediately conclude that the local truncation error is still of order 2.
84 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

Remark 4.3 Assumption 4.2 is satised for any linear control system and any
control system without acceleration measurement, and those categories cover many
practical situations.

Remark 4.4 If the functions fes and feo are not linear but only ane with respect
to q̈ ( i.e., in equations (4.94) and (4.95), feq̈s and feq̈o are not constant), the argu-
ments described in section 3.3.4 about the consequences of a non-constant mass
matrix convince us that the local truncation error might increase by one order.
For instance, such a situation arises when acceleration signals wma are obtained
from a 3-axis accelerometer xed on a moving body. The measurements are the
absolute accelerations in body axes, and they are related to the accelerations in
inertial axes q̈ by the rotation matrix of the body R:

wma = RT q̈ (4.98)

In case of very fast motion, the variations of R may not be negligible over each
time-step, leading to a small loss in accuracy.

Remark 4.5 Control systems involving non-ane acceleration feedbacks are quite
unusual for control practitioners, so that we should not be afraid about the restric-
tions involved by Assumption 4.2.

4.2.6 Optimal choice for the algorithmic parameters

The previous results prove that the time-integration algorithm for mecha-
tronic systems inherits many properties from the purely "mechanical" and "state-
space" cases. The condition
1
γ = + αf m (4.99)
2
leads to second-order accuracy for the coupled system. For given values of αf
and αm , the spurious roots of both subsystems are equal:
−αf
ζ1q = ζ1x = ζ1 = (4.100)
1 − αf

whereas the principal roots are dierent (Figure 4.7):

1 − αf m 1 − 2αf m
q
ζ2,3 =− 6= ζ2x = − (4.101)
1 + αf m 1 + 2αf m
4.2. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS 85

1
|ζq2,3|
0.9 |ζx2|
0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.2 0.4 0.6 0.8 1
αfm

Figure 4.7: Evolution of the roots with respect to the parameter αf m .

This explains why the optimal choices of αf and αm are dierent for a mechanical
model and a state-space model (compare formulae (3.30) and (4.49)).
According to Chung and Hulbert [CH93], the optimal parameters for the
simulation of the mechanical equations are obtained when |ζ1 | = |ζ2,3q
|. From
Figure 4.7, we then conclude

|ζ2x | < |ζ1 | = ρq∞ = ρx∞ (4.102)

√
for αf m < 22 . Since the modulus of the principle root |ζ2x | is lower than the
modulus of the spurious root, this design is non-optimal for the state variables.
Conversely, according to Jansen et al. [JWH00], the optimal parameters for
the state equations are obtained when |ζ1 | = |ζ2x |. From Figure 4.7, we conclude
q
|ζ2,3 | = ρq∞ > |ζ1 | = ρx∞ (4.103)
√
for αf m < 22 . This design is not optimal for the mechanical variables. In
particular, the spectral radius associated with the mechanical variables is higher
than the spectral radius selected for the state variables.
The Hilber-Hughes-Taylor algorithm [HHT77] is obtained for αm = 0, and a
positive value for αf . The spurious root should satisfy the condition:

|ζ2x | > |ζ1 | and q

|ζ2,3 | > |ζ1 | (4.104)
86 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

which imposes the design constraint αf < 41 . In this case, the spectral radius
associated with the mechanical variables, is always higher than the spectral radius
associated with the state variables: ρq∞ > ρx∞ .
Hence, the optimal choice of the algorithmic parameters for the coupled set
of equations is not trivial; in the numerical applications presented in this chapter,
the Hilber-Hughes-Taylor method is systematically applied.

4.3 Systems with discontinuous dynamics

A hybrid system involves both continuous dynamic variables and discrete
events. After a discrete event, it is well-known that a restart procedure is neces-
sary to carry on the simulation. In this section, a restart procedure is developed
for the mechatronic simulation algorithm.
In this work, we have in mind two types of discontinuous behavior:

- the discrete time behavior of a micro-controller or a micro-processor, where

explicit jumps are observed at the output f o (t+n+1 ) 6= f o (t−n+1 ),

- the saturation eect of a software control law or a hardware component,

which is mathematically formulated as a change in the dynamic behavior
n+1 ) 6= f (u, x, tn+1 ) or f (u, x, tn+1 ) 6= f (u, x, tn+1 ).
− −
f s (u, x, t+ s o + o

Those discontinuities may produce discontinuities in the generalized forces.

According to Newton's second law of motion, the discontinuities are thereby
transmitted to the accelerations q̈, and to the reaction forces represented by the
Lagrange multipliers λ [BDG02, BDG03]. The state rates ẋ and the algebraic
variables y are also directly aected:

q̈− +
n+1 6= q̈n+1 , λ− +
n+1 6= λn+1 , ẋ− +
n+1 6= ẋn+1 ,
−
yn+1 +
6= yn+1 (4.105)

Temporarily, we exclude impulse phenomena, so that the energy is continuous,

as well as q, q̇ and x:

q− +
n+1 = qn+1 , q̇− +
n+1 = q̇n+1 , x− +
n+1 = xn+1 (4.106)

The integration scheme developed in the previous section was able to predict
eciently the values at time t−n+1 . Since the integration formulae rely on a Taylor
4.3. SYSTEMS WITH DISCONTINUOUS DYNAMICS 87

expansion of the dynamic variables, they are not applicable over the discontinuous
transition from t−n+1 to t+n+1 , and the correction for q̈, λ, ẋ and y should be
performed independently of q, q̇ and x, which are kept constant.
This correction can be seen as an integration restart procedure, and we know
from earlier convergence results (see section 3.3.4) that the initial conditions at
time t+n+1 should satisfy the residual equations:

M q̈+ T + + +
n+1 + Φq (k λn+1 + p Φ) − g(qn+1 , q̇n+1 , t) − L
qo +
yn+1 = 0 (4.107)
k Φ(q+
n+1 ) = 0 (4.108)
ẋ+ s + + +
n+1 − f (un+1 , xn+1 , tn+1 ) = 0 (4.109)
+
yn+1 − f o (u+ + +
n+1 , xn+1 , tn+1 ) = 0 (4.110)
u+ iq + iq̇ + iq̈ + io +
n+1 − L qn+1 − L q̇n+1 − L q̈n+1 − L yn+1 = 0 (4.111)

otherwise, the order of convergence would be deeply aected. q+n+1 , q̇+n+1 and
n+1 are imposed by (4.106), and these equations can be seen as nonlinear alge-
x+
braic equations for q̈+n+1 , λ+n+1 , ẋ+n+1 , yn+1
+
. They can be solved using a Newton-
Raphson procedure with the initial values given at time t−n+1 . However, since the
constraint equation does not involve any of these unknowns, the system is ill-
dened. In order to overcome this problem, the constraints should be formulated
at the acceleration level:

k Φ̈ = k Φ̇q q̇ + k Φq q̈ = 0 (4.112)

with Φ̇q = Φ̇q (q, q̇) and Φq = Φq (q). If this hidden constraint is satised at
time t−n+1 , it implies:
k Φq (q̈+ −
n+1 − q̈n+1 ) = 0 (4.113)
which may replace equation (4.108) in order to obtain a nonsingular system1 .
The linearized equations follow:
     
M k ΦTq 0 −Lqo 4q̈ −resqk
 4λ   −resΦ
     
 k Φq 0 0 0 k 
(4.114)
 
   =
 0
 0 I −fus Lio 

 4ẋ   −ress 
   k 
−Loq̈ 0 0 I − fuo Lio 4y −resok
1 The mechanical equations are linear in q̈, λ and y, thus, if the control system is also linear,
one Newton iteration yields convergence.
88 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

Time incrementation
- -
t=t+h
no
?
Update Initial prediction
yes
 Discontinuity ?
x, f s , f o t− q0 , q̇0 , q̈0 , x0 , ẋ0
? ?
Evaluation of residues Evaluation of residues
 
resq , resΦ , ress , reso resq , resΦ , ress , reso
? ?
Check for convergence Check for convergence
yes yes
t + kresa k < a , a = {q, Φ, s, o} kresa k < a , a = {q, Φ, s, o}
no ? no ?
Evaluation of corrections Evaluation of corrections
4q̈, 4λ, 4ẋ, 4y 4q, 4λ, 4x, 4y

? ?
Incrementation Incrementation
q̈, λ, ẋ, y q, q̇, q̈, λ, x, ẋ, y

Figure 4.8: Integration algorithm for hybrid systems.

A summary of the integration algorithm is given in Figure 4.8. It is assumed

that the discrete transitions occur at the nodes of the time grid. If the discontinu-
ities are caused by a sampled system, and if a xed time-step strategy is adopted,
this means that the sampling period T s should be a multiple of the time-step h:
Ts
h= (4.115)
p
where p is an integer. In a more general context, a variable step-size time-
integration strategy should be combined with an event detection algorithm in
order to catch exactly the instant of the transition.

4.3.1 Impulse phenomena

In the previous development, the case of impulsive mechanical loading has
been excluded. However, a slight adaptation of the algorithm is sucient to cover
this case, as described in the following. Since discontinuities now appear at the
velocity level:
q̇− +
n+1 6= q̇n+1 (4.116)
4.4. IMPLEMENTATION ISSUES 89

the values of q̇ should simply be updated when the impulse is detected, afterwards
the correction algorithm presented in Figure 4.8 is applicable. Therefore, the
critical point is to compute q̇+n+1 .
If ge denotes the generalized impact forces, their integral eect pe is dened
by
Z t+
pe = ge dt (4.117)
t−

The generalized velocities satisfy the classical impact equations, as described by

De Jalón and Bayo [DJB94]:
" # " # " #
−
M ΦTq q̇+
n+1 − q̇n+1 pe
= (4.118)
Φq 0 λp 0

where λp are Lagrange multipliers that are related to the internal impact forces
(impact reactions). The rst equation represents the conservation of momen-
tum, and the second equation states that the velocity increment has to fulll the
homogeneous velocity constraint equations.
The computation of q̇+n+1 requires the knowledge of pe , which is not always
a trivial problem. If the impulse forces result from an impact between bodies,
pe can be estimated using a physical model of their contact (e.g. an elastic or a
plastic model).

4.4 Implementation issues

The simulation concepts hereby developed for mechatronic systems have been
implemented in the Oofelie C++ Finite Element software [CKG94]. The reason
is that the object oriented technology is very suitable to build a modular, reusable
and ecient code. Let us briey discuss the implementation at the level of the
kernel and of the element library.
At the kernel level, the dof concept is naturally extended to represent input,
state and output variables, and the element concept to represent a generic block
of a block diagram model. All block diagram elements inherit from a few ab-
stract classes, such as System, ContinuousSystem and DiscontinuousSystem,
themselves inheriting from the standard Element class (see Figure 4.9). Those
abstract classes dene
90 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

Element
4

System Mech. FE library

4

DiscontinuousSystem ContinuousSystem
Kernel 4 4
Elements

SampledSystem SaturatedIntegrator LTISystem MatrixGain Saturation

4

LTISampledSystem UserSampledSystem

Figure 4.9: Element library: class diagram. LTI is an acronym for Linear Time
Invariant.

- the generic interface of the blocks,

- the concepts of input, output and state variables,

- the concepts of output and state equations, and their connection with the
Finite Element assembly procedure.
A simplied view of the library of control elements appears in Figure 4.9. The
developments realized in the kernel make the programming eort for an element
minimal, since only the functions f s and f o , and their sensitivity matrices have to
be dened (the implementation of the sensitivities can be avoided using a nite
dierence algorithm).
One may not be surprised to nd the SaturatedIntegrator among the
discontinuous elements, and the Saturation among continuous elements:
- the SaturatedIntegrator is characterized by one state variable x which
integrates the output until the saturation value xmax :

ẋ = u, while x < xmax (4.119)

4.5. APPLICATIONS 91

At the saturation the state equation becomes:

ẋ = 0, x = xmax , while u > 0 (4.120)

The discontinuity of ẋ requires a restart procedure,

- the Saturation element is characterized by one output variable, which is

equal to the input variable until the saturation value y max :

y = u, while u < y max (4.121)

After saturation, the output equation becomes:

y = y max , while u > y max (4.122)

and y is continuous.

4.5 Applications
Three examples with increasing complexity are considered: a rigid four-bar
mechanism, a Scara robot, and a vehicle with semi-active suspensions.

4.5.1 Rigid four-bar mechanism

Let us consider the rigid four-bar mechanism illustrated in Figure 4.10. It is
subject to external loads gext , the lower left hinge is actuated, and the value of
the torque τ is dened by a proportional integral derivative (PID) control law2
(see Figure 4.11). An analog and a digital implementation of the PID law will be
analyzed.
The analog PID controller is represented by the scalar state equations:

ẋ = θ − θref (4.123)
y = −P (θ − θref ) − D θ̇ − I x (4.124)

The digital PID controller generates a piecewise constant output

y = yk ∀ t ∈ [tk , tk+1 ] (4.125)

2 We consider here a regulation problem, so that the derivative action is simply established
from the measurement rate θ̇, and not from the error rate ė.
92 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

Figure 4.10: Four-bar mechanism.

θref - θ -
τ
- PID - Mechanism
θ̇
-

Figure 4.11: PID control of a four-bar mechanism.

with the update equation

ref
xk+1 = xk + T s (θk+1 − θk+1 ) (4.126)
ref
yk+1 = −P (θk+1 − θk+1 ) − D θ̇k+1 − I xk+1 (4.127)

with the sampling period T s = 0.01 s. The discrete state xk is not treated as a dof
in the simulation, it is simply updated at every sampling period. The maximal
value of the time-step is equal to the sampling period: h(max) = T s .
In this example, the mechanism is initially at rest (q0 = qinit , q̇0 = 0),
and the time variation of the external loads and reference inputs is plotted in
Figure 4.12. We may notice that the initial values q̈ = 0 and ẋ = 0 (for the
analog control) are consistent with those initial conditions.
The concise and high level textual model denition in Oofelie is illustrated
in Figure 4.13. For the digital case, the syntax is similar, the user should simply
further specify the sampling period.
4.5. APPLICATIONS 93

g ext 6 θref 6


θtar









θinit





0 h(max) 0 h(max)
 - -
t t

Figure 4.12: Four-bar mechanism with PID control: external loads and reference
input. h(max) = 0.01 s.

[Link](PID,ContinuousLTISystem_E,3,1,1);
ElemSet[PID].addInput(ThetaRefNode,GenDisp);
ElemSet[PID].addInput(HingeAngleNode,GenDisp);
ElemSet[PID].addInput(HingeAngleNode,GenVel);
ElemSet[PID].addOutput(TorqueNode);
ElemSet[PID].setStateSpaceMatrices(A,B,C,D);

Figure 4.13: Oofelie input data le - description of an analog PID controller
(3 inputs, 1 state and 1 output). PID is the element number; ThetaRefNode,
HingeAngleNode, TorqueNode are node numbers; GenDisp, GenVel mean "gen-
eralized displacement" and "generalized velocity" respectively; and A,B,C,D are
the state-space matrices associated with equations (4.123) and (4.124).
94 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

−3
x 10
0
1.55
−1

1.5 −2

−3

PID state
1.45
−4
θ (rad)

−5
1.4
−6

−7
1.35
Analog PID
Digital PID −8
Digital PID without correction
1.3 −9
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
t (s) t (s)

Figure 4.14: Simulation of a four-bar mechanism with PID control.

Simulation results
A small numerical damping is necessary to avoid high frequency oscillations
due to the mechanical algebraic constraints, and we have selected αf = 0.05,
αm = 0 (Hilber-Hughes-Taylor algorithm).
A rst simulation is realized with a time-step h(1) = h(max) = 0.01 s, and the
results are presented in Figure 4.14. A noticeable dierence is observed whether
the system is controlled by an analog or a digital controller. For the digital
control case, a simplied integration scheme has also been tested, where the
specic correction step after discontinuities is omitted. We observe that the
error in the simplied algorithm and the sampling eect have the same order of
magnitude, and we conclude that the correction step is essential for the accurate
representation of the sampling eect.
In order to analyze the inuence of a time-step reduction, ve dierent values
have been considered:
h(max)
h(p) = p = {1, ..., 5} (4.128)
2p−1

In Figure 4.15, the results obtained with h(1) and h(5) are very close to each other
for both the analog and the digital controller. However, for the digital controller,
the simplied algorithm without discontinuity correction leads to important er-
rors, especially for h(1) .
Assuming that the results obtained with h(5) are close to the exact solution,
4.5. APPLICATIONS 95

1.365
Analog PID / h1
1.36 Analog PID / h5
Digital PID / h1
1.355 Digital PID / h5
Digital PID without correction / h1
1.35 Digital PID without correction / h5
θ (rad)

1.345

1.34

1.335

1.33

1.325
0 5 10 15 20 25 30
t (ms)

Figure 4.15: Four-bar mechanism with PID control - inuence of the time-step
(h(1) = 10 ms, h(5) = 0.625 ms).

we dene the error for a dynamic variable as the deviation:

s
1 X (p)
σ(h(p) ) =
(5)
(qn − qn )2 p = {1, ..., 4} (4.129)
n n

where qn(p) is the value of the dynamic variable predicted with a time-step h(p) at
time tn . The evolution of σ with respect to the time-step is plotted in Figure 4.16.
The expected second-order convergence is observed in all cases, excepted for the
simplied algorithm which exhibits a poor rst-order convergence. Therefore, the
simplied algorithm is denitely not acceptable for the simulation of a mecha-
tronic system with a digital controller. For the corrected algorithm, we conclude
that the order of convergence predicted by the theory is veried.

4.5.2 Scara robot

In order to further validate our approach, let us consider a benchmark pro-
posed in the Eurosim framework in order to compare the capabilities of various
software [For98, Sch98, Eck99, Sas04]. The system under consideration is a Scara
robot (see Figure 4.17), and the model accounts for the dynamics of the joint
DC motors, and for their PD control law. The rst two joints are revolute with
vertical axes, and a third prismatic joint yields motion in the vertical direction.
96 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

−3 −2
10 10

−3
−4
10
10

−4
10
−5
σ

σ
10
−5
10

−6
10 −6
10

Analog PID − θ Digital PID

−7
Analog PID − x −7
Digital PID without correction
10 −3 −2
10 −3 −2
10 10 10 10
h (s) h (s)

Figure 4.16: Four-bar mechanism with PID control - convergence analysis. On

the left, the error for the hinge angle and the controller state (analog PID). On
the right, the error for the hinge angle (digital PID).

The electrical current in each motor is controlled independently according to a

collocated PD law. Figure 4.18 presents a block diagram model for the control
law and the actuator associated with one joint. The PD control law computes the
reference voltage U ref , and the applied voltage U a is limited to maximal values:

U a = U ref ∀ |U a | ≤ U max (4.130)

The electrical current satises the following equation

di
L = U a − U em − R i ∀ |i| ≤ imax (4.131)
dt

The electrical current is also limited to a maximal value, and U em is the electro-
motive voltage proportional to the joint velocity:

U em = k em θ̇ (4.132)

The electromechanical equation denes the relation between the torque and the
current intensity:
T = kt i (4.133)

k em and k t are two parameters of the electrical motor.

The behavior of the system is simulated when a step reference command is
4.5. APPLICATIONS 97

Figure 4.17: Scara robot.

θ̇ -
θref -
U a- Electrical Saturated i - Electromech. T -
di
θ - PD U ref
- Limitation dt -

θ̇ - - equation integrator conversion

Figure 4.18: Scara robot - Controller and motor associated with one joint.
98 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

2.5 2.5
θ1
θ2
2 2 θ3
ref ref
θ1 = θ2
ref
1.5 θ3 1.5
(rad and m)

θ (rad and m)
1 1
ref
θ

0.5 0.5

0 0

−0.5 −0.5
0 0.5 1 1.5 2 0 0.5 1 1.5 2
t (s) t (s)

Figure 4.19: Scara robot, joint references (step signal) and joint coordinates.

applied at t = 0 s (Figure 4.19):

Initial position Target position
θ1 = 0 rad θ1 = 2 rad
(4.134)
θ2 = 0 rad θ2 = 2 rad
θ3 = 0 m θ3 = 0.3 m
The results of the simulation obtained for a time-step h = 0.001 s are pre-
sented in Figures 4.19 and 4.20, and are fully equivalent to the results obtained
by other researchers [For98, Sch98, Eck99, Sas04]. Even though the mechanical
motion is rather smooth, the electrical current is subject to signicant transient
phenomena at the beginning and at the end of the motion. In particular, the
simulation predicts saturation eects.
A digital implementation of the control law has also been simulated with a
sampling period T s = 5 ms, leading to the results in Figure 4.21. In motor 1,
a high-frequency electromechanical pole involving the electromotive voltage is
excited at the end of the motion. According to a linearized analysis, if J eq denotes
the equivalent inertia associated with the motion of joint 1, the pulsation of this
pole is: r
k em k t
ω= (4.135)
J eq L
The resulting order of magnitude is in agreement with the simulation results.
Actually, due to the resistor, this pole is also aected by a damping eect. Again,
the simplied simulation method, without discontinuity correction does not lead
to accurate results.
4.5. APPLICATIONS 99

7 7
i i1
1
6 i2 i
6 2
i i
3 3
5
5
4
4
3
i (A)

i (A)
3
2
2
1
1
0

−1 0

−2 −1
0 0.5 1 1.5 2 0 0.02 0.04 0.06 0.08 0.1
t (s) t (s)

Figure 4.20: Scara robot - electrical currents. The picture on the right is a zoom.

7
i1 0.2
6 i2 with correction
0 without correction
i3
5
−0.2
4
−0.4
3
i (A)

−0.6
i (A)

2
−0.8
1 −1

0 −1.2

−1 −1.4

−2 −1.6
0 0.5 1 1.5 2 0.98 0.99 1 1.01 1.02 1.03
t (s) t (s)

Figure 4.21: Scara robot with sampled controller - electrical current in motor 1.
The picture on the right is a zoom centered on t = 1s.
100 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

From this example, we conclude that our simulation method is able to deal sys-
tematically and reliably with the coupled electromechanical equations of a mecha-
tronic system. The block diagram model is modular and convenient for the dy-
namic description of the controller and the actuators. Most methods presented
by other authors were based on an adaptive time-stepping method, which is a
strong advantage over our current xed time-step implementation. It is therefore
hopeless to attempt any comparison with respect to the computational eciency,
and this weakness of our software should be addressed in future investigations.
However, the decisive advantage of our approach certainly comes from its gen-
erality, in particular its ability to deal eciently and accurately with complex
parallel topologies and mechanical deformations; those potentialities are not re-
ally highlighted in this simple example. The objective of the next section is to
demonstrate the relevance of the integrated simulation method for a mechanism
with a far more complex topology, subject to the action of a strongly nonlinear
control system.

4.5.3 Vehicle semi-active suspension

The design of a modern car involves an increasing number of active compo-
nents, e.g. active suspension, anti-lock brake system (ABS), or electronic stability
program (ESP). Therefore, in order to predict the car performances, the dynamics
of those active components has to be considered in the design.
The dynamic simulation of a car equipped with a semi-active suspension
is considered here. This benchmark has been identied in the framework of
the Belgian Inter-University Attraction Pole on Advanced Mechatronic Systems
(AMS-IAP5/06)3 . The objective of the semi-active suspension demonstrator is
to develop new concepts related with the modeling, control and optimization
of mechatronic systems. Our personal contribution to this project concerns the
integrated simulation of the full car including the actuators and controller dy-
namics, using the methodology developed in this work. The various components
that have been considered in the multidisciplinary model of the car are described
in the following sections. We believe that the resulting simulation model is a
3 Website of the AMS-IAP5/06: [Link]
This project is sponsored by the Belgian state, Prime Minister's oce, Science Policy Program-
ming.
4.5. APPLICATIONS 101

r Controller
ab , v- Actuators ga Mechanical ab , lr , vr
iv - - -
model lr , v r - model model

Figure 4.22: Mechatronic model of the car equipped with a semi-active suspen-
sion. ab is the vector of car-body accelerations, lr is the vector of rattle extensions,
vr = l̇r is the vector of rattle velocities, iv is the vector of electrical currents, and
ga is the vector of damper forces.

relevant basis for the numerical optimization of the suspension, which will be
addressed in future research.

Principle of a semi-active suspension

A passive suspension has xed stiness and damping characteristics deter-
mined by their design. Depending on the road conditions, an adaptation of those
characteristics is desirable to optimize the comfort and the road handling, which
is possible if the passive shock absorber is replaced by an active or semi-active ac-
tuator. Active suspensions lead to highly controllable systems, but they require
an external power source (e.g. a hydraulic pump), which makes them costly,
complex and less reliable. Using a semi-active actuator in parallel with a passive
spring, a control action is obtained by varying the restriction in current controlled
valves, or by changing the viscosity of a magneto-rheological uid. Therefore, no
external power is introduced in the system, which makes the design less compli-
cated and safer.
In our case, semi-active hydraulic actuators with current controlled valves are
considered. The model of the mechatronic system is composed of a mechanical
model, a model of the actuators, and a model of the controller, as illustrated in
Figure 4.22.

Mechanical model
The car is an Audi A6 (see Figure 4.23), and its rigid-body model includes
the following components, illustrated in Figure 4.24:
102 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

Figure 4.23: Audi A6. On the right, instrumentation with corner accelerometers.

Figure 4.24: Mechanical model of the car.

- the car-body,

- the suspension mechanisms (including the passive springs),

- the slider-crank direction mechanism for the front wheels,

- the wheel model for the lateral force, the vertical force and the yaw torque.

The current model involves about 600 mechanical dofs, and it could be extended
to include the stiness of the suspension bushings, the exibility of the chassis,
and a longitudinal model for the wheels. However, the current model is sucient
to demonstrate the potentialities of the integrated simulation tool.
4.5. APPLICATIONS 103

Actuator model
A systematic description of the actuator model and of the controller is pre-
sented by Lauwerys et al. [LSS04].
Figure 4.25 compares a passive and a semi-active shock absorber. The passive
absorber contains two passive valves, restricting the oil ow from one chamber to
the other. Due to the motion of the rod in and out of the cylinder, the variation
of the total volume v tot available for the oil is:

v̇ tot = v̇ reb + v̇ comp (4.136)

where v reb and v comp are the volumes of the rebond and compression chambers.
The role of the accumulator is to compensate for the variation of v tot .
In the active shock absorber, the passive valves are replaced by check valves,
and a current controlled CVSA valve4 connects the extreme chambers. When
the rod moves up, the piston check-valve closes and the oil moves out of the
rebond chamber through the CVSA valve. When the rod moves down, the base
check-valve closes and the oil ows to the accumulator through the CVSA valve.
Therefore, both motions are aected by the controlled restriction at the CVSA
valve.
The ow variations reported in Figure 4.25 are related with a quasi-static
motion, where the pressures are in equilibrium at every instant, the uid is not
compressible, and the accumulator is ideal. Actually, the dynamic behavior is
more complex, and a nonlinear model has been calibrated by the manufacturer of
the shock-absorber. This model can be presented in nonlinear state-space format,
dening the inputs, states and output:

u(damp) = [lr l˙r i v ]T x(damp) = [preb pcomp ]T y(damp) = [g a ]T (4.137)

with lr , the rattle extension, iv , the electrical current in the CVSA valve, preb and
pcomp the pressures in the rebond and compression chambers, and g a the force
exerted by the damper. The state equations are:

ẋ(damp) = f s(damp) (u(damp) , x(damp) )

(4.138)
y(damp) = f o(damp) (u(damp) , x(damp) )
4 CVSA: continually variable semi-active.
104 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

Rod q piston q base Rod q piston q base q CV SA

↑ v̇ reb < 0 v̇ tot > 0 ↑ 0 v̇ comp > 0 v̇ reb < 0
↓ v̇ reb > 0 v̇ tot < 0 ↓ −v̇ comp > 0 0 −v̇ tot > 0

Figure 4.25: Passive and semi-active damper. Quasi-static valve ows during
compression and extension sequences. The proportions on the drawings are not
realistic and, for information, typical values for the chamber diameter and the
rod diameter are 32 mm and 22 mm, respectively.
4.5. APPLICATIONS 105

vr1 vr1
vr1
iv1 iv1
gvirt1
gvirt1 vr2
vr Inverse Actuator 1
vr2 vr3
vr2
iv2 iv2
gvirt2
gvirt2 vr4
Inverse Actuator 2
vr3 ab1
vr3
iv3 iv3
gvirt3
gvirt3 ab2
gvirt Inverse Actuator 3
vr4 ab3
vr4
iv4 iv4
gvirt4
gvirt4 ab4
Inverse Actuator 4
Reconnect
Car + Dampers

gvirt Modal Virt Forces Modal Virt Forces Modal Acc Modal Acc Coupled Acc

Coupling Integral Control Decoupling

Figure 4.26: Controller of the semi-active suspension. "vr" stands for the rattle
velocities, "ab" for the car-body accelerations, and "iv" for the valve currents.

The functions f s(damp) and f o(damp) are given by the manufacturer as C-functions,
which are linked with Oofelie and called in a user-element inheriting from the
class ContinuousSystem (see Figure 4.9). The sensitivities of those functions are
obtained using a nite dierence procedure.

Controller model
The control law of the active suspension has been developed by Lauwerys,
Swevers, and Sas [LSS04]. As illustrated in Figure 4.26, it consists of three stages:
a feedback linearization (inverse actuator models), a transformation into modal
space (coupling and decoupling operations), and a linear integral control.
The car and the dampers are represented as a black box, whose inputs are
the four CVSA electrical currents iv , and whose outputs are the rattle velocities
v r = l˙r , and the accelerations measured at the four corners of the car ab (see
106 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

Figure 4.23).
The feedback linearization technique seeks for virtual inputs which have the
property to inuence the outputs in a linear way. If one accepts that the non-
linearity of the mechanism is weak, the main source of nonlinearity lies in the
actuator. According to Lauwerys et al., an ecient feedback linearization law is
obtained by inversion of a simplied quasi-static model of the actuators. From
equation (4.138), it is possible to formulate the quasi-static damper force with
respect to the CVSA electrical current, and the rattle velocity:
g a = g a (iv , v r ) (4.139)
This relation can be inverted for v r 6= 0:
iv = (g a )−1 (g virt , v r ) (4.140)
where g virt is the new virtual input, which can be interpreted as a virtual damper
force. For v r = 0, the controllability is defective, and the singularity of (g a )−1 is
avoided thanks to a regularization strategy. If this inverse model is able to cancel
the nonlinearity of the actuator, the virtual input g virt is actually proportional to
the damper force. Therefore, besides the advantage of a good linearity between
g virt and the output, a force control strategy can be established on this basis.
Since the motion of the car simultaneously involves the forces applied on the
four wheels, the denition of the virtual control forces g virt for the four shock-
absorbers is a linear but coupled multi-input/multi-output problem. This prob-
lem can be simplied by a transformation into a modal space dened by the
heave (pumping), roll and pitch of the car-body. In this modal space, the sys-
tem is represented by three uncoupled single-input/single-output subsystems, for
which three independent integral controllers are designed.
The several stages of the controller are easily described using the block dia-
gram language. The inverse actuator model is implemented as a specic element,
which directly invokes the C-function implemented in the actual controller, and
the sensitivities are obtained by analytical dierentiation. All other blocks are
modeled using the element library mentioned in Figure 4.9.

Simulation results
The simulation of a lane change maneuver has been realized, and the target
trajectory, represented in Figure 4.27, corresponds to a standard qualication
4.5. APPLICATIONS 107

Figure 4.27: Lane change maneuver. The dimensions presented here correspond
to a standard qualication test.

0.05 0.5

0.04 0
slider−crank translation (m)

0.03 −0.5

0.02 −1

0.01 y (m) −1.5

0 −2

−0.01 −2.5

−0.02 −3

−0.03 −3.5

−0.04 −4

−0.05 −4.5
0 1 2 3 4 5 6 0 10 20 30 40 50 60
time (s) x (m)

Figure 4.28: Car - steering command (slider-crank displacement) and horizontal

trajectory of the car-body.

test. The car has a 10 m/s initial velocity, and a driver applies an open-loop
steering command, without any real-time correction (blind driver assumption).
The motor and the brakes do not produce any torque on the wheels. The time-
step for the simulation is 0.01 s, and the algorithmic parameters are αf = 0.05
and αm = 0.
Figure 4.28 illustrates the motion of the slider-crank mechanism actuated by
the driver, and the horizontal trajectory of the car.
Figure 4.29 represents the global motion of the car-body. The yaw, pitch
and roll angles of the car-body are plotted as well as the radius of gyration and
the vertical displacements. The radius of gyration becomes innite when the car
follows a straight line.
The dynamic behavior of the semi-active shock absorbers is illustrated in Fig-
ure 4.30. It is interesting to observe the pressures in the rebond and compression
chambers, depending on the sign of the rattle velocity: during the compression,
108 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

0.5 0.06
yaw roll
0.4 pitch
0.04
0.3

0.2 0.02
angle (rad)

0.1 angle (rad) 0

0

−0.1 −0.02

−0.2 −0.04
−0.3
−0.06
−0.4

−0.5 −0.08
0 1 2 3 4 5 6 0 1 2 3 4 5 6
time (s) time (s)

50
0.02
45

40 0.015
vertical displacement (m)
Gyration radius (m)

35
0.01
30

25
0.005
20

15 0

10
−0.005
5

0 −0.01
0 1 2 3 4 5 6 0 1 2 3 4 5 6
time (s) time (s)

Figure 4.29: Car-body - angle and vertical displacements.

4.5. APPLICATIONS 109

5
x 10
5
0.04
Rear right
Front right 4
0.03
Extension displacements (m)

0.02 3

Pressure (Pa)
0.01 2

0
1
−0.01
0
−0.02

−1
−0.03 Rebond chamber
Compression chamber
−0.04 −2
0 1 2 3 4 5 6 0 1 2 3 4 5 6
time (s) time (s)

1.8
400
1.6
300
1.4
Electrical current (A)

200
1.2
100
Force (N)

0.8 0

0.6 −100

0.4 −200
0.2
Rear right −300
Rear right
0 Front right Front right
−400
0 1 2 3 4 5 6 0 1 2 3 4 5 6
time (s) time (s)

Figure 4.30: Car semi-active dampers - extension, hydraulic pressure (rear right),
electrical current and damper force.

the valve between those chambers is open, and the pressure in the compression
chamber is slightly higher, but during the extension, the valve is blocked, and
the pressure in the rebond chamber can be much higher than in the compres-
sion chamber (see Figure 4.25). The saturation eect dominates the behavior of
the electrical current in the CVSA valves. The static contribution of the forces
produced by the actuators is not zero, due to non-equilibrated pressures in the
dierent chambers when the piston rod is at rest.
In this application, the integrated simulation method was able to predict the
behavior of a complex mechatronic system. The model consists of a full model
of the car and its suspension mechanisms, a nonlinear dynamic model of the hy-
draulic actuator, and the nonlinear control algorithm. The modularity of the
110 CHAPTER 4. INTEGRATED SIMULATION OF MECHATRONIC SYSTEMS

formulation leads to a systematic and convenient description of the mechatronic

model. In particular, the block diagram language allows a direct denition of the
controller model using the algorithm implemented in real-time, and vice versa.
The simulation is carried out within a reasonable computational time (several
minutes on a desk computer). From our experience with the simulation of purely
mechanical systems using the Finite Element method, we believe that this perfor-
mance could benet from implementation optimization and from the development
of an adaptive time-stepping strategy.

4.6 Summary and concluding remarks

In this chapter, well-dened concepts from multibody dynamics, structural
dynamics and system theory have been integrated in a unied formulation for
the simulation of mechatronic systems. In particular, our personal theoretical
contributions concern:

- the adaptation of the block diagram concept for the modular representation
of rst-order state equations within a Finite Element formulation,

- the extension of the generalized-α method for the strongly coupled simula-
tion of a mechatronic system represented by a mechanical Finite Element
model and a block diagram model, with stability and convergence analyses,

- the development of a specic methodology to deal with possible discontin-

uous eects in some subsystems (e.g. sampled systems).

Three examples with increasing complexity have been successfully treated

with this methodology: a four-bar mechanism, a Scara-robot, and a car semi-
active suspension. All these examples involve rigid-body mechanisms. A fourth
mechatronic system involving a exible mechanism will be analyzed in chapter 6.
This systematic approach yields reliable results and it relieves the user from
intricate, time-consuming and dubious implementation of user-elements for the
dynamic description of a control system.
 V

Nonlinear Model Reduction in Flexible

Multibody Dynamics

In the state-of-the-art, section 2.4, new modeling concepts appeared to be

necessary when designing a controller for a exible multibody system with par-
allel topology. Indeed, a Finite Element model or an assumed-mode model are
represented by a relatively high-order set of nonlinear DAEs, which is not con-
venient for this design problem. In this chapter, an original model reduction
method is proposed leading to a suitable compromise between accuracy, order,
structure, computational eciency, systematic formulation and portability. The
reduced model is obtained from an initial nonlinear Finite Element model, but
since the theory assumes linear elasticity, a large part of the developments would
still be valid if another initial formulation had been selected. The Finite Element
method is preferred because of its systematic implementation, especially valuable
when dealing with complex mechanisms.
Before starting the detailed presentation of the reduction method, an overview
is rst presented in order to draw the reader's attention to the key steps and fea-
tures of the approach.

5.1 Introduction
The reduction method is an extension of the component-mode technique
established in structural dynamics, which accounts for the nonlinear kinematics
of the mechanism. Basically, the reduction procedure proceeds in two steps:
112 CHAPTER 5. NONLINEAR MODEL REDUCTION

- an order reduction of the initial Finite Element model, which can be realized
locally for any given conguration,

- an approximation of the reduced-order model in the conguration space.

Let us further develop both steps.

The order reduction comes from a reduced kinematic description with modal
coordinates associated with the motion of the whole mechanism, including all
bodies and joints. The mode shapes are therefore conguration-dependent : for
every conguration, their denition according to a component-mode synthesis is
based on the linearized dynamic equations. Since those modes are not associated
with a particular component of the mechanism, and since they are dened locally
for small motions around a conguration, we refer to them as local modes rather
than component-modes. The local mode shapes characterize a local coordinate
transformation between the initial Finite Element coordinates q and the modal
coordinates η . The reduced-order model is obtained after introduction of this
coordinate transformation into the initial model.
This local reduction procedure, which should be repeated for every cong-
uration change, is computationally expensive. A more suitable model can be
constructed by approximation of the reduced-order model in the conguration
space [BDG04c].
An implementation of this approach is illustrated in Figure 5.1: a number
of reference kinematic congurations θ(i) , (i = 1, ..., r) are selected in the cong-
uration space, the Finite Element model is built and reduced for each of them,
afterwards an approximation algorithm leads to a closed-form model.
The validity and the consistency of this approach is questionable for two
reasons:

- the modal coordinates dened in one conguration may not be naturally

associated with the modal coordinates in another conguration, so that
non-physical discontinuous transitions may appear in the reduced model,

- the modal parameterization is intrinsically nonlinear, and the standard

model reduction procedure dened in linear structural dynamics has to be
reconsidered.
5.1. INTRODUCTION 113

Select a set of cfg {θ(1) ...θ(r) }

?
Next cfg 

?
Kinematic analysis

?
Local mode synthesis

?
Model reduction

no
?
All cfg treated?
yes
?
Approximation

Figure 5.1: Simple approach for the model reduction of a exible mechanism, and
illustration with a four-bar mechanism.
114 CHAPTER 5. NONLINEAR MODEL REDUCTION

In order to guarantee the consistency of the model, we assess the reduction

method on a well-dened nonlinear reduced parameterization, valid in the whole
conguration space, and consistent with the local modes. This parameterization
relies on separated descriptions of the large amplitude rigid motion, and of the
small amplitude exible motion [BDG04a]. Since it is related with mode shapes
of the global mechanism, dened in the global conguration space, it is denoted
the Global Modal Parameterization. The formal denition of this concept is an
important contribution of this dissertation.
A second contribution is associated with the numerical reduction procedure,
which accounts for the nonlinear nature of the Global Modal Parameterization.
Actually, it will be demonstrated that the curvature of the coordinate transfor-
mation leads to a contribution to the reduced inertia forces.
The development of a piecewise approximation strategy in the conguration
space is the third original contribution of this chapter. The method follows an
adaptive conguration space inspection algorithm, which minimizes the compu-
tational resources to satisfy a user-dened tolerance on the approximation error.
The reduced model satises the following specications:

- accuracy in the bandwidth of the actuators and within the limits of the
workspace,

- low-order and free from kinematic constraints, which follows from the Global
Modal Parameterization,

- physical interpretation of the modal coordinates in terms of rigid and ex-

ible modes, and interpretation of the model parameters in terms of mass
matrix, stiness matrix, and gyroscopic tensor,

- available in closed-form, involving inexpensive computations,

- formulated systematically from high-level information,

- dened in a simple standard format, easily exported to other software for

control design.

This model can be exploited for dierent purposes, such as:

- the online implementation of a feedback controller,

5.2. LINEAR COMPONENT-MODE SYNTHESIS 115

- the oine optimization of a control policy,

- the construction of a more structured model, in order to t well-established

control theories.

After this preliminary description, section 5.2 presents component-mode tech-

niques developed in linear structural dynamics. Section 5.3 is devoted to the
characterization of the nonlinear kinematics of parallel mechanisms. The Global
Modal Parameterization and the local reduction procedure are respectively de-
scribed in sections 5.4 and 5.5. The approximation strategy is presented in sec-
tion 5.6. After a summary of the overall algorithm in section 5.7, two applications
are treated in section 5.8: a exible four-bar mechanism, and a rigid parallel-
kinematic machine-tool (a third example is left for chapter 6). The chapter ends
with a few concluding remarks.

5.2 Linear component-mode synthesis

A linear mechanical system is represented by the equations of motion:

M q̈ + K q = g (5.1)

with q, the n × 1 vector of generalized coordinates, M, the mass matrix, K,

the stiness matrix, and g, the vector of applied force. A reduced-order model
oers obvious advantages associated with the simplied treatments required for
its analysis.
In structural dynamics [GR97], the development of model reduction tech-
niques has also been motivated by substructuring applications, where the analy-
sis of a complex dynamic structure exploits a decomposition into several simpler
parts. A reduced-order model is then constructed to represent concisely the dy-
namics of each substructure, and it will later be used to reconstruct the global
model. This approach allows a detailed modeling of components with complex
geometry, while keeping a relatively simple global dynamic model. It also leads
to an increased modularity and reusability in the modeling process. In our con-
text, the mechanical system can be seen as a "substructure" of the mechatronic
system, for which a reduced-order model is desirable.
116 CHAPTER 5. NONLINEAR MODEL REDUCTION

5.2.1 Principle of a linear reduction method

A reduction method relies on the construction of a change of variables from
initial coordinates to modal coordinates according to

q=Ψη (5.2)

where η is the n × 1 vector of modal coordinates, and Ψ is the n × n matrix of

component-modes. The reduction comes from

n<n (5.3)

The variations of q are thus restricted to the subspace spanned by the component-
modes, and the accuracy of the reduced model depends on the ability of the
component-modes to describe the actual motion of the system.
In order to formulate the reduced equations of motion, the coordinate trans-
formation is introduced into the expressions of the kinetic and potential energies,
and of the virtual work of the external forces. Initially, we have:
1
K = q̇T M q̇ (5.4)
2
1
V = qT K q (5.5)
2
δW = gT δq (5.6)

and the transformation into modal coordinates leads to

1
K = η̇ T M η̇ (5.7)
2
1
V = ηT K η (5.8)
2
δW = gT δη (5.9)

with the reduced mass matrix, stiness matrix, and equivalent force vector:

M = ΨT M Ψ, K = ΨT K Ψ, g = ΨT g (5.10)

The equations of motion of the reduced system follow:

M η̈ + K η = g (5.11)

A similar procedure can be applied for a system with m linear constraints:

Φq q = 0 (5.12)
5.2. LINEAR COMPONENT-MODE SYNTHESIS 117

Following an augmented Lagrangian approach, an m × 1 vector of Lagrange mul-

tipliers λ is introduced, and the vector of augmented generalized coordinates is
dened: " #
q
u= (5.13)
λ
as well as the matrices and vector:
" # " # " #
M 0 K + p ΦTq Φq k ΦTq g
Muu = , Kuu = , gu =
0 0 k Φq 0 0
(5.14)
where k and p are the scaling and penalty factors. The constrained equations of
motion are restated:
Muu ü + Kuu u = gu (5.15)
and the component-mode technique can be applied as in the unconstrained case:

u = Ψuη η (5.16)

5.2.2 Component-mode selection

The objective of the component-mode synthesis is to dene the mode shapes
Ψ, leading to a minimized loss in accuracy. Let us consider a partitioning of the
generalized coordinates:  
qr
(5.17)
 
u= g 
 q 
ui
where qr are s rigid dofs strictly able to characterize the rigid modes, qg are ng
constraint dofs where external loads are applied1 , and ui are the ni remaining
internal dofs which are not loaded and are condensed by the reduction procedure.
The set of internal dofs includes the Lagrange multipliers.
On this basis, Hurty [Hur65] proposed three subsets of modes: the rigid-
body modes, the constraint modes, and the internal modes. This choice leads
to an exact representation of the static response of the structure, which is crit-
ical for the consistent assembly of the reduced model with other substructures.
1 In a substructuring framework, qg also includes interface dofs between substructures, sub-
ject to internal forces.
118 CHAPTER 5. NONLINEAR MODEL REDUCTION

In a mechatronic framework, the static response of a structure should also be

accurately represented since it inuences the zeros of the mechanical transfer
functions, as discussed by Preumont [Pre97]. This technique was improved by
Craig and Bampton [CB68] who suggested to treat all boundary degrees of free-
dom (qr and qg ) alike. It can be demonstrated that the modes proposed by Hurty
and by Craig-Bampton span the same subspace, so that the accuracy of the re-
duced model is equivalent for both methods. The Craig-Bampton method oers
several advantages, such as a simplied partitioning, a greater eciency, and the
explicit conservation of all boundary dofs in the reduced model. This is probably
the reason why it is so popular and available in most commercial Finite Element
software. However, the rigid modes do not appear explicitly among the boundary
modes, which is a precluding drawback for us: our nonlinear reduction method
entirely relies on the separation between the rigid and the exible motion.
For this reason, the component-modes dened by Hurty [Hur65] are exploited
here, but this choice is not restrictive. Many other component-modes have been
dened in the literature, see Craig [Cra87] and De Fonseca [DF00] for detailed
reviews, and our nonlinear reduction method could involve any of them, provided
that the rigid-body modes are isolated from the exible modes.
The s rigid-body modes Ψuθ are dened with respect to the rigid dofs qr ,
and they satisfy the equilibrium of internal forces:
     
rr rg ri rθ rθ
K K K I I
(5.18)
    uθ
 
 Kgr Kgg Kgi   Ψgθ  = 0, Ψ =  Ψgθ 

    
ir ig ii iθ iθ
K K K Ψ Ψ

Irθ is the s × s identity matrix.

The ng constraint modes Ψuγ are the static deformations obtained when the
rigid dofs qr are xed and unit displacements are imposed on the constrained
dofs qg :
 
" # " # " # 0rγ
Kgg Kgi Igγ gg
(5.19)
 
= , Ψuγ =  Igγ 
Kig Kii Ψiγ 0  
Ψiγ

Igγ is the ng × ng identity matrix.

The nι internal modes Ψuι are a few eigenmodes when rigid and constraint
5.2. LINEAR COMPONENT-MODE SYNTHESIS 119

dofs are xed (nι < ni ).

 
rι
0
(5.20)
ii 2 ii

iι uι
 
K −ω M Ψ = 0, Ψ =  0gι 


iι
Ψ

The full transformation matrix follows:

       
r rθ
q i  θ   I 0 0 θ
 h
 (5.21)
   
 qg  = Ψuθ Ψuγ Ψuι  η γ  =  Ψgθ Igγ 0   ηγ 
 
    
ui ηι Ψiθ Ψiγ Ψiι ηι

where θ, η γ and η ι denote the modal amplitudes. By construction, the amplitudes

of the rigid modes θ are identied with the rigid dofs qr :

qr = θ (5.22)

This property does not hold for the constraint modes: qg 6= η γ . If the modal
masses of the internal modes are normalized, the reduced matrices and forces
have the following structure:
   θθ θγ θι
    
0 0 0 M M M gθ gr + ΨgθT gg
 γγ   γθ γγ γι    
 ,  gγ  = 

K=
 0 K 0  , M =  M
 M M     gg 

2 ιθ ιγ ιι ι
0 0 Ω M M I g 0
(5.23)
where Ω = diag(ωi ) is the diagonal matrix of internal eigenvalues.
Géradin and Rixen [GR97] interpreted a class of component-mode methods as
a truncation in the modal expansion of the mechanical impedance. In this sense,
the component-modes suggested by Hurty and Craig-Bampton are optimal.
In multibody dynamics, the linear component-mode technique is usually ex-
ploited for the compact kinematic description of an isolated exible body with
respect to a oating frame of reference, see section 2.1.2. In order to obtain a
more drastic reduction, we apply the modal parameterization to the whole mech-
anism, which is seen as a "component" of the mechatronic system. Therefore,
the concept of component-mode is replaced by the concept of local mode dened
around a conguration. The description of the conguration of a mechanism with
a suitable parameterization is addressed in the next section.
120 CHAPTER 5. NONLINEAR MODEL REDUCTION

5.3 Parameterization of the rigid kinematics

The parameterization of the motion of a mechanism has been considered in
section 3.1. Several sets of parameters were described, such as the minimal coor-
dinates, the relative coordinates, the Cartesian coordinates, the absolute coordi-
nates, and the mixed coordinates. The strong advantage of minimal coordinates
comes from their independence, but we mentioned that a minimal parameteriza-
tion is only possible in a restricted part of the conguration space. Thus, depen-
dent coordinates were preferred for the formulation of general analysis methods.
However, for control design, we will demonstrate that the minimal coordi-
nates can be exploited with great benets. Our objective is to formulate the
dynamics of exible mechanisms in terms of minimal coordinates, i.e. free from
kinematic constraints.
Under the assumption of small deformations, the exible motion can be de-
scribed in the neighborhood of the motion of the undeformed mechanism. A rst
problem is then to parameterize the rigid kinematics associated with the exible
mechanism, using independent coordinates. This section specically addresses
this problem.
As for linear reduction methods, the approach is based on a coordinate trans-
formation. Here, we rely on an initial parameterization in terms of dependent
coordinates q (e.g. Finite Element coordinates), and we seek for a transformation
into independent coordinates θ. Therefore, two problems are encountered:

- the denition of the independent parameters θ,

- the characterization of the nonlinear coordinate transformation q = ρ(θ)

and its Jacobian ρθ .

The formulation of the second problem is substantially dierent whether the

initial parameters q account or not for the deformations. For this reason, we
successively consider the case of rigid mechanisms and the case of exible mech-
anisms.
5.3. PARAMETERIZATION OF THE RIGID KINEMATICS 121

5.3.1 Rigid mechanisms

The motion of a rigid mechanism can be described with a vector of dependent
coordinates q ∈ Rn , satisfying m holonomic constraint equations:

Φ(q) = 0 (5.24)

The conguration space Ωrtq is dened as the set of kinematically admissible con-
gurations which satisfy the kinematic constraints:

Ωrt n
q = {q ∈ R | Φ(q) = 0} (5.25)

The superscript r stands for "rigid", and t for the "total" conguration space.
Assuming independent constraints, the number of kinematic modes s satises:

s=n−m (5.26)

We seek for a s × 1 vector of minimal parameters θ, which are related to q

by a coordinate transformation, represented by an invertible mapping ρ = ϕ−1 :

ρ : Ωrt rt
θ → Ωq , θ 7→ q = ρ(θ) (5.27)
ϕ : Ωrt rt
q → Ωθ , q 7→ θ = ϕ(q) (5.28)

where Ωrt θ is the set of possible variations of the parameters θ . We also refer
to Ωrt
θ as the conguration space, whenever no confusion is possible. Figure 5.2
illustrates those denitions.
In general, the existence of a global parameterization with independent coor-
dinates is not possible in the total conguration space Ωrtq , and some restrictions
are necessary to dene a pragmatic solution.
We propose to dene the minimal coordinates θ as the actuated dofs, i.e. the
dofs associated with the generalized forces exerted by the actuators. For instance,
for a motorized hinge, the actuated dof is the angle between the connected links,
whereas for a linear actuator, it is the relative distance between the connected
bodies. As a consequence of this choice, the actuator dofs will appear explicitly
in the reduced model, which is extremely valuable for the design of the control
system. An implicit assumption is that the number of actuators is equal to
s. This does not imply that the reduction method is only applicable to fully
122 CHAPTER 5. NONLINEAR MODEL REDUCTION

Figure 5.2: Coordinate transformation between dependent and independent pa-

rameters.

actuated mechanisms, it only means that in any other situation, the denition of
the independent parameters is left to the user.
The actuated dofs θ are usually relative coordinates, in contrast with the
absolute coordinates q used in our Finite Element formulation. A systematic
implementation of the relation θ = ϕ(q) can be dened using mixed coordinates.
This means that the actuator coordinates θ appear explicitly among the set of n
mixed coordinates q: " #
θ
q= (5.29)
q∗
q∗ are the n − s non-actuated dofs. Since m = n − s, the number of non-actuated
dofs equals the number of kinematic constraints of the mixed formulation.
The mapping ϕ is directly characterized:

θ = ϕ(q) = [I 0] q (5.30)

where I is the s × s identity matrix and 0 is the s × (n − s) null matrix.

The formulation of the inverse mapping ρ requires a deeper investigation.
Mathematically, the n equations:
(
ϕ(q) = θ
(5.31)
Φ(q) = 0
5.3. PARAMETERIZATION OF THE RIGID KINEMATICS 123

can be interpreted as an implicit denition for q = ρ(θ). According to the im-

plicit function theorem, the existence of ρ is guaranteed locally, if the Jacobian
of the nonlinear system (5.31) is non-singular. In other words, the gradient ϕq ,
and the constraint gradient Φq should be linearly independent. Congurations
for which this condition is not satised are either conguration space singular-
ities or actuator singularities [PK99, ZBG01, LLL03]. At a conguration space
singularity, the constraint gradient Φq becomes defective, and the number of
kinematic modes increases suddenly. Those singularities are intrinsic properties
of the conguration space, independent of the parameterization problem. In con-
trast, the location of an actuator singularity (or parameterization singularity)
is conditioned by the choice of the actuated dofs. At an actuator singularity,
the mechanism can move even though all actuators are blocked, which is usually
undesirable from a control point of view.

The regularity of the Jacobian is not sucient to guarantee that the actua-
tor dofs θ are able to parameterize the whole conguration space. A well-dened
parameterization is globally one-to-one: for each given actuator conguration θ,
there exists only one kinematically admissible conguration q (i.e. one solution
to equations (5.31)). This condition is quite dicult to verify; however, in many
practical cases, actuator singularities separate parts of the conguration space
where ϕ is one-to-one, so that the actuator parameterization is valid in a sub-
set Ωrq ⊂ Ωrt q bounded by the singular congurations. All those concepts are
illustrated in Figure 5.3.

In this restricted part of the conguration space, the inverse map ρ is well-
dened:

ρ : Ωrθ → Ωrq , θ 7→ q = ρ(θ) (5.32)

where Ωrθ ⊂ Ωrt θ ⊂ R is the set of possible variations of the actuated dofs θ ,
s

away from the singularities. For usual control applications, actuator singularities
and conguration space singularities are carefully avoided, and the limitation of
the analysis to Ωrq is not restrictive.
124 CHAPTER 5. NONLINEAR MODEL REDUCTION

Figure 5.3: Rigid four bar mechanism, with an actuator at the bottom left hinge.
The conguration space is a 1-dimensional manifold, whose projection in the
plane of the coordinates ze and θ is also represented. Conguration (b) is an
actuator singularity; it separates two sub-domains of the conguration space for
which q = ϕ(θ) is one-to-one. ϕ is not one-to-one in the total conguration
space Ωrt
q , since congurations (a) and (c) are possible for the same angle θ .
5.3. PARAMETERIZATION OF THE RIGID KINEMATICS 125

Figure 5.4: Tangent space and constraint gradient of a rigid mechanism.

Jacobian of the transformation

In the space of dependent coordinates q, a relation exists between the con-
straint gradient and the Jacobian of ρ. Indeed, we have:

Φ(ρ(θ)) = 0, ∀ θ ∈ Ωrθ (5.33)

so that, after dierentiation:

Φq ρθ = 0 (5.34)

ρθ is the Jacobian of the coordinate transformation, which naturally spans the

tangent space of the conguration manifold Tp Ωrq at point p. According to (5.34),
ρθ is orthogonal to the columns of the constraint gradient ΦTq , as represented in
Figure 5.4.
To conclude this discussion, it is worth noticing that the denition of inde-
pendent parameters θ among the generalized coordinates q is formally equivalent
to the constraint elimination technique described in section 3.2.1. Hence, for a
rigid mechanism, the constraint elimination technique can be exploited for model
reduction, as proposed in a previous work [BDG06]. In this dissertation, we focus
on the more general case of exible mechanisms; but the methodology can still be
reinterpreted as a constraint elimination technique if the mechanism is actually
rigid.
126 CHAPTER 5. NONLINEAR MODEL REDUCTION

5.3.2 Flexible mechanisms

In the previous section, the analysis was based on a set of dependent pa-
rameters subject to kinematic constraints. For a exible mechanism, additional
constraints of non-deformation are required for the rigid kinematic analysis.
The n coordinates q still have to satisfy m kinematic constraints:

Φ(q) = 0 (5.35)

The exible manifold Ωfqt is dened as the set of kinematically admissible cong-
urations which satisfy the kinematic constraints:

Ωfqt = {q ∈ Rn | Φ(q) = 0} (5.36)

Moreover, the conguration space Ωrt

q ⊂ Ωq is dened as the set of unde-
ft

formed congurations. The dimension2 s of Ωrtq is smaller than the dimension

n − m of Ωfqt :
s<n−m (5.37)
As for rigid mechanisms, the actuated dofs θ lead to a parameterization of a
restricted part of the conguration space Ωrq ⊂ Ωrt
q . According to the mixed coor-
dinate formulation, those dofs are included among the generalized coordinates q.
The parameterization of the rigid kinematics requires the denition of the
relation between θ and q (or q∗ ) for the undeformed mechanism. The non-
deformation condition is formulated using the elastic potential energy V(q) of the
mechanism. An undeformed conguration satises the combined conditions:
(
∂V
=0
∂q
(5.38)
Φ(q) = 0

This problem can be analyzed numerically using the zero-strain approach. Ac-
cordingly, any equilibrium conguration achieves a minimum of the elastic poten-
tial energy, so that the kinematic problem becomes:
Given θ, nd q∗ :
min
∗
V (5.39)
q

2 The dimension here refers to the intrinsic dimension of the manifolds, as opposed to the
dimension n of the ambient space.
5.3. PARAMETERIZATION OF THE RIGID KINEMATICS 127

subject to
Φ(θ, q∗ ) = 0 (5.40)
According to the augmented Lagrangian method, the following equivalent uncon-
strained problem is considered:
min
∗
V + k λT Φ + p ΦT Φ (5.41)
q ,λ

where λ is the m × 1 vector of Lagrange multipliers, k and p are the scaling and
penalty factors, respectively. The stationarity condition leads to n + m nonlinear
equations: (
∂V
+ ΦTq∗ (k λ + p Φ) = 0
∂q∗
(5.42)
k Φ(q) = 0
Equations (5.42) contains as many equations as unknowns (q∗ ,λ), and can be
solved using standard methods.
The initial problem was to establish q (or q∗ ) from θ; it is equivalent to
construct the relation between u (or u∗ ) and θ, with:
 
θ " #
θ
(5.43)
 
u= ∗ 
 q  = u∗
λ
In the space of augmented coordinates u, the total conguration space is denoted
u . Equation (5.42) gives a condition for static equilibrium which is a necessary
Ωrt
but not sucient condition for an undeformed conguration: some solutions u
of (5.42) might not satisfy (5.38), and be out of Ωrt u . Typically, those solutions
are associated with pre-stressed congurations (λ 6= 0 or ∂V ∂θ
6= 0), which are not
considered in this research.
If the Jacobian of the system (5.42) is not singular, the implicit function
theorem guarantees the local existence of a function ρ such that u = ρ(θ). As
in the rigid case, the global validity of the actuator parameterization can only be
guaranteed in a restricted part of the conguration space Ωru ⊂ Ωrt u , where the
singularities are avoided. The parameterization mapping is dened by:
ρ : Ωrθ → Ωru , θ 7→ u = ρ(θ) (5.44)
where Ωrθ ⊂ Ωrt
θ ⊂ R is the set of authorized variations of the actuated dofs θ .
s

Two important problems are associated with the practical computation of

u = ρ(θ) and its Jacobian for a given θ .
128 CHAPTER 5. NONLINEAR MODEL REDUCTION

Numerical computation of the kinematic mapping

For a given actuator conguration θ, the kinematic problem consists in de-
termining the conguration u = ρ(θ). Starting from a known conguration
uref = ρ(θ ref ), equations (5.42) can be solved for q∗ and λ according to a stan-
dard Newton-Raphson procedure. As in section 3.2.2, the linearized equations
are formulated: " ∗ ∗
# " # " #
Kqt q
k ΦTq∗ ∆q∗ 0
= (5.45)
k Φq∗ 0 ∆λ 0
where Kqt
∗ q∗
is a submatrix of the tangent stiness Kt :


∂ 2 V ∂ ΦTq (k λ + p Φ)
Kt = + (5.46)
∂q2 ∂q

m
2
∂ V ∂ ΦTq λ ∂ 2 Φi
(5.47)
X
T
= + p Φq Φq + k + p Φi
∂q2 ∂q i=1
∂q2

Usually, a good approximation is obtained by neglecting the second-order deriva-

tives of the constraints:
∂2V
Kt ' 2
+ p ΦTq Φq (5.48)
∂q
This approximation is exact at an undeformed conguration, since Φ = 0 and
λ = 0.
The converged solution might be out of Ωru . This problem is avoided if the
reference and the target congurations are suciently close to each other: the
algorithm naturally converges to the unique solution in Ωru , since it is also the
unique solution in the neighborhood of the reference conguration. However, the
danger is important nearby an actuator singularity, where dierent solution sets
to the kinematic problem intersect with each other.
If the reference conguration θref is far from the target θ, the Newton-
Raphson procedure is likely to diverge, or to converge to a solution outside Ωru .
Therefore, a linear homotopy path θp : R → Rs can be dened from θref to θ:

θ p (τ ) = (1 − τ ) θ ref + τ θ τ ∈ [0, 1] (5.49)

The algorithm progresses from τ = 0 to τ = 1 through intermediate congu-

rations. If a failure occurs, additional intermediate congurations are selected
according to a bisection strategy, in order to obtain convergence. The procedure
may not succeed if Ωrθ is non-convex; this rather unusual situation will not be
5.4. THE GLOBAL MODAL PARAMETERIZATION (GMP) 129

detailed here, even though it could be handled by slight customization of the

conguration space inspection algorithm described later in section 5.6.3.

Jacobian of the transformation

This section demonstrates the interpretation of the Jacobian ρθ as a set of
rigid-body modes.
Any undeformed conguration satises the global static equilibrium:
(
∂V
+ ΦTq (k λ + p Φ) = 0
∂q
(5.50)
k Φ(q) = 0

This set of n + m nonlinear equations can be recasted as

F(u) = 0 (5.51)

which is veried by any undeformed conguration:

F(ρ(θ)) = 0, ∀ θ ∈ Ωrθ (5.52)

After dierentiation, we have

Fu ρθ = 0 (5.53)
which can be developed:
" # " # " #
Kt k ΦTq I 0
= (5.54)
k Φq 0 ρ∗θ 0

This expression is equivalent to the denition of the rigid-body modes given

by (5.18) for a linear structure. As a conclusion, the Jacobian ρθ is the set of
rigid-body modes of the linearized static equilibrium.

5.4 The Global Modal Parameterization (GMP)

The concepts dened for the reduced-order parameterization of a deformable
structure and for the minimal parameterization of the rigid kinematics are now
combined for the reduced parameterization of a exible mechanism. The Global
Modal Parameterization results from the construction of a nonlinear coordinate
transformation:
u = χ(η) (5.55)
130 CHAPTER 5. NONLINEAR MODEL REDUCTION

where the n reduced coordinates η are dened in an open set Ωfη ⊂ Rn , and
u = qT λT belongs to a submanifold of the exible manifold: Ωu ⊂ Ωfu . The
 T f

coordinate transformation is thus an invertible mapping:

(5.56)
f
χ : Ωfη → Ωu , η 7→ u = χ(η)

and the reduction comes from

(5.57)
f
dim Ωu = n < dim Ωfu = n − m

The construction of the coordinate transformation χ is based on a partitioned

description of the rigid and of the exible motion.

5.4.1 Partitioned coordinate transformation

Among the reduced coordinates η , we dissociate a subset of s coordinates θ
for the description of the rigid kinematics, and a subset of nδ = n − s coordinates
η δ for the description of the deformations:
" #
θ
η= (5.58)
ηδ
θ is the set of actuator dofs, able to parameterize the conguration space Ωru .
Under the assumption of small deformations, the reduction formula (5.55) is
ane in η δ :
u = ρ(θ) + Ψuδ (θ) η δ (5.59)
ρ is the rigid kinematic mapping dened in section 5.3.2, and Ψuδ is a matrix of
deformation modes which is still to be constructed, and which may depend on the
conguration θ. The possible variations of the reduced coordinates are restricted
to:
and (5.60)
δ
θ ∈ Ωrθ ⊂ Rs η δ ∈ Ωfδ (θ) ⊂ Rn

where Ωrθ has been dened earlier and Ωfδ is a neighborhood of the origin.
Equation (5.59) can be dierentiated:
∂(Ψuδ η δ )
 
δu = χη δη = ρθ + δθ + Ψuδ δη δ (5.61)
∂θ
so that for an undeformed conguration (η δ = 0):

Ψuδ = Ψuη = Ψuθ Ψuδ (5.62)

   
χη = ρθ
5.4. THE GLOBAL MODAL PARAMETERIZATION (GMP) 131

This notation emphasizes the interpretation of the Jacobian ρθ as a set of rigid-

body modes. For consistency, the Jacobian χη should belong to the tangent space
of the exible manifold Tp Ωfu ; this property is important for the construction of
the coordinate transformation, as discussed in the next section.
The reduced parameterization χ of the exible manifold involves the kine-
matic parameterization ρ, which has been analyzed previously, and conguration-
dependent exible mode shapes Ψuδ . In order to exploit component-mode tech-
niques, the denition of this global parameterization starts with a local denition
of mode shapes around an undeformed conguration. The connection of this local
denition with the global parameterization will be demonstrated afterwards.

5.4.2 Denition of the local modes

For a given conguration, the construction of the mode shapes is based on
the linearized dynamic equations. As for the linear component-mode technique,
a partitioning of the initial dofs u is considered
 
θ
(5.63)
 
u= g 
 q 
ui

where θ are the actuator dofs, qg are the constraint dofs, and ui are the internal
dofs, including the Lagrange multipliers.
The mode shapes are selected as an optimal basis to describe the linearized
motion around an undeformed conguration, with zero velocities :
" # " # " # " # " #
M 0 ∆q̈ Kt k ΦTq ∆q 0
+ = (5.64)
0 0 ∆λ̈ k Φq 0 ∆λ 0

The approximated expression of the tangent stiness (5.48) is thus exact and can
be safely exploited.
As in the linear case, three subsets of modes are dened:

Ψuη = Ψuθ Ψuγ Ψuι (5.65)

 

The rigid modes Ψuθ are dened from the static equilibrium (5.18); they are also
the Jacobian of the rigid kinematics: Ψuθ = ρθ . The constraint modes Ψuγ are
the solutions to the static problem (5.19), where the rigid dofs are xed, and
132 CHAPTER 5. NONLINEAR MODEL REDUCTION

Figure 5.5: Local modes of an unconstrained mechanism : 2 rigid modes and one
exible mode.

the internal modes Ψuι are dened according to the internal eigenvalue prob-
lem (5.20), where the rigid and constraint dofs are xed.
For local consistency, those modes should belong to the tangent space of the
exible manifold at point p: Tp Ωfu . Indeed, using (5.18), (5.19), (5.20) and the
structure of the tangent stiness matrix, it is easy to verify that

Φu Ψuη = 0 (5.66)

where Φu = [Φq 0].

Figures 5.5, 5.6 and 5.7 illustrate the denition of the local modes for a few
simple situations.

5.4.3 From local to global modes

The local denition of the exible modes can be reproduced for any con-
guration, leading to conguration-dependent mode shapes. A priori, nothing
prevents spurious discontinuous variations of those mode shapes, and it is neces-
sary to verify that the Global Modal Parameterization χ(η) is well-dened, i.e.

- it is globally one-to-one,
5.4. THE GLOBAL MODAL PARAMETERIZATION (GMP) 133

Figure 5.6: Local modes of an unconstrained mechanism: one rigid mode and
two exible modes.

Figure 5.7: Local modes of a constrained mechanism: one rigid mode, and one
exible mode.
134 CHAPTER 5. NONLINEAR MODEL REDUCTION

- it is continuously dierentiable,
- its rank is maximal (i.e. the Jacobian χη is of rank n),
From the previous developments, χ(η) involves two contributions:
- the kinematic mapping ρ(θ), which is well-dened in the conguration space
of interest, as demonstrated in section 5.3; in particular, it is continuous
and dierentiable,
- the exible modes Ψuδ (θ) have only received a local denition at a given
conguration; their variations in the conguration space deserve a careful
analysis.
The following analysis demonstrates that the three consistency criteria are satis-
ed by the Global Modal Parameterization.

One-to-one criterion
The exible submanifold Ωu is dened as the image of Ωfη mapped by the
f

transformation χ. By construction, for all u ∈ Ωu , there exists a vector η such

f

that u = χ(η). Our purpose is to demonstrate the uniqueness of such a vector η

in Ωfη .
Suppose that another solution η 0 also satises u = χ(η 0 ). Since the actuator
dofs θ appear among the mixed coordinates, we necessarily have θ0 = θ, and the
exible coordinates satisfy
0 0
ρ(θ) + Ψuδ (θ) η δ = ρ(θ) + Ψuδ (θ) η δ ⇒ Ψuδ (η δ − η δ ) = 0 (5.67)
From the linear independence of the exible modes, we deduce that η δ − η δ 0 = 0,
which completes the proof.

Continuity of the constraint modes

The constraint modes Ψuγ are dened from the linear static analysis when
the rigid dofs are xed, see equation (5.19), and we have

−1 ig
Ψiγ (θ) = − Kii (θ) K (θ) (5.68)
Since Kii is invertible3 , and since all the stiness coecients are continuous and
dierentiable, so are the constraint modes.
3 Otherwise, additional rigid modes should be dened.
5.4. THE GLOBAL MODAL PARAMETERIZATION (GMP) 135

Continuity of the internal modes

The internal modes Ψiι are obtained after three steps:

1. the eigenvalue problem is solved:

Kii (θ) − ω 2 Mii (θ) Ψiι∗ (θ) = 0 (5.69)

2. the resulting nι∗ eigenmodes are sorted according to their eigenvalue, and
the rst nι < nι∗ modes are selected,

3. the modes are normalized with respect to the mass matrix so that:

ΨiιT Mii Ψiι = Iιι (5.70)

For one selected eigenvalue ωk and the associated internal mode ψ k , it is

instructive to dierentiate (5.69) with respect to a coordinate θl :
∂Kii ∂Mii ∂ωk2 ii
 

∂ψ k
− ωk2 − M ψ k + Kii − ωk2 Mii =0 (5.71)
∂θl ∂θl ∂θl ∂θl
∂ωk2
A pre-multiplication by ψ Tk leads to the extraction of ∂θl
:
∂ωk2 ∂Kii ∂Mii
 
1 T
= ψ − ωk2 ψk (5.72)
∂θl µk k ∂θl ∂θl

with the modal mass µk = ψ Tk Mii ψ k > 0. Therefore, each eigenvalue ωk exhibits
smooth variations with respect to the coordinate θl .
From the continuity of ωk , and of the mass and stiness matrices, the rst
term in (5.71) is continuous, and so is the second one. However, the derivatives of
the eigenmodes ∂ψ k
∂θl
can take arbitrary large values in the kernel of Kii − ωk2 Mii .
If ωk has a multiplicity κ = 1, ψ k is the single vector in this kernel, which
means that the arbitrary variations are parallel to the initial vector. Those vari-
ations are forbidden due to the normalization condition (5.70), and we conclude
that ∂ψ k
∂θl
is continuous.
This property is no more guaranteed for a multiplicity κ > 1. In this case, the
kernel of Kii −ωk2 Mii can be a multi-dimensional subspace, where the eigenvectors
can take arbitrary variations.
An example is given in Figure 5.8. In the one-dimensional conguration
space, there is one conguration θ1 = θb for which ω1 = ω2 and κ = 2. For θ1 < θb
136 CHAPTER 5. NONLINEAR MODEL REDUCTION

Figure 5.8: Variations of the eigenfrequencies in the conguration space.

or for θ1 > θb , ψ 1 and ψ 2 are thus continuous. By inspection of equation (5.69),

and considering that the eigenvalues are continuous across θb , an eigenmode found
on one side of the boundary θb is still a valid eigenmode on the other side. For
this reason, it is possible to impose ∆ψ 1 = 0 and ∆ψ 2 = 0 from one side of θb
to the other, leading to a continuous parameterization.
This analysis is naturally extended for a s-dimensional conguration space.
The domains such that κ > 1 are b-dimensional manifold Mb , with b < s (the
less usual case where b = s deserves a specic treatment, which is not considered
here). It is still possible to impose ∆ψ k = 0 across the boundary Mb , which
motivates the implementation of a mode tracking strategy at the step 2 of the
construction of the internal modes.
Similar mode-tracking problems arise in topological optimization problems;
for instance, Kim and Kim [KK00] have proposed a method based on the Modal
Assurance Criterion (MAC value) dened in structural identication. The MAC
number is a measure from 0 to 1 of the correlation between two modes, a unitary
value meaning perfect correlation. Our method relies on a correlation of the
nι∗ eigenmodes computed at the current conguration, with nι eigenmodes Ψiιref
associated with a reference conguration. A mode matching algorithm denes
the eigenmodes Ψiι in order to obtain MAC numbers as close as possible to 1.
More precisely, the MAC matrix, which is lled by the MAC numbers of all pairs
of modes, is rendered as close as possible to the identity.
Ideally, all diagonal terms of the resulting MAC matrix are above 0.95, and
the consistency of the internal modes is guaranteed. If the reference conguration
is far from the current conguration, lower MAC values may be obtained, so that
the internal modes Ψiι cannot be validated. Then, the algorithm should restart
the procedure with a closer reference conguration. Therefore, the mode-tracking
5.5. REDUCED-ORDER MODEL 137

problem is partly handled by the conguration space inspection algorithm, which

will be explained in section 5.6.3.

Rank of the Jacobian

The last criterion for a well-dened parameterization is that the Jacobian
χη has full rank. At an undeformed conguration, the columns of χη are the
local modes, see equation (5.62). By construction, their linear independence is
guaranteed, and the rank condition is satised. At a deformed conguration, the
expression of the Jacobian is slightly aected (5.61), but under the assumption
of small deformations, the full-rank requirement is still veried.
From this analysis, we conclude that the Global Modal Parameterization is
well-dened, provided that a mode-tracking strategy is implemented. Hence, it is
a reliable basis for the formulation of a reduced-order model.

5.5 Reduced-order model

From the denition of the Global Modal Parameterization, we shall now for-
mulate the equations of motion in terms of the modal parameters. The Lagrange
equations involve contributions from the elastic potential energy (elastic forces),
the kinetic energy (inertia forces), and the work produced by external forces.

5.5.1 Elastic forces

The elastic potential energy is a strongly nonlinear function V(q), and we
seek for an approximation in terms of the reduced coordinates η . From the dis-
sociation between rigid and exible coordinates, a smoother behavior is expected
for V(θ, η δ ). Around an undeformed conguration p characterized by θ = θ0 and
η δ = 0, let us consider the order 2 Taylor series expansion:
T
∂2V
  
∂V 1
V(η) = Vp + ∆η + ∆η T ∆η (5.73)
∂η p 2 ∂η 2 p

with " #
θ − θ0
∆η = (5.74)
ηδ
138 CHAPTER 5. NONLINEAR MODEL REDUCTION

At the undeformed conguration, the elastic

 potential energy is zero Vp = 0, and
the generalized elastic forces vanish ∂η = 0.
∂V
p
The second order derivatives of V is an equivalent stiness
ηη ∂2V
K = (5.75)
∂η 2
so that
1  ηη 
V = ∆η T
K ∆η (5.76)
2 p

The objective is to establish a connection between the reduced stiness

(K )p and the initial Finite Element formulation. All the following developments
ηη

are realized at the undeformed conguration p. Let us analyze the second-order

derivative of the augmented functional V ∗ = V + k λT Φ + p ΦT Φ:
∂2V ∗ 2 ∗
uηT ∂ V
= Ψ Ψuη (5.77)
∂η 2 ∂u2
" #" #
T qη
h i K k Φ Ψ
(5.78)
t q
= ΨqηT ΨληT
k Φq 0 Ψλη
= ΨqηT Kt Ψqη (5.79)

The property Φq Ψqη = 0 has been used. Since Kt = ∂2V

∂q2
+ p ΦTq Φq , we have:

∂ 2 V qη ∂ 2 V
ΨqηT Kt Ψqη = ΨqηT Ψ = (5.80)
∂q2 ∂η 2
and we conclude:
ηη
K = ΨqηT Kt Ψqη (5.81)
Since by construction of the rigid modes Kt Ψqθ = 0, the equivalent stiness
has the structure " #
ηη 0 0
K = δδ (5.82)
0 K
with
δδ
K = ΨqδT Kt Ψqδ (5.83)
This last formula can be directly exploited in the Finite Element code to compute
the equivalent stiness. Finally, the potential energy is represented by a quadratic
form in the exible coordinates:
1 δT δδ
V(θ, η δ ) = η K (θ) η δ (5.84)
2
5.5. REDUCED-ORDER MODEL 139

The generalized elastic forces are:

∂V δδ
= K (θ) η δ (5.85)
∂η δ

The contribution ∂V
∂θ
is quadratic in the amplitude of deformation; it is neglected
under the small deformation assumption.

5.5.2 Inertia forces

In the initial model, the kinetic energy is a quadratic form of the velocities
1 T
K= q̇ M q̇ (5.86)
2
Considering the coordinate transformation at the velocity level

q̇ = χη η̇ (5.87)

we obtain:
1 T ηη
with
ηη
K= η̇ M η̇ M = χTη M χη (5.88)
2
The generalized inertia forces are dened by:
 
d ∂K ∂K
giner = − (5.89)
dt ∂ η̇ ∂η
A simplied expression is possible if the variations of the reduced mass matrix
M are neglected:
ηη

ηη
giner = M η̈ (5.90)
For a more thorough analysis, let us consider a partitioning of the generalized
coordinates into translation and rotation dofs:
" # " #
x χx (η)
q= = (5.91)
ψ χψ (η)

In the formulation proposed by Géradin and Cardona [GC01], their contributions

to the kinetic energy are uncoupled:
1 T xx 1 T
K = Kx + Kψ = ẋ M ẋ + ψ̇ Mψψ ψ̇ (5.92)
2 2
Mxx is constant, but Mψψ is subject to variations due to the updated Lagrangian
point of view adopted for the rotation parameters.
140 CHAPTER 5. NONLINEAR MODEL REDUCTION

Inertia forces associated with the translation kinetic energy

The contribution of the translation kinetic energy to the generalized inertia
forces is:
d ∂Kx ∂Kx
 
gtrans
iner = − (5.93)
dt ∂ η̇ ∂η
 T     T
∂ ẋ d ∂ ẋ ∂ ẋ
= xx
M ẍ + − Mxx ẋ (5.94)
∂ η̇ dt ∂ η̇ ∂η
We also have
ẋ = χxη η̇ and ẍ = χ̇xη η̇ + χxη η̈ (5.95)
so that
∂(χxη η̇) T xx x
 
gtrans
iner = χxη T M xx
χxη η̈ + χxη T M xx
χ̇xη x
η̇ + χ̇η − M χη η̇ (5.96)
∂η
Using the index summation convention, we develop:
d ∂χxi ∂ 2 χxi
 
χ̇xη ij (5.97)


= = η̇k
dt ∂ηj ∂ηk ∂ηj
∂ χ̇xη η̇ i


∂(χxη η̇)
 x 
∂ 2 χxi
 
∂ ∂χi
= = η̇k = η̇k (5.98)
∂η ij ∂ηj ∂ηj ∂ηk ∂ηj ∂ηk

Thus, the last term in (5.96) vanishes. It is convenient to dene the third order
curvature tensor Γχx :
∂ 2 χxi
with (5.99)
x x x
Γχ
ijk = Γχ χ
ijk = Γikj
∂ηj ∂ηk
so that
(5.100)
x x
χ̇xη = Γχ χ



ij ijk η˙k = Γ . η̇ ij

where the ”.” operator is dened as a product of a third order tensor by a vector.
Hence, the translation inertia forces are concisely formulated

(5.101)
xT x
gtrans xx
χxη η̈ + χxη T Mxx Γχ . η̇ η̇


iner = χη M

This expression is composed of two terms:

- the inertia forces obtained when the variations of the reduced mass matrix
are quasi-static,

- the equivalent forces associated with the curvature (or the nonlinearity) of
the coordinate reduction formula.
5.5. REDUCED-ORDER MODEL 141

Inertia forces associated with the rotation kinetic energy of a rigid

body
The rotation kinetic energy of an isolated rigid body A is expressed by:
1 T
Kψ,A = Ω JΩ (5.102)
2
where Ω is the 3 × 1 vector of material angular velocities, and J is the inertia
tensor. According to Géradin and Cardona [GC01], Ω is connected to the orien-
tation parameters α of the body by the 3 × 3 tangent operator T of the rotation
parameterization:
Ω = T(α) α̇ (5.103)
In the Finite Element formulation, an updated Lagrangian point of view is
adopted, and the reduction procedure is considered at an updated conguration
α = 0. If the rotations are parameterized using the Cartesian rotation vector or
the conformal rotation vector, we have the useful properties:
∂Tij 1
Tij (0) = δij and (0) = ijk (5.104)
∂αk 2
where δij is the Kronecker symbol, and ijk is the alternating symbol, dened by

 1 if (ijk) is a cyclic permutation of (123),


ijk = −1 if (ijk) is an anticyclic permutation of (123), (5.105)
 0 otherwise.



The parameters α are a subset of the orientation parameters ψ . If we dene

χ = Lαψ χψ , where Lαψ is a localization matrix, we have:
α

α = χα (η) and α̇ = χα
η η̇ (5.106)

Introducing the operator χΩη :

χΩη = T χα
η (5.107)

we have
Ω = χΩη η̇ (5.108)
and the developments realized for the translation kinetic energy can be repro-
duced:
T
∂(χΩη η̇)

grot,A
iner =χ Ωη T
Jχ Ωη
η̈+χ Ωη T
J χ̇ Ωη Ωη
η̇+ χ̇ − J χΩη η̇ (5.109)
∂η
142 CHAPTER 5. NONLINEAR MODEL REDUCTION

χΩη is not a Jacobian, and the last term does not vanish in this case. Let us
develop:
∂χα ∂ 2 χα
 
∂Tiq
Ωη q q
(5.110)


χ̇ ij
= + Tiq η̇k
∂ηk ∂ηj ∂ηj ∂ηk
∂χα ∂ 2 χα
 
∂ ∂Tiq
χΩη η̇ i =
q q
(5.111)


+ Tiq η̇k
∂η j ∂ηj ∂ηk ∂ηj ∂ηk

where
∂Tiq ∂Tiq ∂αl 1 ∂χα
= = iql l
(5.112)
∂ηk ∂αl ∂ηk 2 ∂ηk
Introducing the antisymmetric third-order tensor ΓΩ :
1 ∂χα α
q ∂χl
ΓΩ
ijk = iql with ΓΩ Ω
ijk = −Γikj (5.113)
2 ∂ηj ∂ηk
we obtain
 
(5.114)
α
Ωη
ΓΩ Γχ


χ̇ ij
= ijk + ijk η̇k
∂  
(5.115)
χ α
χΩη η̇ i = ΓΩ


ikj + Γijk η̇k
∂ηj

Those expressions can be introduced in (5.109). Using the antisymmetry of ΓΩ ,

and observing that ΓΩ
ljk η̇k η̇j = 0, we get:
  h i
grot,A
iner = (J χα )
η li (χ α
η η̈)l + (J χ α
) Γχα
η li ljk + 2 ΓΩ
lik (J χ α
η lj η̇k η̇j
) (5.116)
i

and in matrix form:

(5.117)
χα

T
grot,A αT α αT
. η̇ η̇ + 2 ΓΩ . η̇ J χα


iner = χη J χη η̈ + χη J Γ η η̇

The gyroscopic tensor of the rigid body hΩ is dened by:

hΩ
ijk = Jjl lik (5.118)

After developments, the inertia forces can be restated:

η η̇) (χη η̇) (5.119)

χα
grot,A αT α αT
. η̇ η̇ + χαT
 Ω
h . (χα α


iner = χη J χη η̈ + χη J Γ η

This expression is composed of three terms:

- the inertia forces obtained when the variations of the reduced mass matrix
are quasi-static,
5.5. REDUCED-ORDER MODEL 143

- the equivalent forces associated with the curvature (or the nonlinearity) of
the coordinate reduction formula,

- the third term is a gyroscopic force:

ggyr,A αT gyr,A
(5.120)
 Ω 
iner = χη h . α̇ α̇ = χαT
η giner

which is interpreted as the projected gyroscopic forces of the initial model.

This generalized formulation for the inertia forces of an isolated rigid body
is also valid in the particular case where no coordinate transformation is applied:

(5.121)
α
η = α, χα
η = I, Γχ = 0

Let us demonstrate the connection of equations (5.119) with the classic Euler
equations in this case. For a rigid body described by a diagonal inertia tensor
 
J1 0 0
(5.122)
 
J=
 0 J2 0 

0 0 J3

the non-zero components of the gyroscopic tensor hΩ are:

hΩ
123 = −J2 , hΩ
132 = J3 ,
hΩ
213 = J1 , hΩ
231 = −J3 , (5.123)
hΩ
312 = −J1 , hΩ
321 = J2

Using η = α, equation (5.119) leads to

 
J1 α̈1 + (J3 − J2 ) α̇3 α̇2
grot,A (5.124)
 
iner =  J2 α̈2 + (J1 − J3 ) α̇1 α̇3 
 
J3 α̈3 + (J2 − J1 ) α̇2 α̇1

The conclusion comes from the observation that α̇ = Ω and α̈ = Ω̇ for an

updated conguration. Indeed, we have:
1
Ṫij α̇j = ijk α̇k α̇j = 0 ⇒ Ω̇ = T α̈ + Ṫ α̇ = α̈ (5.125)
2
144 CHAPTER 5. NONLINEAR MODEL REDUCTION

Inertia forces associated with the global rotation kinetic energy

The inertia forces associated with the rotation dofs ψ involve the contribu-
tions of all rigid-bodies of the mechanism (see previous section), as well as the
slightly more complex contributions of the exible bodies. However, it is easily
veried that the global forces can still be decomposed into three terms:

- a force proportional to the modal accelerations:

χψT
η M
ψψ
χψ
η η̈ (5.126)

- an equivalent force associated with the curvature tensor of the coordinate

transformation:  ψ 
χψT
η M ψψ
Γχ . η̇ η̇ (5.127)

- the projected gyroscopic forces

 
χψT (5.128)
 gyr ψ
 ψ ψT gyr
η h . (χη η̇) (χη η̇) = χη h . ψ̇ ψ̇

where hgyr is the gyroscopic tensor of the initial model.

Assembled inertia forces

As a conclusion of the previous developments, the assembled inertia forces
have the form
(5.129)


giner = M η̈ + h . η̇ η̇

with

M = χTη M χη (5.130)
χ gyr
h=h +h (5.131)

The expression of the reduced mass matrix is classical, and two dierent third-
order tensors are responsible for quadratic forces in the modal rates.
χ
(i) The tensor h is associated with the curvature Γχ of the reduced param-
eterization:
χ
hijk = χTη M il Γχ (5.132)


ljk

Γχ can be estimated by dierentiation of the Jacobian χη using a nite dierence

approach.
5.5. REDUCED-ORDER MODEL 145

gyr
(ii) The tensor h satises
gyr
. η̇) η̇ = χψT (5.133)
 gyr
h . (χψ ψ ψT gyr

(h η η η̇) (χη η̇) = χη g

and it could be computed from hgyr , according to a tensor projection. However,

in the Finite Element code, manipulations of third-order tensors are avoided and
hgyr is not available. For this reason, we propose to proceed by inspection from
a computation of the gyroscopic force vector ggyr . A unit velocity is imposed
to each pair of variables (η̇i , η̇j ), and we deduce the rates of the Finite Element
coordinates q̇ = χη η̇ . Then, the gyroscopic forces ggyr (q, q̇) are obtained us-
ing standard Finite Element routines. This vector is projected to obtain the
equivalent gyroscopic forces ggyr = χTη ggyr , which allows a component-wise iden-
gyr
tication of the gyroscopic tensor h .

Simplication of the inertia forces

Theoretically, the mass matrix and the gyroscopic tensor formulated above
depend on the whole set of reduced parameters η . However, in the following, the
dependence with respect to the exible coordinates η δ is neglected, so that we
have
(5.134)


giner = M(θ) η̈ + h(θ). η̇ η̇

The mass matrix and the gyroscopic tensor are only evaluated at undeformed
congurations.
The curvature tensor is required for the construction of h. From the ex-
pression of the Jacobian (5.61), it has an interesting structure at an undeformed
conguration:

∂ 2 χi ∂ 2 ρi ∂Ψuθ
= =
ij
(5.135)
∂θj ∂θk ∂θj ∂θk ∂θk
2 uδ
∂ χi ∂Ψik
δ
= (5.136)
∂θj ∂ηk ∂θj
2
∂ χi
= 0 (5.137)
∂ηkδ ∂ηjδ

It is associated with the sensitivities of the rigid and exible modes, which are
easily estimated by nite dierence.
146 CHAPTER 5. NONLINEAR MODEL REDUCTION

5.5.3 Reduced equations of motion

After the development of the elastic and inertia forces, the equivalent external
loads are estimated from a virtual work expression:
T
δW = gext T
δq = gext Ψqη δη = gText δη, with gext = Ψqη T gext (5.138)

The reduced equations of motion follow

(5.139)


M(θ) η̈ + h(θ). η̇ η̇ + K(θ) η = gext

with the structure:

 θθ θγ θι
    
a g
M M M 0 0 0 gext + ΨgθT gext
 γθ γγ γι   γγ   g

M=
 M M M ,
 K=
 0 K 0 , gext = 
 gext 

ιθ ιγ
M M I 0 0 Ω2 0
(5.140)
a
gext denotes the actuator forces, and gext
g
the forces applied to the constraint dofs.
This model is fully described by M, K, h and Ψgθ , which smoothly depend on θ.

5.5.4 Algorithm for local model reduction

Assuming that a reference model is available at a conguration θref (with
local modes Ψref ), the construction of the reduced-order model for a conguration
θ involves the following steps:

1. kinematic analysis (from θref to θ),

2. local mode synthesis (Ψuη ),

3. mode matching (Ψ − Ψref ),

4. model reduction (M, K, h and Ψgθ ).

An overview of the local order reduction algorithm is given in Figure 5.9. This
local analysis is combined with a conguration space approximation algorithm,
presented in the next section. Hence, a database is exploited to store and reuse the
local models. As mentioned earlier, the mode matching may fail if the reference
conguration θref is too far from θ, leading to a non-valid model.
5.5. REDUCED-ORDER MODEL 147

Search in database yes

nd θ ?

no
Kinematic analysis
?

u = ρ(θ)

Local mode synthesis Check validity

? ?

no
yes-
Ψ uη
Valid model?

Mode matching
?
no-
MAC(Ψuι , Ψuι Non-valid model
ref )

yes
Model reduction
?

M, K, h

?
Valid model

Figure 5.9: Construction of the reduced model in the neighborhood of a congu-

ration θ.
148 CHAPTER 5. NONLINEAR MODEL REDUCTION

5.6 Approximation in the conguration space

The local construction of the reduced-order model involves a computationally
demanding numerical procedure. However, the variations of the model in the con-
guration space can be described by a simplied and approximated metamodel,
or model of the model. If all relevant coecients associated with M, K, h and
Ψgθ are collected in a single t × 1 output vector f , the numerical reduction proce-
dure is a smooth black box function f (θ), and the metamodel is an approximation
f (θ) for this function:
b

f = f (θ) ⇒ f 'b
f (θ) ∀ θ ∈ Ωrθ (5.141)

Obviously, there is a trade-o between the accuracy of the approximation and

the complexity of the metamodel.
This standard approximation problem received a major interest in the litera-
ture. In particular, a strong connection can be established with response surface
methods proposed in optimization [MM95], which take advantage of an approxi-
mated model to reduce the number of runs of a full model.
In the state of the art (section 2.4.2), several approximation techniques were
reported, such as neural networks, polynomial approximations, rational approxi-
mations, kriging functions and the local model network approach. In general, the
metamodel is elaborated according to the following procedure:

1. select a set of inputs θ(1) , ..., θ(r) ;

2. run the local reduction algorithm for each input and get back f (1) ...f (r) ;

3. dene a generic approximated function bf (W, θ), where W is a matrix of

free parameters;

4. select W to t bf on the data obtained in step 2.

Step 1 is also referred to as the experimental design problem. Step 2 was the
subject of our previous developments. In step 3, many choices are possible, and
we restrict the study to linear approximations:
nv
(5.142)
X
fbi (θ) = Wij vj (θ)
j=1
5.6. APPROXIMATION IN THE CONFIGURATION SPACE 149

where vj (j = 1, ..., nv ) are xed basis functions, and Wij are the free weights
which allow to t the model. The linearity is with respect to the weights, and
not to the input variables. Step 4 is then a linear regression problem which can
be solved in the least-square sense using standard methods.
Radial basis functions are an interesting candidate for the basis functions:
vj = vj (kθ − θ (j) k) j = 1, ..., r (5.143)
If the number of radial functions equals the number of given congurations θ(j) ,
it is possible to interpolate exactly any set of scattered data. However, for large
data sets, the resulting function bf may become computationally inecient, and
the number of basis functions should be reduced following a non-trivial selection
procedure. Moreover, the choice of appropriate radial functions is not systematic,
and sometimes requires nonlinear optimization strategies.
In this research, we prefer low-order polynomial basis functions, which are
more simple and ecient for our problem. In order to increase the exibility of
the approximation, a piecewise strategy is adopted as proposed in [BDG06], so
that the above procedure is applied for several non-overlapping subsets of the
conguration space Ωrθ . The decomposition into subsets can be realized adap-
tively in order to satisfy a specication on the approximation error. Thus, a
general, ecient and systematic approximation procedure is developed leading to
a portable and computationally ecient nonlinear model.
The next section presents the piecewise strategy for approximation. After-
wards, the polynomial basis functions are dened and an algorithm is proposed
for automatic conguration space decomposition.

5.6.1 Piecewise strategy

A piecewise approximation is dened over a collection of subsets that cover
the conguration space. In computational mechanics, several methodologies have
been proposed to enforce the consistency of a piecewise approximation, e.g. in a
Finite Element context. Most of them are dedicated to 2 or 3 dimensional spaces,
and they are not applicable for our problem if the conguration space has a higher
dimension. This section describes a more general approach for the decomposition
of the conguration space using the concept of subpaving, developed in the theory
of interval analysis.
150 CHAPTER 5. NONLINEAR MODEL REDUCTION

Figure 5.10: Inner and outer approximation of the conguration space using
subpavings.

Interval analysis theory [JKDW01] oers systematic methods for the approxi-
mation of a complex set using simple subsets. If [θ] ⊂ R represents a real interval,
with a lower and an upper bounds:
[θ] = [θ θ] (5.144)
an interval vector [θ] ⊂ Rs is dened by the cartesian product of s real inter-
vals [θi ]:
[θ] = [θ1 ] × [θ2 ] × . . . × [θs ] (5.145)
Roughly speaking, an interval vector is a s-dimensional box that can be repre-
sented by two corners: the lower bound θ = [θ1 θ2 . . . θs ]T and the upper bound
θ = [θ1 θ2 . . . θs ]T .
A subpaving X is a set of nb non-overlapping boxes4 :
X = {[θ]1 , . . . , [θ]nb } (5.146)
The conguration space Ωrθ can be approximated from inside and outside by two
subpavings, see Figure 5.10:
X − ⊂ Ωrθ ⊂ X + (5.147)

A subpaving is an appropriate basis for the denition of piecewise functions.

If B = {0, 1} denotes the set of booleans, we dene the activation functions
τ k : X → B:
τ k (θ) = 1 ∀ θ ∈ [θ]k
k = 1, ..., nb (5.148)
τ k (θ) = 0 / [θ]k
∀θ∈
4 The interested reader may nd a more accurate denition in [JKDW01].
5.6. APPROXIMATION IN THE CONFIGURATION SPACE 151

and a piecewise approximation function can be formulated

nb X
nv
(5.149)
X
fbi (θ) = Wijk τ k (θ) vj (θ − θ k )
k=1 j=1

where vj are a low-order polynomials, dened for each boxes in local coordinates.
Hence, the construction of an approximated piecewise model in the conguration
space involves two problems:

- the denition of an appropriate inner subpaving X − ⊂ Ωrθ ,

- the selection of local basis functions vj .

In the following, possible choices for polynomial approximation are rst dis-
cussed; the construction of an appropriate subpaving is addressed later.

5.6.2 Local polynomial approximation

Two approaches for the local polynomial approximation are hereby consid-
ered and compared: quadratic polynomials and Lagrange polynomials. For nota-
tional convenience, the presentation focuses on a single component i of the model
f = fi which depends on a s-dimensional conguration vector θ . The interpo-
lating polynomial is denoted Ps = bfi , and the weight matrix becomes a vector
wj = Wij .

Quadratic polynomials
For a one-dimensional problem, a quadratic polynomial P1 is dened by

P1 = a + b θ + c θ 2 (5.150)

This formula can be generalized to s-dimensional problems:

s s X
s
(5.151)
X X
Ps = a + bi θi + cij θi θj
i=1 i=1 j=i

The basis functions vk (k = 1, ..., nv ) are the monomials 1, θi (i = 1, ..., s) and θi θj

(i = 1, ..., s, j = i, ..., s), and their weights are the coecients of the polynomial.
152 CHAPTER 5. NONLINEAR MODEL REDUCTION

Full factorial Fractional Factorial Central Composite Design

Figure 5.11: Experimental designs in 3D [Mon97].

If one assumes that the products θi θj are pre-computed5 , the computation of Ps

requires nop oating-point operations, with:

nop (s) = s2 + 3s (5.152)

Considering a set of r data f (k) = f (θ(k) ), (k = 1, ..., r and r > nv ), the

denition of the weights is a linear regression problem:
     
(1) (1)
f (1) v1 . . . vnv w1
..   .
 . ... ..   . 
 .  (5.153)
  .
.  . 
  . 


 =
(r) (r)
f (r) v1 . . . vnv wnv

where vj(k) = vj (θ(k) ). This overdetermined set of equations can be solved in the
least square sense, using standard regression algorithms (normal equations, QR
algorithm or Singular Value Decomposition).
Several authors (e.g. Montgomery [Mon97]) argued that the data points θ(k)
should be selected according to an experimental design in order to improve the
quality of the approximation. Figure 5.11 illustrates classical denitions: the full
factorial design, the fractional factorial design, and the central composite design.
More sophisticated methods, such as the Taguchi approach, seem unnecessarily
complicated for our problem.
The central composite design is retained here, since it provides sucient
information to estimate the quadratic eects required for a second-order poly-
5 For a function f : Rs → Rt , the values of θi θj are computed once for the t output com-
ponents, and the associated computational cost is thus far less signicant than for the other
operations.
5.6. APPROXIMATION IN THE CONFIGURATION SPACE 153

nomial approximation. The axial position of the "star points" is selected at the
intersection with the border of the box.6
Due to the piecewise strategy, the approximation function exhibits a discon-
tinuous behavior at every boundary between boxes. This important drawback
can be overcome using the family of Lagrange polynomials (see [ZT89, PTVF92]
for detailed presentations), so that a C0 continuity is obtained at every boundary
between matching boxes.

Lagrange polynomials
Let us consider the approximation of a real function f : R → R, from p data:
f (k) = f (θ(k) ), k = 1, ..., p. It is well-known that an order p − 1 polynomial can
be tted exactly on those data, for instance using the p Lagrange polynomials
L(i) , which are constructed to satisfy the condition:

L(i) (θ(k) ) = δik (5.154)

δik is the kronecker symbol. This condition leads to the Lagrange interpolation
formula: p p
θ − θ(k)
(5.155)
Y X
(i)
L (θ) = = Aik θk−1
θ(i) − θ(k) k=1
k=1
(k 6= i)

The last equality denes the components of the matrix A, which is implemented
in the numerical code for ecient computations. Therefore, the interpolating
polynomial of the function f is simply
p p
(5.156)
X X
P1 (θ) = f (i) (i)
L (θ) = f ∗(k) θk
i=1 k=1

with p
(5.157)
X
f ∗(k) = f (i) Aik
i=1

This idea can be recursively implemented for s-dimensional problems, if the

data points are placed in a s-dimensional grid. Suppose that an interpolating
polynomial Ps−1 is available for s−1 dimensional problems, and consider ps values
6 For one-dimensional problems, in order to avoid superimposation on the other data points,
the star points are xed at intermediate positions.
154 CHAPTER 5. NONLINEAR MODEL REDUCTION

θ2 6

(3)
θ2 × × ×

(2)
θ2 × × ×

(1)
θ2 × × ×

-
(1) (2) (3)
θ1 θ1 θ1 θ1

Figure 5.12: Two-dimensional conguration space with a quadratic interpolation

formula (s = 2, p = 3).

for the additional coordinate θs . The s-dimensional interpolating polynomial is

obtained in two steps:

- for each xed value θs(i) (i = 1, ..., ps ), dene an interpolating polynomial

Ps−1 for the s − 1 dimensional problem,
(i)

- exploit 1-dimensional polynomials L(i) (i = 1, ..., ps ) for interpolation in

direction θs .

This is summarized in the following formula:

ps ps
∗(k)
(5.158)
X (i)
X
(i)
Ps (θ) = Ps−1 L (θs ) = Ps−1 θsk
i=1 k=1

with
p
∗(k)
(5.159)
X (i)
Ps−1 = Ps−1 Aik
i=1

According to this last equation, the construction of Ps−1∗(k)

relies on a recur-
sive implementation of addition and scalar multiplication for polynomials. In
Figure 5.12, a two-dimensional problem with a quadratic interpolation formula
(s = 2, p = 3) is considered, and equation (5.158) becomes
3
3 X  
(5.160)
X (i) (j)
P2 (θ1 , θ2 ) = f θ1 , θ2 Li (θ1 ) Lj (θ2 )
i=1 j=1
5.6. APPROXIMATION IN THE CONFIGURATION SPACE 155

Figure 5.13: Discontinuity of Lagrange polynomials between boxes with non-

matching grids: the circled data points do not belong to the grid of the left
box.

A remarkable properties of the Lagrange interpolation, is that the s − 1

dimensional function obtained by xing θs = θs(k) is not inuenced by the data
associated with θs(i) , i 6= k. Therefore, if two neighbor boxes share the same s − 1
dimensional grid at their boundary, the piecewise interpolation is continuous.
The continuity is lost when the subpaving is not made of matching boxes, for
instance due to an adaptive strategy for the conguration space decomposition,
as illustrated in Figure 5.13. Such an adaptive strategy is usually adopted to
improve the compromise between accuracy and memory storage. Hence, the
continuity of the model is a third conicting criterion to be considered for the
denition of an appropriate approximation strategy.
In this work, the second-order Lagrange interpolation has been selected,
which relies on p = 3 data in the 1-dimensional case:

P1 = f ∗(1) + f ∗(2) θ + f ∗(3) θ2 (5.161)

If θ2 is pre-computed7 , each component of bf requires nop (1) = 4 operations (2

additions and 2 multiplications).
An s-dimensional interpolation involves 3s−1 1-dimensional polynomials, and
we get:
∗(1) ∗(2) ∗(3)
Ps (θ) = Ps−1 + Ps−1 θs + Ps−1 θs2 (5.162)
7 As for quadratic polynomials, the computation of θ2 is computed once for the t output
components, and the associated computational cost is far less signicant than for the other
operations.
156 CHAPTER 5. NONLINEAR MODEL REDUCTION

the number of operations per component nop satises

nop (s) = 3 nop (s − 1) + 4 (5.163)

This dierence equations has the solution

nop (s) = 2 (3s − 1) (5.164)

In order to make a connection with standard approximation techniques, equa-

tion (5.162) could be rewritten as a sum of 3s monoms or basis functions, with
3s weights allowing to t exactly the 3s data.

Comparison Quadratic/Lagrange polynomials

The Lagrange polynomials have more exibility, with the advantages of ex-
act interpolation, and C0 continuity for neighboring boxes with matching grids.
Moreover, recursive concepts make the implementation reliable and ecient. The
price to pay is the manipulation of higher-order polynomials, whose construc-
tion requires more data points. For dimensions s = 1 to 6, Table 5.1 compares
quadratic and Lagrange approximation for a function f : Rs → R with respect
to

- the number of operations nop to estimate Ps ,

- the number of weights ndb , that should be eventually stored in a database

for each component of the model,

- the number of data points ncf g required to compute the weights.

From this analysis, Lagrange polynomials are less attractive for high-dimensional
problems. Anyway, a compromise has to be dened between:

- the accuracy,

- the continuity,

- the computational load for the approximation function,

- the memory storage requirement,

- the number of data point necessary to construct the approximation.

5.6. APPROXIMATION IN THE CONFIGURATION SPACE 157

s nop ndb ncf g

Quadratic Lagrange Quadratic Lagrange Quadratic Lagrange
1 4 4 3 3 5 3
2 10 16 6 9 9 9
3 18 52 10 27 15 27
4 28 160 15 81 25 81
5 40 484 21 243 43 243
6 54 1456 28 729 77 729
s2 +3s+2
s s2 + 3s 2(3s − 1) 2
3s 2s + 2s + 1 3s

Table 5.1: Comparison between quadratic and Lagrange polynomials.

5.6.3 Automatic conguration space decomposition

Quadratic and Lagrange polynomials are dened in a box [θ], on the basis
of information collected at the data points. The approximation error within the
box is directly connected with

- the smoothness of f in [θ],

- the size of [θ].

Ideally, the size of the boxes should be adapted according to the behavior of f , in
order to improve the compromise between accuracy, continuity, memory storage
and ecient construction of the approximation function. For instance, close to
singularities, a renement is suitable to track the strong variations of the dynamic
parameters. The adaptive algorithm presented here relies on the representation
of the subpaving as tree data structure and it allows a high-level denition of the
approximation error.
Before the description of the algorithm, let us dene the algorithmic object
model, at the basis of the conguration space inspection procedure. A model is
a data structure containing:

- the box [θ], dened by θ and θ,

- a pointer to the approximation function bf ,

- a boolean valid, true if the approximated function is a valid model,

158 CHAPTER 5. NONLINEAR MODEL REDUCTION

- two pointers leftModel, and rightModel to child models that might be de-
ned after the bisection of the box [θ], thus, a model is binary tree structure
which represents a subpaving,

- an integer bisectDirection, that species the possible bisected dimension.

Temporary variables are also dened

- eval and ebi (i = 1, ..., s) are representative of the approximation error in the
box, they will be dened later,

- cfgSet is the set of congurations dened in the box according to the

experimental design,

- θref is a reference conguration extracted from the database, for which a

valid local model is available; it may belong to cfgSet, otherwise it should
be as close as possible to the box [θ],

- θ is the current conguration.

Figure 5.14 illustrates the construction of a model associated with a box [θ].
Local models are dened for every data points of the conguration set cfgSet, and
after, the approximation function is constructed. The model is bisected if the local
analysis failed at any data point or if the approximation error is not acceptable.
In those cases, a bisection of the model leads to the recursive denition of two
child models. In the following, the error analysis, and the selection of the bisection
direction are presented with more details.

Error analysis
The relative approximation error is dened by:

f (θ) − f (θ)
b
e(θ) = >0 (5.165)
kf (θ)k

Since the components of the vector f do not have the same physical meaning
and the same order of magnitude, it is recommended to replace the norm by a
weighted norm:
√
kf kD = f T D f (5.166)
5.6. APPROXIMATION IN THE CONFIGURATION SPACE 159

Initialize
valid, ebi , eval , cfgSet, θ ref , θ

Local model construction Next conguration

?

(Figure 5.9) cfgSet → θ


θ→f
6

Update status
?

valid

All congurations treated?

?
no
cfgSet

yes
Valid model? Approximation
?
yes-
valid f
b

no
Error analysis
?

eval , ebi

Bisection direction Accuracy criterion

? ?

no
yes-
ebi , toliw eval
< tol e

Recursive procedure
?

Build leftModel
Build rightModel
?

Figure 5.14: Construction of the global model in a box [θ].

160 CHAPTER 5. NONLINEAR MODEL REDUCTION

where D = diag(d1 , ..., dt ) is a diagonal matrix of positive weights. Each weight

di should be adjusted according to the physical meaning of the component fi (e.g.
component of M , component of K , etc.).
θθ γγ

The validation of the approximation relies on the analysis of the error for a
set of validation congurations θval(i) (i = 1, ..., nval ). The average error of the
validation set is: n val
1 X  
eval
= e θ val(i)
(5.167)
nval i=1
The criterion for the validation of the approximation is dened by

eval < tole (5.168)

where tole is a tolerance. For critical applications, eval can be dened as the
maximal error observed in the validation set.
In order to select the optimal bisection direction, the sensitivity of the relative
error with respect to the conguration parameters θ is analyzed. In particular,
the sensitivity of e with respect to a component θi is dened by
∂e
(5.169)
∂θi
Assuming that the error vanishes at the center of the box, a rst order approx-
imation of the maximal error encountered when moving to the side of the box
along direction i is given by:
  
 ∂e θi − θi 
ebi = 
 (5.170)
∂θi 2 

Therefore, the optimal bisection direction is dened as the direction which maxi-
mizes ebi .
The sensitivities of the error are computed using a nite dierence method.
The validation set can be exploited for this purpose, for instance, by the de-
nition of validation congurations at mid-distance star-points, as illustrated in
Figure 5.15.

Selection of the bisection direction

As described in Figure 5.14, if the denition of a local model failed, or if the
approximation error eval is not acceptable, the algorithm achieves a bisection of
5.6. APPROXIMATION IN THE CONFIGURATION SPACE 161

θ2 6

θ2
b

θ2center b × b
b

θ2

-
θ1 θ1center θ1 θ1

Figure 5.15: Validation congurations for error and sensitivity analysis in a two
dimensional conguration space. The validation points are represented by the
circles.

the box into two child boxes. The choice of the bisection direction has a critical
inuence on the eciency of the procedure.
First, it is forbidden to bisect in any direction i for which

∆i < 2 toliw (5.171)

where toliw (i = 1, ..., s) is a minimal tolerance on the width in direction i, nec-

essary to ensure the termination of the algorithm. If no error information is
available, the direction associated with the largest normalized width ∆i /toliw is
bisected. If the error analysis has been performed, the maximal value of ebi denes
the bisection direction.
To conclude, the adaptive decomposition of the conguration space limits
the diculties associated with the curse of dimensionality, and it is especially
valuable for complex mechanisms with high-dimensional conguration space.

5.6.4 Additional parameters of the model

The vector f can collect the components of M, h, K, and Ψgθ , necessary
to describe the dynamics associated with the modal coordinates. However, any
other relevant information can also be included in this vector, for instance, the
modal contributions of the gravity forces, the amplitudes of the mode shapes at
a sensor coordinate, etc.
162 CHAPTER 5. NONLINEAR MODEL REDUCTION

5.7 Summary of the reduction procedure

After the detailed description of the reduction algorithm, and before consid-
ering a couple of applications, let us present an overview of the concepts proposed
in this chapter, which are illustrated in Figure 5.16.
At an undeformed conguration with zero velocities, the linearized equations
are formulated for the assembled mechanism, and exploited for the local mode
synthesis. The local modes are consistent with a nonlinear Global Modal Pa-
rameterization (GMP), in the sense that they are connected with its Jacobian.
From an analysis of the potential and kinetic energies, the Lagrange equations are
formulated, and the components of the stiness matrix, of the mass matrix, and
of the gyroscopic tensor are identied using the local reduction algorithm. The
approximation technique allows a nonlinear, simplied and portable expression
of the reduced model.

5.8 Applications
Two applications are treated in this chapter: an academic exible four-bar
mechanism, and a rigid parallel-kinematic machine-tool, called Orthoglide.

5.8.1 Four-bar mechanism

Figure 5.17 illustrates a four-bar mechanism with large conguration changes.
The mechanism is controlled by a motor at the lower left hinge, so that the rigid
motion is parameterized with the hinge angle θ. Congurations (a) and (e) are
associated with actuator singularities, and they dene the limits of our study in
the conguration space (actually, the conguration space of interest is restricted
to θ ∈ [−1.75, 1.75] rad). At the singular congurations, the actuator is not able
to control the motion of the mechanism, and the aligned links may bend upward
or downward, depending on other external forces. It is noticeable that another
actuator location would lead to another actuator singularity.
We consider that the operations of a tool cause vertical loads on the upper
right hinge, and one constraint mode is therefore associated with the vertical
displacement ze of that point. Three internal modes are also selected to represent
the deformations of the mechanism. The initial Finite Element model contains
5.8. APPLICATIONS 163

GLOBAL LEVEL LOCAL LEVEL

undeformed conguration, zero velocity

" # " #
M 0 4q̈
M(q) q̈ + ΦTq λ − g(q, q̇, t) = 0
0 0 4λ̈
→ " # " # " #
Kt k ΦTq 4q gext
Φ(q) = 0 + =
k Φq 0 4λ 0
↓

" #
h i ∆θ
∆u = Ψuθ Ψuδ
u = ρ(θ) + Ψuδ (θ) η δ ↔ ηδ
= Ψuη ∆η

↓ ↓

1 δδ
V= 2
η δT K (θ) η δ M = ΨqηT M Ψqη
δδ
→ K = ΨqδT Kt Ψqδ
χ gyr
K= 1
2
η̇ T M(η) η̇ h=h +h

.
↓
Approximation
" #
0
M η̈ + (h.η̇) η̇ + δδ = ΨqηT gext
K ηδ

Figure 5.16: Overview of the model reduction method.

164 CHAPTER 5. NONLINEAR MODEL REDUCTION

Figure 5.17: Large conguration changes of the four-bar mechanism

91 dofs, whereas the reduced model involves 5 modal coordinates (1 rigid mode,
1 constraint mode, 3 internal modes), which are represented for conguration (d)
in Figure 5.18. Two internal modes have an out-of-plane deformation.

For this one-dimensional conguration space, the Lagrange interpolation

technique is selected and dierent versions of the reduction algorithm are tested.

First, the conguration space is decomposed into 32 boxes according to a

regular grid, and the mode tracking algorithm is disabled. The natural frequencies
of the selected internal modes are plotted in Figure 5.19. On the left plot, non-
smooth variations are observed around θ = 0.7 rad. They can be attributed to the
presence of additional low-order modes which interfere with the selected internal
modes. The algorithm selects the three modes with the lowest frequencies, so
that discontinuities are observed whenever an eigenvalue crossing occurs.

This situation is of course not acceptable, since it leads to a non-consistent

parameterization of the motion. The mode tracking strategy is able to remedy
this situation, as attested by the results on the right plot in Figure 5.19. The
selected modes are tracked in the conguration space, despite the presence of
other low-frequency modes.

For the rigid mode, Figure 5.20 illustrates the variations of the equivalent
mass and gyroscopic tensor in the conguration space. A vertical asymptote
is expected close to the extreme singular congurations. Mathematically, the
components of the rigid modes grow to innity at those points, which explains
the phenomenon. From the actuator point of view, the mechanical blocking at
the singularity is equivalent to an innite inertia. The relative error resulting
5.8. APPLICATIONS 165

Figure 5.18: Mode shapes for conguration (d). From top to bottom: rigid mode,
constraint mode (θ is xed), and 3 internal modes (θ and ze are xed).
166 CHAPTER 5. NONLINEAR MODEL REDUCTION

Without tracking With tracking

8 8
ω1 ω1
ω2 ω2
7 7
ω3 ω3

6 6
Frequencies (Hz)

Frequencies (Hz)
5 5

4 4

3 3

2 2

1 1

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
θ (rad) θ (rad)

Figure 5.19: Natural frequencies in the conguration space - importance of the

tracking strategy (obtained with a regular grid discretization).

from the approximation strategy is also presented:

fi (θ) − fbi (θ)

erri (θ) = (5.172)
kfi (θ)k
A remarkable accuracy is obtained away from the singularities: the error is around
0.1% for the mass, and around 0.5% for the gyroscopic tensor. In particular,
the errors vanish at every grid points. However, at the singularity, low-order
polynomials dened on a regular grid are not able to represent the sti behavior
of the system. It is also observed that the gyroscopic tensor is connected with
the gradient of the equivalent mass in the conguration space, in agreement with
well-known principles of mechanics.
The equivalent stiness associated with the constraint mode is analyzed in
Figure 5.21. The stiness takes very large values around conguration (c) repre-
sented in Figure 5.17. In this conguration, the two left links are aligned, and any
vertical motion ze induces traction or compression eorts, involving tremendous
strain energy. As a result of the large stiness of the constraint mode, the vertical
deection of the eector is almost blocked.
In order to achieve a better trade-o between the number of boxes and the
accuracy of the approximation, the adaptive strategy for the grid denition has
been applied. The tolerance for the relative error is set to tole = 0.003 and the
minimal box width tol1w = 0.03 rad. The comparison between the regular and
5.8. APPLICATIONS 167

700
2500

600 2000

1500

Gyr. tensor (kg.m2)

Inertia (kg.m 2)

500
1000

400 500

0
300
−500
200
−1000

100 −1500
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
θ (rad) θ (rad)

0.015

Rel error − gyr tensor

Rel error − inertia

0.01 0.1

0.005 0.05

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
θ (rad) θ (rad)

Figure 5.20: Regular grid with tracking strategy: variations of the equivalent
inertia associated with the rigid mode Mθθ (1, 1), and of the gyroscopic tensor
h(1, 1, 1).

6
x 10
3
9000

8000
2.5
7000
Stiffness (N/m)

Stiffness (N/m)

2 6000

5000
1.5
4000

1 3000

2000
0.5
1000

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
θ (rad) θ (rad)

Figure 5.21: Regular grid with tracking strategy: equivalent stiness associated
with the constraint mode. The picture on the right is a zoom.
168 CHAPTER 5. NONLINEAR MODEL REDUCTION

Regular grid

−2 −1.5 −1 −0.5 0 0.5 1 1.5 2

θ (rad)
Adaptive grid

−2 −1.5 −1 −0.5 0 0.5 1 1.5 2

θ (rad)

Figure 5.22: Regular and adaptive grid.

the adaptive grid is presented in Figure 5.22. The adaptive grid consists in 37
boxes (only 5 more than the regular grid), and it is rened close to the singu-
lar congurations, and close to conguration (c). The benets of this strategy
clearly appears in Figure 5.23, where inertia characteristics and relative errors
are compared. For the adaptive strategy, the accuracy is strongly increased near
the singularities, and small errors are tolerated in other parts of the conguration
space, allowing a coarser discretization. Figure 5.24 reveals the ability of this
approach to deal with the non-smooth behavior of the stiness of the constraint
mode around conguration (c). The regular grid leads to a bad approximation
in a large domain around the peak, which may strongly aect the quality of the
model.
Additional information can be recorded in the reduced model. For instance,
the modal contribution of the gravity forces ggrav and the amplitude of the rigid
modes for the constraint dof (Ψgθ ) are plotted in Figure 5.25.
From this example, we conclude that the reduced-order modeling technique
is able to capture accurately the nonlinear changes in the stiness and inertia
properties. The mode-tracking strategy is essential to guarantee the consistency
of the results. Close to the singular congurations, the adaptive conguration
space decomposition is valuable to reduce the approximation error and to op-
timize the computational resources. In order to demonstrate the generality of
5.8. APPLICATIONS 169

700
2500

600 2000

1500

Gyr. tensor (kg.m2)

Inertia (kg.m 2)

500
1000

400 500

0
300
−500
200
−1000

100 −1500
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
θ (rad) θ (rad)

0.015
regular regular
adaptive adaptive
Rel error − gyr tensor
Rel error − inertia

0.01 0.1

0.005 0.05

0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
θ (rad) θ (rad)

Figure 5.23: Comparison adaptive/regular grid: inertia components.

6 5
x 10 x 10
15 10
regular regular
adaptive 9 adaptive
exact exact
8

7
Stiffness (N/m)

Stiffness (N/m)

10
6

4
5
3

0 0
0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
θ (rad) θ (rad)

Figure 5.24: Comparison adaptive/regular grid: stiness of the constraint mode

for θ ∈ [0.6 1] rad.
170 CHAPTER 5. NONLINEAR MODEL REDUCTION

1000 6

800 4
Gravity torque (N.m)

600
2
400
0

Ψgθ
200
−2
0

−200 −4

−400
−6
−600
−8
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
θ (rad) θ (rad)

Figure 5.25: Additional information recorded in the model: equivalent gravity

force exerted on the rigid dof, and amplitude of the rigid mode for the constraint
dof Ψgθ (1, 1).

the approach, a mechanical system with a 3-dimensional conguration space is

considered in the next example.

5.8.2 The Orthoglide

The Orthoglide (Figure 5.26) is a parallel kinematic machine-tool designed at
the IRCCyN research center (Nantes, France) for high-speed machining [WC00,
CWA00, GKCW02, CWM02, CW03]. Three orthogonal linear actuators, xed
to the frame, control the 3-dimensional motion of the eector, and an isotropic
behavior is obtained in the center of the workspace, where the 3 links are parallel
to the inertial frame axes. The eector is connected to the actuators through a
set of three parallelogram mechanisms, so that its orientation is kept constant.
A simplied model is constructed, assuming rigid links, and ideal joints.
Since we are especially interested in the dynamic relation between the actuators
and the translation of the eector, the parallelogram mechanism is modeled as a
rigid link, connected to the actuator by a universal joint, and to the eector by
a spherical joint, see Figure 5.27. Hence, the initial model involves 27 dofs: 15
generalized coordinates, and 12 Lagrange multipliers. In this case, the objective of
the reduction is to formulate an equivalent model in term of the three independent
actuator coordinates. This problem could be equivalently solved according to an
extended constraint elimination method [BDG06]. The resulting inverse dynamic
5.8. APPLICATIONS 171

Figure 5.26: Prototype of the Orthoglide and conceptual kinematic representa-

tion. The length of the links is approximately 45 cm.

model is especially well-suited for the design of a computed-torque controller: the

model predicts the required actuator forces and torques to produce a pre-specied
motion.
Chablat et al. [CWM02] analyzed the workspace of the Orthoglide using
interval analysis concepts. Here, the motion of each actuator is limited to the
interval [-0.08, 0.26] m and the corresponding workspace does not contain any
singular conguration; two extreme congurations are illustrated in Figure 5.28.

The reduced mass matrix and gyroscopic tensor are constructed. The ap-
proximation in the conguration space is done with quadratic and Lagrange poly-
nomials. The tolerance on the relative error tole is xed to 0.5%, the minimal box
width toliw is xed to 0.05 m (i = 1, 2, 3). The discretization in the conguration
is illustrated in Figure 5.29. Quadratic polynomials require 22 boxes, whereas 8
boxes are sucient for Lagrange polynomials. Therefore, each component of the
reduced model requires the storage of 220 oating-point numbers for quadratic
polynomials, and of 216 numbers for Lagrange polynomials.
The variations of the inertia parameters along a straight line in the con-
172 CHAPTER 5. NONLINEAR MODEL REDUCTION

Figure 5.27: Model of the Orthoglide. The xation of the links on the actuator
slider is modeled as a universal joint to prevent their axial rotation.

guration space are presented in Figure 5.30. The quadratic polynomial is not
appropriate, since signicant discontinuities are observed. The results would be
improved using smaller tolerances tole and toliw , but it would lead to an exagger-
ated discretization of the conguration space. Using Lagrange polynomials, the
relative error is kept below 0.5% for the mass, and below 3% for the gyroscopic
tensor.
As a conclusion, the reduction method is applicable to a spatial mechanism
with 3 kinematic dofs. The nonlinear variations of the inertia properties are sig-
nicant but smooth, since the actuator singularities are avoided. Thus, piecewise
polynomials lead to an ecient representation of the dynamic model.

5.9 Summary and concluding remarks

This chapter proposed a modeling tool for the compact and closed-form repre-
sentation of complex exible multibody systems with parallel topology. Following
a model reduction approach, the method relies on the following assumptions:

- the system is conservative and holonomic, no pre-stressing eect is consid-

ered,

- the singularities are avoided in the conguration space,

- the elastic behavior is linear, the reduced stiness only depends on the
kinematic conguration θ, and centrifugal stiening eects are neglected,
5.9. SUMMARY AND CONCLUDING REMARKS 173

θ = [−0.08 − 0.08 − 0.08]T m :

θ = [0.26 0.26 0.26]T m :

Figure 5.28: Extreme congurations of the Orthoglide. In grey lines, the cong-
uration θ = [0 0 0]T m.
174 CHAPTER 5. NONLINEAR MODEL REDUCTION

0.3 0.3

0.25 0.25

0.2 0.2

0.15 0.15
θ2 (m)

θ (m)
0.1 0.1

2
0.05 0.05

0 0

−0.05 −0.05

−0.1 −0.1
−0.1 −0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 −0.1 −0.05 0 0.05 0.1 0.15 0.2 0.25 0.3
θ1 (m) θ1 (m)

Figure 5.29: Conguration space discretization: Quadratic and Lagrange poly-

nomials. Conguration space cut in the plane θ3 = −0.08 m.

5 16
Quadratic Quadratic
Lagrange 14 Lagrange
4.5
12
Gyr. tensor (kg.m2)
Inertia (kg.m 2)

4 10

8
3.5
6

3 4

2
2.5
0

2 −2
−0.1 −0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 −0.1 −0.05 0 0.05 0.1 0.15 0.2 0.25 0.3
θ1 (m) θ1 (m)

0.035 0.1
Quadratic Quadratic
Lagrange 0.09 Lagrange
0.03
0.08
Rel error − gyr tensor
Rel error − inertia

0.025 0.07

0.06
0.02
0.05
0.015
0.04

0.01 0.03

0.02
0.005
0.01

0 0
−0.1 −0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 −0.1 −0.05 0 0.05 0.1 0.15 0.2 0.25 0.3
θ1 (m) θ1 (m)

Figure 5.30: Variations of the inertia components from the extreme conguration
θ = [−0.08 − 0.08 0.26]T m to [0.26 0.26 − 0.08]T m. On the left,
the equivalent mass associated with the rst rigid mode M 11 , on the right the
θθ

component h111 .
5.9. SUMMARY AND CONCLUDING REMARKS 175

- the reduced parameterization is dened in an optimal way on the basis

of the linearized equations around any kinematic conguration with zero
velocities, i.e., it is optimized for quasi-static conguration changes.

The resulting model is described concisely according to a Global Modal Pa-

rameterization, which is free from kinematic constraints. The clear physical inter-
pretation of the modal coordinates guarantees the consistency of the formulation.
The reduced model is represented by a piecewise polynomial function, with the
advantages of computational eciency and portability. In the reduction proce-
dure, the accuracy loss is localized at three levels:

- the truncation of the modal basis (this source of error disappears when
considering a rigid mechanism),

- the approximation in the conguration space,

- the elimination of nonlinear phenomena associated with large deformations.

The procedure is systematic, and the user only needs to provide high-level
information:

- a standard Finite Element model of the mechanism,

- the partitioning into rigid, constraint, and internal dofs,

- the number of internal modes,

- an inner subpaving approximation of the conguration space,

- the tolerance on the approximation error, and the minimum size of the
boxes.

The mode shapes are deduced automatically, even if the mechanical topology is
complex and if the distribution of physical properties in the exible bodies is non-
uniform. This is a denite advantage over assumed-mode techniques, where the
appropriate denition of the modes and boundary conditions is often intricate.
In this chapter, two examples have been successfully analyzed: a exible
four-bar mechanism, and a rigid parallel kinematic machine-tool. In chapter 6,
the reduced-order model of a exible manipulator will be exploited for the design
of a control algorithm.
176 CHAPTER 5. NONLINEAR MODEL REDUCTION

The dofs appearing in the nal model are directly associated with the actua-
tors, the sensors, and other components of the mechatronic system. The method
can be exploited for control design in (at least) three dierent ways:

- online implementation of a feedback controller: computed torque control of

a rigid mechanism, active vibration control (see chapter 6),

- oine numerical optimization of a control law: trajectory generation, in-

verse dynamics, training of a feedback action,

- construction of a more structured model, in order to t well-established

control theories: polytopic linear model, model based on linear fractional
transformation, etc. Then, the reduction procedure makes the database
θ (i) − f (i) , i = 1, ..., r available for the external approximation algorithm,
and the conguration space decomposition can be interpreted as an adap-
tive inspection algorithm, which optimizes the information content of the
database.

A priori knowledge for improved conguration space inspection

One may object that the size of the database and the computational eort to
build the reduced-order model increase exponentially with respect to the dimen-
sion of the conguration space. At the end of this chapter, two direct extensions
of the original method are discussed to overcome this problem.
The reduction method has been presented assuming that the model should be
available in the whole conguration space. But, in many control applications, the
trajectory (or the desired trajectory) of the mechanism is known in advance. The
inspection strategy can be restricted to the congurations along the trajectory,
which is parameterized using the arclength, the time, or a pseudo-time variable.
The approximation then becomes a one-dimensional problem, which can be solved
with a high eciency. The resulting model is a linear time-varying model about
the reference trajectory, if the gyroscopic eects are neglected.
The approximation strategy can benet from an analysis of the congura-
tion space symmetries. Indeed, some properties of the dynamic model may be
invariant with respect to a group of transformations, so that the inspection can
be restricted to a lower-dimensional part of the conguration space. For instance,
5.9. SUMMARY AND CONCLUDING REMARKS 177

if a mechanism is mounted on a straight slider, the whole dynamic model is cer-

tainly not aected by a translation on the slider. Our algorithm would behave
nicely in such circumstances, since the interpolating polynomials are able to rep-
resent exactly the constant behavior along the translation coordinate, and no
bisection shall occur in this direction (provided slight customization). Likewise,
a mechanism mounted on a moving basis, e.g. a free-ying manipulator, has six
rigid dofs associated with the motion of the basis, but the dynamic model rep-
resented in local axes is not aected by the overall motion. In this case, if an
updated Lagrangian point of view is adopted for the orientation of the basis in
the reduced model, the construction of the model can be restricted to the relative
conguration changes with respect to the basis.
178 CHAPTER 5. NONLINEAR MODEL REDUCTION
 VI

Modeling, Simulation and Control of an

Experimental Manipulator

In this chapter, the theoretical concepts presented previously are applied to

the dynamic analysis and the control design of a lightweight manipulator called
RALF, an acronym for Robot Arm Large and Flexible.

Dierent models are elaborated and utilized for frequency domain analysis,
simulation and control design, as detailed in Table 6.1. Those models are dened
either for the mechanism, for the actuated mechanism or for the whole controlled
mechanism. An experimental validation is also conducted.
Guided by those models, a control strategy is developed and implemented for
the exible manipulator. According to a two-time-scale approach, a traditional
joint-tracking controller is complemented with a fast indirect vibration controller.
The vibration controller exploits the inertia forces of the manipulator in order to
damp the exible motion. Hence, it is an extension of inertial damping schemes
developed for the control of macro/micro-manipulators [Geo02].

After a description of the test-bed, the modeling of the actuated mecha-

nism is addressed in section 6.2, and an experimental validation is realized in
the frequency domain. In section 6.3, the composite two-time-scale controller
is designed. Section 6.4 deals with the performance analysis of the controlled
mechanism: the predictions of various models are compared with experimental
results.
180 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

Actuated Controlled
Mechanism
mechanism mechanism
Full-order
Time simulation
nonlinear
Reduced-order
Control design Time simulation
nonlinear
Reduced-order Root locus
Frequency response
linear Frequency response

Table 6.1: Dierent models and their use for dynamic analysis and control design.
The lines of the tabular are associated with the type of mechanical model, the
columns with the additional components of the mechatronic system.

6.1 Description of the test-bed

The manipulator Ralf, shown in Figure 6.1, is a long-reach manipulator oper-
ating in a vertical plane, that has been developed at the Georgia Institute of Tech-
nology. It was designed by Wilson [Wil86], and completed by Huggins [Hug88].
Ralf has two kinematic dofs, and it is hydraulically actuated.
The structure consists of two main links and a parallel actuation mechanism.
The two main links are 3.05 m long and constructed from aluminium pipes,
whereas the lighter actuation link is constructed from a rectangular aluminium
tube. Thick sleeves of aluminium tubing connect the links to each other, to the
actuators, and to the base. The joints are constructed from bronze bushings
and steel shafts. The assembled manipulator structure without actuators and
base weights approximately 45 kg, with its payload capacity of 27 kg. Hence,
Ralf has a high payload to weight ratio, and it is sti enough to achieve real-
world applications. However, exible eects in the aluminimum tubes aect the
positioning accuracy.
Two hydraulic cylinders are used for the actuation of Ralf. The rst moves
the rst link relative to the base, the second moves the second link relative to the
rst through the parallel mechanism. Each cylinder is controlled by a hydraulic
servovalve. The oil is supplied at 1500 psi. The maximum cylinder velocities are
0.156 m/s for extension and 0.208 m/s for retraction.
Two linear position sensing transducers, xed to the hydraulic cylinder, mea-
6.1. DESCRIPTION OF THE TEST-BED 181

Figure 6.1: Ralf, coordinates of the actuators and the sensors.

182 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

120 120

110 110

100 100

90 90
α1 (deg)

α2 (deg)
80 80

70 70

60 60

50 50

40 40

30 30
0.7 0.8 0.9 1 1.1 1.2 1.3 0.7 0.8 0.9 1 1.1 1.2 1.3
θ1 (m) θ2 (m)

Figure 6.2: Relation between actuator coordinates θi and link angles αi , i = 1, 2.

The circle represents the home conguration.

sure the cylinder extension. Moreover, in order to detect the vibrations of the
mechanism, two accelerometers have been placed at the tip, in two orthogonal
directions.
The control law is implemented on a real-time computer, equipped with a
National Instruments A/D board. The control law is developed in the Simulink
environment, built using Real-Time workshop, and loaded on the real-time com-
puter thanks to the XPC Target architecture.
Ampliers and power supplies are in charge of generating the electrical signals
for the hydraulic valves, as well as to amplify the accelerometer signals.

6.1.1 Workspace description

As seen in Figure 6.1, the kinematic conguration of Ralf can be described
either using the relative angles α1 and α2 of the main links, or the actuator
extensions θ1 and θ2 . The relation between those coordinates is illustrated in
Figure 6.2. The tip-accelerometers measure the absolute accelerations aX and aY
in body axes.
The allowed conguration space is dened by a rectangular box represented
in Figure 6.3. In the same gure, 6 congurations of interest are mentioned;
conguration 0 is the so-called home position.
6.1. DESCRIPTION OF THE TEST-BED 183

Figure 6.3: Conguration space of Ralf.

Ralf
v - - wm
Actuators + Mechanism

Figure 6.4: Ralf - input/output representation.

6.1.2 Transfer functions acquisition

In the experimental system, the inputs are the voltages of the actuators, and
the outputs are the sensor measurements, i.e. the extensions of the hydraulic
actuators and the tip accelerations. We dene the voltage vector v and the
measurement vector wm :
 
θ1m
" #  m  " m #
v1  θ2
= θ (6.1)
m

v= , w =
 am
v2  X

 am
am
Y

An input/output view of the actuated mechanism is illustrated in Figure 6.4.

In order to establish the linearized transfer functions of the mechanism
around dierent operating points, the actuators are independently controlled by
low gain proportional controllers, and a swept-sine signal excites the system, as il-
lustrated in Figure 6.5. The resulting information shall be used for the validation
184 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

θiexcitation

+ vi θim
+ ?
j - Gs
θiref -
i
- Ralf -
−6

Figure 6.5: Independent SISO controller for each actuator (i = 1, 2). Low gain
values G1 = G2 = 100.

of the mechanical model.

6.2 Model of the manipulator

6.2.1 Previous models of Ralf
Huggins [Hug88] and Lee [Lee90] achieved an experimental validation of an
assumed-mode model based on Lagrangian dynamics. They considered a two-link
model, with two vibration modes per beam. The denition of appropriate bound-
ary conditions for each link turned out to be a dicult task. In order to verify the
quality of the assumed-mode model, they developed and validated a linear Finite
Element model. On this basis, improved shapes of the assumed-modes were ob-
tained by component-mode synthesis. Despite small discrepancies observed with
respect to experimental data, the assumed-mode model was suciently accurate
for control design. The dynamic behavior of the actuators was investigated, it
appeared to play a signicant role.
Lee [Lee90] developed a Lagrangian dynamic model including the parallel
drive mechanism, represented by nonlinear constraint equations. For simulation
purpose, a constraint elimination technique based on a singular value decom-
position was implemented. Through experimental work, Lee observed that the
importance of the inertia forces quadratic in velocity is minor, due to the speed
limitation of the hydraulic cylinder.
Magee [Mag91] achieved an experimental analysis of Ralf in the whole workspace.
The structure was excited by the hydraulic actuators, and the tip motion was mea-
sured using accelerometers. The underlying assumed-mode model was based on
6.2. MODEL OF THE MANIPULATOR 185

the Lagrangian method formulated by Book [Boo84]. In order to develop an in-

put shaping strategy, a curve tting method was applied to get the experimental
values of the eigenfrequencies and damping ratio in the conguration space.
The design of Ralf evolved over the years, which explains substantial dif-
ferences in the results obtained by those researchers. More recent investigations
were realized by Obergfell [Obe98], who modeled Ralf as a serial two-link manip-
ulator. Surprisingly low natural frequencies were observed, that could be hardly
represented by the mechanical model. Obergfell concluded that the actuator dy-
namics should be responsible for those discrepancies. Introducing a more realistic
model of the hydraulic actuators, signicant improvements were obtained, even
though the natural frequencies predicted by the model were still 20 % too high.

6.2.2 Structural model

Following the formalism presented in chapter 3, a nonlinear Finite Element
model has been elaborated for Ralf. Compared with previous models, this ap-
proach allows to represent all the details of the mechanism, including the parallel
actuation mechanism. The two main links and the actuator link are modeled us-
ing exible beam elements, all other components are considered as rigid-bodies or
lumped masses. For the details about the geometrical data and the computations
of the dierent masses, we refer to the drawings in [Wil86] and to the numerous
reports on Ralf, such as [Obe98]. The gravity forces are taken into account.
Various other components are present on Ralf, and their masses have been
estimated and introduced in the model. The structural damping has been ad-
justed to a value around 1 % (actually, the dominant damping eect in the overall
system comes from the actuators and the control system). The Finite Element
model involves 164 dofs.
Around a xed conguration, the dynamics of Ralf is reasonably represented
by a linearized model, and a frequency-domain analysis is possible. The linearized
model is reduced according to the standard component-mode technique: 2 rigid
modes and 2 exible modes are able to capture the essential dynamic behavior
of the system in the frequency range 0-15 Hz . Higher order modes are above 30
Hz , and their participation to the dynamic response is neglected.
The experimental validation of this model relies on transfer functions from
186 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

Input: v ; Output: θ Input: v ; Output: θ

1 1 2 2
−40 −40
Mag Ratio (dB)

Mag Ratio (dB)

−50 −50

−60 −60

−70 −70

−80 0 1
−80 0 1
10 10 10 10

−80 −80

−100
Phase (deg)

Phase (deg)
−100
−120
−120
−140

Model −140 Model

−160
Experiment Experiment
−180 0 1
−160 0 1
10 10 10 10
Freq (Hz) Freq (Hz)
Input: v1 ; Output: aY Input: v2 ; Output: aY
60 60
Mag Ratio (dB)

40 Mag Ratio (dB) 40

20 20

0 0

−20 0 1
−20 0 1
10 10 10 10

600 0
Phase (deg)

Phase (deg)

400
−100
200
−200
0 Model Model
Experiment Experiment
−200 0 1
−300 0 1
10 10 10 10
Freq (Hz) Freq (Hz)

Figure 6.6: First actuator model (home conguration).

input voltages to linear sensors, and to accelerometers established for dierent

congurations. A model of the hydraulic actuators is also needed at this level.

6.2.3 Validation and actuators models

First actuator model: velocity source

The experimental transfer functions from voltages to actuator displacements

are plotted Figure 6.6. A velocity source model of the actuator is also considered,
which behaves as an integrator:

θ̇i = Ghydr
i vi , i = 1, 2 (6.2)
6.2. MODEL OF THE MANIPULATOR 187

where θi is the displacement of the actuator i, vi is the applied voltage, and Ghydr
i
is the hydraulic gain, whose value has been tted on the experimental data:

Ghydr
1 = 0.0236, Ghydr
2 = 0.0265 (6.3)

This ideal model does not adequately capture the dynamics of the physical
system in two ways:

- the actuator does not behave as a velocity source near the natural frequen-
cies of the system,

- the predicted frequency responses from voltages to tip accelerations, also

plotted in Figure 6.6, are not satisfactory.

For these reasons, and according to the idea of Obergfell [Obe98], a more elabo-
rated model of the actuator is investigated, accounting for the limited dynamics
of an intrinsic velocity feedback.

Second actuator model: intrinsic velocity feedback

In order to account for the intrinsic velocity feedback that regulates the
ow through the valve of a hydraulic actuator, the following actuator model is
proposed:
gia = Gvi (Ghydr
i vi − θ̇i ), i = 1, 2 (6.4)
where gia is the force applied by the actuator, and Gvi is a new constant to be
determined. If Gvi → ∞, this model is equivalent to the velocity source model.
For nite Gvi , the transfer function between vi and θi is aected by the dynamics
of the mechanical system, especially near the resonances.
The value of Gvi inuences the transfer functions from voltages to actua-
tor displacements, but unfortunately, the transfer function from voltages to tip
acceleration is not improved, as shown in Figure 6.7, with the values:

Gv1 = 3.5e5, Gv2 = 3.0e5 (6.5)

In particular, the predicted transfer functions are shifted to the high frequencies,
which means that the model is probably too sti. Therefore, let us consider the
eect of a lumped exibility at the level of the actuators.
188 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

Input: v ; Output: θ Input: v ; Output: θ

1 1 2 2
−40 −40
Mag Ratio (dB)

Mag Ratio (dB)

−50
−50
−60
−60
−70
−70
−80

−90 0 1
−80 0 1
10 10 10 10

0 −60

−80
Phase (deg)

Phase (deg)
−50
−100
−100
−120
−150 Model Model
−140
Experiment Experiment
−200 0 1
−160 0 1
10 10 10 10
Freq (Hz) Freq (Hz)
Input: v1 ; Output: aY Input: v2 ; Output: aY
60 40
Mag Ratio (dB)

Mag Ratio (dB)

40
20
20
0
0

−20 0 1
−20 0 1
10 10 10 10

300 0

200
Phase (deg)

Phase (deg)

−100
100

0
−200
−100 Model Model
Experiment Experiment
−200 0 1
−300 0 1
10 10 10 10
Freq (Hz) Freq (Hz)

Figure 6.7: Second actuator model (home conguration).

Figure 6.8: Actuator model including a lumped exibility. gia denotes the gen-
erated force, kia is the nite stiness of the actuator, θi and Li the actuated dof
and the total length.
6.2. MODEL OF THE MANIPULATOR 189

Third actuator model: intrinsic velocity feedback and exibility

In order to obtain a better matching between the model and the experiments,
lumped exible eects are considered in the actuators, see Figure 6.8. As in the
previous section, the actuator forces are modeled according to equation (6.4), but
the length Li between the attachment points of the actuator is now dierent from
the actuated dof θi .
It is not clear whether the measured displacement θim corresponds to Li (non-
collocated sensor/actuator conguration) or to θi (collocated sensor/actuator con-
guration). Hence, we follow a two-step identication procedure: (i) select the
stinesses kia in order to improve the transfer function vi - a, (ii) compare the
experimental transfer functions vi - θim with the model transfer functions vi - θi
and vi -Li .
(i) The local stinesses are easily adjusted by the observation that actuator
1 has a dominant eect on the rst mode, whereas actuator 2 essentially excites
the second one. An iterative procedure results in the selection of the following
stiness values:
k1a = 3.5e6 N/m2 , k2a = 3.0e6 N/m2 (6.6)
For the set of congurations mentioned in Figure 6.3, the model and experimental
transfer functions are plotted in Figure 6.9, 6.10 and 6.11. The correspondence is
noteworthy, despite strong variations from one conguration to the other. In some
congurations, the predicted frequency response are still slightly shifted to the
high frequencies, but the model captures the dynamics with sucient accuracy
for control design.
(ii) The actuator transfer functions are plotted for the collocated and non-
collocated cases in Figure 6.12. It seems that the model with collocated mea-
surements is closer to the experimental results; it is selected for the following
developments.

6.2.4 Nonlinear model reduction

The model of the actuated mechanism can be splitted into one model for the
actuator and one model for the mechanical system, as illustrated in Figure 6.13,
where the vector of actuator forces ga = [g1a g2a ]T is computed according to for-
mula (6.4). The actuator exibility is included in the mechanical model.
190 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

Conguration 0 Conguration 1

Input: v ; Output: a Input: v ; Output: a

1 X 1 X
40 40
Mag Ratio (dB)

Mag Ratio (dB)

30
20
20

10
0
0

−20 0 1
−10 0 1
10 10 10 10

200 200
Phase (deg)

Phase (deg)
100 100

0 0

−100 Model −100 Model

Experiment Experiment
−200 0 1
−200 0 1
10 10 10 10
Freq (Hz) Freq (Hz)

Input: v1 ; Output: aY Input: v1 ; Output: aY

40 40
Mag Ratio (dB)

Mag Ratio (dB)

20 20

0 0

−20 0 1
−20 0 1
10 10 10 10

200 200
Phase (deg)

Phase (deg)

100 100

0 0

−100 Model −100 Model

Experiment Experiment
−200 0 1
−200 0 1
10 10 10 10
Freq (Hz) Freq (Hz)

Input: v ; Output: a Input: v ; Output: a

2 Y 2 Y
60 60
Mag Ratio (dB)

Mag Ratio (dB)

40 40

20 20

0 0

−20 0 1
−20 0 1
10 10 10 10

0 0
Phase (deg)

Phase (deg)

−100 −100

−200 −200
Model Model
Experiment Experiment
−300 0 1
−300 0 1
10 10 10 10
Freq (Hz) Freq (Hz)

Figure 6.9: Experimental and model transfer functions, congurations 0 (home)

and 1.
6.2. MODEL OF THE MANIPULATOR 191

Conguration 2 Conguration 3

Input: v ; Output: a Input: v ; Output: a

1 X 1 X
30 40
Mag Ratio (dB)

Mag Ratio (dB)

30
20
20
10
10
0
0

−10 0 1
−10 0 1
10 10 10 10

200 200
Phase (deg)

Phase (deg)
100 100

0 0

−100 Model −100 Model

Experiment Experiment
−200 0 1
−200 0 1
10 10 10 10
Freq (Hz) Freq (Hz)

Input: v1 ; Output: aY Input: v1 ; Output: aY

40 60
Mag Ratio (dB)

Mag Ratio (dB)

20 40

0 20

−20 0

−40 0 1
−20 0 1
10 10 10 10

200 300

200
Phase (deg)

Phase (deg)

100
100
0
0
−100 Model Model
−100
Experiment Experiment
−200 0 1
−200 0 1
10 10 10 10
Freq (Hz) Freq (Hz)

Input: v2 ; Output: aY Input: v2 ; Output: aY

60 60
Mag Ratio (dB)

Mag Ratio (dB)

40 40

20 20

0 0

−20 0 1
−20 0 1
10 10 10 10

0 −50

−100
Phase (deg)

Phase (deg)

−100
−150

−200
−200
Model −250 Model
Experiment Experiment
−300 0 1
−300 0 1
10 10 10 10
Freq (Hz) Freq (Hz)

Figure 6.10: Experimental and model transfer functions, congurations 2 and 3.

192 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

Conguration 4 Conguration 5

Input: v ; Output: a Input: v ; Output: a

1 X 1 X
30 25
Mag Ratio (dB)

Mag Ratio (dB)

20
20
15

10
10
5

0 0 1
0 0 1
10 10 10 10

200 200
Phase (deg)

Phase (deg)
100 100

0 0

−100 Model −100 Model

Experiment Experiment
−200 0 1
−200 0 1
10 10 10 10
Freq (Hz) Freq (Hz)

Input: v1 ; Output: aY Input: v1 ; Output: aY

60 40
Mag Ratio (dB)

Mag Ratio (dB)

40 20

20 0

0 −20

−20 0 1
−40 0 1
10 10 10 10

200 200
Phase (deg)

Phase (deg)

100 100

0 0

−100 Model −100 Model

Experiment Experiment
−200 0 1
−200 0 1
10 10 10 10
Freq (Hz) Freq (Hz)

Input: v2 ; Output: aY Input: v2 ; Output: aY

60 60
Mag Ratio (dB)

Mag Ratio (dB)

40 40

20 20

0 0

−20 0 1
−20 0 1
10 10 10 10

100 −50

−100
Phase (deg)

Phase (deg)

0
−150
−100
−200
−200 Model Model
−250
Experiment Experiment
−300 0 1
−300 0 1
10 10 10 10
Freq (Hz) Freq (Hz)

Figure 6.11: Experimental and model transfer functions, congurations 4 and 5.

6.2. MODEL OF THE MANIPULATOR 193

Collocated sensors/actuators (θim = θi ) Noncollocated sensors/actuators (θim = Li )

m m
Input: v1 ; Output: θ1 Input: v1 ; Output: θ1
−40 −40
Mag Ratio (dB)

Mag Ratio (dB)

−60
−60
−80
−80
−100

−100 0 1
−120 0 1
10 10 10 10

0 −50

−100
Phase (deg)

Phase (deg)
−50
−150
−100
−200
−150 Model Model
−250
Experiment Experiment
−200 0 1
−300 0 1
10 10 10 10
Freq (Hz) Freq (Hz)
m m
Input: v2 ; Output: θ2 Input: v2 ; Output: θ2
−40 −40
Mag Ratio (dB)

Mag Ratio (dB)

−50
−60
−60
−80
−70

−80 0 1
−100 0 1
10 10 10 10

−60 −50

−80
Phase (deg)

Phase (deg)

−100
−100
−150
−120

Model −200 Model

−140
Experiment Experiment
−160 0 1
−250 0 1
10 10 10 10
Freq (Hz) Freq (Hz)

Figure 6.12: Third actuator model (home conguration).

ga
v - Actuators - Mechanism - wm

Figure 6.13: Mechanical and actuator model. The model of the mechanism in-
cludes the lumped stinesses at the level of the actuators.
194 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

In this section, a reduced mechanical model, valid in the whole conguration

space is constructed according to the method presented in chapter 5. The two
parameters of the rigid motion are naturally the actuated dofs θ = [θ1 θ2 ]T .
Two internal modes, with coordinates η δ = [η1δ η2δ ] are sucient to represent the
dynamic behavior in the bandwidth of interest (0-15 Hz ). The rigid and exible
mode shapes are illustrated in Figure 6.14.
The gyroscopic forces are neglected, which is a reasonable assumption ac-
cording to Lee [Lee90]. Therefore, the reduced equations of motion are:
" # " # " # " # " #
Mθθ Mθδ θ̈ 0 0 θ ga
+ = (6.7)
Mδθ I η̈ δ 0 Ω2 ηδ 0

Ω2 = diag(ωi2 ) is the diagonal matrix of the square eigenvalues. The mass matrix
and Ω2 depend nonlinearly on the conguration. This dependence is approxi-
mated by piecewise low-order polynomials. The conguration space is decom-
posed into boxes where the local polynomials are dened, in order to achieve
a relative error below the tolerance tole = 0.001. The set of boxes are repre-
sented in Figure 6.15, and the variations of the model parameters are illustrated
in Figure 6.16. The dierent orders of magnitude of the equivalent masses come
from the dierent normalization of the rigid and exible modes, as discussed in
Figure 6.14.
The relation between the end eector position xe and the modal coordinates
is:
xe = ρe (θ) + Ψeδ (θ) η δ (6.8)

where ρe is the nonlinear kinematic transformation of the rigid mechanism, and

Ψeδ is the 2 × 2 matrix of the exible modes at the end eector. At the velocity
level, we have
" #
h i θ̇
ẋe = Ψeθ Ψeδ δ
= Ψeη (θ) η̇ (6.9)
η̇

Ψeθ is the 2 × 2 matrix of rigid modes at the eector. In this equation, the
eη
nonlinear contribution Ψ̇ (θ) η δ has been neglected under the assumption of
small deformations.
6.2. MODEL OF THE MANIPULATOR 195

Rigid modes Flexible modes

Figure 6.14: Mode shapes, home conguration. The amplitudes of the modes
are δθ1 = 0.05 and δθ2 = 0.08, η1δ = 2.8, and η2δ = 1.45. Those dierent orders
of magnitude come from the dierent normalization of the modal cordinates:
the rigid modes are associated with unitary actuator displacements, whereas the
exible modes are normalized with respect to the mass matrix.
196 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

0.3

0.2

0.1
θ2 (m)

−0.1

−0.2

−0.2 −0.1 0 0.1 0.2 0.3

θ1 (m)

Figure 6.15: Conguration space discretization.

6.3 Control design

Most control algorithms tested on Ralf in the past were relying on a two-link
assumed-mode model. Our modeling approach is more systematic and general,
and it allows to account for the exact topology of the mechanism. In this section,
the reduced-order model is exploited for the elaboration of a control strategy.
First, let us summarize previous researches on the control of Ralf. For a more
general review of the literature on the control of exible mechanisms, we refer to
section 2.5. Lee [Lee90] complemented a PID joint controller with a decentral-
ized strain feedback, in order to damp the vibration modes and to improve the
performances of Ralf. Obergfell [Obe98] exploited the concept of output rede-
nition, reinterpreted in terms of combined positive deection feedback with PD
joint control. He developed an independent controller for each link/joint of the
serial manipulator. The deections were measured by optical sensors. In order
to compensate for static deections, an end-point position feedback was added as
an external loop on the rst controller. The absolute position was measured with
a landmark tracking system, using a machine vision system. Even though good
performances were obtained by Obergfell, the sensors are rather sophisticated,
and their implementation imposes restrictive operational and design constraints.
Alternatively, we consider an accelerometer feedback, whose hardware im-
6.3. CONTROL DESIGN 197

6.5

5.5

ω (Hz)
5

4.5

3.5
ω1
3
ω2
2.5
0.7 0.8 0.9 1 1.1 1.2 1.3 1.4
θ1 (m)

9000 90

8000
80
7000
70
6000
M11

M13

5000 60

4000
50
3000
40
2000

1000 30
0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4
θ1 (m) θ1 (m)

Figure 6.16: Variations of the model parameters along a diagonal line in the
conguration space from θ (lower left corner) to θ (upper right corner). The
variations of the natural frequencies are in the interval [3, 7] Hz . We recall
that the modal masses of the exible modes equals 1, as a consequence of their
normalization.

θm
θ ref +
- h - vs - hv - Mechanism
+
Slow ctrl
−6 +
am
6

vf

Fast ctrl 

Figure 6.17: Two-time-scale control.

198 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

plementation is fairly simple and convenient. In particular, we investigate a

composite two-time-scale control strategy (see Figure 6.17), which involves:
- a slow controller based on a feedback of the collocated measurements, re-
sponsible for the trajectory tracking,

- a fast controller, which relies on both collocated measurements and tip

acceleration measurements, in order to increase the damping in the exible
modes.
The time-scale separation is a necessary condition to avoid interferences between
the two controllers; in our case, we shall see that additional lters need to be
implemented in order to fulll this requirement.
In the literature, numerous authors applied two-time-scale control schemes
for exible mechanisms. Singular perturbation theory formulates a dynamic
model in terms of a slow and a fast subsystem, which naturally conducts to
the design of a composite controller [SB88, MC95, GLL98, MPK00, GS00, SV01,
CYC02]. The HAC/LAC strategy [Pre97, VBS02, Sym04] is another pragmatic
dual approach. The Low-Authority Controller (LAC) refers to the active damp-
ing, which is eective near the resonances, and which modies the poles of the
system only slightly. The High-Authority Controller (HAC) refers to the motion
controller which dominates the low-frequency dynamics.
Our contribution especially focuses on the fast active damping controller;
before a more detailed investigation, let us briey describe the design of the slow
controller.

6.3.1 Slow controller

In a serial robot, the actuated dofs are located at the joints, and roboticists
are familiar with decentralized control schemes : each joint is controlled indepen-
dently, using collocated feedback of the joint angle or position. This technique
is opposed to the tip-tracking strategy, where one feeds back measurements or
estimates of the tip position.
The slow controller presented here is a conventional joint-tracking controller,
illustrated in Figure 6.18:
 
vis = Gsi θiref − θim , i = 1, 2 (6.10)
6.3. CONTROL DESIGN 199

+ j - Gs vi θim
θiref -
i
- Ralf -
−6

Figure 6.18: Joint-tracking controller for actuator i (i = 1, 2).

The gains Gsi have to satisfy a double requirement:

- the joint-tracking control should be stable and performant,

- the bandwidth of the slow controller is limited by the natural frequencies

of vibration, and it should be separated from the bandwidth of the fast
controller.
At home conguration, the experimental transfer functions of the open-loop
system are plotted in Figure 6.12. In order to guarantee sucient phase margin
near the resonance, the value of the gain is limited to 52 dB for both joints, so
that
Gs1 = Gs2 = 400 (6.11)
If one accepts the velocity source actuator model, the open-loop system is an
integrator with gain Ghydr
i , and the slow controller leads to the closed-loop band-
width:
ωis = Ghydr
i Gsi
⇒ ω1s = 9.4 rad/s (1.5 Hz), (6.12)
⇒ ω2s = 10.6 rad/s (1.7 Hz).
A parametric analysis in the conguration space has shown that those gain values
lead to a stable behavior for any conguration of the mechanism. This has been
conrmed experimentally.
When applying a conventional joint motion control which ignores exibility,
a rule-of-thumb is to restrict its bandwidth below the half [VBSN+ 01] or the
third [Boo93] of the lowest natural frequency ω1 of the mechanical system. In our
case, ω1 is about 4 Hz (see Figure 6.16), and the ratio ω s /ω1 is about 1/3. As
we shall see, this ratio is suciently low to limit the interferences with the fast
controller.
200 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

Inverse




θm
Mode Inertial
-
cδ
η̇

des
θ̈ vf -
actuator

-
-
- observer

damping


am

model

Figure 6.19: Fast controller. The arrows indicate a model-based scheduling ac-
cording to the conguration changes.

6.3.2 Fast controller

The dynamics of the mechanism is conguration dependent. However, when
designing the fast controller, the time-scale separation assumption allows to con-
sider the nonlinear conguration changes as quasi-static variations. Concepts
from linear system dynamics can thus be exploited to build a linear controller,
whose parameters may be conguration dependent.
Thanks to the reduction procedure, the dynamic model of the mechanism (6.7)
is in linear parameter-varying format, and it can be directly exploited for the
design of such a fast controller. In this model, the action of the actuator forces
ga on the exible modes comes indirectly from the inertial coupling. Therefore,
in order to control the exible modes, our idea is to excite the rigid modes so
that the inertia forces produce a damping eect. This approach is equivalent to
the inertial damping method developed for macro/micro-manipulators [Geo02].
The fast controller is composed of three stages illustrated in Figure 6.19:

- the mode observer extracts the rates of the exible modes η̇

cδ from the

measurements,

- the inertial damping control law denes the desired accelerations of the rigid
des
modes θ̈ that would produce appropriate inertial forces,

- the inverse actuator model establishes the voltages vf to be applied to the

actuators.

In the following, the inertial damping control law is rst presented; the ob-
server and the inverse actuator model will be described afterwards.
6.3. CONTROL DESIGN 201

Inertial damping
In order to analyze the possible damping eect of the coupling inertia forces,
the dynamic equation associated with the exible modes is extracted from the
reduced-order model (6.7):

η̈ δ + Ω2 η δ = −Mδθ θ̈ (6.13)

The inertial damping consists in controlling the rigid variables θ in order to

produce a velocity feedback. Indeed, if θ̈ satises:

− Mδθ θ̈ = −Gf η̇ δ (6.14)

the closed-loop equation of motion becomes:

η̈ δ + Gf η̇ δ + Ω2 η δ = 0 (6.15)

and Gf η̇ δ has the clear interpretation of a damping term, provided that the gain
matrix Gf is positive denite. The stability of the closed-loop system is thereby
guaranteed. It is convenient to dene the gain matrix with respect to the desired
modal damping ratios ξi :
Gf = diag(2 ξi ωi ) (6.16)
As will be motivated in the root locus analysis (section 6.4.2), our design is based
on the choice
ξ1 = ξ2 = 0.1 (6.17)
In order to generate this damping eect, the subsequent problem is to control
the rigid modes so that equation (6.14) is satised. If the number of rigid modes
equals the number of exible modes, and if the inertia coupling matrix Mδθ is
not singular, this equation can be inverted, leading to the denition of the desired
des
acceleration of the rigid modes θ̈ :
des
−1
θ̈ = Mδθ Gf η̇ δ (6.18)

−1
The matrix Mδθ depends on the conguration. Recorded in the reduced-
order model, it is represented by piecewise polynomial functions, which can be

−1
exported from the modeling environment to the real-time controller. Mδθ is
thus scheduled online, according to the slow conguration changes θ detected
m

by the linear sensors.

202 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

6
x 10
3.5
θ2 = −0.206 m
θ2 = −0.0807 m
3
θ2 = 0.0445 m
θ2 = 0.17 m
2.5

Perf. Index
2

1.5

0.5

0
−0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4
θ1 (m)

Figure 6.20: Performance index of the inertial damping control.

If the number of exible modes to be controlled is lower than the number

of rigid modes, this strategy is still applicable using the pseudo-inverse of the
rectangular matrix Mδθ .
It is necessary to check that no mechanical conguration is associated with
an inertial singularity, i.e. a singularity of the coupling mass matrix Mδθ . At
an inertial singularity, it may be impossible to generate the required inertial
forces. For this purpose, George [Geo02] proposed to analyze the variations of a
performance index in the conguration space:
 
Perf. Index = det Mδθ Mδθ (6.19)
T

This denition is still applicable if the coupling matrix Mδθ is not square. Since
the determinant is the product of the singular values, the higher the performance
index, the farther from inertial singularities. In Figure 6.20, we observe that no
singularity is present in the conguration space of Ralf.

Mode observer
des
In the inertial damping control law (6.18), θ̈ is computed from the modal
velocities η̇ δ , which are not directly available. The observer aims at reconstructing
η̇ δ from the measurements θ m and am . First, let us characterize the relation
between the accelerometer signals and the modal coordinates.
The accelerometers measure the absolute accelerations in body axes, and
those axes rotate during the motion. If R denotes the rotation matrix of the end
6.3. CONTROL DESIGN 203

eector node, and xe its position referred to an inertial frame, we have:

ẍe = R am (6.20)

Under the assumption of small deformations, R only depends on the conguration

θ , and not on the exible modes. Thus, in the fast time-scale, the variations of
R are quasi-static and the integration of (6.20) leads to
Z t Z t
e
ẋ = m
R a dτ ' R am dτ (6.21)
0 0

Dening the integrated accelerometer signal cma

Z t
c ma
= am dτ (6.22)
0

we obtain
ẋe ' R cma (6.23)

Inverting this expression, and introducing the modal amplitudes of the end eec-
tor displacements ẋe = Ψeη η̇ , we obtain:

cma = Ψaη η̇ with Ψaη (θ) = RT (θ) Ψeη (θ) (6.24)

The relation between the measurements and the modal coordinates is sum-
marized by:
" m # " # " #
θ̇ I 0 θ̇
= (6.25)
cma Ψaθ Ψaδ η̇ δ

Since there are as many sensors as modal coordinates, this expression can be
inverted, leading to a static estimation formula for the modal velocities:
" m #
h
i θ̇
cδ = aδ −1 aδ −1 (6.26)


η̇ − Ψ Ψaθ Ψ
| {z } cma
=Ψδm

The estimation algorithm is illustrated in Figure 6.21, and it requires the

2 × 4 modal matrix Ψδm to be available online. The importance of the lters
shall be discussed later on.
204 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

m 
θ̇ -
θm Filter d
- - 
dt 

Ψδm - cδ

 η̇
Rt cma 
am - Filter -
0 dτ - 



Figure 6.21: Observer. The signals of the sensors are rst ltered, then numer-
ically dierentiated/integrated with respect to time, and nally, equation (6.26)
is used to estimate the modal velocity.

Inverse actuator model

The velocity source actuator model (6.2) can be inverted, leading to the
denition of the applied voltage:
Z t
1
vif = θ̈ides dτ, i = 1, 2 (6.27)
Ghydr
i 0

In order to avoid any drift in the numerical integration, the integrator is imple-
mented as a high-pass integrator, i.e. an integrator in series with a high-pass
lter.
The construction of the inverse model from the rst actuator model has been
motivated by the direct connection that it denes between the input voltage vi and
the actuator motion θi . On the contrary, the second actuator model involves the
actuator force, so that the input/output relation between vi and θi is aected by
the whole dynamics, in particular, by the mechanical resonances. The inversion
of this model would be more complicated and sensitive to modeling error.
One may be afraid that the velocity source model is not satisfactory near the
resonances, where the fast controller is expected to be performant. However, the
root locus analysis presented in section 6.4.2 will assess the stabilization eect of
the resulting control law.

6.3.3 Summary and rst results

In essence, the inertial damping strategy dened by (6.18) is conicting with
the slow controller which tries to achieve θ = θref . This paradox is resolved by
6.3. CONTROL DESIGN 205

Figure 6.22: Signal-ow graph of the controlled mechanism: slow and fast control
loops.

imposing that the inertial damping is only active at high-frequencies, near the
resonances of the exible modes η δ .
In order to clarify this point, let us summarize the interactions of both con-
trollers with the exible mechanism. In Figure 6.22, the signal-ow graph illus-
trates the dynamics of the controlled mechanism; in particular, the slow and fast
control loops are clearly represented. Let us briey comment the graph. In the
second actuator model (6.4), the actuator forces ga are inuenced by the applied
voltage v and by the extension rate θ̇. According to the mechanical model (6.7),
these forces excite the rigid dynamics described by the actuator coordinates θ,
which is coupled with the exible dynamics described by the modal amplitudes η δ .
The joint-tracking controller is based on collocated measurements θm , whereas
the fast controller feeds back an estimation of the exible modal rates η̇cδ .

The interaction between both control loops is limited under the time-scale
separation assumption. Indeed, the closed-loop bandwidth of slow controller is
limited to ω s ' 1/3 ω1 , whereas the fast controller is a feedback of the exi-
ble modal coordinates, which are mainly excited near the resonances. At rst
sight, the time-scale separation seems to be satised, and good performances are
expected from the composite controller.
However, we argue that various disturbances may be responsible for unde-
206 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

Simulation results
Fast control o Fast control on / η̇ δ = η̇ δ Fast control on / η̇ δ ' η̇ δ
c c
1.2
1.08 1.08
1.06 1.06 1.15
1.04 1.04
1.1

θ2 (m)
θ2 (m)

θ2 (m)
1.02 1.02
1 1 1.05
0.98 0.98
1
0.96 0.96
response response response
0.94 reference 0.94 reference 0.95 reference
0 1 2 3 4 0 1 2 3 4 0 1 2 3 4
t (s) t (s) t (s)
5 5 5
aY (m/s2)

aY (m/s2)

aY (m/s2)
0 0 0

−5 −5 −5
1 2 3 4 1 2 3 4 1 2 3 4
t (s) t (s) t (s)

Figure 6.23: Point-to-point trajectory. Sensitivity to an observer error.

sirable interactions between the slow and the fast controller. To illustrate this,
Figure 6.23 presents simulation results for a point-to-point trajectory (more de-
tails about the model will be given later). The left plot shows the slow controller
working alone. Good tracking performances are observed for θ2 , but the tip vibra-
tions decay slowly after reaching the nal conguration. In the middle plot, the
fast control law is activated, assuming a perfect estimation of the modal veloci-
ties. The improvement of the tip response is remarkable, and the joint-tracking
performances are barely aected. However, when one accounts for the non-perfect
estimation of η̇ δ due to the approximation in equation (6.23), the fast controller
is still ecient, but the joint-tracking controller is completely disturbed. This
is easily explained, since the approximation (6.23) is only justied in the fast
time-scale. In the slow time-scale, a signicant error is introduced and amplied
by the inverse actuator model, which directly aects the slow controller.
It is possible to improve the observer in order to eliminate this particular
source of error. However, in the experimental system, other disturbances may
aect the observer at low-frequencies, such as the accelerometer errors. Therefore,
there is a need to eliminate the low-frequency components induced in the fast
6.4. CLOSED-LOOP SYSTEM: SIMULATIONS AND EXPERIMENTS 207

controller by appropriate ltering. The design of those lters is presented in the

following section.

Filters design
The fast controller is expected to be active around the rst two natural
frequencies, between 3 Hz and 7 Hz. Both the lower and higher frequency distur-
bances should be attenuated. Two Butterworth band-pass lters are thus consid-
ered, whose properties are illustrated in Figure 6.24. Filter A is a 4th order lter,
with corner frequencies at 1 and 25 Hz , whereas lter B is a 2nd order lter with
corner frequencies at 1.6 and 16 Hz . Both lters have been designed in order
to keep the phase shift below 20 deg in the interval [3, 7] Hz . Filter A has the
advantage of a larger roll-o, leading to an ecient attenuation of low-frequency
disturbances. However, its lower corner frequency is smaller than for lter B, so
that its performances at intermediate frequencies, around ω s = 1.5 Hz , are less
interesting. Both lters will be tested in simulations and experiments.
Two less critical lters can also be mentioned:

- an output high-pass lter associated with the inverse actuator model (4th
order Butterworth lter, with corner frequency at 0.1 Hz ),

- a low-pass lter for the slow controller in order to reduce the measurement
noise (2nd order Butterworh lter, with corner frequency at 100 Hz ).

6.4 Closed-loop system: simulations and experi-

ments
In this section, the closed-loop system is simulated and tested experimentally.
The dynamic model includes the model of the mechanism, of the actuator, and
of the controller. As mentioned in Table 6.1 at the beginning of this chapter,
dierent models of the mechanism are considered for dierent purposes:

- a reduced-order linear model is used for the construction of root loci, and
the analysis of frequency responses,
208 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

10

Mag Ratio (dB)

0
−10
−20
−30
−40
−50 0 1 2
10 10 10
ω ,ω
1 2
300
Slow controller
200
Phase (deg)

Band−pass filter A
100 Band−pass filter B
0
−100
−200
−300 0 1 2
10 10 10
Freq (Hz)

Band-pass lter A Band-pass lter B

150 150

100 100
Imaginary axis (rad/s)

Imaginary axis (rad/s)

50 50

0 0

−50 −50

−100 −100

−150 −150
−150 −100 −50 0 −150 −100 −50 0
Real axis (rad/s) Real axis (rad/s)

10 10

8 8

6 6
Imaginary axis (rad/s)

Imaginary axis (rad/s)

4 4

2 2

0 0

−2 −2

−4 −4

−6 −6

−8 −8

−10 −10
−10 −8 −6 −4 −2 0 −10 −8 −6 −4 −2 0
Real axis (rad/s) Real axis (rad/s)

Figure 6.24: Band-pass lters at the inputs of the fast controller. In the Bode plot,
the closed-loop transfer function of the slow controller is represented, assuming a
velocity source actuator model (actuator 1).
6.4. CLOSED-LOOP SYSTEM: SIMULATIONS AND EXPERIMENTS 209

Figure 6.25: Model of the mechatronic system.

- a full-order nonlinear Finite Element model is considered for simulations in

the time-domain, i.e. the response to a tip force, point-to-point trajectories
and a square trajectory,

- a reduced-order nonlinear model is exploited to reduce the computational

load associated with time-domain simulations.
Two categories of analyses are dened: small motion around the home con-
guration, and large amplitude trajectories in the conguration space. Before
the presentation of the experimental and simulation results, the model of the
mechatronic system is described.

6.4.1 Model of the mechatronic system

In the experimental setup, the control law is implemented on a digital com-
puter, with a sampling rate of 1 kHz . The sampling eect has thus little eects
on the dynamic behavior in the frequency range of interest; we assume that it
can be neglected in the model.
Figure 6.25 represents the model of the mechatronic system. Each block is
implemented according to the formalism dened in chapter 4:
- Mechanism. Dierent models of the mechanism are used for the dierent
analyses, as mentioned above.

- Actuators. The actuator model accounts for the intrinsic velocity feedback
dened by equation (6.4).
210 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

- Accelerometers. In the Finite Element model, the tip accelerations ẍe are
available in inertial axes1 . In body axes, the accelerometer signals am are
computed by inversion of the formula (6.20) 2 .

- Filters and integrator. The lters are continuous linear state-space systems.

- Inertial damping + observer. A nonlinear algebraic equation is dened by

combination of (6.18) and (6.26):
" m #
des δθ −1 θ̇
Gf Ψδm (6.28)


θ̈ = M
|
d{z
=Gf (θ)
} cma

cf (θ) is implemented in the controller as a C-function, which is directly

G
reused in the simulation model.

- Inverse actuator model. Formula (6.27) is combined with a high-pass lter.

6.4.2 Dynamic behavior at home conguration

First of all, the closed-loop system is analyzed when Ralf is at home cong-
uration: root loci, bode plots, and step responses will be successively analyzed.
The linearized model of the mechatronic system is sucient, and the reduced-
order mechanical model is exploited for analysis in the frequency-domain. The
simulation of the step response in the time-domain is based on the full Finite
Element model, according to the methodology proposed in chapter 4.
For xed boundary conditions at the level of the actuators, the model predicts
the natural frequencies of the exible modes at home conguration:

ω1 = 4.4 Hz, ω2 = 6.7 Hz (6.29)

In the model, the exible modes are characterized by a small structural damping
(around 1 %). Obviously, in the closed-loop system, those poles are aected by
the behavior of the control system.
1 According to Remark 4.2 page 75, an "observer equation" is added for the accurate esti-
mation of ẍ .
e
2 Due to rather slow variations of the rotation matrix R with respect to the time-step

(h = 0.01 s), this formula is not strongly nonlinear, and does not represent a problem for
the numerical integration scheme (see Remark 4.4, page 84)
6.4. CLOSED-LOOP SYSTEM: SIMULATIONS AND EXPERIMENTS 211

Root locus
The root loci for increasing gains of the fast controller ξi (i = 1, 2) are
presented in Figure 6.26. Due to interactions with the actuators and the slow
controller, the exible modes have initially a signicant non-negligible damping.
In the absence of band-pass ltering, the vibration controller leads to signicant
damping improvements. Considering the position of the poles for ξi = 0.1, there
is room for further improvements. However, in the presence of ltering, a smaller
active damping can be reached for the rst exible mode. For large gain values,
unstable behaviors are predicted, and the value ξi = 0.1 appears as a reasonable
choice.

Frequency response
The tracking performances of the composite controller can be analyzed in
the transfer functions between the reference inputs and the dynamic responses.
Generally, a joint-tracking controller tries to optimize the responses of the actu-
ator dofs θ, but in practice one is more interested in the positioning of the end
eector. The impact of the fast controller on both types of transfer functions is
illustrated in Figure 6.27.
First, let us consider the predictions of the model when the fast controller is
turned o. In the actuator transfer function, the corner frequency is about 1.5 Hz ,
and the roll-o is -20 dB/decade. 4.4 Hz corresponds to the rst resonance of the
mechanism with clamped actuators. At this frequency, any motion of the actuator
is amplied in the whole structure and requires a huge amount of energy. This
phenomenon appears as an anti-resonance in the collocated transfer function. In
the tip transfer function, the exible modes are excited at the resonance despite
the roll-o of the joint-tracking controller.
When the fast controller is turned on, the actuator transfer function is mod-
ied around the resonance, where the inertial damping is active. The benets
mostly appear in the end eector transfer function, where the resonance peaks
are eciently attenuated. Due to phase lags, the tracking behavior is still not
ideal around the resonance, but the system is now isolated from disturbances
in this frequency range. This clearly illustrates the objective of the composite
control law: performances at low frequencies, and stabilization at the resonances.
212 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

Without ltering

Increasing ξ1 Increasing ξ2
50 50
45 45
40 40
35 35
Imaginary axis (rad/s)

Imaginary axis (rad/s)

30 30
25 25
20 20
15 15
10 10
5 5
0 0
−30 −25 −20 −15 −10 −5 0 5 −30 −25 −20 −15 −10 −5 0 5
Real axis (rad/s) Real axis (rad/s)

With lter A
(the dashed lines still refer to the unltered case)

Increasing ξ Increasing ξ
1 2
50 50
45 45
40 40
35 35
Imaginary axis (rad/s)

Imaginary axis (rad/s)

30 30
25 25
20 20
15 15
10 10
5 5
0 0
−30 −25 −20 −15 −10 −5 0 5 −30 −25 −20 −15 −10 −5 0 5
Real axis (rad/s) Real axis (rad/s)

Figure 6.26: Root loci. The arrows point to ξi = 0.1. In the second plot, remote
poles and loci associated with the higher corner frequency of the lters exist out
of the axes scales.
6.4. CLOSED-LOOP SYSTEM: SIMULATIONS AND EXPERIMENTS 213

Model

Input: ref θ ; Output: θ Input: ref θ ; Output: Ye

1 1 1
0 40
Mag Ratio (dB)

Mag Ratio (dB)

−10
20
−20
0
−30
−20
−40

−50 0 1
−40 0 1
10 10 10 10

0 0

−100
Phase (deg)

Phase (deg)

−100
−200

−300
−200
−400 Fast ctrl off Fast ctrl off
Fast ctrl on Fast ctrl on
−500 0 1
−300 0 1
10 10 10 10
Freq (Hz) Freq (Hz)

Experiments

Input: ref θ1 ; Output: θ1

10
Mag Ratio (dB)

−10

−20

−30 0 1
10 10

0
Phase (deg)

−50

−100

−150 Fast ctrl off

Fast ctrl on
−200 0 1
10 10
Freq (Hz)

Figure 6.27: Closed-loop transfer function (home conguration). Y e represents

the displacements in body axes, in the direction of the accelerometer measurement
Y.
am
214 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

Figure 6.28: Step force applied to the tip of Ralf (home conguration).

An experimental transfer function is also represented; at least from a quali-

tative point of view, the predictions of the model are conrmed.

Response to a tip force

The actively damped system should be less sensitive to external disturbances.
In order to demonstrate this capability, the response of the system is analyzed
when an external force is suddenly applied at the tip, as illustrated in Figure 6.28.
The step of force is exerted at time t = 0 s, with an amplitude g e = 89 N .
First, the simulation of this problem is considered. The full Finite Element
model is used for the modeling of the mechanism, and the control system is
modeled using the block diagram formalism developed in chapter 4. The Hilber-
Hughes-Taylor time-integration scheme is exploited with a time-step h = 0.01 s
and a numerical damping αf = 0.02. The step of force is a discontinuous phe-
nomenon, and consistent initial conditions have been carefully computed.
This problem has also been analyzed experimentally. The step of force is
emulated using a payload xed to the tip of Ralf with a cable: the instantaneous
cutting of the cable gives an equivalent upright force. In practice, the cutting
operation is replaced by a trigger mechanism.
In Figure 6.29, the correspondence between simulation and experimental re-
sults is impressive. The limited resolution of the measured signal θ2m appears
in the experimental data. A tiny shift is noticed in the static value of θ2m ; it is
6.4. CLOSED-LOOP SYSTEM: SIMULATIONS AND EXPERIMENTS 215

Simulation results
Fast control o Fast control on
15 15

10 10

5 5
aY (m/s2)

aY (m/s2)
0 0

−5 −5

−10 −10

−15 −15
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
t (s) t (s)

0.9518 0.9518

0.9516 0.9516
θ2 (m)

θ (m)

0.9514 0.9514
2

0.9512 0.9512

0.951 0.951

0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5

t (s) t (s)

Experimental results
Fast control o Fast control on
15 15

10 10

5 5
aY (m/s2)

aY (m/s2)

0 0

−5 −5

−10 −10

−15 −15
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
t (s) t (s)
0.9512 0.9512

0.951 0.951

0.9508 0.9508
θ2 (m)

θ (m)
2

0.9506 0.9506

0.9504 0.9504

0.9502 0.9502
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
t (s) t (s)

Figure 6.29: Tip disturbance. Filter A is used in the control algorithm.

216 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

caused by the osets of the hydraulic valves.

When the fast controller is not active, the actuator remains almost motion-
less, and the tip response is dominated by the second mode of vibration. In the
model, the main source of damping comes from the actuators. The decay of the
experimental response is faster, which might be due to unmodeled phenomena in
the joints, such as friction eects.
The active vibration controller eciently reduces the tip vibrations. This
performance is obtained at the price of a perturbed actuator motion, dominated
by a pole associated with the band-pass lter (in this case, lter A, with corner
frequency at 1 Hz ). In the next section, improvements will be discussed for the
attenuation of this phenomenon.

6.4.3 Trajectories in the conguration space

After the analysis of the controller performances around the home congura-
tion, the next step is to consider large amplitude motions, where several nonlinear
eects may arise, due to conguration changes, to the scheduled control strategy,
and to observer errors. The composite controller should also cope with the dual
task trajectory control/active damping.
When θ1 increases, the system gets closer to a singular conguration, where
actuator 1 is aligned with the rst link3 . For instance, in Figure 6.2, the slope
∂α1
∂θ1
increases with θ1 . As a consequence, any motion of the actuator is amplied
by the mechanism, and for a desired transmitted power, higher forces g1a are
required. The controllability measures the ability of the actuator forces to excite
the system; it is thus optimal when actuator 1 is orthogonal to the rst link4 and
it vanishes at the actuator singularity.
Therefore, the implementation of the control law is analyzed in a restricted
part of the conguration space θ1 < 1.12 m (α1 < 84 deg ), where the controlla-
bility is sucient for the design of an ecient controller. The trajectories follow
straight paths dened in the space of joint angles (Figure 6.30). Any of these
motions involves both actuators 1 and 2. A point-to-point motion and a square
3 Rigorously, at the singularity, actuator 1 is orthogonal to the virtual displacement of its
attachment point.
4 Rigorously, the controllability is optimal when actuator 1 is parallel to the virtual displace-

ment of its attachment point.

6.4. CLOSED-LOOP SYSTEM: SIMULATIONS AND EXPERIMENTS 217

Figure 6.30: Point-to-point motions in the conguration space.

trajectory will be successively considered.

The mechatronic system is modeled and simulated as in the previous case
(full Finite Element model, block-diagram model of the control system, Hilber-
Hughes-Taylor algorithm, h = 0.01 s, αf = 0.02).

Point-to-point motion
A smooth point-to-point trajectory from conguration S4 to conguration
S1 is analyzed. The trajectory θ ref (t) follows bang-bang acceleration proles, i.e.
ref ref
θ̈ is piecewise constant and the velocity θ̇ is limited. At the end conguration
(t = 0.85 s), the transition is not perfectly smooth in order to emulate possible
high-frequency disturbances.
Simulation and experimental results are compared in Figures 6.31. In the
simulation results, three subcases are studied. When the vibration controller is
turned o, the slow controller is able to track eciently the trajectory, with a
small delay. However, tip vibrations continue for a long time after the end of the
trajectory. Those vibrations are dominated by the rst exible mode. Without
ltering, the fast controller leads to an important reduction of those transient
vibrations, but the tracking controller is disturbed in an unacceptable way, as
discussed earlier. Filter A leads to a better trade-o between vibration control
and trajectory tracking; the slow oscillations at the actuator level are dominated
by the poles of the lter.
Qualitatively, the experiments agree with those observations (the non-ltered
218 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

Simulation results
Fast control o Fast control on / lter o Fast control on / lter A
1.2
1.08 1.08
1.06 1.15 1.06
1.04 1.04
1.1

θ2 (m)
θ2 (m)

θ2 (m)
1.02 1.02
1 1.05 1
0.98 0.98
1
0.96 0.96
response response response
0.94 reference 0.95 reference 0.94 reference
0 1 2 3 4 0 1 2 3 4 0 1 2 3 4
t (s) t (s) t (s)
30 30 30

20 20 20

10 10 10
aY (m/s2)

aY (m/s2)

aY (m/s )
2
0 0 0

−10 −10 −10

−20 −20 −20

−30 −30 −30

0 1 2 3 4 0 1 2 3 4 0 1 2 3 4
t (s) t (s) t (s)

Experimental results
Fast control o Fast control on / lter A
1.08 1.08
1.06 1.06
1.04 1.04
θ2 (m)

θ2 (m)

1.02 1.02
1 1
0.98 0.98
0.96 0.96
response response
0.94 reference 0.94 reference
0 1 2 3 4 0 1 2 3 4
t (s) t (s)
30 30

20 20

10 10
aY (m/s )

aY (m/s2)
2

0 0

−10 −10

−20 −20

−30 −30
0 1 2 3 4 0 1 2 3 4
t (s) t (s)

Figure 6.31: Point-to-point trajectory.

6.4. CLOSED-LOOP SYSTEM: SIMULATIONS AND EXPERIMENTS 219

fast controller has not been tested, fearing unpredictable motions of the mecha-
nism). However, the amplitude level of the transient vibrations after the motion
is higher than the prediction of the simulation, even when the fast controller is
disabled.
Part of this discrepancy may be attributed to unmodeled phenomena and
disturbances. For instance, clearance eects in the joints may excite the exible
motion at the end of the trajectory. In the tip-force experiment, the predictions
of the model were far closer to the experimental results, which can be explained
as follows. The response to a tip force is dominated by the second mode of
vibration, which is almost a pure bending mode of the second link, as represented
in Figure 6.14. Thus, the joints do not participate in the motion, and no clearance
phenomenon is observed. The situation is quite dierent for the point-to-point
motion, since the rst mode of vibration is then excited.
Another explanation may come from the non-perfect actuator model around
the rst natural frequency. Indeed, in Figure 6.12, it was dicult to make a deci-
sion whether the sensor measurements and the actuators were collocated or not.
For our developments, we assumed collocated measurements, but, a posteriori, an
additional simulation has been realized for the non-collocated case. The compar-
ison is given in Figure 6.32. For the noncollocated case, without active vibration
control, the vibrations detected by the linear sensors have a destabilizing eect on
the slow controller. When the vibration controller is activated, the eciency of
the active damping is overestimated by the model. We may conclude that the true
reality is between the collocated and the non-collocated situations. Experimental
data could be exploited for the optimization of an "intermediate" model. In the
following, this identication problem is left aside, and the simulation results rely
on the assumption of collocated measurements.
Figure 6.33 focuses on the behavior of the system at the end of the trajectory,
when the nal conguration has been reached. The damping performances of the
fast controller appear more clearly. Filter B turns out to be more appropriate in
this case: the slow motion of the actuator is more eciently damped.
In order to improve the computational eciency of the simulation algorithm,
the full Finite Element model of the mechanism can be replaced by the nonlinear
reduced model. The reduced model describes the dynamics in terms of modal
coordinates. The tip position is connected with the modal amplitudes by the
220 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

Simulation results (collocated sensors)

Fast control o Fast control on / lter A
30 30

20 20

10 10
aY (m/s2)

a (m/s2)
0 0

Y
−10 −10

−20 −20

−30 −30
0 1 2 3 4 0 1 2 3 4
t (s) t (s)

Simulation results (noncollocated sensors)

Fast control o Fast control on / lter A
30 30

20 20

10 10
aY (m/s2)

a (m/s2)

0 0
Y

−10 −10

−20 −20

−30 −30
0 1 2 3 4 0 1 2 3 4
t (s) t (s)

Experimental results
Fast control o Fast control on / lter A
30 30

20 20

10 10
aY (m/s2)

aY (m/s2)

0 0

−10 −10

−20 −20

−30 −30
0 1 2 3 4 0 1 2 3 4
t (s) t (s)

Figure 6.32: Comparison: collocated / noncollocated measurements.

6.4. CLOSED-LOOP SYSTEM: SIMULATIONS AND EXPERIMENTS 221

Simulation results
Fast control o Fast control on / lter A Fast control on / lter B
1.085 1.085 1.085

1.08 1.08 1.08

θ2 (m)

θ2 (m)

θ2 (m)
1.075 1.075 1.075

1.07 1.07 1.07

response response response
reference reference reference
1.065 1.065 1.065
1 2 3 4 1 2 3 4 1 2 3 4
t (s) t (s) t (s)
5 5 5
aY (m/s2)

aY (m/s2)

aY (m/s2)
0 0 0

−5 −5 −5
1 2 3 4 1 2 3 4 1 2 3 4
t (s) t (s) t (s)

Experimental results
Fast control o Fast control on / lter A Fast control on / lter B
1.085 1.085 1.085
response response response
reference reference reference
1.08 1.08 1.08
θ2 (m)

θ2 (m)

θ2 (m)

1.075 1.075 1.075

1.07 1.07 1.07

1.065 1.065 1.065

1 2 3 4 1 2 3 4 1 2 3 4
t (s) t (s) t (s)
30 30 30

20 20 20

10 10 10
aY (m/s2)

aY (m/s2)

aY (m/s )
2

0 0 0

−10 −10 −10

−20 −20 −20

−30 −30 −30

1 2 3 4 1 2 3 4 1 2 3 4
t (s) t (s) t (s)

Figure 6.33: Point-to-point trajectory: zoom at the end of the trajectory, after
reaching the end point.
222 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

Simulation results
Full model Nonlinear reduced model
1.08
1.06 1.05
1.04

θ (m)
θ2 (m)

1.02
1

2
1
0.98
0.96 response
response 0.95
0.94 reference reference
0 1 2 3 4 0 1 2 3 4
t (s) t (s)
30 30

20 20

10 10
aY (m/s2)

a (m/s2)

0 0
Y

−10 −10

−20 −20

−30 −30
0 1 2 3 4 0 1 2 3 4
t (s) t (s)

Full model Nonlinear reduced model

1.085
1.085

1.08 1.08
θ2 (m)
θ2 (m)

1.075 1.075

1.07 1.07
response response
reference reference
1.065 1.065
1 2 3 4 1 2 3 4
t (s) t (s)
5 5
aY (m/s2)

aY (m/s2)

0 0

−5 −5
1 2 3 4 1 2 3 4
t (s) t (s)

Figure 6.34: Comparison of the reduced model and the full model. Point-to-point
trajectory, normal bandwidth of the slow controller, fast controller o.
6.5. CONCLUDING REMARKS 223

relation ẋe = Ψeη η̇ , which is implemented in a special ModesToDofs element. In

this element, the dependence of the 2 × 4 mode-shape matrix Ψeη with respect to
the conguration θ is represented by a polynomial approximation obtained from
the model reduction procedure.
The results are compared with the full model in Figure 6.34. Discrepancies
mostly concern the high-frequency content of the response, but the correspon-
dence is good for the slow component of the motion. We conclude that the model
reduction technique can be utilized to speed up the simulation of a (controlled)
mechanism.

Square trajectory
For a point-to-point trajectory, we are especially interested in the perfor-
mances at the nal conguration. Since the performances of the composite con-
troller might depend on the conguration, a square trajectory is analyzed in the
conguration space. The cyclic sequence S1 , S2 , S3 , S4 , S1 was illustrated in Fig-
ure 6.30. The experimental results are presented in Figures 6.35 and 6.36, and the
performances of the vibration controller are appreciable for each conguration.

6.5 Concluding remarks

In this chapter, the theoretical concepts developed previously have been ap-
plied to a large and lightweight manipulator, with a parallel actuation mechanism.
The dynamic modeling of the mechanism, of the actuators and of the controller
has been realized, validated and exploited at the dierent stages of a control
design procedure. In summary, dierent models have been constructed:

- a low-order linearized model of the actuated mechanism, for experimental

validation,

- a low-order and nonlinear model of the mechanism, for the design of the
vibration controller, and for ecient simulation,

- a linearized model of the controlled mechanism, for root locus and frequency
response analyses,
224 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

Experimental results
Control o Control on / lter B
1.1 1.1
response response
1.05 S1 reference 1.05 S1 reference

1 1
S2 S4 S2 S4
θ2 (m)

θ2 (m)
0.95 0.95

0.9 0.9

0.85 0.85
S3 S3
0.8 0.8

0.75 0.75
0 5 10 15 0 5 10 15
t (s) t (s)

30 30

20 20

10 10
aY (m/s )

aY (m/s2)
2

0 0

−10 −10

−20 −20

−30 −30
0 5 10 15 0 5 10 15
t (s) t (s)

Figure 6.35: Square trajectory.

6.5. CONCLUDING REMARKS 225

Experimental results
Fast control o Fast control on / lter B
30 30

20 20

10 10
aY (m/s2)

aY (m/s )
2
S1
0 0

−10 −10

−20 −20

−30 −30
1 2 3 4 1 2 3 4
t (s) t (s)

30 30

20 20

10 10
aY (m/s2)

aY (m/s )
2

S2
0 0

−10 −10

−20 −20

−30 −30
5 6 7 8 5 6 7 8
t (s) t (s)

30 30

20 20

10 10
aY (m/s2)

a (m/s2)

S3
0 0
Y

−10 −10

−20 −20

−30 −30
9 10 11 12 9 10 11 12
t (s) t (s)

30 30

20 20

10 10
aY (m/s2)

a (m/s2)

S4
0 0
Y

−10 −10

−20 −20

−30 −30
13 14 15 16 13 14 15 16
t (s) t (s)

Figure 6.36: Square trajectory: various zooms.

226 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR

- a full model of the mechatronic system, for the simulation of the closed-loop
dynamics.

The predictions of those models were generally in good agreement with the exper-
imental results. However, the dynamic parameters of the actuators are critical for
the quality of the model. Their estimation was based on experimental data, but
a more systematic identication procedure would be helpful at this level. Some
discrepancies are also attributed to friction and clearance eects in the joints of
the mechanism.
A two-time-scale strategy has been selected for the design of the controller. A
slow controller is responsible for the tracking performances, and a fast vibration
controller damps the exible motion. The vibration controller is based on the
inertial damping concept, and we conclude that:

- the inertial damping concept, initially developed for macro/micro manipu-

lators is also applicable for any exible manipulator,

- the implementation of this control strategy is fairly simple: only two ac-
celerometers complement the hardware of the joint-tracking controller,

- a low-order model is necessary for the design and the implementation of the
control law, our reduced-order modeling technique is appropriate for this
purpose,

- one diculty is associated with the reliable estimation of the rates of the
exible coordinates, especially at low frequencies,

- in order to prevent interferences between both controllers, ltering is nec-

essary,

- nally, a compromise can be obtained between the eciency of the fast

vibration controller, and the performances of the slow tracking controller,

The modeling and design concepts were intimately connected throughout

the chapter. The major conclusion is that the modeling and simulation tools
developed in this dissertation are highly relevant for the control design of a complex
experimental mechanism.
6.5. CONCLUDING REMARKS 227

In future research, signicant improvements of the controller could be ob-

tained using a more accurate estimation formula, or a dynamic observer. Replac-
ing the accelerometers by strain gauges would also lead to a more reliable signal
at low frequencies, less sensitive to the rigid motion.
228 CHAPTER 6. CONTROL OF AN EXPERIMENTAL MANIPULATOR
 VII

Conclusion

The integrated design of mechatronic systems is a multidisciplinary challenge.

In particular, our objective was to develop modeling and simulation methods
to assist the engineer at the dierent stages of the design of controlled exible
mechanisms. The dissertation investigated two main research orientations: the
simulation of mechatronic systems, and the reduced-order modeling of exible
mechanisms. Those concepts were also exploited for the control design of an
experimental manipulator.
The specic conclusions associated with the dierent parts of this work are
reported hereafter.

1. Integrated simulation of mechatronic systems

A modular formulation has been presented for the dynamic analysis of mecha-
tronic systems composed of a exible mechanism and a control system. The cou-
pled dierential and algebraic equations are constructed numerically, and their
time-integration is performed according to a strongly coupled approach.
The methodology relies on several strategic choices and personal contri-
butions:
- The selection of the Finite Element method as a modular modeling frame-
work for the analysis of mechatronic systems, with the advantage of an
accurate representation of the mechanical subsystem.

- The use of the block diagram language for the description of control systems
in the Finite Element environment. This leads to a general, systematic and
230 CHAPTER 7. CONCLUSION

modular formulation, familiar to control engineers.

- The implementation of the generalized-α method for the simulation of the

coupled system. The ability of this method to deal with constrained me-
chanical equations or with rst-order equations is well-known, and the pos-
sibility to adjust the high-frequency numerical damping is very convenient
for sti problems.

- The analysis of stability and convergence properties of the generalized-α

method when applied to the coupled dynamic equations.

- The extension of the time-integration algorithm to deal with discontinu-

ous dynamic eects in some subsystems, for instance due to sampling or
saturation phenomena.

Well-established formalisms from exible multibody dynamics and from sys-

tem theory have thus been unied for the reliable and modular simulation of
controlled exible mechanisms. The most important conclusions are:

- The theoretical convergence and stability results are conrmed by several

test-cases.

- The proposed methodology is able to predict the dynamic behavior of com-

plex mechatronic systems, e.g. a car equipped with a complex semi-active
suspension system, and a controlled exible manipulator. For this last sys-
tem, the simulation results are validated experimentally.

- The simulation method is a relevant pre-prototyping tool, useful for control

design.

Future investigations could focus on the following issues:

- The analysis and the optimization of the computational eciency, in par-

ticular the development of an adaptive time-stepping method.

- The combination of the approach with a multi-rate or a subcycling strategy,

in order to improve the computational eciency when very fast dynamic
phenomena have to be represented within specic subsystems.
 231

- The extension of the functional simulation concept to include subsystem

models developed in formalisms dedicated to specic technologies (e.g. elec-
trical, hydraulic or pneumatic systems), which would reinforce the modu-
larity and the generality of the methodology.

- The development of a user-friendly graphical interface, and of drivers with

software specialized in block diagram modeling.

2. Model reduction in exible multibody dynamics

In order to represent concisely the conguration-dependent dynamics of a
exible mechanism, a systematic model reduction technique has been proposed,
which transforms an initial high-order model into a low-order and explicit model.
The reduction procedure relies on (i) an order reduction, which can be realized
locally for any given conguration, (ii) a consistent approximation of the reduced-
order model in the conguration space.
Since similar approaches are reported in the literature, we need to draw the
attention to the originality of our methodology:

- As the Finite Element method is selected to formulate the dynamics of the

initial model, complex mechanical topologies can be systematically repre-
sented.

- The Global Modal Parameterization concept is formally dened in order

to guarantee the consistency of the reduction procedure and the physical
interpretation of the reduced coordinates. This parameterization is based
on an extension of the component-mode method, and it accounts for the
nonlinear kinematics. The Global Modal Parameterization is well-dened in
the conguration space of interest, and it represents exactly the rigid-body
motion.

- The local reduction procedure accounts for the nonlinearity of the Global
Modal Parameterization; in particular, the curvature of the parameteriza-
tion contributes to the equivalent inertia forces. Nonlinear eects caused
by large deformations are neglected.
232 CHAPTER 7. CONCLUSION

- A piecewise polynomial approximation of the reduced-order model in the

conguration space follows an adaptive strategy, which achieves a user-
specied tolerance on the approximation error.
A complex exible mechanism can thus be represented by a compact and
closed-form dynamic model, formulated in a systematic way. From the develop-
ment of the methodology and its application to various problems, we conclude
that:
- The model leads to a deep understanding of the dynamic system under
consideration. Valuable insights are obtained into the inertia, stiness and
kinematic properties all over the conguration space.
- The degree of nonlinearity of the model is associated with the degree of
nonlinearity of the dynamics seen by the actuators; the reduction method
breaks down at the actuator singularities.
- Low-order polynomials are able to capture eciently the smooth variations
of the model in the conguration space. A compromise is necessary between
accuracy, continuity, memory storage requirements and computational time
to build the model. Lagrange polynomials dened on a regular grid have
ideal continuity properties, but this strategy requires important resources to
capture a nonlinear behavior in a high-dimensional conguration space. For
such specic problems, adaptive conguration space decomposition and/or
quadratic polynomials may be preferred.
- The method is applicable for rigid mechanisms with parallel topology; the
reduced model gives the solution to the inverse dynamics problem, and it
is directly exploitable for computed-torque control techniques.
- The conguration-dependent dynamics of a exible manipulator is e-
ciently captured by a reduced model, as conrmed by experimental results.
- A reduced model is an appropriate basis for the design of a vibration con-
troller, e.g. using the inertial damping concept. It is portable and it can
be implemented online on a real-time controller.
- The computational eciency of a simulation algorithm is improved using a
reduced model.
 233

Several outcomes are foreseen for the model reduction technique:

- The reduction method can be a basis for the extraction of a more structured
model (e.g., a polytopic linear model or a model based on linear fractional
transformation), for which specialized control theories are applicable.

- In the literature related with component-mode synthesis, various enhanced

mode shapes have been dened; any of them could be associated with the
Global Modal Parameterization.

- A priori knowledge on the conguration space geometry (specied trajec-

tory, symmetries, etc) could be exploited for an optimized approximation
of the nonlinear model.

- As such, our formalism is applicable to mechanisms without pre-stressing

eects. The extension to include those eects could also be considered in
future investigations.

3. Control of a exible manipulator

The proposed simulation and modeling methods have been applied for the
motion and vibration control of a exible manipulator.
The design of the controller is based on the following personal develop-
ments:
- A two-time-scale control strategy is elaborated from the structure of the
dynamic model: the rigid modes are controlled by a slow joint-tracking
controller, and the exible modes by a fast vibration controller.

- The inertial damping concept is applied for the vibration control of the
exible manipulator.

- The interaction problem caused by a non-perfect time-scale separation is

analyzed, and solved by appropriate ltering.

Various models are involved in the control design procedure, which conrms
the relevance of our original modeling and simulation tools. From the simulation
and experimental results, we conclude that:
234 CHAPTER 7. CONCLUSION

- The identication of the parameters of the mechanical model is important

to obtain a good agreement with the experimental data. This motivates the
future development of systematic identication techniques in exible multi-
body dynamics, which can take advantage of the model reduction method.

- Inertial damping leads to a simple, pragmatic and ecient control strategy,

able to improve the dynamic behavior of a mechanism proned to vibration
problems.

- Due to the necessary time-scale separation, a compromise is necessary be-

tween motion performances and vibration suppression.

- The reliable estimation of the modal velocities is critical for the performance
of the controller. Improvements are expected if the acceleration feedback is
replaced by a strain feedback, less sensitive to low-frequency disturbances.

4. General perspective and closure

The modeling and simulation tools presented in this dissertation appear as a
relevant basis for constructive methods in mechatronics. In particular, trajectory
generation, inverse dynamics and concurrent design are challenging problems for
controlled mechanisms, for which optimization techniques can be of great assis-
tance. The utilization of our methods in a numerical optimization approach is
thus a recommended direction for future research.
Mechatronics is a multidisciplinary framework; at the closure of this the-
sis, we strongly believe that our contribution naturally resulted from the cross-
fertilization of concepts encountered in several elds of science and technology.
More than ever, discarding inter-domain barriers appears as a fruitful guideline
in engineering.

Acknowledgements
This work has been supported by the Belgian National Fund for Scientic Research (FNRS)
which is gratefully acknowledged. It also presents research results of the Belgian Program on
Inter-University Poles of Attraction initiated by the Belgian state, Prime Minister's oce,
Science Policy Programming (AMS IAP V/06). The scientic responsibility is assumed by its
author.
 Bibliography

[Ang01] G.Z Angelis. System Analysis, Modelling and Control with Polytopic Linear
Models. PhD thesis, Technische Universiteit Eindhoven, 2001.
[Arn01] M. Arnold. Numerically stable modular time integration of multiphysical sys-
tems. In K.J. Bathe, editor,Proceedings of the First MIT Conference on Com-
putational Fluid and Solid Mechanics, pages 10621064, Cambridge, MA, June
2001. Elsevier, Amsterdam.

[Arn04] M. Arnold. Numerical methods in the simulation of vehicle-guideway interaction.

In A. Buikis, R. Ciegis, and A.D. Fitt, editors, Progress in Industrial Mathemat-
ics at ECMI 2002, pages 115120, Berlin Heidelberg, 2004. SpringerVerlag.
[AS86] H. Asada and J.-J.E. Slotine. Robot Analysis and Control. Wiley-Interscience,
1986.

[Bau72] J. Baumgarte. Stabilization of constraints and integrals of motion in dynamical

systems. Comp. Meth. Appl. Mech. Eng., 1:116, 1972.
[Bay87] E. Bayo. A nite element approach to control the end-point motion of a single-
link exible robot. Int. J. Robotic Systems, 4(1):6375, 1987.
[BBT03] O.A. Bauchau, C.L. Bottasso, and L. Trainelli. Robust integration schemes for
exible multibody systems. Comp. Meth. Appl. Mech. Eng., 192:395420, 2003.
[BC04] C.L. Bottasso and A. Croce. Optimal control of multibody systems using an en-
ergy preserving direct transcription method. Multibody System Dynamics, 12:17
45, 2004.

[BCGF04] C.L. Bottasso, A. Croce, L. Ghezzi, and P. Faure. On the solution of inverse
dynamics and trajectory optimization problems for multibody systems. Multibody
System Dynamics, 11:122, 2004.
[BCP96] K.E. Brenan, S.L. Campbell, and L.R. Petzold. Numerical Solution of Initial-
Value Problems in Dierential-Algebraic Equations. SIAM, Philadelphia, 2nd
edition, 1996.

[BDG02] O. Brüls, P. Duysinx, and J.-C. Golinval. An adaptation of the Newmark scheme
for the integrated simulation of mechatronic systems. In Proc. of the 6th Int.
236 BIBLIOGRAPHY

Conf. on Motion and Vibration Control (MOVIC), volume 1, pages 484489,

Saitama, Japan, August 2002.

[BDG03] O. Brüls, P. Duysinx, and J.-C. Golinval. Computational environment for the
design of exible mechanisms with feedback control. In Proc. of the 6th National
Congress on Theoretical and Applied Mechanics, Ghent, Belgium, May 2003.
[BDG04a] O. Brüls, P. Duysinx, and J.-C. Golinval. Generation of closed-form models for
the control of exible mechanisms: a numerical approach. In Proc. of the 7th Int.
Conf. on Motion and Vibration Control (MOVIC), Saint-Louis, United States,
August 2004.

[BDG04b] O. Brüls, P. Duysinx, and J.-C. Golinval. Reliable simulation of mechatronic sys-
tems using newmark algorithms. In Colloquium of the European Mechanics Soci-
ety on Advances in Simulation Techniques for Applied Dynamics (EUROMECH
452) - Book of Abstracts, Halle, Germany, March 2004.
[BDG04c] O. Brüls, P. Duysinx, and J.-C. Golinval. A systematic model reduction method
for the control of exible multibody systems. In Proc. of the 21st Int. Congress
of Theoretical and Applied Mechanics (ICTAM), Warsaw, Poland, August 2004.
[BDG06] O. Brüls, P. Duysinx, and J.-C. Golinval. A model reduction method for the
control of rigid mechanisms. Multibody System Dynamics, in press:, 2006.
[BG03] O. Brüls and J.-C. Golinval. Simulation of an active control system in a hot-
Virtual Nonlinear
dip galvanizing line. In W. Schiehlen and M. Valásek, editors,
Multibody Systems, volume 103 of NATO Science Series II: Mathematics, Physics
and Chemistry, pages 325332. Kluwer Academic Publishers, 2003.
[Boo84] W.J. Book. Recursive Lagrangian dynamics of exible manipulator arms. Int.
J. Robotics Res., 3(3):87101, 1984.
[Boo93] W.J. Book. Controlled motion in an elastic world. ASME J. Dyn. Syst., Meas.,
and Control, 115(2B):252261, 1993.
[BP87] E. Bayo and B. Paden. On trajectory generation for exible robots. Int. J.
Robotic Systems, 4(2):229235, 1987.
[Car89] A. Cardona. An Integrated Approach to Mechanism Analysis. PhD thesis, Uni-
versité de Liège, 1989.

[CB68] R. Craig and M. Bampton. Coupling of substructures for dynamic analysis.

AIAA J., 6(7):13131319, 1968.
[CG88] A. Cardona and M. Géradin. A beam nite element non-linear theory with nite
rotations. Int. J. Num. Meth. Eng., 26:24032438, 1988.
[CG89] A. Cardona and M. Géradin. Time integration of the equations of motion in
mechanism analysis. Computers and Structures, 33:801820, 1989.
BIBLIOGRAPHY 237

[CG91] A. Cardona and M. Géradin. A superelement formulation for mechanism analysis.

Int. J. Num. Meth. Eng., 32(8):15651594, 1991.
[CH93] J. Chung and G.M. Hulbert. A time integration algorithm for structural dynam-
ics with improved numerical dissipation: The generalized-α method. ASME J.
Appl. Mech., 60:371375, 1993.
[CKG94] A. Cardona, I. Klapka, and M. Géradin. Design of a new nite element program-
ming environment. Engineering Computations, 11:365381, 1994.
[Cra87] R. Craig. A review of time-domain and frequency domain component-mode syn-
thesis methods. Int. J. Analytical and Experimental Modal Analysis, 11(6):562
570, 1987.

[CS84] H.R. Cannon and E. Schmitz. Initial experiments on the end-point control of a
exible one-link robot. Int. J. Robotics Res., 3(3):6275, 1984.
[CU02] F.M. Caswara and H. Unbehauen. A neurofuzzy approach to the control of a
exible-link manipulator. IEEE Trans. Robotics and Automation, 18(6):932944,
December 2002.

[CW03] D. Chablat and Ph. Wenger. Architecture optimization of a 3-dof translational

parallel mechanism for machining applications, the Orthoglide. IEEE Trans.
Robotics and Automation, 19(3):403 410, June 2003.
[CWA00] D. Chablat, Ph. Wenger, and J. Angeles. Conception isotropique d'une morpholo-
gie parallèle : Application à l'usinage. In Proc. of the 3rd Int. Conf. On Integrated
Design and Manufacturing in Mechanical Engineering, Montreal, Canada, May
2000.

[CWM02] D. Chablat, Ph. Wenger, and J-P Merlet. Workspace analysis of the orthoglide
using interval analysis. In Proc. of the 8th Int. Symp. on Advances in Robot
Kinematics, Caldes de Malavella, Spain, June 2002. Kluwer Academic Publishers.
[CYC02] J. Cheong, Y. Youm, and W.K. Chung. Joint tracking controller for multi-link
exible robot using disturbance observer and parameter adaptation scheme. Int.
J. Robotic Systems, 19(8):401417, 2002.
[DB94] S. Devasia and E. Bayo. Inverse dynamics of articulated exible structures : Si-
multaneous trajectory tracking and vibration reduction. J. Dynamics and Con-
trol, 4(3):299309, 1994.
[DF00] P. De Fonseca. Simulation and Optimisation of the Dynamic Behaviour of
Mechatronic Systems. PhD thesis, Katholieke Universiteit Leuven, 2000.
[DJB94] J.G. De Jalón and E. Bayo. Kinematic and Dynamic Simulation of Multibody
Systems - The Real-Time Challenge. Springer Verlag, 1994.
[DLS93] A. De Luca and B. Siciliano. Inverse-based nonlinear control of robot arms with
exible links. AIAA J. Guidance, Control, and Dynamics, 16(6):11691176, 1993.
238 BIBLIOGRAPHY

[DV76] B.F. De Veubeke. The dynamics of exible bodies. Int. J. Engineering Science,
14:895913, 1976.

[EBB02] S. Erlicher, L. Bonaventura, and O.S. Bursi. The analysis of the generalized-α
method for non-linear dynamic problems. Computational Mechanics, 28:83104,
2002.

[Eck99] H. Ecker. Comparison 11 (scara-robot) with acsl, full hybrid modelling approach
- environment level. Simulation News Europe, 26:43, July 1999.
[EMO98] H. Elmqvist, S.E. Mattson, and M. Otter. Modelica - the new object-oriented
modeling language. In Proc. of the 12th European Simulation Multiconference
(ESM'98), June 1998.
[ES98] P. Eberhard and W. Schiehlen. Hierarchical modeling in multibody dynamics.
Arch. of Appl. Mech., 68:237246, 1998.
[Fav97] W. Favre. Contribution à la Représentation Bond Graph des Systèmes Mé-
caniques Multicorps. PhD thesis, INSA de Lyon, 1997.
[FLMR99] M. Fliess, J. Lévine, P. Martin, and P. Rouchon. A Lie-Bäcklund approach to
equivalence and atness of nonlinear systems. IEEE Trans. Autom. Control,
44(5):922937, May 1999.

[FLOR95] M. Fliess, J. Lévine, F. Ollivier, and P. Rouchon. Flatness and dynamic feedback
linearizability: Two approaches. In Proc. 3rd European Control Conference, 1995.

[For98] E. Forsthuber. Comparison 11 (scara-robot) with dymola, classical mechani-

cal approach, automatic-symbolical and numerical inversion. Simulation News
Europe, 23:49, July 1998.
[Gaw98] W. K. Gawronski. Dynamics and Control of Structures - A Modal Approach.
Springer, New York, 1998.

[GC01] M. Géradin and A. Cardona. Flexible Multibody Dynamics: A Finite Element

Approach. John Wiley & Sons, New York, 2001.
[Gea71] C.W. Gear. Numerical Initial Value Problems in Ordinary Dierential Equations.
Prentice-Hall, New Jersey, 1971.

[Geo02] L.E. George. Active Vibration Control of a Flexible Base Manipulator. PhD
thesis, Georgia Institute of Technology, 2002.

[GKCW02] S. Guegan, W. Khalil, D. Chablat, and P. Wenger. Modélisation dynamique d'un

robot parallèle à 3-ddl : l'Orthoglide. In Conférence Internationale Francophone
d'Automatique, Nantes, July 2002.
[GLL98] L.B. Gutiérrez, F.L. Lewis, and J.A. Lowe. Implementation of a neural net-
work tracking controller for a single exible link: Comparison with PD and PID
controllers. IEEE Trans. Industrial Electronics, 45(2):307318, april 1998.
BIBLIOGRAPHY 239

[GLZ96] S.S. Ge, T.H. Lee, and G. Zhu. Energy-based robust controller design for multi-
link exible robots. Mechatronics, 6(7):779798, 1996.
[GR97] M. Géradin and D. Rixen. Mechanical Vibrations: Theory and Application to
Structural Dynamics. John Wiley & Sons, New York, 2nd edition, 1997.
[GS00] F. Ghorbel and M.W. Spong. Integral manifolds of singularly perturbed systems
with application to rigid-link exible-joint multibody systems. Int. J. of Non-
Linear Mechanics, 35:133155, 2000.
[GVVD04] K. Gallivan, A. Vandendorpe, and P. Van Dooren. Model reduction of MIMO
systems via tangential interpolation. SIAM J. Matrix Analysis and Applications,
26(2):328349, 2004.

[GW84] C.W. Gear and D.R. Wells. Multirate linear multistep methods. BIT, 24:482
502, 1984.

[HAV03] A. Heckmann, M. Arnold, and O. Vaculin. Distributed multiphysical phenom-

ena in multibody dynamics. In Proc. of the ECCOMAS Conf. on Advances in
Computational Multibody Systems, Lisbon, Portugal, July 2003.
[Her79] D.N. Herting. A general purpose, multi-stage, component-mode synthesis
method. In Proceedings
of the AIAA/ASME/ASCE/AHS 20th Structure, Struc-
tural Dynamics and Material Conference, 1979.
[HHT77] H. Hilber, T. Hughes, and R. Taylor. Improved numerical dissipation for time in-
tegration algorithms in structural dynamics. Earthquake Engineering and Struc-
tural Dynamics, 5:283292, 1977.
[HNW87] E. Hairer, S.P. Norsett, and G. Wanner. Solving Ordinary Dierential Equations
I - Nonsti Problems. Springer-Verlag, 1987.
[Hol80] J.M. Hollerbach. A recursive formulation of lagrangian manipulator dynamics.
IEEE Trans. Syst., Man, Cybern., 10(11):730736, 1980.
[HP95] V. Hadwitch and K. Pfeier. The principle of virtual work in mechanical and
electromechanical systems. Arch. of Appl. Mech., 65:390400, 1995.
[Hug87] T.J.R. [Link] Finite Element Method, Linear Static and Dynamics Finite
Element Analysis. Prentice Hall, 1987.
[Hug88] J.D. Huggins. Experimental verication of a model of a two-link exible,
lightweight manipulator. Master's thesis, Georgia Institute of Technology, 1988.

[Hur65] W.C. Hurty. Dynamic analysis of structural systems using component modes.
AIAA J., 3(4):678685, 1965.
[Hus91] R.L. Huston. Computer methods in exible multibody dynamics. Int. J. Num.
Meth. Eng., 32:16571668, 1991.
240 BIBLIOGRAPHY

[HW91] E. Hairer and G. Wanner. Solving Ordinary Dierential Equations II - Sti and
Dierential-Algebraic Problems. Springer-Verlag, 1991.
[Isi95] Isidori. Nonlinear Control System. Springer, 1995.
[JKDW01] L. Jaulin, M. Kieer, O. Didrit, and E. Walter. Applied Interval Analysis.
Springer Verlag, London, 2001.

[JWH00] K.E. Jansen, C.H. Whiting, and G.M. Hulbert. A generalized-a method for
integrating the ltered navier-stokes equations with a stabilized nite element
method. Comp. Meth. Appl. Mech. Eng., 190:305319, 2000.
[Kar97] D. Karnopp. Understanding multibody dynamics using bond graph representa-
tions. J. Franklin Inst., 334B(4):631642, 1997.
[KB94] D.S. Kwon and W.J. Book. A time-domain inverse dynamic tracking control
of a single-link exible manipulator. ASME J. Dyn. Syst., Meas., and Control,
116(2):193200, 1994.

[KJT95] F. Khorrami, S. Jain, and A. Tzes. Experimental results on adaptive nonlinear

control and input preshaping for multi-link exible manipulators. Automatica,
31(3):8397, 1995.

[KK00] T.S. Kim and Y.Y. Kim. MAC-based mode-tracking in structural topology op-
timization. Computers and Structures, 74:375383, 2000.
[KMR00] D. Karnopp, D.L. Margolis, and R.C. Rosenberg. System Dynamics: Modeling
and Simulation of Mechatronic Systems. John Wiley & Sons Inc., 3rd edition,
2000.

[KS00] R. Kübler and W. Schiehlen. Two methods of simulator coupling. Mathematical

and Computer Modelling of Dynamical Systems, 6(2):93113, 2000.
[KTKH67] H. E. Koenig, Y. Tokad, H. K. Kesavan, and H. G. Hedges. Analysis of Discrete
Physical Systems. McGraw-Hill Book Company, New York NY, 1967.
[Lan79] Y.D. Landau. Adaptive Control - The Model Reference Approach. Marcel Dekker,
1979.

[LCG04] E.V. Lens, A. Cardona, and M. Géradin. Energy preserving time integration for
constrained multibody systems. Multibody System Dynamics, 11(1):4161, 2004.
[Lee90] J.W. Lee. Dynamic Analysis and Control of Lightweight Manipulators With
Flexible Parallel Link Mechanism. PhD thesis, Georgia Institute of Technology,
1990.

[LHKB91] H.-J. Lai, E.J. Haug, S.S. Kim, and D.-S. Bae. A decoupled exible-relative
co-ordinate recursive approach for exible multibody dynamics. Int. J. Num.
Meth. Eng., 32:16691690, 1991.
BIBLIOGRAPHY 241

[LLL03] G. Liu, Y. Lou, and Z. Li. Singularities of parallel manipulators: A geometric

treatment. IEEE Trans. Robotics and Automation, 19(4):579594, 2003.
[LSS04] C. Lauwerys, J. Swevers, and P. Sas. Model free design for a semi-active suspen-
sion of a passenger car. In Proc. of ISMA 2004, Leuven, Belgium, September
2004.

[LWP80] J.Y.S. Luh, M.W. Walker, and R.P.C. Paul. On-line computational scheme for
mechanical manipulators. ASME J. Dyn. Syst., Meas., and Control, 102:6976,
1980.

[Mag91] D.P. Magee. Dynamic control in teleoperation of a exible manipulator. Master's

thesis, Georgia Institute of Technology, 1991.

[MC95] L. Meirovitch and Y. Chen. Trajectory and control optimization for exible space
robots. AIAA J. Guidance, Control, and Dynamics, 18(3):493502, 1995.
[McP96] J.J. McPhee. On the use of linear graph theory in multibody system dynamics.
Nonlinear Dynamics, 9:7390, 1996.
[McP98] J.J. McPhee. Automatic generation of motion equations for planar mechani-
cal systems using the new set of "branch coordinates". Mech. Mach. Theory,
33(6):805823, 1998.

[McP03] J.J McPhee. Virtual prototyping of multibody systems with linear graph theory
and symbolic computing. In W. Schiehlen and M. Valásek, editors, Virtual Non-
linear Multibody Systems, volume 103 of NATO Science Series II : Mathematics,
Physics and Chemistry, pages 3756. Kluwer Academic Publishers, 2003.
[MEFK97] P. Maisser, O. Enge, G. Freundenberg, and G. Kielau. Electromechanical in-
teractions in multibody systems containing electromechanical drives. Multibody
System Dynamics, 1(3):281302, 1997.
[Mei90] L. Meirovitch. Dynamics and Control of Structures. John Wiley & Sons, New-
York, 1990.

[MK91] L. Meirovitch and M.-K. Kwak. Rayleigh-ritz based substructure synthesis for
exible multibody dynamics. AIAA J., 28:17091719, 1991.
[MM95] R.H. Myers and D.C. Montgomery. Response Surface Methodology. Wiley Series
in Probability and Statistics, 1995.

[Mon97] D.C. Montgomery. Design and Analysis of Experiments. John Wiley & Sons,
New York, 4th edition, 1997.

[MPK00] M. Moallem, R.V. Patel, and K. Khorasani. Flexible-link Robot Manipulators -

Control Techniques and Structural Design. Springer, 2000.
[MS94] R. Murray-Smith. A Local Model Network Approach to Nonlinear Modelling. PhD
thesis, University of Strathclyde, Department of Computer Science, Glasgow,
Scotland, November 1994.
242 BIBLIOGRAPHY

[Nel99] O. Nelles. Nonlinear System Identication with Local Linear Neuro-Fuzzy Models.
PhD thesis, TU Darmstadt, 1999.

[New59] N. Newmark. A method of computation for structural dynamics. ASCE J. Eng.

Mech. Div., 85:6794, 1959.
[Nik03] P. Nikravesh. The use of principal axes as oating reference frame for a moving
deformable body. In Proc. of the ECCOMAS Conf. on Advances in Computa-
tional Multibody Systems, Lisbon, Portugal, July 2003.
[Obe98] Obergfell. End-point position sensing and control of exible multi-link manipu-
lators. Master's thesis, Georgia Institute of Technology, 1998.

[OC89] C.M. Oakley and R.H. Cannon. End-point control of a two-link manipulator with
a very exible forearm: Issues and experiments. In Proc. of the 1989 American
Control Conference, volume 2, pages 13811388, Pittsburgh, PA, USA, June
1989.

[OE97] M. Otter and H. Elmqvist. Energy ow modeling of mechatronic systems via
object diagrams. InProc. of the 2nd Mathmod Vienna, IMACS Symposium on
Mathematical Modelling, pages 705710, 1997.
[OL99] J.H. Oh and J.S. Lee. Control of exible joint robot system by backstepping
design approach. Automation and Soft Computing, 5(4):267278, 1999.
[OS95] B. Owren and H.H. Simonsen. Alternative integration methods for problems in
structural dynamics. Comp. Meth. Appl. Mech. Eng., 122:110, 1995.
[OS01] F. Ollivier and A. Sedoglavic. A generalization of atness to nonlinear systems
of partial dierential equations. application to the command of a exible rod. In
Proceedings of the 5th IFAC Symposium on Nonlinear Control Systems, volume 1,
pages 196200, Saint Petersburg, Russia, July 2001. Elsevier.

[PK99] F.C. Park and J.W. Kim. Singularity analysis of closed kinematic chains. ASME
J. Mech. Des., 121(1):3238, 1999.
[PLS03] J.R. Phillips, D. Luca, and L.M. Silveira. Guaranteed passive balancing trans-
formations for model order reduction. IEEE Trans. Computer-Aided Design of
Integrated Circuits and Systems, 22(8):1027 1041, August 2003.
[Pre97] A. Preumont. Vibration Control of Active Structures, An Introduction. Kluwer,
1997.

[PTVF92] W.H. Press, S.A. Teukolsky, W.T. Vetterling, and B.P. Flannery. Numerical
Recipes in C - The Art of Scientic Computing. Cambridge University Press,
2nd edition, 1992.

[Rai77] M.H. Raibert. Analytical equations vs. table look-up for manipulation: a unifying
concept. In Proc. of the IEEE Conf. Decision and Control, New Orleans, 1977.
BIBLIOGRAPHY 243

[RS98] A. Rükgauer and W. Schiehlen. Simulation of modular dynamic systems. Com-

puters and Structures, 46:535542, 1998.
[RV02] I. Romero and R. Vignjevic. Deployable space manipulator closed loop con-
trol, ideas and possibilities of using gps as sensor. Advances in Space Research,
30(2):419425, 2002.

[SAM99] SAMTECH, Liège, Belgium. SAMCEF Users Manual - v 8.0, 1999.

[Sas04] L. Sass. Symbolic Modeling of Electromechanical Multibody Systems. PhD thesis,
Université Catholique de Louvain, 2004.

[SB88] B. Siciliano and W.J. Book. A singular perturbation approach to control of

lightweight fexible manipulators. Int. J. Robotics Res., 7(4):7990, 1988.
[Sch97] W. Schiehlen. Multibody system dynamics: Roots and perspectives. Multibody
System Dynamics, 1:149188, 1997.
[Sch98] J. Scheikl. Comparison 11 (scara-robot) with matlab/simulink, hybrid modelling
approach, model level. Simulation News Europe, 24:44, November 1998.
[Sch01] C.W. Scherer. LPV control and full block multipliers. Automatica, 37:361375,
2001.

[Sco95] M.A. Scott. Time Varying Compensator Design for Recongurable Structures
using Non-Collocated Feedback. PhD thesis, University of Colorado, 1995.
[SF03] J.-C. Samin and P. Fisette. Symbolic Modeling of Multibody Systems. Kluwer
Academic Publishers, 2003.

[SH04] K. Seto and M. Hon. Modeling strategy suited to motion and vibration control
for exible structures. In Proc. of the 7th Int. Conf. on Motion and Vibration
Control (MOVIC), St-Louis, USA, August 2004.
[Sha98] A.A. Shabana. Dynamics of Multibody Systems. Cambridge University Press,
2nd edition, 1998.

[Sim85] J.C. Simo. A nite strain beam formulation. the three-dimensional dynamic
problem. Part I. Comp. Meth. Appl. Mech. Eng., 49:5570, 1985.
[SJK97] R. Sepulchre, M. Jankovic, and P. Kokotovic. Constructive Nonlinear Control.
Springer-Verlag Series in Communications and Control Engineering, 1997.

[SL91] J.-J.E. Slotine and W. Li. Applied Nonlinear Control. Prentice Hall, 1991.
[SLP03] R. Serban, D. Li, and L.R. Petzold. Adaptive algorithms for optimal control
of time-dependent partial dierential-algebraic equation systems. Int. J. Num.
Meth. Eng., 57(2):14571469, 2003.
[SM00] P. Shi and J. McPhee. Dynamics of exible multibody systems using virtual
work and linear graph theory. Multibody System Dynamics, 4:355381, 2000.
244 BIBLIOGRAPHY

[SM03a] M. Scherrer and J. McPhee. Dynamic modelling of electromechanical multibody

systems. Multibody System Dynamics, 9:87115, 2003.
[SM03b] C. Schmitke and J. McPhee. Modelling mechatronic multibody systems using
symbolic subsystems models. InProc. of the ECCOMAS Conf. on Advances in
Computational Multibody Systems, Lisbon, Portugal, July 2003.
[SMH01] P. Shi, J. McPhee, and G. Heppler. A deformation eld for Euler-Bernoulli beams
with applications to exible multibody dynamics. Multibody System Dynamics,
5:79104, 2001.

[SP01] R. Serban and L.R. Petzold. COOPT - a software package for optimal control of
large-scale dierential-algebraic equation systems. Mathematics and Computers
in Simulation, 56(2):187203, 2001.
[SS90] N.C. Singer and W.P. Seering. Preshaping command inputs to reduce system
vibration. ASME J. Dyn. Syst., Meas., and Control, 112:7682, 1990.
[Str98] O. von Stryk. Optimal control of multibody systems in minimal coordinates. Z.
Angew. Math. Mech, 78:11171120, 1998.
[STW92] J.C. Simo, N. Tarnow, and K.K. Wong. Exact energy-momentum conserving
algorithms and symplectic schemes for nonlinear dynamics. Comp. Meth. Appl.
Mech. Eng., 100:63116, 1992.
[SV01] B. Siciliano and L. Villani. Two-time scale force and position control of exible
manipulators. In Proc. of the 2001 IEEE Int. Conf. on Robotics and Automation,
pages 27292734, Seoul, Korea, 2001.

[SVQ86] J.C. Simo and L. Vu-Quoc. A three-dimensional nite-strain rod model. Part II:
Computational aspects. Comp. Meth. Appl. Mech. Eng., 58:79116, 1986.
[SW83] A.A. Shabana and R.A. Wehage. A coordinate reduction technique for dynamic
analysis of spatial substructures with large angular rotations. J. Struct. Mech.,
11:401431, 1983.

[Sym04] W. Symens. Motion and Vibration Control of Mechatronic Systems with Variable
Conguration and Local Non-Linear Friction. PhD thesis, Katholieke Univer-
siteit Leuven, 2004.

[Til01] M.M. Tiller. Introduction to Physical Modeling with Modelica. Kluwer Academic
Publishers, 2001.

[TOB01] M. Thümmel, M. Otter, and J. Bals. Control of robots with elastic joints based
on automatic generation of inverse dynamics models. In Proc. of the IEEE/RSJ
Int. Conf. on Intelligent Robots and Systems, Maui, Hawaii, USA, October 2001.
[TS85] T. Takagi and M. Sugeno. Fuzzy identication of systems and its applications
to modeling and control. IEEE Trans. Syst., Man, Cybern., 15:116132, 1985.
BIBLIOGRAPHY 245

[VA99] A. Veitl and M. Arnold. Coupled simulation of multibody systems and elastic
structures. In J.A.C. Ambrosio and W.O. Schiehlen, editors, Advances in Com-
putational Multibody Dynamics, pages 635644, IDMEC/IST, Lisbon, Portugal,
September 1999.

[VB96] H. Van Brussel. Mechatronics - a powerful concurrent engineering framework.

IEEE/ASME Trans. Mechatronics, 1(2):127136, 1996.
[VBS02] H. Van Brussel and W. Symens. Vibration and motion control of mechatronic
systems. InProc. of the 2nd Int. Conf. on Advanced Computational Methods in
Engineering (ACOMEN), Liège, Belgium, May 2002.
[VBSN+ 01] H. Van Brussel, P. Sas, I. Németh, P. De Fonseca, and P. Van den Braembussche.
Towards a mechatronic compiler. IEEE/ASME Trans. Mechatronics, 6(1):90
105, 2001.

[VBSV99] M. Valásek, P. Breedveld, Z. Sika, and T. Vampola. Software tools for mecha-
tronic vehicles: Design through modelling and simulation. Vehicle System Dy-
namics, 33:214230, 1999.
[VGVDS+ 99] A. Veitl, T. Gordon, A. Van De Sand, M. Howell, M. Valásek, O. Vaculin, and
P. Steinbauer. Methodologies for coupling simulation models and codes in mecha-
tronic system analysis and design. Vehicle System Dynamics, 33:231243, 1999.
[VNM98] M.J. Van Nieuwstadt and R.M. Murray. Real-time trajectory generation for
dierentially at systems. Int. J. Robust Nonlinear Control, 18(11):9951020,
1998.

[WC00] P. Wenger and D. Chablat. Kinematic analysis of a new parallel machine tool:
the Orthoglide. In Proc. of the 7th International Symposium on Advances in
Robot Kinematics, Slovenia, June 2000.
[WH82] R.A. Wehage and E.J. Haug. Generalized coordinate partitioning for dimension
reduction in analysis of constrained dynamic systems. ASME J. Mech. Des.,
104:247255, 1982.

[Wil86] T.R. Wilson. The design and construction of exible manipulators. Master's
thesis, Georgia Institute of Technology, 1986.

[WV91a] D. Wang and M. Vidyasagar. Control of a class of manipulators with a single

exible link : Part I. feedback linearization. ASME J. Dyn. Syst., Meas., and
Control, 113:655661, 1991.
[WV91b] D. Wang and M. Vidyasagar. Transfer functions for a single exible link. Int. J.
Robotics Res., 10:540549, 1991.
[WW95] E. Wittbrodt and S. Wojciech. Application of rigid nite element method to
dynamic analysis of spatial systems. AIAA J. Guidance, Control, and Dynamics,
18(4):891898, 1995.
246 BIBLIOGRAPHY

[YBS89] B.-S. Yuan, W.J. Book, and B. Siciliano. Direct adaptive control of a one-link
exible arm with tracking. Int. J. Robotic Systems, 6(6):663680, 1989.
[ZBG01] D. Zlatanov, I. Bonev, and C. Gosselin. Constraint singularities as conguration
space singularities, 2001. [Link]

[ZT89] O.C. Zienkiewicz and R.L. Taylor. The nite element method. Vol. I. Basic
formulations and linear problems. McGraw-Hill, London, 4th edition, 1989.