eri

DISTURBANCE COMPENSATION FOR MACHINE

TOOLS WITH LINEAR MOTOR DRIVES

Promotoren: Proefschrift voorgedragen tot

Prof. Dr. ir. H. Van Brussel het bekomen van de graad
Prof. Dr. ir. J. Swevers van Doctor in de
Ingenieurswetenschappen

door
Zamberi JAMALUDIN

2008D09 September 2008

 DISTURBANCE COMPENSATION FOR MACHINE
TOOLS WITH LINEAR MOTOR DRIVES

Jury:
Proefschrift voorgedragen tot
Prof. Dr. ir. A. Haegemans, voorzitter
het bekomen van de graad
Prof. Dr. ir. H. Van Brussel, promotor
van Doctor in de
Prof. Dr. ir. J. Swevers, promotor
Ingenieurswetenschappen
Prof. Dr. ir. J. De Schutter
Prof. Dr. ir. H. Ramon
door
Prof. Dr. ir. P. Sas
Prof. Dr. ir. G. Pritschow (University of Stuttgart) Zamberi JAMALUDIN

2008D09 September 2008

© Katholieke Universiteit Leuven
Faculteit Ingenieurswetenschappen
Arenbergkasteel, B-3001 Heverlee (Leuven), Belgium

Alle rechten voorbehouden. Niets uit deze uitgave mag worden

verveelvoudigd en/of openbaar gemaakt worden door middel van druk,
fotokopie, microfilm, elektronisch of op welke andere wijze ook zonder
voorafgaandelijke schriftelijke toestemming van de uitgever.

All rights reserved. No part of this publication may be reproduced in any

form, by print, photoprint, microfilm or any other means without written
permission from the publisher.

D/2008/7515/85
ISBN 978-90-5682-975-9
U.D.C. 681.5
My Parents

My Wife
 The beginning of knowledge is the intention, then listening, then
understanding, then action, then preservation, and then spreading it.
Acknowledgements

This journey begins in 2003. The common question at that time is “Why
Belgium?” This is apparent especially because of the weak link in the
educational relationship between Malaysia and Belgium. A nice and
encouraging introduction by Dr. Indra Tanaya (a graduate of K.U.Leuven
and who was once my colleague in Malacca) about PMA, its strong lists of
academicians and researchers with dedicated research activities and the
beautiful and peaceful city of Leuven have motivate me enough to begin this
new chapter of my life here in PMA, Leuven. I am grateful that I have made
this important choice because after these five wonderful and challenging
years, both PMA and Leuven have never failed to contribute positively to
my academic and personal development.

There are two very important persons that have been very influential to my
academic development here in Leuven. First and foremost, I would like to
express my deepest gratitude to Prof. Van Brussel for accepting me into the
PMA community and for the trust that he has put on me in realising this
work. It is a great honour and pleasure to be able to work with him, a person
of great mechatronics background, experience, success, ideas, and stature. It
is certainly be a great challenge to me to emulate your success and to be as
good as you are. Your kind advice, time, attention, and dedication towards
realising this work are greatly appreciated. Finally, I would like to extend my
deepest appreciation for your kind attentions on the well-being of myself
(visiting me in the hospital on just my third week in Belgium) and my wife
and your warm hospitality towards my parents.

Secondly, I would like to express my special gratitude to Prof. Swevers – the

control specialist, for nurturing me into the challenging world of control.
Your knowledge in advanced control theory certainly amazed me and I wish
that there is somebody of your calibre in Malacca (there is none at the
moment) so that I can keep learning from the best mind in this field. Your
guidance, attention, and time are greatly appreciated. I would like to thank
you, your wife and family for being such a great and wonderful host during
your yearly summer bbq sessions.

I would like to extend my deepest gratitude to the thesis jury committee

members; Prof. Dr ir. Herman Ramon, Prof. Dr ir. Joris De Schutter, Prof.
Dr ir. Paul Sas, and Prof Dr ir. Gunther Pritshow of the Stuttgart University.
Thank you for your valuable time, suggestions, comments, and critical
i
Acknowledgements

thoughts that have enable me to improve the quality of this manuscript. I

would also like to express my thanks to Prof. Dr Ann Haegemans for being
the chairperson.

I admired the great knowledge and experience of Prof. Vanherck, a person

whom I regard as a walking technical library. Your help, assistance,
teachings, and explanation on many technical issues have helped me
tremendously throughout my work in PMA. Thank you for your dreams (I
will always remember you saying this when there are problems with the xy
table and the Ferraris sensor).

I would like to extend my special thanks to the administrators of PMA - Luc

Haine, Lieve Notre, Karin Dewit, Carine Coosemans, and Ann Letelier for
their support, assistance and warm hearts. I am also grateful to all the
technical personnel at the workshop, the electronic and the IT department for
making my life easier at the workshop. My special thanks to Dirk
Bastiaensen, Eddy Smets, Paul Van Cauwenbergh, Luc De Simpelaere,
Bertram Van Soom, Jean Pierre Merkcx, Ronny Moreas, and Jan
Thielemans.

My life here at PMA is made easier by Dr T. Tjahjowidodo who has given

me a nice introduction to life at PMA, its surrounding and its research
environment. I still remember the confusions and funs that we had during my
early days in PMA trying to understand each other in Bahasa Melayu (some
similar Malaysian and Indonesian words can have totally different
meanings). I would also like to express my appreciation for your time and
assistance in explaining the world of friction to me. The nice explanations
have certainly helped in my work on friction compensation.

It is a great pleasure to be in the environment of the control group people and

of my office colleagues – Bram, Christophe, David Vaes, Bart Paijmans,
Kris Smolders, Goele, Jan De Caigny, Dimitri, Lieboud, Diederick, Bert
Stalleart, Myriam, Leopoldo, Marnix and Maarten. Thank you for the
company, assistance, and fruitful discussions. I would like to express my
special thanks to Goele Pipeleers for assisting me in the design of the
repetitive controller.

The mechatronics group meetings have exposed me to many wonderful

people. It is a pleasure to be in the company of these distinguished
colleagues of various backgrounds and origins - Prof. Farid Al-Bender, Wim
Symens, Brecht, Hsiao-Wei.Tang, Emmanuel Vander Poorten, Wim van de
ii
 Acknowledgements

Vijver, Tri, Gorka, Mohamed El-Said (also Neny and Zaharah for your
kindness and warm hearts), Maira, Thierry, Kris, Pauwel, Bert, and
Agusmian.

Life in Belgium for a Malaysian can be a lonely experience (there are not
many of us here). However, since the last two years, the students community
has grown with the arrival of other Malaysian students and families
(although there are all not in Leuven). I would like to extend my sincere
appreciation to: Zaini and Ros, Saifullah, Yeong and family, Helmi and
Natrah, Am and Nurul, Dr Razak and family, abang Zam, kak Zu and family,
and Kelvin and family. Thank you for the friendship and hospitalities. I
would also like to thank Akmal and Zura for their warm heart and hospitality
towards me and my wife and for all the trips that you have taken us to.
Finally, to the Embassy of Malaysia in Brussels, thank you for all the nice
receptions and your kind treatment of myself and my wife for all the five
years that we are in Belgium.

I am also grateful to have known many kind and warm hearted Indonesian
friends and families especially Pak Oemar, Mbak Leen and Esti, Gandjar and
family, Tegoeh, Ira and family, Singgih, Uly and the kids, Arie and Sarah,
and Freddy, Novita and family.

This work is made possible with the financial supports of the Ministry of
Higher Education Malaysia (SLAB) and the University Technical of
Malaysia-Malacca (UTeM). These financial supports are greatly appreciated
and indebted. I also wish to express my sincere gratitude towards the
International Student Office of K.U.Leuven for the financial support that I
have received during the final three months of my stay in Leuven. This
financial support has enabled me to complete my study here uninterruptedly
and successfully.

Finally and most importantly, I wish to express my deepest gratitude to both

my parents, Hj. Jamaludin and Hjh Khatijah Bee, my brother Saifullah, and
the families in Kedah and Singapore for their prayers, loves, cares, and
support. I am greatly indebted to them and may God repay all their deeds
and sacrifices. To my wife Zila, there are no words to describe how grateful I
am for your company during both the easy and the difficult times of this
journey. I salute your patient, dedication, love, and care that have brought
happiness and serenity to our life.
Zamberi Jamaludin
Leuven, Sept. 2008.
iii
iv
Abstract

Machining processes have evolved significantly over time in order to adapt

to the increasing demand for speed, accuracy, and efficiency. This evolution
or paradigm shift has created new and highly critical challenges. This thesis
aims at addressing some of these issues, namely the compensation of the
effect of friction and cutting forces on the accuracy of the machining
process.

Issues regarding friction effects in machining process have been studied

extensively in the past and various techniques and ideas have been proposed
and validated. Simple linear feedback control techniques such as PI, PID, or
cascade control alone are insufficient to compensate the nonlinear friction
behaviour. In this thesis compensating elements are proposed, developed and
validated that can be added to existing feedback controllers in order to
improve accuracy. Friction-model-based and friction-model-free approaches
are considered to supplement the cascade P/PI position controller. The
compensation performance is measured based on the magnitude of the
“quadrant glitch” – a product of highly nonlinear complex friction behaviour
near zero velocity or motion reversal identified by the appearance of
“spikes” at each quadrant of a circle. In this thesis, the recently developed
Generalized Maxwell-slip (GMS) friction model is used as a feedforward
element in combination with the well-known and widely applied inverse-
model-based disturbance observer for friction compensation on a linear drive
based xy feed table of a high-speed milling machine. This combined
approach almost completely compensates all friction effect.

Besides friction forces, the effect of cutting forces on machining accuracy is

significant. Several techniques described in literature are studied and their
applicability to compensate cutting forces in machining process is evaluated.
First, the application of the inverse-model-based disturbance observer is
further extended to cutting force compensation. However, its performance is
critically influenced by the limited bandwidth of a low pass filter often
referred to as the Q-filter that is necessary to preserve the system stability.
Second, cutting forces are estimated from the force balance acting on the
drive using a Ferraris relative acceleration sensor measurement. The
bandwidth is again restricted by a stability preserving low-pass filter, similar
as for the inverse-model based disturbance observer.

v
Abstract

Finally, a method that is renown for its excellent compensation of periodic

disturbance signals is applied, namely, the repetitive controller (RC). A
repetitive controller is developed for the considered linear drive based xy
feed table.

To validate the performance of this RC, an actual cutting process is

performed on the test setup. It is shown that the developed RC is able to
compensate almost completely the tracking errors introduced by the cutting
forces. The repetitive controller, when combined with the previous friction
compensation elements such as the GMS friction model feedfoward and the
disturbance observer, almost completely removed the cutting forces during
an actual cutting process.

This thesis has successfully demonstrated that the tracking performance of a

machine tool can be increased significantly by adding dedicated
compensation elements to the simple and widely used cascade P/PI position
controller. However, further studies are desired to include adaptive measures
in both friction and cutting forces compensation using the advanced GMS
friction model and the RC. This will ensure a robust friction compensation
approach to changing friction behaviour and characteristics over time due to
the influence of lubrication, heating and etc. An adaptive RC will
compensate against changes in the cutting conditions, for example, changes
in the spindle speed, tools diameter, tracking speed, and etc.

vi
Beknopte Samenvatting

De technologische evolutie van conventionele materiaalbewerkingsmachines

wordt gedreven door de vraag naar steeds hogere snelheden,
nauwkeurigheden en efficiëntie. Deze evolutie creëert nieuwe uitdagingen.
Deze thesis richt zich op enkele van deze uitdagingen, namelijk de actieve
compensatie van de effecten van wrijving en snijkrachten op de
nauwkeurigheid van het bewerkingsproces.

Het effect van wrijving op de nauwkeurigheid van bewerkingsprocessen

werd in het verleden reeds uitgebreid onderzocht. Verschillende technieken
om het effect van wrijving te compenseren werden ontwikkeld en
experimenteel gevalideerd. Eenvoudige lineaire regelaars op basis van
terugkoppeling, zoals de PI-, PID-, en de cascaderegelaars voldoen niet om
het effect van het niet-lineaire wrijvingsgedrag te compenseren.

Deze thesis stelt een aantal wrijvingscompensatietechnieken voor die kunnen

toegevoegd worden aan bestaande (lineaire) sturingen. Zowel technieken
gebaseerd op een wrijvingsmodel als technieken die geen wrijvingsmodel
vereisen, worden beschouwd.

De performantie van deze technieken wordt gemeten aan de hand van de

grootte van de “quadrant glitch”, het resultaat van het sterk niet-lineair
gedrag van wrijving dat optreedt bij omkering van de bewegingrichting en
bij zeer lage bewegingssnelheden, en dat zich uit als spikes in de volgfout.
De compensatietechnieken die werden onderzocht worden ook
experimenteel gevalideerd op een met lineaire motoren aangedreven XY
voedingstafel van een hoge-snelheidsfreesmachine. De combinatie van
voorwaartse koppeling aan de hand van het recent ontwikkelde “Generalized
Maxwell-slip (GMS)” wrijvingsmodel en de gekende en vaak toegepaste
storingsschatter op basis van een invers systeemmodel resulteert in de
grootste verbetering van de volgnauwkeurigheid: wrijvingseffecten bij elke
kwadrant van een circulaire baan worden bijna volledig gecompenseerd.

Naast wrijving kunnen ook de snijkrachten een belangrijke invloed hebben

op de bewerkingsnauwkeurigheid van de machine. Verschillende
compensatietechnieken beschreven in de literatuur worden bestudeerd en
hun toepasbaarheid om snijkrachten te compenseren wordt onderzocht en
vii
Beknopte Samenvetting

experimenteel gevalideerd. Eerst wordt de storingsschatter op basis van een

invers systeemmodel verder uitgebreid voor snijkrachtcompensatie. De
performantie van deze storingsschatter wordt kritisch beïnvloed door de
beperkte bandbreedte van de laagdoorlaatfilter (Q-filter) die aanwezig is in
deze schatter om de systeemstabiliteit te vrijwaren.

Daarnaast wordt een snijkrachtschatter bestudeerd die gebruik maakt van

metingen van de relatieve versnelling van de lineaire motor op basis van een
Ferraris sensor. Ook voor deze schatter gelden dezelfde beperkingen met
betrekking tot bandbreedte en performantie.

Tenslotte wordt de repetitieve regelaar bestudeerd. Deze regelaar richt zich

uitsluitend op periodische storingen waarvan de periode gekend is of kan
geschat worden. Deze regelaar maakt gebruik van niet-causale filters
waardoor een aanzienlijk hogere bandbreedte kan bereikt worden. Een
repetitieve regelaar wordt ontworpen en toegepast op de beschouwde
testopstelling.

Om de performantie van de ontwikkelde repetitieve regelaar te valideren

wordt een freesbewerking uitgevoerd op de testopstelling. De testen tonen
aan dat de ontworpen repetitieve regelaar in staat is om volgfout ten gevolge
van het freesproces bijna volledig te elimineren. Deze repetitieve regelaar
wordt toegepast in combinatie met de ontwikkelde wrijvingscompensatie op
basis van het GMS model en de storingsonderdrukker zodat ook wrijving
tijdens deze bewerking gecompenseerd wordt. De overblijvende volgfout is
verwaarloosbaar.

Deze thesis heeft met succes aangetoond dat de volgnauwkeurigheid van een
werktuigmachine significant kan verbeterd worden door gerichte
compensatietechnieken toe te voegen aan de eenvoudige klassieke cascade
P/PI positiesturing van de machine. Echter, verder onderzoek is vereist om
deze compensatie adaptief te maken om zo efficiënt te kunnen inspelen op
veranderingen in de tijd van de wrijvingskarakteristieken, die beïnvloed
worden door smering, temperatuur, en slijtage, en van de snijkarakteristieken,
die o.a. beïnvloed worden door de spilsnelheid en de freesdiameter.

viii
Symbols and Abbreviations

Symbols
Control:

d(t) disturbance force signal [N]

~
d (t ) estimated disturbance force [N]
ep(t) position tracking error signal [µm]
ev(t) velocity tracking error signal [µm/s]
f frequency [Hz]
γp,∆ robust periodic performance index
γnp Non-periodic performance index
kf motor force constant [N/V]
kp proportional gain controller in velocity loop [volt·s/µm]
ki integral gain controller in velocity loop [volt/µm]
kv velocity gain factor, or the Proportional gain
[1/s]
controller in cascade position loop
n(t) noise
u(t) control command signal [volt]
vref desired reference velocity [µm/s]
z(t) Output position signal [µm]
zref(t) Desired reference position signal [µm]
ωo Undamped natural frequency [rad/s]
ζ damping ratio
F Force acting on linear motor [N]
G considered system
Ĝ FRF of the considered system
Gm system model transfer function
Gn nominal model system transfer function
Gm’ system model with GMS friction term
M mass of the sliding table [kg]
N(s) notch filter
S sensitivity function
T complementary sensitivity function
T0 sampling period [s]
Td time delay [s]
Vest(s) estimated velocity signal [µm/s]

ix
Symbols & Abbreviations

Friction:

δ Stribeck function shape factor

σ viscous friction force [N·s]
αi elementary normalized friction force [N]
ki elementary spring constant of the Maxwell slip [N/m]
blocks element
s(v) Stribeck function
Ff total friction force [N]
Fc Coulomb friction force [N]
Fs static friction force [N]
Vs Stribeck velocity [µm/s]
Wi maximum elementary friction force [N]

Abbreviations
rms root mean square
rpm revolution per minute
DOB disturbance observer
FRF frequency response function
FF feedforward
GMS Generalized Maxwell-slip
H∞ H-infinity
I/O input-output
PI proportional plus integrator
PID proportional plus integrator plus differentiator
RC repetitive controller
RHP right half plane
SISO single input single output
SMC sliding mode control

x
Table of Contents
Acknowledgements i
Abstract v
Beknopte Samenvetting vii
Symbols and Abbreviations ix
Table of Contents xi
List of Figures xv
List of Tables xxiii

1 Introduction 1
1.1 Motivation ……………………………………………... 1
1.2 State of the Art on Motion Control ……………………. 3
1.2.1 Mechanical Drive Systems ……………………. 3
1.2.2 Disturbance Forces and Compensation Methods 6
1.3 Scope, Objective and Approaches ………….…………. 9
1.4 Contributions …………………………………………... 10
1.5 Outlines ………………………………………………... 11

2 Classical Motion Control 13

2.1 Introduction ……………………………………………. 13
2.2 High-Speed XY Milling Machine ……………………... 14
2.3 System Identification ………………………………….. 17
2.4 Cascade Control Structure and Analysis …………….. 20
2.4.1 Cascade Controller Structure and Configuration 20
2.4.2 Analysis of the Closed Loops Behaviour with
Proportional (P) Velocity Loop ……………….. 22

xi
Table of Contents

2.4.3 Analysis of the Closed Loops Behaviour with

Proportional (P) and Integrator (I) Velocity
Loop …………………….……………………... 25
2.5 Design and Validation of Cascade P/PI Controller
Based on Measured FRFs ……………………….….…. 27
2.5.1 Design and Analysis of the Velocity Loop...….. 27
2.5.2 Design and Analysis of the Position Loop ….… 34
2.5.3 Tracking Performance Numerical Validation … 39
2.5.4 Cascade P/PI with Feedforward …………...….. 40
2.6 Correlation between Velocity and Position Closed Loop
Characteristics …………………………………………. 41
2.6 Summary ………………………………………………. 44

3 Disturbance Forces in Servo Drives System 45

3.1 Introduction ………………………………………….… 45
3.2 Friction Characterization and Model Structures ………. 45
3.3 The Dahl Model ……………………………………….. 46
3.4 The LuGre Model ……………………………………… 47
3.5 Static Friction Model ……………………………….…. 48
3.2.1 Model Structure ………………………….…..... 48
3.2.2 Identification of Static Friction Model ………... 49
3.6 Generalized Maxwell-slip Model (GMS) ……………... 50
3.2.1 Model Structure ……………………………….. 50
3.2.2 Identification of GMS friction model …………. 52
3.7 Cutting Forces …………………………………………. 56
3.7.1 Cutting Force Measurements and Analysis ….... 56
3.7.2 Artificial Cutting Force ……………….………. 60
3.8 Summary ………………………………………………. 62

xii
 Table of Contents

4 Friction Forces Compensation 63

4.1 Introduction ……………………………………………. 63
4.2 Friction Model-Based Feedforward ………………...…. 64
4.2.1 System Transfer Function with a GMS Friction
Term for Friction Simulation ……………… 64
4.2.2 Friction Feedforward: Numerical Results …… 67
4.2.3 Friction Feedforward: Experimental Results 69
4.3 Inverse-Model Based Disturbance Observer ……...…... 73
4.3.1 Q-filter Design and Stability Analysis ……...… 75
4.3.2 Loops Characteristic with Disturbance Observer 78
4.3.3 Numerical Validations……………………......... 87
4.3.4 Experimental Validations………...……………. 90
4.4 Summary ………………………………………………. 94

5 Cutting Force Compensation 97

5.1 Introduction ………………………………………….… 97
5.2 Inverse-Model-Based Disturbance Observer ………….. 98
5.2.1 Numerical Validations…...…………………….. 98
5.2.2 Experimental Validations...…….……...………. 100
5.3 Explicit Estimation of Cutting Force using Ferraris
Relative Acceleration Sensor Measurements ………….. 103
5.3.1 The Ferraris Principle ……………….………… 104
5.3.2 Cutting Force Estimator ………………………. 105
5.3.3 Experimental Validations...……………………. 106
5.4 Repetitive Controller …………………………………... 112
5.4.1 Design Structure of a Repetitive Controller…… 112
5.4.2 Experimental Validations...……………………. 116

xiii
Table of Contents

5.5 Cutting Force Compensation during Actual Cutting

Process ……………………………………………….... 120
5.6 Summary ……………………….……………………… 123

6 Conclusions and Future Studies 125

7 Bibliography 129

8 Curriculum Vitae 135

9 List of Publications 137

10 Appendices 139
A System Dynamic Analysis 139
B Calibration of Ferraris Sensor 140
C Cutting Force Estimation from Ferraris Acceleration
Sensor Measurement 143
D General State Force Observer Design 147

xiv
List of Figures

Chapter 1
1.1 Quadrant glitches in circular test ………………………... 2
1.2 Electromechanical ball-screw drive structure…………… 4
1.3 Structure of an iron-core linear motor…………………… 5
1.4 Schematic diagram for friction compensation techniques. 6

Chapter 2
2.1 A linear-driven xy feed table of a high-speed milling
machine …………………………………………………. 14
2.2 Schematic diagram of the xy table ….………………….. 15
2.3 Motion controller structure of a xy feed table with three
linear drives for high speed milling application…………. 17
2.4 FRFs measurement of the x and y axes………………….. 18
2.5 X-axis: FRF measurement and proposed model………… 19
2.6 Y-axis: FRF measurement and proposed model………… 19
2.7 General scheme of a cascade control structure………….. 20
2.8 Schematic diagram of the open loop system for force
constant estimation………………………………………. 21
2.9 An ideal cascade control structure for a linear motor
position control…………………………………………... 22
2.10 Frequency domain scheme of the cascade P/PI controller
for control of a linear motor drive……………………….. 28
2.11 Bode diagram of Vest(s)………………………………….. 28
2.12 X-axis: Bode diagram of a notch filter N(s) for frequency
between 320 Hz and 340 Hz…………………………….. 29
2.13 X-axis: Theoretical bode of the velocity open loop
transfer function based on measured FRF of the system... 31

xv
List of Figures

2.14 Y-axis: X-axis: Theoretical bode of the velocity open

loop transfer function based on measured FRF of the
system……………………………………………………. 31
2.15 Nyquist plots of the x-axis (left) and the y-axis (right)
velocity open loop transfer functions based on measured
FRFs of the system………………………………………. 32
2.16 Sensitivity function of x (left) and y (right) axes for the
velocity loop……………………………………………... 33
2.17 Velocity closed loop transfer function for (a) x-axis and
(b) y-axis ………………………………………………... 34
2.18 Position open loop transfer function for (a) x-axis and (b)
y-axis based on measured FRF of the system ………...… 36
2.19 Nyquist plots of the x-axis (left) and the y-axis (right) of
the position open loop transfer functions………………... 37
2.20 Sensitivity function of x (left) and y (right) axes for the
position loop……………………………………………... 37
2.21 Position closed loop transfer functions for x and y axes… 38
2.22 Measured dynamic stiffness of x (left) and y (right) axes
using cascade P/P and cascade P/PI controller………….. 39
2.23 Simulated tracking error of the x-axis (left) and the y-
axis (right) for sinusoidal reference signal of an
amplitude of 1mm and frequency of 10Hz……………… 39
2.24 Position error transfer functions of the x-axis (left) and
the y-axis (right)…………………………………………. 40
2.25 Cascade P/PI with feedforward for improved tracking
performance……………………………………………… 40
2.26 Simulated tracking error of the x-axis (left) and the y-
axis (right) for sinusoidal reference signal of amplitude
1mm and frequency of 10Hz…………………………….. 41
2.27 Schematic diagram of a velocity closed loop……………. 42
2.28 Y-axis: Simulated speed closed loop based on measured
system FRF ………............................................................ 43

xvi
 List of Figures

2.29 Measured FRF of the y-axis position closed loop with

kv =150 s-1………………………………………………... 44

Chapter 3
3.1 Pre-sliding and sliding friction regimes…………………. 46
3.2 Friction components in static friction model (3.5)………. 48
3.3 Friction force from control command signals at constant
velocity motion of 2.0 mm/s (y-axis)……………………. 49
3.4 Friction force-velocity mapping and the manually fitted
static friction model using equation (3.5) ……………..... 50
3.5 N-elementary configuration of the Maxwell-slip structure 51
3.6 Y-axis: Friction force and position for sinusoidal
reference signal of 0.1Hz and amplitudes of (a) 15µm
and (b) 450µm………………………………………….... 53
3.7 Y-axis: (a) virgin curve and (b) virgin curve with
selected knots and slopes………………………………... 53
3.8 Y-axis GMS model parameters identification…………… 54
3.9 X-axis: Friction force and position for sinusoidal
reference signal of 0.1Hz and amplitudes of (a) 15µm
and (b) 450µm…………………………………………… 55
3.10 X-axis: (a) virgin curve (b) virgin curve with selected
knots and slopes (right)………………………………….. 55
3.11 X-axis: GMS model parameters identification………….. 56
3.12 Cutting force in x and y direction for cutting depth 1mm.. 57
3.13 Cutting force in x and y direction for cutting depth of
3mm …………………………………………………….. 58
3.14 Cutting force measurements on a work-piece using
Kistler dynamometer force sensor along the x-axis……... 58
3.15 Spectral of cutting force along x and y direction, for
cutting depth of 1mm……………………………………. 59
3.16 Spectral of cutting force along x and y direction, for
cutting depth of 3mm……………………………………. 59

xvii
List of Figures

3.17 X-axis: Comparison between the measured cutting force

and the synthesized cutting force for 1mm depth of cut… 62

Chapter 4
4.1 Friction compensation scheme using friction model-
based feedforward (kf is the force constant)……………... 64
4.2 System transfer function with a GMS friction model to
characterize the complex friction behaviour…………….. 65
4.3 System model parameters adjustment for system
structure that includes the GMS friction term: (a) using
the original model parameters and (b) using updated
model parameters………………………………………... 66
4.4 Y-axis: Simulated position, velocity and tracking error
for three different cases of friction compensation
techniques (a) no friction feedforward, (b) Stribeck
friction model feedforward, (c) GMS model feedforward. 67
4.5 Y-axis: Simulated tracking error using GMS model
feedforward for (a) system model with delay and (b) with
the delay removed.…………….………………………… 68
4.6 Y-axis: Simulated tracking error using (a) modified GMS
model feedforward and (b) matched GMS model for
system both without delay……………………………….. 69
4.7 Y-axis: Measured position and tracking error for (a) no
friction feedforward, (b) static friction model
feedforward, (c) GMS feedforward …...………………… 70
4.8 X-axis: Measured quadrant glitches for maximum
tracking velocity of 10mm/s: (a) static friction model
feedforward and (b) GMS model feedforward…………... 71
4.9 Y-axis: Measured quadrant glitches for maximum
tracking velocity of 10mm/s: (a) static friction model
feedforward and (b) GMS model feedforward…………... 71
4.10 Measured contour and radial tracking error at tangential
tracking velocity of 100mm/s: (a) no friction
feedforward, (b) static friction model feedforward, (c)
GMS model feedforward………………………………… 72
xviii
 List of Figures

4.11 Measured contour and radial tracking errors for

tangential tracking velocity of 10mm/s: (a) static friction
model feedforward and (b) GMS model feedforward…… 73
4.12 Block diagram of a system with an inverse-model- based
disturbance observer……………………………………... 74
4.13 Equivalent block diagram of a system with an inverse
model-based disturbance observer………………………. 74
4.14 Bandwidth limitation of the filter Q for (a) x-axis and (b)
y-axis…………………………………………………….. 77
4.15 Cascade P/PI position control with inverse model-based
disturbance observer …...................................................... 78
4.16 Cascade P/PI position control with equivalent block
diagram of the inverse model-based disturbance observer 78
4.17 Velocity open loop transfer function for (a) x-axis and
(b) y-axis for system with and without a disturbance
observer………………………………………………….. 80
4.18 Velocity closed loop transfer functions for (a) x-axis and
(b) y-axis for system with and without a disturbance
observer………………………………………………….. 81
4.19 Effect of the disturbance observer on the sensitivity
function of the velocity loop for x-axis (left) and y-axis
(right) …………………………………………….……… 82
4.20 Nyquist plots of the velocity loops with and without the
disturbance observer……………………………………... 82
4.21 Position open loop transfer function for (a) x-axis and (b)
y-axis for system with and without a disturbance
observer………………………………………………….. 84
4.22 Position closed loop transfer functions for (a) x-axis and
(b) y-axis for system with and without a disturbance
observer………………………………………………….. 85
4.23 Effect of the disturbance observer on the sensitivity
function of the position loop for the (a) x-axis and (b) y-
axis …..………………………………………………….. 86

xix
List of Figures

4.24 Nyquist plots of the position loops with and without the
disturbance observer……………………………………... 87
4.25 Simulated control scheme for friction compensation
using friction model feedforward and a disturbance
observer………………………………………………….. 88
4.26 Y-axis: Simulated position and tracking error for
configuration (a), (b) and (c)…………………………….. 89
4.27 X-axis: Measured position and tracking errors for
configuration (a), (b) and (c)…………………………….. 90
4.28 X-axis: Measured position and tracking errors for slower
reference trajectory of 10mm/s for configuration (a) and
(b)………………………………………………………... 91
4.29 Y-axis: Measured position and tracking errors for
configuration (a), (b) and (c)…………………………….. 92
4.30 Measured contours and radial errors for configuration
(a), (b) and (c)…………………………………………… 93
4.31 Circular tracking tests at tangential tracking velocity of
10mm/s…………………………………………………... 93
4.32 Position and radial tracking error for different friction
compensation approaches………………………………... 94

Chapter 5
5.1 MATLAB/Simulink diagram of a cascade P/PI controller
and a disturbance observer with a sinusoidal based
disturbance input signal………………………………….. 98
5.2 Y-axis: The effect of a disturbance observer on tracking
errors for disturbance input frequencies of (i) 1 Hz, (ii)
58 Hz, and (iii) 100 Hz…………………………………... 99
5.3 Y-axis: Position errors for system with and without a
disturbance observer……………………………………... 100
5.4 Cutting force compensation using inverse-model-based
disturbance observer……………………………………... 101

xx
 List of Figures

5.5 Y-axis: Measured position errors with inverse model-

based disturbance observer for synthesized cutting forces
disturbance with different harmonics component……….. 101
5.6 Y-axis: Spectral analysis of the position errors without
the disturbance observer (top) and with the disturbance
observer (bottom) for disturbance force signals with
different harmonic contents……………………………… 102
5.7 Measured contours and radial errors for circular tests
with (right) and without (left) the inverse model based
disturbance observer …………………………………….. 103
5.8 Schematic diagram of the Ferraris principle and the
actual sensor used in measurement……………………… 104
5.9 Schematic diagram of a cascade P/PI position control
with cutting force estimator (without feedback)………… 107
5.10 Reference and estimated cutting forces…………………. 107
5.11 Schematic diagram of a cascade P/PI position control
with cutting force estimator (with feedback)……………. 108
5.12 Position errors for different cutting force harmonic
components using estimator…………………………….. 108
5.13 Spectral analysis of the position errors for different
cutting force harmonic components using estimator……. 109
5.14 Tracking errors for different cutting force harmonic
contents for cases with and without the estimator………. 110
5.15 Spectral analysis of the tracking errors for different
cutting force harmonic components for cases with and
without the estimator…………………………………….. 110
5.16 Cutting Forces Estimation Analysis (without reference
trajectory)………………………………………………... 111
5.17 Standard RC as an add-on module to a closed loop
control scheme ……………………………..…………… 113
5.18 Standard memory loop with periodic signal generator
with period T0[s] …………………………….…………... 113
5.19 Discrete time implementation of a typical RC…………... 114

xxi
List of Figures

5.20 Schematic diagram of (a) nth order RC, (b) 2nd order
RC …………………………………………………….. 115
5.21 Y-axis: Bode plots of (a) the loop gain, (b) sensitivity
function and (c) complementary sensitivity function of
the second order RC …………………………………….. 116
5.22 A schematic diagram of a cascade P/PI controller with a
RC module and friction compensation elements ...…….. 117
5.23 Y-axis: Measured position tracking errors with and
without the RC for different harmonic component of the
cutting forces…………………………………………….. 118
5.24 X-axis: Measured position tracking errors with and
without the RC for different harmonic components of the
cutting forces. …………………………………….……... 118
5.25 Circle tests: Measured position and radial tracking errors
for cases (i), (ii), and (iii) ………………..……………… 119
5.26 Cutting force compensation during actual cutting process 121
5.27 Measured position tracking errors and spectral analyses
for control configurations (i), (ii), and (iii) ...…………… 122
5.28 Contour tracking measurement and radial tracking errors
for control configurations (i), (ii), and (iii) …………… 122

xxii
List of Tables
Chapter 2
2.1 System model parameters for x and y axes. ……………. 19
2.2 Velocity loop PI controller parameters of x and y axes … 30
Gain margin and phase margin of x and y axes velocity
2.3
open loop………………………………………………… 32
2.4 Bandwidths of the velocity loop………………………… 33
Gain margin and phase margin of x and y axes position
2.5
open loop………………………………………………… 35
2.6 Bandwidths of the position loop………………………… 37

Chapter 3
3.1 Static friction model parameters for x and y axes……….. 49
3.2 GMS slip-blocks model parameters for the y-axis………. 54
3.3 GMS slip-blocks model parameters for the x-axis………. 55
3.4 Cutting process parameters……………………………… 57
3.5 X-axis: Harmonic contents of the measured cutting force
for 1mm depth of cut……………………………………. 60
3.6 Y-axis: Harmonic contents of the measured cutting force
for 1mm depth of cut……………………………………. 61

Chapter 4
4.1 Original and revised system transfer function parameters. 65
4.2 Simulated friction compensation performance for
reference sinusoidal signal with an amplitude of 30mm
and a maximum tracking velocity of 100mm/s………….. 68
4.3a Y-axis: Original GMS slip-blocks model parameters…… 69
4.3b Y-axis: Adjusted GMS slip-blocks model parameters…... 69

xxiii
List of Tables

4.4 Measured friction compensation performance for

reference sinusoidal signal with an amplitude of 30mm
and a maximum tracking velocity of 100mm/s………….. 70
4.5 Characteristics of the low pass filter Q…………………... 77
Gain margin and phase margin of x and y axes velocity
4.6
open loop………………………………………………… 80
Gain margin and phase margin of x and y axes position
4.7
open loop………………………………………………… 83
4.8 Bandwidth of x and y axes position loops………………. 86
4.9 Simulated friction compensation performance for
reference sinusoidal signal with an amplitude of 30mm
and a maximum tracking velocity of 100mm/s………….. 89
4.10 Measured friction compensation performance for
reference sinusoidal signal with an amplitude of 30mm
and a maximum tracking velocity of 100mm/s………….. 92
4.11 Magnitude of the quadrant glitches for reference
sinusoidal signal with an amplitude of 30mm and a
maximum tracking velocity of 100mm/s………………... 95
Chapter 5
5.1 rms of position errors for different cutting force
harmonics content using inverse-model-based
disturbance observer ……………………………………. 102
5.2 rms of position errors for different cutting force
harmonics rms of position errors for different cutting
force harmonics content using the explicit estimation of
cutting force technique…………………………………... 109
5.3 rms of tracking errors for different cutting force
harmonics content using the explicit estimation of cutting
force technique…………………………………………... 111
5.4 rms of the radial tracking errors for cases (i), (ii),and (iii) 119
5.5 Cutting test characteristics used in cutting force
compensation…………………………………………….. 120
5.6 rms of the radial tracking errors for cases (i), (ii), and (iii)
during actual cutting process 123

xxiv
 Chapter 1

Introduction

1.1 Motivation
The constant demand for higher speed and accuracy in machine tools
stimulates the development of machine tool technology and design
methodologies. An integrated part of this technology and design is the
machine tool controller. A coordinated and concurrent development of the
different technology fields and a good knowledge and understanding of the
factors that contribute to the machine speed and accuracy are essential. One
of the factors that contribute to the accuracy of a machine tool is the tracking
performance of its drive system, which is critically influenced by the
following factors:

• The Mechanical structure can limit the system tracking

performance. Mechanical resonances influence the dynamic and
frequency response function (FRF) of a system. The resonances
can be excited during motion and can limit the potential
bandwidth and reduce the stability margins of the control system.
The excitation of the system resonances is often accompanied by
mechanical vibration of the moving structure which can then
influence the tracking accuracy. This vibration can be damped
with a good control design but this will result in a more complex
control structure. The influence of mechanical structural vibration
and mechanical resonance on positioning and tracking
performances can be limited by a well-balanced and integrated
mechanical design and good control design strategies – a
mechatronics approach.

1
Introduction

• Friction forces, a nonlinear phenomenon that originates from the

motor and the support bearings of electromechanical servo drive
systems. Friction is an undesired nonlinear phenomenon in
electromechanical servo-drives systems. Friction compensation is
therefore a prerequisite for accurate motion control applications.
A distinct phenomenon that is caused by friction is “spikes” at
quadrant locations during circular motion (see fig. 1.1). These are
widely known as “quadrant glitches”. Quadrant glitches in
circular motion [1] are the product of complex nonlinear
behaviour of friction at motion reversal or near zero velocity on
each axis of a motion system.
Circle, 15mm radius; Controller: Cascade
15

10
position y [mm]

-5

-10

-15
radial error x 90
-15 -10 -5 0 5 10 15
position x [mm]

Fig. 1.1 Quadrant glitches in circular test

• Cutting forces, in addition to friction force, act as an input

disturbance on the motion control system during cutting
processes. The cutting force disturbance signal, which is periodic,
reduces the tracking performance of a system.

• The work-piece mass can vary from a fraction of a table mass to

several times the table mass. Robust controllers are needed to
make the machine tool performing well for all these masses. This
issue is not considered in this thesis, but has been treated
extensively in [2].

Active and efficient compensation of both friction and cutting forces are
required to achieve accurate tracking and positioning. This thesis focuses on
the development and experimental validation of control techniques to
2
Introduction

actively compensate the influence of friction at motion reversal and cutting

forces on the tracking error of drive systems. Both friction feedforward
compensation based on the advanced nonlinear Generalized Maxwell Slip
(GMS) friction model and various simple and advanced linear control
techniques are added to a conventional machine tool controller.
Compensation techniques that are highly practical and simple in their
application are desired.

1.2 State of the Art on Motion Control

Mechanical drive systems have evolved to meet the present demands for
high speed (i.e. shorter transient response time) and high accuracy
applications. This evolution however has created a new challenge to the
control community with regard to the complexity in effectively rejecting
disturbance forces and achieving the best possible tracking accuracy. This
section presents a discussion of the paradigm shift in mechatronic drive
system technology and a literature review on disturbance forces and their
compensation techniques.

1.2.1 Mechanical Drive Systems

Mechanical drives system technology has seen a shift from
electromechanical drive systems to direct drive systems. The emergence of
direct drive systems over conventional electromechanical drive systems has
provided the industry with their speed and high tracking performance
requirements while eliminating some disadvantages of electromechanical
drives systems, as discussed in the following paragraph.

Pritschow [3] discusses the principle differences between linear motors and
the more conventional and still widely used ball-screw drives, and explains
that a shift to linear motors is required to further increase the productivity
level of machine tools. Fig 1.2 shows a schematic diagram of the
conventional ball-screw drive system. The main characteristic of a
conventional ball-screw drives structure is the transmission mechanism that
converts rotary motion of the motor to linear motion. The transmission
mechanism includes the gearing elements and the lead-screw. The lead-
screw element contributes negatively to the drives’ performance. The pitch
tolerances of the lead-screw generate transmission errors that reduce the
tracking accuracy. Tracking accuracy is also compromised by the backlash
3
Introduction

effects that originate from the dead zones and the large friction forces that
are generated in a high stiffness electromechanical ball-screw and bearing
structure. The large friction force results in limit cycles and reduced tracking
accuracy. High-gain feedback controllers reduce this effect but their
bandwidth is often limited by the resonance mode of the structure. In
addition, position, velocity, and acceleration are limited by the mechanical
characteristics of the lead-screw (e.g. stiffness) itself. Also, the lead-screw
adds large additional inertias to the system and reduces the first natural
frequency of the system. This reduces the bandwidth of the system and
represents a critical limitation to the system tracking accuracy.

motor
milling table
Position /
motor
Velocity
Controller
ball-screw

encoder position measurement

Fig. 1.2. Electromechanical ball-screw drive structure

Direct drives techniques with linear motors have been introduced as an

enhancement over the ball-screw drives system. Direct drives system has no
mechanical transmission between motor and load. This important
characteristic removes friction and backlash that are previously present in the
transmission mechanisms (gear boxes, harmonic drives, rack-and-pinion
systems, …), such that the total system friction is reduced to the (low)
friction in the guideways. Also, without transmission errors, better tracking
performance can then be achieved. In addition, the absence of mechanical
transmission elements enables higher velocities and accelerations. The first
natural frequency normally associated with the ball-screw drives is removed
thus extending the bandwidth of the system.

The structure of an iron-core linear motor is shown in fig. 1.3. Linear motors
are a special class of synchronous brushless servo motor. The structure
consists mainly of lamination stacks, coils, and magnets. Electrical energy is
converted to linear mechanical energy as a result of the electromagnetic
interaction between a coil assembly (primary part) and a permanent magnet
assembly (secondary part).

4
Introduction

Fig. 1.3. Structure of an iron-core linear motor (reproduced from the website of
ETEL, www.etel.ch)

Direct drives have several disadvantages. First, because of the absent of the
transmission mechanism, cogging forces, external disturbances and load
variations have a direct influence on the positioning and tracking accuracy.
As a consequence, the design of controllers that minimize these effects
becomes more challenging. However, the potential to achieve controllers
with better performance for direct drive systems is higher because of their
simpler mechanical structure and consequently more favourable dynamics.

Second, direct drives have a cost disadvantage compared to the conventional

electromechanical drive. Pritschow [3] states that for equivalent power
rating, the overall size of the linear drive is larger than the electromechanical
drive. The larger size is caused by the unused velocity potential and the
proportional relationship between the drive rated force and the surface area
of the moving drive parts. The additional cost is required for cooling the high
level of heat generation (in iron-core linear drive).

Linear motors applications are justified for applications where positioning or

tracking performance requirements out-weight the total initial cost. An
example of such application is high-speed machining. Efficient control
strategies are critical in machine tool direct drive applications as both
friction force and cutting force are acting directly on the motor and could
reduce the overall tracking performance. Successful compensation of these
disturbance forces is essential for accurate and effective motion control.

The next section investigates various disturbance compensation techniques

and approaches that have been proposed and applied in literature.

5
Introduction

1.2.2 Disturbance Forces and Compensation Methods

Two different types of disturbance forces that can critically affect tracking
performance and work-piece finishing quality in a linear motor based high-
speed milling cutting process are friction and cutting forces.

Hard non-linear friction behaviour can only be partly compensated using

linear feedback control strategies such as PID, cascade P/PI or state-
feedback control. The performance of these control techniques are restricted
by their limited bandwidth. More advanced techniques must be incorporated
in order to achieve sufficiently high path and tracking accuracy. Various
different approaches for friction compensation have been identified in the
literature. In general, these techniques can be summarized as indicated in fig.
1.4.
Friction Compensation Techniques

Friction-Model Friction-Model
Based Free

feedback feedforward nonlinear linear

Fig. 1.4. Schematic diagram for friction compensation techniques

Friction is compensated using friction-model based and friction-model free

approaches. Friction-model based compensation is either implemented in a
feedback or feedforward configuration. The difference between both
configurations is further discussed below. Friction-model free techniques can
be nonlinear or linear. Various model and non-model based friction
compensation techniques have been proposed and discussed in the literature.
Armstrong et al. [4] give an overview of the existing friction models and
compensation approaches. The recently developed friction models such as
the LuGre model, the Leuven model, and the Generalized Maxwell-slip
model are obviously not mentioned in this survey.

In friction-model based compensation, the accuracy of the model determines

directly the efficiency of the friction compensation. The accuracy of the
model depends on the selected model structure and accuracy of the model
parameters. Several simple and more advanced friction models have been
proposed in literature. Friction is categorized according to its pre-sliding and
6
Introduction

sliding regimes, and the most simple friction models consider the sliding
regime only. These models are a static map between friction force and
velocity, e.g. viscous, Coulomb and Stribeck effect friction models. A first
attempt in describing the more complex friction behaviour in pre-sliding
regime was accomplished by Dahl [5]. The Dahl model was applied
extensively for systems with ball-bearing friction. In 1995, Canudas et al. [6]
have proposed a new improved friction model, the LuGre model, for control
of systems with friction. The model captures most of the observed frictional
behaviours that include Coulomb friction, Stribeck effect, and hysteresis.
The LuGre friction model is widely applied and accepted for its simplicity
and relatively good performance. However, the LuGre friction model fails to
describe the hysteresis non-local memory behaviour of friction force in pre-
sliding regime. Swevers et al. [7] have improved the LuGre model yielding
the Leuven integrated friction model, which is further modified by Lampaert
et al. [8]. Recently, Al-Bender et al. [9] developed the so-called Generalized
Maxwell-slip (GMS) friction model and illustrate the superiority of the
model with respect to simulation of friction behaviour in both the pre-sliding
and sliding regimes. The main disadvantage of the GMS model is its
complexity and large number of parameters, which complicates its
application in control.

In a feedback configuration, the measured position signal is used to generate

a friction compensation force based on the available friction model. In a
feedforward configuration, the reference position is used instead of the
measured position signal. The feedback approach is generally designed on
the basis of the Coulomb model and the Maxwell-slip model [10]. A careful
design strategy is necessary to avoid instability problems. In this thesis,
feedforward friction compensation is applied: friction forces predicted by the
applied friction model are added to the control command signal to
compensate friction forces in the system. The effectiveness of this approach,
however, depends on the accuracy of the friction models. Changes in the
system environment and structure (lubrication, wear, etc.) could change the
friction behaviour and re-identification of the friction model will be
necessary to avoid deterioration in the friction compensation performance.

Besides model based friction compensation, various friction-model free

compensation methods have been suggested and discussed in the literature.
These include linear and non-linear control strategies. In a linear control
approach, Tung, Anwar, and Tomizuka [11] have demonstrated the
effectiveness of a repetitive controller, first introduced by Hara et al. [12], in
improving tracking performance and quadrant glitches compensation.
7
Introduction

In addition, Lampaert et al. [13] have proposed a control approach that

includes a disturbance observer based on a Kalman filter using a second
order random walk model.

Since friction is a highly nonlinear phenomenon it can be expected that

nonlinear control approaches are more appropriate. Several nonlinear control
strategies have been adapted for friction compensation. Tjahjowidodo et al.
[14] have shown that a Maxwell-slip-model-based nonlinear gain scheduling
controller yields fast response and low steady-state error for friction
compensation in electro-mechanical systems. Sliding mode control (SMC)
[15, 16] is an example of an important robust control design approach for
linear and nonlinear systems. Two main features of SMC are the finite
reaching time and the complete disturbance rejection of matched
uncertainties. In SMC applications, the system states approach the switching
line (or surface) and slide along this line to reach the final states. During
sliding, the control behaviour is independent of the system dynamics and
hence is independent of the influence of the acting disturbance forces.
Altintas [17] has described the SMC design for high-speed feed drives. A
tracking performance comparison between SMC and the classical cascade
controller is discussed in [18]. SMC however, is widely known for its
“chattering” problem [19, 20], a high-frequency switching of the SMC that is
a result of the imperfect switching mechanism and discontinuous control
signal around the switching surface. Chattering is highly undesirable since it
involves large control activity and possible excitation of the resonance
frequencies of the system. Many different techniques have been proposed in
literature to eliminate chattering in SMC related applications [21, 22].

Besides friction force compensation, attenuation of other external

disturbance forces, such as the cutting force that acts on a motion control
system, is of equal importance. Various compensation methods for different
applications have been proposed in literature. A repetitive controller is ideal
for periodic reference command signals and disturbance inputs. The
repetitive controller is only active in small frequency ranges around the
harmonics and system phase lag is compensated for in its design [23]. The
performance of a repetitive controller is highly influenced by the accuracy of
the period of the harmonics signal. Disturbances can also be compensated
using a robust control approach such as the H-infinity control [24] and SMC.
Van Brussel and Van den Braembussche have investigated the robustness of
both H∞ and SMC in linear motor systems [25].

8
Introduction

Ohnishi et al. present an inverse model-based disturbance observer [26, 27]

control structure that can be applied for any type of input disturbance. It adds
a high gain loop to an existing control configuration, thereby improving the
disturbance rejection with its bandwidth, which depends on the accuracy of
the available system model. The tuning of this bandwidth is a trade-off
between performance and stability margin. Kempf et al. [28] have added the
disturbance observer into an existing control system structure and have
shown enhanced disturbance attenuation performance of a track following
optical disk drive system against shock and vibration influence. The
disturbance attenuation property is removed from the design specifications
of the feedback controller resulting in simplified control design and greater
stability. This approach can be applied to any disturbance and is effective up
to a specified bandwidth which depends on the accuracy of the available
system model.

An explicit estimation of the cutting force has been attempted by Pritschow

et al. [29]. The estimation is based on the balance of forces acting on the
system and uses a relative acceleration sensor measurement. A reduced order
state observer estimates the velocity and the cutting force. However, to the
best of our knowledge, no report on the application of this estimator to
effectively compensate the influence of the cutting force is available.

1.3 Scope, Objective, and Approaches

Effective compensation of friction forces and cutting forces are a
prerequisite for accurate tracking performance in machine tool drive
applications. Classical motion controllers (discussed in chapter 2) are not
able to fully reduce the effect of these disturbances sufficiently. Various
other techniques have been reported as discussed in the previous section.
However, very limited knowledge on the application of some of these
techniques for friction and cutting forces compensation in high-speed milling
machine direct drives is available.

The objective of this research can be summarized as follow:

“Improve the positioning and tracking accuracy of motion

systems controlled by the classical cascade P/PI feedback
control structure by adding simple but effective friction and
cutting force compensation mechanisms.”

9
Introduction

These compensation mechanisms are developed for a linear drive based xy-
table of a high-speed milling machine. The modelling and experimental
identification of this system, the design of the cascade P/PI feedback
controller and various compensation mechanisms and their simulation-based
and experimental validation are discussed in detail in this thesis.

Friction forces are compensated using both friction-model based feedforward

and the inverse-model-based disturbance observer. These techniques have
been selected because of their simplicity. Feedforward is preferred over
feedback in order to avoid stability issues. Two friction models are
compared: a static friction model and the recent and more advanced
Generalised Maxwell-slip model. The inverse model based disturbance
observer [26, 27] is selected because of its simple structure and design
methodology that accounts for model uncertainty.

To compensate cutting forces, several techniques are implemented and

compared. These include: the inverse model based disturbance observer, a
repetitive controller, a force disturbance observer, and Pritschow’s explicit
estimation of cutting forces using the Ferraris relative acceleration sensor
measurement.

1.4 Contributions
The thesis attempts to enrich the understanding and knowledge regarding
accurate motion control of machine tool feed drive systems. The following
contributions are presented:

i. Development of simple motion control structure based on the

classical cascade P/PI controller combined with modules to improve
the friction and cutting forces compensation.
ii. Application of a simple identification procedure for the Generalized
Maxwell-slip (GMS) friction model that produces superior friction
compensation performance, especially at lower tracking velocity
iii. Application of an inverse model based disturbance observer in
combination with feedforward of the GMS friction model for
efficient friction compensation.
iv. Comparison of the tracking performance of different cutting force
compensation techniques during actual circular cutting tests.

10
Introduction

1.5 Outline
This thesis focuses on friction and cutting forces compensation of a linear-
motor based xy feed drive system. The above mentioned friction
feedforward, inverse model based disturbance observer, and repetitive
controller are add-on devices, implying that they are added to an existing
feedback control system, which is a classical cascade P/PI controller. The
structure, design procedures, and analysis of a classical cascade P/PI position
controller that is widely applied in many mechatronic systems are discussed
first in Chapter 2. The minimization of the effect of disturbance forces on
the system position and tracking performance requires precise knowledge
and complete understanding of the characteristics of these disturbance forces.
Chapter 3 discusses the modelling and identification of friction, and the
spectral analysis of cutting force measurements of an actual milling cutting
process. Chapter 4 discusses the implementation of the friction force
compensation based on the feedforward of friction models. Simulations and
experimental validations of friction compensation performances using a
simple static friction model and the more advanced and recent GMS friction
model are performed and compared. An inverse model-based disturbance
observer module is introduced in combination with feedforward of the
friction models and its influence on the magnitude of the quadrant glitches
from circular tests are studied and analysed. Chapter 5 describes and
discusses compensation of the cutting force using various selected control
techniques. The experimental validation of the different compensation
techniques are performed first using artificial cutting force synthesized from
actual milling cutting force measurements and second during actual circular
cutting processes. The practical implementation of friction and cutting forces
compensation during these cutting tests and the tracking performances of the
different techniques are discussed and compared. Finally, Chapter 6 presents
the conclusions regarding the design of the cascade P/PI controller, the
performance of friction and cutting forces compensation and some
recommendations of future work.

11
12
 Chapter 2

Classical Motion Control

2.1 Introduction
Motion tracking controllers are designed with the objective of achieving
maximal tracking accuracy and robustness against disturbances and plant
uncertainties. The size of the tracking errors and actuator input signals are
important indicators to validate a control system design. A good control
design ensures that these indicators remain below some pre-specified
conditions and hence exploiting the system to its fullest potential. Feedback
control strategy is the basic principle of many control systems. The desired
reference signals are compared with the actual output of the system and
corrective measures are implemented to compensate these errors. Feedback
control linearizes nonlinear elements and partly compensates the effects of
disturbances and system variations. Due to its simplicity, PI and PID control
form 90% of the practical control applications. Feedforward control
strategies are normally introduced to compliment feedback control [30]. A
combined feedforward and feedback control strategy improves tracking
performance especially in systems with pre-knowledge of the inputs
reference and disturbance signals.

A prominent and classical tracking controller that exists in the majority of

servo motion control systems is the classical cascade controller. Cascade
control is widely applied for its simple structure and transparent design. In
addition, the loop controllers are based on the familiar classical proportional
P, and integrator I controllers. Various modifications to this controller have
been suggested in literature with the aim of improving its tracking
performance. For example, Doenitz [31] analysed the tracking error of a
combination of a cascade controller with a disturbance observer. In addition,

13
Classical Motion Control

Boucher et al. [32] have presented a generalized predictive cascade control

for control of machine tool drives.

This chapter discusses the analysis and design of a classical cascade P/PI
controller for position control of a linear drive based xy feed table. First, a
detailed description of the experimental setup, that is, a high-speed xy
milling machine is presented.

2.2 High-Speed XY Milling Machine

The test setup that is considered for the experimental validation is a linear-
drive based xy feed table of a high-speed milling machine (see fig. 2.1). This
xy table is the result of a European project named MOTION (Modular
Technologies for Intelligent Motion Unit with linear motors and axis control)
that was implemented from 1996-2000. The partners included Philips CFT,
Fidia S.p.A, SMT, Schneeberger, ETEL, Royal Institute of Technology
(KTH), and K.U.Leuven (Div. PMA).

y-axis
Linear
x-axis Motor

Linear
Motor
Linear
Motor

Fig. 2.1. A linear-driven xy feed table of a high-speed

milling machine

14
 Classical Motion Control

The axes are driven by three ETEL’s LMC 22-070-3TA linear motors. The
upper stage y-axis is driven by a single ETEL iron-core linear motor while
the bottom stage x-axis is driven by two ETEL iron-core linear motors.
ETEL patented anti-cogging design ensures very low cogging and force
ripple effects. The motions are driven in current mode using ETEL DSB2-S
amplifiers. The schematic representation of the two axes is shown in fig.2.2.

Encoder Y Motor Y

Motor X1
Milling
Guideways Table
Encoder X

Motor X2

y-axis

x-axis

Guideways
Fig. 2.2. Schematic diagram of the xy table

Both axes are equipped with a 0.25µm resolution Heidenhain LF481C linear
encoder, mounted in the middle of the stages. The encoders have a signal
period of 4µm and generate two sinusoidal signals of 1 Vpp (peak to peak)
and a phase difference of π/2. A third index signal provides the motor
absolute position. The signals are linked to the ETEL DSB2-S amplifiers for
the stages position measurement. A dSPACE DS1103 controller board
interfaced to the ETEL DSB2-S amplifiers receives an interpolated TTL
analog position signal. The resolution of the encoder signal is determined by
the interpolation setting factor of the DSB2-S amplifier. For an interpolation
factor of 4, the encoder resolution is

encoder period 4 μm
RES = = = 0.25μm . (2.1)
4 ⋅ interpolation factor 4 ⋅ 4

15
Classical Motion Control

In the absence of a velocity sensor, the velocity signals are calculated by

numerical differentiation of the position signals in combination with a first-
order low-pass Butterworth filter with a bandwidth of 150Hz. This filter is
added to attenuate amplified measurement noise associated with the
derivative action (see section 2.5).

The stages are supported by 4 Schneeberger MRD35 preloaded roller

carriages on 2 rails that are fixed on the middle part. For the y-axis, the stator
of the linear motor is fixed to the middle part and the motor’s forcer is
connected to the table. The total mass of the y-axis is 170kg. The x-axis of
the table is supported in the middle by 6 MRD35 carriages on 3 rails that are
fixed to the base. The stators of the two linear motors are fixed to the base
while the forces are connected to the middle part. The total mass for the x-
axis that includes the y-axis, the middle part, and the stators of the linear
motors is 340kg. “Rubber wipers” protect the rolling elements in the linear
bearing from any harsh environment. A u-shaped shield covers each linear
motor, protecting the motors magnets from the cutting chips. The shields
move in a resin profile for improved encapsulation.

Fig. 2.3 shows a schematic diagram of the motion controller structure. The
controller of each axis is implemented on a dSPACE 1103 DSP controller
board, using the ControlDesk software from dSPACE to link the host
computer to the ETEL drives. An additional DS1102 controller board
provides the communications between the DSP controller board and the
ETEL drive amplifiers using digital I/O interface. The digital I/O interface
provides basic communications between the host computer and the linear
drives. These actions include the drives power on/off, drive enable/disable
commands, table absolute positioning, and homing procedures.

The dSPACE 1103 controller board is responsible for position control of the
drives. The controller settings and tracking commands are uploaded to the
drives from the host computer via the dSPACE and tracking performances
are monitored from the measured encoder signals.

16
 Classical Motion Control

Fig. 2.3. Motion controller structure of a xy feed table with three

linear drives for high speed milling application

The system identification is described in the next section.

2.3 System Identification

A linear time-invariant model of the dynamic behaviour of each axis is
obtained from frequency response function (FRF) measurements of the
system. The dynamic coupling between both axes is negligible and the
system dynamics can be described by two single-input single-output (SISO)
models. In order to measure the two SISO FRFs of the system, the system is
excited with band-limited white noise signals. The output encoder
measurements and the excitation signal, that is the signal that is sent to the
ETEL drive amplifiers, are recorded. The sampling frequency is 2000Hz and
the total duration of the measurement is 5 min. A Hanning window is
applied. The number of samples per window is 2048. This yields a sampling
resolution of 1Hz. The SISO FRFs of the system are estimated using the H1
estimator [33].

17
Classical Motion Control

Fig. 2.4 shows the measured FRFs for the x and the y axes.

6
Bode Diagram
10
G =out. pos. / in. voltage

x-axis
4
10 y-axis
[µm]/[volt]

2
10

0
10

-2
10

0
phase [degrees]

-90

-180

-270

-360
1 2
10 10
Frequency [Hz]

Fig. 2.4. FRFs measurement of the x and y axes

The FRF of each axis contains an anti-resonance and resonance combination

near 45Hz. This is due to the relative motion between the base of the
machine and the ground. A study of this is presented in detailed in Appendix
A. The study shows a left shift in the anti-resonance and resonance
frequencies as the tightness of the bolts that hold the base to the ground are
loosened. Further theoretical explanation of this phenomenon is presented in
detail in [34].

Parametric models are fitted on the FRFs measurement using the nonlinear
least squares frequency domain identification method [33], yielding the
following second order model with time delay:

Z (s ) B
G m (s ) = = ⋅ e − sTd , (2.2)
U (s ) s (s + A)

Equation (2.2) relates the linear dynamic relation between input voltage u
and table position z[m], with z = x and z = y for the x and y axes respectively.
A summary of the estimated model parameters, A, B, and Td, for each axis is
given in Table 2.1.

18
 Classical Motion Control

Table 2.1. System model parameters for x and y axes

A B Td
x-axis 28.57 [s-1] 4.526 [m/volt.s2] 0.00065s
y-axis 20.00 [s-1] 8.916 [m/volt.s2] 0.00065s

Fig. 2.5 and fig. 2.6 show the frequency responses of the obtained models
and corresponding measured FRFs for the x and y axes respectively.
Bode Diagram
4
10 FRF
G= out. pos / in. volt

model
[µm]/[volt]

2
10

0
10

-2
10
-90
-135
Phase (deg)

-180
-225
-270
-315
-360
1 2
10 Frequency (Hz) 10

Fig. 2.5. X-axis: FRF measurement and proposed model

Bode Diagram
4
G= out. pos / in. volt

10 FRF
model
[µm]/[volt]

2
10

0
10

-2
10
-90
-135
Phase (deg)

-180
-225
-270
-315
-360
1 2
10 10
Frequency (Hz)

Fig. 2.6. Y-axis: FRF measurement and proposed model

19
Classical Motion Control

The parametric model (2.2) estimates the models as a mass line with
damping characteristic at lower frequency range. These models do not
include the resonance and anti-resonance peaks at frequencies beyond 40Hz
as shown in fig. 2.5 (e.g. the anti-resonance frequency at 47Hz for the x-axis,
and at 44 Hz for the y-axis).

The next section discusses the control analysis and design of a cascade
controller that is applied for position control of the xy table. Analysis of the
cascade controller based on an ideal configuration of the system is first
considered.

2.4 Cascade Control Structure & Analysis

2.4.1 Cascade Controller Structure & Configuration

A cascade control structure is used in many mechatronics applications. The
control structure consists of a velocity loop and a position loop. The position
loop is built around the velocity loop. Fig. 2.7 shows a general scheme of a
cascade control structure. A disturbance force d(t) enters at the input of the
system.

Fig. 2.7. General scheme of a cascade control structure

Tracking performance of a cascade controller is analysed based on the closed

loop bandwidths, damping ratio, and dynamic stiffness. The compliance
function is the transfer function from input disturbance d to the measured
output position z. The dynamic stiffness is the inverse of this transfer
function. Dynamic stiffness is an important characteristic of a tracking
controller because it indicates the controller disturbance rejection capacity.

20
 Classical Motion Control

This section analyses the characteristics of the cascade P/PI controller by

first considering an ideal theoretical design configuration. The following
idealized configuration is considered:

¾ the system behaves as an ideal rigid body system

without friction,
¾ velocity v and position signals z are assumed to be
available,
¾ measurement noise n(t) is only present on the
velocity signal, not on the position signal,
¾ disturbances d(t) enter at the input of the system, that
is, these disturbances are forces.

The analysis (in section 2.4.2 and 2.4.3) is based on a model that is a further
simplification of model (2.2), that is, the time delay Td and real pole s = - A
are removed. The resulting model is a double integrator, or mass-line model,
and is presented in fig. 2.8. M represents the mass of the considered system.
kf [N/volt] represents in this model the motor constant relating the motor
force F to the input voltage to the motor drive amplifier. Figure 2.7 indicates
that this simplified model equals:

Z (s) G kf
= = (2.3a)
U (s) s2 Ms 2

kf = G ⋅M (2.3b)

Based on the estimate of the mass M of each axis and the gain G [m/volts.s2]
derived from the model (2.2) and Table 2.1, the motor constant of the x and
y axes are estimated at 1450[N/volt] and 1470 [N/volt] respectively.

Fig. 2.8. Schematic diagram of the open loop system for force constant estimation

21
Classical Motion Control

Fig. 2.9 shows an ideal cascade control configuration.

Fig. 2.9. An ideal cascade control structure for a linear motor position control

The tracking error ep(t) is the difference between the reference position
zref(t) and the measured position z(t). Structural resonances are not
considered here. In practical applications, they are often present, and
compensated in the control structure by adding notch filters [35].

The velocity loop controller is usually either a proportional (P) controller or

a proportional and integral (PI) controller. Both a P and a PI velocity loop
controllers are considered in the following theoretical analysis. The position
loop controller is usually a proportional controller. First, the analysis of a
system with a proportional velocity controller and a proportional position
controller is considered.

2.4.2 Analysis of the Closed Loops Behaviour with

Proportional (P) Velocity Control
A velocity controller with a proportional gain kp and a position controller
with a proportional gain kv is considered. The system dynamic open loop is:

Z (s) kf
= (2.4)
U (s) Ms 2

The transfer functions relating input reference signal zref (t), velocity noise
signal n(t), and input force disturbance d(t) to output position signal z(t) are:

22
 Classical Motion Control

kv k p k f
Z (s) = 2
⋅ Z ref ( s )
Ms + k p k f s + kv k p k f
k pk f
+ 2
⋅ N (s) (2.5)
Ms + k p k f s + kv k p k f
1
+ ⋅ D(s)
Ms 2 + k p k f s + kv k p k f

Z (s)
The dynamic stiffness of the controller, which is the inverse of in (2.5),
D(s)
is

D(s)
= Ms 2 + k p k f s + kv k p k f . (2.6)
Z (s)

The position error ep(t) can be expressed as a function of the reference signal
zref (t), velocity noise signal n(t), and the input disturbance d(t),

E p ( s ) = Z ref ( s ) − Z ( s )
Ms 2 + k p k f s
Ep (s) = ⋅ Z ref ( s )
Ms 2 + k p k f s + kv k p k f
k pk f (2.7)
- ⋅ N (s)
Ms 2 + k p k f s + kv k p k f
1
- 2
⋅ D(s).
Ms + k p k f s + kv k p k f

The undamped natural frequency ω0 and the damping ratio ς of the system
Z (s )
are obtained from the transfer function in (2.5),
Z ref (s )

kv k p k f
ω0 = (2.8)
M

1 k pk f
ς = . (2.9)
2 Mk v

23
Classical Motion Control

The tracking performance of the cascade P/P controller is analysed as

follows:

For a reference signal with a constant velocity,

v
Z ref (s ) = , (2.10)
s2

the steady-state tracking error obtained using the final-value theorem to

equation (2.7) (ignoring noise signal n(t) and input disturbance d(t)) equals:

v
ep (∞) = (2.11)
kv

The steady-state error is inversely proportional to the proportional gain kv

and is directly proportional to the constant tracking velocity v.

Next, the positioning performance is analysed for a step reference input

signal zref(t)=k and for a static disturbance signal d(t) = k. The steady state
position error for a step reference input signal and for a static disturbance
signal, obtained by applying the final value theorem to equation (2.7) are:

e pz _ step = 0 (2.12a)

k
ep =− . (2.12b)
d _ step kv k p k f

The steady-state error is proportional to the static disturbance force k

(equation 2.12b) and will always be a non-zero value. This also means that,
in the absence of a disturbance signal, the steady-state error is zero (equation
2.12a).

24
 Classical Motion Control

2.4.3 Analysis of the Closed Loops Behaviour with PI

Velocity Control
A velocity loop with proportional-integral (PI) controller and a position
controller with a proportional gain kv is considered. The velocity loop PI
controller has the following structure,

ki
PI = k p + (2.13)
s

With reference to previous fig. 2.9, with velocity controller (2.13) and a
proportional position controller kv, the transfer functions relating input
reference signal zref (t), noise signal n(t), and the input force disturbance f(t)
to the output position signal z(t) are:

kv k p k f s + kv ki k f
Z (s) = ⋅ Z ref ( s )
( )
Ms + k p k f s 2 + ki k f + kv k p k f s + kv ki k f
3

k p k f s + ki k f
+ ⋅ N (s) (2.14)
3 2
( )
Ms + k p k f s + ki k f + kv k p k f s + kv ki k f
s
+ ⋅ D(s)
3 2
(
Ms + k p k f s + ki k f + kv k p k f s + kv ki k f)
The system dynamic stiffness, which is the inverse of Z(s)/D(s) in (2.14)
equals:

D(s)
Z (s)
=
1
s
( ( )
Ms 3 + k p k f s 2 + ki k f + kv k p k f s + kv ki k f ) (2.15)

The integrator in (2.15) gives an infinite static stiffness for the system at low
frequency range (see fig. 2.22).

The relationship between the position error ep(t) and the reference signal
zref (t), the velocity noise signal n(t), and the input force disturbance d(t) are:

25
Classical Motion Control

E p ( s ) = Z ref ( s ) − Z ( s )
Ms 3 + k p k f s 2 + ki k f s
Ep (s) = ⋅ Z ref ( s )
( )
Ms 3 + k p k f s 2 + ki k f + kv k p k f s + kv ki k f
k p k f s + ki k f (2.16)
- ⋅ N (s)
3 2
( )
Ms + k p k f s + ki k f + kv k p k f s + kv ki k f
s
- ⋅ D(s).
( )
Ms 3 + k p k f s 2 + ki k f + kv k p k f s + kv ki k f

The undamped natural frequency ω0, and the damping ratio ς cannot be
obtained directly from (2.14) because the system has 3 poles. An
approximation of the transfer function is desired. J. Wang [36] approximated
the transfer function Z(s)/Zref(s) in (2.14) as a second order model and
suggested the following approximate relationships relating the controller
parameters kv, kp, ki, with the damping ratio ς and the undamped natural
frequency ω0:

kv k p k f
ω0 = (2.17)
M

1 k pk f 1 kf kp
ς= − 0.05 ≈ (2.18)
2 Mk v 2 Mk v

The tracking performance of the cascade P/PI controller is analysed as

follows:

For a reference signal with a constant velocity,

v
Z ref ( s ) = , (2.19)
s2

the steady-state tracking error (with no disturbance and noise signals)

obtained by applying the final-value theorem to equation 2.16 equals:

v
ep (∞) = . (2.20)
kv

26
 Classical Motion Control

Similar to the previous cascade P/P analysis, the steady-state error for
constant velocity reference tracking is inversely proportional to the
proportional gain kv and is directly proportional to the constant tracking
velocity v.

Next, the positioning performance is analysed for a step reference input

signal zref(t) = k and for a static disturbance signal d(t) = k. The steady-state
position error for the respective cases, obtained by applying the final value
theorem to equation 2.16 equals:

e pz _ step = 0 (2.21a)

e pd _ step = 0. (2.21b)

Unlike the cascade P/P analysis, the integrator in the velocity controller
eliminates position error for a step reference input signal and compensates
the static disturbance force.

Next, the design for the considered cascade P/PI controller is discussed. The
design is based on the measured FRFs of the system shown in fig. 2.4.

2.5 Design and Validation of Cascade P/PI

Controller Based on Measured FRFs
The parameters of the cascade P/PI controllers are selected and tuned based
on gain margin and phase margin considerations of the open loop transfer
functions, using measured FRFs of the system Ĝ (fig. 2.4). First, the PI
velocity loop controller parameters are designed and the loop characteristics
are analysed based on the control scheme shown in Fig. 2.10.

2.5.1 Design and Analysis of the Velocity Loop

The velocity loop of the cascade P/PI controller scheme includes the PI
velocity controller (equation 2.13), a notch filter N(s), and the transfer
function Vest(s) of the operator that estimates a velocity signal from the
measured position encoder signal.

27
Classical Motion Control

Fig. 2.10. Frequency domain scheme of the cascade P/PI controller for control of a
linear motor drive

This transfer function (2.22) is a combination of a continuous-time

derivative, s, and a first order low-pass Butterworth filter with a cut-off
frequency of 150 Hz for both axes. The filter is added to attenuate amplified
measurement noise associated with the derivative action.

942.5s
Vest ( s ) = (2.22)
s + 942.5

Fig. 2.11 shows the Bode diagram of Vest(s).

60
Magnitude (dB)

40

20

-20
90

45
Phase (deg)

-45 low pass filter

Vest(s)
-90
0 1 2 3
10 10 10 10

Fig. 2.11. Bode diagram of Vest(s)

28
 Classical Motion Control

High frequency structural resonances can reduce the performance and

stability of a system. Often, notch filters are added to reduce the influence of
these resonances on the closed loop behaviour. With reference to the FRFs
measurement shown in fig. 2.4, notable high structural resonance frequencies
are observed between 320 Hz and 340 Hz and between 270Hz and 320 Hz
for the x and the y axes respectively. Equation (2.23) describes a notch filter
transfer function for the x-axis, based on a Butterworth band-stop filter. Fig.
2.12 shows the Bode diagram of (2.23).

s 2 + 4.3e6
N (s ) = (2.23)
s 2 + 125.7 s + 4.3e6

Bode Diagram
0
Magnitude (dB)

-50

-100

-150
450

405
Phase (deg)

360

315

270
250 300 350 400
Frequency (Hz)

Fig. 2.12. X-axis: Bode diagram of a notch filter N(s) for frequency
between 320 Hz and 340 Hz

The transfer function of the notch filter selected for the y-axis equals:

s 2 + 3.41e6
N (s ) = (2.24)
s 2 + 314.2s + 3.41e6

Next, the velocity open and closed loop transfer functions are analysed based
on the cascade control scheme in fig. 2.10. The velocity open and closed
loop transfer functions are:

29
Classical Motion Control

Z& est (s )
v ol = = PI ⋅ Gˆ ⋅ N ⋅ V est (2.25)
E v (s )

Z& ( s ) ˆ ⋅ N ⋅V
PI ⋅ G
vcl = est = est (2.26)
U p ( s ) 1 + PI ⋅ G
ˆ ⋅ N ⋅V
est

The compliance function and the sensitivity function of the velocity loop are:

Z (s ) Gˆ / k f
= (2.27)
D(s ) 1 + PI ⋅ Gˆ ⋅ N ⋅ Vest

1
S v (s ) = (2.28)
ˆ
1 + PI ⋅ G ⋅ N ⋅ Vest

The velocity loop PI (equation (2.13)) controller parameters, kp and ki are

selected based on gain margin and phase margin consideration of the
velocity open loop transfer function. A reasonable value of gain margin and
phase margin ensures stability and good transient response characteristics
[32]. A minimum gain margin of 5dB and a minimum phase margin of 50
degree are desired. The velocity loop PI controller parameters of each axis
are listed in Table 2.2.

Table 2.2. Velocity loop PI controller parameters of x and y axes

kp [volt·s ] ki [volt·s2]
x-axis 0.00006 0.00045
y-axis 0.00003 0.00025

Fig. 2.13 and fig. 2.14 show the velocity open loop transfer functions
(equation (2.25)) for the x and y axes respectively, based on the PI
controller’s parameters in Table 2.2. The influence of the structural
resonance frequencies near 320 Hz is removed as indicated by the circle in
the magnitude plots of fig. 2.13 and 2.14.

Table 2.3 summarizes the gain margins and the phase margin of the x and y
axes. These values ensure good transient response and stability margin for
each axis velocity loops.

30
 Classical Motion Control

Bode Diagram
40
without notch filter
Magnitude (dB)

20 with notch filter

-20

-40
0
-45
-90
Phase (deg)

-180

-270

-360
1 2
10 frequency [Hz] 10

Fig. 2.13. X-axis: Theoretical bode plot of the velocity open loop
transfer function based on measured FRF of the system

Bode Diagram
40
without notch filter
Magnitude (dB)

20
with notch filter

-20

-40
0

-90
Phase (deg)

-180

-270

-360
1 2
10 10
Frequency (Hz)
Fig. 2.14. Y-axis: Theoretical bode plot of the velocity open loop
transfer function based on measured FRF of the system

31
Classical Motion Control

Table 2.3. Gain margin and phase margin of x and y axes velocity open loop
Gain Margin Phase Margin
x-axis 12.332 dB (at 315.08 Hz) 74.514 deg (at 65.012 Hz)
y-axis 10.311 dB (at 307.63 Hz) 73.059 deg (at 34.922 Hz)

The magnitudes of the gain margin and the phase margin for the x and y axes
velocity loops are conservative to ensure a good loop stability margin. The
stability of both axes velocity loops is analysed from the Nyquist plots of the
velocity open loop transfer function, that is, the point [-1,0] is not encircled.
The Nyquist plots, each drawn with a unit circle, are shown in fig. 2.15. The
positive influence of the notch filter in increasing the stability margin of the
velocity loop is emphasized.

Nyquist Plot (x-axis) Nyquist Plot (y-axis)

2 2
1.5

1 1
Imaginary Axis

0.5
Imaginary Axis

0 0
-0.5
-1 -1
-1.5
w ith notch filter w ith notch filter
no notch filter no notch filter
-2 -2
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
real axis real Axis

Fig. 2.15. Nyquist plots of the x-axis (left) and the y-axis (right) velocity open
loop transfer functions based on measured FRFs of the system

The bandwidth of the velocity loop is determined from the magnitude plot of
the sensitivity function Sv(s), that is the transfer function (2.28), shown in fig.
2.16. The bandwidth is defined as the frequency of which the magnitude of
the sensitivity function crosses the -3dB line from below. Table 2.4
summarizes the values of the velocity loop bandwidth for both axes. Fig 2.16
also emphasizes the influence for the removal of the structural resonance
frequency near 320 Hz, as indicated by the two circles.

32
 Classical Motion Control

Bode Magnitude Diagram Bode Magnitude Diagram

x-axis y-axis

2 2

Sv=pos err. / disturbance force

Sv=pos err. / disturbance force

1 1

[µm] / [N]
[µm] / [N]

w ith notch filter w ith notch filter

no notch filter no notch filter
0.1 0.1
0 1 2 0 1 2
10 10 10 10 10 10
Frequency [Hz] Frequency [Hz]
Fig. 2.16. Sensitivity function of x (left) and y (right) axes for the
velocity loop

Table 2.4. Bandwidths of the velocity loop

x-axis y-axis
bandwidth 28.9Hz 28.8Hz

This definition of a system bandwidth gives a more accurate representation

of the system performance compared to the traditional definition based on
the complimentary sensitivity function (equation 2.26) as indicated in [34].
Bode diagrams of the velocity closed loop transfer function of each axis
based on (2.26) are shown in fig. 2.17.

Bode Diagram
10

0
Magnitude (dB)

-10

-20
without notch filter
-30
with notch filter
-40
90

0
Phase (deg)

-90

-180

-270

-360
1 2
10 10
Frequency (Hz)
(a)

33
Classical Motion Control

Bode Diagram
10

0
Magnitude (dB)

-10

-20

-30 without notch filter

with notch filter
-40
90

0
Phase (deg)

-90

-180

-270

-360
1 2
10 10
Frequency (Hz)
(b)

Fig. 2.17. Velocity closed loop transfer function for (a) x-

axis and (b) y-axis

The positive effect of the notch filters is illustrated by the removal of the
peak near 320Hz, as indicated by the two circles. This positive effect is
specifically important because it removes the possibility of deterioration in
performance at high frequency.

2.5.2 Design and Analysis of the Position Loop

The second step of the design procedure includes the design and analysis of
the position loop that is built around the velocity loop. The position open
loop and closed loop characteristics are examined for a proportional gain
controller kv equals:

P = kv = 150 s −1 (2.29)

Based on fig. 2.10, the position open and closed loop transfer function are:

Z (s) ˆ
P ⋅ PI ⋅ G
posol = = (2.30)
Ep (s) ˆ ⋅ N ⋅V
1 + PI ⋅ G est

34
 Classical Motion Control

Z (s) ˆ
P ⋅ PI ⋅ G
poscl = = . (2.31)
Z ref ( s ) ˆ ˆ
1 + PI ⋅ G ⋅ N ⋅ Vest + P ⋅ PI ⋅ G

The compliance function, the sensitivity function, and the position steady-
state error transfer function equals:

Z (s ) Gˆ / k f
= (2.32)
D(s ) 1 + PI ⋅ Gˆ ⋅ N ⋅ Vest + P ⋅ PI ⋅ Gˆ

1
S p (s ) = (2.33)
1 + PI ⋅ Gˆ ⋅ N ⋅ Vest + P ⋅ PI ⋅ Gˆ

1 + PI ⋅ Gˆ ⋅ N ⋅V
Ep (s) =
ˆ ref ( )
est Z s
ˆ
1 + PI ⋅ G ⋅ N ⋅ V + P ⋅ PI ⋅ G
est
(2.34)
Ĝ / k f
- D(s).
ˆ ⋅ N ⋅ V + P ⋅ PI ⋅ G
1 + PI ⋅ G ˆ
est

The gain margin and the phase margin of the x and the y axes are identified
from the position open loop transfer functions shown in fig. 2.18. These
values are summarized in Table 2.5.

Table 2.5. Gain margin and phase margin of x and y axes position open loop
Gain Margin Phase Margin
x-axis 15.1 dB (at 100 Hz) 68 deg (at 21.4 Hz)
y-axis 16.49 dB (at 100.62 Hz) 67.18 deg (at 22.47 Hz)

The gain and phase margins are conservative. A higher value of kv increases
the sensitivity peak (refer to fig. 2.20) near and above the 6dB line (or the
absolute magnitude value of 2), that is the normal design criteria for the
sensitivity peak value. Higher sensitivity peak value deteriorates the system
performance at high frequencies.

35
Classical Motion Control

x-axis
40
Magnitude (dB)

20
0
-20
-40
-60

-90
Phase (deg)

-180

-270

-360
1 2
10 Frequency [Hz] 10

(a)
y-axis
40
Magnitude (dB)

20
0
-20
-40
-60

-90
Phase (deg)

-180

-270

-360
1 2
10 Frequency [Hz] 10

(b)
Fig. 2.18. Position open loop transfer function for (a) x-axis and (b) y-axis
based on measured FRF of the system

The stability of the x-axis and the y-axis position loops is analysed from the
Nyquist plots of the position open loop transfer function, that is, the point
[-1,0] is not encircled. Fig. 2.19 shows the Nyquist plots, each drawn with a
unit circle. In both cases, the [-1, 0] point is not encircled, thus ensuring
system stability.

36
 Classical Motion Control

Nyquist Plot (x-axis) Nyquist Plot (y-axis)

2 2

1 1
Imaginary Axis

Imaginary Axis
0 0

-1 -1

-2 -2
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
Real Axis Real Axis

Fig. 2.19. Nyquist plots of the x-axis (left) and the y-axis (right) of
the position open loop transfer functions

Next, the bandwidth of the position loop is identified from the -3dB crossing
(from below) on the magnitude plot of the sensitivity function Sp(s), equation
(2.33). Fig. 2.20 shows the magnitude plots of the sensitivity functions for
the x and y axes. The bandwidths of both axes are summarized in Table 2.6.
Bode Magnitude Diagram Bode Magnitude Diagram
x-axis y-axis
Sp=pos err. / disturbance force
Sp=pos err. / disturbance force

0 0
10 10
[µm] / [N]
[µm] / [N]

w ith notch filter w ith notch filter

no notch filter no notch filter
-1 -1
10 10

1 2 1 2
10 10 10 10
[Hz] [Hz]

Fig. 2.20. Sensitivity function of x (left) and y (right) axes for the
position loop

Table 2.6. Bandwidths of the position loop

x-axis y-axis
bandwidth 26.6Hz 27.1Hz

The positive influence of the notch filters are emphasized from the removal
of the structural resonance peak near 320 Hz, as indicated by the two circles.
37
Classical Motion Control

The peaks of the sensitivity magnitude curves lie below the conventional
guideline of 6dB (or abs. value 2). This ensures good transient response
characteristics. Also, the results of Table 2.6 confirm the fundamental
characteristic of a cascade controller, that is, the bandwidths of the position
loop is smaller than the bandwidth of the inner velocity loop. Next, the
position closed loops of both axes are analysed and their Bode plots are
shown in fig. 2.21.

Bode Diagram
10
0 x-axis
Magnitude (dB)

-10 y-axis
-20
-30
-40
-50
-60
0

-90
Phase (deg)

-180

-270

-360
1 2
10 10
Frequency (Hz)
Fig. 2.21. Position closed loop transfer functions for x and y axes

In addition to the system bandwidth, dynamic stiffness gives a good measure

of the controller disturbance rejection performance. Dynamic stiffness is the
inverse of the transfer function from the input disturbance to the output
position signals. In order to identify the system dynamic stiffness, a band-
limited white noise signal is introduced as a disturbance input to a system,
and the position is measured. Based on these measurements, the dynamic
stiffness is estimated using the H1 estimator [33].

The dynamic stiffness FRF measurements for the x and the y axes using a
cascade P/PI controller, in comparison to the cascade P/P controller are
shown in fig. 2.22. These measurements confirm the previous assertion that
the cascade P/PI control structure has an infinite stiffness at lower frequency
(equation 2.15) as a result of the integrator in the velocity loop. The
minimum stiffness of the cascade P/PI controller is 15N/µm and 8N/µm for
the x and the y axis respectively.

38
 Classical Motion Control

x-axis y-axis

d istu rb an ce fo rce / p o sitio n

3 3
10 10
dis turba nc e forc e / pos ition

[N] / [µ m ]
cascade P/PI cascade P/PI
[N ] / [µ m ]

2 2
10 10

1 1
10 10

0
cascade P/P 0 cascade P/P
10 10
0 1 2 0 1 2
10 10 10 10 10 10
frequency [Hz] frequency [Hz]

Fig. 2.22. Measured dynamic stiffness of x (left) and y (right) axes

using cascade P/P and cascade P/PI controller

2.5.3 Tracking Performance Numerical Validation

In addition to the disturbance forces rejection property, a good tracking
performance is of equal importance in control design. The tracking
performance is evaluated based on position and tracking errors.

The tracking performance of the cascade P/PI controller is analysed for a

sinusoidal reference input signal. A control scheme similar to fig. 2.10 with
the system equals the system model Gm(s) is applied. Fig. 2.23 shows for
both x and y axes, the simulated tracking errors for a sinusoidal reference
signal with amplitude of 1mm and a frequency of 10Hz. The amplitude of
these tracking errors corresponds to the magnitude of the error transfer
function Ep(s)/Zref(s) (see equation 2.34) shown in fig. 2.24 (indicated by
block square).
x-axis y-axis
0.5 0.5
0.44 0.42
position error [µm]

position error [µm]

0.2 0.2

0 0

-0.2 -0,2

-0.44 0,42
-0.5 -0.5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
time [s] time [s]
Fig. 2.23. Simulated tracking error of the x-axis (left) and the y-axis (right) for
sinusoidal reference signal of an amplitude of 1mm and frequency of 10Hz
39
Classical Motion Control

Bode Magnitude Diagram Bode Magnitude Diagram

0 0
10 10
Magnitude (abs)

Magnitude (abs)
System: errorX System: errorY
Frequency [Hz]: 10 Frequency [Hz]: 10
-1
Magnitude (abs): 0.441 -1 Magnitude (abs): 0.423
10 10

-2 -2
10 10
0 1 2 0 1 2
10 10 10 10 10 10
[Hz] [Hz]

Fig. 2.24. Position error transfer functions for the x (left) and y axes (right)

Significant tracking errors are still present. Tracking performance can be

significantly improved by adding feedforward. This is discussed in the
following section.

2.5.4 Cascade P/PI with Feedforward

Various feedforward techniques are used to improve tracking performance in
motion control system. Velocity feedforward, applied at the velocity loop of
the cascade P/PI controller structure is considered. The tracking performance
is further improved with an additional feedforward, based on the inverse
open loop model of the system G-1(s) and position reference. Fig. 2.25 shows
the resulting cascade configuration.

Fig. 2.25. Cascade P/PI with feedforwards for improved tracking

performance

40
 Classical Motion Control

The tracking performance analysis in 2.5.3 is repeated for cascade P/PI

control with feedforward. A control scheme similar to fig. 2.25 is applied. A
system model Gm(s) and the inverse open loop model feedforward Gm-1(s) are
considered for this numerical validation purposes. Fig. 2.26 compares the
tracking errors between the initial cascade P/PI control structure and the
cascade P/PI with feedforward.

x-axis y-axis
0.5 0.5
position error [µm]

position error [µm]

0.25 0.25

0 0

-0.25 -0,25

with feedforward with feedforward

-0.5 -0,5
0 0.5 1 0 0.5 1
time [s] time [s]

Fig. 2.26. Simulated tracking error of the x-axis (left) and the y-axis
(right) for sinusoidal reference signal of an amplitude 1mm and
frequency of 10Hz

With feedforward, a zero tracking error is obtained (as indicated above by

the thicker line). A complete removal of the tracking errors is accomplished
using feedforwards of the velocity signal and of Gm-1(s).

In the next section, correlation between a second order model time domain
performance specifications and its frequency domain characteristics are
analysed.

2.6 Correlation between Velocity and

Position Closed Loops Characteristic
Consider for the y-axis, a cascade control with PI velocity controller and a P
position controller. A speed closed loop transfer function relating the actual
velocity v and the reference velocity vref can be described according to the
following form [35]:

41
Classical Motion Control

v 1
= (2.35)
vref . 1
s +1
ωos

ωos is the velocity closed loop bandwidth. Fig. 2.27 shows the velocity loop
with a PI controller and the system model that is based on equation (2.2).
The system model parameters (B and A) and the PI controller parameters are
defined in Table 2.1 and Table 2.2 respectively.

Fig. 2.27. Schematic diagram of a velocity closed loop

The velocity closed loop transfer function based on the schematic diagram
shown above is:

B
PI ⋅
v s + A
= (2.36)
vref . B
1 + PI ⋅
s+A

A comparison between (2.35) and (2.36) is performed in the next paragraph.

1 PI ⋅ B 1
= ≈ (2.37)
1 s + ( A + PI ⋅ B ) 1
s +1 s +1
ωos PI ⋅ B

ωos = PI ⋅ B ≈ 267.5 rad / s = 42.6 Hz (2.38)

A bode plot of the velocity closed loop transfer function based on the control
scheme in fig. 2.10 and transfer function 2.26 is shown in fig. 2.28. The
velocity closed loop shown in fig. 2.28 has a bandwidth of 41 Hz. This value
is closely estimated in (2.38).

42
 Classical Motion Control

Bode Diagram

0
System: spdtf_cl
Magnitude (dB)

-20 Frequency (Hz): 40.9

Magnitude (dB): -3.03

-40

-60

-80
0

-180
Phase (deg)

-360

-540

-720
0 1 2
10 10 10
Frequency (Hz)

Fig. 2.28. Y-axis: Simulated speed closed loop based on measured

system FRF

Next, the position closed loop bandwidth ωop is estimated based on the
following relationship:

1 ωos
ζ=
2 kv (2.39)
ωop = kv ωos

For kv = 150s-1 and ωos = 267.5rad/s,

ωop = 200 rad / s ≈ 31 Hz

(2.40)
ζ = 0.67

Fig 2.29 shows the measured FRF of the position closed loop obtained using
band-limited white noise as the reference input. The FRF relates the
reference position to actual position. The measured bandwidth of the system
is identified at 28 Hz. This is nearly identical to the estimated bandwidth of
31Hz obtained in (2.40). A definite correlation between theoretical
estimation and actual measurement result concerning the velocity and
position closed loop characteristic has been established.

43
Classical Motion Control

Bode Diagram
20

0
Magnitude (dB)

System : frdFRF
-20 Frequency (Hz): 28
Magnitude (dB): -3
-40

-60

-80

-100

0
Phase (deg)

-90

-180

-270

-360
0 1 2
10 10 10
Frequency (Hz)

Fig. 2.29. Measured FRF of the y-axis position close loop with kv =150 s-1

2.7 Summary
This chapter discusses classical feedback control of motion systems. A
cascade controller structure with a PI velocity loop controller and a P
position loop controller is considered. Theoretical analysis of an ideal system
using cascade P/P and cascade P/PI is presented. A cascade P/PI is then
designed for the linear drive-based xy feed table considered in this thesis
based on measured FRFs of the system. The cascade P/PI controller
parameters are tuned according to gain margin and phase margin
considerations of the model open loop transfer function. Velocity
feedforward and feedforward of the inverse open loop model of the system
G -1(s) are added to improve the tracking performance. A known correlation
between velocity and position closed loop characteristics is successfully
established. This correlation is confirmed using an actual experimental
measurement of the position closed loop.

44
 Chapter 3

Disturbance Forces in Servo

Drives System

3.1 Introduction
Disturbance forces directly affect the positioning and tracking accuracy of
linear drive systems because of the lack of transmission. Two important
disturbance forces are the highly nonlinear friction forces and cutting forces
that occur during milling operation. Effective reduction of the effects of
these disturbances is highly desirable. This requires a deep and detailed
knowledge of the characteristics of these disturbances.

This chapter focuses on the understanding, characterization and modelling of

friction behaviour and cutting forces. Two different friction models are
considered and the identification of the parameters of these models is
discussed. A frequency domain analysis of the spectrum of cutting forces
based on a sequence of cutting force measurements on a milling machine is
provided. First, the friction force characterization and model structures are
discussed and explained in the next section.

3.2 Friction Characterization and Model

Structures
Friction, by definition, refers to the resistance to motion during sliding of
two opposing objects against one another. Friction is present in various
applications and influences many aspects of our daily life. It brings benefit in
various forms. Braking systems and clutches are two examples of friction
effects that benefit us. However, friction is highly undesirable in most
45
Disturbance Forces in Servo Drives System

industrial and mechanical applications. Uncompensated friction force

increases energy consumption and reduces efficiency of these processes. In a
positioning system, at near zero velocity and during motion reversal, the
highly complex non-linear friction behaviour introduces a unique form of
contour tracking error, namely, the quadrant glitch error (fig. 1.1).

Friction is categorized according to its pre-sliding and sliding regimes (see

fig. 3.1). In pre-sliding regime, friction force is predominantly dependent on
displacement. In sliding regime, the friction force is predominantly
dependent on the sliding velocity. The most simple friction models consider
the friction sliding regime only. These models are a static map between
friction force and velocity, e.g. viscous, Coulomb and Stribeck friction
models (see section 3.3).

Fig. 3.1. Pre-sliding and sliding friction regimes

(reproduced from [14])

A first attempt in describing the more complex friction behaviour in pre-

sliding regime was accomplished by Dahl [5]. Since then, several other
advanced and complex friction models have been proposed. This includes
the LuGre friction model [6], the Leuven model [7, 8], and the Generalized
Maxwell-slip (GMS) model [9]. The basic structures of the Dahl and the
LuGre friction models are summarized in the following sections.

3.2.1 The Dahl Model

The Dahl model is used extensively for simulating friction behaviour in ball-
bearing system. The model is an extension to the classic Coulomb friction,
with smooth transitions around the critical zero velocity regions. A
generalized first order differential equation of the position that is a function
of the sign of the velocity v, approximates the hysteresis at presliding
regime. The function is describes below:

46
 Disturbance Forces in Servo Drives System

δd
dF f Ff ⎛ Ff ⎞
= σo 1 − sgn ( v ) sgn ⎜⎜1 − sgn ( v ) ⎟⎟ . (3.1)
dx Fs ⎝ Fs ⎠

Ff and Fs are the total friction force and the static friction force respectively,
σo is the initial stiffness of the contact at velocity reversal, and δd determines
the shape of the hysteresis.

The LuGre friction model is an extension and an improvement over the Dahl
model. The model combines the pre-sliding friction behaviour of the Dahl
model with the steady-state friction characteristic of the sliding regime.

3.2.2 The LuGre Model

The model is based on the concept of averaging deformation of the contact
asperities. The friction force is defined as:

dz
F f = σo z + σ1 + σ 2v. (3.2)
dt

σo, σ1, and σ2 are the asperity stiffness, micro-viscous friction coefficient, and
viscous friction coefficient respectively. The state variable z represents the
average deflection of the asperities and v is the velocity.

dz v
= v − σo z. (3.3)
dt s (v)

s(v) is the Stribeck curve, that is, a decreasing function for increasing
velocity with upper limit and lower limit bounds corresponding to the static
friction force Fs and Coulomb friction force Fc respectively (see section
3.3.1). Vs is the Stribeck velocity and δ is the Stribeck shape factor.

⎛ v
δ ⎞
⎜ − ⎟
s ( v ) = sgn ( v ) ⎜ Fc + ( Fs − Fc ) e
Vs
⎟. (3.4)
⎜ ⎟
⎜ ⎟
⎝ ⎠

The LuGre model fails to characterize the non-local memory behaviour of

the presliding regime. Further improvement to this model is accomplished by
the Leuven model that incorporates a hysteresis function with non local
memory behaviour in presliding regime using a Maxwell-slip (see section
47
Disturbance Forces in Servo Drives System

3.3.3) approximation. The Leuven model is unattractive due to its difficult

implementation. A recent improvement of the Leuven model is the GMS
friction model [9]. This model is discussed in detailed in section 3.3.2.

This thesis considers the widely known static friction model and the recent
and more advanced GMS friction model for friction compensation. First, the
static friction model is discussed in the following section.

3.3 Static Friction Model

3.3.1 Model Structure

Static friction models describe the steady-state friction behaviour in sliding
regime and hence are dependent on the sliding velocity v. The well-known
static friction model considers Coulomb, viscous, and Stribeck friction. This
is described in equation (3.5). Fc, Fs, and σ represent the Coulomb, static and
viscous friction forces respectively. The Stribeck effect represents a
decreasing effect of friction forces with increasing velocity. The Stribeck
friction model parameters are the Stribeck velocity Vs and the Stribeck shape
factor δ.

⎧ ⎛ v δ ⎞ ⎫
⎪ ⎟ + σ ⋅ v ⎪⎬ ⋅ sign ( v ) .
F f ( v ) = ⎨ Fc + ( Fs -Fc ) ⋅ exp ⎜ − (3.5)
⎪⎩ ⎜ Vs ⎟ ⎪⎭
⎝ ⎠

Fig. 3.2 illustrates the different friction components in the static friction
model.

Fig. 3.2. Different friction components in static friction model (3.5).

The next section discusses the estimation of the model parameters for the
considered linear drive xy feed table.
48
 Disturbance Forces in Servo Drives System

3.3.2 Identification of the Static Friction Model

At constant velocity, the motor force equals the friction force and is
represented by the force control command signal. An illustration of a control
command signal measurement during a constant velocity motion is shown in
fig. 3.3.
160
140
control command signal [N]

120
this average
is equivalent
100
to the friction
80 force
60

40

20

-20
0 5 10 15 20 25 30 35 40 45 50
time [s]
Fig. 3.3. Friction force from control command signals at
constant velocity motion of 2.0 mm/s (y-axis)

Constant velocity is enforced using a manually tuned PID controller with a

constant velocity reference signal (a ramp signal). The control command
signal measurements for constant reference velocities of 0.010, 0.040, 0.080,
0.2, 0.4, 0.5, 1, 2, 4, 5, 8, 10, 14, 16, 20, 25, and 30 mm/s were averaged and
recorded. Fig. 3.4 shows these averaged measured friction forces and the
manually fitted friction-velocity map of the static friction force model (3.5).

The identified parameters of the static friction model for both the x and y
axes are tabulated in Table 3.1.

Table 3.1. Static friction model parameters for x and y axes

Fc Fs 1/Vs σ δ
x-axis 122 N 177 N 0.00045 s/µm 0.00095Ns/µm 1
y-axis 105 N 165 N 0.001s/µm 0.00004Ns/µm 1

49
Disturbance Forces in Servo Drives System

Y-Axis X-Axis
200 200
static friction,Fs static friction,Fs
150 150
Coulomb friction,Fc
Coulomb friction,Fc
100 100
friction force [N]

friction force [N]

50 50

0 0

-50 -50

-100 -100

-150 measured -150 measured

model model
-200 -200
-40 -20 0 20 40 -40 -20 0 20 40
velocity [mm/s] velocity [mm/s]

Fig. 3.4. Friction force-velocity mapping and the manually

fitted static friction model using equation (3.5).

The friction forces that act on the bottom x-axis are considerably larger than
the y-axis friction forces. This is because the larger mass x-axis has more
influence on the bearings than the top stage y-axis (the x-axis is supported by
6 Schneeberger MRD 35 carriages on 3 rails compared to the y-axis that is
supported by 4 Schneeberger MRD 35 carriages on 2 rails).

The static friction model fails to describe friction behaviour in pre-sliding

regime. Therefore, a more advanced friction model, that is, the recently
developed Generalized Maxwell-slip (GMS) model is considered.

3.4 Generalized Maxwell-slip Model (GMS)

3.4.1 Model Structure

The GMS friction model is a relatively new model that has great potential
for efficient friction compensation applications. Its main advantage is that it
contains a hysteresis function with non-local memory behaviour to describe
the pre-sliding regime. The GMS friction model incorporates the following
components: (i) the Stribeck curve for constant velocity, (ii) hysteresis
function with non-local memory for the pre-sliding regime, and (iii)
frictional memory for the sliding regime. The structure of this model is
similar to the Maxwell-slip structure [37, 38], that is, it consists of a parallel
connection of N different elementary slip-blocks and springs (see fig. 3.5).

50
 Disturbance Forces in Servo Drives System

Fig. 3.5. N-elementary configuration of the Maxwell-slip structure

Each block represents a generalized asperity of the contact surface that can
either stick or slip and each element i has an elementary stiffness ki, a state
variable zi that describes the element position, a maximum elementary
Coulomb force Wi and a friction output Fi. A new state equation that
characterizes sliding dynamics of each elementary slip-block replaces the
original Coulomb law in the Maxwell-slip friction model structure. The
dynamic behaviour of an elementary model is captured and described
mathematically based on its current condition.

At the time of sticking, which occurs during motion reversal and as velocity
approaches zero, the dynamic behaviour of an elementary slip-block is
described by a spring model with stiffness ki:

dFi
= ki v. . (3.6)
dt

On the other hand, slipping occurs if the elementary friction force Fi equals a
maximum value Wi = αis(v), αi is the normalized sustainable maximum
friction force of each element during sticking and s(v) is the Stribeck curve.
During slipping, the state equation describing the dynamic behaviour of the
elementary slip-block is represented as

dFi ⎛ F ⎞
= sign ( v ) ⋅ C ⋅ ⎜ αi − i ⎟ . (3.7)
dt ⎜
⎝ s ( v ) ⎟⎠

The constant parameter C (equals to 1/Vs in 3.5) indicates the rate at which
the friction force follows the Stribeck effect in sliding. The total friction
force is the summation of the output of all elementary state models and a
viscous term σ (if viscous friction is present).

51
Disturbance Forces in Servo Drives System

N
F f ( v ) = ∑ Fi ( v ) + σ ⋅ v ( t ). (3.8)
i =1

A more detail description of the GMS friction model can be found in [10,
14]. The following section discusses the identification of the GMS model
parameters.

3.4.2 Identification of GMS Model

For the considered test setup, a GMS model with 4 elementary slip-blocks is
selected, yielding a total of 13 model parameters: two parameters (αi‘s and
ki’s) from each of the four elements and another five parameters from the
state equations (3.5) in sliding regime. Increasing the number of elementary
slip-block has minimal effect on the compensation performance of the GMS
model [14]. The identification of the parameters of the state equation (3.5) is
discussed in previous section 3.3.2. These values, for both the x and y axes,
are presented in Table 3.1.

In pre-sliding regime, friction is dominated by the displacement and behaves

as a hysteretic function of displacement with non-local memory behaviour.
This behaviour is characterized by the so-called virgin curve [39]. The virgin
curve is derived from a sinusoidal excitation of the system. The frequency
and amplitude of the sinusoidal input signal are selected to minimize inertia
effect and to remain in the pre-sliding regime. Virgin curves for both axes
are then constructed based on these measurements, and their parameters are
identified. The identification procedures for the construction of the y-axis
virgin curve are explained in the following paragraph.

The virgin curve is constructed using two measurements obtained with

different excitation amplitude. The measurement obtained with the small
excitation amplitude (fig. 3.6(a)) provides detailed information about the
friction-displacement behaviour in pre-sliding regime away from breakaway,
while the measurement obtained with the larger excitation amplitude (fig.
3.6(b)) provides information close to breakaway or in the transition to
sliding. Fig. 3.6(a) and 3.6(b) clearly indicate which part of the measurement
is selected to compose the virgin curve shown in fig. 3.7(a). The selected
parts are combined, reduced by a factor of 2 (the combined parts constitute a
double-stretched version of the virgin curve), and shifted to the origin (0μm,
0N) to generate the virgin curve.

52
 Disturbance Forces in Servo Drives System

80 100
80
60 A B
60
40 A
40

friction force,[N]
friction force,[N]

20 20
0 0
-20
-20
-40
-40
-60
-60 -80
-80 -100
-20 -10 0 10 20 -600 -400 -200 0 200 400 600
position,[µm] position,[µm]
(a) (b)
Fig. 3.6. Y-axis: measured friction force and position for sinusoidal reference signal
of 0.1Hz and amplitudes of (a) 15µm and (b) 450µm

90 90 α4
Kc
80 80 α3
Kb
70 70
friction force,[N]
friction force,[N]

60 60 α2
Ka
50 50
40 40 α1
30 30
Ko
20 20
10 10
0 0
-10 0 10 20 30 40 50 60 -20 0 20 40 60
position,[µm] position,[µm]

(a) (b)
Fig. 3.7. Y-axis: (a) virgin curve and (b) virgin curve with
selected knots and slopes

The four αi’s and ki’s slip-block parameters are estimated by approximating
this virgin curve with a piecewise linear function. The knots and slopes of
this piecewise linear function determine the αi’s and ki’s parameters. First,
the knots of the piecewise linear curve are selected manually. Fig 3.7b shows
the selected knots αi’s and the corresponding slopes Ki’s. Equation (3.9)
summarizes the relationship between [K0, Ka, Kb, Kc], [α1, α2, α3, α4], and [k1,
k2, k3, k4] based on fig. 3.8 and the theory of superposition.

53
Disturbance Forces in Servo Drives System

100
friction force,[N]

80 Kc
Kb
60
Ka
40
Ko
20
virgin curve
0
0 10 20 30 40 50
position,[µm]
friction force,[N]

50 α1
40
k1
30
20 α2
k2 k3
10 α3
k4
α4
0
0 10 20 30 40 50
position,[µm]

Fig. 3.8. Y-axis GMS model parameters identification

α1 + α 2 + α3 + α 4 = ∑ Wi
k1 + k2 + k3 + k4 = K 0
k2 + k3 + k4 = K a (3.9)
k3 + k4 = Kb
k4 = K c .

ΣWi is the friction force at the moment of breakaway. First, the values of
[K0, Ka, Kb, Kc] are obtained from the measured virgin curve in fig. 3.7b for
the manually selected knots [α1, α2, α3, α4]. Finally, based on these values
and using (3.9), the values for [k1, k2, k3, k4] are obtained. Table 3.2
summarizes the identified GMS model parameters of the y-axis.

Table 3.2. GMS slip-blocks model parameters for the y-axis

αi [N] α1 = 50 α2 = 20 α3= 15 α4 = 3.4

ki [N/μm] k1 = 99.94 k2 = 1.364 k3 = 1.081 k4 = 0.119

This technique is repeated for the GMS friction identification of the x-axis.
The virgin curve is first constructed (see fig. 3.9 and 3.10) based on
sinusoidal excitation at amplitudes of 15µm and 450µm, and a frequency of
0.1Hz.

54
 Disturbance Forces in Servo Drives System

200 200
B
150
150 A
100 100

friction force [N]

friction force [N]

50 50

0 0

-50 -50
A
-100 -100

-150 -150

-200 -200
-15 -10 -5 0 5 10 15 -400 -200 0 200 400
position [µm] position [µm]

(a) (b)
Fig. 3.9. X-axis: Measured friction force and position for sinusoidal
reference signal of 0.1Hz and amplitudes of (a) 15µm and (b) 450µm

160 160
α4
140 Kc
140 α3
Kb
120 120
friction force [N]
friction force [N]

α2
100 100
Ka
80 80

60 60
α1 Ko
40 40

20 20

0 0
0 10 20 30 40 50 -20 -10 0 10 20 30 40 50
position [µm] position [µm]

(a) (b)
Fig. 3.10. X-axis: (a) virgin curve (b) virgin curve with
selected knots and slopes (right)

The four αi’s and ki’s slip-block parameters are obtained based on fig. 3.10,
fig. 3.11, and equation (3.9). The identified GMS model parameters for the
x-axis are summarized in Table 3.3.

Table 3.3. GMS slip-blocks model parameters for the x-axis

αi [N] α1 = 65 α2 = 55 α3= 26 α4 = 6.5
ki [N/μm] k1 = 61.80 k2 = 9.852 k3 = 4.490 k4 = 0.325

55
Disturbance Forces in Servo Drives System

150
friction force, [N]

Kc
Kb
100
Ka
50
Ko
0
0 5 10 15 20 25 30 35 40
position [µm]
80
friction force[N]

α1
60
α2
k1
40
k2 α3
20
k3 k4
0
α4
0 5 10 15 20 25 30 35 40
position [µm]

Fig. 3.11. X-axis: GMS model parameters identification

3.5 Cutting Forces

Cutting forces act directly on the linear motor. Uncompensated cutting
forces generate position and contour tracking errors. Knowledge of the
cutting force characteristics provides valuable information on possible
approaches and strategies for its effective compensation. For this purpose, a
series of milling cutting processes were executed to measure and analyse the
resulting cutting forces. These cutting processes were performed on an actual
industrial milling machine and not on the experimental test setup itself.
Details of these experiments are discussed in the following section.

3.5.1 Cutting Force Measurements and Analysis

Cutting force measurements were performed using a Kistler 3-component
dynamometer (type 9265B) force measurement unit, yielding cutting forces
measurements in two horizontal directions, Fx and Fy. The experimental
conditions of the milling process are summarized in Table 3.4.

56
 Disturbance Forces in Servo Drives System

Table 3.4. Cutting process parameters

Parameters Value Parameters Value
Work piece
Aluminium Spindle speed 1360 rpm
material
Cutter diameter 10mm Table feed rate 320mm/min
No. of edges 3 Depth of cut 1mm and 3mm
Sampling
Feed per tooth 0.1 mm per tooth 3000Hz
frequency

The cutting force measurements Fx and Fy for cutting depth of 1mm and
3mm are illustrated in fig. 3.12 and fig. 3.13 respectively. The cutting
processes were performed along the x-axis of the milling table (see fig.
3.14).

x-axis y-axis

200 200
start of cutting
150 150

100 100

50 50
force [N]

force [N]

0 0

-50 -50

-100 -100

-150 -150

0 2 4 6 0 2 4 6
time [s] time [s]

Fig. 3.12. Cutting force in x and y direction for cutting depth of 1mm

57
Disturbance Forces in Servo Drives System

x-axis y-axis
400 400
start of cutting
300 300

200 200
force [N]

force [N]
100 100

0 0

-100 -100

-200 -200
0 2 4 6 0 2 4 6
time [s] time [s]

Fig. 3.13. Cutting force in x and y direction for cutting depth of 3mm

cutting tool

workpiece

Kiestler
measurement
unit

Fig. 3.14. Cutting force measurements on a workpiece

using Kistler dynamometer force sensor along the x-axis

Spectral analysis of the measured cutting forces for 1mm and 3mm depth of
cut is shown in fig. 3.15 and fig. 3.16 respectively.

58
 Disturbance Forces in Servo Drives System

x-axis y-axis

120 120

100 100

80 80
force [N]

force [N]
60 60

40 40

20 20

0 0
0 100 200 300 0 100 200 300
frequency [Hz] frequency [Hz]

Fig. 3.15. Spectral of cutting force along x and y direction,

for cutting depth of 1mm

x-axis y-axis

120 120

100 100

80 80
force [N]

force [N]

60 60

40 40

20 20

0 0
0 100 200 300 0 100 200 300
frequency [Hz] frequency [Hz]

Fig. 3.16. Spectral of cutting force along x and y direction,

for cutting depth of 3mm

The fundamental frequency is 23.3Hz, that is, the frequency of the spindle
rotation. Table 3.5 and Table 3.6 summarized the first 15 harmonic
components of the measured cutting forces for 1mm depth of cut for x and y
axes.

59
Disturbance Forces in Servo Drives System

3.5.2 Artificial Cutting Force

An artificial cutting force, based on the first fifteen harmonic components of
the measured cutting force, is synthesized for simulation and experimental
validation purposes. The synthesized artificial cutting force is generated
according the following equation (z = x or y axis):

15
( )
F%z ( t ) = ∑ Ai ⋅ cos ( ωi t ) − Bi sin ( ωi t ) .
i =1
(3.10)

Ai and Bi are the amplitude and the phase of the ith harmonic component of
the measured cutting forces. These values are represented in Table3.5 and
3.6.

Table 3.5. X-axis: Harmonic contents of the measured cutting force for 1mm
depth of cut
Harmonics Freq. (Hz) Ai Bi

1 23.29 -2.147e+004 6.615e+004

2 46.59 1.45e+004 7.434e+004
3 69.88 1.544e+005 -1.353e+005
4 93.18 -2.491e+004 -4.705e+004
5 116.47 9.209e+003 1.860e+004
6 139.76 -7.812e+004 -1.512e+004
7 163.06 9.260e+002 1.460e+004
8 186.35 -1.589e+004 5.733e+002
9 209.65 3.384e+004 7.720e+003
10 232.94 8.626e+003 -1.410e+003
11 256.24 1.836e+003 1.618e+003
12 279.53 6.953e+003 9.00e+002
13 302.83 -4.101e+003 -1.294e+003
14 326.12 1.305e+004 2.9406e+002
15 349.41 8.234e+003 3.321e+003

60
 Disturbance Forces in Servo Drives System

Table 3.6. Y-axis: Harmonic contents of the measured cutting force for 1mm
depth of cut
Harmonics Freq. (Hz) Ai Bi

1 23.29 -6.792e+004 -1.263e+005

2 46.59 4.497e+003 -7.883e+004
3 69.88 4.833e+004 1.453e+005
4 93.18 4.546e+004 2.417e+004
5 116.47 -1.020e+004 1.563e+004
6 139.76 9.306e+004 2.657e+004
7 163.06 -4.213e+003 -1.180e+004
8 186.35 1.061e+004 3.850e+003
9 209.65 -6.261e+004 -1.425e+004
10 232.94 -4.878e+003 6.895e+003
11 256.24 -6.341e+003 -2.983e+003
12 279.53 1.574e+004 1.280e+004
13 302.83 -5.417e+003 4.901e+003
14 326.12 -3.133e+003 2.225e+003
15 349.41 -8.592e+003 -2.971e+003

A comparison between the measured cutting force and the synthesized

artificial cutting force is shown in fig. 3.17. The synthesized cutting force
signal closely resembles the actual cutting force.

61
Disturbance Forces in Servo Drives System

100
actual cutting force
80 synthesized cutting force
error
60

40

20
Force [N]

-20

-40

-60

-80

-100
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
time [s]

Fig. 3.17. X-axis: Comparison between the measured cutting

force and the synthesized cutting force for 1mm depth of cut

3.6 Summary
This chapter discusses the characteristics of the disturbance forces
considered in the thesis; namely, the friction and the cutting forces. Two
friction models, the widely known static friction model, and the recent more
advanced GMS friction model are discussed. Unlike static friction model, the
GMS friction model describes friction behaviour in both the sliding and the
pre-sliding regimes. The GMS friction model is more complex and more
difficult to identify. The identification of the two friction models for the two
axes of the experimental test setup considered in this thesis is presented in
detail. The characteristics of cutting forces based on cutting force
measurements performed during an actual milling cutting process is
presented. The spectral analysis of the cutting force reveals its harmonics
component. Artificial cutting forces are synthesized from these measured
cutting forces for simulation and experimental validation of cutting forces
compensation performance of different techniques discussed in Chapter 5.

62
 Chapter 4

Friction Force Compensation

Design

4.1 Introduction
Friction forces reduce the positioning and tracking accuracy in mechanical
systems. Various friction compensation techniques have been described and
presented in literature and have been highlighted in section 1.2.2. This
chapter focuses on the application of two compensation techniques on the
considered test setup: friction model feedforward and a disturbance observer
that does not rely on a friction model.

Feedforward friction compensation is validated for two different friction

models: a static friction model (section 3.3) and a Generalized Maxwell-slip
model (section 3.4). The disturbance observer is based on the inverse-model
based disturbance observer design by Ohnishi [26]. The design and the
analysis of this disturbance observer for the considered system are presented.
The model applied in the disturbance observer is an approximate second
order linear model of the considered system (equation 2.2). Finally, both
techniques are combined and the friction compensation performances are
compared [40]. Both numerical and experimental validations are performed.
The magnitude of the quadrant glitch that appears near zero velocity or
during motion reversal becomes the basis of performance comparison.

63
Friction Force Compensation Design

4.2 Friction Model-Based Feedforward

The performance of feedforward friction compensation based on a static and
a GMS friction model is evaluated numerically and experimentally. The
general control scheme is shown in fig. 4.1 and is based on the control
structure presented in section 2.5.4 (fig. 2.25). The control scheme of each
individual axis consists of a cascade P/PI position controller that includes a
proportional (P) position and a PI velocity feedback controllers, a notch filter
N(s), a first order velocity estimator Vest(s) (2.22), a velocity feedforward,
and a feedforward Gm-1 based on the inverse system model Gm (equation 2.2)
and reference position. The friction model feedforward is inserted at the
input of the plant.

Fig. 4.1. Friction compensation scheme using friction model-

based feedforward (kf is the force constant).

The following section discusses the numerical validation of friction

feedforward compensation using two different friction models.

4.2.1 System Transfer Function with a GMS Friction

Term for Friction Simulation
The friction compensation performance is validated numerically in
MATLAB/ Simulink. The system model Gm’ (see figure 4.1) consists of a
linear second order model with delay (equation 2.2) in combination with the
GMS friction model. Fig. 4.2 shows this system model in detail.

64
 Friction Force Compensation Design

Fig. 4.2. System transfer function with a GMS friction model to

characterize the complex friction behaviour.

The second order model parameters (parameters A and B in equation (2.2)

and Table 2.1) are adapted manually such that the FRF of this system model
Gm’ corresponds to that of the original linear model Gm (equation 2.2 with
parameters in Table 2.1). Extending this original linear model with the GMS
model yields a FRF that differs significantly from the FRF of the original
linear model only. This is shown in figure 4.3a for the y-axis. The FRF of the
original linear model is obtained by evaluating equation 2.2 with parameters
in Table 2.1 in the frequency domain (replacing s by jω).

The FRF of the linear model in combination with the GMS model is
obtained by exciting this combined model with a band-limited white noise
and estimating the FRF based on input and simulated output using the H1
[33] estimator. Fig. 4.3a clearly shows a difference in the phase diagram
between both FRF's at low frequencies. In order to improve the
correspondence between both FRF's, the A and B parameters of the linear
model (2.2) were adapted manually. Better correspondence is obtained with
the A' and B' parameter values shown in Table 4.1. This improved
correspondence is clearly visible in fig. 4.3b. This procedure is applied to the
models of both the x and y axes.

To conclude, the system model used for numerical analysis in this section is
model equation (2.2) with the A' and B' parameters of Table 4.1 in
combination with the GMS friction model (equation 3.6, 3.7, 3.8) and GMS
parameter values of Tables 3.2 and 3.3.

Table 4.1. Original and revised system transfer function parameters

A [s-1] A’ [s-1] B [m/volt.s2] B’ [m/volt.s2]
x-axis 28.57 2.5 4.526 4.526
y-axis 20 0.095 8.916 8.916

65
Friction Force Compensation Design

Bode Diagram
150
Magnitude (dB)

100

50

-50 original linear plant tf. based on measured FRF

original linear plant tf. linear model+ GMS model
-100
0
-90
Phase (deg)

-180
-270
-360
-450
-540
1 2
10 Frequency (Hz) 10

(a)

Bode Diagram
150
Magnitude (dB)

100

50

-50 original linear plant tf. based on measured FRF

revised linear plant tf + GMS model
-100
0
-100
Phase (deg)

-200
-300

-400
-500
1 2
10 Frequency (Hz) 10

(b)

Fig. 4.3. System model parameters adjustment for system structure that
includes the GMS friction term: (a) using the original model
parameters and (b) using updated model parameters.

66
 Friction Force Compensation Design

4.2.2 Friction Feedforward: Numerical Results

Feedforward friction compensation performance for both axes is validated
numerically using MATLAB/Simulink, based on the control scheme shown
in fig. 4.1. The reference trajectory is a sinusoidal signal which corresponds
to a frequency of 0.53 Hz with an amplitude of 30mm and a maximum
tracking velocity of 100mm/s. Friction model feedforward using the static
friction model (see equation 3.5 and Table 3.1) and the GMS friction model
(see equation 3.6-3.8 and Table 3.2) are applied. The position signal and the
tracking errors for the y-axis are shown in fig. 4.4. Significant quadrant
glitch magnitudes are observed for the case where no friction compensation
is applied (indicated by the circle in fig. 4.4a).
4 No Friction 4
Static Friction 4
GMS Friction
x 10 Model FF. x 10 Model FF. x 10 Model FF.
4 4 4
position [µm]

position [µm]

position [µm]
2 2 2

0 0 0

-2 -2 -2

-4 -4 -4
0 5
0.5 1 1.5 0 5
0.5 1 1.5 0 5
0.5 1 1.5
x 10 x 10 x 10
velocity [µm/s]

velocity [µm/s]
velocity [µm/s]

1 1 1

0 0 0

-1 -1 -1

0 0.5 1 1.5 0 0.5 1 1.5 0 0.5 1 1.5

40 20 20
tracking error,[µm]
tracking error,[µm]
tracking error,[µm]

20 10 10

0 0 0

-20 -10 -10

-40 -20 -20

0 0.5 1 1.5 0 0.5 1 1.5 0 0.5 1 1.5
time [s] time [s] time [s]

(a) (b) (c)

Fig. 4.4. Y-axis: Simulated position, velocity and tracking error for three
different cases of friction compensation techniques (a) no friction feedforward,
(b) Stribeck friction model feedforward, (c) GMS model feedforward.

The results show a significant reduction of the quadrant glitch magnitude

using the static and GMS friction model feedforward. The GMS friction
model feedforward almost completely compensates the hard non-linear
friction behaviour. A complete compensation of friction using the GMS
friction model feedforward is not possible as a result of the delay term in the
model structure. Fig. 4.5 shows that complete friction compensation is
obtained for system where the delay term is removed.
67
Friction Force Compensation Design

3 3

2 2
tracking error [µm]

tracking error [µm]

1 1

0 0

-1 -1

-2 -2

-3 -3
0 0.5 1 1.5 0 0.5 1 1.5
tim e [s] tim e [s]
(a) (b)
Fig. 4.5. Y-axis: Simulated tracking error using GMS model feedforward for (a)
system model with delay and (b) with the delay removed.

Similar simulations are performed for the x-axis. Table 4.2 summarizes the
results of friction compensation for both axes based on the magnitude of the
quadrant glitches (shown in circle). The magnitude of the quadrant glitches
are reduced from a maximum of 31 micrometer (no friction compensation)
to near zero (GMS model feedforward).

Table 4.2. Simulated friction compensation performance for a reference

sinusoidal signal with an amplitude of 30mm and a maximum tracking
velocity of 100mm/s
Friction Compensation Quadrant Glitch [µm]
Methods x-axis y-axis
No friction feedforward 15 31
Static friction feedforward 5 8.3
GMS friction feedforward 4 1.2

In feedforward friction compensation, the performance depends on the

accuracy of the friction model. The performance robustness against friction
model parameters variation is analysed next. A simulation using a
feedforward of an adjusted GMS friction model parameters is performed
(system has no delay). Fig. 4.6 shows the position error in comparison to the
result of fig. 4.5(b). Table 4.3a and 4.3b list the original GMS friction model
parameters and the adjusted model parameters respectively.

68
 Friction Force Compensation Design

4 4

2 2
position error [µm]

position error [µm]

0 0

-2

-4 -4

0 0.5 1 1.5 2 0 0.5 1 1.5 2

time [s] time [s]

(a) (b)
Fig. 4.6. Y-axis: Simulated tracking error using feedforward of (a) adjusted
GMS model and (b) matched GMS model, for system both without delay.

Table 4.3a. Y-axis: Original GMS slip-blocks model parameters

αi [N] α1 = 50 α2 = 20 α3 =15 α4 = 3.4

ki [N/μm] k1 = 99.94 k2 = 1.364 k3 = 1.081 k4 = 0.119

Table 4.3b. Y-axis: Adjusted GMS slip-blocks model parameters

αi [N] α1 = 57 α2 = 26.5 α3 = 15 α4 = 3.4

ki [N/μm] k1 =110.0 k2 =2.0 k3 = 1.5 k4 = 0.119

The experimental validation of the friction model-based feedforward

compensation is discussed in the next section.

4.2.3 Friction Feedforward: Experimental Results

Experimental validation of the feedforward friction model compensation is
performed on both axes individually. The reference is again a sinusoidal
signal with an amplitude of 30mm and a maximum tracking velocity of
100mm/s. Friction model feedforward, using the static friction model (see
equation 3.5 and Table 3.1) and the GMS (see equation 3.6-3.8 and Table
3.2) friction model are applied. The measured position and the position
tracking errors are recorded. Fig. 4.7 shows for the y-axis, a significant
improvement in tracking accuracy when applying the static friction model
and GMS friction model feedforward. Table 4.4 lists a summary of the
friction compensation results for both axes.
69
Friction Force Compensation Design

4 No Friction 4 4 GMS
x 10 Feedforward x 10 Static Model FF. x 10 Friction Model
3 3 3
position,[µm]

position,[µm]

position,[µm]
0 0 0

-3 -3 -3

0 0.5 1 1.5 0 0.5 1 1.5 0 0.5 1 1.5

time,[s] time,[s] time,[s]
60 20 20
tracking error,[µm]

tracking error,[µm]

tracking error,[µm]
40
10 10
20
0 0 0
-20
-10 -10
-40
-60 -20 -20
0 0.5 1 1.5 0 0.5 1 1.5 0 0.5 1 1.5
time,[s] time,[s] time,[s]

(a) (b) (c)

Fig. 4.7. Y-axis: Measured position and tracking error for (a) no friction
feedforward, (b) static friction model feedforward, (c) GMS feedforward.

Table 4.4. Measured friction compensation performance for a reference

sinusoidal signal with an amplitude of 30mm and a maximum tracking
velocity of 100mm/s
Friction Compensation Quadrant Glitch [µm]
Methods x-axis y-axis
No friction feedforward 15 35
Static friction feedforward 5 13
GMS friction feedforward 4 10

The GMS friction model feedforward yields the least quadrant glitch
magnitudes on both x and y axes. A 71% reduction in the magnitudes of the
quadrant glitches are observed with GMS friction model feedforward
(compared to the uncompensated case). The improvement obtained using the
GMS friction model over static friction model is small for this reference
trajectory which has a maximum tracking velocity of 100mm/s. The use of a
complex GMS friction model cannot be motivated for this relatively fast
motion. However, by further reducing the tracking velocity to 10mm/s, the
pre-sliding regime becomes more dominant and the benefit of using the
GMS friction model is more pronounced. This is shown explicitly in fig. 4.8
and fig. 4.9 for the x and y-axis respectively.
70
 Friction Force Compensation Design

4 4
x 10 Static friction FF. x 10 GMS friction Model FF.
4 4
position [µm]

2 2

position
0 0

-2 -2

-4 -4
0 5 10 15 0 5 10 15
time [s] time [s]
20 20

10 10
glitch [µm]

glitch
0 0

-10 -10

-20 -20
0 5 10 15 0 5 10 15
time [s] time [s]

(a) (b)
Fig. 4.8. X-axis: Measured quadrant glitches for maximum tracking
velocity of 10mm/s: (a) static friction model feedforward and (b)
GMS model feedforward.

4 4
x 10 Static Model FF. x 10 GMS Friction Model FF.

3 3
position,[µm]

position,[µm]

0 0

-3 -3

0 5 10 15 20 0 5 10 15 20
time [s] time [s]
20 20
tracking error,[µm]

tracking error,[µm]

10 10

0 0

-10 -10

-20 -20
0 5 10 15 20 0 5 10 15 20
time [s] time [s]

(a) (b)
Fig. 4.9. Y-axis: Measured quadrant glitches for maximum tracking
velocity of 10mm/s: (a) static friction model feedforward and (b)
GMS model feedforward.
71
Friction Force Compensation Design

Next, circular tests are performed, where friction is compensated with static
and GMS friction models feedforward, simultaneously on the x and y axes.
These results are shown in fig. 4.10 for a tangential tracking velocity of
100mm/s. The tracking errors are amplified by a factor of 400 for graphical
presentation purposes.

no friction feedforward static friction model ff. GMS friction model ff.
40 40 40
position Y [mm]

position Y [mm]
position Y [mm]

20 20 20

0 0 0

-20 -20 -20

-40 -40 -40

-40 -20 0 20 40 -40 -20 0 20 40 -40 -20 0 20 40
position X [mm] position X [mm] position X [mm]
30 30 30
radial error [µm]

radial error [µm]

radial error [µm]

20 20 20

10 10 10

0 0 0

-10 -10 -10

0 90 180 270 360 0 90 180 270 360 0 90 180 270 360
angle [degree] angle [degree] angle [degree]

(a) (b) (c)

Fig. 4.10. Measured contour and radial tracking error at tangential
tracking velocity of 100mm/s: (a) no friction feedforward, (b) static
friction model feedforward, (c) GMS model feedforward.

Previous results have shown that the advantage of using the GMS friction
model is not clearly demonstrated at higher tracking velocity. The circular
tests are repeated at a lower tangential tracking velocity of 10mm/s. The
measured contour and the radial errors for the two friction models are shown
in fig. 4.11.

72
 Friction Force Compensation Design

static model (error x400) GMS model (error x400)

30 30
position Y [mm]

position Y [mm]
15 15

0 0

-15 -15

-30 -30
-30 -15 0 15 30 -30 -15 0 15 30
position X [mm] position X [mm]
20 20
radial error [µm]

radial error [µm]

10 10

0 0

-10 -10

-20 -20
0 90 180 270 360 0 90 180 270 360
angle [degree] angle [degree]

(a) (b)
Fig. 4.11. Measured contour and radial tracking errors for tangential
tracking velocity of 10mm/s: (a) static friction model feedforward and
(b) GMS model feedforward.

The experimental results presented thus far have identified the advantage of
the GMS friction model over the more simple static friction model for
friction compensation especially at lower tracking velocity, where the pre-
sliding regime becomes more dominant. The results are not perfect because
of the imperfection of the GMS friction model.

The next section considers a friction model-free approach for friction

compensation, namely, the inverse-model-based disturbance observer [26,
27].

4.3 Inverse- Model- Based Disturbance

Observer
Inverse-model-based disturbance observer has been studied and applied in
various high accuracy positioning systems; namely the hard disk drive and
high-speed servo systems [41, 42]. In this thesis, the disturbance observer is
applied as an add-on module to the cascade P/PI controller for friction force
and cutting force compensation. Its structure and design is based on the work
of Ohnishi [26] that has been further refined by Umeno and Hori [43]. The
73
Friction Force Compensation Design

disturbance observer estimates disturbance forces that can be either real

disturbances acting on the system, for example cutting forces if present, or
forces representing the linear model and friction model inaccuracies. The
disturbance force estimate equals the difference between the control
command signal and the input obtained by filtering the system output signal
z with the inverse of a selected plant model Gn(s). Fig. 4.12 shows a
schematic diagram of the system G(s) and the inverse-model-based
disturbance observer. A low pass filter, known as the Q-filter [28], is added
to preserve stability. Fig. 4.13 represents an equivalent block diagram to the
control scheme shown in fig. 4.12.

Fig 4.12. Block diagram of a system with an inverse-model-based

disturbance observer.

Fig 4.13. Equivalent block diagram of a system with an inverse

model-based disturbance observer.

The disturbance observer design freedom is restricted to the selection of the

inverse model Gn-1 and the characteristics of the filter Q. For the considered
application, the inverse model Gn-1 is the inverse of the nominal second order
plant model which corresponds to model (2.2) without delay and parameter
values presented in Table 2.1. The delay is omitted from model (2.2) because
this would yield a non-causal inverse. The selection of the filter Q is
discussed in the following section 4.3.1.
74
 Friction Force Compensation Design

4.3.1 Q-filter Design and Stability Analysis

Consider the equivalent control scheme shown in fig. 4.13. The following
transfer functions can be derived:

Z (s) G ⋅ Gn
GUpi − z ( s ) = = ; (4.1)
U pi ( s ) Gn + Q ( G − Gn )

Z (s) G ⋅ Gn (1 − Q )
Gd − z ( s ) = = ; (4.2)
D(s) Gn + Q ( G − Gn )

Z (s) G ⋅Q
Gn − z ( s ) = = . (4.3)
N (s) Gn + Q ( G − Gn )

The above transfer functions show that if Q → 1, the followings are true:

GUpi − z ( s ) = Gn ;
(4.4)
Gd − z ( s ) = 0.

Secondly, if Q → 0, then,

Gn ( s ) = 0. (4.5)

Equation (4.4) shows that as Q → 1 at low frequency, the closed loop

behaves as the nominal plant Gn(s), and the effect of the disturbance force on
the output position z is zero. In addition, as Q → 0 at higher frequency, the
effect of noise on the output position z becomes zero (equation 4.5). These
analyses confirm the necessity of the transfer function Q being a low pass
filter. Q is also necessary in order to realize G-1(s). Ryoo, Doh, and Chung
[44] have suggested the following structure for the low pass filter Q:
m
i
∑ qi s
Q(s) = i =0 (4.6)
( s + ωc ) n

where; qi’s are the filter numerator coefficients, ωc is the cut-off frequency
and m and n are the order of the numerator and the order of the denominator
respectively, with n > m. The difference in n-m is used to make Q/Gn proper.

75
Friction Force Compensation Design

According to Kwon [45], the coefficients qi’s can be determined from

binomial coefficients of the following form:

n!
qi = ⋅ ωcn −i (4.7)
( ( n − i )!i! )
Kempf and Kobayashi [24] have illustrated the significance of ωc on the
performance of the disturbance observer. A higher cut-off frequency is
desired for better disturbance attenuation. However, the bandwidth of the
filter Q is limited by the un-modelled dynamics of the system. These un-
modelled dynamics are expressed as a multiplicative perturbation Δ(f):

(
Ĝ ( f ) = Gn ( f ) 1 + Δ ( f ) ) (4.8)

Ĝ ( f ) − Gn ( f )
Δ( f ) = (4.9)
Gn ( f )

Ĝ(f) represents the real system dynamics. Since the real system dynamics are
nonlinear and the expression is only valid for linear systems, the measured
FRF of the system, presented in fig. 2.4 for the x and y axes respectively, are
used for G(f) in equation (4.9). In the absence of any un-modelled dynamics,
the disturbance observer loop, described by the equivalent diagram shown in
fig. 4.13, has an open loop transfer function equal to:

D% ( s ) Q(s)
Gol ( s ) = = .. (4.10)
U PI ( s ) 1 − Q(s)

The complementary sensitivity transfer function T(s) of the disturbance

observer loop is Q(s). Based on H∞ theory, it can be stated that the robust
stability of the disturbance observer loop is guaranteed only if,

T ( jω ) ⋅ Δ ( jω ) ≤1 (4.11)
∞

or if,

1
Q(s) ≤ (4.12)
Δ(s)

76
 Friction Force Compensation Design

This means that the magnitude of the filter Q has to lie below the |1/Δ| line.
The |1/Δ| line is constructed based on equation (4.9). Fig. 4.14 shows the
|1/Δ| curves for both axes and the designed Q filters of which the parameters
are presented in Table 4.5. The bandwidth of the filter Q is 60 Hz and 90 Hz
for the x and y axes respectively. Figure 4.14 illustrates that further
increasing the bandwidth of the filter Q reduces the stability margins and
increases the possibility of the filter Q line intersecting the |1/Δ| curve. A
reasonable margin is required to ensure stability during actual practical
implementation of the disturbance observer on the considered test setup.
Bode Magnitude Diagram (x-axis) Bode Magnitude Diagram (y-axis)
40 30

20 1/∆
20 1/∆
10
Magnitude (dB)

Magnitude (dB)

0 0
Q(s)
Q(s) -10

-20
-20

-30
-40
-40

-60 -50
1 2 1 2
10 10 10 10
Frequency [Hz] Frequency [Hz]

(a) (b)
Fig. 4.14. Bandwidth limitation of the filter Q for (a) x-axis
and (b) y-axis.

Table 4.5. Characteristics of the low pass filter Q

m n wc Q
1.421e005
x-axis 0 2 60Hz Q(s ) = 2
s + 754 s + 1.421e005
3.198e005
y-axis 0 2 90Hz Q(s ) = 2
s + 1131s + 3.198e005

The order of the filter numerator m and denominator n are selected such that
the transfer function Q(s)/Gn(s) is proper that is n-m ≤ 2 (since the relative
degree of the nominal plant model Gn(s) is two).

77
Friction Force Compensation Design

4.3.2 Loops Characteristic with Disturbance Observer

Adding this disturbance observer changes the characteristic of the velocity
loop and the position loop. This section analyses these loops. The loops
transfer functions are analysed to determine the influence of the added
disturbance observer on the loops characteristics (i.e. gain margin, phase
margin, and stability). Fig. 4.15 shows the schematic diagram of the control
scheme consisting of the cascade P/PI controller and the disturbance
observer.

Fig. 4.15. Cascade P/PI position control with an inverse model-based

disturbance observer.

In order to calculate these loops’ characteristics, G(s) is replaced by the

measured system FRFs (Ĝ) shown in figure 2.4, and the different transfer
functions in this control scheme are evaluated at the corresponding
frequencies, that is s is replaced by jω. Fig. 4.16 shows an equivalent control
scheme that is based on fig. 4.13 and 4.15.

Fig. 4.16. Cascade P/PI position control with equivalent block

diagram of the inverse model-based disturbance observer.

78
 Friction Force Compensation Design

First, the velocity loop with the added disturbance observer is considered.
The velocity open loop, closed loop, and sensitivity transfer functions equal:

Z& ( s ) ˆ ⋅ G ⋅V ( s ) ⋅ N ( s )
PI ⋅ G n est
Vdob _ ol = = , (4.13)
Ev ( s ) (
Q G ˆ −G +G
n ) n

Z& ( s ) ˆ ⋅ G ⋅V ( s ) ⋅ N ( s )
PI ⋅ G n est
Vdob _ cl = = , (4.14)
U p (s) ˆ
( )
ˆ ⋅ G ⋅V ( s ) ⋅ N ( s)
Gn + Q G − Gn + PI ⋅ G n est

1
Sv _ dob ( s ) = . (4.15)
ˆ
(
Gn + Q G − Gn + PI ⋅ G n est )
ˆ ⋅ G ⋅V ( s ) ⋅ N ( s )

Fig. 4.17 and 4.18 illustrate the influence of the disturbance observer on the
velocity open loop and closed loop characteristics for the x and y axes
respectively.
x-axis
40
Magnitude (dB)

20

-20

-40

-60
90

0
Phase (deg)

-90

-180

-270 no observer
with observer

-360
1 2
10 10
Frequency (Hz)

(a)

79
Friction Force Compensation Design

y-axis
40
Magnitude (dB)

20

-20

-40

-60
90

0
Phase (deg)

-90

-180
no observer
-270
with observer

-360
1 2
10 10
Frequency (Hz)

(b)
Fig. 4.17. Velocity open loop transfer function for (a) x-axis and
(b) y-axis for system with and without a disturbance observer.

Table 4.6 summarizes the gain margin and phase margin for the system with
and without the disturbance observer, based on fig. 4.17.

Table 4.6. Gain margin and phase margin of x and y axes velocity open loop

Gain Margin Phase Margin

no observer with observer no observer with observer
x-axis 12.332 dB 11.02 dB 74.514 deg 67.345 deg
y-axis 14.622 dB 12.938 dB 71.888 deg 65.11 deg

The inverse model-based disturbance observer reduces the gain margin and
the phase margin of the cascade P/PI controller. The gain margin and the
phase margin of both axes were reduced by nearly 1.5dB and 7 degrees
respectively. For both axes, the gain margin and the phase margin remain at
acceptable values and ensure good transient response and stability margin.

80
 Friction Force Compensation Design

x-axis
0
Magnitude (dB) -10

-20

-30

-40

-50
0
Phase (deg)

-90

-180

-270 no observer
with observer

-360
1 2
10 Frequency (Hz) 10

(a)
y-axis
0

-10
Magnitude (dB)

-20

-30

-40

-50
0
Phase (deg)

-90

-180

-270 no observer
with observer

-360
1 2
10 10
Frequency (Hz)

(b)
Fig. 4.18. Velocity closed loop transfer functions for (a) x-axis and
(b) y-axis for system with and without a disturbance observer.

Fig. 4.18 shows the bandwidth improvement for velocity loops of both axes.
A nearly 5 Hz increase in bandwidth is recorded for each of the axes. Similar
conclusion follows from the velocity loops sensitivity curves (equation 4.15,
shown in fig. 4.19). Here, an average bandwidth improvement of 2Hz is
observed. The arrows clearly indicate the reduction in the sensitivity
magnitude at lower frequency and the corresponding increases in the
magnitude at the higher frequency- the waterbed effect [46].

81
Friction Force Compensation Design

Bode Magnitude Diagram (x-axis) Bode Magnitude Diagram (y-axis)

5 5
Sspd Sspd
0 Freq(Hz): 28.9 0 Freq (Hz): 28.9
Mag(dB): -3.01 Mag(dB): -3.02

Magnitude (dB)
Magnitude (dB)

-5 -5
Sspd_dob Sspd_dob
Freq(Hz): 29.9 Freq (Hz): 32
Mag(dB): -3.01 Mag(dB): -3.02
-10 -10

-15 -15
no dist. observer no dist. observer
w ith dist. observer w ith dist. observer
-20 -20
1 2
10 10 1 2
10 10
Frequency [Hz] Frequency [Hz]

Fig. 4.19. Effect of the disturbance observer on the velocity loop

sensitivity transfer functions for x-axis (left) and y-axis (right)

Next, the velocity loop stability margin with the disturbance observer is
identified. The Nyquist plot of equation (4.13) (see fig. 4.20) indicates the
velocity loop stability and robustness.
Nyquist Diagram (x-axis) Nyquist Diagram (y-axis)
3 3

2 2

1 1
Imaginary Axis

Imaginary Axis

0 0

-1 -1

-2 -2
no observer no observer
with observer with observer
-3 -3
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5
Real Axis Real Axis
Fig. 4.20. Nyquist plots of the velocity loops with and without the disturbance
observer.

This analysis is further extended to include the position loop. The position
open loop and closed loop transfer functions (based on control scheme
shown in fig. 4.16) with the disturbance observer equals:

Z (s) P ⋅ PI ⋅ GGn
L pos = = , (4.16)
Ep (s) Q ( G − Gn ) + Gn + PI ⋅ GGn ⋅ Vest ( s ) ⋅ N ( s )

82
 Friction Force Compensation Design

Z (s) P ⋅ PI ⋅ G ⋅ Gn
T pos = = . (4.17)
Z ref ( s ) Q ( G − Gn ) + Gn + PI ⋅ GGn ⋅ Vest ( s ) ⋅ N ( s ) + P ⋅ PI ⋅ G ⋅ Gn

The compliance transfer function is

Z (s) G (1 − Q )
= . (4.18)
D(s) ⎛ G ⎞
Q⎜ − 1⎟ + P ⋅ PI ⋅ G + GPI ⋅ Vest ( s ) ⋅ N ( s ) + 1
G
⎝ n ⎠

The sensitivity function is

S p _ DOB ( s ) =
(1 − Q ) . (4.19)
⎛ G ⎞
Q⎜ − 1⎟ + P ⋅ PI ⋅ G + G ⋅ PI ⋅ Vest ( s ) ⋅ N ( s ) + 1
G
⎝ n ⎠

The tracking error is

⎛ ⎛ G ⎞ ⎞
− ⎜1 + Q ⎜ − 1⎟ + PI ⋅ G ⋅ Vest ( s ) ⋅ N ( s ) ⎟
⎜ G ⎟
⎝ ⎝ n ⎠ ⎠
E p _ DOB ( s ) = Z ref ( s )
⎛ G ⎞
1+ Q⎜ − 1⎟ + PI ⋅ G ⋅ Vest ( s ) ⋅ N ( s ) + P ⋅ PI ⋅ G
⎝ Gn ⎠ (4.20)
G (1 − Q )
_ D(s).
⎛ G ⎞
1+ Q⎜ − 1⎟ + PI ⋅ G ⋅ Vest ( s ) ⋅ N ( s ) + P ⋅ PI ⋅ G
G
⎝ n ⎠

Fig. 4.21 and 4.22 show the Bode plots of the position open loop and closed
loop transfer functions based on the measured system FRFs (fig. 2.4)
respectively. The plots compare the loop characteristics with and without the
disturbance observer. Table 4.7 summarizes the gain margin and the phase
margin of the position open loop transfer function (4.16). These values
ensure good transient response characteristic and stability margin of the
position loop.

Table 4.7. Gain margin and phase margin of x and y axes position open loop
Gain Margin Phase Margin
no observer with observer no observer with observer
x-axis 15.071 dB 13.621 dB 67.959 deg 69.561 deg
y-axis 16.331 dB 13.773 dB 67.295 deg 70.113 deg
83
Friction Force Compensation Design

The gain margin and the phase margin of both axes were reduced by about
1.5-2 dB and 3 degrees respectively. As with velocity loop, the resulting gain
and phase margins are appropriate to ensure good system transient response
and stability margin.
x-axis

20
Magnitude (dB)

-20

-40

-60

-90
Phase (deg)

-180

-270
no observer
with observer

-360
1 2
10 Frequency (Hz) 10

(a)

20

-20

-40

-60

1
10

84
 Friction Force Compensation Design

Next, the position closed loop transfer functions of both axes are analysed.
Fig. 4.22 shows the Bode diagram of the position closed loop transfer
functions.
x-axis
0
Magnitude (dB)

-20

-40

-60
0
Phase (deg)

-90

-180

-270
no observer
wth observer
-360
1 2
10 Frequency (Hz) 10

(a)
y-axis
0
Magnitude (dB)

-20

-40

-60
0
Phase (deg)

-90

-180

-270
no observer
with observer
-360
1 2
10 Frequency (Hz) 10

(b)
Fig. 4.22. Position closed loop transfer functions for (a) x-axis and (b)
y-axis for system with and without a disturbance observer.

The disturbance observer increases the bandwidth of both axes position

loops by nearly 5Hz. The updated bandwidths of the x-axis and the y-axis
are 40.6Hz and 43.2Hz respectively.
85
Friction Force Compensation Design

A comparison in the sensitivity magnitude Bode plots based on equation

(4.19) and (2.33) for a system with and without the disturbance observer
respectively is shown in fig. 4.23 below.

Bode Magnitude Diagram (x-axis) Bode Magnitude Diagram (y-axis)

10 10

0 0

-10 -10
Magnitude (dB)

Magnitude (dB)
-20 -20

-30 -30

-40 -40

-50 -50

-60 no observer -60 no observer

with observer with observer
-70 1 2
-70 1 2
10 10 10 10
Frequency (Hz) Frequency (Hz)

(a) (b)

Fig. 4.23. Effect of the disturbance observer on the sensitivity function of the
position loop for the (a) x-axis and (b) y-axis

Table 4.8 summarizes the bandwidths of the position loop based on the
sensitivity curve shown in fig. 4.23. Moderate improvements in the
bandwidths of both axes are observed.

Table 4.8. Bandwidth of x and y axes position loops

Bandwidth
no observer with observer
x-axis 26.6 Hz 29.2 Hz
y-axis 27.2 Hz 34 Hz

The disturbance observer improves tracking and disturbance rejection

performances in the lower frequency range (below the system bandwidth), as
seen by shifted sensitivity curves (illustrated by the arrows). This, however,
increases the magnitude of the sensitivity function at higher frequency range.
This phenomenon is generally known as the “waterbed effect” [46].

86
 Friction Force Compensation Design

Finally, the stability of each axis position loop is analysed. Fig. 4.24 shows
the Nyquist plots of each axis position open loop transfer function Lpos(s) that
confirm the loop stability and robustness.
x-axis y-axis
2 2

1.5 1.5

1 1
Imaginary Axis

Imaginary Axis
0.5 0.5

0 0

-0.5 -0.5

-1 -1

-1.5 -1.5
no observer no observer
-2
with observer -2
with observer
-1.5 -1 -0.5 0 0.5 1 -1.5 -1 -0.5 0 0.5 1
Real Axis Real Axis

Fig. 4.24. Nyquist plots of the position loops with and

without the disturbance observer.

Based on the presented Bode and Nyquist plots of the open-loop transfer
functions and on the comparison of the Q filter frequency characteristic and
the inverse of the amplitude of the multiplicative uncertainty (fig. 4.14), it
can be concluded that there is still room to increase the bandwidth of the
disturbance observer and closed loop transfer function. However,
implementing this higher bandwidth inverse model-based disturbance
observer in the experimental test setup resulted in actuator saturation and
thus limits the performance potential of the disturbance observer.

Next, the friction compensation performance of the inverse-model-based

disturbance observer is analysed numerically using MATLAB/Simulink.

4.3.3 Numerical Validations

The friction compensation performance of the inverse-model-based
disturbance observer is validated numerically with MATLAB/Simulink,
using the simulation model presented in section 4.2.1. The trajectory is a
sinusoidal signal with an amplitude of 30 mm, a frequency of 0.53Hz, and a
maximum tracking velocity of 100 mm/s. The simulated control scheme is
shown in fig. 4.25.

87
Friction Force Compensation Design

Fig. 4.25. Simulated control scheme for friction compensation using

friction model feedforward and a disturbance observer.

The tracking and friction compensation performances are analysed using 3

different compensation configurations:

a) An inverse-model-based disturbance observer only

b) An inverse-model-based disturbance observer with a static friction
model feedforward
c) An inverse-model-based disturbance observer with a GMS model
feedforward

Fig. 4.26 shows the simulated results for the y-axis for these different
compensation configurations. Similar simulations are performed for the x-
axis. The results are tabulated in Table 4.9.

88
 Friction Force Compensation Design

4 4 Observer + 4 Observer +
x 10 Observer Only x 10 x 10
Static Model GMS Model
3 3 3
position [µm]

position [µm]

position [µm]
2 2 2

0 0 0

-2 -2 -2
-3 -3 -3

0 0.5 1 1.5 0 0.5 1 1.5 0 0.5 1 1.5

time [s] time [s] time [s]
tracking error [µm]

tracking error [µm]

tracking error [µm]

10 10 10

0 0 0

-10 -10 -10

0 0.5 1 1.5 0 0.5 1 1.5 0 0.5 1 1.5

time [s] time [s] time [s]

(a) (b) (c)

Fig. 4.26. Y-axis: Simulated position and tracking error
for configuration (a), (b) and (c).

Table 4.9. Simulated friction compensation performance for reference

sinusoidal signal with an amplitude of 30mm and a maximum tracking
velocity of 100mm/s
Friction Compensation Quadrant Glitch [µm]
Methods x-axis y-axis
Disturbance observer only 9 8.5
Disturbance observer and 5.5 5
static friction feedforward
Disturbance observer and 1 1
GMS friction feedforward

A combination of GMS friction model feedforward and the inverse model-

based disturbance observer yields a near zero quadrant glitch magnitude. As
discussed previously in section 4.2.2, the system delay has prevented a
complete removal of the quadrant glitch, as expected in an ideal case. Next,
experimental validations are performed on the test setup. The results are
analysed and discussed in the next section.

89
Friction Force Compensation Design

4.3.4 Experimental Validations

Fig. 4.27 shows the measurement results on the real test setup (x-axis)
obtained by applying the inverse model-based disturbance observer with
static and GMS friction model feedforward for a similar sinusoidal reference
trajectory as applied in the previous section.

4 Observer 4
Observer + 4
Observer +
x 10 Only x 10 Static Model FF. x 10 GMS Model FF.

3 3 3
position,[µm]

position,[µm]

position,[µm]
0 0 0

-3 -3 -3

0 0.5 1 1.5 0 0.5 1 1.5 0 0.5 1 1.5

time [s] time [s] time [s]

20 20 20
15 15 15
tracking error,[µm]

tracking error,[µm]

tracking error,[µm]

7
3 3
0 0 0
-3 -3
-7

-15 -15 -15

-20 -20 -20
0 0.5 1 1.5 0 0.5 1 1.5 0 0.5 1 1.5
time [s] time [s] time [s]

(a) (b) (c)

Fig. 4.27. X-axis: Measured position and tracking errors for
configuration (a), (b) and (c).

The combination of the inverse model-based disturbance observer and the

GMS friction model feedforward yield the most significant reduction of the
magnitude of the quadrant glitches (as concluded previously using the
simulated results). The results however do not clearly indicate the difference
in friction compensation performance between conditions (b) and (c). The
advantage of the GMS friction model is not clearly demonstrated. Similar
experiments for configuration (b) and (c) are repeated with a reduced
maximum tracking velocity of 10mm/s instead of 100mm/s. Fig. 4.28 shows
the result for this slower reference trajectory.

90
 Friction Force Compensation Design

Observer + Observer +
4 4
x 10 Static Friction FF. x 10 GMS Friction FF.
4 4

2 2

position,[µm]
position,[µm]

0 0

-2 -2

-4 -4
0 5 10 15 0 5 10 15
time [s] time [s]

20 20

tracking error,[µm]
tracking error,[µm]

10 10
5 5
0 0
-5 -5
-10 -10

-20 -20
0 5 10 15 0 5 10 15
time [s] time [s]

(a) (b)
Fig. 4.28. X-axis: Measured position and tracking errors for slower
reference trajectory of 10mm/s for configuration (a) and (b).

This figure clearly demonstrates the advantage of using a GMS friction

model at lower tracking velocity. At a lower tracking velocity, the pre-
sliding regime is more dominant and the benefit of using the GMS friction
model that includes hysteresis with non-local memory is more pronounced.

A similar experimental validation was repeated for the y-axis. The results are
shown in fig. 4.29. Similar to the x-axis, the combination of the disturbance
observer and the GMS friction model feedforward yields the smallest
quadrant glitch magnitude. A quadrant glitch magnitude of about 3µm was
recorded. The tracking error that remains besides the quadrant glitches is due
to the imperfection of the inverse model feedforward which result from
modelling errors. The measurement results for both the axes are summarized
in Table 4.10.

91
Friction Force Compensation Design

4 4
Observer + 4
Observer +
x 10 Observer Only x 10 Static Friction FF. x 10 GMS Friction FF.

3 3 3
position [µm]

position [µm]

position [µm]
0 0 0

-3 -3 -3

0 0.5 1 1.5 0 0.5 1 1.5 0 0.5 1 1.5

time [s] time [s] time [s]

20 20 20
tracking error [µm]

tracking error [µm]

tracking error [µm]

10 10 10
5
3
0 0 0
-3
-5
-10 -10 -10

-20 -20 -20

0 0.5 1 1.5 0 0.5 1 1.5 0 0.5 1 1.5
time [s] time [s] time [s]

(a) (b) (c)

Fig. 4.29. Y-axis: Measured position and tracking errors
for configuration (a), (b) and (c).

Table 4.10. Measured friction compensation performance for a reference

sinusoidal signal with an amplitude of 30mm and a maximum tracking
velocity of 100mm/s
Axial Quadrant Glitch [µm]
Friction Compensation Methods
x-axis y-axis
Disturbance observer only 7 11
Disturbance observer and static friction 3 5
feedforward
Disturbance observer and GMS friction 2.5 3
feedforward

Next, circular tests using the three different compensation configurations are
performed on the xy table when both axes are controlled simultaneously.
Fig. 4.30 shows the contour measurements and the radial tracking errors.

92
 Friction Force Compensation Design

observer only observer + static model observer + GMS model

(error x400) (error x400) (error x400)

30 30 30
position Y [mm]

position Y [mm]

position Y [mm]
0 0 0

-30 -30 -30

-30 0 30 -30 0 30 -30 0 30
position X [mm] position X [mm] position X [mm]

10 10 10

5 5 5
radial error [um]

radial error [um]

radial error [um]

0 0 0

-5 -5 -5
rmsX=1.13um rmsX=0.666um rmsX=0.84um
rmsY=1.02um rmsY=0.675um rmsY=0.62um
-10 -10 -10
0 90 180 270 360 0 90 180 270 360 0 90 180 270 360
angle [degree] angle [degree] angle [degree]

(a) (b) (c)

Fig. 4.30. Measured contours and radial errors for configuration (a), (b) and (c).

Circular tests at tangential tracking velocity of 10mm/s are performed to

illustrate more clearly the advantage of the GMS model over the static
friction model. The results are shown in fig. 4.31.

Observer + Static Friction Model Observer + GMS Friction Model

30 30
position, y-axis [mm]

position, y-axis [mm]

20 20
10 10
0 0
-10 -10
-20 -20
-30 error: x400 -30 error: x400
-20 0 20 -20 0 20
position, x-axis [mm] position, x-axis [mm]

20 20
radial tracking error [µm]

radial tracking error [µm]

8 8

0 0

-8 -8

-20 -20
0 90 180 270 360 0 90 180 270 360
angle [ degree] angle [degree ]

Fig. 4.31. Circular tracking tests at tangential tracking velocity of 10mm/s.

93
Friction Force Compensation Design

4.4 Summary
This chapter discusses friction compensation on a direct driven linear motors
XY feed table using friction model based and friction model-free
approaches. Static and GMS friction model feedforward, and an inverse-
model-based disturbance observer are compared and combined. The friction
compensation performances are compared based on the magnitude of the
quadrant glitches. Quadrant glitches, which are caused by the complex non-
linear behaviour of friction at velocity reversal, can be compensated
effectively using a combination of the GMS friction model feedforward and
an inverse-model based disturbance observer. The benefits of using an
advanced friction model like the Generalized Maxwell-slip (GMS) friction
model are especially clear at slow motions where the pre-sliding friction is
dominant. Fig. 4.32 and Table 4.11 summarize the friction compensation
performances for all six different compensation configurations tested on the
experimental test setup using axial and circular trajectory with an amplitude
of 30mm, a frequency of 0.53Hz, and a maximum tracking velocity of
100mm/s.

no static model GMS model disturbance static model + GMS model +

compensation feedforward feedforward observer only observer observer
position Y [mm]

30 30 30 30 30 30

0 0 0 0 0 0

-30 -30 -30 -30 -30 -30

error x400 error x400 error x400 error x400 error x400 error x400
-30 0 30 -30 0 30 -30 0 30 -30 0 30 -30 0 30 -30 0 30
position X [mm] position X [mm] position X [mm] position X [mm] position X [mm] position X [mm]

30 15 15 15 15 15
25
10 10 10 10 10
radial error [µm]

20
5 5 5 5 5
15
10 0 0 0 0 0
5
-5 -5 -5 -5 -5
0
-10 -10 -10 -10 -10
-5
-15 -15 -15 -15 -15
0 90 180 270 360 0 90 180 270 360 0 90 180 270 360 0 90 180 270 360 0 90 180 270 360 0 90 180 270 360
angle [degree] angle [degree] angle [degree] angle [degree] angle [degree] angle [degree]

Fig. 4.32. Position and radial tracking error for different friction compensation
approaches.

94
 Friction Force Compensation Design

Table 4.11. Measured magnitude of the quadrant glitches for reference

sinusoidal signal with an amplitude of 30mm and a maximum tracking
velocity of 100mm/s.

x-axis y-axis
Friction Compensation Radial
Methods Quadrant Quadrant Error (µm)
Glitch (µm) Glitch (µm)

No friction feedforward 15 35 25

Static friction feedforward 5 13 7

GMS friction feedforward 4 10 6

Disturbance observer only 7 11 7

Disturbance observer + static

3 5 5
friction model feedforward
Disturbance observer + GMS
2.5 3 3
friction model feedforward

95
96
 Chapter 5

Cutting Force Compensation

Design

5.1 Introduction
Friction force compensation alone is insufficient to obtain high tracking
accuracy in high-speed linear drive based milling processes. Cutting forces
generated by the cutting tool and work-piece interaction act directly on the
linear motors and influence positioning and tracking accuracy. Cutting force
compensation is essential and is traditionally accomplished with high gain
feedback control yielding high dynamic stiffness. High dynamic stiffness
means high disturbance rejection capacity. High controller gain however
reduces the stability margins. This chapter analyses several disturbance
compensation methods that have been proposed in literature for various
applications but limited knowledge is available with regards to their
application and performance for compensation of cutting forces. The
different compensation methods will supplement the existing cascade P/PI
position controller in a modular based approach. Three compensation
methods are considered, namely; (i) an inverse-model-based disturbance
observer (section 5.2), (ii) an explicit estimation of cutting force using
relative acceleration measurement (section 5.3), and (iii) a repetitive
controller (section 5.4).

First, the performance of the inverse-model-based disturbance observer is

analysed for cutting force compensation, an extension from its previous
friction force compensation.

97
Cutting Force Compensation Design

5.2 Inverse-Model-Based Disturbance

Observer
The structure, design and analysis of the inverse-model-based disturbance
observer [26] for friction compensation have been discussed extensively in
the previous chapter (section 4.3). Here, the application of this disturbance
observer to compensate cutting forces is analysed.

Cutting forces are composed of several harmonic components as shown

previously in fig. 3.13 and 3.14. In section 4.3 it is shown that the
performance of the inverse- model-based disturbance observer depends on
the bandwidth of the low pass filter Q. The number of cutting force harmonic
components that are compensated is therefore limited by the bandwidth of
the filter Q. The disturbance observer cutting force compensation
performance is first validated numerically using MATLAB/Simulink.

5.2.1 Numerical Validations

Fig. 5.1 shows a MATLAB/Simulink diagram of the control structure,
including the cascade P/PI position controller and the inverse-model-based
disturbance observer. In this analysis, only the cutting force compensation is
considered (no friction force compensation) and no reference tracking is
applied. First, a simulated disturbance force d(t) is applied. Its amplitude is 1
volt which corresponds to a force of 1450 N. The performance of the
disturbance observer is compared based on the position error ep(t), for three
different frequencies: (i) 1 Hz, (ii) 58 Hz, and (iii) 100 Hz.

Fig. 5.1. MATLAB/Simulink diagram of a cascade P/PI controller and a

disturbance observer with a sinusoidal based disturbance input signal

98
 Cutting Force Compensation Design

Fig. 5.2 shows the position tracking error of the system, before and after the
disturbance observer is activated.
1 Hz 58Hz 100Hz
150 150 150
133.9µm 109µm
100 100 100
position error [µm]

position error [µm]

position error [µm]

56.1 µm
34.91µm
50 50 50

0 0 0
2.97 µm

-50 -50 -50

-100 -100 -100

observer on
observer on
observer on
-150 -150 -150
0 2 4 6 0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
time [s] time [s] time [s]

(i) (ii) (iii)

Fig. 5.2. Y-axis: The effect of a disturbance observer on tracking errors for
disturbance input frequencies of (i) 1 Hz, (ii) 58 Hz, and (iii) 100 Hz

The position tracking error of the system without a disturbance observer is

described in equation (2.34). In the absence of a reference input signal, the
tracking error from the input disturbance signal is simply,

Gm
E p ( s )noDOB = − D(s)
1 + Gm ⋅ P ⋅ PI + Gm ⋅ PI ⋅ Vest ⋅ N
. (5.1)
E p ( s )noDOB = −Gm ( s ) ⋅ S ( s ) ⋅ D ( s ) .

S(s) is the sensitivity function. Equation (4.20) describes the position

tracking error of the system with a disturbance observer. In the absence of a
reference input signal, the tracking error equals:

Gm (1 − Q )
E p ( s ) DOB = − D(s)
⎛G ⎞
1 + Q ⎜ m − 1⎟ + PI ⋅ Gm ⋅ Vest ⋅ N + P ⋅ PI ⋅ Gm (5.2)
⎝ nG ⎠
E p ( s ) DOB = − Gm ⋅ S ( s ) ⋅ D ( s ) .

Fig. 5.3 shows the Bode magnitude plot of the position error transfer
functions (5.1) and (5.2). Indicated on the two curves are the theoretical
positions tracking errors for frequencies of (i) 1Hz, (ii) 58Hz, and (iii)
100Hz.

99
Cutting Force Compensation Design

3 Bode Magnitude Diagram

10
Freq (Hz): 58 Freq (Hz): 58.2
Freq (Hz): 1 Mag (abs): 109 Mag (abs): 112
Mag(abs): 133
2
10 Freq (Hz): 100
Mag (abs): 57.8
Freq (Hz): 100
Magnitude (abs)

1 Mag (abs): 36.4

10
Freq (Hz): 1
Mag(abs): 2.96

0
10

-1
10
no observer
w ith observer
-2
10
-1 0 1 2 3
10 10 10 10 10
Frequency (Hz)

Fig 5.3. Y-axis: Position errors for system with and

without a disturbance observer

Fig. 5.3 shows an almost exact match of position errors for each of the three
different frequencies, for cases with and without the disturbance observer,
when compared to the simulated position errors shown in fig. 5.2. The
system is able to partly compensate the purely sinusoidal signal until a
frequency of nearly 60Hz. The position error is amplified beyond this
frequency, as shown in fig. 5.2 (iii) and fig. 5.3.

5.2.2 Experimental Validations

Experimental validation is performed on the test setup using a synthesized
cutting force with different harmonic contents. Fig. 5.4 shows the control
scheme that includes the cascade P/PI controller, GMS friction model
feedforward, the inverse-model reference feedforward, and the inverse-
model-based disturbance observer.

First, a synthesized cutting force described in section 3.4.2 is applied at the

input of the system with no reference trajectory motion. The position errors
for a synthesized cutting force with 1, 2, 3, and 15 harmonics content is
shown in fig. 5.5 (the fundamental frequency is 23.3Hz, see Table 3.5).

100
 Cutting Force Compensation Design

Fig. 5.4. Cutting force disturbance compensation using inverse-

model-based disturbance observer

1 harmonic 2 harmonics 3 harmonics 15 harmonics

30 30 30 30

20 20 20 20
position error [µm]

position error [µm]

position error [µm]

10 10 10 position error [µm] 10

0 0 0 0

-10 -10 -10 -10

-20 -20 -20 -20

observer on observer on observer on observer on

-30 -30 -30 -30

0 0.5 1 1.5 0 0.5 time1[s] 1.5 0 0.5 1 1.5 0 0.5 1 1.5
time [s] time [s] time [s]

Fig. 5.5. Y-axis: Measured position errors with inverse model-based

disturbance observer for synthesized cutting forces disturbance with
different harmonics component

The mean value of the position errors for input disturbance signals with two
or more harmonic components are not zero due to the differences in the
amplitudes and phases of the various harmonics. The position errors are
reduced by applying the inverse model-based disturbance observer even if
the disturbance input signal contains the full fifteen harmonics (the third
harmonic is already beyond the Q-filter bandwidth). This is possible because
of the much smaller amplitudes of the high frequencies harmonic
components.

A detailed analysis of the frequency content and the root mean square (rms)
values of the position errors are shown in fig. 5.6 and Table 5.1 respectively.

101
Cutting Force Compensation Design

1 harmonic 2 harmonics 3 harmonics 15 harmonics

20 20 20 20
(no observer)

15 15 15 15
spectrum of
pos. error

10 10 10 10

5 5 5 5

0 0 0 0
0 50 100 150 200 0 50 100 150 200 0 50 100 150 200 0 50 100 150 200

20 20 20 20
(with obse)rver

15 15 15 15
spectrum of
pos. error

10 10 10 10

5 5 5 5

0 0 0 0
0 50 100 150 200 0 50 100 150 200 0 50 100 150 200 0 50 100 150 200
freq. [Hz] freq. [Hz] freq. [Hz] freq. [Hz]

Fig. 5.6. Y-axis: Spectral analysis of the position errors without the disturbance
observer (top) and with the disturbance observer (bottom) for disturbance force
signals with different harmonic content.

The spectral analyses show a significant amplitude reduction (nearly 60%)

only for the first harmonic (23.3Hz). This is consistent with the fact that the
performance of the disturbance observer is limited by the bandwidth of the
filter Q.

Table 5.1. rms of position errors for different cutting force harmonics content
using the explicit estimation of cutting force technique
1 harmonic 2 harmonics 3 harmonics 15 harmonics
no feedback 9.62µm 10.83 µm 13.99 µm 14.67 µm
with
5.63 um 6.34 µm 6.98 µm 11.11 µm
compensation
% reduction 41.47% 41.45% 50.00% 24.26%

Next, circular tests are performed to measure both the x-axis and the y-axis
cutting force compensation performance using the inverse-model-based
disturbance observer. A synthesized cutting force with 15 harmonics is
applied at both inputs of the system. The control scheme includes also the
GMS friction model feedforward and the inverse model reference
feedforward. Fig. 5.7 compares the circular test results with and without the
disturbance observer.

102
 Cutting Force Compensation Design

No Cutting Force Compensation Inverse-Model-Based Disturbance Observer

30 30
position, y-axis [mm]

position, y-axis [mm]

20 20
10 10
0 0
-10 -10
-20 -20
error: x200 -30 error: x200
-30
-40 -20 0 20 40 -40 -20 0 20 40
position, x-axis [mm] position, x-axis [mm]

30 30

20 20
radial error [µm]

10 radial error [µm] 10

0 0

-10 -10

-20 -20

-30 -30
0 90 180 270 360 0 90 180 270 360
angle [degree] angle [degree]

Fig. 5.7. Measured contours and radial errors of circular tests with
(right) and without (left) the inverse model based disturbance observer

The results show a reduction in the magnitude of the radial tracking errors
for the system with disturbance observer. The radial tracking error is reduced
from a maximum of 20µm to about 10 µm. A complete compensation of the
cutting force is not possible due to the limited bandwidth of the filter Q.

In the next section, an approach to explicitly estimate the cutting forces from
relative acceleration measurements is described.

5.3 Explicit Estimation of Cutting Force via

Ferraris Sensor Measurements
Pritschow [29] has presented an approach to reconstruct process forces in
direct drives using relative acceleration measurements. However, there is
limited information on the application of this estimator to effectively
compensate the influence of cutting forces. This section analyses the
proposed estimation method on the experimental test setup.

103
Cutting Force Compensation Design

5.3.1 The Ferraris Principle

The relative acceleration signal, which is the main element of this approach,
is captured by a sensor that is known as a Ferraris sensor, named after the
Ferraris principle [47]. Fig. 5.8 shows a schematic diagram of the Ferraris
sensor.

copper strip

Ferraris
sensor

Fig. 5.8. Schematic diagram of the Ferraris principle (* republished from

[36] ) and the actual sensor used in measurement

The sensor has three main elements, namely:

• A block consisting of the lamination, permanent magnet, and the
windings.
• A non-magnetic eddy current strip that is either Aluminium or
copper (in this application a copper strip is used)
• An amplifier for the sensor signal

104
 Cutting Force Compensation Design

The eddy current strip that lies within the slot of the block is attached to the
moving object while the other parts are fixed to the frame. A relative motion
between the eddy current strip and the block generates eddy current in the
eddy current strip due to the permanent magnet that induces a voltage. The
resulting eddy current generates magnetic fields and the strength of this field
changes as the velocity of motion changes. This results in a voltage being
induced in the winding. This voltage is proportional to the change of
velocity, that is, the acceleration of the motion. The amplifier then amplifies
the signal.

A Ferraris sensor is attractive in mechatronic application because it provides

an alternative method for generating a velocity signal from the acceleration
signal [48]. Noise amplification and quantization of velocity signals that are
generated by numerically differentiating position encoder signals can be
avoided.

J. Wang [36] has described the effect of eddy current strip characteristics,
placement, and structure on the quality of the sensor signals. The
disadvantage of this sensor is that, it has to be calibrated on-site. This is
because the installation of the sensor part can affect the sensor sensitivity.
The calibration of the sensor is based on the comparison between the
position encoder measurement and the position signal that is generated from
the Ferraris sensor measurements. The calibration of the Ferraris sensor
sensitivity and offset value are described in Appendix B.

5.3.2 Cutting Force Estimator

The Pritschow’s cutting force estimator [29] is designed based on a reduced
order observer [35] that estimates state variables that cannot be measured
directly. The observer is designed by considering the following model
equation that describes the balance between forces that acts on a direct drive
system (here, the &x& term refers to absolute acceleration).

kf 1 1
x (t ) =
&& u (t ) + d (t ) − Ff (v) (5.3)
M M M

x is the position, d is the disturbance force, kf is the motor force constant, M

is the mass, Ff is the friction force, and u is the voltage to the drive
amplifier. The state space representations of the system and the measurement
equation are:

105
Cutting Force Compensation Design

⎡ 0 0 ⎤
⎡ x& ⎤ ⎡0 1 0 ⎤ ⎡x⎤ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
1 ⎥ ⎢v ⎥ + ⎢kf −
1 ⎥ ⎡ u ⎤
⎢ v& ⎥ = ⎢0 0 M ⎢ ⎥ ⎢M M ⎥ ⎢⎢ F f ( v ) ⎥⎥
(5.4)
⎢ d& ⎥ ⎢ ⎥ ⎥ ⎣ ⎦
⎣ ⎦ ⎣0 0 0 ⎦ ⎢⎣ d ⎥⎦ ⎢
⎢⎣ 0 0
⎦⎥

0 ⎤ ⎢⎡ ⎥⎤ ⎡ 0
x 0 ⎤
⎡ x ⎤ ⎡1 0 ⎥ ⎡ u ⎤
y=⎢ ⎥=⎢ ⎥ ⎢v⎥ + ⎢ k f 1 ⎢ ⎥ (5.5)
⎣ a ⎦ ⎢⎣
0 0 1 ⎥ ⎢
M ⎦ ⎢ d ⎥ ⎣⎢ M − ⎥ ⎣⎢ F f ( v ) ⎦⎥
⎣ ⎦ M ⎦⎥

The detailed derivation of the reduced order observer for the estimation of
position x, velocity v, and the disturbance forces d is presented in Appendix
C.

In fact, the disturbance force d can be estimated directly from (5.3) and is
described as:

⎛ kf 1 ⎞
d̂ = M ⎜⎜ a − u+ F f ( v ) ⎟⎟ (5.6)
⎝ M M ⎠

a is the Ferraris relative acceleration signal measurement.

Next, the estimator is designed based on (5.6) to compensate cutting forces

and its experimental validation is performed on the considered test setup.

5.3.3 Experimental Validations

The first part of the experimental validation analysed the quality of the
estimated disturbance forces only, without feedback. Fig. 5.9 shows the
schematic diagram of the control structure that includes the cascade P/PI
position controller and the estimator. The estimator is similar to equation
(5.6) without the friction term. The effect of friction is not considered at this
moment. The disturbance force is the only input to the system. The estimator
is experimentally validated on the y-axis where the Ferraris sensor is
attached.

106
 Cutting Force Compensation Design

Fig. 5.9. Schematic diagram of a cascade P/PI position control

with cutting force estimator (without feedback)

The inputs to the estimator are the control command signal u and the
calibrated relative acceleration signal a. A synthesized cutting force signal
with 1, 3, and 15 harmonic contents are introduced at the input of the system
G(s). The axial position errors ep(t) are recorded. Fig. 5.10 compares the
reference cutting force d and the estimated cutting force d̂ . The estimated
cutting force is delayed due to the delay in the drive.
1 harmonic 3 harmonics 15 harmonics
600 600 600

400 400 400

cutting force [N]

200 200 200

0 0 0

-200 -200 -200

-400 -400 -400

reference
-600 estimated -600 -600

0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2
time [s] time [s] time [s]

Fig 5.10. Reference and estimated cutting forces

In the second part of the experimental validation, a feedback of the estimated

cutting force is applied. Fig. 5.11 shows the applied control scheme. A low
pass filter Q similar to equation (4.7) and desc

107
Cutting Force Compensation Design

Fig. 5.11. Schematic diagram of a cascade P/PI position control

with cutting force estimator (with feedback)

Synthesized cutting force signals with 1, 2, 3, and 15 harmonic contents are

applied at the input of the system G(s). The axial position errors ep(t) are
recorded. Fig. 5.12 illustrates the synthesized cutting force and compares the
position tracking errors with and without the estimator feedback. The
spectral analysis of the position errors is shown in fig. 5.13.

1 harmonic 2 harmonics 3 harmonics 15 harmonics

600 600 600 600
cutting force [N]

400 400 400 400

synthesized

200 200 200 200

0 0 0 0
-200 -200 -200 -200
-400 -400 -400 -400
-600 -600 -600 -600
0 0.2 0.4 0 0.2 0.4 0 0.2 0.4 0 0.2 0.4
time [s] time [s] time [s] time [s]
40 40 observer on 40 40 observer on
observer on observer on
position error [um]

position error [um]

position error [um]

20 20 20 20
error [um]
position

0 0 0 0

-20 -20 -20 -20

-40 -40 -40 -40

0 0.2 0.4 0 0.2 0.4 0 0.2 0.4 0 0.2 0.4
time [s] time [s] time [s] time [s]

Fig. 5.12. Position errors for different cutting force harmonic

components using estimator

108
 Cutting Force Compensation Design

without compensation 1 harmonic 2 harmonics 3 harmonics 15 harmonics

20 20 20 20
position error [um]

position error [um]

position error [um]

position error [um]
10 10 10 10

0 0 0 0
0 100 200 0 100 200 0 100 200 0 100 200
freq. [Hz] freq. [Hz] freq. [Hz] freq. [Hz]
15 15 15 15
position error [um]

position error [um]

with compensation

position error [um]

position error [um]

10 10 10 10

5 5 5 5

0 0 0 0
0 100 200 0 100 200 0 100 200 0 100 200
freq. [Hz] freq. [Hz] freq. [Hz] freq. [Hz]

Fig. 5.13. Spectral analysis of the position errors for different cutting
force harmonic components using an estimator

Table 5.1 summarizes the rms of the position errors for cases with and
without the estimator compensation.

The results show limited performance of the estimator. An average of 33%

reduction in the magnitude of the position errors is achieved. The limited
bandwidth of the filter Q reduces the overall compensation performance
(similar argument as previously discussed in section 4.3). The estimator has
an almost equal percentage reduction in the position errors for all the
different cutting force harmonic contents (see Table 5.2). However, the
magnitude of the rms values are lower compared to the results obtained
using the inverse-model based disturbance observer shown in Table 5.1.

Table 5.2. rms of position errors for different cutting force harmonics content
using the explicit estimation of cutting force technique
1 harmonic 2 harmonics 3 harmonics 15 harmonics
no feedback 10.76µm 12.04 µm 18.27 µm 18.53 µm
with
7.16 um 8.15 µm 11.57 µm 11.89 µm
compensation
% reduction 33.46% 32.37% 36.67% 29.77%

109
Cutting Force Compensation Design

Finally, the cutting force is compensated during a reference tracking of a

sinusoidal signal with an amplitude of 30mm and a maximum tracking
velocity of 100mm/s. Fig. 5.14 shows the tracking errors with and without
the estimator feedback. The spectral analysis of these tracking errors is
shown in fig. 5.15.
1 harmonic 3 harmonics 15 harmonics
50 50 50
position [mm]

position [mm]

position [mm]
0 0 0

-50 -50 -50

0 0.5 1 1.5 0 0.5 1 1.5 0 0.5 1 1.5
position error [um] position error [um]

position error [um] position error [um]

40 40 40
(no estimator)
error [um]

20 20 20
position

0 0 0
-20 -20 -20
-40 -40 -40
0 0.5 1 1.5 0 0.5 1 1.5 0 0.5 1 1.5
40 40 40
(with estimator)
error [um]

20 20 20
position

0 0 0
-20 -20 -20
-40 -40 -40
0 0.5 1 1.5 0 0.5 1 1.5 0 0.5 1 1.5
time [s] time [s] time [s]

Fig. 5.14. Tracking errors for different cutting force harmonic

contents for cases with and without the estimator

1 harmonic 3 harmonics 15 harmonics

20 20 20
position error [um]

position error [um]

position error [um]

(no estimator)

15 15 15

10 10 10

5 5 5

0 0 0
0 50 100 150 200 0 50 100 150 200 0 50 100 150 200
freq. [Hz] freq. [Hz] freq. [Hz]
20 20 20
position error [um]

position error [um]

position error [um]

(with estimator)

15 15 15

10 10 10

5 5 5

0 0 0
0 50 100 150 200 0 50 100 150 200 0 50 100 150 200
freq. [Hz] freq. [Hz] freq. [Hz]

Fig. 5.15. Spectral analysis of the tracking errors for different cutting
force harmonic components for cases with and without the estimator

110
 Cutting Force Compensation Design

Table 5.3 summarizes the rms of the tracking errors.

Table 5.3. rms of tracking errors for different cutting force harmonics content
using the explicit estimation of cutting force technique
1 harmonic 3 harmonics 15 harmonics

no compensation 8.96µm 10.06 µm 10.11 µm

compensation 5.67 um µm 6.60 µm 6.73 µm

% reduction 36.72% 34.39% 33.43%

The estimator is able to partly compensate the cutting forces that act on the
system during tracking. An average of 33% reduction in the magnitude of
the position errors is observed.

Fig. 5.16 shows a comparison between the reference cutting forces and the
estimated feedback cutting force. The results illustrate the poor performance
of the estimator. This is a stark contrast to the estimator performance when
the estimated cutting force is not fed back into the system as previously
shown in fig. 5.11. The performance is degraded during feedback due to the
limited bandwidth of the filter Q. The delay that has already influenced the
estimated cutting force further reduces the estimator performance.
1 harmonic 3 harmonics 15 harmonics
600 600 600

400 400 400

cutting force [N]

200
200 200
0
0 0
-200
-200 -200
-400
-400 reference -400 -600
estimated
-600 -600 -800
0.2 0.25 0.3 0.35 0.4 0.2 0.25 0.3 0.35 0.4 0.2 0.25 0.3 0.35 0.4
time [s] time [s] time [s]

Fig 5.16. Cutting Forces Estimation Analysis (without

reference trajectory)

A more efficient and complete compensation of the cutting force is desirable.

An approach that incorporates the characteristic of the cutting force signals
in its design structure is presented next. This approach is based on the
design of a general order state observer.

111
Cutting Force Compensation Design

Cutting forces consist mainly of harmonic components with a fundamental

frequency equal to the frequency of rotation of the tool (as indicated in Table
3.5 and 3.6). It is expected that by taking this property into account,
improved cutting force rejection can be obtained. Two approaches are
considered: a disturbance force observer based on a state estimator that
includes the knowledge of the harmonic contents of the cutting force and a
repetitive controller. The first approach has been designed and validated
numerically and experimentally on the considered test setup. Preliminary
results indicate that the method has good potential in compensating cutting
forces. However, due to practical limitations (drive saturation etc.) of the test
setup, compensation of cutting forces beyond the first three harmonic
contents was not possible. The design, analysis, and preliminary results of
this technique are presented in Appendix D.

The second approach that is using the repetitive controller is discussed in the
following section.

5.4 Repetitive Controller

Repetitive control (RC) is an appropriate control strategy if the dominant
disturbances or the tasks to be performed are periodic. For example, in the
track-following servo system of optical disk drives, peristaltic pumps used in
medical devices, active noise control, robotized laparoscopic surgery, or
tracking of periodic trajectories as, for instance, in robots performing
repetitive tasks and during non-circular machining. RC is also appropriate to
compensate the effect of the cutting forces since these forces are mainly
periodic.

5.4.1 Design Structure of a Repetitive Controller

Including a repetitive controller (RC) in the feedback loop improves the
attenuation of periodic disturbance inputs. RC is based on the internal model
principle [49], which states that, if a disturbance signal can be regarded as
the output of an autonomous system, including this system in a stable
feedback loop guarantees asymptotically perfect rejection. A memory loop is
the most frequently used generator of periodic signals. Fig. 5.17 shows a RC
(indicated in blue) as an add-on module in a closed loop of a general control
scheme for a plant with a periodic disturbance input d of known period time
T0[s].
112
 Cutting Force Compensation Design

Fig. 5.17. Standard RC as an add-on module to a closed

loop control scheme

The memory loop, a basic element of a RC, is a delay line of T0 [s], in a

positive feedback loop. T0 equals the period time of the disturbance signal d
(T0 = 2π/ωo, where ωo is the period frequency). Figure 5.18 shows the
memory loop diagram. ep(t) is the input to the closed loop and can include
both periodic and non-periodic components.

Fig. 5.18. Standard memory loop with periodic signal

generator with period T0[s].

In most practical implementations of a RC, this delay is implemented in

discrete-time for a selected sampling period Ts. The sampling period is
chosen such that T0 = N·Ts, where N ∈ . If the ratio of T0 /Ts is not exactly
an integer number, it is rounded to the lower integer value.

The transfer function from ep to ep’ equals:

ep ' ( s) e −T0 s
= (5.7)
ep ( s) 1 − e −T0 s

For frequency ω = kωo (k an integer and ωo=2π/T0), the magnitude of the

denominator of equation (5.7) becomes zero and the transfer function is
infinite. This means that an infinite loop gain at all the harmonics of the
disturbance signal is realised using this memory loop. High loop gains,
however, can cause instability. Stability analysis in RC design is thoroughly
113
Cutting Force Compensation Design

discussed in [50, 51]. Generally, two discrete time filters are added: L(z) and
Q(z). The filters L(z) and Q(z) are necessary to preserve stability, modulus
margin and high-frequency rolls-off of the original feedback controller [52].
L(z) and Q(z) are usually designed as follows [50, 51]:

• L(z) is set equal to the inverse of the complementary sensitivity

function of the system, augmented with a low pass filter with a cut-
off frequency above or equal to that of the Q filter. L(z) is
implemented as a FIR filter.
• Q(z) is a low-pass linear-phase FIR filter with unity dc-gain and cut-
off frequency ωQ.

The RC performance is limited by the cut-off frequency ωQ. The RC is not

active beyond the pass band of the filter Q(z) (that is when Q(ω) ≈ 0). It is
important to note that due to the presence of the delay z-N, L(z) and Q(z)
filters can be designed as non-causal filters. For the implementation of these
filters, the non-causal part is absorbed into the delay z-N. Fig. 5.19 shows the
discrete-time implementation of a typical RC with the delay line z-N and the
filters L(z) and Q(z).

Fig. 5.19. Discrete time implementation of a typical RC.

This typical RC is often referred to as a first order RC, a special form of a so

called high order RC where the delay z-N is replaced by a polynomial of the
delay:

n
∑ wi z −iN (5.8)
i =1

n equals the order of the RC. High order RC gives the designer more degrees
of freedom (wi parameter) to achieve a better robustness against uncertainties
of the period T0 of the disturbance signal. For the considered application, i.e.
the cutting force compensation, a second order RC (n = 2) is designed. Fig.
5.20 shows the schematic diagrams of an nth order RC and a second order
RC. RC‘s with an order higher than two were not considered because of
processor hardware limitations.
114
 Cutting Force Compensation Design

Fig. 5.20. Schematic diagram of (a) nth order RC, (b) 2nd order RC

The two parameters w1 and w2 of the second order RC were designed using
the method developed by G.Pipeleers et. al. [52, 53] as a trade-off between
periodic performance considering a relative uncertainty of 1% on the period
of the disturbance and non-periodic performance degradation, that is the
amplification of non-periodic disturbances such as measurement noise that
are present between the harmonics of the cutting force. The periodic and
non-periodic performances of a RC are quantified by the robust periodic and
non-periodic performance indices respectively [51]. The robust periodic
performance index γp,∆ [51] is defined as the smallest reduction of the
sensitivity function over all considered harmonics of the disturbance and
over all potential frequencies within the specified (in this case 1%) relative
uncertainty interval. Attenuation of the periodic disturbances corresponds to
γp,∆ < 1, where γp,∆ = 0 indicates perfect rejection. The non-periodic
performance index γnp [51] is defined as the highest amplification of the
sensitivity function over all frequencies within the bandwidth of the RC.

The RC is designed to have a bandwidth of 200 Hz (ωQ =200Hz), such that

the first eight harmonics (the fundamental frequency is 23.3 Hz, for a spindle
speed of 1360rpm) of the synthesised cutting force are compensated. This
bandwidth is chosen due to the model accuracy limitation beyond a
frequency of 200Hz. This limitation will have a minimal effect on the
controller compensation performance because the influence of the 8th or
higher harmonic components of the cutting forces is not significant
(Table 3.5 and 3.6). The designed RC has a periodic performance index γp,∆
of 0.063 and a non-periodic performance index γnp of 6dB.

115
Cutting Force Compensation Design

Fig. 5.21 shows the Bode diagrams of the loop gain, sensitivity function, and
the complementary sensitivity function of the y-axis based on the designed
second order RC and a general control scheme that is the cascade P/PI
controller (designed in section 2.5). The high loop gain magnitudes at the
harmonic frequencies (fig. 5.21(a)) are translated to the reduction in the
magnitude of the sensitivity function at the corresponding frequencies.
Fig. 5.21(b) compares the sensitivity curves between the cascade P/PI
control scheme and the control scheme with the RC added. The non-periodic
performance degradation is identified by the arrows A, and is smaller than
6dB as predicted by γnp.
Bode Plot of Loop Gain

80

60

40
Loop Gain [dB]

20

-20

-40

-60
0 1 2
10 10 10
frequency (Hz)
(a)
Sensitivity: S(z) Complementary Sensitivity Function

10 10

0 A 0

-10 -10
Magnitude [dB]
Magnitude (dB)

-20 -20
S-RC
-30 -30
S-Cascade
-40 -40

-50 -50

-60 -60

-70 -70
0 1 2 0 1 2
10 10 10 10 10 10
Frequency (Hz) Frequency (Hz)

(b) (c)

Fig. 5.21. Y-axis: Bode plots of (a) the loop gain, (b) sensitivity function and
(c) complementary sensitivity function with second order RC

116
 Cutting Force Compensation Design

5.4.2 Experimental Validations

The repetitive controller is implemented on the test setup to compensate the
synthesized cutting forces applied at the input of the system. Fig. 5.22 shows
a schematic diagram of the overall control structure. The RC (the blue line)
is structured as an add-on module to the existing cascade P/PI position
controller. A GMS friction model feedforward (green line) and an inverse-
model-based disturbance observer (red line) compensate mainly the friction
forces.

Fig. 5.22. A schematic diagram of a cascade P/PI controller with a RC

module and friction compensation elements.

First, the cutting forces compensation is validated on each individual axis

during reference tracking. The reference trajectory is a sinusoidal signal with
amplitude of 30mm and a maximum tracking velocity of 100mm/s.
Synthesized cutting forces with 1, 2, 3 and 15 harmonic contents are applied
at the input of the system. Fig. 5.23 and fig. 5.24 show the position tracking
errors for cases with and without the repetitive controller along the x-axis
and the y-axis respectively. The RC is able to compensate these repetitive
disturbances effectively up to the 8th harmonic. The influence of the
remaining harmonics (9-15) on the tracking error is negligible.

The experimental validation is extended to include simultaneous

compensation of cutting forces in both axes. Three different control
configurations are compared, namely:
i. no friction and no cutting force compensation
ii. inverse-model-based disturbance observer with GMS friction model
feedforward
iii. repetitive controller with the inverse-model-based disturbance
observer and with the GMS friction model feedforward
117
Cutting Force Compensation Design

1 harmonic 2 harmonics 3 harmonics 15 harmonics

400 400 400 400
cutting force [N]

200 200 200 200

0 0 0 0

-200 -200 -200 -200

-400 -400 -400 -400

0 0.1 0.2 0 0.1 0.2 0 0.1 0.2 0 0.1 0.2
20 20 20 20
position error [µm]

10 10 10 10
no RC

0 0 0 0

-10 -10 -10 -10

-20 -20 -20 -20

0 1 2 0 1 2 0 1 2 0 1 2
20 20 20 20
position error [µm]

10 10 10 10
with RC

0 0 0 0

-10 -10 -10 -10

-20 -20 -20 -20

0 1 2 0 1 2 0 1 2 0 1 2
time [s] time [s] time [s] time [s]

Fig. 5.23. Y-axis: Measured position tracking errors with and without
the RC for different harmonic component of the cutting forces.

1 harmonic 2 harmonic 3 harmonic 15 harmonics

400 400 400 400
cutting force [N]

200 200 200 200

0 0 0 0
-200 -200 -200 -200
-400 -400 -400 -400
0 0.1 0.2 0 0.1 0.2 0 0.1 0.2 0 0.1 0.2
position error [µm]

20 20 20 20
10 10 10 10
no RC

0 0 0 0
-10 -10 -10 -10
-20 -20 -20 -20
0 1 2 0 1 2 0 1 2 0 1 2
20 20 20 20
position error [µm]

10 10 10 10
with RC

0 0 0 0
-10 -10 -10 -10

-20 -20 -20 -20

0 1 2 0 1 2 0 1 2 0 1 2
time [s] time [s] time [s] time [s]

Fig. 5.24. X-axis: Measured position tracking errors with and without
the RC for different harmonic components of the cutting forces.

118
 Cutting Force Compensation Design

Fig. 5.25 shows for each case the reference contour, the measured contour,
and the radial tracking errors. The measured contours have been amplified
by a factor of 200 for display purposes. The root mean square (rms) values
of the radial tracking errors for the three cases are summarized in Table 5.4.
No Friction + Cutting Force Inverse-Model-Based RC + Inv. Mod. Based
Compensation Disturbance Observer + GMS FF. Dist. Observer + GMS FF.
40 40 40

position y [mm]
position y [mm]

position y [mm]
20 20 20

0 0 0

-20 -20 -20

measured pos.
error = x200
error = x200 -40 reference pos. error = x200
-40 -40 -20 0 20 40 -40
-40 -20 0 20 40 -40 -20 0 20 40
position x [mm] position x [mm] position x [mm]

20 20 20
radial error [µm]

radial error [µm]

10 10 radial error [µm] 10

0 0 0

-10 -10 -10

-20 -20 -20

0 90 180 270 360 0 90 180 270 360 0 90 180 270 360
angle [degree] angle [degree] angle [degree]

(i) (ii) (iii)

Fig. 5.25. Circle tests: Measured position and radial tracking errors
for cases (i), (ii), and (iii)

Table 5.4. rms of the radial tracking errors for cases (i), (ii), and (iii)
case (i) case (ii) case (iii)
rms 7.68µm 4.43 µm 1.09 µm

The repetitive controller yields the lowest rms of the measured radial
tracking errors. A reduction of 85% is obtained compared to case (i) where
no friction and cutting forces compensation are applied. These results show
the ability of the repetitive controller to effectively eliminate the first 8
harmonics of the cutting force. Higher harmonics have limited influence on
the tracking error. The 8th harmonic (186 Hz) lies well beyond the
bandwidth of the position controller (45Hz) and of the inverse model
disturbance observer (60Hz). This is possible since the repetitive controller
is only active in small frequency ranges around these harmonics and the
system phase lag is compensated for in its design [21].
119
Cutting Force Compensation Design

The disturbance observer performance is limited by its inherent phase lag. A

higher bandwidth is possible with a more accurate model, but will always be
lower than the bandwidth of the repetitive controller due to uncompensated
system delay.
In the next section, actual cutting operations are performed on the considered
test setup. Because of the superior performance of the RC, only this
technique is implemented and validated.

5.5 Cutting Forces Compensation during an

Actual Cutting Process
Actual cutting tests are performed on the test setup. The cutting test
characteristics are summarized in Table 5.5.

Table 5.5. Cutting test characteristics used in cutting force compensation

Cutting conditions Descriptions

Work piece material Aluminium

Cutter diameter 16mm
No. of edges 4
Feed per tooth 0.1mm/tooth
Spindle speed 1360rpm
Axial depth of cut 4mm
Radial depth of cut 1mm
Sampling frequency 2000Hz

The following control configurations are considered for the evaluation of the
cutting force compensation:
(i) no friction and no cutting force compensation
(ii) inverse-model-based disturbance observer + GMS friction model
feedforward
(iii) inverse-model-based disturbance observer and GMS friction
model feedforward with cutting force compensation using a
second order repetitive controller
120
 Cutting Force Compensation Design

The cutting tests are performed on the considered test setup for a circle with
a radius of 25mm and at a tangential tracking velocity of 5mm/s (see fig.
5.26). The slower tracking velocity will emphasize the advantage of the
GMS friction model in compensating the highly nonlinear friction behaviour
during velocity removal.

Fig. 5.26. Cutting force compensation during actual cutting process

Fig. 5.27 compares the table axial tracking errors and their spectral contents
for the three control configurations. A gradual improvement in tracking
performance is observed from case (i) to case (iii). The spectral analysis
results shown in fig. 5.27 indicate an almost complete removal of the cutting
force harmonics using the repetitive controller. However, due to structural
vibration during the cutting process, the magnitudes of some non-harmonic
frequencies are amplified, especially between the 250 Hz and 300 Hz
frequency range (indicated by the circle A) and also at the lower frequency
range between 1Hz and 20 Hz (indicated by the circle B). These effects are
clearly identified by the circles A and B on the spectral analysis results in fig
5.27.

Fig. 5.28 compares the reference contour and the measured contour of each
of the three cases. The tracking performance is analysed from the magnitude
of the radial tracking errors. The control configuration using the repetitive
controller yields the least radial tracking error and the most accurate contour.
This is clearly indicated by the rms measurements of the radial tracking
errors for the three cases shown in Table 5.6.

121
Cutting Force Compensation Design

No Friction and No GMS + GMS + Inv. Model Observer +

Cutting Force Compensation Inv. Model Observer Repetitive Controller
Position Error
(X-Axis) [µm]

20 20 20

0 0 0

-20 -20 -20

0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
Position Error
(Y-Axis) [µm]

20 20 20

0 0 0

-20 -20 -20

0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
Spectral Content Spectral Content

1 1 1
(X-Axis)

0.5 0.5 0.5

A A
0 0 0
0 100 200 300 0 100 200 300 0 100 200 300

1 1 1
(Y-Axis)

0.5 0.5 0.5

B
0 0 0
0 100 200 300 0 100 200 300 0 100 200 300
[Hz] [Hz] [Hz]

(i) (ii) (iii)

Fig. 5.27. Measured position tracking errors and spectral analyses
for control configurations (i), (ii), and (iii).
No Friction + No GMS + GMS +Inv. Model
Cutting Forces Comp. Inv. Model Observer Observer + RC
40 40 40
position Y [mm]

position Y [mm]

position Y [mm]

20 20 20

0 0 0

-20 -20 -20

error x 200 error x 200

error x 200 -40 -40
-40 -40 -20 0 20 40 -40 -20 0 20 40
-40 -20 0 20 40
position X [mm] position X [mm] position X [mm]

30 30 30

20 20 20
radial error [µm]

position X [mm]

position X [mm]

10 10 10

0 0 0

-10 -10 -10

-20 -20 -20

-30 -30 -30

0 90 180 270 360 0 90 180 270 360 0 90 180 270 360
angle [degrees] angle [degrees] angle [degrees]

(i) (ii) (iii)

Fig. 5.28. Contour tracking measurement and radial tracking errors
for control configurations (i), (ii), and (iii).
122
 Cutting Force Compensation Design

Table 5.6. rms of the radial tracking errors for cases (i), (ii), and (iii) during
actual cutting process
radial error case (i) case (ii) case (iii)
rms 3.04 µm 1.72 µm 0.86 µm

5.6 Summary
This chapter discusses the compensation of the cutting forces. Several
techniques that are available in literature but have not been applied
specifically for cutting forces compensation in milling machine have been
discussed and presented. These techniques include the inverse-model-based
disturbance observer, the explicit estimation and compensation of cutting
forces using the Ferraris sensor relative acceleration measurements, and
finally the repetitive controller. The compensation performances are first
validated experimentally using synthesized cutting forces that are applied at
the input of the system. The performance of the inverse-model-based
disturbance observer and the second method that is based on the Ferraris
sensor measurement are restricted by the bandwidth limitation of the low
pass filter. The best performance is observed with the repetitive controller.
An almost complete removal of the cutting forces harmonics is achieved.
Finally, actual cutting process was performed on the test setup. It is shown
that the developed second order repetitive controller is able to almost
completely compensate the cutting forces.

123
124
 Chapter 6

Conclusions & Future Study

This thesis investigates friction and cutting forces compensation in a linear

drive xy feed table based on several techniques that have been discussed in
literature. A simple and highly practical cascade control configuration is
selected as the primary controller. The design of a cascade P/PI controller is
first discussed. Careful attention is given during the design steps to minimize
the influence of the system natural resonance frequencies on system stability
and tracking behavior of the drives. A PI velocity controller and a P position
controller for each axis are chosen. The system bandwidth of each axis is
limited by the first anti-resonance and resonance frequency at 40 Hz and 45
Hz for x and y axes respectively. These anti-resonance and resonance
frequencies have been shown to be caused by the relative motion between
the base of the machine and the floor. During the design analysis, notch
filters are added to suppress the high resonance frequencies. The advantages
of adding the notch filters have been clearly demonstrated in the design
analysis discussed in Chapter 2. Velocity feedforward and an inverse model
position feedforward compliment the cascade P/PI structure in order to
eliminate the reference tracking errors.

Compensation of disturbance forces acting on the linear drives is critical for

high tracking performance. The non-linear friction behaviour is compensated
using both friction model-based and friction model-free approaches. The
friction compensation performances of these methods are compared based on
the magnitude of the quadrant glitches – that are “spikes” at each quadrant of
a circle, resulting from the high and complex nonlinear friction behaviour
during velocity reversal. The combination of friction feedforward based on
the recently developed Generalized Maxwell-slip (GMS) friction model and
the inverse-model-based disturbance observer yields the smallest quadrant
glitch magnitudes. The magnitude of the quadrant glitches were reduced
from a maximum of 27 µm to only 3µm. This thesis has demonstrated the
125
Conclusions and Future Study

effectiveness of the GMS friction model especially at lower tracking

velocity. This can be explained from the model’s ability to accurately
describe the friction pre-sliding regime that is dominant at slower motion. In
addition, the accuracy of the friction model especially in describing the pre-
sliding behaviour of the friction regime determines the degree of
performance. This thesis does not consider an adaptive friction model that
will be necessary to improve the robustness of this approach. Further study
on the application of such technique is desired.

On the other hand, the performance of the inverse-model-based disturbance

observer is limited by the bandwidth of the applied low pass filter Q. The
filter Q is necessary for system stability. A low pass filter Q with a
bandwidth of 60 Hz and 90 Hz are designed for the x and y axes. These
bandwidths are selected to ensure a reasonable stability margin and are
restricted by the unmodeled system dynamics. The inverse-model-based
disturbance observer is able to compliment the GMS friction model
feedforward approach because it estimates the disturbance forces for
example cutting forces if present, or forces representing the linear model and
friction model inaccuracies that cannot be explained by the system output
and the available system model.

Several different techniques are studied for cutting force compensation. The
application of the inverse-model-based disturbance observer is further
extended to include cutting forces compensation. As discussed earlier, the
performance is again limited by the bandwidth of the low pass filter Q. Next,
a method that is based on explicit estimation of the cutting forces from
relative acceleration measurement using the Ferraris sensor is applied. The
compensation performance is negatively influenced by the delay in the
estimator caused by the delay in the drive. In addition, this thesis has shown
that the application of a low pass filter in the feedback loop degrades the
compensation performance. A successful compensation of the synthesized
cutting forces is observed using a repetitive controller. The spectral analysis
of the tracking errors shows complete removal of the cutting force up to its
eighth harmonic. However, due to the waterbed effect, an improvement in
the performance at the harmonic frequencies has yielded the degradation of
performance in other frequencies.

126
Conclusions and Future Study

A study on other compensation techniques is desirable. This could include

the application of the nonlinear sliding mode control (SMC) or the robust H∞
approach. The application of a sliding mode controller is attractive because
SMC is widely known for its robustness against matched disturbances,
which means that the disturbance forces are completely rejected once sliding
has occurred.

An actual cutting process was performed on the test setup. The performance
of the inverse-model-based disturbance observer and the repetitive controller
are compared. Analysis of the drive position tracking errors indicates a
successful compensation of the cutting forces using the repetitive controller.
However, the roundness measurements of the finished product (a circular
shape work-piece) indicate some structural problem of the test setup. An
additional study is necessary to clearly identify the cause of this deficiency.
In addition, a surface roughness analysis on the finished products (circles) is
desired in order to establish the effect of the high frequencies component of
the cutting forces on surface finish quality.

In the test setup, the tracking performance is highly influenced by the high
friction and actuator saturation. It is believed that the full potential of these
techniques has not yet been fully explored. Also, a future study on the
application of adaptive repetitive controller is desired to compensate
changing frequencies of the cutter drive because of the changes in the set-
point reference and also due to the cutting tools and work-piece interactions.

127
128
Bibliography

[1] T. Prajogo, “Experimental study of pre-rolling friction for motion-

reversal error compensation on machine tool drive systems”, PhD
thesis, Department of Werktuigkunde, Katholieke Universiteit Leuven,
Belgium, 1999.

[2] P. Van den Braembussche, Robust motion control of high-performance

machine tools with linear motors, PhD thesis, Department of
Werktuigkunde, Katholieke Universiteit Leuven, Belgium, 1998.

[3] G. Pritschow, “A comparison of linear and conventional

electromechanical drives”, Annals of the CIRP, vol. 47, no2, pp. 541-
548, 1998.

[4] B. P. Armstrong-Hélouvry, Dupont and C. Canudas de Wit, “A survey

of models, analysis tools, and compensation methods for the control of
machines with friction,” Automatica, vol. 30(7), pp. 1083-1138, 1994.

[5] P.R. Dahl, “Measurement of solid friction parameters of ball bearings,”

Proc. of 6th Annual Symp. on Incremental Motion, Control System and
Devices, 1977.

[6] Canudas de Wit, C., H. Olsson, K. Astrom, and P. Lishinsky. “A

new model for control system with friction”, IEEE Trans. on
Automatic Control, 40, 5, pp. 419-425. 1995.

[7] J. Swevers, F. Al Bender, C. Ganseman and T. Prajogo, “An integrated

friction model structure with improved pre-sliding behaviour for
accurate friction compensation,” IEEE Trans. on Automatic
Control, vol. 45, no. 4, pp. 675-686, 2000.

[8] Lampaert, V., J. Swevers, F. Al Bender, C. Ganseman and T. Prajogo,

“Modification of the Leuven integrated friction model structure”,
IEEE Trans. on Automatic Control, 47, 4, pp. 683-687. 2002.

[9] F. Al Bender, V. Lampaert and J. Swevers, “The generalized Maxwell-

slip friction model: a novel model for friction simulation and

129
Bibliography

compensation,” IEEE Trans. on Automatic Control, vol. 50, no. 11, pp.
1883-1887, 2005.

[10] T. Tjahjowidodo, F. Al Bender, H. Van Brussel, and W. Symens,

“Friction characterization and compensation in electro-mechanical
systems,” Journal of Sound and Vibration. Article in press, 2007.

[11] E.D. Tung, G. Anwar and M. Tomizuka, “Low-velocity friction

compensation and feedforward solution based on repetitive
control”, Journal Dynamics System Measurement and Control, vol.
115, pp. 279-284. 1993.

[12] S. Hara., , Y. Yamamoto, T. Omata, , M. Nakano, “Repetitive control

system: a new type servo system for periodic exogenous signals”,
IEEE Transactions on Automatic Control, Vol. 33, No. 7, pp. 659-
668, 1988.

[13] V. Lampaert, J. Swevers, F. Al Bender, “Comparison of model and

non-model based friction compensation techniques in the
neighbourhood of pre-sliding regime,” Proc. of the 2004 American
Control Conference, Boston. pp. 1121-1126, 2004.

[14] T. Tjahjowidodo, Characterization, modelling and control of

mechanical systems comprising material and geometrical
nonlinearities, PhD thesis, Department of Werktuigkunde, Katholieke
Universiteit Leuven, Belgium, 2006.

[15] V.I. Utkin, “Variable structure system with sliding mode”, IEEE
Trans. on Automatic Control, 22, pp 212-222, 1977.

[16] K.D. Young, V.I. Utkin, U. Ozguner, “A control engineer’s guide to

sliding mode control”, IEEE Transactions on Control Systems
Technology, vol 7, no.3, May 1999.

[17] Y. Altintas, “Sliding mode controller design for high-speed feed

drives”, Annals of the CIRP, 49 (1), pp 265-270, 2000.

[18] Z. Jamaludin, H. Van Brussel, J. Swevers, “Classical cascade and

sliding mode control tracking performances for a x-y feed table of a
high-speed machine tools”, Int. J. Precision Technology, vol.1/1,
pp. 65-74, 2007.
130
 Bibliography

[19] V.I. Utkin, “Sliding mode control in discrete-time and difference

systems”, Variable Structure and Lyapunov Control, London, UK,
Springer-Verlag, 1993.

[20] S.K. Spurgeon, C. Edwards, Sliding Mode Control, Taylor and France,
London, pp 15-16, 1998.

[21] V.I. Utkin, J. Guldner, J. Shi, Sliding Mode Control in Electro-

Mechanical Systems, Taylor & Francis, London, 1999.

[22] J.J. Slotine, “Sliding controller design for non-linear systems”, Int.
Journal for Control, 40, pp. 421-434, 1984.

[23] M. Steinbuch, “Repetitive control for systems with uncertain period-

time”, Automatica, 38, pp. 2103-2109, 2002.

[24] H. Kwakernaak, “Robust control and H∞ Optimization”, Automatica,

vol 29, pp. 255-273, 1993.

[25] H. Van Brussel, and P. Van den Braembussche, “Robust control of feed
drives with linear motors”, Annals of CIRP, 47 (1), PP 325-328, 1998.

[26] K. Ohnishi, “Advanced motion control in robotics”, Proc. IEEE

IECON89, Philadelphia, PA, 1989.

[27] T. Murakami, K. Ohnishi, “Advanced motion control in mechatronics

– A tutorial”, Proc. IEEE Int. Workshop Intelligent Control, vol. 1,
Istanbul Turkey, pp 9-17, Aug. 1990.

[28] C.J. Kempf, S. Kobayashi, “Disturbance observer and feedforward

design for a high-speed direct drive positioning table”, IEEE Trans. on
Control System Tech., vol. 7, no. 5, pp 513-526, 1999.

[29] G.Pritschow, S. Fritz, P. Pruschek, “Reconstruction of process forces

of direct drives using the Ferraris sensor”, Proceedings of the VIIth
Int. Conference on Monitoring and Automatic Supervision in
Manufacturing AC 04, Zakopane, Poland, 19-21 August 2004.

131
Bibliography

[30] G. Pritschow, W. Phillip, “Research on the efficiency of feedforward

controllers on machine direct drives”, Annals of the CIRP, vol. 41, no1,
pp. 411-415, 1992.

[31] S. Doenitz, “Comparison of disturbance suppression for servo

drives”, 8th European Conference on Power Electronics and
Applications, Lausanne (1999).

[32] P. Boucher, D. Dumur, K.F. Rahmani, “ Generalized predictive cascade

control (GPCC) for machine tool drives”, Annals of the CIRP, vol. 39,
no1, pp. 357-360, 1990.

[33] R. Pintelon and J. Schoukens, “System identification – a frequency

domain approach”, IEEE Press, New Jersey, pp. 199-227, 2000.

[34] A.M. Rankers, Machine dynamics in mechatronic systems – an

engineering approach, PhD Dissertation, Universiteit Twente, 1997.

[35] G.F. Franklin, J.D. Powell, A.E. Naeni, “Feedback control of

dynamic system”, 4ed., Addison-Wesley, New Jersey, 2002.

[36] J. Wang, “Robust tracking controller design with application to the

motion control of an x-y feed table for high-speed machining”, PhD
Thesis, Faculty of Engineering, Department of Mechanical,
Division of Production, Machine Design and Automation (P.M.A),
Katholieke Universiteit Leuven, Belgium, 2004.

[37] W.D. Iwan, “A Distributed-element model for hysteresis and its

steady-state dynamic response”, Journal of Appl. Mech, 33, pp. 893-
900, 1966.

[38] F. Al Bender, V. Lampaert, S.D. Fassois, D.D. Rizos, K. Worden, D.

Engster, A. Hornstein, and U. Parlitz, “Measurement and identification
of pre-sliding friction dynamics”, Nonlinear Dynamics of Production
Processes, Wiley-VCH, 2003.

[39] F. Al Bender, W. Symens, J. Swevers, and H. Van Brussel,

“Theoretical analysis of the dynamic behaviour of hysteresis elements
in mechanical system,” Int . Journal of Nonlinear Mechanics, vol. 39,
pp. 1721-1735, 2004.

132
 Bibliography

[40] Z. Jamaludin, H. Van Brussel, J. Swevers, “Quadrant glitch

compensation using friction model-based feedforward and an
inverse-model-based disturbance observer”, Proc. Advanced Motion
Control, Trento, Italy, 2008.

[41] H.S. Lee and M. Tomizuka, “Robust motion controller design for
high-accuracy positioning systems”, IEEE Trans. on Ind. Electronics,
vol. 43, 1, pp. 48-55, 1996.

[42] L. Yi and M. Tomizuka, “Two-degree-of-freedom control with robust

feedback control for hard disk servo systems”, IEEE Trans. on
Mechatronics, vol. 4, 1, pp. 17-24, 1999.

[43] T. Umeno, Y. Hori, “Robust speed control of DC servomotors using

modern two-degree-of freedom controller design”, IEEE Trans. of Ind.
Electronic, vol. 38, pp 363-368, 1991.

[44] J.R. Ryoo, T.Y. Doh, M.J. Chung, “Robust disturbance observer for
the track-following control system of an optical disk drive”, Control
Engineering Practice, 12, pp. 577-585. 2004.

[45] S.J. Kwon, “Robust tracking control and state estimation of

mechanical systems: a pertubation compensator based approach”, PhD
Dissertation, POSTECH, (2002).

[46] S. Skogestad, I. Posletthwaite, Multivariable Feedback Control -

Analysis and Design, John Wiley & Sons, Chichester, West Sessex,
England, 2005.

[47] B. Hiller, “Ferraris acceleration sensor –Principle and field of

application in servo drive”, Hübner and Elektromaschinen AG, Berlin,
Germany.

[48] G. Pritschow, “Ferraris sensor – the key for advanced dynamic drives”,
Institute for Control Engineering for Machine Tool and Manufacturing
Units, University of Stuttgart, Germany.

[49] B.A. Francis and W.M. Wonham, “The internal model principle for
linear multivariable regulators”, Applied Mathematics and Optics, 2,
pp. 170-194, 1975.

133
Bibliography

[50] M. Tomizuka, T.C. Tsao, and K.K. Chew, “Discrete-time domain

analysis and synthesis of repetitive controllers”, Proc. of the 1998
American Control Conference, pp. 860-868, 1988.

[51] G. Hillerström, “Adaptive suppression of vibration – a repetitive

control approach”, IEEE Trans. on Control Systems Technology, 4(1),
pp. 72-77, 1996.

[52] G. Pipeleers, B. Demeulenaere, J. De Schutter, and J. Swevers, “Robust

high-order repetitive control”, accepted for the 2008 American Control
Conference, Seattle, June 11-13, 2008.

[53] G. Pipeleers, B. Demeulenaere, J. De Schutter, and J. Swevers, “Robust

High-Order Repetitive Control: Optimal Performance Trade-Offs”,
Accepted for Automatica. 2008.

134
Curriculum Vitae

Name : Zamberi Jamaludin

Date of Birth : 4th September 1972
Place of Birth : Malacca, Malaysia
Nationality : Malaysian
Home Address : 22 Jalan Burung Pikau,
Taman Bukit Maluri,
52100 Kepong, Kuala Lumpur,
Malaysia.
Telephone No. : (60) 12-3911912
Position : Lecturer
Office Address : Faculty of Manufacturing Engineering,
Department of Robotics and Automation,
P.O.Box 1200, 75450 Ayer Keroh, Malacca,
Malaysia
Office Tel. No. : (60) 6-233 2421
Email Address : zamberi@utem.edu.my

Academic Backgrounds

1. Bachelor of Engineering (Chemical Engineering)

Lakehead University (Department of Chemical Engineering),
955 Oliver Road, Thunder Bay, Ontario, Canada.
1993-1997.
Thesis: Study on Optimal Parameters for Pulp and Paper Sludge
Dewatering Process Using Polymer Based Material.
Supervisor: Prof. Dr. Robert Rosehart

2. Master Engineering (Manufacturing Eng.)

National University of Malaysia (U.K.M)
Bangi, Selangor, Malaysia.
1999-2001.
Thesis: Analysis for Optimal Parameters For Filling Process In
Metal Injection Molding Utilising the Moldflow Flow Simulation
Software.
Supervisor: Associate Prof. Dr. Norhamidi Muhammad
135
Curriculum Vitae

Working Experiences
1. Japan-Malaysia Technical Institute,
Vocational Training Officer,
Manpower Dept.,
Ministry of Human Resources.
01/08/1997 – 09/06/2002

2. Malaysia-Malacca Technical University (UTeM)

Lecturer,
Ministry of Higher Education.
10/06/2002 - present

Awards
1. Malaysian Federal Government Scholarship
High School and Bachelor Studies in Canada
1990-1997.
2. Japan Int. Cooperation Agency (JICA) Scholarship
Malaysian-Japanese Counterpart Training in Japan
27/10/1997 – 08/02/1998.
3. Japan Int. Cooperation Agency (JICA) Scholarship
Seminar on Enhancing Vocational Training, Japan
23/10/2000 – 08/12/2000.
4. Malaysian Ministry of Higher Education Scholarship
PhD Studies at K.U.Leuven, Belgium
29 Sept. 2003 – 28 Sept. 2008.

136
List of Publications
Journals
1. Z. Jamaludin, H. Van Brussel, J. Swevers, “Friction Compensation
of a XY Feed Table using Friction Model-Based Feedforward and an
Inverse-Model-Based Disturbance Observer”, Submitted for
publication in IEEE Trans. on Industrial Electronics- Advances in
Motion Control . 2008.

2. Z. Jamaludin, H. Van Brussel (1), G. Pipeleers, J. Swevers,

“Accurate Motion Control of XY High-Speed Linear Drives using
Friction Model Feedforward and Cutting Forces Estimation”, Annals
of the CIRP-Manufacturing Technology, vol. 57/1, pp 403-406,
2008.

3. Z. Jamaludin, H. Van Brussel, J. Swevers, “Classical Cascade and

Sliding Mode Control Tracking Performances for a X-Y Feed Table
of a High-Speed Machine Tools”, Int. J. Precision Technology,
vol.1/1, pp. 65-74, 2007.

Proceedings
1. Z. Jamaludin, H. Van Brussel, J. Swevers, "Design of a Disturbance
Observer and Model-Based Friction Feedforward to Compensate
Quadrant Glitches", The 9th International Conference on Motion and
Vibration Control - MOVIC, Munich, Sept. 15-18, accepted, 2008.

2. Z. Jamaludin, H. Van Brussel, J. Swevers, “Quadrant Glitch

Compensation using Friction Model-based Feedforward and an
Inverse-Model-Based Disturbance Observer”, Proc. of the 10th
International Workshop on Advanced Motion Control, Trento, Italy,
pp. 212-217, 2008.

3. Z. Jamaludin, H. Van Brussel, J. Swevers, “Tracking Performances

of Cascade and Sliding Mode Controllers with Application to a XY
Milling Table”, Proc. on Int. Conference on Noise and Vibration
Engineering, Leuven, Belgium, Sept. 18-20, 2006.

137
List of Publications

4. Z. Jamaludin, H. Van Brussel, J. Swevers, “Comparison of Classical

Cascade and Sliding Mode Control of a X-Y Feed Table of a High-
Speed Machine Tools”, Proc. of Int. Conf. on Manufacturing
Science and Technology-ICOMAST, August 28-30, Malacca,
Malaysia, 2006.

5. Z. Jamaludin, H. Van Brussel, J. Swevers, “Comparison of Classical

Cascade and Sliding Mode Control for a X-Y Feed Table of a High-
Speed Machine Tools”, Book of Abstracts - 25th Benelux Meeting
on System and Control, March 15-18, Heeze, The Netherlands,
2006.

138
Appendix A

Machine Dynamics Analysis

The machine conditions are adjusted and the frequency response function
relating the input voltage to the drive amplifier to output position is
measured. Fig. A1 shows the FRF measurement for two different machine
conditions:

(i) The bolts that hold the machine to the ground are removed
(indicated by the solid line).
(ii) The bolts remained in place (indicated by the dash line)

Bode Diagram of TF from Voltage to Drive Amplifier to Encoder (micrometer)

120
100 no mat, unbolted
no mat, bolted
80 A
Magnitude (dB)

60
40
20
A
0
-20
0
-45
-90
Phase (deg)

B
-135 B
-180
-225
-270
-315
0 1 2
10 10 10
Frequency (Hz)

Fig. A1. Y-axis: FRF measurement for different machine

conditions (bolted and unbolted)

The anti-resonance and resonance frequencies shifted from 45Hz to 12Hz

with the removal of the bolts (indicated by the circle A and B).
*Two bolts are located along the y-direction and 1 bolt each along the x-direction of
the machine

139
Appendix B

Ferraris Acceleration Sensor Calibration

This section describes the procedures and the results of the identification of
the Ferraris relative acceleration sensor. The sensor gain sensitivity and
offset values are determined. The Ferraris sensor is graphically shown in fig.
B.1.

Fig B.1: Ferraris relative acceleration sensor

Procedures

¾ The milling table was excited with sinusoidal signals at selected

frequencies between 0.2-100Hz.

¾ Large gain was applied at lower frequencies (until 25Hz for

displacement of ~10mm) and the gain was reduced for higher
frequencies excitation to avoid saturation effect.

¾ The encoder measurement and the Ferraris sensor signals were

recorded using the dSPACE.

Gain Sensitivity and Zero Offset Determination

The calibrated relative acceleration signal, a(t) is defined as follow:

a (t ) = m(t ) ⋅ G + Z (B.1)

140
 Appendix B

m(t) is the Ferraris measurement signal, G(t) is the sensor sensitivity, and z(t)
is the offset. Velocity signal v(t) can derived from the integration of (1),

v(t ) = G ⋅ ∫ m(t ) dt ⋅ + Zt + v0 (B.2)

v0 isthe initial velocity. Finally, position, p(t) is obtained from integration of

the velocity function:

z
p(t ) = G ⋅ ∫ ∫ m(t ) dτdt ⋅ + t 2 + v0 t + p0 (B.3)
2

p0 is the initial position.The sensor sensitivity, G and offset Z are then

obtained by comparing equation (B.3) to the actual position measurement
recorded by the encoder, e (t).

In matrix form:

⎡G ⎤
⎡ t2 ⎤⎢ Z ⎥
e ( t ) = ⎢ ∫ ∫ m dτdt t 1⎥ ⎢ ⎥ (B.4)
⎢⎣ 2 ⎥⎦ ⎢ v0 ⎥
⎢ ⎥
⎣ p0 ⎦

At t=kT, T being the sampling time, ∫ ∫ m dτdt = M

⎡G ⎤
⎡ 2
k T 2 ⎤⎢ Z ⎥
e (t ) = ⎢M kT 1⎥ ⎢ ⎥ (B.5)
⎣⎢ 2 ⎦⎥ ⎢⎢ v0 ⎥⎥
⎣ p0 ⎦

A least square method is then applied to solve for G and Z.

Results:
The Ferraris sensor sensitivity, G and the zero offset Z as a function of
frequency are illustrated below.

141
 Appendix B

7 4
x 10 Ferraris sensor sensitivity x 10 Ferraris sensor zero offset
3.7 2

3.6
1

3.5
0
sensitivity,

zero offset,
3.4
-1
3.3

-2
3.2

-3
3.1

3 -4
0 20 40 60 80 100 0 20 40 60 80 100
frequency,[Hz] frequency,[Hz]

Fig B.2. Gain sensitivity and offset value of Ferraris

sensor as a function of frequency

The sensitivity and the zero offset values are rather uniform until 40Hz
before fluctuating drastically. The mean sensitivity and zero offset values
within this frequency range are

µm µm
G = 3.298e7 2
; Z = -8580 (B.6)
volt ⋅ s s2

The following figure illustrates the validity of the gain sensitivity and offset
in (B.6). The actual encoder measurement was compared to the position
signal generated from the double integration of the calibrated Ferraris signal.
A close comparison was observed.
4 4
x 10 0.4Hz x 10 2Hz 10Hz
1 1 2000
position,[micrometer]

1000

0 0 0

-1000

-1 -1 -2000

0 1 2 3 4 0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5

time,[s] time,[s] time,[s]
20Hz 40Hz 50Hz
500 200 100
position,[micrometer]

300
100 50

0 0 0

-100 -50
-300 encoder
pos. from Ferr. sensor
-500 -200 -100
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
time,[s] time,[s] time,[s]

Fig. B3. Comparison between encoder signal and position signals acquired
from the calibrated Ferraris relative acceleration signal
142
Appendix C

Cutting Force Estimation from Ferraris

Acceleration Sensor Measurement

The state space representations of the system, including the measurement

equation are:

⎡ x& ⎤ ⎡0 1 0 ⎤ ⎡ x⎤ ⎡ 0 0 ⎤
⎢ v& ⎥ = ⎢0 0 1 / M ⎥ ⎢ v ⎥ + ⎢ k f 1 ⎥⎡
− ⎥⎢
u ⎤
⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢M F (v
M ⎥ ⎣ friction ⎦ )⎥
⎢⎣d& ⎥⎦ ⎢⎣0 0 0 ⎥⎦ ⎢⎣d ⎥⎦ ⎢⎢ 0 0 ⎥⎦
⎣
(C.1)

⎡ x⎤
⎡ x ⎤ ⎡1 0 0 ⎤⎢ ⎥ ⎡ 0 0 ⎤⎡ u ⎤
y=⎢ ⎥=⎢ ⎥⎢v ⎥ + ⎢ k f 1 ⎥⎢
− ⎥ ⎣ F friction (v )⎥⎦
⎣a ⎦ ⎣0 0 1 / M ⎦ ⎢d ⎥ ⎢⎣ M M⎦
⎣ ⎦
(C.2)

Next, a reduced order observer is designed. Consider a general state equation

of the form:

x& = Ax + Bu
y = Cx, C = ⎣⎡ E 0 ⎦⎤
(C.3)

E is a unit matrix. A state of the system that contains a combination of the

measured variable y and the non-measured state variable w are:

⎡ y⎤
x=⎢ ⎥
⎣⎢ w⎦⎥
(C.4)

The system equation then becomes,

143
Appendix C

⎡ y& ⎤ ⎡ A11 A12 ⎤ ⎡ y ⎤ ⎡ B1 ⎤

⎢ ⎥=⎢ ⎥ ⎢ ⎥ + ⎢ ⎥u (C.5)
⎣⎢ w& ⎦⎥ ⎣ A21 A22 ⎦ ⎣⎢ w⎦⎥ ⎣ B2 ⎦

or,

y& = A11 y + A12 w + B1u (C.6a)

w& = A21 y + A22 w + B2u (C.6b)

w is the only unknown in (C.6a). Therefore, a measuring equation for w

becomes:

y = y& − A11 y − B1u = A12 w (C.7)

The observer equation for w then equals:

ˆ& = ( A22 − LA12 ) w

w ˆ + A21 y + B2u + Ly ,
(C.8)
ˆ = ( A22 − LA12 ) w
w& ˆ + A21 y + B2u + L ( y& − A11 y − B1u )

Equation (C.8) is modified to avoid using the derivative of y. Let:

z=w
ˆ − Ly
(C.9)
∴ z& = ( A22 − LA12 ) z + ( A22 − LA12 ) Ly + ( A21 − LA11 ) y + ( B2 − LB1 ) u

The matrix L is the observer gain. A transformation matrix T is required

since the measured acceleration signal is not a state variable. Pritschow [29]
has suggested a transformation matrix T that is equal to:

⎡1 0 0⎤
⎢ ⎥
T = ⎢0 0 1⎥ (C.10)
⎢⎣0 M 0 ⎥⎦

The state equation for the transformed system equals (with x’=Tx):

⎡ ⎤
⎡0 0 1 ⎤ ⎢ 0 0 ⎥
⎢ ⎥ ⎢ ⎥ ⎡ u ⎤
& = 0 0 0 x' + ⎢ 0
x' 0 ⎥ ⎢ ⎥ (C.11a)
⎥ ⎣⎢ friction ( ) ⎦⎥
⎢ ⎥ F v
⎢⎣0 1 0 ⎥⎦ ⎢
⎢kf 1
− ⎥
⎢⎣ M M ⎥⎦

144
 Appendix D

⎡ x ⎤ ⎡1 0 0 ⎤ ⎡ 0 0 ⎤⎡ u ⎤
y=⎢ ⎥=⎢ x'+ ⎢ k f 1⎥ (C.11b)
⎥
⎣ a ⎦ ⎣0 1 0 ⎦ ⎢ − ⎥ ⎢⎣ F friction (v )⎥⎦
⎣M M⎦

By definition,

⎡ y '⎤ ⎡1 0 0⎤ ⎡ x ⎤
y = ⎢ 1 ⎥ = y − Du = ⎢ x ' = ⎢ k 1 ⎥ (C.12a)
⎥ F friction (v )⎥
f
⎣ y 2 '⎦ ⎣0 1 0 ⎦ ⎢a − u+
⎣ M M ⎦

and,

⎡ x1 ' ⎤
⎢ ⎥
x' = ⎢ x2 ' ⎥ (C.12b)
⎢⎣ x3 ' ⎥⎦

The matrices A11, A12, A21, A22, B1, B2 (based on (B.11a) and the observer
gain L for the designed of the reduced order observer of the transformed
system are:

⎡0 0 ⎤ ⎡1⎤
A11 = ⎢ ⎥; A12 = ⎢ ⎥; A21 = [0 1]; A22 = [0];
⎣ 0 0 ⎦ ⎣0 ⎦
⎡0 0 ⎤ ⎡kf 1 ⎤
B1 = ⎢ ⎥; B2 = ⎢ − ⎥; (C.13)
⎣ 0 0 ⎦ ⎣ M M ⎦
L = [l1 0]

Substituting (C.13) into the reduced order observer of equation (C.9) yields,

[ ]
⎡ y '⎤ ⎡ k f
zˆ& = −l1 &z& + − l1 2 1 ⎢ 1 ⎥ + ⎢
⎣ y 2 '⎦ ⎣⎢ M
−
1 ⎤⎡
⎥⎢
F
u
(v
M ⎦⎥ ⎣ friction ⎦ )
⎤
⎥ (C.14a)

⎡y '⎤
ˆ = ˆz + ⎣⎡l1 0⎦⎤ ⎢ 1 ⎥ = ˆz + l1 y1 '
w (C.14b)
⎣ y2 ' ⎦

y1' = x, ˆx3 = w,
ˆ ∴ ˆx3 ' = ˆz + l1x (C.14c)

⎡ y⎤
Based on x = ⎢ ⎥ ,
⎢⎣ w⎥⎦

145
Appendix C

⎡ x̂1 ⎤ ⎡ y1' ⎤ ⎡ 0 ⎤ ⎡ 1 0⎤ ⎡ y1' ⎤

⎢ˆ ⎥ ⎢ ⎥ ⎢ ⎥ ˆ ⎢ ⎥ ⎡ y1 ' ⎤ ⎢ ⎥
x =
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
2 y 2 ' = 0 z + ⎢ 0 1 ⎢
⎥ y ' ⎥ = ⎢ y2 ' ⎥ (C.15)
⎢⎣ ˆx3 ⎥⎦ ⎢⎣ w ⎥⎦ ⎢⎣1 ⎥⎦ ⎢⎣l1 0 ⎥⎦ ⎣ 2 ⎦ ⎢⎣ ˆz + l1 y1 ' ⎥⎦

Equations (C.14a) and (C.15) can be simplified as follows:

ˆz& = A' ˆz + B1' y' + B2 ' u

(C.16)
ˆx' = D1' ˆz + D2 ' y'

Next, re-transformation into original physical states are accomplished, using

the transformation matrix T and by replacing y’ with y:

ˆz& = A' ˆz + B1 ' ( y − Du ) + B2 ' u

(C.17)
ˆx = T ( D1' ˆz + D2 ' y' )

or,

[ ]
⎡ x ⎤ ⎡k 1 ⎤⎡ u ⎤
z&ˆ = −l1&z& + − l12 1 ⎢ kf 1 ⎥+⎢ f − ⎥ ⎢F ⎥ (C.18a)
⎢ a − u + F (v )
friction ⎥ ⎣ M M ⎦ ⎣ friction (v )⎦
⎣ M M ⎦

⎡ ⎤
⎢ ⎥
⎡1 0 0⎤ ⎡ 1 ' ⎤ ⎢
x x ⎥ ⎡ xˆ ⎤
xˆ = ⎢⎢0 0 ⎥⎢x ⎥ = ⎢
1⎥ ⎢ 2 '⎥ zˆ + l1 x ⎥ = ⎢ vˆ ⎥ (C.18b)
⎢ ⎥ ⎢ ⎥
⎢⎣0 M 0⎥⎦ ⎢⎣ x3 '⎥⎦ ⎢ ⎛⎜ kf 1 ⎞⎥ ⎢⎣dˆ ⎥⎦
M a− u+ F friction (v )⎟
⎢ ⎜ M M ⎟⎥
⎣ ⎝ ⎠⎦

Fig. C.1 shows the structure of this reduced order observer based on (C.16)

Fig. C.1. Schematic diagram of the cutting forces estimator

146
Appendix D

General State Force Observer

Cutting forces contain harmonic components (as shown in Table 3.5 and 3.6)
3.4). A cutting force state observer estimates and compensates the cutting
forces based on these harmonic frequencies. The fundamental frequency
equals the frequency of the spindle speed (1360Hz).

Structure of a Disturbance Force Observer

Consider a sinusoidal disturbance force input d (t), its first derivative d& ( t ) ,
and the second derivative d&&( t ) :

d ( t ) = q1 = A' sin ( ωt + θ )
d& ( t ) = q&1 = q2 = A' ω cos ( ωt + θ ) (D.1)
d&& ( t ) = q&2 = − A' ω sin ( ωt + θ ) = −ω q1
2 2

A’ is the amplitude of the sinusoidal signal, t is the time, ω and θ are the
signal frequency and offset respectively. Equation D.1 shows that the second
derivative of the reference disturbance force is directly proportional to the
reference signal itself. A force disturbance observer can then be designed
according to this observation. The following is the state space representation
of a general order state observer:

ˆx& = Fx + Gu + L ( y − ˆy )
(D.2)
y = Hx

x is the state variable matrix, y is the output, and L is the observer gain
matrix. Equation (D.2) is extended based on the system transfer function
(2.2) for cutting force compensation of a single harmonic content.

147
Appendix D

⎡ x̂& ⎤ ⎡0 1 0 0⎤ ⎡ x̂ ⎤ ⎡ 0 ⎤ ⎡ l1 ⎤
⎢ ⎥ ⎢ ⎥ ⎢ vˆ ⎥ ⎢ ⎥ ⎢l ⎥
&
⎢v ⎥ ⎢
ˆ 0 − A B 0⎥
⎢ & ⎥ = ⎢0 0 ⎢ ⎥ + ⎢ B ⎥ [u ] + ⎢ 2 ⎥ ( y − ˆy ) (D.3a)
⎢ q̂1 ⎥ ⎢ 0 1⎥ ⎢ q̂1 ⎥ ⎢ 0 ⎥ ⎢ l3 ⎥
⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ && ⎥ ⎢0 0 −ω2 0 ⎥⎦ ⎣ q̂2 ⎦ ⎣ 0 ⎦ ⎣l4 ⎦
⎣ q̂2 ⎦ ⎣

⎡ x̂ ⎤
⎢ v̂ ⎥
y = [1 0 0 0] ⎢ ⎥ (D.3b)
⎢ q̂1 ⎥
⎢ ⎥
⎣ q̂2 ⎦

The output y is equal to z in (2.2), where with z=x and z=y for the x and y
axes respectively. The term A and B in matrix F and G are the system model
parameters (Table 2.1). Fig. D.1 shows a schematics diagram of the system
and the observer, with disturbance force input signal d, and a disturbance
force.

Fig. D.1. Schematic diagram of a force observer structure

148
 Appendix D

The fundamental equation (D.3) can be further extended to include other

cutting forces harmonic contents. The following is an example of a
disturbance force observer structure for compensation of cutting forces with
three harmonics components:

d1 ( t ) = q1 = A1 ' sin ( ω1t + θ1 )

d&1 ( t ) = q&1 = q2 = A1 ' ω1 cos ( ω1t + θ1 ) (D.4a)
d&&1 ( t ) = q&2 = − A1 ' ω12 sin ( ω1t + θ1 ) = −ω12 q1

d 2 ( t ) = q3 = A2 ' sin ( ω2 t + θ 2 )
d&2 ( t ) = q&3 = q4 = A2 ' ω2 cos ( ω2 t + θ 2 ) (D.4b)
d&&2 ( t ) = q&4 = − A2 ' ω22 sin ( ω2 t + θ 2 ) = −ω2 2 q3

d3 ( t ) = q5 = A3 ' sin ( ω3t + θ3 )

d&3 ( t ) = q&5 = q6 = A3 ' ω3 cos ( ω3t + θ3 ) (D.4c)
d&&3 ( t ) = q&6 = − A3 ' ω32 sin ( ω3t + θ3 ) = −ω3 q5 2

The total forces equal:

dtotal = d1 + d 2 + d3
or (D.5)
dtotal = q1 + q3 + q5 .

The structure of this observer is an extension to the fundamental structure

shown in fig. D.1. The disturbance observer design is based on the selection
of the design parameters, that is, the observer gains l1, l2, l3, l4. Large LHP
poles results in faster response but with reduced stability margin while small
LHP poles have slower response but better stability margin.

149
Appendix D

Design of a Disturbance Force Observer

The disturbance force observer is designed as an add-on module to the
primary cascade P/PI position controller, discussed in Chapter 2. The
schematic diagram of a control scheme that includes the cascade P/PI
position controller and the disturbance force observer is shown in Fig. D.2.

Fig. D.2. A schematic diagram of a cascade P/PI controller with a

general state force observer.

For analyses purposes, a single harmonic cutting force is considered. The

disturbance observer state space model equals:

ˆx& = Fxˆ + Gu + L ( y − ˆy ) ; but ˆy = Hxˆ

(D.6)
ˆx& = Fxˆ + Gu + Ly − LHxˆ

With,

⎡0 1 0 0⎤ ⎡0⎤
⎢ ⎥ ⎢ ⎥
0 − A B 0⎥ B
⎡⎣F ⎤⎦ = ⎢⎢ ; ⎡
⎣G ⎤
⎦ = ⎢ ⎥ ; ⎡⎣ H ⎤⎦ = ⎡⎣1 0 0 0⎤⎦ ; (D.7a)
0 0 0 1⎥ ⎢0⎥
⎢ ⎥ ⎢ ⎥
⎢⎣0 0 −ω2 0 ⎥⎦ ⎣⎢ 0 ⎦⎥

⎡ x̂ ⎤
⎢ ⎥
v̂
dˆ = Hxˆ = ⎡⎣0 0 1 0⎤⎦ ⎢ ⎥ (D.7b)
⎢ q̂1 ⎥
⎢ ⎥
⎢⎣ q̂2 ⎥⎦

150
 Appendix D

Taking the Laplace Transform,

sXˆ ( s ) = FXˆ ( s ) + GU ( s ) + LY ( s ) − LHXˆ ( s ) (D.8)

However,

U ( s ) = U PI ( s ) − HXˆ ( s )
(D.9)
( )
U ( s ) = P ⋅ PI Z ref − Y − PI ⋅ Vest ⋅ Y − HXˆ

Substituting (D.9) into (D.8),

( ( ) )
sXˆ ( s ) = FXˆ ( s ) + G P ⋅ PI Z ref − Y − PI ⋅ Vest ⋅ Y − HXˆ + LY ( s ) − LHXˆ ( s ) (D.10)

G ⋅ P ⋅ PI ( L − G ⋅ P ⋅ PI − G ⋅ PI ⋅Vest )
X̂ ( s ) = Z ref ( s ) + Y (s) (D.11)
sI − F + GH + LH sI − F + GH + LH

Also, (let the system be Gsys)

Y ( s ) = Gsys (U + D ) (D.12)

( ( )
Y ( s ) = Gsys P ⋅ PI Z ref − Y − PI ⋅ Vest ⋅ Y − HXˆ + D ) (D.13)

Substituting (D.11) into (D.13) and with further simplification,

⎛ Gsys H ⋅ GP ⋅ PI ⎞
⎜ Gsys PPI ⋅ − ⎟
⎜ sI − F + GH + LH ⎟⎠
Y (s) = ⎝ Z (s)
L − GPPI − G ⋅ PI ⋅ Vest ref
1 + Gsys PPI + Gsys PI ⋅ Vest + Gsys H (D.14)
sI − F + GH + LH
Gp
+ D(s)
L − G ⋅ P ⋅ PI − G ⋅ PI ⋅ Vest
1 + Gsys PPI + Gsys PI ⋅ Vest + Gsys H
sI − F + GH + LH

The sensitivity function, S(s) equals:

1
S (s) = (D.15)
L − G ⋅ P ⋅ PI − G ⋅ PI ⋅ Vest
1 + Gsys P ⋅ PI + Gsys PI ⋅ Vest + Gsys H
sI − F + GH + LH

151
Appendix D

Fig. D.3(a) and (b) show sensitivity functions based on equation (D.15) for a
force observer that is designed for a single harmonic (46Hz) and a multiple
harmonics (46Hz and 69Hz) component.

Large gain is applied at each specific frequency.

Bode Diagram
100

0
Magnitude (dB)

-100

-200
Freq(Hz): 46
Mag (dB): -307
-300

180

135
Phase (deg)

90

45

-45
0 1 2
10 10 10
Frequency (Hz)
(a)
Bode Diagram
100

0
Magnitude (dB)

-100

-200 System: frdy_d

System: frdy_d Frequency (Hz): 69
Frequency (Hz): 46 Magnitude (dB): -310
-300 Magnitude (dB): -346

180
90
0
Phase (deg)

-90
-180
-270
-360
-450
0 1 2
10 10 10
Frequency (Hz)

(b)

Fig. D.3. Y-axis: Sensitivity function for (a) one harmonic (46Hz)
and (b) two harmonics (46Hz and 69Hz)

152
 Appendix D

Next, the transfer function (D.14) and (D.15) are validated numerically using
MATLAB/Simulink based on the control scheme in fig. D.2. Fig. D.4(a)
shows the simulated position error for sinusoidal disturbance signal at 23Hz.
Fig. D.4(b) shows the estimated position error obtained using the sensitivity
function (D.15).

1000

782
position error,
[µm]

-782

-1000
2 2.02 2.04 2.06 2.08 2.1
tim e,[s]

(a)
5
Bode Magnitude Diagram
10

0
10 Freq (Hz): 23
Mag (abs): 783
Position error (abs)

-5
10
[µm]

-10
10

-15
10

-20
10
0 1 2
10 10 10
Frequency (Hz)

(b)
Fig. D.4. Y-axis: (a) Simulated position error and (b) estimated
position error for force observer designed for 46Hz and 69Hz
harmonics

An exact match between the simulated position error and the estimated
position error are observed.. This confirms the transfer functions (D.14) and
(D.15).
153
Appendix D

A reduction in position error at a harmonic frequency, however, inflates the

position errors of neighbouring frequencies in the opposite direction. This
effect is widely known as the waterbed effect [46]. Fig. D.5 illustrates this
waterbed effect.

Bode Magnitude Diagram

0
10
Position error (abs)

-5
10
[µm]

-10
10

-15
10
1 harm onic
2 harm onics
-20
10
1 2
10 10
Frequency (Hz)

Fig. D.5. Waterbed effect observed from additional harmonic

component

The effectiveness of the force observer depends on the accuracy of the

predicted frequencies used in the observer model. Fig. D.6 shows the force
observer performance when a significant mismatch exists between the
designed frequencies and the actual frequencies. The figure shows the effect
of the disturbance force observer on the position error of the system. The
input disturbance force contains harmonics at 23.3Hz and 40Hz. However,
the observer is designed for disturbance harmonics at 23.3Hz and 46.6Hz.

154
 Appendix D

15

10
Force observer is activated
position error [µm]

-5

-10

-15
0 0.2 0.4 0.6 0.8 1 1.2
time [s]

Fig. D.6. Position error for cutting force signal with 23.3Hz and
46.6Hz harmonic components associated with force disturbance
observer with unmatched harmonic (23.3Hz and 40Hz)

Fig. D.7 shows a spectral analysis of the position error.

Without Force Observer With Force Observer
10 10

8 8
position error [µm]

position error [µm]

6 6

4 4

2 2

0 0
0 20 40 60 80 0 20 40 60 80
frequency [Hz] frequency [Hz]

Fig. D.7. Spectral analysis of position error for force observer

with unmatched harmonics

The spectral analysis shows total compensation of the matched harmonic

alone. The mismatched harmonic remained uncompensated. This illustrates
the non-robustness property of the disturbance force observer.

155
Appendix D

Experimental Validation
The disturbance force observer is validated experimentally on the test setup.
Synthesized disturbance forces with 1, 2, and 3 harmonics content are
applied at the input of the system. Fig. D.8 illustrates the control scheme.
The control scheme includes the inverse-model-based disturbance observer
that is added to compensate friction during reference tracking.

Fig. D.8. Schematic diagram of a control scheme with a general state

force disturbance observer and friction compensation

The performance is validated for two different cases, namely, (i) without
reference tracking and (ii) with reference tracking. Fig. D.9 shows for case
(i), the position errors of the y-axis when synthesized disturbance forces with
1, 2, and 3 harmonics are inserted at the input of the system.
1 harmonic 2 harmonics 3 harmonics
[23.25Hz] [23.25Hz 46.6Hz] [23.25Hz 46.6Hz, 69.9Hz]
500 500 500
cutting force [N]
synthesized

200 200 200

0 0 0
-200 -200 -200

-500 -500 -500

0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4

observer activated
position error [µm]

observer activated observer activated

position error [µm]

position error [µm]

20 20 20
10 10 10
0 0 0
-10 -10 -10
-20 -20 -20

0 0.2 0.4 0 0.2 0.4 0 0.2 0.4

time [s] time [s] time [s]

Fig. D.9. Y-axis: Measured position errors for 1, 2, and 3 harmonic

synthesized cutting force.
156
 Appendix D

The above results indicate a complete removal of the cutting forces influence
on the position errors. Fig. D.10 shows the compensation results during
reference tracking.
4 1 harmonic 4 3 harmonics
x 10 x 10
4 4
reference [µm]

reference [µm]
2 2
0 0
-2 -2
-4 -4
0 0.5 1 1.5 0 0.5 1 1.5
time [s] time [s]
position error [µm]

20 20

position error [µm]

no observer

no observer
10 10
0 0
-10 -10
rms =3.1µm rms = 7.12µm
-20 -20
0 0.5 1 1.5 0 0.5 1 1.5
time [s]
position error [µm]

position error [µm]

10 10
with observer
with observer

0 0

-10 -10
rms = 0.66µm rms = 3.90µm
0 0.5 1 1.5 0 0.5 1 1.5
time [s] time [s]

Fig. D.10. Y-axis: Measured tracking errors for synthesized cutting

force with 1 (left) and 3 (right) harmonics.

Fig. D.11 shows the spectral content of the tracking errors.

1 harmonic 3 harmonics
10 10
position error [µm]
position error [µm]

8 8
no observer

no observer

6 6

4 4

2 2

0 0
0 20 40 60 80 100 0 20 40 60 80 100
frequency [hz] frequency [hz]

10 10
position error [µm]

position error [µm]

8 8
with observer

with observer

6 6

4 4

2 2

0 0
0 20 40 60 80 100 0 20 40 60 80 100
frequency [hz] frequency [hz]

Fig. D.11. Spectral analysis of the tracking error for synthesized

cutting force with 1 harmonic (left) and 3 harmonics (right).
157
Appendix D

Tracking error is completely eliminated for disturbance force with one

harmonics content. In case (ii), the harmonics components are removed from
the tracking errors, as indicated by the spectral analysis result in shown in
fig. D.11. However, the lower frequency components of the signal is
amplified. This is shown by the non-zero tracking error in fig. D.11
(indicated by the circle). This amplification is described as a waterbed effect.

Next, circular tests are performed using a synthesized cutting force with 1
harmonic component. Fig. D.12 compares between the reference and the
actual contours, and the radial tracking errors between two cases, (i) without
a disturbance force observer, (ii) with a disturbance force observer. The
actual contours have been amplified by a factor of 200 for graphical display
purposes. The results show that the synthesized cutting force that contributes
to the radial tracking errors is completely compensated by the disturbance
force observer.

During reference tracking, the experimental validations for both the radial
and circular tracking tests are restricted to compensation of cutting force of
lower harmonic components. Compensation of synthesized cutting forces
with 3 or more harmonics components is not possible due to the physical
limitation of the drives. This is explained in fig. D.13.
No Observer (1 harmonic) With Observer ( 1 harmonic)
position, y-axis [mm]
position, y-axis [mm]

20 20

0 0

-20 -20
error: x200 error: x200
-40 -20 0 20 40 -40 -20 0 20 40
position, x-axis [mm] position, x-axis [mm]

15 15

10 10
radial error [µm]

radial error [µm]

5 5

0 0

-5 -5

-10 -10

-15 -15
0 100 200 300 0 100 200 300
angle [degree] angle [degree]

Fig. D.12. Circular test results for synthesized cutting force with 1 harmonic
component for system without (left) and with (right) the force observer.

158
 Appendix D

200

100 no observer with observer

-100
position error [µm]

-200

-300

-400

-500

-600

-700

-800
0 1 2 3 4 5 6 7 8 9 10
time [s]

Fig. D.13. Position errors signal when the disturbance force observer
is activated

This large transient response saturates the motor and result in stability. As a
result of this limitation, a complete cutting force compensation analysis
using synthesized cutting forces with 15 harmonics components is not
possible. This preliminary results, however, indicates the effectiveness of
this method for compensation of cutting forces of known frequency contents.

159