Brake force control and judder

compensation of an automotive
electromechanical brake

Chih Feng Lee

Submitted in total fulfilment of the requirements of the degree of

Doctor of Philosophy

July 2013

Department of Mechanical Engineering

The University of Melbourne
Australia

Produced on archival quality paper

Brake force control and judder compensation of an automotive
electromechanical brake
Chih Feng Lee

Copyright © 2013 by Chih Feng Lee

All rights reserved. No part of this thesis may be reproduced in any form or by any means,
without permission in writing from the author.

Department of Mechanical Engineering

The University of Melbourne
Victoria 3010 Australia

Typeset in LATEX 2ε
This thesis is printed on archival quality paper
Melbourne, 2013
 Abstract

Automotive electromechanical brakes (EMBs) offer several advantages over

their hydraulic counterparts in terms of manufacturability, environmental con-
siderations and performance. The latter is made possible through the EMB
offering accurate continuous brake force tracking and higher closed-loop band-
width. Naturally, full realisation of these benefits depends on the performance
and robustness of the EMB control algorithms. Previous existing EMB control
designs have either been based on cascaded proportional-integral (PI) architec-
ture or designed around a feedback linearised model where performance across
the whole operating range is not addressed.
To support model-based controller developments and post-production qual-
ity control checks, this thesis develops a rapid parameter identification proce-
dure for a control-oriented EMB model. The proposed routine minimises the
time spent for experiments while ensuring an accurate estimation of the param-
eters. Compared to the incumbent approach, the proposed rapid parameter
identification procedure significantly reduces the time required for experiments.
Three model-based clamp force controller designs are then investigated.
First, the time-optimal tracking control that achieves the minimum rise-time
without overshoot is derived from Pontryagin’s minimum principle. The time-
optimal controller is very sensitive to modelling and measurement discrepan-
cies, and can lead to control chatter, which is undesirable from an operational
perspective. However, it represents the quickest possible response without
overshoot and is used as a benchmark for system performance.
To overcome the limitations of the proposed time-optimal control, a robust
near-time-optimal clamp force tracking control (NTOC) is developed. Exper-

iii
iv

imental results show that the rise time of the NTOC is typically within 5%
of the time-optimal control without control chatter. The NTOC also demon-
strates significantly higher bandwidths than the industry-tuned PI controller,
which makes it viable for advanced algorithm deployment.
Furthermore, a holistic control design approach that considers tracking
error, control effort and system constraints is examined under the model pre-
diction control (MPC) framework. This design approach exhibits greater flex-
ibility and slight control performance improvements compared to the NTOC.
However, the intensive computational and large memory demands are pro-
hibitive under current hardware limits.
By utilising the high bandwidth clamp force tracking control of the NTOC,
a novel method for attenuating cold judder directly from the source of vibra-
tion is proposed. Two compensator designs are developed, whereby the EMB is
utilised to generate a compensating clamp force to counteract the judder caus-
ing brake torque variation (BTV). The first approach is the linear parameter-
varying (LPV) BTV compensator, where it is scheduled using the measured
wheel speed and estimated acceleration. Experimental investigations show fa-
vorable results, including a first-order implementation reducing BTV by up to
55%.
To mitigate the limitations arising from relatively low sampling rates, an
alternative approach examines adaptive feedforward compensation, where the
compensator is scheduled using the wheel position, speed and acceleration. Ex-
perimental investigations demonstrate improved BTV attenuation compared
to the LPV compensator, albeit with the additional requirement of wheel po-
sition measurements.
 Declaration
This is to certify that:

1. the thesis comprises only my original work towards the PhD except where
indicated,
2. due acknowledgement has been made in the text to all other material
used,
3. the thesis is less than 100,000 words in length, exclusive of tables, maps,
bibliographies and appendices.

Chih Feng Lee

25 July 2013

v
 Acknowledgements
First and foremost, I would like to express my deep gratitude to my supervisor,
Associate Professor Chris Manzie for accepting me into the group, patient
guidance, enthusiastic encouragement, continuous support and useful critiques
throughout the project.
I would like to thank Professor Malcolm Good, who headed the Melbourne
University node of the Research Centre for Advanced By-Wire Technologies
(RABiT) for providing a pleasant working environment. I wish to acknowledge
the help provided by the engineers at Pacifica Group Technologies, Sam Oliver,
Tony Rocco and Peter Harding for giving technical support and troubleshoot-
ing tips in regards to the control implementation on the electromechanical
brake (EMB), and to the technician of the Advanced Centre for Automotive
Research and Testing (ACART), Don Halpin for his excellent machining skill.
I would also like to thank Chris Line for sharing his parameter identifica-
tion script. The EMB parameter identification script used in Section 3.2 was
based on the work of Chris Line, contributed by Caroline Jekelius (a master ex-
change student from TU Munich working under my supervision in Melbourne
University), and ported to Vector CANape scripting language by myself.
I am indebted to my undergraduate supervisor, Dr Janusz Krodkiewski
for introducing me to the joy of research during my undergraduate summer
research project and final year project. I am also grateful to Professor Dragan
Nešić, Rahul Sharmar, Alan Chang and Sei Zhen Khong for various discussions
that have broaden my understanding in control theory and mathematics in
general. Many thanks to Professor Zhihong Man, Associate Professor Edwin
Tan and Mark Ng for their cordial reception of my visit to Monash University
in Malaysia.
Special thanks should be given to my hyung in the Dynamics and Control
Group, Brett Bishop and Tae Soo (Terence) Kim, and to my contemporaries
in the group, Alireza Mohammadi, Jalil Sharafi, Kuan Waey Lee, Michael
Stephens and Ronny Kutadinata for their assistance and encouragement at

vii
viii

different phases of the project. In particular, I would like to thank Terence for
enjoyable and inspirational soju sessions. I wish to acknowledge the friendship
and encouragement of Chuck Yong Kong, Darwin Lau, Sei Zhen Khong, Thi-
laksiri Bandara, Tien Mun Foong, Xin Fu Tan, Yee Khoon Tee and Yee Wei
Law. I would also like to express my very great appreciation to Ming Hui Siew
for her love, encouragement and support.
Last, but certainly not least, I wish to thank my parents and sister for their
guidance, encouragement and support. My mother and sister have been cook-
ing delicious meals for me and my mechanical engineer father has demonstrated
to me that choosing engineering as a career can both be fun and rewarding.
A Graduate Certificate in Commercialisation has been completed during
the course of this work, and is subsidised under the Federal Government’s
Commercialisation Training Scheme (CTS). The PhD research has been fi-
nancially supported by the Australian Postgraduate Award and Special Stu-
dentships from the Mechanical Engineering Department. Travel is funded by
the Melbourne School of Engineering Research Training Conference Assistance
Scheme.
 Contents

1 Introduction 1
1.1 Brake-by-wire systems . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Brake judder . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Literature review and research aims 11

2.1 Electromechanical brake control . . . . . . . . . . . . . . . . . . 12
2.1.1 Current industry practice . . . . . . . . . . . . . . . . . 12
2.1.2 Model-based control approaches . . . . . . . . . . . . . . 14
2.2 Time-optimal control . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.1 Pontryagin’s minimum principle . . . . . . . . . . . . . . 19
2.2.2 Implementation issues . . . . . . . . . . . . . . . . . . . 22
2.2.3 Approaches to chatter reduction . . . . . . . . . . . . . . 25
2.3 Active vibration attenuation . . . . . . . . . . . . . . . . . . . . 30
2.3.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3.2 Feedback compensation approaches . . . . . . . . . . . . 32
2.3.3 Learning compensation approaches . . . . . . . . . . . . 37
2.3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.4 Literature review conclusions . . . . . . . . . . . . . . . . . . . 40
2.5 Research aims . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.6 Thesis layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3 Experimental setup and model descriptions 45

3.1 Experimental facilities . . . . . . . . . . . . . . . . . . . . . . . 46
3.1.1 Electromechanical brake . . . . . . . . . . . . . . . . . . 46
3.1.2 Bench top test rig . . . . . . . . . . . . . . . . . . . . . . 48
3.1.3 Brake dynamometer . . . . . . . . . . . . . . . . . . . . 49
3.2 Modelling of an electromechanical brake . . . . . . . . . . . . . 50
3.2.1 Overview of high-order models . . . . . . . . . . . . . . . 51
3.2.2 Development of a simulation model . . . . . . . . . . . . 52

ix
x Contents

3.2.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.3 Development of a control-oriented EMB model . . . . . . . . . . 63
3.4 Modelling of the characteristics of brake torque variation . . . . 68
3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

4 Rapid EMB model parameter identification 77

4.1 Design of an optimal trajectory for parameter identification . . . 78
4.1.1 Problem formulation . . . . . . . . . . . . . . . . . . . . 78
4.1.2 Results and discussions . . . . . . . . . . . . . . . . . . . 82
4.2 Parameter estimation . . . . . . . . . . . . . . . . . . . . . . . . 84
4.2.1 Output error method . . . . . . . . . . . . . . . . . . . . 85
4.2.2 Prediction error method . . . . . . . . . . . . . . . . . . 86
4.2.3 Experimental results and discussions . . . . . . . . . . . 87
4.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

5 High performance clamp force tracking control 97

5.1 Time-optimal tracking control . . . . . . . . . . . . . . . . . . . 98
5.1.1 Problem formulation . . . . . . . . . . . . . . . . . . . . 98
5.1.2 Time-optimal control using PMP . . . . . . . . . . . . . 100
5.1.3 Time-optimal feedback controller design . . . . . . . . . 105
5.1.4 Results and discussions . . . . . . . . . . . . . . . . . . . 107
5.2 Robust near-time-optimal tracking control . . . . . . . . . . . . 109
5.2.1 Problem formulation . . . . . . . . . . . . . . . . . . . . 110
5.2.2 Near-time-optimal controller design . . . . . . . . . . . . 112
5.2.3 Stability analysis of the near-time-optimal control . . . . 114
5.2.4 Practical implementation and gain tuning . . . . . . . . 120
5.2.5 Results and discussions . . . . . . . . . . . . . . . . . . . 126
5.3 Explicit model predictive control . . . . . . . . . . . . . . . . . 141
5.3.1 Problem formulation . . . . . . . . . . . . . . . . . . . . 142
5.3.2 Explicit model predictive controller design . . . . . . . . 143
5.3.3 Results and discussions . . . . . . . . . . . . . . . . . . . 145
5.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

6 BTV compensation with wheel speed measurements 155

6.1 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . 156
6.2 Design of a BTV compensator with wheel speed measurements . 162
6.2.1 BTV compensation using output regulation theory . . . 162
6.2.2 Tuning for compensator gains . . . . . . . . . . . . . . . 166
6.2.3 Discretisation of the compensator . . . . . . . . . . . . . 167
6.3 Results and discussions . . . . . . . . . . . . . . . . . . . . . . . 168
6.3.1 Simulation results . . . . . . . . . . . . . . . . . . . . . . 168
6.3.2 Experimental results . . . . . . . . . . . . . . . . . . . . 173
6.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
Contents xi

7 BTV compensation with wheel position measurements 183

7.1 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . 184
7.2 Design of a BTV compensator with wheel position measurements186
7.2.1 BTV compensation using adaptive feedforward cancel-
lation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
7.2.2 Stability and decay of error dynamics . . . . . . . . . . . 188
7.3 Results and discussions . . . . . . . . . . . . . . . . . . . . . . . 190
7.3.1 Simulation results . . . . . . . . . . . . . . . . . . . . . . 190
7.3.2 Experimental results . . . . . . . . . . . . . . . . . . . . 193
7.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

8 Conclusions 201
8.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
8.2 Directions for further work . . . . . . . . . . . . . . . . . . . . . 203

A Periodic signals 205

B Expansion of equation 4.11 209

Bibliography 211
 List of Figures
1.1 Growing pace of the electronics content of vehicles, with ongoing
adoption of by-wire technologies. . . . . . . . . . . . . . . . . . . . 2
1.2 Brake-by-wire system. . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Cutaway view of the PGT electromechanical brake. . . . . . . . . . 5
1.4 Disc runout and disc thickness variation. . . . . . . . . . . . . . . . 7
1.5 Brake judder transfer path. . . . . . . . . . . . . . . . . . . . . . . 8
1.6 Strategies for brake judder attenuation. . . . . . . . . . . . . . . . . 9

2.1 EMB clamp force control with cascaded control structure. . . . . . 13

2.2 EMB clamp force control with cascaded control structure and feed-
back linearisation. . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 EMB clamp force control with linear robust controller or model
predictive control with feedback linearisation. . . . . . . . . . . . . 16
2.4 Phase portrait of time-optimal regulation trajectories of a double
integrator system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.5 Control and state trajectories of the double integrator system. . . . 24
2.6 Phase portrait of regulation trajectories of a double integrator sys-
tem using NTOC. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.7 Phase portrait of regulation trajectories of a double integrator sys-
tem using PTOS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.8 Comparison of control and state trajectories of a double integrator
system using time-optimal control, NTOC and PTOS. . . . . . . . 28
2.9 Systems susceptible to vibration. . . . . . . . . . . . . . . . . . . . 31
2.10 Block diagram of a disturbance filter structure with parallel reali-
sation added on a baseline servo system. . . . . . . . . . . . . . . . 33
2.11 Feedback interconnection structure of generalised plant, controller
and exosystem in an output regulation problem. . . . . . . . . . . . 35
2.12 Plug-in repetitive control system. . . . . . . . . . . . . . . . . . . . 37
2.13 Plug-in adaptive feedforward cancellation scheme. . . . . . . . . . . 39
2.14 Thesis structure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

xii
List of Figures xiii

3.1 PBR T5.1 series Smart Calliper. . . . . . . . . . . . . . . . . . . . . 47

3.2 Cross-sectional view of a PBR T5.1 series Smart Calliper. . . . . . 47
3.3 Controller board of a PBR T5.1 series Smart Calliper. . . . . . . . 48
3.4 Bench top experimental setup. . . . . . . . . . . . . . . . . . . . . . 49
3.5 Brake dynamometer. . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.6 Comparison between the clamp force and motor position measure-
ments with the fitted stiffness curve. . . . . . . . . . . . . . . . . . 54
3.7 Static friction measurements from break-away tests in the air-gap. . 56
3.8 Measurements from break-away tests with applied clamp force. . . . 57
3.9 Example of experimentally obtained friction time series data for
motor velocity reference at 1 rad/s. . . . . . . . . . . . . . . . . . . 58
3.10 Friction torque measurements obtained during constant speed ref-
erence tracking in air-gap. . . . . . . . . . . . . . . . . . . . . . . . 59
3.11 Simulink model of an EMB. . . . . . . . . . . . . . . . . . . . . . 61
3.12 Comparison of experimental measurements and model predictions
for a 10 kN to 20 kN step response. . . . . . . . . . . . . . . . . . . 64
3.13 Approximating the dynamics of the current control loop and the
EMB actuator using a control-oriented model. . . . . . . . . . . . . 64
3.14 Simulated responses of the nonlinear simulation model, the
reduced-order control-oriented model and the discretised model. . . 69
3.15 Setup for disc rotor runout measurements. . . . . . . . . . . . . . . 70
3.16 Runout measurements for inner and outer faces of the brake disc. . 71
3.17 Disc thickness variation measurements and approximations. . . . . 72
3.18 Brake torque variation measurements and approximations. . . . . . 73
3.19 A block diagram illustrating the models that contribute to brake
torque generation. . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

4.1 Optimal trajectories for parameter identification versus various

length of experiments. . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.2 Performance measures of optimal trajectories versus the length of
experiments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.3 A total of 40 sets of measurements obtained using the optimal ex-
perimental design and non-optimal experimental design. . . . . . . 88
4.4 The normalised estimated model parameters obtained using the
output error method. . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.5 The normalised estimated model parameters obtained using the
prediction error method. . . . . . . . . . . . . . . . . . . . . . . . . 90
4.6 Validation of parameterised model using the nominal parameters
from Table 3.1 and the estimated parameters from Tables 4.1 and
4.2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.7 Validation of parameterised models with sinusoidal trajectory. . . . 92
4.8 Kalman estimates of mean and variance for sinusoidal tracking re-
sponse. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
xiv List of Figures

4.9 The normalised estimated model parameters obtained using a non-

optimal experimental design and the output error method. . . . . . 94

5.1 Time-optimal step response without singular interval. . . . . . . . . 103

5.2 Time-optimal tracking response with singular interval. . . . . . . . 104
5.3 Time-optimal switching curves. . . . . . . . . . . . . . . . . . . . . 106
5.4 Overlaid switching curve plotted in the error state plane. . . . . . . 107
5.5 Responses for 1 kN (thin lines) and 5 kN (think lines) steps obtained
using the time-optimal feedback control law. . . . . . . . . . . . . . 108
5.6 Control magnitude in different regions of the error state plane. . . . 114
5.7 Comparison of switching curve with its first order linear approxi-
mation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.8 Controller parameters ssat and ωm sat are plotted against ρ and γ
with ks = 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.9 Comparison of unsaturated control region with different ρ and γ. . 125
5.10 Approximation of the switching curve using a 5th order polynomial. 127
5.11 Block diagram of the simplified NTOC structure. . . . . . . . . . . 127
5.12 Implementation procedure for NTOC of an EMB. . . . . . . . . . . 128
5.13 Comparison of 2 kN setpoint tracking responses with different
NTOC gains. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.14 Rise time and energy usage for a 2 kN step using the time-optimal
control and NTOC with γ = {0.02, 0.03} and ρ ∈ [1, 10] × 10−8 . . . 131
5.15 Gain and phase-shift of the clamp force output relative to the ref-
erence with NTOCs. . . . . . . . . . . . . . . . . . . . . . . . . . . 132
5.16 Gain and phase-shift of the clamp force output relative to the ref-
erence with baseline PI controller. . . . . . . . . . . . . . . . . . . . 133
5.17 Responses of the nominal and the perturbed plant models. . . . . . 135
5.18 Experimental and simulation setpoint tracking responses. . . . . . . 137
5.19 Experimental measurements presented in state plane for 2 kN and
20 kN setpoint tracking. . . . . . . . . . . . . . . . . . . . . . . . . 138
5.20 Rise time for step responses initiated from 0.4 kN. . . . . . . . . . . 138
5.21 Experimentally obtained reference tracking responses using trian-
gular wave with 1 second period and 20 kN magnitude. . . . . . . . 139
5.22 Experimentally obtained reference tracking responses using sine
wave with 1 second period, 1 kN mean and 0.5 kN amplitude. . . . 140
5.23 Experimentally obtained energy usage for step responses initiated
from 0 N clamp force. . . . . . . . . . . . . . . . . . . . . . . . . . 141
5.24 Motor torque deviation map with fixed reference at 13 kN. . . . . . 144
5.25 Comparison of simulated setpoint tracking performance between
MPCs with different prediction and control horizons for 2 kN and
20 kN steps. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
5.26 Rise time and energy usage for a 2 kN step using the time-optimal
control and MPC. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
List of Figures xv

5.27 Comparison of rise time and energy usage of the time-optimal con-
trol, NTOC and MPC. . . . . . . . . . . . . . . . . . . . . . . . . . 149
5.28 Responses of 20 kN step using the time-optimal control, NTOC and
MPC. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
5.29 EMB closed-loop system and clamp force controllers developed in
this chapter. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

6.1 Block diagram of a brake torque variation compensator scheduled

by wheel speed and acceleration. . . . . . . . . . . . . . . . . . . . 156
6.2 Brake torque variation profile. . . . . . . . . . . . . . . . . . . . . . 169
6.3 Simulated responses for BTV compensators with different orders of
harmonic. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
6.4 Simulated brake torque and clamp force trajectories for different
Q/R ratios. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
6.5 Simulated brake torque and clamp force trajectories for different
Dw magnitudes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
6.6 Simulated brake torque and clamp force trajectories for differ-
ent ν magnitudes, where the process noise is defined as Bw =
[0, 1, 1, 0, ν, 0, ν]T . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
6.7 Comparison of discretisation methods and sampling period for a
second-order compensator on different wheel speeds. . . . . . . . . . 174
6.8 Comparison of discretisation methods and sampling period for a
first-order compensator on different wheel speeds. . . . . . . . . . . 175
6.9 Experimentally obtained brake torque variation compensation us-
ing the first-order compensator discretised with forward-Euler
method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
6.10 Amplitude spectrum of brake torque trajectories for different wheel
speeds. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
6.11 Experimentally obtained brake torque responses with and without
compensation under varying wheel speed. . . . . . . . . . . . . . . . 179
6.12 Experimentally obtained brake torque responses with and without
compensation under varying wheel speed and brake torque reference.180

7.1 Block diagram of a brake torque variation compensator scheduled

by wheel position, speed and acceleration. . . . . . . . . . . . . . . 184
7.2 The lower bound of the decay rate, α versus the compensator gain, g.191
7.3 Simulated responses of BTV compensation for compensators with
different gains. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
7.4 Simulated responses of BTV compensation for compensators with
different orders of harmonic. . . . . . . . . . . . . . . . . . . . . . . 193
7.5 Simulated responses of continuous-time and discretised compensators.194
7.6 Experimentally obtained BTV compensation for compensators with
different orders of harmonic. The brake torque reference is fixed at
150 Nm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
xvi List of Figures

7.7 Amplitude spectrum of brake torque trajectories for different wheel

speeds. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
7.8 Experimentally obtained brake torque responses with and without
compensation under varying wheel speed. . . . . . . . . . . . . . . . 199

A.1 Generator of periodic signals. . . . . . . . . . . . . . . . . . . . . . 206

 List of Tables
3.1 EMB model parameters obtained using Procedure 3.1. . . . . . . . 63

4.1 Mean and variance of the normalised estimated parameters ob-

tained using the output error method (OEM) and the prediction
error method (PEM) with the optimal experimental design. . . . . 89
4.2 Mean and variance of the normalised estimated parameters ob-
tained using the output error method with the non-optimal ex-
perimental design. . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.3 Optimised EMB model parameters. . . . . . . . . . . . . . . . . . . 95

5.1 Control laws for regions in Figure 5.6. . . . . . . . . . . . . . . . . 115

5.2 Tracking performance %nIAE of the baseline PI and NTOC. . . . . 140
5.3 Time required for simulating a 0.2 seconds trajectory with a 13 kN
setpoint. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

6.1 Compensation performance and power usage of compensators with

different orders of harmonic. . . . . . . . . . . . . . . . . . . . . . . 170
6.2 Normalised RMSE of compensated responses. . . . . . . . . . . . . 178

7.1 Compensation performance of compensators with different orders

of harmonic. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

xvii
 Notations

Acronyms
ABS Anti-lock braking system
BBW Brake-by-wire
BPV Brake pressure variation
BTV Brake torque variation
CAN Controller area network
EHB Electrohydraulic brake
EKF Extended Kalman filter
EMB Electromechanical brake
EMF Electromotive force
ESC Electronic Stability Control
EWB Electronic wedge brake
FFT Fast Fourier transform
LMI Linear matrix inequality
LPV Linear parameter-varying
LTV Linear time-varying
MIMO Multiple-input multiple-output
MPC Model predictive control
NTOC Near-time-optimal control
PMP Pontryagin’s minimum principle
PMSM Permanent Magnet Synchronous Motor
RMSE Root Mean Square Error
SISO Single-input single-output

xix
xx Notations

Functions

Diagonal matrices:

0
 
A1
diag(A1 , . . . , An ) = 
 ... 
 (1)
 
0 An

Sign function:

1,
if x > 0,



sgn(x) = 0, if x = 0, (2)

−1, if x < 0,



Unit saturation function:


x, if |x| ≤ 1,
sat(x) = (3)
sgn(x), if |x| > 1.

Symbols
θd Disc rotor angular position
θm EMB motor angular position
θm,r EMB motor angular position reference
ωm EMB motor angular velocity
ωm,r EMB motor angular velocity reference
ωd Disc rotor angular speed
ωs Characteristic velocity of the Stribeck friction
µ Coefficient of friction between brake pads and rotor
µs Coefficient of load-dependent static friction

H Hamiltonian
L Lagrangian

Fcl Clamp force

Fcl,r Clamp force reference
iq EMB motor quadrature current
Kt EMB motor torque constant
N Gear ratio
rd Effective disc rotor radius
Tb Brake torque
Tb,r Brake torque reference
Tbtv Brake torque variation
Notations xxi

TC Coulomb friction torque

Te Net external non-friction torque
Tf Friction torque
Tl Load torque
Tm EMB motor torque
Ts Load-dependent static friction torque
Ts0 Load-independent static friction torque
Tv Viscous friction coefficient
x State vector
xr State vector reference

*Symbols that are not listed here are defined close to its use in the body of the thesis.
 Publications
The material presented in this thesis is based on a number of refereed papers,
which have appeared during the course of my PhD research. In particular,
Chapter 4 is based on:

1. Lee, C. F. and Manzie, C. (2013), Rapid Parameter Identification for

an Electromechanical Brake, in ‘Proceedings of the Australian Control
Conference’, pp. 391–396.

Chapter 5 is based on:

2. Lee, C. F., Manzie, C. and Line, C. (2008), Explicit nonlinear MPC of an

automotive electromechanical brake, in ‘Proceedings of the 17th IFAC
World Congress’, pp. 10758–10763.
3. Lee, C. F. and Manzie, C. (2012), ‘Near-time-optimal tracking controller
design for an automotive electromechanical brake’, Proceedings of the In-
stitution of Mechanical Engineers, Part I: Journal of Systems and Con-
trol Engineering 226(4), pp. 537–549.
4. Lee, C. F. and Manzie, C. (2012), ‘High-Bandwidth Clamp Force Control
for an Electromechanical Brake’, SAE International Journal of Passen-
ger Cars - Electronic and Electrical Systems 5(2), pp. 590–599. SAE
Paper 2012-01-1799.

Chapter 7 is based on:

5. Lee, C. F. and Manzie, C. (2012), ‘Adaptive Brake Torque Variation

Compensation for an Electromechanical Brake’, SAE International Jour-
nal of Passenger Cars - Electronic and Electrical Systems 5(2), pp. 600–
606. SAE Paper 2012-01-1840.

Several papers are in preparation at the time of writing. Additionally, a pro-

visional patent has been filed as a result of this research:

xxiii
xxiv Publications

6. Lee, C. F. and Manzie, C. (2012), ‘Automotive brake judder compensa-

tion using electromechanical brakes’, Australian Provisional Application
no. 2012904156, filed on 24 September 2012.
 1

Introduction

erospace and automotive manufacturers are constantly chal-

A lenged by both regulatory bodies and consumers to produce aircrafts and
automobiles that are more reliable and perform better at lower cost. One ap-
proach to addressing these conflicting objectives involves replacing mechanical
and hydro-mechanical control systems by computer-controlled systems. This
has led to purely electrically-signalled control systems, designated as by-wire
systems.
The fly-by-wire system was first introduced to the aerospace industry in the
1950’s (Whitcomb, 2004), and is now a standard feature in commercial pas-
senger airplanes (Briere and Traverse, 1993; Bartley, 2000). Instead of using
mechanical or hydraulic linkages, a fly-by-wire system sends a control sig-
nal from the computer to the actuators for the ailerons, flaps, and other flight
control surfaces through electric wiring. Additional functionalities, such as au-
tomatic stability systems and computerised navigation can then be added or
improved by amending the flight control computer. The brake-by-wire system
(BBW) has also been recently introduced (in 2011) to the Boeing 787 air-
liner, replacing the traditional hydraulic brakes with electromechanical brakes
(EMBs). The elimination of brake fluid and modularity in BBW systems has
shortened the assembly time, improved diagnostics capability and increased
dispatch reliability (Danielson, 2005).
The benefits of by-wire technologies in aerospace have proven to be sub-
stantial, and has encouraged its adoption in the automotive industry. The
uptake of by-wire systems in the automotive industry is a natural extension of

1
2 Chapter 1. Introduction
Electrical/Electronics Content of Vehicles (%)

• Keyless vehicle
30 • Occupant sensing
• Rollover sensing
• Remote keyless entry
• Integrated powertrain control
25 • Electronic instrument cluster • Multiplexing
• Traction control • Diagnostics • Throttle-by-wire
• Adaptive suspension • Navigation systems • Automated parking
20 • Crash sensor diagnostics
• Theft deterrent systems • Adaptive cruise control
• Stability control
• Cruise control
15 • Side impact sensor
• Electronic fuel injection
• Tyre pressure sensor
• Electronic transmission control
• Single point crash sensor
• Lamp
10 • Horn • Breakerless ignition • Vehicle immobiliser
• Alternator • Electrical power steering
• Radio
• Generator • Trip computer
5 • Starter • Service indicator
• Antilock braking
• Cellular phone

0
1950 1960 1970 1980 1990 2000 2010
Year

Figure 1.1: Growing pace of the electronics content of vehicles, with ongoing
adoption of by-wire technologies (adapted from Winters et al. (2004)).

the trend observed in the last few decades, whereby a steady increase of electri-
cal, electronics and mechatronics content in vehicles, augmenting or replacing
the mechanical components that had been traditionally dominating the share
of vehicle parts (Figure 1.1). The major adoption of advanced vehicular elec-
tronics began in the 1970’s in order to comply with the Clean Air Act in the
USA (passed in 1963 and amended in 1970). This motivated the use oxygen of
sensors and powertrain control technologies to reduce emission pollution and to
mitigate the smog problem (Klier and Rubenstein, 2011). More recently with
the introduction of more stringent emission standards, increasing demands of
fuel efficiency and high engine performance, the throttle-by-wire system has
became a standard feature in many vehicles (Bretz, 2001).
To fully realise the benefits of drive-by-wire requires the integration of three
primary vehicle control systems, namely throttle, steering, and braking. Steer-
ing system technology has evolved from hydraulic to electric power-assisted
steering, whereby assisting torque is supplied by an electric motor to reduce
the effort required to turn the wheels. Modern advancements including the
speed sensitive steering and automated parallel parking are also available in
many production vehicles. The full steer-by-wire provides the greatest design
1.1. Brake-by-wire systems 3

flexibility that facilitates additional functionalities, and was demonstrated in

concept vehicles such as GM Sequel.
In terms of brake refinement, the debut of anti-lock braking system (ABS)
in 1978 initiated the intelligent control of braking system, leading to the volume
production of electronic stability control (ESC) in 1995 (Bosch, 2004). The
ESC is highly effective in maintaining the control of the vehicle by detecting
and minimising skids, thus reduces crashes. A study carried out by the USA
National Highway Traffic Safety Administration (NHTSA) showed that ESC is
highly effective in reducing single-vehicle crashes, where passenger car crashes
were reduced by 35 percent, and Sport Utility Vehicles crashes by 67 percent
(Dang, 2007). These safety benefits have promoted lawmakers to implement
legislations to mandate the adoption of this technology in new vehicles. Among
many countries, Australia requires all new vehicles to be equipped with ESC
from 2011, the USA from 2012 and the European Union from 2014. This
demonstrates the market demands and legislative requirements for a safer and
more intelligent braking system, and motivates the development of a complete
BBW system.

1.1 Brake-by-wire systems

The automotive BBW system consists of brake callipers, brake pedal feel emu-
lator, central supervisory control unit, fault-tolerant power supply and sensors
(yaw rate, lateral acceleration and wheel speed) (Figure 1.2). There are two
generations of BBW actuators, namely the electro-hydraulic brake (EHB) and
the EMB. The EHB retained many of the conventional hydraulic braking com-
ponents while offering by-wire capabilities by equipping with the brake pedal
feel emulator. It served as a stepping stone towards the electromechanical
braking system (Reuter et al., 2003).
There are currently two distinctive EMB designs, the first one is the con-
ventional design without self-energisation (Taig et al., 1988) and the other
is the Electronic Wedge Brake (EWB) with self-energisation (Dietrich et al.,
2001; Roberts et al., 2003; Baumann et al., 2011). The EWB requires less en-
ergy input from the electric motor to generate clamp force during braking by
utilising the wedge shaped mechanism to transfer a force vector perpendicular
to the pads from the actual braking drag force. This leads to smaller require-
4 Chapter 1. Introduction

1 EMB wheel modules, 1 2 1

2 Brake pedal feel emulator,
3 Wheel speed sensor,
4 Central control unit,
5 Yaw rate and lateral Power Supply
3 3
acceleration sensors,
6 Parking brake switch.

PBS CCU

3 3

6 5 4

Figure 1.2: Brake-by-wire system.

ments for motor and power sizings. However, the amount of actuator force
that can be exerted by the motor without self-energising may not be sufficient
to use as a park brake. On the other hand, the conventional EMB is the more
popular design, whereby many automotive Original Equipment Manufactur-
ers (OEMs), including Bosch (Schumann, 2001), Continental (Halasy-Wimmer
and Völkel, 2005), Delphi (Siler and Drennen, 2002) and Hitachi (Watanabe
et al., 2007) have their respective models. This thesis employed the EMB de-
signed and manufactured by Pacifica Group Technologies (PGT) (Wang et al.,
2005). It comprised an electric motor, a gear reduction stage, a ball screw, a
floating calliper and control electronics (Figure 1.3). Its main features include
continuous brake force actuation and near elimination of the residual drag1 .
Compared to the hydraulic braking systems, the absence of brake booster,
vacuum pump and hydraulic lines in the EMB-based BBW systems leads to
the reduction in component number, installation space, assembly time and
maintenance costs. Environmental benefits are achieved by avoiding brake
fluid usage, which is corrosive and toxic.
Furthermore, the pedal no longer vibrates during ABS operation, thus
lessening the risk of an inexperienced driver erroneously reducing the brake
pressure and prematurely terminating the ABS control mode. Other safety
features include redundant wiring, extra batteries to offset current peaks and
lessen the risk of power failure under a defective generator2 , and fault-tolerant
1
Brake pads contacting with the disc during off-brake running.
2
In line with the regulatory requirements to split the hydraulic brake systems into two
1.2. Brake judder 5

1 Calliper, 1 2 3 4 5 6
2 Ball screw shaft,
3 Motor rotor,
4 Motor stator,
5 Planetary gears,
6 Control board.

Figure 1.3: Cutaway view of the PGT electromechanical brake.

time-triggered communication protocol.

The flexibility inherent from the computer-controlled architecture facili-
tates the implementation of ABS, ESC, traction control, acceleration slip reg-
ulation, and also opens up further opportunities for additional functions such
as diagnostic, error reporting, adaptive cruise control and judder compensa-
tion. Additionally, the pedal feel and branding-specific vehicle dynamics can
be fine-tuned through software updates rather than changing the hardware
on a vehicle, therefore potentially reducing production costs and development
time (SAE-Australasia, 2007). Inspection agency TÜV Rheinland found that
the ergonomically designed brake pedal feel emulator on an EMB-enabled ve-
hicle is capable of reducing the stopping distance by almost 20 percent (Lang,
2008). This indicates the potential additional performance capability of EMB
over hydraulics, and allows other braking-related issues to be investigated.

1.2 Brake judder

One major problem in automotive braking system is brake judder, which has
attracted growing interest from the automotive industry. In a typical occur-
rence of brake judder, vibration is perceived by the driver at the brake pedal,
vehicle floor and steering wheel during light braking (approximately 0.2g) at
separate circuits.
6 Chapter 1. Introduction

mid to high speed (30 to 140 km/h) (Jacobsson, 2001; Lee and Brooks, 2003).
Brake judder is defined as a braking-induced forced vibration occurring at the
steering wheel, brake pedal or vehicle body, where the frequency of vibration
is proportional to the wheel speed3 (Jacobsson, 2001). It has resulted in major
recalls, warranty claims (Nissan North America, 2005) and detrimental effects
to brand image. Furthermore, the associated vibrations also induce additional
wear and prompt components fracture, hence resulting in shorter service life
and higher operating costs.
Brake judder is caused by the irregularities in the brake disc, and can be
categorised into cold and hot (or thermal) judder (Abdelhamid, 1997). Cold
judder is induced by geometrical irregularities due to mounting, machining,
uneven wear, uneven corrosion and uneven friction film generation (Jacobs-
son, 2003a). The irregularities associated with cold judder occur fewer than
five times per disc revolution, and correspond to vibration frequencies which
are lower than five times of the wheel rotational speed (Jacobsson, 2001). Geo-
metrically, these irregularities resemble disc runout or disc thickness variation
(DTV) (Figure 1.4). Disc runout may induce uneven wear and further gen-
erates DTV (Haigh et al., 1993a), which contributes to the brake pressure
variation (BPV) and brake torque variation (BTV). For most manufacturers,
the manufacturing tolerances for disc runout and DTV are less than 80 µm
and 10 µm respectively (de Varies and Wagner, 1992). However, a slight in-
crease in DTV to 15 µm can cause brake judder in vehicles that are sensitive to
judder (Jacobsson, 2003a), therefore raising the need for better manufacturing
tolerances.
Hot judder is induced by thermal deformation (e.g. coning and waving),
uneven thermal expansion, or phase transformation of the disc material (Jacob-
sson, 2003a). Unlike cold judder, hot judder does not require large amplitude
permanent disc irregularities for its initialisation. When exposed to braking
pressure, permanent irregularities (both with low and high amplitudes) cause
a localised rise in contact pressure field and temperature, leading to hot spots
generation (Kubota et al., 1998; Little et al., 1998; Jacobsson, 2001; Sury-
atama et al., 2001). The hot spots resulted in instantaneous DTV and spatial
friction variation, causing brake pressure and brake torque pulsations during

3
It is usual to relate judder frequencies to wheel speed, e.g. frequency at twice the
number of wheel speed is called second order judder.
1.2. Brake judder 7

1 Brake pad, 1 2
2 Brake disc,
3 Wear due to pad rubbing. Disc Disc
2 Rotation Rotation

3 Rotation
Axis

1
3

Disc Runout Disc Thickness Variation

Figure 1.4: Disc runout and disc thickness variation.

braking. Additionally, large thermal gradients may eventuate disc cracking,

warping and buckling (Little et al., 1998; Fieldhouse and Beveridge, 2001),
further deteriorate the disc condition. Compared to cold judder, thermally
excited judder is of higher order (with respect to the wheel speed), typically
between six to twenty times the wheel rotational speed (Kao et al., 2000).
Both hot and cold judder lead to brake pressure variation (BPV) and brake
torque variation (BTV), which are the primary excitation mechanisms for jud-
der vibrations (Figure 1.5). The generated BTV are further transmitted to
the pedal via hydraulic systems, resulting in pedal vibration. Furthermore,
the vibration excited by the BTV is transmitted to the vehicle chassis, vehicle
body and the steering system through the wheel suspension system, causing
noise and vibrations in the vehicle compartment and steering wheel. The vi-
bration is also amplified when its frequency is within the resonance of the
suspension and steering systems (Stringham et al., 1993; Kim et al., 1996).
In particular, the front wheel suspensions of most passenger cars have fore-aft
vibration mode resonance in the range of 6 Hz to 20 Hz. This frequency range
corresponds to a maximal amplitude of vibrations at a critical vehicle speed
of between 60 km/h to 140 km/h for the first-order judder, and 30 km/h to
70 km/h for the second-order judder (Jacobsson, 2003b).
Gassmann and Engel (1993) investigated the excitation and transfer mech-
anism of brake judder experimentally in street-going vehicles, and determined
the interactions between excitation and transfer mechanism. Based on this
result, a holistic approach to brake judder was suggested in Engel et al. (1994)
by considering the total system, including the brakes, the suspension and the
steering system. Brake judder attenuation strategies can be categorised into
8 Chapter 1. Introduction

Excitation
Brake disc irregularities

BPV BTV

Transfer
Hydraulic system Wheel suspension & Amplification

Amplification
Vibration Car body Steering system & Perception

Noise Vibration Vibration

Figure 1.5: Brake judder transfer path (Jacobsson, 2003a).

minimisation of the judder excitation and modification of the transfer path, as

shown in Figure 1.6.
Minimisation of the judder excitation can be achieved through BPV and
BTV reduction, attained by reducing disc runout, DTV and thermal effects.
Initial disc runout and DTV can be improved by adopting better assembly and
machining tolerances, albeit resulting in higher manufacturing costs. Subse-
quent disc resurfacing can be performed if the DTV level exceeds the disc
specifications4 due to wear. However, this leads to higher operating costs.
Furthermore, in order to minimise DTV generation, residual drag can be re-
duced by increasing clearance, optimising sliding pin and piston seal rollback
characteristics of the calliper (Doi et al., 2000; Tamasho et al., 2000), which
may adversely affect pedal feel characteristics. Jacobsson (1997) analytically
showed that a heavier calliper reduced judder, but this creates conflicting de-
sign requirements for improved assembly weight and fuel economy. Other
methods include: reducing thermal deformation of the disc through improved
air flow and heat transfer characteristics, reducing uneven disc wear and re-
ducing nonuniform deposition of transfer film on the disc (Haigh et al., 1993b;
Doi et al., 2000). Decreasing pad stiffness was also suggested in (Jacobsson,
2003b; Leslie, 2004; Duan and Singh, 2011), although leading to the increase
of pedal travel, deteriorating ABS performance and decreasing pad wear life.
Lower coefficient of friction was suggested also by Lee and Dinwiddie (1998),
4
Most manufacturers require the disc runout and DTV to be less than 80 µm and 10 µm
respectively (de Varies and Wagner, 1992).
1.2. Brake judder 9

Passive Approaches Active Approaches

• Shift resonance • Compensation using

frequency active steering

• Shift resonance • Shift resonance

frequency frequency using
active suspension

• Reduced residual • Compensation using

drag EMB (investigated in
• Increased weight this thesis)
• Stiffer pads

• Improved
manufacturing
tolerance
• Resurface disc
• Improved heat
transfer

Figure 1.6: Strategies for brake judder attenuation.

albeit compromising braking effectiveness.

The judder transfer path can be optimised by shifting the resonance fre-
quencies of suspension and steering systems outside of the first wheel order
excitation (Engel et al., 1994). In particular, Zhang et al. (2007) reported that
the suspension bush bearings characteristics greatly affect the resonance am-
plitude. Judder performance can be improved by increasing the rubber bush
bearing stiffness (Duan and Singh, 2011) and adopting a hydraulic bush bear-
ing with higher damping (Meyer, 2005). Furthermore, Gruber et al. (2002)
suggested using an active wheel suspension to shift the resonance frequency
online to dampen judder, while von Groll et al. (2006) proposed using an active
steering system to compensate for judder.
Apart from engaging in lengthy testing procedures during design iterations
and mandating high manufacturing tolerance (which is expensive monetarily
and time-wise), the aforementioned resolutions are either corrective measures
(as it is executed after the driver has had the unpleasant driving experience
caused by judder), or are locally isolating the vibration (as the source of vibra-
tion is still exists). Ideally, judder should be eradicated at the source, which is
the brake unit. The brake unit should be insensitive to DTV, able to eradicate
judder before occurrence, and to provide prompting signal for brake service.
In this regard, there is much potential in the development of an intelligent
braking system.
The utilisation of a BBW system for brake judder attenuation is yet to be
10 Chapter 1. Introduction

explored. While the brake pedal feel emulator equipped in BBW system is
capable of eliminating pedal pulsation due to BPV, the resultant BTV that
causes judder is not dealt with. It is still uncertain whether the additional
bandwidth offered by EMBs can be used to actively attenuate the BTV. This
would reduce the need for a better manufacturing tolerance and extend service
life of components.
 2

Literature review and

research aims

As motivated in Introduction, the overall goal of this thesis is to utilise an

electromechanical brake (EMB) to compensate the brake torque variation that
induces brake judder. In order to achieve this objective, a number of research
areas anticipated to contribute to the research topic are reviewed. The review
of these areas has been broken down into three main sections. First, exist-
ing clamp force control techniques used in EMB are examined including the
limitations of such approaches. The next section presents an overview of time-
optimal control theory and the design adjustment for practical implementation.
The motivation for this section of the review has two parts, firstly to obtain
a model-based controller design that provides the best possible performance
given the input and state constraints on the system, and secondly to develop
an implementable high gain controller that achieves high closed-loop band-
width for reference tracking. In the third section, various vibration control
approaches are reviewed for their potential to solve the brake torque variation
problem, and their suitability for implementation on an EMB. Subsequently,
the review identifies the gap in the literature for active brake torque variation
compensation and the research aims are stated. Finally, the remaining thesis
structure is laid out.

11
12 Chapter 2. Literature review and research aims

2.1 Electromechanical brake control

As discussed in Chapter 1, compared with traditional hydraulic alternatives,
electromechanical brakes (EMBs) offer reductions in components, space re-
quirements, environmental impact, and increased design flexibility through
software upgrades. Clamping force generation in EMBs is facilitated by an
electric motor, therefore potentially offer finer clamp force control actuation
with high closed-loop bandwidth relative to hydraulic systems. This capa-
bility is beneficial for maintaining high performance anti-lock braking (ABS)
and electronic stability control (ESC). However, the clamp force control per-
formance is dependent on the adopted control architecture and the selected
gains. Additionally, the EMB actuator comprises electric motor, gear reduc-
tion state, ball screw and floating calliper with brake pads. These components
experience nonlinear friction during motion and possess nonlinear stiffness,
therefore inducing a nonlinear response and complicating the controller de-
sign. Furthermore, the calliper has operational limit on the maximum clamp
force while the electric motor has constraints on torque and velocity. These
limitations have to be taken into consideration during controller design and
implementation. Many control architectures have been suggested to address
(or partially address) the nonlinear response and the actuator saturation, and
they will be reviewed in this section.

2.1.1 Current industry practice

The main role of the EMB controller is to track the clamp force reference
given by higher level controllers such as ABS and ESC. Founded on a standard
motion control architecture, the most popular control structure adopted in the
industry for clamp force control is the cascaded proportional-integral-derivative
(PID) control, as illustrated in Figure 2.1.
A typical cascaded control structure used in EMB consists of three-loops,
where the nested loops follow the sequence of force, speed and current. The
outmost loop contains the force controller that accepts the clamp force ref-
erence and compares with the measured clamp force. Using the difference
between the reference and the measured value, motor speed reference is then
calculated and fed to the speed controller. The hierarchical structure is then
continued with the speed control loop, where the same idea is again extended
2.1. Electromechanical brake control 13

Fcl
Fcl,r Force ωm,r Speed iq,r Current EMB
+ Controller + Controller + Controller Actuator
− − −
Fcl ωm iq

Figure 2.1: EMB clamp force control with cascaded control structure (Schwarz
et al., 1998).

with the speed controller generating the current reference. The innermost loop
contains the current controller, where the motor voltage is calculated and ap-
plied to the motor via a Pulse Width Modulation (PWM) driven inverter. The
purpose of the current controller is to deal with the dynamics of the power sup-
ply and the armature, and to cancel the effect of the induced armature voltage
or Back Electromotive Force (back EMF). Since the motor torque is propor-
tional to the current, the current loop can also be called the torque loop. Note
that this multi-loop control structure requires the bandwidth of the control
increases towards the inner loops with the current loop being the fastest and
the force loop the slowest. This is to ensure reference generated by the outer
loop can be handled well by the inner loop (Leonhard, 2001).
To address the motor speed limit, Maron et al. (1997) design the force con-
troller using a proportional controller with output saturation, and the velocity
controller is implemented using a PI controller. While the presence of current
controller is inevitable in implementation, it is omitted in the model presented
this paper, and the measured motor current is assumed to be identical to the
reference. This assumption is based on the fact that the current loop has a
relatively fast dynamics compared to the dynamics of the mechanical compo-
nents.
Similar work was done by Schwarz et al. (1998), albeit with more details on
the current controller design. Likewise, the force controller is a proportional
controller implemented on a microcontroller. However, the speed and current
controllers are both PI controllers with series-connected first-order lag elements
implemented using analog circuits. The task of the first-order lag elements
are to attenuate measurement noise. Additionally, the motor current limit is
addressed in this work by the inclusion of an output saturation component in
the speed controller.
Note that compared to the standard motion control problem, the EMB
clamp force control is differentiated by a large operating range in which non-
14 Chapter 2. Literature review and research aims

linear friction and nonlinear stiffness become significantly high at large clamp
force (of up to 30 kN). In order to improve the dynamic response, Line et al.
(2004) suggest augmenting a static friction compensation to the current refer-
ence. The compensating current reference introduces extra motor current in
order to break away the stiction. This improvement reduces the tracking error
when the motor velocity switches sign and crosses zero.
In a follow up paper, Line et al. (2006) show that the dynamic response can
be further improved by applying feedback linearisation techniques. Given the
block diagram of the implementation shown in Figure 2.2, compensating cur-
rent reference is augmented to cancel the nonlinear response due to Coulomb
friction as well as nonlinear stiffness. This technique transforms the nonlinear
EMB actuator system into an equivalent linear system (as seen by the cas-
caded PID controller), hence allows for to a more intuitive gain tuning. The
use of gain scheduling for the force and velocity controllers is also suggested
to enhance tracking performance over the entire working envelop of the EMB.
Nevertheless, scheduling the entire set of gains in real-time increases memory
and processing requirements. By taking a compromising step, Jo et al. (2010)
tackles this impediment by scheduling only the proportional gain in the force
controller, and the derivative and integral gains remain fixed.
Despite their simplicity, the cascaded PID control structure has several lim-
itations associated with high performance clamp force tracking of EMB. The
key limitations of cascaded PID controllers associated with this research are:
control in the presence of nonlinear friction and nonlinear stiffness, maintaining
performance against actuator saturations, and tuning for optimal performance.
Although techniques to include auxiliary mechanisms for controller integration
have been developed, they appear to be ad-hoc additions to the standard cas-
caded PID controller and the optimal performance is not guaranteed. Utilising
a suitable model of the plant that is to be controlled can assist with the design
of a high performance controller, and this approach will be reviewed in the
following subsection.

2.1.2 Model-based control approaches

The knowledge of system characteristics of the plant to be controlled can be
employed to find a suitable model that sufficiently represents the plant. As a
mathematical depiction of the real plant, the plant model is instrumental in
2.1. Electromechanical brake control 15

iq,f Feedback
Linearisation

+ Fcl
Fcl,r Force ωm,r Speed iq,r Current EMB
+ Controller + Controller + Controller Actuator
− − −
Fcl ωm iq

Figure 2.2: EMB clamp force control with cascaded control structure and
feedback linearisation (Line et al., 2006).

designing a satisfactory control structure and tuning the gains. The control
design paradigm that utilises a model is called model-based control design.
Holweg et al. (2000) make use of the plant model to develop an inverse-
model controller, where nonlinear terms can be dealt with over the whole
operating region. However, since the controller is the inverse of the model,
the controller only performs well if the model is accurate. Generally, there is
no model that can exactly represent the plant and modelling discrepancy is
inevitable. Furthermore, the plant dynamics are likely to change with time due
to factors such as wear of mechanical components and thermal effects. When
inverse models are used in controller synthesis, even the slightest amount of
modelling inaccuracy can yield significant steady state error. In other words,
an inverse-model controller is not robust against modelling error.
To increase the robustness of the controller against modelling error, Krish-
namurthy et al. (2005, 2009) propose the application of backstepping design
method in EMB control. While the controller is analytically shown to be robust
to bounded parametric uncertainties, the design method lacks a systematic
routine for picking the controller gains to achieve good tracking performance.
Another approach to handle modelling error is offered by Line et al. (2007),
in which the parametric uncertainty is dealt with using linear robust control.
Figure 2.3 displays the block diagram of the proposed control scheme. Com-
pared to the cascaded PID control structure, it can be noted that the linear
robust controller has now taken over the role of force and speed control loops.
The controller synthesis is first started with the implementation of feedback
linearisation in order to transforms the nonlinear plant model into an equiv-
alent linear model. This is followed by gain selection for the linear robust
controller. Note that a systematic gain tuning procedure is available based on
the H∞ control theory (Zhou and Doyle, 1998). Since the controller is synthe-
16 Chapter 2. Literature review and research aims

iq,f Feedback
Linearisation

+ Fcl
Fcl,r Linear Robust iq,r
Controller Current EMB
+ Controller Actuator
(or MPC)
−
iq
Fcl ωm

Figure 2.3: EMB clamp force control with linear robust controller or model
predictive control with feedback linearisation (Line et al., 2006, 2007).

sised upon the feedback linearised model, the set of gains is only optimal with
respect to the linear model. In other words, the optimal performance cannot
be guaranteed for the original plant model. Additionally, the input and state
saturations are not explicitly considered during the synthesis. Based on simu-
lation and experimental results, the linear robust control is found to be overly
conservative, where the tracking performance is sluggish (Line, 2007).
Notice that backstepping control and robust linear control contain fixed sets
of controller parameters which are synthesised to be robustly stabilising for a
bounded set of parametric uncertainty. Kwak (2005) takes the idea of robust
control further, where the robust controller is augmented with an auxiliary
mechanism to adaptively tune the controller parameters online. While this
presents an interesting concept, the controller structure is complex and requires
differentiation of measured values — all of these may lead to difficulties in
implementation1 .
To address the sluggish performance of the linear robust control, robustness
is compromised by replacing it with a linear unconstrained model predictive
control (MPC) (Line et al., 2008). MPC employs a model of the plant to
predict how the system will respond to a set of control inputs, and then op-
timises these inputs so that the predicted output closely coincides with the
reference. A notable advantage of the linear unconstrained MPC is that a
systematic synthesis procedure can be applied offine to optimally choose the
controller parameters (Maciejowski, 2002). Furthermore, this control struc-
ture resembles that of a state feedback controller with simple implementation.
However, similar to the linear robust controller presented previously, the in-
put and state constraints are not explicitly considered during the controller
synthesis. To handle these constraints, post-saturation with varying limits is
1
Experimental results are not included in this work.
2.1. Electromechanical brake control 17

applied on the controller output during implementation. Apart from the ad-
hoc constraints management, optimality of the controller with respect to the
original nonlinear plant model is also not guaranteed. Note that additional
motor current is demanded by the feedback linearisation block. Hence, the
current control available to the MPC is reduced in the face of current limit.
Since the controller is optimally tuned for the feedback linearised model, it
may not perform optimally on the original plant. Nevertheless, experimental
results demonstrate that this controller is simple to implement with improved
performance compared to the cascaded PID control.

The benefits of model-based control in the application of clamp force con-

trol on EMB have been well established. First, the knowledge of plant char-
acteristics, especially the nonlinear properties can be utilised to construct an
appropriate control structure, as illustrated in the inverse-model controller and
the feedback linearising compensator. Furthermore, having a plant model al-
lows systematic controller design procedures to be devised. The same controller
tuning process can be carried out, or even automated for different sets of model
parameters. Third, rather than relying on experimentation alone, robustness
analysis can first be performed on the model perturbed from a nominal value
by some bounded measures, therefore reducing controller development time
and cost.

On the other hand, the application of model-based controller on EMB suf-

fers from several issues. Most notably, the performance of the model-based
controller is dependent on how accurate the model represents the real plant.
Additionally, many controllers are optimised upon feedback linearised model
and do not explicitly account for system constraints. Constraints, where con-
sidered, are often addressed using ad-hoc methods such as post-saturation of
the controller output. For these reasons, there is strong motivation to investi-
gate further and it is the gap in the literature which forms the basis for another
of the research questions posed in this thesis. The topic of the next section is
time-optimal control, which is a type of model-based control that provides the
best possible performance while accounting explicitly for system constraints.
18 Chapter 2. Literature review and research aims

2.2 Time-optimal control

Briefly alluded in Section 2.1.2, the construction of some model-based con-
trollers such as linear robust control and MPC involve optimal tuning of the
controller parameters. The optimisation of the gain is preceded by careful a
selection of control performance measure, which varies according to applica-
tion. For example, control objectives of many motion control applications are
dominated by the need to minimise tracking error. This is to ensure responsive
tracking for the EMB and increased throughput for the motion controlled ma-
chines. Some spacecraft problems, on the other hand, are so dominated by the
need to reduce fuel consumption. Regardless of the difference, the performance
of these systems can be evaluated by a measure of the form:
  Z tf  
J := h x(tf ), tf + g x(t), u, t dt, (2.1)
t0

where t0 , tf are the initial and final time respectively; and g, h are scalar func-
tions. As usual, x, u are the plant state and input respectively. The problem
of finding a control policy u(·) that optimises the performance measure (2.1)
is called the optimal control problem.
In many applications, such as robot arm control, automatic assembly sta-
tions and disc drive heads, controllers are designed to make optimal use of the
maximum torque available for rapid maneuvers (Newman and Souccar, 1991;
Zhang, 2000). Rapid clamp force reference tracking in EMB is also desired
for a responsive braking response and shorter braking distance. The require-
ment of time minimisation are considered in the time-optimal control design,
whereby the aim is to find a control policy that transfers the system state from
an arbitrary initial condition, x(t0 ) = x0 to a specified target set S in minimum
time. If the target set S is given by a reference, xr (·), then the performance
measure to be minimised is given by
Z tf
Jto := tf − t0 = 1 dt,
t0

where tf is the first instant of time when the trajectory of x coincides with
xr (·).
To facilitate the following discussions in this section, a general nonlinear
plant model is considered:
 
ẋ(t) = f x(t), u(t), t ,
2.2. Time-optimal control 19

where x(t) ∈ X ⊆ Rnx and u(t) ∈ U ⊆ Rnu . Furthermore, the assumption that
X and U are compact allows state and input constraints to be considered. This
is commonly applicable in practical situations as state and input constraints
are usually found in actual plants. The time-optimal control for the general
nonlinear plant model with constraints can be formulated as:
Z tf
min 1 dt, (2.2a)
u(t) t0
 
subject to ẋ(t) = f x(t), u(t), t , (2.2b)
x(t) ∈ X , (2.2c)
u(t) ∈ U, (2.2d)
x(t0 ) = x0 , (2.2e)
x(tf ) = xr (tf ). (2.2f)

The following subsections are led by a brief introduction to the time-optimal

control theory, followed by a survey of issues faced in implementation and the
required control modifications to alleviate these difficulties.

2.2.1 Pontryagin’s minimum principle

The history of optimal control theory traced back to 1697 when Johann
Bernoulli published his solution of the brachistochrone problem (Sussmann
and Willems, 1997). Since then, it has been studied under the subject of
calculus of variations and contributed by many greatest minds in history, in-
cluding Euler, Lagrange, Legendre and Weierstrass. In the 1950’s, the cold
war and the space race necessitated a means of predicting the optimal trajec-
tory for aeronautical applications, such as spacecrafts and missiles. Further
supported by the advent of computers, the interest in optimal control theory
and application mushroomed ever since.
One notable development in the optimal control theory is brought by Lev
S. Pontryagin and his coauthors from the USSR. Their work provides the
necessary conditions for optimality, named as the maximum principle in the
book (Pontryagin et al., 1962). This classical version is improved and modified
by many other authors, including (Kirk, 2004) whose version is quoted in this
subsection2 .
2
The definition of Hamlitonian (2.3) presented in Kirk (2004) is different compared to the
20 Chapter 2. Literature review and research aims

Time-optimal control without state constraint. The time-optimal con-

trol problem (2.2) where the input is bounded, but the allowable state spans
the whole Euclidean space, that is X = Rnx is first considered. The negli-
gence of the state constraint in control design can be be applied in practical
situations where the bounds of allowable state trajectories are arbitrary large.
The corresponding Hamiltonian is given by
   
H x(t), u(t), p(t), t := 1 + p(t)T f x(t), u(t), t , (2.3)

for all t ∈ [t0 , tf ] and p ∈ Rnx denotes the costate vector. Note that optimal
terms will be superscribed by ∗ in the following. According to Pontryagin’s
minimum principle (PMP) presented in Kirk (2004), necessary conditions for
the time-optimal control, u∗ (t) are
   
H x∗ (t), u∗ (t), p∗ (t), t ≤ H x∗ (t), u(t), p∗ (t), t (2.4)

and
 
− p∗ (tf )T δxf + H x∗ (tf ), u∗ (tf ), p∗ (tf ), tf δtf = 0. (2.5)
h i
dx (t )
It should be pointed out that δxf = 0 for fixed xr , while δxf = r f
dt
δtf for
time-varying state reference xr (t). Additionally, the optimal state and costate
are governed by
∂H  ∗ 
ẋ∗ (t) = x (t), u∗ (t), p∗ (t), t , (2.6a)
∂p
∂H  ∗ 
ṗ∗ (t) = − x (t), u∗ (t), p∗ (t), t , (2.6b)
∂x
and the boundary conditions are given by

x∗ (t0 ) = x0 , x∗ (tf ) = xr (tf ). (2.7)

Furthermore, if the Hamiltonian does not explicitly depend on time, then the
following additional necessary condition can be shown:
dH  ∗ 
x (t), u∗ (t), p∗ (t) = 0, t ∈ [t0 , tf ]. (2.8)
dt
Note that PMP only provides necessary condition for optimality. For cases
where input is unbounded, or the bounds on input are conceived as arbitrar-
ily large, the following additional necessary and sufficient conditions for local
maximum principle (Pontryagin et al., 1962), and this resulted in the equivalent Pontryagin’s
minimum principle.
2.2. Time-optimal control 21

optimality can be obtained:

∂H  
x∗ (t), u∗ (t), p∗ (t), t = 0,
∂u
∂ H
2  
x∗ (t), u∗ (t), p∗ (t), t > 0.
∂u2
This result can be extended to the global optimality if the Hamiltonian is
convex and can be expressed in the form
 
H x(t), u(t), p(t), t
   T 1
= c x(t), p(t), t + m x(t), p(t), t u(t) + uT (t)R(t)u(t),
2
where m(·) does not have any terms containing u(t).

Time-optimal control with state constraint. In high performance ap-

plications, such as the EMB and many motion control machineries, state tra-
jectories are often driven close to the limits of components. This necessitates
the inclusion of state constraint in the time-optimal control, which has been
studied by Pontryagin et al. (1962), McGill (1965), Jacobson et al. (1971) and
Hamilton (1972). A comprehensive review on the subject is also provided by
Hartl et al. (1995). The approach presented here is called the direct adjoining
approach (Hartl et al., 1995). In this method, the Lagrangian L is introduced
and is defined as follows:
   
L x(t), u(t), p(t), t = H x(t), u(t), p(t), t
   
+ µu (t)fsu u(t), t + µx (t)fsx x(t), t ,

where the input constraint and state constraint are given respectively by
 
fsu u(t), t ≤ 0, (2.9a)
 
fsx x(t), t ≥ 0, (2.9b)

and µu (·), µx (·) are the multipliers. Using the newly defined Lagrangian, the
costate dynamics (2.6b) are modified as

∂L  ∗ 
ṗ∗ (t) = − x (t), u∗ (t), p∗ (t), t .
∂x
22 Chapter 2. Literature review and research aims

Furthermore, in addition to the minimum principle (2.4), other necessary op-

timality conditions are introduced as follows:

∂L  ∗ 
x (t), u∗ (t), p∗ (t), t = 0,
∂u  
µu (t)fsu u∗ (t), t = 0, µu (t) ≥ 0
 
µx (t)fsx x∗ (t), t = 0, µx (t) ≥ 0

The change of Hamiltonian in time is also constrained by

dH  ∗  dL  ∗ 
x (t), u∗ (t), p∗ (t), t = x (t), u∗ (t), p∗ (t), t
dt dt (2.10)
∂L  ∗ 
= x (t), u∗ (t), p∗ (t), t .
∂t

2.2.2 Implementation issues

The time-optimal control theory reviewed so far has been focused on
continuous-time systems. However, modern control systems rely on digital
microcontrollers that operate in discrete time steps. In order to implement
the controller, one of the two common digital controller design methods has to
be employed. The first one is the emulation approach, which involves the de-
sign of a continuous controller to satisfy performance and robustness objective,
followed by the conversion of the continuous controller into a digital controller.
Alternatively, a digital controller can be design directly from a discretised plant
model. Although time-optimal control for discrete-time systems has been pro-
posed, for example in (Desoer and Wing, 1961; Keerthi and Gilbert, 1987),
it requires recursive online computation and large data storage requirements,
which add complexity in implementation. Therefore, the emulation approach
for time-optimal control is favoured for practical applications.
Note that the requirement of time minimisation is typically achieved using
extreme control efforts at all time, provided that the state constraint is not
activated. Consequently, time-optimal control is sometimes referred as bang-
bang control in literatures.
One method for solving the time-optimal control problem is to resolve
conditions given by (2.4)–(2.7) as a two point boundary value problem using
numerical iterative methods (Rosen, 1966; Kirk, 2004). The obtained optimal
control solution, which is a function of time is then implemented in open loop.
2.2. Time-optimal control 23

Clearly, this approach is not robust against disturbance and is likely to yield
large discrepancy from the intended optimal trajectory.
Another approach is to determine the number of control switching (Pon-
tryagin et al., 1962; Sussmann, 1979) and to compute the associated switching
function, which the latter provides the instant of control switching as a func-
tion of the state. A switching curve that is computed prior to online imple-
mentation facilitates the construction of a feedback control structure. As an
example, the switching function for the time-optimal control that regulates a
double integrator system:
ẋ1 = x2
(2.11)
ẋ2 = u, |u| ≤ 1,
to the origin can be analytically calculated as
q
fto (x1 ) = − sgn(x1 ) 2|x1 |, (2.12)

which is shown in Figure 2.4. Furthermore, using the predetermined switching

curve (2.12), the time-optimal control can be implemented as

− sgn (x2 )

if sgn (x2 − fto (x1 )) = 0,
uto = (2.13)
− sgn (x2 − fto (x1 ))

otherwise,

where the optimal state trajectories using control law (2.13) for several initial
values are also shown in Figure 2.4.
Implementation of time-optimal control is challenged by several issues.
First, the time-optimal control changes sign across the switching curve, resem-
bling a control with infinite gain. The sharp transition of control magnitude
may excite structural modes during maneuvers (Albassam, 2002). Further-
more, the switching function is sensitive to modelling errors, whereby impos-
ing an arbitrarily small perturbation to the model can lead to control chatter
(Ryan, 1980). In addition to modelling uncertainties, feedback delays and mea-
surement noises also induce chatter. The resulting chatter leads to vibration,
excessive heat lost in power electronics and wear in mechanical components
— all clearly undesirable in a braking application. Figure 2.5 demonstrates
the occurrence of control chatter in the double integrator system (2.11) when
the state is in vicinity of the origin. Third, apart from some simple low-order
systems, the control switching function is in general difficult to obtain and
usually does not have a analytic expression.
24 Chapter 2. Literature review and research aims

10
B u*=−1
u*=+1
5
u* = −1
dx/dt

−5
A
u* = +1
−10
−25 −20 −15 −10 −5 0 5 10 15 20 25
x

Figure 2.4: Phase portrait of time-optimal regulation trajectories of a double

integrator system (2.11). Control switching occurs along segment A-B, where
the optimal state trajectories for several initial values are shown.

3
Control Switching Control Chatter
2

1
u

−1

−2
0 1 2 3 4 5 6 7 8 9 10

0
x, dx/dt

−5
x
dx/dt
−10
0 1 2 3 4 5 6 7 8 9 10
Time (s)

Figure 2.5: Control and state trajectories of the double integrator system
(2.11) with initial condition at (−10, 0.5). Control chatters when the state is
in vicinity of the origin.
2.2. Time-optimal control 25

2.2.3 Approaches to chatter reduction

To overcome the chattering problem in time-optimal control, numerous tech-
niques have been suggested to aid practical implementations. Conceptually,
majority of these methods share a common thought, that is to enforce a gradual
change in control input across the switching curve in spite of a sharp transition.
In the following, a number of techniques that are able to achieve near-time-
optimal performance for rigid systems, such as the EMB and many robotic
manipulators in a practical feedback structure are reviewed. Note that there
are also numerous methods that are specifically tailored for flexible structures,
for example (Pao, 1994; Pao and Singhose, 1998; Albassam, 2002) deviate from
the main focus of this subsection and will not be examined here.
It can be shown that the double integrator system (2.11) and the time-
optimal feedback control (2.12), (2.13) form a discontinuous nonlinear system
that resembles a variable structure system. Inspired by sliding mode control,
which is a variable structure control method, Newman (1987, 1990) proposes
the near-time-optimal control (NTOC) for a double integrator system, whereby
the concept of boundary layer is adopted to define the control across the switch-
ing curve. Using the sliding mode terminology, the pre-computed switching
curve is treated as a sliding manifold with s = 0, the trajectories starting off
the manifold move toward it and reach it in finite time are called the reaching
phase, and the motion with state confined to the manifold is denoted the slid-
ing phase (Khalil, 2002). In the reaching phase, the distance to the manifold
is assigned to be
s(x1 , x2 ) = x1 − fsm (x2 ),
where the manifold is given by
1
fsm (x2 ) = − x2 |x2 |. (2.14)
2
Note that (2.14) is the inverse function of (2.12), and can be used to define
another form of time-optimal feedback control law equivalent to (2.13):

− sgn (x2 )

if sgn(s(x1 , x2 )) = 0,
uto = (2.15)
− sgn (s(x1 , x2 ))

if sgn(s(x1 , x2 )) 6= 0.
To suppress chatter, the NTOC incorporates a continuous function described
by " !#
s x2
  
untoc = − sat α ks sat + sat , (2.16)
ssat x2 sat
26 Chapter 2. Literature review and research aims

10

TOC
5
u = −1
dx/dt

−5
u = +1

−10
−25 −20 −15 −10 −5 0 5 10 15 20 25
x

Figure 2.6: Phase portrait of regulation trajectories of a double integrator

system using NTOC. Note that the shaded region has unsaturated control
magnitude.

where the parameters α, ks , ssat and x2 sat can be tuned according to the guide-
lines provided in (Newman, 1990). This algorithm comprises two operating
range, a reaching phase where maximum allowable control is used, and a
boundary layer close to the manifold when the control is not saturated (Fig-
ure 2.6). Furthermore, it reduces the switching structure with two cases found
in the original time-optimal control law (2.15) to a single case only, there-
fore is simpler to implement. Another novelty is presented when the state
approaches the origin, the controller becomes a proportional-derivative (PD)
controller with a velocity-squared dissipative term:
αks x1 αx2 αks x2 |x2 |
untoc,unsat. = − − − (2.17)
ssat x2 sat ssat
The derivative term 1/x2 sat can be adjusted to reduce the magnitude of over-
shoot. This approach results in a very high performance control suitable for
high speed tracking with simple tuning and implementation, while free of chat-
ter.
Further extension of the NTOC is presented in (Newman and Souccar,
1991), where a nonlinear second-order system is considered. Additionally, the
expression of the manifold (2.14) is generalised. For example, (2.14) can be
modified such that its curvature in vicinity of the origin is linear in order
to eliminate the velocity-squared dissipative term in (2.17) when the state is
approaching the origin.
An alternative approach towards chatter reduction for the time-optimal
control of a double integrator system is proposed by Workman et al. (1987),
2.2. Time-optimal control 27

and come to be called proximate time-optimal servomechanism (PTOS). Over-

all, the PTOS control law comprises three operating region, an acceleration
region where maximum allowable control is used, a region where the control is
unsaturated, and a linear zone where PD control is used. In order to generate
the linear zone in vicinity of the origin represented as |x1 | ≤ yl , the switching
curve is defined as

− k1 x1

if |x1 | ≤ yl ,
fptos (x1 ) = k2
q  (2.18)
− sgn(x1 )

2α|x1 | − 1
if |x1 | > yl ,
k1

where k1 and k2 are the control gains, and α is the acceleration discount factor
such that 0 < α ≤ 1. Furthermore, using the saturation function as the finite
slope approximation to the signum function, the PTOS control law is given by
 
uptos = − sat k2 (x2 − fptos (x1 )) . (2.19)

Additional design guidelines are provided in (Workman et al., 1987) to tune

α, k1 and k2 for system performance. Specifically, to maintain the continuity
of magnitude and slope of the switching profile, it can be shown that

1 2
s
k1 = , k2 = ,
yl αyl

with α and yl being the free parameters. Furthermore, comparing with the
characteristic equation of a second-order system

s2 + 2ζωn s + ωn2 = 0,

the natural frequency and the damping ratio of the closed-loop system in the
linear region near the origin are given by
s
1 1
s
ωn = , ζ= . (2.20)
yl 2α

A notable difference between PTOS and NTOC is the definition of their

respective switching functions, where the former is defined as a function of x1 ,
and the latter, as a function of x2 . Furthermore, comparing Figure 2.6 and
Figure 2.7 (control parameters are tuned such that both methods have almost
identical gains in the linear region), the unsaturated region of the NTOC is
significantly smaller when the state is far from the goal state. Therefore, the
28 Chapter 2. Literature review and research aims

10
PTOS
TOC
5
u = −1
dx/dt

−5
u = +1

−10
−25 −20 −15 −10 −5 0 5 10 15 20 25
x

Figure 2.7: Phase portrait of regulation trajectories of a double integrator

system using PTOS. Note that the shaded region has unsaturated control
magnitude.

0
u

−1
0 1 2 3 4 5 6 7 8 9 10

0
x

−5
TOC NTOC PTOS
−10
0 1 2 3 4 5 6 7 8 9 10
4

2
dx/dt

−2
0 1 2 3 4 5 6 7 8 9 10
Time [s]

Figure 2.8: Comparison of control and state trajectories of a double integrator

system using time-optimal control, NTOC and PTOS.
2.2. Time-optimal control 29

NTOC better utilises the control capacity during the deceleration stage of a
set-point tracking manoeuver shown in Figure 2.8.
Additional performance improvement to PTOS algorithm, called the mod-
ified PTOS or MPTOS is proposed in Dhanda and Franklin (2009). This work
introduces extra flexibility in the choice of switching curve, given by

− k1 x1

if |x1 | ≤ yl ,
fmptos (x1 ) = k2
q 
− sgn(x1 )

2α|x1 | − J if |x1 | > yl .

It can be shown that PTOS is a special case of MPTOS. From the continuity
constraints of the switching function, the gains are obtained as

2αζ 2 2α
s
k1 = , k2 = 2ζ 2
,
yl yl

with α, yl and ζ being the free parameters. Compare to PTOS, the addition
of the closed-loop damping coefficient ζ as a calibration parameter lead to a
more effective way of shaping the system response. The closed-loop natural
frequency in vicinity of the origin is given as

2αζ 2
s
ωn = .
yl

Note that both PTOS and MPTOS have the acceleration discount factor
α as a tuning parameter. While it is desirable to set α arbitrarily close to 1 to
make full use of the actuation capacity, it is typically assigned from 0.5 < α <
0.8 in practise to reduce overshoot. It is evident from (2.20) that a larger α
will yield a smaller damping, and consequently a larger overshoot. Salton et al.
(2012) proposed two solutions to remove this impediment. The first approach
involves modification of the control law in the linear region by the introduction
of a dynamic damping scheme. The second approach engages modification of
the switching curve by the replacement of α with an exponentially varying
factor scheduled by x1 .
Another approach that build on the PTOS algorithm but uses a sliding
mode controller in the linear region is proposed in (You and Lee, 2000) and
(You and Kwak, 2006). Yet another alternative is presented in (Zhou et al.,
2001) and (Zhang and Guo, 2000), whereby the switching profile in PTOS
is approximated using a third-order polynomial, and the feedback control is
30 Chapter 2. Literature review and research aims

implemented as a sliding mode controller with a time-varying sliding man-

ifold (parameterised by linearising the third-order polynomial at each time
instance).

From the survey in this subsection, it is evident that there are three main
ideas in eliminating chatter in time-optimal control while preserving high track-
ing performance. First, the abrupt change in control input across the switching
curve is replaced by a gradual transition. Furthermore, overshoot can be re-
duced by constructing a control law with high damping near the goal state.
Third, performance improvement can be achieved by reducing the acceleration
discount factor of the switching curve, for example by the introduction of a
scheduled acceleration discount factor, or approximating the switching curve
using a third order polynomial.

Nevertheless, it is noticed that majority of these approaches focus on the

double integrator problem, with a few extending beyond second-order nonlin-
ear systems. Moreover, there does not have been any substantial work done
considering state constraint. For these reasons, neither of the methods men-
tioned before are well-suited to the EMB control problem. There is a strong
motivation to investigate further and it is this gap in the literature which forms
the first research question posed in this thesis.

2.3 Active vibration attenuation

As motivated in Introduction, brake torque variation (BTV) is the main cause

of brake judder, which is a major problem in braking systems associated with
vibrations in suspension, chassis and steering wheel (Jacobsson, 2001). Util-
isation of an electromechanical brake (EMB) to actively attenuate the BTV
forms a large component of the work in this research project. To facilitate
further examination of the idea, active vibration attenuation literatures are
surveyed in the view of determining suitable algorithms to be used. It is also
aimed to identify the gaps in existing algorithms that hinder implementation,
and the resolution of these gaps will dictate further research directions.
2.3. Active vibration attenuation 31

2500
R(s) + C(s) G1 (s) = s(s+1)(s2 +s+2500) Y (s)
−

(a) A system with resonance.

di (t + pi ) = di (t) do (t + po ) = do (t)
+ +
+ 1 +
R(s) + C(s) G2 (s) = s(s+1) Y (s)
−

(b) A system with periodic forcing terms.

Figure 2.9: Systems susceptible to vibration.

2.3.1 Background
It is well-known that vibration is commonly undesirable in many engineering
systems, as it hinders normal operation of machines, deters performance and
induces fatigue failure in mechanical components. One paradigm of solution
to suppress vibration is to employ actuators available in the system to cancel
the oscillation. This is exemplified by many successful implementations in a
broad range of systems, including flexible towers (Kar et al., 2000), magnetic
resonance imaging (MRI) systems (Roozen et al., 2008), nanopositioning stages
(Fleming, 2010) and hard disc drives (Semba and White, 2008).
From the system point of view, vibrations can be categorised into resonance
and forced vibration. Figure 2.9 shows two examples of systems susceptible to
vibration, where the top plant, G1 (s) has a resonance mode at 500 rad/s, while
the bottom plant G2 (s) endures external disturbances at the input and output.
Note that brake judder is categorised as a forced-vibration excited by BTV,
which is cyclical with respect to wheel spins (Jacobsson, 2001). Therefore, it
is the compensation of external periodic disturbances that will be the focus of
this section.
A periodic disturbance, d(·) cyclical in the domain x can be characterised
by
d(x + p) = d(x), ∀x.

The domain can be either time or spatial position, with respective examples
including the oscillation of offshore landing platforms due to sea waves (Mar-
coni et al., 2002) and force ripple in linear motors (Zhao and Tan, 2005).
32 Chapter 2. Literature review and research aims

Although not all periodic disturbance has an a priori known period, for exam-
ple, the sea waves that hit on the offshore landing platforms, there are many
disturbances that occur in rotational machineries where the period can be as-
sumed to be known or measurable, with examples including the eccentricity
in disc drives (Sacks et al., 1996), rotor vibration in helicopters (Arcara et al.,
2000) and mold-level fluctuation in steel castings (Manayathara et al., 1996).
Since the wheel rotational speed in modern vehicles are measurable and readily
available online, the period of oscillation for BTV can be calculated. In the
following, both the feedback and feedforward methods that mitigate periodic
disturbances with known or measurable period are reviewed.

2.3.2 Feedback compensation approaches

The most common approach for the rejection of periodic disturbances is based
on the internal model principle, which requires a model of the disturbance gen-
eration system to be included in the feedback system (Francis and Wonham,
1975). It is well-known that a periodic disturbance can be decomposed into a
sum of sinusoidal disturbances with distinctive frequencies using Fourier trans-
form. Furthermore, the generating polynomial for each decomposed sinusoidal
disturbances is known to have a pair of poles on the imaginary axis, where
the pole locations correspond to the frequency of oscillation (see Appendix A).
Incorporating these facts to the IMP, it can be postulated that the open-loop
transfer function must contains multiple pairs of poles on the imaginary axis
matching the frequencies of the decomposed disturbances. This results in very
high gain amplitudes at these pole locations.

Filtering

One method to achieve a high gain within a frequency range is to augment an

appropriately tuned peak filter to the baseline controller. A peak filter is a
band-pass filter with amplitude response flat at all frequencies except for the
stop band on either side of the centre frequency, which can be modelled as

s2 + 2ζ1 ωs + ω 2
F (s) = ,
s2 + 2ζ2 ωs + ω 2
where ω is the centre frequency and ζ1 , ζ2 are the damping ratios with ζ1 > ζ2 .
The filter structure with parallel realisation augmented to a baseline controller
2.3. Active vibration attenuation 33

Filter

Baseline
do
F (s)
controller Plant
+ +
+ u +
r + C(s) G(s) y
−

Figure 2.10: Block diagram of a disturbance filter structure with parallel real-
isation added on a baseline servo system (Zheng et al., 2006).

is a popular configuration, especially in the reduction of harmonic disturbances

in hard disc drives (HDDs) (Wu et al., 2002; Zheng et al., 2006). The block
diagram of the overall system is depicted in Figure 2.10. The overall controller
design typically involves a two-stage approach. In the first stage, the baseline
controller is designed for basic closed-loop stability, reference tracking and
noise rejection performance. This is followed by the second stage, where the
filter is designed to reject the periodic disturbances.
In many applications, the filter state is often cycle between on and off
state and its value is re-initialised between each cycle. For example in HDD
operation, the peak filter is switched on only during the track following op-
eration, and conversely, is switched off during the track seeking operation —
to avoid filter computation excited by the rapidly changing positional error
signal. Note that similar procedure can be expected in the BTV reduction
operation, where the filter will be switched on during braking, and switched
off during cruising due to the absence of brake torque tracking error signal.
As the filter is repeatedly re-initialised during operation, although the peak
filter can effectively attenuate the tracking error after some time, a lengthy
transient after each initialisation is detrimental to the overall operational ef-
ficiency. Therefore, Sri-Jayantha et al. (1997, 2001) propose an initialisation
algorithm to provide an initial value to the filter that can reduce the duration
of the transient dynamics.
Zheng et al. (2006) argue that the peak filter is only used in low-frequency
(100-600 Hz) narrow-band disturbance in HDD, and it is hardly used to reject
the mid- and high-frequency disturbances, which can be caused by windage.
The main reason is the intrinsic phase loss due to the peak filter and causes
adverse effect on the phase margin of the open-loop system and distorts the
sensitivity gain around the disturbance frequency. Subsequently, a phase-lead
34 Chapter 2. Literature review and research aims

peak filter is proposed, given by the following form:

s[ω cos (ϕ) − sin (ϕ)s]
Fpl (s) = K ,
s2 + 2ζωs + ω 2
where ϕ is the phase angle of the complementary sensitivity function of the
baseline servo system, ω is the centre frequency and ζ is the damping ratio.
It is shown that this modified filter can reject the narrow-band disturbance at
a broader frequency range. While the two-stage design approach employing
a peak filter is very versatile and facilitates the construction of a low-order
implementable controller, it is only applicable to single-input single-output
(SISO) system.

Output regulation

In order to handle disturbance rejection for broader class of systems, such

as a multiple-input multiple-output (MIMO) system or a linear time-varying
system, a more holistic controller design facilitated by the output regulation
theory is considered. The output regulation problem is to design a feedback
controller that ensures disturbance rejection (or reference tracking) when the
disturbance (or reference) signal is produced by some known autonomous ex-
ogenous dynamic system, called an exosystem, with unknown initial conditions.
Many authors have contributed to this important problem including the pi-
oneering work by Smith and Davison (1972), Francis and Wonham (1975),
Francis (1977) and Wonham (1979). Additionally, recent developments are
also summarised in books by Byrnes et al. (1997), Saberi et al. (2000) and
Huang (2004). The essence of output regulation theory will be briefly reviewed
in the following, where the discussion will be based heavily on the results from
Saberi et al. (2000).
To facilitate further discussion, a LTI system is introduced, given by
ẋ = Ax + Bu u + Bw w,
y = Cy x + Dyw w, (2.21)
e = Ce x + Deu u + Dew w,
with state x ∈ Rnx , control input u ∈ Rnu , exogenous input w ∈ Rnw , measured
output y ∈ Rny and tracking error e ∈ Rne . Additionally, the exogenous input
is generated by an exosystem, given in the form of

ẇ = Sw. (2.22)
2.3. Active vibration attenuation 35

Exosystem
w e
Generalised Plant

u y

Controller

Figure 2.11: Feedback interconnection structure of generalised plant, controller

and exosystem in an output regulation problem.

When both the states x and w are available for feedback, that is
 
x
y =  ,
w

a state feedback controller will be considered, given by

u = Kx x + Kw w. (2.23)

In the general case of any measurement y is available for feedback, then the
output feedback controller is considered, given by

ẋc = Ac xc + Bc y,
(2.24)
u = Cc xc + Dc y,

where controller state xc ∈ Rnc . The interconnection of the plant (2.21),

exosystem (2.22) and controller (2.23) (or (2.24)) form a closed-loop system,
which can be depicted in Figure 2.11.
It is common to translate the requirements of the output regulation problem
into two parts. First is to achieve internal stability of the feedback system,
in the absence of exogenous input. Additionally, the second requirement is
to asymptotically regulate the tracking error to zero. A well-known necessary
and sufficient condition for the existence of a controller that solves the output
regulation problem is given by the existence of two matrices, Γ and Π which
solve the following Sylvester equation:

ΠS = AΠ + Bu Γ + Bw ,
(2.25)
0 = Ce Π + Deu Γ + Dew .

To obtain the controller gains ((Kx , Kw ) in (2.23) or (Ac , Bc , Cc , Dc ) in (2.24)),

the reader is referred to, for example Saberi et al. (2000) and Huang (2004).
36 Chapter 2. Literature review and research aims

In many engineering systems, such as the HDD or the vehicle wheel, their
rotational speeds are time-varying and may result in periodic disturbances with
time-varying frequencies. This type of disturbance is called non-stationary
sinusoidal disturbance, and can be modelled using a linear time-varying (LTV)
exosystem:
ẇ = S(t)w. (2.26)
Note the difference between (2.26) and (2.22) is the time-varying nature of S
in the former case. Output regulation for LTV system has rarely been consid-
ered until the last decade in (Ichikawa and Katayama, 2006; Shim et al., 2006;
Zhang and Serrani, 2006; Shim et al., 2010). While most output regulation
results for the LTI system can be extended to the LTV system, one notable dif-
ference between the two is the introduction of a differential regulator equation
(Ichikawa and Katayama, 2006), given in the form of

Π̇ + ΠS = AΠ + Bu Γ + Bw ,
(2.27)
lim (Ce Π + Deu Γ + Dew ) = 0.
t→∞

Compare to the linear matrix equality form of (2.25), obtaining a solution for
the differential equation (2.27) is much more involving. Nevertheless, existing
algorithms are available, such as those provided in (Shim et al., 2006) to obtain
solution for certain type of minimum phase system, and in (Shim et al., 2010)
for certain type non-minimum phase system.
Further extension is proposed recently to the output regulation of linear
parameter-varying (LPV) systems, whereby the rejection of non-stationary si-
nusoidal disturbance is treated in (Köroǧlu and Scherer, 2011b) for measurable
frequencies and in (Köroǧlu and Scherer, 2011a) for uncertain frequencies. Sim-
ilar to the treatise on LTV systems, the necessary and sufficient condition for
the existence of a controller is governed by the differential regulator equation
(2.27). Furthermore, the controller is in the form of a LPV system, with a set
of gains scheduled by the frequency.
Although the derivation of a scheduled controller that achieves output reg-
ulation is investigated, the implementation of such controller has not been
extensively investigated. For practical reasons, there exists a strong motiva-
tion to investigate the discretisation of the LPV output regulation controller
and thus forms the basis for another research question to be addressed in this
thesis.
2.3. Active vibration attenuation 37

Repetitive Controller

+
Q(s)e−ps B(s) Generalised
do
+ Plant
+ +
+ + y
r + G(s)
−

Figure 2.12: Plug-in repetitive control system.

2.3.3 Learning compensation approaches

Based on a different concept, learning compensation methods rely on the online
learning of a compensating input to the plant that negates the effect of the
disturbance. They are distinguished by their simple implementation and little
performance dependency on system parameters.
One popular learning algorithm is called the repetitive control method. A
comprehensive survey paper by Li et al. (2004) and a more recent paper by
Wang et al. (2009) presents excellent summary of the work done in this area.
The repetitive control is based on the internal model principle, combined with
the fact that any periodic signal with a known period, p can be generated by
a time-delay positive feedback system

e−ps
M (s) =
1 − e−ps

with an appropriate initial function (see Appendix A). By the inclusion of

M (s) into the feedback loop and stabilising it, complete rejection of any peri-
odic signals with the period, p can be achieved. One choice of repetitive control
configuration is the plug-in repetitive control, which is depicted in Figure 2.12,
where the dynamic compensator, B(s) and the low-pass filter, Q(s) are tuning
parameters. Note that the repetitive control system based on M (s) is infinite
dimensional, whereby asymptotic rejection of arbitrary periodic signal is pos-
sible. However, stabilising the repetitive control system is a non-trivial task
due to the inclusion of the time-delay element in the positive feedback loop
(Li et al., 2004). Furthermore, the unrealistic demand for perfect rejection
of arbitrary periodic signal including the high frequency components is not
only restrictive from a practical sense, but also limits its application to proper
plants only (Hara et al., 1985).
38 Chapter 2. Literature review and research aims

Although modified repetitive control approaches are provided by many au-

thors, including Inoue et al. (1981) and Hara et al. (1988) to mitigate the
stability issue, it comes at the cost of degraded performances at high frequen-
cies. Furthermore, since the classical framework of repetitive control is based
on a fixed known period, it is not suitable for compensation of disturbance
with time-varying frequencies. Additionally, its application is also restricted
to SISO systems.
Most periodic disturbances have a finite number of harmonics that are
significant. To take advantage of this fact, another learning compensation
method, called the adaptive feedforward cancellation (AFC) scheme is pro-
posed in (Chen and Paden, 1990; Bodson et al., 1994). In this approach,
adaptive control law is utilised to estimate the Fourier coefficients of the com-
pensating input signal that negates the disturbances. Figure 2.13 shows the
feedback structure of a plant with AFC and input disturbance. Bodson et al.
(1994) show that a disturbance with single harmonic can be cancelled using
the following compensation law:

γ̇ = −gy$
(2.28)
uaf c = γ T $,

where y is the plant output and

   
γ1 cos (ωt) 
γ =  , $= .
γ2 sin (ωt).

Furthermore, they also demonstrate that the adaptive compensation law is

equivalent to a linear time-invariant compensator of the form
s
,
s2 + ω2
which is a special case of the peak filter. Additionally, the equivalence is
also extended to input disturbances with time-varying frequencies in (Bodson,
2004; Guo and Bodson, 2010), whereby the w matrix is scheduled on the
basis of frequency. Nonetheless, the equivalence has its limits in practise,
where the preference on which method to use depends on implementation
considerations. For instance, Bodson et al. (1994) postulate that the AFC
can be utilised to compensate the eccentricity in HDDs even during the track
seeking operation (when the tracking error signal is either not available or is
2.3. Active vibration attenuation 39

Adaptive Feedforward Cancellation

R θ1
× g dt ×
+
cos (ωt) cos (ωt)

R θ2 +
× g dt ×
di Generalised
sin (ωt) sin (ωt) Plant
+ +
+ + y
r + G(s)
−

Figure 2.13: Plug-in adaptive feedforward cancellation scheme.

rapidly changing due to tracking transient). It is this case where the AFC
demonstrates its advantage by simply freezing the parameter updates while
keeping the cancellation scheme in operation. Furthermore, in processes where
the compensator is frequently re-initialised, it is easy to store the parameters
values for initialisation at the next startup.
However, it is noticed that the AFC literatures mainly focus on the rejection
of input disturbances. While the complete cancellation of input disturbances
with time-varying frequencies can be achieved by scheduling the AFC scheme
with the frequency, Kinney and de Callafon (2011) indicate that the same ap-
proach is not sufficient for output additive disturbances when the frequency
is rapidly varying. This limitation hinders its implementation on the brake
judder compensation problem as the BTV is an additive output disturbance
with a varying frequency, which corresponds to the changing vehicle speed.
Extending the AFC for output disturbances with time-varying frequencies is
essential for its implementation in BTV compensation. Therefore, this pro-
vides the motivation for further investigation and forms the basis for the final
research question to be addressed in this thesis.

2.3.4 Summary
Rejection of periodic disturbance is a classical field that has been examined
using a variety of methods, including the approaches based on internal model
principle and learning-type compensation methods. While active vibration
attenuation is a mature area of research, active reduction of BTV using an
EMB is a novel idea that has not been investigated before in the literature.
40 Chapter 2. Literature review and research aims

A survey on existing literatures provides the basis for determining suitable

algorithms to be used.
The output regulation theory is the direct result of the internal model prin-
ciple and appears to be a versatile design tool for synthesising a disturbance
compensator. While existing results are available for compensating aperiodic
disturbances, they are mainly derived in continuous-time framework. Further-
more, the resultant controller may be in the form of a LPV system, scheduled
by the frequency. Discretisation of a LPV output regulation controller remains
a challenge and further work has to be done.
Furthermore, the AFC scheme offers several notable advantages from a
practical sense. First, it is relatively simple to implement and tune, with little
performance dependency on system parameters. Second, in the case where the
feedback error signal is temporarily unavailable, the compensator parameter
updates can be stopped without discontinuing the generation of cancellation
signal. This feature is useful in processes where the scheme is frequently re-
initialised. However, existing literatures primarily focus on the rejection of
input disturbances and there remains gaps in the field of implementing AFC
for output disturbances with varying-frequencies.

2.4 Literature review conclusions

Two aspects of electromechanical brake (EMB) control are reviewed in this
chapter, namely the clamp force control design and active vibration attenua-
tion algorithms that are suitable for the compensation of brake torque variation
(BTV).
First, the survey of existing EMB control designs has revealed that the cas-
caded proportional-integral-derivative (PID) control remains the most popular
choice in the industry. However, there are several performance setbacks with
this design, notably the difficulty in gain tuning and the inconsistent perfor-
mance over the whole clamp force operating envelop. Although performance
improvement are obtained through the application of various model-based con-
trol design techniques, it is found that the majority of these technique rely on
a feedback linearised model, where the optimal performance cannot be guaran-
teed on the real system which has nonlinear characteristics. Additionally, ac-
tuator and state constraints are often handled using ad-hoc approaches. This
2.5. Research aims 41

highlights the need to investigate further an optimal control design method

that considers the nonlinear characteristic as well as the constraints.
The second topic of the review centred on active vibration attenuation
algorithms suitable for the application of brake torque variation (BTV) atten-
uation. This topic is motivated by the survey of existing approaches to brake
judder in Introduction. The survey identify the BTV as the major cause of
brake judder, and highlights that the current resolutions either fall in the cate-
gory of corrective measures, or are only locally isolating the vibration. Ideally,
judder should be attenuated at the brake, which is the source of vibration.
The EMB offers a promising alternative to judder attenuation. However, to
the author’s best knowledge, the idea of utilising an EMB to actively attenuate
the BTV has not been reported in the literature, and remains a novel concept
to be examined.
Due to the geometrical relationship of BTV and brake rotor, the resultant
oscillation has a time-varying frequency in normal driving conditions. Fur-
thermore, the BTV can be regarded as an additive output disturbance that
contributes to the overall brake torque. From the survey of available active
vibration attenuation techniques, output regulation and adaptive feedforward
compensation methods emerge as two promising choices for BTV compensa-
tion. However, there does appear to be scope for further research and the
potential to make contributions in the following areas:

• A corresponding compensator designed using the linear output regulation

theory is a linear parameter-varying (LPV) controller. To aid implemen-
tation on a digital controller, discretisation of the output regulation LPV
controller is a challenging problem that requires further research.
• Extension of existing Adaptive Feedforward Cancellation (AFC) scheme
to additive output periodic disturbances with varying-frequencies is a
necessary step for its implementation in the BTV compensation problem.

2.5 Research aims

The aim of this thesis is to design a brake torque controller for an automotive
electromechanical brake (EMB) that provides excellent response in the pres-
ence of brake torque variation (BTV). This requires the construction of a high
bandwidth clamp force controller that best utilises available motor torque in
42 Chapter 2. Literature review and research aims

the EMB. Furthermore, an outer BTV compensation loop is proposed to de-

sign a compensating clamp force signal that attenuates the BTV. In response
to this, the research is broken up into the following specific tasks that are
aimed to be achieved in this thesis:

To study the feasibility of shortening the time required for

parameter identification of an EMB

A suitable problem formulation for parameter identification and subsequently

solving it is proposed, where the aim is to decrease the time required for
parameter identification while maximising estimation accuracy.

To develop a benchmark for clamp force tracking and to devise an

implementable controller that approaches the benchmark

To identify the best possible clamp force tracking trajectory in the face of
nonlinear dynamics and system constraints of the EMB, and to use it as a
benchmark for system performance. The established benchmark can be utilised
for subsequent clamp force controller developments to fully utilise the hardware
capability of the EMB. Experimental validations are required to demonstrate
that the proposed controller is practically implementable and to assess its
performance.

To study the feasibility of attenuating brake torque variation using

an EMB

The idea of utilising an EMB to actively compensate for BTV is investigated.

To demonstrate the developed compensators are practically feasible and deliver
the performance predicted, experimental validation is necessary.

2.6 Thesis layout

This thesis proposes utilising an electromechanical brake (EMB) to compen-
sate for brake torque variation, which is responsible for causing brake jud-
der. Firstly, the control architecture that best utilises the available control
actuation is devised, therefore renders higher closed-loop bandwidth for the
2.6. Thesis layout 43

force control loop. By utilising the achieved bandwidth, brake torque varia-
tion compensation schemes are designed to mitigate the brake judder problem.
The overall structure of the thesis is depicted in Figure 2.14, while the specific
goals of each chapter are detailed in the following.
Chapter 2 reviews recent developments in the control of an automotive
electromechanical brake, existing solutions for brake judder attenuation and
various controller designs that will be further developed in the context of EMB
application. The review highlights the need to actively attenuate judder di-
rectly from the source. Furthermore, two topics on the control design warrant
further investigation: the development of a constrained time-optimal controller
and the oscillatory disturbance compensation schemes.
In Chapter 3, experimental setup used for this study is introduced. Utilising
the test rigs, EMB models with different levels of complexity are developed and
parameterised, and the brake judder phenomenon is examined.
In Chapter 4, a fast and accurate parameter identification procedure is pro-
posed to support quick quality control checks and rapid model-based controller
tunings post-production.
In Chapter 5, time-optimal tracking control of an EMB is first investigated
using Pontryagin’s minimum principle. To facilitate practical implementation,
robust near-time-optimal tracking control is developed. The performance of
this controller is experimentally validated against the current industry prac-
tise, which is the cascaded proportional-integral controller. Finally, the model
predictive controller is presented as the generalisation of preceding designs.
In Chapter 6, a linear parameter-varying torque variation compensator is
developed to attenuate brake torque variation (BTV) accountable for judder
development. Emulation design is investigated to aid practical implementation
of this scheme in a digital controller. Furthermore, experimental results of BTV
compensation are presented and discussed.
In Chapter 7, an adaptive feedforward torque variation compensator is pro-
posed to attenuate BTV. This design is shown experimentally to be effective
for BTV compensation under both fixed and varying wheel speed, and demon-
strates the advantage of having wheel position measurements that are typically
absent in production vehicles.
The last chapter summarises the major contributions of this thesis and sug-
gests further investigations that would further supplement the present results.
44 Chapter 2. Literature review and research aims

Introduction (Ch. 1)

Literature review and

research aims (Ch. 2)

Experimental setup and model

descriptions (Ch. 3)

Rapid EMB model parameter

identification (Ch. 4)

High performance clamp

force tracking control (Ch. 5)

BTV compensation with wheel BTV compensation with wheel

speed measurements (Ch. 6) position measurements (Ch. 7)

Conclusions (Ch. 8)

Figure 2.14: Thesis structure.

 3

Experimental setup and

model descriptions
The chapter begins by introducing the production-ready electromechanical
brake (EMB) prototype which is the focus of this thesis. Also presented is
the bench top experimental setup used to identify the EMB system dynamics
and to measure the clamp force tracking performance of new control algo-
rithms. This is followed by a description of a dedicated brake dynamometer
which is available for the testing of braking performance under rotating disc
rotor and the examination of the novel brake torque variation compensation
algorithms.
Sections 3.2 to 3.3 are devoted to the development of EMB models used for
simulation and controller design. In Section 3.2, a short review on the EMB
modelling is presented, leading to the development of a simulation environment
with suitable fidelity to be used for controllers testing. Order reduction on the
simulation model and the current controller is presented in Section 3.3, where
a control-oriented model is developed.
Section 3.4 presents a grey-box model of the brake torque variation (BTV)
caused by geometry irregularities of a rotating disc rotor. Experimental results
are obtained from a dedicated dynamometer by applying clamp force on a
rotating disc rotor with runout and thickness variation. The resultant brake
torque and its variation due to the geometry irregularities are measured and
analysed, with the aim of characterising the brake torque variation (BTV) as
a function of disc rotation.

45
46 Chapter 3. Experimental setup and model descriptions

3.1 Experimental facilities

Access to production-ready EMB prototypes and test rigs were provided
through the Research Centre for Advanced By-Wire Technologies (RABiT),
a joint venture between the University of Melbourne, Pacifica Group Tech-
nologies (PGT) and Swinburne University of Technology. The EMB and the
test rigs utilised for experimental work in this thesis are described here.

3.1.1 Electromechanical brake

Figure 3.1 shows the production-ready EMB prototype used in this thesis, the
PBR T5.1 series Smart Calliper patented in (Wang et al., 2005) and manufac-
tured by Pacifica Group Technologies (PGT). The EMB is powered by a 42 V
DC power supply and is capable of producing 40 kN maximum clamp force.
Actuation is provided by a 1 kW three-phase, permanent magnet syn-
chronous motor (PMSM), which is capable of producing 3 Nm maximum
torque and 300 rad/s maximum rotational speed. Figure 3.2 shows the cross-
sectional view (as a 3D CAD model) of the EMB, which comprises a PMSM,
a planetary gear set, a ball screw and a floating calliper. Note that the ac-
tuation is solely provided by the PMSM without self-energisation1 , and the
power transmission from the motor to brake piston is achieved by the planetary
gear unit and ball screw (Wang et al., 2005). This arrangement is commonly
found in many other EMB designs with minor distinctions, including (Taig
et al., 1988), (Schumann, 2001), (Halasy-Wimmer and Völkel, 2005), (Siler
and Drennen, 2002) and (Watanabe et al., 2007).
The controller board attached to the EMB can be seen in Figure 3.3. It in-
cludes a 16-bit hybrid controller, signal conditioning circuits and an integrated
power stage. The motor current control loop and the clamp force control loop
are run at 5000 Hz and 250 Hz respectively. Additionally, the clamp force
reference is updated at 125 Hz. For the purpose of feedback control, internal
sensors are equipped to measure motor position, three-phase motor currents
and clamp force. The motor position is measured using an inductive position

1
A self-energising brake, such as the Electronic Wedge Brake (EWB) uses the actual
braking drag force to increase the force with which the brake is applied, where wedge-shape
mechanisms are typically employed in the design (Dietrich et al., 2001).
3.1. Experimental facilities 47

Figure 3.1: PBR T5.1 series Smart Calliper.

1 Planetary gears, 1 2 3 4 5 6 7
2 Motor stator,
3 Motor rotor,
4 Ball screw shaft,
5 Piston,
6 Pad,
7 Calliper.

Figure 3.2: Cross-sectional view of a PBR T5.1 series Smart Calliper.

48 Chapter 3. Experimental setup and model descriptions

Figure 3.3: Controller board of a PBR T5.1 series Smart Calliper.

sensor mounted about a toothed gear wheel on the motor rotor2 . Numer-
ical differentiation is utilised to calculate the instantaneous motor velocity
from the motor position measurements. The three phase motor currents are
evaluated from the DC link current measurements obtained using a DC link
shunt resistor. The three phase motor currents are used to evaluate the motor
quadrature current, which is then utilised to calculate the instantaneous motor
torque. Additionally, the clamp force is measured using a rosette strain gauge
mounted on the outside of the EMB piston sleeve. Lastly, communication with
the EMB is accomplished via the controller area network (CAN) bus interface.

3.1.2 Bench top test rig

Clamp force reference tracking experiments are conducted on the static test
rig shown in Figure 3.4. The bench top facility includes a PC equipped with
Vector CANcardXL for CAN communication, a data acquisition system com-
prising a National Instruments PCI-MIO-16XE-10 and a SC-2345 Carrier, and
2
The measured motor position is not used for clamp force feedback control due to the
unknown zero offset, which is defined when the brake pads are in contact with the rotor.
3.1. Experimental facilities 49

1 2 3 4

Figure 3.4: Bench top experimental setup. The individual components are:
1. Data acquisition board; 2. PC; 3. EMB; 4. Brake calliper stand with load
cells.

three HBM C9B load cells used to ascertain the clamp force measurements ob-
tained from the internal sensor of the EMB. Not shown in the figure is a Delta
Elektronika SM 45-70 power supply that provides 42 V DC to the EMB.
Vector CANape, a software package with scripting capability is utilised to
update clamp force reference and log measurements. Furthermore, embedded
programming for the EMB is coded in Matlab/Simulink with Real-Time
Workshop Toolbox.

3.1.3 Brake dynamometer

A dedicated brake dynamometer depicted in Figure 3.5 is utilised to test the
braking performance of an EMB under a rotating disc rotor. The dynamometer
has the capacity of delivering 440 Nm torque at 375 rpm to the rotating brake
disc. Actuation is provided by a Siemens three-phase brushless servomotor
(model 1FT6108-8WB71-1AH0-Z) with 110 Nm rated torque and 1500 rpm
maximum speed. The motor is equipped with a step-down gearbox with 4:1
ratio, where a disc rotor is attached to the output shaft. At the far end of
the dynamometer, there is an independent shaft with one end rigidly secured
to the structural frame of the dynamometer, while the other end connected to
the EMB. A Crane Electronics CheckStar torque sensor is equipped on this
shaft to measure the instantaneous brake torque during braking.
50 Chapter 3. Experimental setup and model descriptions

1 2 3 4 5 6

Figure 3.5: Brake dynamometer. The individual components are: 1. Dy-

namometer motor; 2. CAN cable; 3. Disc rotor; 4. EMB; 5. Thermometer;
6. Torque sensor.

A Siemens SIMODRIVE 611 converter system is employed to drive the

dynamometer motor. It is configured to accept dynamometer motor speed
reference as input and provides the instantaneous motor position or speed
measurement. Communications between the PC and the brake dynamome-
ter are established using the analog input and output ports of the National
Instruments data acquisition system.

3.2 Modelling of an electromechanical brake

Mathematical models that sufficiently represent the EMB and capture the
relevant dynamics are essential for simulation studies and the construction
of model-based controllers. This section begins with a survey of high-order
3.2. Modelling of an electromechanical brake 51

physical models of the EMB in the view of choosing a low-order model that
is adequate for simulation studies. This is followed by the development and
parameterisation of the simulation model.

3.2.1 Overview of high-order models

The complete model of the EMB consists of two main elements: the mechanical
plant and the electrical system. Using modern computer aided design (CAD)
softwares, models with very high-order are developed for the mechanical plant
using finite element analysis (FEA) during the design phase. However, eval-
uation of FEA models are time-consuming due to the enormous degrees of
freedom. Therefore, a lower-order model is preferred during stages of con-
troller development to reduce complexity in parameter estimation and time
involved during simulations.
Kwak et al. (2004) derive a ten degree-of-freedom physical model of an
EMB with planetary gear transmission similar to the EMB used in this work.
This model includes the dynamics of the sun gear, planetary gears, nut carrier,
spindle and brake calliper, whereby nonlinear effects such as disc gap clearance,
Coulomb frictions and gear backlashes are considered. To study the possibility
of further reducing the system order, low frequency dynamics of the mechan-
ical plant are obtained by linearising the full nonlinear model, proceeded by
the calculation of natural frequencies and mode shapes. The results of model
analysis show that only the first mode is critical for the brake operation, there-
fore suggesting the use of one degree-of-freedom lumped parameter model to
represent the mechanical plant. The lumped parameter model can be derived
using the torque equation around the motor rotational axis, whereby the over-
all inertia, friction and stiffness are individually lumped into three terms. This
model is given in the form of

J θ̈m = −Tl (θm ) − Tf (θm , θ̇m ) + Tm , (3.1)

where θm , J, Tf , Tl , Tm are the motor angular position, lumped inertial, friction

torque, load torque and motor torque respectively. The work by Kwak et al.
(2004) has not only provided a low-order mechanical plant model that reduces
complexity in parameter identification and controller development, it also offers
analytical justification that reaffirms the choice of the model, which has been
previously proposed by Maron et al. (1997) and Lüdemann (2002).
52 Chapter 3. Experimental setup and model descriptions

The torque that drives the mechanical plant is provided by a PMSM, which
motion is governed by three-phase AC voltage supplied by the electrical system.
Detailed modelling of the PMSM circuitry can be found in (Lyshevski, 1999).
Line (2007) shows that the application of the direct-quadrature-zero (or dq0)
transformation on the three-phase circuit of an EMB reduces the three AC
quantities to two DC quantities, where simplified analysis can be carried out.
The equivalent DC circuit equations are given by
did
Lm = −Rm id + Lm iq θ̇r + vd ,
dt (3.2)
diq
Lm = −Rm iq − Lm id θ̇r − λm θ̇r + vq ,
dt
where id , iq are the currents in the d- and q-axis respectively, similarly vd , vq
are the voltages in the d- and q-axis respectively, Lm , Rm are the motor syn-
chronous inductance and winding resistance respectively, λm is the magnitude
of the flux linkage, and θr is the electrical angle. Furthermore, the motor
torque is given by
Tm = K t i q , (3.3)

where Kt is the motor torque constant.

It can be shown that for a given magnitude of phase current, the maximum
motor torque can be attained if the direct current is set to zero. To achieve
this, a designated feedback controller is used to regulate the direct current to
zero. Hence, the order of (3.2) can be reduced under the assumption that
id = 0, therefore arriving at
diq
Lm = −Rm iq − Kb ωm + vq , (3.4)
dt
where the following relationships of Back Electromotive Force (back EMF) are
used:
2
λm θ̇r = Kb ωm , Kb = Kt . (3.5)
3
Note that the quadrature voltage is regulated using the quadrature current
controller, typically implemented using a proportional-integral (PI) controller.

3.2.2 Development of a simulation model

The simulation model adopted in this work includes the dynamics of mechan-
ical components (3.1) and electrical circuits (3.4), (3.3). The resultant simu-
3.2. Modelling of an electromechanical brake 53

lation model is proposed as

dθm


 = ωm ,
dt





 dω
Σemb,sim : J m = −Tl (θm ) − Tf (θm , ωm ) + Kt iq , (3.6)


 dt
di


 Lm q = −Rm iq − Kb ωm + vq ,


dt
where ωm is the motor velocity. In order to attain a complete simulation
model, the characteristics of load torque Tl (·) and friction torque Tf (·) need
to be identified.

Characteristics of load torque

The load torque, Tl accounts for the stiffness of brake pads, calliper and gears,
and is described by
Tl (θm ) = N Fcl (θm ), (3.7)

where N is the gear ratio and Fcl (θm ) is the clamp force as a function of motor
position.
The lumped stiffness of the EMB calliper can be characterised by the
amount of motor angular displacement needed in order to generate a set
amount of clamp force. To determine the stiffness curve, the motor position
was incrementally varied stepwise from the contact position until the maximum
clamp force was reached, where the contact position is defined when the brake
pads and disc have just came into contact. The clamp force measurements are
plotted against the motor position in Figure 3.6. A third-order polynomial of
the motor position is fitted to the measured clamp force using the least squares
method. Since the polynomial visually overlaps with the measurements, this
demonstrates that the polynomial approximates the clamp force well under
the loading condition when the applied clamp force is increasing. With θm = 0
defined at the contact position, the stiffness characteristics is modelled using:

k1 θ 3 + k2 θ 2

if θm ≥ 0,
m m
Fcl (θm ) = (3.8)
0

if θm < 0.

Furthermore, to study the hysteresis effect, the stiffness curve identifica-

tion procedure was repeated, albeit with the motor position reversed stepwise
54 Chapter 3. Experimental setup and model descriptions

4
x 10

3 Measurement (Apply)
Measurement (Release)
Model (Fcl = −0.3535θ3m + 35.19θ2m)
2.5

2
Clamp Force [N]

Apply
1.5

1
Release

0.5

0
0 5 10 15 20 25 30 35
Motor Position [rad]

Figure 3.6: Comparison between the clamp force and motor position measure-
ments with the fitted stiffness curve (3.8).

starting from the maximum clamp force to the contact position. The hystere-
sis effect can be observed from Figure 3.6. This indicates that the clamp force
is dependent on not only the motor position, but also the direction or rate of
loading (or unloading). To capture the effect of hysteresis, a rate-dependent
stiffness model is suggested by Jo et al. (2010). However, hysteresis is not
considered here as the modelling error due to hysteresis is less substantial
in practise compared to the effect due to variations in pad thickness or pad
temperature (Schwarz et al., 1998).

Characteristics of friction torque

The friction torque, Tf (θm , ωm ) comprises load-dependent static, Coulomb and

viscous sources of friction acting on the motor’s rotational axis. Comprehensive
reviews on friction models are provided in (Armstrong-Hélouvry et al., 1994)
and (Olsson et al., 1998). The friction model adopted here is based on the
friction model presented in (Hess and Soom, 1990), but modified to also include
the Karnopp friction model (Karnopp, 1985). The Karnopp friction model
3.2. Modelling of an electromechanical brake 55

incorporates a zero velocity interval, |ωm | <  as a remedy to mitigate the

problem of zero velocity detection during simulations and to avoid switching
between different state equations for sticking and sliding. Hence, the friction
model is given in the form of
  
Tv ωm + TC +
 Ts −TC
sgn (ωm ) if |ωm | > ,
1+(ωm /ωs )2
Tf (θm , ωm ) =  (3.9)
min (|Te |, Ts ) sgn (Te ) if |ωm | ≤ ,

where TC , Ts , Tv and ωs are the Coulomb friction torque, load-dependent static

friction torque, viscous friction coefficient and the characteristic velocity of the
Stribeck friction respectively. Additionally, Te is the net external non-friction
torque, given by
Te = Tm − Tl (θm ). (3.10)

The friction model parameters were identified using three set of experi-
ments. The first set of experiment identifies the load-dependent static friction
torque, Ts , where a series of break-away tests was executed in air-gap, featuring
small increments of motor torque from stationary until motion is detected. The
motor torque that initiated the motion corresponded to the torque that was
required to overcome the stiction. This set of experiment was executed in the
air-gap to isolate the effect of clamp force. From 1000 observations, the stiction
torque resembles closely to a normally distributed probability density function
with mean 4.68 × 10−2 N·m and standard deviation 0.90 × 10−2 N·m, as shown
in Figure 3.7. In the absence of clamp force, the static friction torque solely
comprises of the load-independent component, given by Ts0 = 4.68×10−2 N·m.
The second set of experiment for identifying the friction model parameters
was the break-away tests executed with applied clamp force. Due to the pres-
ence of clamp force, the torque that started the motion corresponded to the
sum of load torque (3.7) and static friction torque, given by

Tm = Tl + Ts = N Fcl + Ts .

Experimental results are shown in Figure 3.8, where each data point represents
the mean of 20 repeated trials. The measured break-away torque increases with
clamp force, and can be well-approximated using the linear regression model
fitted using the least squares approach. Therefore, the load-dependent static
friction torque, Ts can modelled as a linear function of clamp force, in the form
56 Chapter 3. Experimental setup and model descriptions

200
Avg: 0.0468
Static Friction, Ts [Nm]
0.06
Std: 0.0090

Number of samples
150

0.04
100

0.02 50
Measurement
Mean
0 0
0 500 1000 0.02 0.04 0.06
Sample number Static Friction, Ts [Nm]

Figure 3.7: Static friction measurements from break-away tests in the air-gap,
where the mean is 0.0468 Nm

of
Ts (Fcl ) = Ts0 + µs Fcl , (3.11)

where Ts0 and µs are the load-independent static friction torque and the coeffi-
cient of friction respectively. Employing the y-interception of the linear regres-
sion, the load-independent static friction force is found to be 4.95 × 10−2 N·m.
Furthermore, using the slope of the linear regression and the gear ratio,
N = 2.63 × 10−5 m/rad, the load-dependent static friction coefficient is calcu-
lated as 3.53 N·m/N.
Note that the routine used in the break-away tests with applied clamp force
possesses several drawbacks. As one could expect, the process of incremental
increasing motor torque until motion is detected can be lengthy. In order to
speed up the process, larger torque increments were used, which could possibly
have resulted in a slightly larger load-independent static friction torque esti-
mate (4.95 × 10−2 N·m) compared to the previously obtained value from the
break-away tests in air-gap (4.68 × 10−2 N·m). Therefore, to resolve the loss of
accuracy in the estimation of Ts0 , the measurement data is refitted using the
least squares approach, however with Ts0 constrained to 4.68 × 10−2 N·m. The
constrained linear fit is depicted in Figure 3.8, which visually overlaps with
the linear regression obtained without constraints. From the slope of the con-
strained linear fit, the coefficient of friction is recalculated as µs = 3.55 N·m/N.
Additionally, the data point distribution along the clamp force axis is sparse
at high clamp force (> 18 kN), and measurement data for the maximum al-
lowable clamp force (30 kN) was not acquired. Since high clamp force leads
3.2. Modelling of an electromechanical brake 57

1.6
Measurement
Least Squares Linear Fit (y = 6.16e−05*x + 0.0495)
1.4 Constrained Linear Fit (y = 6.18e−05*x + 0.0468)

1.2

1
Torque [Nm]

0.8

0.6

0.4

0.2

0
0 0.5 1 1.5 2 2.5
Clamp Force [N] 4
x 10

Figure 3.8: Measurements from break-away tests with applied clamp force.

to large load torque and friction torque, a larger motor torque is required to
balance the motion. However, a prolonged high torque demand results in high
temperature that can potentially damage the power electronics. This prohib-
ited the execution of this test at high clamp force region. Therefore, the data
collected does not cover the whole operating envelope, and fitting a model us-
ing this data implicitly presume the extrapolation of data. This highlights the
need of a quicker identification procedure that can be performed throughout
the whole operating capability of the EMB.
The final parameters to be identified in the friction model were the
Coulomb friction, viscous friction coefficient and the characteristic velocity of
the Stribeck friction, represented by TC , Tv and ωs respectively. To achieve this,
the last set of experiment was executed by running the motor with a constant
velocity reference. In order to isolate the identification of Coulomb friction and
viscous friction coefficients, the tests were done in the air-gap. Due to a combi-
nation of cogging torque and torque ripple, as shown in Figure 3.9, the motor
torque required to maintain constant motor velocity varies sinusoidally. By us-
ing the average values of motor velocity and torque obtained from each test, a
58 Chapter 3. Experimental setup and model descriptions

Motor Velocity [rad/s] 2.5

Measurement
2
Mean
Mean: 0.948
1.5

0.5

0
0 2 4 6 8 10 12 14 16 18

0.08
Mean: 0.0377
0.06
Torque [Nm]

0.04

0.02

0
0 2 4 6 8 10 12 14 16 18
Time [s]

Figure 3.9: Example of experimentally obtained friction time series data for
motor velocity reference at 1 rad/s.

plot of friction torque versus motor velocity can be constructed, as depicted in

Figure 3.10. The previously identified load-independent static friction torque
is also plotted on the figure for comparison, showing a slightly higher value
compared to the measured dynamic friction at very low speed. The friction
model (3.9) is fitted to the measurement data using the least squares approach,
where the parameters TC = 7.67 × 10−2 N·m, Tv = 4.46 × 10−4 N·m·s/rad and
ωs = 28.85 rad/s are found. The parameterised model is plotted on Fig-
ure 3.10, demonstrating a good fit to data. Note that the load-independent
static friction is found lower than the Coulomb friction, typically encountered
when way lubricants are used (Armstrong-Hélouvry et al., 1994).
To gain a qualitative assessment of the combined effect of temperature
and friction memory on the total friction torque, constant velocity tests were
performed in cold start condition, where each test was started when the EMB
components were approximately at room temperature, and were allowed to cool
down after the test. The results obtained in the cold start condition are plotted
in Figure 3.10, which generally have higher values compared to the normal
3.2. Modelling of an electromechanical brake 59

Measurement (Normal Operation)

0.25
Load−Independent Static Friction
Friction Model
Measurement (Cold Start)
0.2
Torque [Nm]

0.15

0.1

0.05

0
0 50 100 150 200 250
Motor Velocity [rad/s]

Figure 3.10: Friction torque measurements obtained during constant speed

reference tracking in air-gap.

operation. Although exact component temperature was not available due to

the lack of accurate temperature measurements inside the EMB housing, the
results are suffice to indicate the potential variations in friction characteristics
due to changes in operating conditions. Therefore, robustness analysis of the
as-yet-undiscussed controllers against parameter variations is essential.
Note that the friction model adopted here does not capture dynamical
phenomena such as hysteresis and rate-dependent break-away force, as offered
by the bristle-based dynamical model (or known as the LuGre model) proposed
by Canudas de Wit et al. (1995). However, the simplicity offered by this model
allows relatively straight forward parameter identification procedures to be
carried out, while it also accurately predicts the steady-state behaviour. These
two advantages have supported the adoption of the current friction model.

System constraints

The operating limitations presence in the plant shall be included in the simula-
tion model or addressed in the controller design. The clamp force is limited at
60 Chapter 3. Experimental setup and model descriptions

or below3 30 kN and the motor torque is constrained at ±3 Nm to avoid over-

heating the motor windings. Furthermore, the maximum quadrature voltage
is constrained at ±42 V, and consequently imposes a limit for the maximum
motor velocity at ±300 rad/s. Finally, the simulation model (3.6)–(3.11) in-
corporated with the quadrature voltage constraint is realised using Simulink
as shown in Figure 3.11.

3.2.3 Summary
This section begins with a brief survey on high-order EMB models in the
view of selecting a suitable model fidelity used for controller designs. This is
followed by the development of a simulation model, presented in (3.6)–(3.11).
Subsequently, model parameter identification is executed, in which the routines
are summarised in Procedure 3.1.

Procedure 3.1 (Estimation of EMB Parameters). The parameters for the

EMB simulation model (3.6)–(3.11) are identified using the following steps:

1. The gear ratio, N , lumped moment of inertia, J, motor torque constant,

Kt and armature circuit parameters Lm , Rm are taken from manufac-
turer’s data.
2. Determine the stiffness curve, Fcl (θm ) as a function of motor position.
3. Determine the friction parameters by performing three distinctive tests4 :

a) Break-away tests without clamping: resolve the torque required to

overcome static friction, Ts0 ,
b) Break-away tests with clamping: resolve the load-dependent static
friction coefficient, µs .
c) Constant velocity tests in air-gap: resolve Coulomb friction coeffi-
cient, TC , viscous friction coefficient, Tv and the characteristic ve-
locity of the Stribeck friction, ωs in the absence of load torque.


3
While the design limit for clamp force is at 40 kN, the 30 kN limit is deliberately chosen
in this work to avoid overheating the motor during the initial development stage.
4
Some of the preliminary tests were performed with assistance of Caroline Jekelius, an
exchange student studying towards a Dipl.-Ing. at TU Munich, Germany.
3.2. Modelling of an electromechanical brake 61

Mechanical

Electrical

Figure 3.11: Simulink model of an EMB.

62 Chapter 3. Experimental setup and model descriptions

The EMB power transmission unit incorporates a planetary gear set and a
ball screw. Using the components’ geometry obtained from proprietary design
information, the gear ratio can be found. Furthermore, the lumped moment
of inertia about the motor’s axis of rotation is computed using the CAD data
of the EMB. The motor torque constant, armature inductance and armature
resistance are inferred from manufacturer’s data. Furthermore, the parameters
for the stiffness curve and friction model are identified using Step 2 and Step 3
of Procedure 3.1. The identified EMB parameters are tabulated in Table 3.1.
Note that the model parameters for stiffness and friction are estimated
separately using specifically designed control maneuvers that allow isolation of
a particular set of parameters. For example, the procedures are divided into
four different tests: the stiffness curve identification, two sets of break-away
tests and the constant velocity tests, where the first test identifies the stiffness
curve parameters, and the latter three tests identify the friction model param-
eters individually. Also notice that these experiments are run in steady state
conditions. While these routines have successfully characterised the stiffness
and friction properties of the EMB, they are lengthy and time consuming. For
instance, the break-away tests without clamping and with clamping each took
over four hours to complete, while the constant velocity tests required approx-
imately half an hour to run. Although human labour can be reduced consider-
ably by using computer scripts that automates the tests5 , it is envisaged that
a fast parameter identification procedure is required during production and
maintenance, where rapid quality control checks and model-based controller
tuning are conducted.
To check the accuracy of the estimated parameters and the simulation
model, the parameterised model for the mechanical plant is validated against
independently obtained experimental data. The measured motor torque is
used as the input for the simulation model to produce outputs in clamp force
and motor velocity. Comparison of experimental measurements and model
predictions is illustrated in Figure 3.12. The experimental and simulation
results are qualitatively matching, however, slight overshoots are observed in
the simulated responses. Note that the measured motor torque, which is also
the model input is contaminated with measurement noise. Since the input
measurement noise is not pre-filtered, it is not surprising to find the small
5
All tests are automated using scripts written in Vector CANape.
3.3. Development of a control-oriented EMB model 63

Table 3.1: EMB model parameters obtained using Procedure 3.1.

Parameter Value Unit
J 2.91 × 10−4 kg·m2
Kt 4.65 × 10−2 N·m/A
Lm 5.64 × 10−5 H
N 2.63 × 10−5 m/rad
Rm 5.00 × 10−2 Ω
Stiffness:
k1 −3.54 × 10−1 N/rad3
k2 3.52 × 10 N/rad2
Friction:
TC 7.67 × 10−2 N·m
Ts0 4.68 × 10−2 N·m
Tv 4.46 × 10−4 N·m·s/rad
µs 3.55 × 10−5 N·m/N
ωs 2.89 × 10 rad/s
Constraints:
Fcl max 30000 N
Tm max 3 N·m
vq max 42 V
ωm max 300 rad/s

discrepancies between the experimental and simulation results. Furthermore,

since this is an open-loop comparison, any modelling error will also lead to an
offset. The discrepancy may be reduced by recalibrating the model parameters
using transient responses.

3.3 Development of a control-oriented

electromechanical brake model
Motivated by the need of a low-order continuous model for controller design, a
control-oriented model that approximates the dynamics of the current control
loop and the EMB actuator is proposed. The approximation is illustrated by
Figure 3.13.
Note that the friction model (3.9) is discontinuous at ωm = . Furthermore,
if  → 0, then it is discontinuous at zero velocity due to the presence of
the sgn(·) function. To impose the smooth property on the friction model, a
64 Chapter 3. Experimental setup and model descriptions

4
x 10
3

2
Fcl [N]

1 Experiment
Model
0
0 0.05 0.1 0.15 0.2 0.25 0.3
400
ωm [rad/s]

200

−200
0 0.05 0.1 0.15 0.2 0.25 0.3
4
Tm [Nm]

−2
0 0.05 0.1 0.15 0.2 0.25 0.3
Time [s]

Figure 3.12: Comparison of experimental measurements and model predictions

for a 10 kN to 20 kN step response.

Approximated using the control-oriented EMB model (Σemb )

Current Controller

1
s kci +
EMB Fcl
1 iq,r vq
Tm,r Kt +
Actuator
(Σemb,sim )
−
iq kcp +

Figure 3.13: Approximating the dynamics of the current control loop and the
EMB actuator using a control-oriented model.
3.3. Development of a control-oriented EMB model 65

smooth friction model (Makkar et al., 2005) is adopted, given by

!
Ts − TC
T̄f (θm , ωm ) = Tv ωm + TC + tanh(εωm ), (3.12)
1 + (ωm /ωs )2

where ε is a sufficiently large number chosen to match the friction model to

the experimentally obtained measurements. Compared to the original friction
model Tf (·) (3.9), the smooth friction model T̄f (·) replaces the discontinuity
of the sgn(·) function by the tanh(·) function of finite slope, hence T̄f (·) is
continuous and continuously differentiable with respect to the states.
As commonly observed, the mechanical dynamics are relatively slow com-
pared to the electrical dynamics. In the simulation model (3.6), the first two
state equations are the mechanical torque equation, which have slow dynamics
compared to the third state equation corresponds to the electric transient in
the armature circuit. In addition, the current control loop dynamics has sim-
ilar time-scale as the armature circuit. This suggests a possibility of utilising
the time-scale separation to obtain a reduced-order model, approximating the
closed-loop dynamics of the current controller and the EMB actuator. To this
end, a two-step model order reduction approach based on the singular pertur-
bation argument is employed, which is led by the approximation of the electric
transient dynamics using a quasi-static model, followed by the reduction of the
current controller dynamics.
The dynamics of the current controller coupled with the EMB actuator
model with smooth functions can be represented by a fourth-order state equa-
tion
dθm
= ωm ,
dt
dωm
J = −Tl (θm ) − T̄f (θm , ωm ) + Kt iq ,
dt (3.13)
diq
Lm = −Rm iq − Kb ωm + vq ,
dt
deo
= iq,r − iq ,
dt
where the load torque, Tl (·) is given by (3.7)–(3.8) and the friction torque,
T̄f (·) is given by (3.12). The quadrature voltage, vq is regulated by the current
controller, which is in the form of a PI controller with integrator state, e0 , and
the controller output
vq = kci e0 + kcp (iq,r − iq ). (3.14)
66 Chapter 3. Experimental setup and model descriptions

Prior to the application of singular perturbation technique, the closed-loop

model (3.13) has to be in standard form (Khalil, 2002). To this end, the
following dimensionless variables are defined:
ωm Rm Rm
ωr = , ir = iq , ir,r = iq,r . (3.15)
Ω Kb Ω Kb Ω
With substitution of (3.14) and (3.15), the state equation (3.13) is rewrote as

dθm
tmech = tmech ωm ,
dt
dωr Rm
tmech =− (Tl + T̄f ) + ir ,
dt Kt Kb Ω
(3.16)
dir Rm + kcp kci kcp
telec =− ir − ωr + e0 + ir,r ,
dt Rm Kb Ω Rm
deo tmech Kb Ω
tmech = (ir,r − ir ),
dt Rm
where tmech = JRm /(Kt Kb ) is the mechanical time constant and telec = Lm /Rm
is the electrical time-constant. Using the model parameters given in Table 3.1,
tmech and telec are 1.01 × 10−2 and 1.13 × 10−3 respectively. Since tmech  telec ,
tmech is considered as the time unit and the state equation (3.16) is rewrote
with respect to a dimensionless time variable τr = t/tmech as

dθm
= tmech Ωωr , (3.17a)
dτr
dωr Rm
=− (Tl + T̄f ) + ir , (3.17b)
dτr Kt Kb Ω
telec dir Rm + kcp kci kcp
=− ir − ωr + e0 + ir,r , (3.17c)
tmech dτr Rm Kb Ω Rm
deo tmech Kb Ω
= (ir,r − ir ). (3.17d)
dτr Rm
Equation (3.17) is in the standard form if the following dimensionless pa-
rameter is introduced:
telec Lm Kt Kb
ε1 = = 2
tmech JRm
Note that the dynamics of dir /dτr is fast when ε1 is small, therefore leading to
a rapid convergence of ir to the equilibrium of (3.17c). By setting ε1 to zero,
the equilibrium of (3.17c) is procured as
!
Rm kci kcp
ir = −ωr + e0 + ir,r . (3.18)
Rm + kcp Kb Ω Rm
3.3. Development of a control-oriented EMB model 67

Hence, the third-order reduced model is therefore given by (3.17a), (3.17b),

(3.17d) and (3.18).
The second stage of the reduction considers the elimination of the current
controller dynamics. By substituting (3.18) into (3.17d) and rearranging terms,
the current controller dynamics is represented by

tcc deo Kb Ω
= −e0 + (ωr + ir,r ), (3.19)
tmech dτr kci

where tcc = (Rm + kcp )/kci is the current controller time constant. Using the
Rm value provided in Table 4.3, the proportional gain kcp = 0.3385 and the
integral gains kci = 199, tcc can be calculated as 1.95 × 10−3 , thus tcc  tmech .
Introducing a dimensionless parameter

tcc (Rm + kcp )Kt Kb

ε2 = = ,
tmech JRm kci

which can be made small by proper selection of kcp and kci , the equilibrium of
(3.19) can be computed as

Kb Ω
e0 = (ωr + ir,r ). (3.20)
kci

By substituting (3.20) into (3.18), it can be shown that

ir = ir,r .

Furthermore, follow from (3.15), it can be shown that

iq = iq,r , (3.21)

hence leading to the approximation of Tm = Tm,r .

Finally, the resultant model is a second-order state equation consists of
(3.17a), (3.17b) and (3.21). It has been shown using singular perturbation
arguments that the current controller dynamics and the electrical transients
can be neglected. In terms of the physical state variables, the control-oriented
EMB model is given by

 ẋ = f (x, Tm ), x(0) = x0
Σemb :  (3.22)
y = h(x),
68 Chapter 3. Experimental setup and model descriptions

where x := [θm , ωm ]T is the state vector, x0 is a given initial condition of the

state vector, and f (·), h(·) are given by
 
ωm
f (x, Tm ) =  1   ,
J
−N F cl (θm ) − T̄ f (θm , ωm ) + T m
  (3.23)
Fcl (θm )
h(x) =  .
ωm
Additionally, the input constraint is |Tm | ≤ Tm max and the output constraints
are Fcl (θm ) ≤ Fcl max and |ωm | ≤ ωm max .

Discrete model. To facilitate analysis in discrete-time, a discretised version

of the control-oriented EMB model (3.22) is proposed:
  
 xd (k + 1) = fd xd (k), Tm (k) ,
 xd (1) = x0
Σemb,d :   (3.24)

 yd (k) = h xd (k) , k = 1, 2, . . . , n

where xd (k) = [θm (k), ωm (k)]T is the state vector at time step k and x0 is
a given initial condition of the state vector. Employing the forward-Euler
method, fd (·) can be calculated as
 
fd = xd (k) + ts f xd (k), Tm (k), θ , (3.25)

where ts is the step size. The function f (·) and h(·) are given in (3.23).
Alternative discretisation methods, such as the family of Runge-Kutta methods
can also be used to derive alternative forms of (3.25).
Note that three EMB models have been presented from Section 3.2 to Sec-
tion 3.3, namely the simulation model Σemb,sim in (3.6), the control-oriented
model Σemb in (3.22), and the discretised model Σemb,d in (3.24). The step
responses and sinusoidal responses of these three models are compared in Fig-
ure 3.14. The responses of the three models visually overlap with each other.

3.4 Modelling of the characteristics of brake

torque variation
The aim of this section is to construct a grey-box model of the brake torque
variation (BTV). As introduced in Chapter 1, BTV is caused by the geo-
metrical irregularities in the brake rotor arise from imperfect mounting, bad
3.4. Modelling of the characteristics of brake torque variation 69

4
x 10
4
Full−order nonlinear model

Fcl [N]
Reduced−order model
2 Discrete model

0
0 0.05 0.1 0.15 0.2 0.25 0.3
400
30 kN Step
ωm [rad/s]

200 20 kN Step
0
10 kN Step
−200
0 0.05 0.1 0.15 0.2 0.25 0.3
5
Tm,r [Nm]

0
10 kN Step 30 kN Step
20 kN Step
−5
0 0.05 0.1 0.15 0.2 0.25 0.3
Time [s]

(a) Step responses for 10 kN, 20 kN and 30 kN.

4
x 10
4
Full−order nonlinear Reduced−order Discrete
Fcl [N]

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
400
ωm [rad/s]

200

−200
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
5
Tm,r [Nm]

−5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time [s]

(b) Sinusoidal responses. Grey lines denote responses for 10 + sin (20π) kN and
black lines denote responses for 20 + sin (20π) kN.

Figure 3.14: Simulated responses of the nonlinear simulation model, the

reduced-order control-oriented model and the discretised model.
70 Chapter 3. Experimental setup and model descriptions

Figure 3.15: Setup for disc rotor runout measurements. The individual parts
are: 1. Brake disc rotor with runout and thickness variation; 2. Dial indicator.

machining, uneven wear, uneven corrosion and uneven friction film generation.
Typically, these irregularities occur fewer than five times per rotor revolution,
with the most threatening case occurs in rotor with one or two blemishes.
This thesis deals with the compensation of low-order BTV that causes
the most common type of brake judder, which is called cold judder. The
cold judder typically occurs when the brake temperature is less than 100◦ C.
Additionally, the low-order BTV is due to the uneven wear of disc rotor, where
the level of evenness can be determined from runout and thickness variation
(DTV). For most manufacturers, the manufacturing tolerances for runout and
DTV in a disc rotor are less than 80 µm and 10 µm respectively (de Varies
and Wagner, 1992).
The disc rotor employed for experiments in this work is illustrated in Fig-
ure 3.15 and was obtained from an in-service vehicle with brake judder. Using
a dial gauge, runout for both surfaces of the rotor was measured, where the
measurements are shown in Figure 3.16. The maximum runout for the inner
surface and the outer surface are 87 µm and 97 µm respectively, demonstrating
higher values than the designed tolerance.
DTV is computed from the difference in runout of both surfaces, which is
3.4. Modelling of the characteristics of brake torque variation 71

0.1

Inner face runout [mm] 0.05

−0.05

−0.1
0 50 100 150 200 250 300 360

0.1
Outer face runout [mm]

0.05

−0.05

−0.1
0 50 100 150 200 250 300 360
Disc Rotation [deg.]

Figure 3.16: Runout measurements for inner and outer faces of the brake disc.

shown in Figure 3.17. The maximum DTV is approximately 18 µm, which is

significant enough to cause judder when included in a vehicle chassis. When
brake force is applied to the rotating brake rotor, it is inevitable that the DTV
will produce a variation in braking torque. Additionally, the effect due DTV
on a rotating rotor is clearly periodic with respect to the angular position. By
formulating the DTV xdtv (θd ) as a function of the disc rotor position θd , it can
be modelled using the following n-th order Fourier series:
n
xdtv (θd ) = ad,k cos (kθd + φd,k ), (3.26)
X

k=1

where the amplitude of the k-th harmonic is ad,k and the phase-shift is φd,k .
The components of different frequencies in the DTV waveform can be de-
termined using the fast Fourier transform (FFT) method. Employing the fft
function in Matlab, the Fourier coefficients are computed, which are depicted
in Figure 3.17. It is evident that the first harmonic has the largest component,
followed by smaller amplitudes of the second and the third harmonics. Util-
ising the Fourier coefficients, various orders of approximation can be made.
For example, the m-th order of approximation is computed using (3.26) up to
72 Chapter 3. Experimental setup and model descriptions

(a) 0.02

0.015

0.01

0.005
DTV [mm]

−0.005
Measurement
−0.01 1st approx.
−0.015 2nd approx.
3rd approx.
−0.02
0 50 100 150 200 250 300 360
(b) Disc Rotation [deg.]
0.015
Amplitude [mm]

0.01

0.005

0
0 1 2 3 4 5 6
Harmonics

Figure 3.17: (a) Disc thickness variation measurements versus the first three or-
ders of approximation; (b) Fourier transformation of the measurements shows
that the signal is dominated by the first three harmonics.

n = m. These approximations are plotted and compared with the measured

DTV, where all approximations resemble closely to the measurements.
The brake torque experienced by a rotating disc rotor, Tb is modelled as

Tb = 2µrd Fcl + Tbtv (θd ), (3.27)

where 2µrd Fcl represents the mean brake torque, and Tbtv (θd ) represents the
variation in brake torque induced by rotor geometrical irregularities. In order
to examine the variations of brake torque, the disc rotor with runout and DTV
was rotated at one hertz with 750 kN of clamp force applied. The brake torque
measurements are shown in Figure 3.18, demonstrating a periodic oscillation
in brake torque. Note that the brake torque can be approximated using an
n-order Fourier series, given by
n
Tb = ab,0 + ab,k cos (kθd + φb,k ), (3.28)
X

k=1

where ab,0 is the average brake torque, which corresponds to

ab,0 = 2µrd Fcl ,

3.4. Modelling of the characteristics of brake torque variation 73

(a) 160
Mea. DC 1st. 2nd. 3rd.
150
Brake Torque [Nm]
140

130

120

110

100

90
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
(b) Disc Rotations
15
Amplitude [Nm]

10

0
0 1 2 3 4 5 6
Harmonics

Figure 3.18: (a) Brake torque variation measurement versus the mean and
the first three orders of approximations; (b) Fourier transformation of the
measurements shows that the signal is dominated by the mean and the first
three harmonics.

while the sum of the periodic components is the BTV, given by

n
Tbtv (θd ) = ab,k cos (kθd + φb,k ). (3.29)
X

k=1

Similar to the analysis conducted on the DTV waveform, the FFT is applied
to the brake torque measurements to calculate the components of different
frequencies. Apart from the mean, which has the largest amplitude given by
ab,0 = 112.85 Nm, the first harmonic has the second largest component at ap-
proximately 15 Nm. This is followed by smaller components of the second-order
and the third-order harmonics. The second-order and the third-order approxi-
mations are very close and visually coincide, and they match the measurements
well. Furthermore, utilising the effective rotor radius of rd = 127 mm and the
mean of measured clamp force, Fcl = 732.28 N, the effective coefficient of
friction can be calculated as µ = 0.607.
The measurements of the DTV and BTV clearly indicate that they are peri-
odic with respect to disc rotor rotations. Therefore, the periodic BTV function
74 Chapter 3. Experimental setup and model descriptions

established in (3.29) will be used in the latter chapters for compensators syn-
thesis. However, the amplitude and phase-shift of the BTV are unknown a
priori but are assumed to be time-invariant, and they have to be adapted by
the proposed compensators. In practice, these quantities may vary slowly rel-
ative to the disc rotations. Since both of the DTV waveform and the BTV
measurements can be well approximated by their respective second-order ap-
proximations, therefore BTV compensators of up to two orders are sufficient
in implementation.

3.5 Conclusions
An overview of the experimental facilities that are used in this research are
presented in this chapter. Subsequently, the modelling work is started in Sec-
tion 3.2 with a review on high-order models of EMB in the interest of deciding
a suitable model fidelity for control performance simulations. An EMB model
used for simulation is then presented in (3.6), which comprises of the torque
equation evaluated around the motor rotational axis and the electrical circuit
dynamics for quadrature current. The overall nonlinear stiffness characteristics
and nonlinear friction of the EMB calliper are included in the model. A model
identification routine summarised in Procedure 3.1 is presented to determine
the characteristics of stiffness and friction, and also to estimate the parame-
ters contained in these models. The routine is designed to run in steady-state
conditions, and therefore it is lengthy and time-consuming. It demanded nine
hours of data collection even with the scripted automatic testings, plus addi-
tional hours for post-processing of data. Although this routine has successfully
parameterised the stiffness model and the friction model, it is not applicable
for post-production quality control checks and quick model-based controller
tunings.
To facilitate controller developments, a reduced-order nonlinear control-
oriented model for the EMB is presented in Section 3.3. The proposed control-
oriented model incorporates the dynamics of the current controller and the
dynamics of the EMB actuator. Since the dynamics of the current controller
and the dynamics of the electrical circuits in the EMB actuator are relatively
fast compared to the mechanical components, singular perturbation technique
is employed to justify the reduction of the current controller and electrical
3.5. Conclusions 75

Discretised control-oriented model Σemb,d , (3.24) in Sec. 3.3

Control-oriented model Σemb , (3.22) in Sec. 3.3
Brake Torque
Input from or
θd Variation Tbtv
output to the Tbtv , (3.29)
controllers in Sec. 3.4

+
EMB Fcl Brake Disc +
1 Current vq Actuator with Nominal Tb
Tm,r Kt + Controller
Σemb,sim , Thickness
− (3.6) in
iq Sec. 3.2

ωm
Fcl
Tb

Figure 3.19: A block diagram illustrating the models that contribute to brake
torque generation.

dynamics from the model. The resultant model is reduced to a second-order

nonlinear model, and will be used as the core model for model-based controller
design in later chapters. Additionally, a discretised version of the control-
oriented model is also presented to facilitate analysis in discrete-time.
In Section 3.4, the braking characteristics of a disc rotor with geometry
irregularities is investigated using the dedicated brake dynamometer. The
runout and the disc thickness variation (DTV) are measured, which indicate
high values than the permissible manufacturing tolerances and are significant
to cause judder when included in a vehicle chassis. The brake torque variation
(BTV) resulted from applying a clamping force on the rotating brake disc is
also measured and analysed. The BTV measurements are clearly periodic with
respect to the rotor rotations. Therefore, a grey-box BTV model in the form
of a Fourier series is suggest in (3.29). This grey-box model will be used in
BTV compensators design, where the amplitude and phase-shift of the model
are adapted. Furthermore, the BTV waveform can be well-approximated using
the first two harmonics, indicating that the grey-box model with an order to
two is sufficient in implementations.
Finally, the models developed in this chapter that contribute to the brake
torque generation are summarised in a block diagram illustrated in Figure 3.19.
The relevant section numbers and equation numbers are notated in the dia-
gram.
 4

Rapid electromechanical
brake model parameter
identification

In order to support quick quality control checks and rapid model-based con-
troller tunings post-production, a fast and accurate parameter identification
procedure is desirable. However, the parameter identification routine sum-
marised in Procedure 3.1 is lengthy and time consuming, where it required
nine hours of data collection even with automated testings, in addition to
extra hours for post-processing of data. Additionally, measurement noise is
inevitable in parameter identification experiments, which leads to estimation
inaccuracies. These considerations motivate the development of a rapid model
parameter identification procedure for an electromechanical brake (EMB) that
incorporates noise models.
The design of the rapid parameter identification procedure is commenced in
two stages. First, the optimal experimental design is conducted in Section 4.1.
This involves planning an optimal output trajectory to minimise the time re-
quired for experiments needed to infer model parameters from input-output
measurements. This trajectory has to be achievable by complying to the input
and output constraints. In the second stage, parameter estimation algorithms
are devised in Section 4.2 to identify model parameters from the input-output
measurements. Two estimation methods, namely the output error method

77
78 Chapter 4. Rapid EMB model parameter identification

and the prediction error method are formulated. Their accuracy and ease of
implementation are also compared and discussed. Finally, the parameterised
EMB model is validated against independently obtained experimental mea-
surements.

4.1 Design of an optimal trajectory for

parameter identification
4.1.1 Problem formulation
An optimal experimental design is performed in order to plan an output tra-
jectory utilised for inferring the model parameters. The changes of gear ratio
N and the lump inertia J are expected to be insignificant over time. On the
other hand, the stiffness and friction characteristics are prone to large changes
due to wear and operating conditions. Therefore, during the rapid parameter
estimation process, N and J are assumed to be fixed to reduce dimensionality
and only the parameters in the stiffness model and friction model are to be
estimated.
Since experimental measurements are sampled at discrete time, a discrete
mechanical model is considered. In order to take account of measurement
noises, the discrete model (3.24) is modified as follows:
 
(k + 1) = (k), (k), + Bv vi (k), xd (1) = x0



 x d f d x d Tm θ

  
Σemb,d : yd (k) = h xd (k), θ , k = 1, 2, . . . , n (4.1)

  
ym (k) = h xd (k), θ + vo (k),




where xd (k) := [θm (k), ωm (k)]T is the state vector at time step k and x0 is
a given initial condition of the state vector. The functions fd (·) and h(·) are
given in (3.25) and (3.23) respectively. The matrix, Bv associated with the
process noise is given by  
0
Bv =  1  . (4.2)
J
The plant dynamics are affected by the process noise, vi (k) and the associated
matrix Bv , while the output measurement ym (k) is corrupted by the output
measurement noise, vo (k). Note that the EMB parameters,
h i h i
θ = θ1 θ2 · · · θ7 = k1 k2 TC Ts0 Tv µs ωs .
4.1. Design of an optimal trajectory for parameter identification 79

are now included as function arguments in the discretised model (4.1).

For the purpose of experimental design, it is assumed that there exists an
a-priori estimation of the model parameters, θ. The values listed in Table 3.1
are used as the initial estimates of θ in the section. Furthermore, the process
noise, vi (k) is assumed to be zero during experimental design, while the output
measurement noise, vo (k) is assumed to be zero-mean and has a Gaussian
distribution with covariance R, that is vo ∼ N (0, R). Note that the covariance
can be defined as:
n o
E v(k1 )v(k2 )T = R(k1 , k2 ), k1 = 1, 2, . . . , n
(4.3)
k2 = 1, 2, . . . , n

where E{·} denotes the expected value. By exploiting (4.3), an estimate for
the variance of the parameter estimation error can be calculated from (Raol
et al., 2004):
1X n   T
R= ym (k) − yd (k) ym (k) − yd (k) . (4.4)
n k=1
To obtain ym (·) and yd (·), an effort can be made to drive the plant using
an arbitrary chosen input trajectory {Tm (k), k = 1, . . . , n}, where the output
ym (k) are measured. Then, the model output, yd (·) can be computed from
(4.1) using the input sequence.
The task of parameter identification is to estimate θ from the knowledge
of the input and output sequences, which are denoted by U = {Tm (k), k =
1, . . . , n} and Y = {ym (k), k = 1, . . . , n} respectively. Since the true parame-
ters, θ∗ are unknown in real-life scenario, the exact parameter estimation errors
cannot be obtained. Alternatively, the relative accuracy of the estimation is
evaluated from the bias and the covariance of the estimated parameter. An
unbiased estimator with maximum likelihood is often assumed during the ex-
perimental design phase so that the it can be accomplished independently of
the estimator used (Mehra, 1974). Note that the covariance of any unbiased
estimator is bounded by the Cramér-Rao lower bound, given by
n o
E (θ − θ∗ )(θ − θ∗ )T ≥ M −1 , (4.5)

where the Fisher Information Matrix, M is defined by

 !T 
 ∂ log p(Y |U, θ) ∂ log p(Y |U, θ)
!

M =E , (4.6)
 ∂θ ∂θ 
80 Chapter 4. Rapid EMB model parameter identification

and p(Y |U, θ) denotes the conditional probability density, also known as the
likelihood function. Furthermore, the lower bound of the covariance can be
achieved by utilising the maximum likelihood estimator, which maximises the
conditional probability density:

p(Y |U, θ̂) = max p(Y |U, θ). (4.7)

θ

Assumption 4.1. The Fisher information matrix for the input output pairings
in the trajectory, M is nonsingular.

Remark 4.1. Note that a nonsingular M is necessary for the existence of the
Cramér-Rao lower bound matrix, and can be provided through appropriate
design of the experimental trajectory.
The experimental design can be approached by optimising the input se-
quence, U such that the resultant output trajectory yields a small covariance
in the estimated parameters. As established by the Cramér-Rao lower bound
(4.5), the lower bound of the maximum likelihood estimator covariance is given
by the inverse of the Fisher Information Matrix, M −1 . Therefore, a scalar
measure of M −1 can be used as a performance metric for optimising the tra-
jectory. To this end, the following scalar measures of performance based on M
are commonly considered (Mehra, 1974):

• A-optimality (average): minU tr (M −1 ) results in minimising the average

variance of the parameters, where tr(·) denotes the trace.
• D-optimality (determinant): minU det (M −1 ) results in minimising the
volume of the uncertainty ellipsoids.

The D-optimal design has a notable advantage over the A-optimal design, in
which it is invariant under scale changes in the parameters and linear trans-
formations of the output (Mehra, 1974). This attractive feature has motivated
the adoption of D-optimal design in this work.
Note that the EMB model have magnitudes differing by factors of up to
106 . To facilitate meaningful comparison between variances and to avoid nu-
merical sensitivity during optimisation, parameters and outputs are scaled.
The normalised parameter vector, θ̄ = [θ̄1 , θ̄2 , . . . , θ̄7 ]T is introduced, which is
defined by
θi
θ̄i = , i = 1, 2, . . . , 7 (4.8)
θnom,i
4.1. Design of an optimal trajectory for parameter identification 81

where the original parameters in SI units, θ := [k1 , k2 , TC , Ts0 , Tv , µs , ωs ]T are

normalised to their respective nominal values, θnom,i taken from Table 3.1.
The output ranges in SI units are given by
n o
h := [Fcl , ωm ]T : Fcl ∈ [0, 30000], ωm ∈ [−300, 300] , where they differ by
two orders of magnitude. To avoid heavy weighting on one of the measured
outputs, h(·) is scaled using the following expression:
 
10000−1 0 
h̄ =  h. (4.9)
0 100−1

The scaled outputs h̄ have reduced ranges with the same order of magnitude,
n o
given by h̄ := [h̄1 , h̄2 ]T : h̄1 ∈ [0, 3] .
Furthermore, the scalar performance measure of the D-optimal design can
be converted into a convex function using log det (M −1 ) (Boyd and Vanden-
berghe, 2004). It can be shown that M is a positive semidefinite symmetric
matrix. If M is nonsingular, then − log det (M ) = log det (M −1 ) can be em-
ployed in order to avoid numerical issues during the matrix inversion when M
has a bad condition number.

Problem 4.1 (Optimal experimental design). Given a discretised system (4.1)

without process noise, a length of of experimental trajectory n and the Assump-
tion 4.1 is satisfied, find an input sequence U := [Tm (1), . . . , Tm (n)]T for the
optimal experimental trajectory that satisfies the following expressions:

min − log det (M ), (4.10a)

U

subject to xd (1) = x0 , (4.10b)

 
xd (k + 1) = fd xd (k), Tm (k), θ , k = 1, . . . , n
|Tm (k)| ≤ Tm max ,
Fcl (k) ≤ Fcl max , (4.10c)
|ωm (k)| ≤ ωm max . (4.10d)

Alternatively, the output constraints (4.10c) and (4.10d) can be incorpo-

rated in a penalty function, which is then augmented to the objective function
(Goodwin, 1971).
82 Chapter 4. Rapid EMB model parameter identification

The Fisher Information Matrix, M in (4.10) can be numerically approxi-

mated using the following expressions (Mehra, 1974; Raol et al., 2004):

n
!T !
dh̄ dh̄
M= −1
k = 1, 2, . . . , n (4.11a)
X
R ,
k=1 dθ̄ k dθ̄ k
!
dh̄ h i
= dh̄
dθ̄1
dh̄
dθ̄2
··· dh̄
dθ̄7
, (4.11b)
dθ̄ k
! ! !
dh̄ ∂ h̄ ∂ h̄
= xθ̄i ,k + , i = 1, 2, . . . , 7 (4.11c)
dθ̄i k
∂xTd k ∂ θ̄i k
! !
∂fd ∂fd
xθ̄i ,k+1 = x + , xθ̄i ,1 = 0. (4.11d)
∂xTd k θ̄i ,k ∂ θ̄i k

The arguments of fd (·) and h(·) are omitted and the time steps are written in
the subscripts for clarity in presentation. For completeness, the expressions in
(4.11) are expanded in Appendix B.

4.1.2 Results and discussions

The optimisation problem (4.10) can be solved using a wide range of optimi-
sation packages. For example in Matlab, derivative-based optimisation is
offered by the fmincon function and derivative-free optimisation is offered by
the ga function. A sufficiently small sampling period is chosen to reduce the
discretisation error while large enough to shorten evaluation time. The value
of ts = 0.001 seconds seems to achieve a balance between both requirements
and thus it is chosen in this work. The input is updated every 20th sample
to reduce the number of optimised variables. For the ease of repeatability, the
initial conditions of the experiment are selected to be zero clamp force and
zero velocity, where xd (1) = [0, 0]T . Additionally, the length of experiment
can also be freely chosen, which can be reflected by varying the n in (4.10).
The input and output trajectories of four optimal experimental designs
with varying length from 0.1 seconds to 0.4 seconds are plotted in Figure 4.1.
Regardless of the length of experiments, the clamp force trajectories resemble
the step responses, while the velocity transients consist of initial accelera-
tion, operation at maximum speed and deceleration to zero. The performance
measures of the optimal trajectories with varying lengths are illustrated in
Figure 4.2, where the criteria for A-optimal design and D-optimal design are
4.1. Design of an optimal trajectory for parameter identification 83

4
x 10
3

2
Fcl [N] 1
0.1 0.2 0.3 0.4
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
400
ωm [rad/s]

200
0
−200
−400
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
4
2
Tm [Nm]

0
−2
−4
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Time [s]

Figure 4.1: Optimal trajectories for parameter identification versus various

length of experiments.

presented. It is observed that the values of both criteria1 decrease with the
experimental length, implying a longer experiment provides better parameter
estimation accuracy. This is because a longer experiment allows more time for
full system dynamics to be excited and more data to be collected. However, a
decreased rate of improvement is observed, where the reduction in the perfor-
mance criteria is more significant for experiments shorter than 0.18 seconds.
This is because for shorter experiments, the maximum values of clamp force
transients increase significantly with a slight increase in experimental length.
However, for experimental length longer than 0.18 seconds, the maximum val-
ues of clamp force trajectories remain at approximately 23 kN, as shown in
Figure 4.1. To strive a balance between the estimation accuracy and the time
required for experiments, 0.3 seconds is chosen as the length of experiment,
and the corresponding optimal trajectory will be applied in the next subsec-
tion for parameter estimation. While the experimental length of 0.2 seconds
is adequate, a slightly longer experiment is chosen to account for situations
1
A few points for tr (M −1 ) are missing due to numerical problem faced during the
evaluation of M −1 as M has bad condition number.
84 Chapter 4. Rapid EMB model parameter identification

10 800
−log det (M)
0 tr (M−1) 700

−10 600

−20 500
−log det (M)

tr (M−1)
−30 400

−40 300

−50 200

−60 100

−70 0
0.1 0.15 0.2 0.25 0.3 0.35 0.4
Length of Experiment [s]

Figure 4.2: Performance measures of optimal trajectories versus the length of

experiments.

where softer brake pads are used and longer rise-times are expected compared
to the nominal case.
Since the proposed rapid parameter identification procedure is intended for
quick post-production quality control checks and model-based controller turn-
ing during maintenance, it is therefore presumed that a feedback controller,
possibly with unknown structure or gains has already been implemented. To
reduce the risk of damaging the hardware, the parameter identification is ex-
ecuted with feedback, where the identified optimal clamp force trajectory is
employed as a reference trajectory and tracking is handled by the existing feed-
back controller. The experimentally obtained input and output measurements
will then be used to infer the model parameters in the following subsection.

4.2 Parameter estimation

Following the determination of the optimal experiment, parameter estimation
methods are investigated in order to identify the model parameters from the
4.2. Parameter estimation 85

experimental measurements. Two parameter estimation methods are consid-

ered, namely the output error method and the prediction error method. These
two methods will be briefly introduced in the following, and then proceed to
the estimation results and discussions.

4.2.1 Output error method

The output error method assumes that the noise only contaminates the output
measurements, while there is no process noise. Similar to Section 4.1, the
output measurement noise, vo given in the discretised EMB model (4.1) is
assumed to be zero-mean with covariance matrix R, that is vo ∼ N (0, R).
The aim of the output error method is to minimise the error between
the measured and the model outputs by changing the model parameters θ
contained in fd (·) and h(·) in (4.1). Additionally, since the measured initial
condition is disturbed by noise, it has to be adjusted in order to reduce the
estimation error.

Procedure 4.1 (Parameter estimation using output error method). Given the
plant model (4.1) and a sequence of measured input U := [Tm (1), . . . , Tm (n)]T
and measured output Y := [ym (1), . . . , ym (n)]T , the OEM finds the model
parameters θ and the initial condition x0 in order to minimise the error between
the measured and the model outputs, which can be formulated as the following
(Raol et al., 2004):

1X n  T   n
min ym (k) − yd (k) R−1 ym (k) − yd (k) + log det (R),
x0 ,θ 2 k=1 2
subject to xd (1) = x0 ,
 
xd (k + 1) = fd xd (k), Tm (k), θ , k = 1, . . . , n
 
yd (k) = h xd (k), θ ,
1X n   T
R= ym (k) − yd (k) ym (k) − yd (k) .
n k=1
(4.12)


Note that the optimisation problem (4.12) can be regarded as a maximum

likelihood method, as the cost function corresponds to the negative logarithmic
of the likelihood function presented in (4.7). To reduce numerical difficulties
86 Chapter 4. Rapid EMB model parameter identification

during optimisation, the normalised parameters θ̄ defined in (4.8) and the

scaled output defined in h̄ (4.9) can also be adopted in (4.12).

4.2.2 Prediction error method

Similar to the output error method, the prediction error method also adopts
the view of maximum likelihood method, albeit with the additional considera-
tion of the process noise. Note that the effect of input measurement noise can
be included in the process noise. The noises are assumed to be Gaussian, given
by vo ∼ N (0, R) and vi ∼ N (0, Q). To identify the likelihood function in the
case of sequential observation, it is necessary to resolve the conditional distri-
bution of the output at time step k + 1 based on data available at time step
k. The determination of this conditional distribution is essentially a predic-
tion problem, where a prediction model has to be formulated to postulate the
prediction error. The estimation method that is based on a criteria that uses
the postulated prediction error is called the Prediction Error Method (Åström,
1980).
A one-step ahead prediction model can be constructed using the extended
Kalman filter (EKF). The EKF is the nonlinear version of the Kalman filter
that involves repeated linearisation of the model during evaluation. It can
be evaluated at two stages, namely the predict phase and the update phase.
During the predict phase, the states x̂ and the covariance of the states P are
predicted one-step ahead using measurements at k − 1, given by the following
equations:
 
x̂(k|k − 1) = fd x̂d (k − 1|k − 1), Tm (k − 1), θ ,
(4.13)
P (k|k − 1) = A(k − 1)P (k − 1|k − 1)A(k − 1)T + Bv QBvT ,

where A(k − 1) is the state transition matrix. These predictions are subse-
quently corrected in the update phase using measurements at k. This involves
the calculation of the Kalman gain, given by

K(k) = P (k|k − 1)C(k)T S(k)−1 , (4.14)

where S(k) denotes the residual covariance, obtained from

S(k) = C(k)P (k|k − 1)C(k)T + R. (4.15)

4.2. Parameter estimation 87

The states and covariance are then updated using the following expressions
with the Kalman gain:
h i
x̂(k|k) = x̂(k|k − 1) + K(k) ym (k) − h(x̂d (k|k − 1), θ) ,
h i (4.16)
P (k|k) = I − K(k)C(k) P (k|k − 1).

The state transition matrix A(k − 1) and the observation matrix C(k) are
defined to be the following Jacobians:
∂fd ∂h
A(k − 1) = , C(k) = , (4.17)
∂xTd x̂(k−1|k−1),Tm (k−1)
∂xTd x̂(k|k−1)

where their expansions can be found in Appendix B.

Procedure 4.2 (Parameter estimation using prediction error method).

Given the EKF (4.13)–(4.17) and a sequence of measured input U :=
[Tm (1), . . . , Tm (n)]T and measured output Y := [ym (1), . . . , ym (n)]T , the PEM
finds the model parameters θ, the initial condition x0 and the process noise
covariance Q in order to minimise the error between the measured and the pre-
dicted outputs, which can be formulated as the following (Raol et al., 2004):
1X n  T   n
min ym (k) − ŷd (k) S −1
ym (k) − ŷd (k) + log det (S),
x0 ,θ,Q 2 k=1 2
subject to x̂d (1|1) = x0 ,
P (1|1) = P0 ,
(4.13)–(4.17),
 
ŷd (k) = h x̂(k|k), θ , k = 1, . . . , n
(4.18)
where P0 is the estimate of the variance of the initial condition xd (1). 

4.2.3 Experimental results and discussions

Utilising the optimal experiment design derived in Section 4.1, the experiment
was repeated 20 times using the static test rig of the prototype EMB described
in Chapter 3. Employing the experimentally measured clamp force, motor
velocity and motor torque illustrated in Figure 4.3, model parameters are
estimated using the output error method and the prediction error method.
In order to provide a fair comparison, the parameters are normalised using the
expression (4.8) and the nominal values given in Table 3.1.
88 Chapter 4. Rapid EMB model parameter identification

4
x 10
3
Non−optimal experimental design
2
Fcl [N]

Optimal experimental design

1

0
0 0.05 0.1 0.15 0.2 0.25 0.3
400
ωm [rad/s]

200

−200
0 0.05 0.1 0.15 0.2 0.25 0.3
4
Tm [Nm]

−2
0 0.05 0.1 0.15 0.2 0.25 0.3
Time [s]

Figure 4.3: A total of 40 sets of measurements obtained using the optimal

experimental design and non-optimal experimental design are presented. The
accuracy of parameter estimation obtained using these two experimental de-
signs will be compared.

Output error method versus prediction error method. The nor-

malised estimated parameters are listed in Table 4.1 and also plotted in Fig-
ure 4.4 and Figure 4.5. Due to the presence of noise, the estimated parameters
obtained from each set of experiment are not exactly identical. Hence, the
means are calculated and compared. The means of the estimations given by
both methods match closely.
However, there are discrepancies between the estimations and the nominal
values. To check the validity of the estimations and the nominal values, simula-
tions results are compared against independently obtained experimental data.
From the step responses illustrated in Figure 4.6, the simulated responses using
parameters obtained from the output error method and the prediction error
method visually overlap with the experimentally obtained responses, indicat-
ing a satisfactory estimation.
Additionally, simulated responses obtained using the nominal parameter
values are also plotted in Figure 4.6, and are compared to responses obtained
4.2. Parameter estimation 89

Table 4.1: Mean and variance of the normalised estimated parameters obtained
using the output error method (OEM) and the prediction error method (PEM)
with the optimal experimental design.
Normalised OEM PEM
Parameter Mean Variance Mean Variance
k1 /k1,nom 0.7985 0.004162 0.8438 0.004331
k2 /k2,nom 0.9015 0.003235 0.9363 0.000898
TC /TC,nom 1.6624 0.014324 1.6682 0.006034
Ts0 /Ts0,nom 0.9298 0.014876 0.9212 0.008480
Tv /Tv,nom 2.2158 0.059713 2.4498 0.065986
µs /µs,nom 1.0508 0.018334 0.8747 0.002355
ωs /ωs,nom 0.4261 0.012167 0.4063 0.026712

1.25 1.9
1.2 Est. Parameter
Mean 1.8
TC / TC,nom

1.1 50% Conf. Region

k2 / k2,nom

1.7
1
1.6
0.9
1.5
0.8
0.75 1.4
0.5 1 1.5 1.8 2 2.2 2.4 2.6 2.8
k1 / k1,nom Tv / Tv,nom

1.25 0.5
1.2
0.4
ωs / ωs,nom

1.1
µs / µs,nom

0.3
1
0.2
0.9
0.1
0.8
0.75 0
0.5 1 1.5 1.8 2 2.2 2.4 2.6 2.8
Ts0 / Ts0,nom Tv / Tv,nom

Figure 4.4: The normalised estimated model parameters obtained using the
output error method.
90 Chapter 4. Rapid EMB model parameter identification

1.25 1.9
1.2 Est. Parameter
Mean 1.8

TC / TC,nom
1.1 50% Conf. Region
k2 / k2,nom

1.7
1
1.6
0.9
1.5
0.8
0.75 1.4
0.5 1 1.5 1.8 2 2.2 2.4 2.6 2.8
k1 / k1,nom Tv / Tv,nom

1.25 0.5
1.2
0.4
ωs / ωs,nom
1.1
µs / µs,nom

0.3
1
0.2
0.9
0.1
0.8
0.75 0
0.5 1 1.5 1.8 2 2.2 2.4 2.6 2.8
Ts0 / Ts0,nom Tv / Tv,nom

Figure 4.5: The normalised estimated model parameters obtained using the
prediction error method.

using the newly acquired optimised parameters. It is evident that the simulated
responses obtained using the optimised parameters match the experimental
results better. This result indicates that the optimised parameters are more
accurate than the nominal values.
The performance of estimation can be compared using the variance of the
estimated parameters. The sum of variances given by the output error method
and the prediction error method are 0.0048024 and 0.0038791 respectively.
Since the sum of variances given by the prediction error method is smaller,
this suggests that it renders better estimations. This can also be examined by
comparing the 50% confidence region depicted in Figure 4.4 and Figure 4.5,
where the estimations provided by the prediction error method have smaller
50% confidence regions. This improvement is resulted from the consideration
of process noise in the prediction error method, where the covariance Q is
estimated to be 0.00020154, suggesting that the motor torque measurements
have errors with standard deviation of 0.014196 Nm. Note that the estimated
standard deviation in torque measurement error is relatively substantial com-
4.2. Parameter estimation 91

4
x 10
6
Exp. Nom. OEM PEM Non−opt.

Fcl [N]
4

0
0 0.05 0.1 0.15 0.2 0.25 0.3
400
ωm [rad/s]

200

−200
0 0.05 0.1 0.15 0.2 0.25 0.3
4
Tm [Nm]

−2
0 0.05 0.1 0.15 0.2 0.25 0.3
Time [s]

Figure 4.6: Validation of parameterised model using the nominal parameters

from Table 3.1 and the estimated parameters from Tables 4.1 and 4.2.

pared to the friction torque at very slow speed, as it has the same order of
magnitude compared to the break-away friction torque.
Further model validation is performed using experiment with sinusoidal
trajectory, as shown in Figure 4.7. There are some disagreements in the simu-
lated and the experimental responses. A closer look at the velocity responses
indicates that the model shows lockups when the velocity crosses zero while
the clamp force transient is at its peak. However, the velocity lockups are not
observed in the experimental measurements. This suggests that the response
is very sensitive at very low speed, as a small error in motor torque measure-
ment or minor estimation error in friction characteristics at very low speed
may result in velocity lookups, which then lead to significant deviation in the
clamp force trajectory. Nevertheless, in the period where velocity lookups do
not happen, for example in between 1.5 seconds to 1.8 seconds, the simulated
responses are qualitatively matching the measurements well.
The correction mechanism provided by the extended Kalman filter is signif-
icant for model response sensitive to the input measurement noise. Figure 4.8
92 Chapter 4. Rapid EMB model parameter identification

4
x 10
3
Experiment OEM PEM
2
Fcl [N]

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
200
ωm [rad/s]

−200
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
2
Tm [Nm]

−1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time [s]

Figure 4.7: Validation of parameterised models with sinusoidal trajectory.

shows the mean and covariance estimates provided by the extended Kalman
filter for the sinusoidal trajectory. In the clamp force response, the mean and
the region bounded by the variance are visually overlapping, implying a good
estimate is obtained. On the other hand, the variance of the estimated velocity
increases significantly at zero-crossings. However, due to the correction fea-
ture, lookup is not noticeable in the estimated velocity mean. Therefore, the
parameter estimation provided by the prediction error method is more robust
against input measurement noise. On the other hand, due to the inclusion of
extended Kalman filter, the prediction error method requires ten times more
computational time compared to output error method in this application.

Optimal experimental design versus non-optimal experimental de-

sign. As postulated in Section 4.1, improved estimation accuracy is expected
from the optimal experimental design. Utilising the non-optimal experiment
transients with equal experimental length, as shown in Figure 4.3, the output
error method is employed to estimate the model parameters. The estimates
are listed in Table 4.2 and plotted in Figure 4.9. Simulations are also run using
4.2. Parameter estimation 93

Figure 4.8: Kalman estimates of mean and variance for sinusoidal tracking
response.

Table 4.2: Mean and variance of the normalised estimated parameters obtained
using the output error method with the non-optimal experimental design.
Normalised Mean Variance
Parameter
k1 /k1,nom 0.8147 0.004884
k2 /k2,nom 0.8912 0.004399
TC /TC,nom 1.6093 0.009592
Ts0 /Ts0,nom 0.9251 0.019993
Tv /Tv,nom 1.6597 0.018759
µs /µs,nom 0.9616 0.021593
ωs /ωs,nom 0.3837 0.013261

the estimated parameters and plotted in Figure 4.6. Compared to the experi-
mental measurements, an overshoot of velocity is observed. Additionally, the
trajectories of clamp force and velocity deviate from the measurements after
0.2 seconds, hinting an underestimation of friction torque. The sum of vari-
ances of the estimated parameters is 0.0072916, which is significantly higher
compared to the optimal experimental design.
94 Chapter 4. Rapid EMB model parameter identification

1.25 1.9
1.2 Est. Parameter
Mean 1.8

TC / TC,nom
1.1 50% Conf. Region
k2 / k2,nom

1.7
1
1.6
0.9
1.5
0.8
0.75 1.4
0.5 1 1.5 1.4 1.6 1.8 2 2.2 2.4
k1 / k1,nom Tv / Tv,nom

1.25 0.5
1.2
0.4
ωs / ωs,nom
1.1
µs / µs,nom

1
0.3
0.9

0.8 0.2
0.75
0.5 1 1.5 1.4 1.6 1.8 2 2.2 2.4
Ts0 / Ts0,nom Tv / Tv,nom

Figure 4.9: The normalised estimated model parameters obtained using a non-
optimal experimental design and the output error method.

The major attractive feature offered by the optimal experimental design

and the parameter estimation proposed in this section is the speed and accu-
racy in obtaining a useful set of model parameters for an EMB. Compared to
the nine hours of data collection needed for the routine summarised in Proce-
dure 3.1, the rapid procedure proposed in this section only requires one minute
of data collection. This is because the full system dynamics is exploited in the
rapid procedure and transient responses are used to infer the model param-
eters, as opposed to the steady state responses used in the slower routine.
Moreover, the parameter estimations provided by the rapid procedure are also
more accurate, as demonstrated by the validation results in Figure 4.6.

Finally, the optimised model parameters provided in Table 4.3 will be used
in the following chapters for simulations and controller designs.
4.3. Conclusions 95

Table 4.3: Optimised EMB model parameters.

Parameter Nominal Optimised Unit
J 2.91 × 10−4 kg·m2
Kt 4.65 × 10−2 N·m/A
Lm 5.64 × 10−5 H
N 2.63 × 10−5 m/rad
Rm 5.00 × 10−2 Ω
Stiffness:
k1 −3.54 × 10−1 −2.98 × 10−1 N/rad3
k2 3.52 × 10 3.29 × 10 N/rad2
Friction:
TC 7.67 × 10−2 1.28 × 10−1 N·m
Ts0 4.68 × 10−2 4.31 × 10−2 N·m
Tv 4.46 × 10−4 1.09 × 10−3 N·m·s/rad
µs 3.55 × 10−5 5.40 × 10−5 N·m/N
ωs 2.89 × 10 1.17 × 10 rad/s
Constraints:
Fcl max 30000 N
Tm max 3 N·m
vq max 42 V
ωm max 300 rad/s

4.3 Conclusions
A rapid parameter identification procedure for the prototype automotive elec-
tromechanical brake (EMB) is proposed in this chapter, which is intended to
support post-production quality control checks and quick model-based con-
troller tunings. The proposed procedure parameterised the EMB model in-
troduced in Chapter 3. Compared to the incumbent approach summarised in
Procedure 3.1, which uses a few different experimental manoeuvres that isolate
the identification of certain sets of model parameters, the rapid approach uses
a single optimised experimental trajectory, therefore greatly reduces the time
required for parameter identification. Specific outcomes of this work include:

• The optimal experimental design is formulated in Procedure 4.1 and the

optimal experimental trajectory is determined with the aim of minimis-
ing the time spent for data collection while ensuring the measurements
provide sufficiently rich information about the plant dynamics — needed
for accurate estimation of parameters. The resultant optimal experimen-
96 Chapter 4. Rapid EMB model parameter identification

tal design has reduced the time required for experiments from nine hours
to one minute.
• The parameter estimation problem is formulated using both output er-
ror method and prediction error method to infer model parameters from
experimental measurements. Both methods explicitly considers the out-
put measurement noise, but the prediction error method also takes into
account of the process noise. It is observed that the prediction error
method provides a marginally better estimation accuracy compared to
the output error method but requires significantly longer evaluation time.
This may be an important consideration for practical implementation of
the EMB model parameter estimation.
• The parameterised model is validated against independently obtained ex-
perimental data. Compared to the parameters estimated in Section 3.2,
the simulated responses of the newly estimated parameters provide a
better match with real measurements. This indicates that the proposed
rapid parameter identification procedure is able to obtain more accurate
parameter estimations in shorter time. The optimised model parameters
are listed in Table 4.3 and they will be used for controller development
in subsequent chapters.
 5

High performance clamp

force tracking control
In this chapter, high performance clamp force tracking controllers for an elec-
tromechanical brake (EMB) are developed. The notion of high performance
refers to the ability to track a given reference closely. A high performance
clamp force controller is crucial for fulfilling demanding requests from safety
critical systems, such as the anti-lock braking system (ABS) and the electronic
stability control (ESC).
To establish a benchmark for the best tracking performance achievable by
the EMB, time-optimal control is examined in Section 5.1. The time-optimal
controller is expected to provide the quickest possible response while taking
into account of system constraints, including the motor torque saturation and
motor velocity limits. However, the time-optimal control is prone to chatter,
and as will be discussed shortly, is not suitable for implementation on the actual
hardware. The reason for its inclusion here is to provide a benchmark for the
system performance, and is used as a basis for a more robust implementation.
The robust near-time-optimal controller (NTOC) is introduced in Sec-
tion 5.2, and can be implemented on the actual hardware. An assessment
is made as to whether or not this technique gets sufficiently close to the per-
formance of the time-optimal control. Additionally, tracking performance of
the NTOC and the industry-tuned cascaded PI control are compared. Apart
from the standard tracking test, the NTOC is also assessed in terms of robust-
ness to changing plant dynamics and suitability for implementation on the

97
98 Chapter 5. High performance clamp force tracking control

application under investigation.

In order to facilitate a more general performance criteria that includes both
the tracking speed and control effort, an alternative control paradigm called
the model predictive control is investigated in Section 5.3. The conflicting
objectives between tracking error and control effort are discussed, and the
online implementability of the model predictive clamp force controller is also
assessed.

5.1 Time-optimal tracking control

In this section, the time-optimal clamp force control for an EMB is examined.
The motivation behind this investigation is to establish a benchmark for the
best possible tracking performance without overshoot in the face of actuator
constraints. To this end, the EMB-specific time-optimal tracking problem
with input and state constraints is formulated, followed by the derivation of
necessary conditions for time-optimal using Pontryagin’s minimum principle.
The feedback implementation of the controller is also presented, followed by
the simulation results. Finally, this section concludes with a discussion on the
practicality of this design.

5.1.1 Problem formulation

The controller design presented in this section is based on the control-oriented
model (3.22) described in Chapter 3. The model is reproduced below for
convenience:
ẋ = f (x, Tm ), x(0) = x0 (5.1)

where  
ωm
f (x, Tm ) =  1   . (5.2)
J
−N F (θ
cl m ) − T̄ (θ
f m , ωm ) + T m

The stiffness characteristics Fcl (·) and friction model T̄f (·) are given by (3.8)
and (3.12) respectively. The state vector consists of the motor position and ve-
locity, x := [θm , ωm ]T ∈ X and the input comprises the motor torque, Tm ∈ U,
where U is compact. Furthermore, f : X × U → R2 with f (0, 0) = [0, 0]T .
Note that the regions of X and U are dictated by the state and input con-
straints. The motor position is limited by the clamp force range. Additionally,
5.1. Time-optimal tracking control 99

the velocity saturates due to the maximum voltage and the maximum rotary-
current-frequency of the servo amplifier output. Therefore, the state region is
given by
n o
X = x = (θm , ωm ) ∈ R2 : Fcl (θm ) ∈ [0, Fcl max ], |ωm | ≤ ωm max . (5.3)

Furthermore, the torque is limited by the maximum current, hence

n o
U = Tm ∈ R : |Tm | ≤ Tm max . (5.4)

The control problem dealt with is the clamp force feedback control. It is
studied under the state feedback control framework. However, since only the
velocity and clamp force are available for feedback, the measured clamp force
is used to deduce the motor position. By utilising the one-to-one relationship
of the clamp force function to the motor position, the clamp force reference
and clamp force measurement are represented by the equivalent motor position
reference and measurement respectively. The corresponding state reference is
given by xr = (θm,r , ωm,r ).
In order to deliver demanding clamp force requests from higher level con-
trollers, such as the ABS and ESC during safety-critical operations, the clamp
force controller is expected to have quick tracking performance. This moti-
vates the utilisation of the time-optimal control, as it is expected to provide
the quickest possible tracking response.

Problem 5.1 (Time-optimal tracking). Given an initial state x(t0 ) = x0 and

a time-varying state reference xr (t), find a torque trajectory Tm (t) such that
the state coincides to the reference at tf , that is x(tf ) = xr (tf ), subjected
to the state and input constraints. The time-optimal control solution Tm∗ (t)
minimises the time taken to reach the reference from the initial condition. This
problem can be formulated as the following:
Z tf
min 1 dt,
Tm (t) t0

subject to (5.1)–(5.2),
Tm (t) ∈ U, (5.5)
x(t) ∈ X ,
x(t0 ) = x0 ,
x(tf ) = xr (tf ).
100 Chapter 5. High performance clamp force tracking control

The state constraint X and control constraint U are defined by (5.3) and (5.4)
respectively.

5.1.2 Time-optimal control using Pontryagin’s

minimum principle
The EMB-specific time-optimal clamp force tracking problem (5.5) is examined
using the Pontryagin’s minimum principle (PMP). To this end, the Hamilto-
nian is introduced as follows:
 
H x(t), Tm (t), p(t) = 1 + p1 (t)θ̇m (t) + p2 (t)ω̇m (t), (5.6)

where the dynamics of the states, θm and ωm are expressed in (5.1) and (5.2).
Furthermore, the Lagrangian L is defined by
   
L x(t), u(t), p(t) = H x(t), u(t), p(t)
   
+ µu1 (t)fcu1 Tm (t) + µu2 (t)fcu2 Tm (t)
   
+ µx1 (t)fcx1 ωm (t) + µx2 (t)fcx2 ωm (t) ,

where the multipliers µu1 , µu2 , µx1 , µx2 are non-negative. The constraints on
the motor torque, fcu1 , fcu2 can be expressed as
 
fcu1 Tm = Tm − Tm max ≤ 0,
  (5.7)
fcu2 Tm = −Tm − Tm max ≤ 0,

where t is omitted from the arguments for ease of presentation. Additionally,

the constraints corresponding to the velocity limits, fcx1 , fcx2 are
 
fcx1 ωm = −ωm + ωm max ≥ 0,
  (5.8)
fcx2 ωm = ωm + ωm max ≥ 0.

Furthermore, the dynamics of the costates, p1 , p2 are governed by

∂L
ṗ1 = −
∂θm
µs tanh (εωm )
!
p2  
= 3k1 θm
2
+ 2k2 θm N + , (5.9a)
J 1 + (ωm /ωs )2
∂L
ṗ2 = −
∂ωm
5.1. Time-optimal tracking control 101

" !
p2 Ts − TC  
= −p1 + Tv + ε TC + 1 − tanh (εω m ) 2
J 1 + (ωm /ωs )2
2ωm (Ts − TC ) tanh (εωm )
#
− − µx1 + µx2 . (5.9b)
ωs2 (1 + ωm
2 /ω 2 )2
s

The following proposition shows that the time-optimal solution comprises

intervals of either maximum acceleration or maximum deceleration when the
states are within the constraint boundaries.

Proposition 5.1 (Bang-bang control for unconstrained states). Consider the

time-optimal tracking problem (5.5) and assume that the state constraints are
not violated, |ωm | < ωm max . If Tm∗ (t) is the time-optimal control, it is necessary
 
that Tm∗ (t) = −Tm max sgn p∗2 (t) for p∗2 (t) 6= 0.

Proof. By substituting (5.1) and (5.2) into the Hamiltonian (5.6), proceeded
by the application of PMP (2.4), the following expression can be obtained:

p∗2  
1 + p∗1 ωm
∗
+ −N Fcl (θm
∗
) − T̄f (θm
∗ ∗
, ωm ) + Tm∗
J
p∗  
≤ 1 + p∗1 ωm + 2 −N Fcl (θm ∗
) − T̄f (θm
∗ ∗
, ωm ) + Tm , (5.10)
J
where the superscript ∗ denotes the optimal trajectory.
When p∗2 (t) > 0, the optimal control has to be Tm∗ (t) = −Tm max for (5.10)
to hold for all Tm (t) ∈ U. Conversely, when p∗2 (t) < 0, Tm∗ (t) = Tm max . Hence,
the time-optimal control input is given by
 
Tm∗ (t) = −Tm max sgn p∗2 (t) . (5.11)

Note that switchings of the time-optimal control (5.11) occur at isolated

instants when p∗2 (t) = 0. If there is a time interval t ∈ [t1 , t2 ], t0 ≤ t1 < t2 ≤ tf
during which p∗2 (t) = 0, then this interval is called the singular interval. It
can be shown that the singular interval does not exist for problem (5.5) in
the absence of state constraints (Kirk, 2004). However, the singular problem
may arise when the time-optimal trajectory reaches one of the state constraint
boundaries (Hartl et al., 1995). In the following, the constraint boundary is
shown to be a part of the optimal trajectory within the singular interval.
102 Chapter 5. High performance clamp force tracking control

Proposition 5.2 (Optimal trajectories which lie on the boundary of the state
constraint). Consider the time-optimal tracking problem (5.5). Assume that
the velocity constraint is reached, ωm (t) = ±ωm max and the optimal costate
p∗2 (t) = 0, for t ∈ [t1 , t2 ], t0 ≤ t1 < t2 ≤ tf . If Tm∗ (t) is the time-optimal control,
   
it is necessary that Tm∗ (t) = N Fcl θm ∗
(t) + T̄f θm ∗
(t), ωm
∗
(t) for t ∈ [t1 , t2 ].

Proof. Consider the singular interval case, where

p∗2 (t) = ṗ∗2 (t) = 0, t ∈ [t1 , t2 ], t1 < t2 (5.12)

It then follows from (5.9a) that

ṗ∗1 (t) = 0.

Furthermore, consider that the positive motor velocity constraint is reached,

ωm (t) = ωm max , then it can be shown using (2.9b) that

µx1 (t) 6= 0, µx2 (t) = 0. (5.13)

Substituting (5.12) and (5.13) to (5.9b) yields

p∗1 (t) = −µx1 (t) 6= 0.

In addition, evaluation of (2.10) gives

dH  ∗ ∗ ∗ 
x , Tm , p = ṗ∗1 ωm
∗
+ p∗1 ω̇m
∗
+ ṗ∗2 ω̇m + p∗2 ω̈m = 0. (5.14)
dt
Since p∗1 6= 0 and ṗ∗1 = p∗2 = ṗ∗2 = 0, the motor acceleration has to be zero, that
is ω̇m = 0 for (5.14) to hold. The corresponding motor torque that allows for
zero acceleration can be obtained from (5.1) and (5.2), given by
   
Tm∗ (t) = N Fcl θm
∗
(t) + T̄f θm
∗
(t), ωm
∗
(t) . (5.15)

When negative motor velocity constraint is reached, ωm = −ωm max , the proof
can be repeated to obtain the same time-optimal control solution (5.15).

It has been shown in Proposition 5.1 and Proposition 5.2 that the time-
optimal trajectory consists of maximum acceleration and/or deceleration, and
possibly an intermediate singular arc with constant maximum velocity. This
demonstrates that the available motor torque capacity is fully utilised, and the
motor velocity is pushed to the limit in order to achieve a rapid response.
5.1. Time-optimal tracking control 103

1500 −1.405
Reference
1000 −1.41
Fcl [N]

p1
500 −1.415

0 −1.42
0 0.02 0.04 0 0.02 0.04
6 0.05
θm [rad]

p2
0
2

0 −0.05
0 0.02 0.04 0 0.02 0.04
300
3
ωm [rad/s]

200

Tm [Nm]
100 0

0 −3

0 0.02 0.04 0 0.02 0.04

Time [s] Time [s]

Figure 5.1: Time-optimal step response without singular interval.

Solving the time-optimal control as a boundary value problem. The

process of finding the time-optimal control trajectory involves solving a bound-
ary value problem, where the conditions in (5.5), (5.9) and (5.11) have to be
satisfied by choosing the initial values p∗1 (t0 ) and p∗2 (t0 ). Additionally if the ve-
locity constraint is reached, then the period of the singular interval tsi = t2 −t1
has to be chosen, where the input during the singular interval is given by (5.15).
The boundary value problem can be tackled using the shooting point method
discussed in (Kirk, 2004).
A time-optimal response without singular interval is shown in Figure 5.1,
where state reference is given by (θm,r , ωm,r ) = (5.12, 0). The corresponding
clamp force reference is a constant 822 kN. Using the EMB model parameters
 
taken from Table 4.3 and the initial condition at θm (0), ωm (0) = (0, 10),
the time-optimal control trajectory are calculated. Both states meet their re-
spective references at the final time, tf = 0.0436 seconds. The costate p∗2
crosses zero at approximately 0.0232 seconds, where the control switching oc-
curs. Since the motor velocity is well within the constraints, singular interval
does not exist in this case.
104 Chapter 5. High performance clamp force tracking control

6000 −0.826
Reference
4000
Fcl [N]

p1
−0.828
2000

0 −0.83
0 0.02 0.04 0.06 0 0.02 0.04 0.06
15 0.02
θm [rad]

10 0

p2
5 −0.02

0 −0.04
0 0.02 0.04 0.06 0 0.02 0.04 0.06

300 3
ωm [rad/s]

Tm [Nm]
200 0
100
−3
0
0 0.02 0.04 0.06 0 0.02 0.04 0.06
Time [s] Time [s]

Figure 5.2: Time-optimal tracking response with singular interval.

A time-optimal response with a singular interval is illustrated in Figure 5.2,

where a time-varying state reference is given by (θm,r , ωm,r ) = (1.92+185t, 185).
 
Starting from the initial condition at θm (0), ωm (0) = (0, 10), both states
reach the reference at the final time, tf = 0.06 seconds. The motor veloc-
ity reaches the constraint at approximately 0.0312 seconds, where the time-
optimal control is switched from the maximum limit to a smaller magnitude
that yields constant velocity. During the singular interval, between 0.0312 sec-
onds and 0.0506 seconds, the optimal costate p∗2 remains zero, and p∗1 remains
constant. Finally at 0.0506 seconds, the motor torque switches to the min-
imum extremity and maximum deceleration occurs. Note that the costates
are continuous in time even though the trajectory of the multiplier µx1 (t) is
discontinuous at 0.0312 seconds and 0.0506 seconds, where µx1 (t) 6= 0 for
t ∈ [0.0312, 0.0506] seconds. As opposed to the step response where the fi-
nal velocity is zero, the final velocity of a general tracking response is not
necessarily zero, and it is required to meet the given velocity reference.
5.1. Time-optimal tracking control 105

5.1.3 Time-optimal feedback controller design

The time-optimal control trajectories have been solved numerically as a bound-

ary value problem utilising the shooting point method (Kirk, 2004). However,
this approach is impractical in real applications because the computational
time for solving a boundary value problem is likely to be long. Furthermore,
even if a solution is obtained, the open-loop implementation of the optimal
solution is sensitive to time delays and discrepancies in modelling and mea-
surements. This motivates the feedback implementation of the time-optimal
control, whereby the time-dependency of the optimal solution is rolled out,
replacing it with a state-dependent feedback control law. This property en-
hances the closed-loop robustness as the time delays are no longer affecting
the control performance.
The time-optimal clamp force tracking consists of at most one segment
of maximum acceleration or deceleration that leads to the final state. For a
given final state, xf = (θm,f , ωm,f ), the terminal segment can be calculated
by solving the equations of motion (5.1) and (5.2) backwards in time, which
is attainable by considering the negation of (5.2) and evaluating the ordinary
differential equation (5.1) for t ∈ [−ta , 0], ta > 0 using a standard numerical
software package, such as Matlab. The final state is employed as the initial
condition and the maximum motor torque Tm = ±Tm max is taken as input.
Several terminal segments are plotted in Figure 5.3. It is noted that the control
switching takes place when the time-optimal trajectory coincides to the ter-
minal segment. Henceforth, the terminal segment is also called the switching
curve. The control magnitude on the area to the left of the switching curve is
Tm = Tm max , while Tm = −Tm max for the area to the right of the switching
curve.
The time-optimal feedback control law can be devised utilising the switch-
ing curve and the instantaneous states. To this end, the motor position offset
before control switching is defined as

se = θm − φe (ωm , θm,f , ωm,f ), (5.16)

where φe (ωm , θm,f , ωm,f ) denotes the motor position where control switching
should occur, as illustrated in Figure 5.3. Additionally, the instantaneous
106 Chapter 5. High performance clamp force tracking control

300

200

100
ωm [rad/s]

−100

θm,f = 0 10 20 30 40 50
−200

−300

−10 0 10 20 30 40 50 60
φe [rad]

Figure 5.3: Time-optimal switching curves for θm,f = {0, 10, · · · , 50} rad. Solid
lines represent cases for ωm,f = 0 rad/s, and dash-dotted lines represent cases
for ωm,f = 200 rad/s. The velocity constraints are indicated by dash lines.

tracking errors for motor position and velocity are respectively defined by:

e1 = θm − θm,r , (5.17a)
e2 = ωm − ωm,r . (5.17b)

Note that the switching curves can be moved to a common origin as depicted
in Figure 5.4. Using (5.16) and (5.17), the time-optimal feedback control law
can be stated as:





Tm,cv if |ωm | = ωm max and se e2 < 0,

Tm,to = −Tm max sgn (e2 ) if sgn (se ) = 0, (5.18)


sgn (s ) otherwise,

−T

m max e

where Tm,cv represents the motor torque (5.15) that maintains constant velocity
at the velocity limit, given by

Tm,cv = N Fcl (θm ) + T̄f (θm , ωm ). (5.19)

5.1. Time-optimal tracking control 107

600
θm,f = 0, 10, …, 50

400

200
e2 [rad/s]

−200

−400

−600
−20 −15 −10 −5 0 5 10 15 20 25
e1 [rad]

Figure 5.4: Overlaid switching curve plotted in the error state plane: θm,f =
{0, 10, · · · , 50} rad for ramp tracking with ωm,f = 0 (solid) and ωm,f =
200 rad/s (dash-dot). The thick solid curve corresponds to (θmf , ωm,f ) = (0, 0).

The feedback implementation (5.18) is state-dependent, therefore mitigat-

ing the robustness issue due to the time-dependency of the open-loop im-
plementation. However, the control magnitude switches abruptly across the
switching curve, and leads to control chatter in the presence of modelling and
measurement errors. As will be shown in the next subsection, control chatter
is noticeable in simulation, which may be caused by numerical error during
time integration.

5.1.4 Results and discussions

Simulated tracking responses of a small step and a larger step of 1 kN and
5 kN respectively are compared in Figure 5.5. Beginning at t = 0 when
θm = ωm = 0, a clamp force setpoint reference is given, demanding the clamp
force and motor velocity to reach the targeted states in minimum time. The
time-optimal transient consists of maximum initial acceleration, operation at
maximum speed and controlled deceleration to the prescribed target point,
108 Chapter 5. High performance clamp force tracking control

6000
Reference
4000
Fcl [N]

2000

0
0 0.02 0.04 0.06 0.08 0.1
400
ωm [rad/s]

200

−200
0 0.02 0.04 0.06 0.08 0.1
5
Tm [Nm]

Control Chatter
−5
0 0.02 0.04 0.06 0.08 0.1
Time [s]

(a) Responses versus time.

350

300

250

200
ωm [rad/s]

150

100
Control Chatter
50

−50
0 2 4 6 8 10 12 14
θm [rad]

(b) Phase portraits of the state trajectories.

Figure 5.5: Responses for 1 kN (thin lines) and 5 kN (think lines) steps ob-
tained using the time-optimal feedback control law.
5.2. Robust near-time-optimal tracking control 109

as illustrated in the 5 kN step response. Note that the intervals of maximum

acceleration and deceleration are characterized by arcs in the state plane, while
the interval at constant velocity is portrayed by a straight line parallel to the
horizontal axis.
As the clamp force reaches the reference, control chatter is detected. A
closer examination of the phase portrait depicted in Figure 5.5 reveals the
state trajectory does not asymptotically converge to the targeted states, but
continuously moving in a bounded region close to the targeted states. The
finite discrepancy between the targeted and the current states are induced by
numerical time step issues. Even though the state tracking errors are small,
the maximum control magnitude is commanded by the time-optimal feedback
control law (5.18), which then leads to the control chatter. In implementation
on real systems, modelling and measurement errors will also induce control
chatter. For a small step response, as illustrated by the 1 kN case, the max-
imum velocity may not be reached. The acceleration is immediately followed
by deceleration with the switch-over taking place at fractions of the distance
travelled. Nevertheless, control chatter occurs for both the large and small
step commands when the references are reached.
The idealised time-optimal solution represents the theoretical upper-bound
performance subject to the power rating and velocity limit of the motor. How-
ever, this control law is impractical, as the smallest modelling error and mea-
surement noise will cause the control to chatter between the maximum and the
minimum values across the switching curve in vicinity of the zero state error.
The idealised control is useful however as the basis for further development
of a practical high-bandwidth controller robust against modelling error and
noises.

5.2 Robust near-time-optimal tracking

control
The major drawback of implementing the time-optimal feedback control in
an EMB is the associated control chatter, which leads to vibration, excessive
wear to the mechanical components and overheating of the power electronics.
To alleviate control chatter, the robust near-time-optimal control (NTOC) is
110 Chapter 5. High performance clamp force tracking control

developed. The key features of this design are simple implementation and
rapid tracking performance.
In this section, the perturbed EMB model is first presented, followed by
the NTOC design considering tracking and state constraints. Stability anal-
ysis of the closed-loop error system under modelling and measurement errors
is then executed. In order to aid in practical implementation, a gain tuning
procedure is proposed. This procedure is intuitive to implement and allows a
balanced choice between tracking speed and energy usage to be made. Then,
the controller implementation routines are summarised. Lastly, the controller
is implemented on the production-ready prototype EMB, and tested over dif-
ferent braking scenarios to assess the performance and robustness relative to
the benchmark controllers.

5.2.1 Problem formulation

Control chatter arises from uncertainty, inherent in any mathematical mod-
els that can at best approximate the behaviour of a real world process. In
order to take account of uncertainty, the control-oriented EMB model (5.1)
is augmented with a bounded, but potentially time-varying perturbation,
|d(t)| ≤ δ ∈ R. This perturbation encapsulates the effects due to modelling
error, model order reduction, imprecise numerical values of model parameters,
input disturbance and measurement noise. The perturbed control-oriented
EMB model is given by

Σemb,p : ẋ = f (x, Tm ) + Bd d(t), x(0) = x0 (5.20)

where f (·) is defined in (5.2) and Bd is given by

 
0
Bd =   . (5.21)
1

In addition to the state constraint x ∈ X and input constraint Tm ∈ U, the

available control authority is strictly greater than N Fcl (θm ) + T̄f (θm , ωm ) +Jδ
over the whole operating range.

Assumption 5.1 (Sufficient control authority in the plant). For a given max-
imum motor torque Tm max , the perturbation δ is small such that there exist a
5.2. Robust near-time-optimal tracking control 111

positive ap > δ defined by

1
!
Tm max − N Fcl (θm ) + T̄f (θm , ωm ) − δ ≥ ap > 0, ∀x ∈ X .
J
Remark 5.1. Assumption 5.1 implies sufficient actuation authority to always
control the sign of ω̇m . The practical implication can be understood from
the motor control problem, where ap simply represents the minimum of the
maximum achievable motor acceleration over the whole operating range x ∈ X ,
when the motor torque is set to its highest.
The second assumption is made on the state reference, where a class of
bounded tracking reference is considered.

Assumption 5.2 (Bounded tracking reference). xr (t) ∈ {(θm,r , ωm,r ) ∈ X :

|ωm,r | ≤ vr < ωm max , |ω̇m,r | ≤ ar < ap }.

Remark 5.2. To allow asymptotically converging tracking under state con-

straints, the upper-bounds of |ωm,r | and |ω̇m,r | are strictly less than the velocity
constraint, ωm max , and the achievable acceleration of the plant, ap respectively.
The tracking error is defined by e = [e1 , e2 ]T := x − xr (t) as introduced in
(5.17). Furthermore, the instantaneous maximum and minimum values of e2
are given by e2 max := ωm max − ωm,r and e2 min := −ωm,max − ωm,r respectively.

Problem 5.2 (Robust time-optimal tracking). Given an initial state, x(t0 ) =

x0 , a time-varying state reference, xr (t), an upper-bound for the disturbance,
δ, and a tracking error bound, β, find a state-feedback control Tm (x, xr ) such
that the tracking error is uniformly ultimately bounded by β when t ≥ tf ,
subjected to the input and state constraints. The robust time-optimal control
solution Tm∗ (x, xr ) minimises the time taken to reduce the tracking error within
β from the initial condition. This problem can be formulated as the following:
Z tf
min 1 dt,
Tm (x,xr ) t0

subject to (5.20), (5.21), (5.2),

Tm (t) ∈ U,
x(t) ∈ X , (5.22)

x(t0 ) = x0 ,
e(t) := x(t) − xr (t),
|e(τ )| ≤ β, τ ≥ tf .
 112 Chapter 5. High performance clamp force tracking control

Compared to the time-optimal control problem (5.5) which considers the

unperturbed model (5.1), the robust time-optimal control problem (5.22) con-
siders the perturbed system (5.20). Furthermore, the final state of the time-
optimal problem is constrained to a single point, while the final state of the ro-
bust time-optimal problem is constrained to a bounded region β. The optimal
control solution Tm∗ (x, xr ) is often hard to obtain. Nevertheless, a suboptimal
solution is available by taking advantage of the time-optimal switching curve
obtained from Section 5.1. The corresponding suboptimal controller will be
presented next.

5.2.2 Near-time-optimal controller design

The near-time-optimal control design is conducted in two stages, namely the
reduction of storage requirement for switching curve function and the intro-
duction of gradual transition in control magnitude across the switching curve.

Reduction of switching curve storage. The exact switching curve func-

tion, φe (ωm , θm,f , ωm,f ) is a continuous function of three variables, and sam-
ples of the switching curve are plotted in Figure 5.3. Online implementation
involves the storage and interpolation of a three dimensional map. If the
switching curve function can be approximated by a reduced-order function,
the storage requirement will be reduced and search speed will be increased.
It is observed in Figure 5.3 that the exact time-optimal switching curves
possess similar shape. When these curves are overlaid in Figure 5.4 after a
change of variable (represented in the error state), it is noticed that the exact
switching curve differs due to the plant nonlinearity, but may be approximated
using a single switching curve, φa (e2 ) that satisfies the following assumption.

Assumption 5.3 (Properties of switching curve). The following properties

hold for φa (e2 ):

(SC-1) φa (0) = 0,
(SC-2) e2 φa (e2 ) < 0, ∀e2 6= 0,
(SC-3) ξ(e2 ) < 0, where ξ(e2 ) = ∂φ∂e
a (e2 )
2
,
(SC-4) For a given ab , there exist ar , e2δ ≥ 0 such that |e2 | + (ar + ab )ξ(e2 ) > 0
for all |e2 | ≥ e2δ .
5.2. Robust near-time-optimal tracking control 113

Remark 5.3. Figure 5.4 shows the exact switching curves approximated using
a single switching curve that corresponds to xf = (0, 0). The approximated
switching curve crosses the origin, therefore (SC-1) holds. Furthermore, (SC-
2) and (SC-3) hold because the curve passes through the second and fourth
quadrant in the phase plane, and has negative gradient everywhere. Condition
(SC-4) can be used to determine the limiting characteristics of the reference.
The design variable, ab represents the motor acceleration when the velocity is
in vicinity of its constraint. For a particular set of ab and ar , vr has to be
chosen such that ωm max − δ − vr ≥ e2δ . This is to ensure the convergence of
e1 − φa (e2 ) in vicinity of the state boundary.
Having the approximated switching curve defined, the motor position offset
before control switching is given by

sa = e1 − φa (e2 ). (5.23)

The implementation of (5.23) is simpler than (5.16) because the dimension of

the lookup table has reduced from three to one, therefore resulting in a faster
table interpolation and reduced storage requirement.

Gradual transition of control magnitude. One approach to alleviate

chatter is to define a high slope saturation function in the vicinity of the
switching curve, whereby a gradual transition of control magnitude across the
switching curve is introduced (Workman et al., 1987; Zhou et al., 2001; New-
man, 1990). To this end, the sign function that is reminiscent of an infinite gain
operator and governs the abrupt control switching is replaced by the saturation
function over a transition region. It provides a finite slope approximation and
gives the system a finite bandwidth, therefore it is much more practical for the
EMB application. When the state is unconstrained, the method proposed in
Newman (1990) is extended to nonlinear plants with tracking, resulting in the
following form:
 !
sa e2 N Fcl + T̄f + J ω̇m,r 
 
Tm,trac = −Tm max sat ks sat + sat − ,
ssat e2 sat Tm max
(5.24)
where ks ≥ 2, ssat > 0 and e2 sat > 0 are tunable gains.
When the motor velocity is at its constraint boundary, application of the
feedback linearising controller (5.19) may lead to constraint violation due to
114 Chapter 5. High performance clamp force tracking control

e2
II(a): Tm = Tm,cvs

e2 max
e2 max − ∆
IV(b): Tm = −Tm max
e2 (t)

sa = e1 − φa (x2 ) e2sat

I: |Tm | < Tm max

III(a): Tm = +Tm max
e1
e1 (t) φa (e2 ) −ssat 0 ssat
III(b): Tm = −Tm max

−e2sat

IV(a): Tm = +Tm max

switching curve
e2 min + ∆
e2 min

II(b): Tm = Tm,cvs

Figure 5.6: Control magnitude in different regions of the error state plane.

the presence of perturbation. Therefore, (5.19) is adapted by augmentation of

a stabilising component, and is given by
 
∆
!
Tm,cvs = N Fcl + T̄f − Jkc ωm − ωm max − sgn (ωm ). (5.25)
2
where kc > 0 and 0 < ∆ ≤ δ are tunable gains.
Remark 5.4. When the feedback linearising-stabilising control law (5.25) is
applied in vicinity of the velocity constraint, |ωm | ∈ [ωm max − ∆, ωm max ], the
acceleration is given by |ω̇m | ≤ kc2∆ ≤ ab .
The overall controller can be summarised as:

Tm,cvs

if |ωm | ∈ [ωm max − ∆, ωm max ] and sa ωm ≤ 0,
Tm,ntoc = (5.26)
Tm,trac otherwise.


The control magnitude over different region in the state plane is shown in
Figure 5.6, and the regions in error state space where control laws (5.24) and
(5.25) operate are listed in Table 5.1.

5.2.3 Stability analysis of the near-time-optimal

control
In the following, the tracking error of the perturbed system (5.20) under the
control law (5.26) is shown to be uniformly ultimately bounded. To this end,
5.2. Robust near-time-optimal tracking control 115

Table 5.1: Control laws for regions in Figure 5.6.

Tm
Region sgn (sa ) sgn (sa e2 ) |ωm | Ctrl. Tm max
I {0, ±1} {0, ±1} < ωm max − ∆ Tm,trac ≤1
II(a) {0, −1} {0, −1} ∈ [ωm max − ∆, ωm max ] Tm,cvs ≤1
II(b) {0, +1} {0, −1} ∈ [ωm max − ∆, ωm max ] Tm,cvs ≤1
III(a) −1 {0, −1} < ωm max − ∆ Tm,trac +1
III(b) +1 {0, −1} < ωm max − ∆ Tm,trac −1
IV(a) −1 +1 ≤ ωm max Tm,trac +1
IV(b) +1 +1 ≤ ωm max Tm,trac −1

it is shown that any states in Regions IV(a) and IV(b) (refer to Figure 5.6)
will converge towards e2 = 0, i.e. falling into either Regions III(a) or III(b)
or I. Furthermore, any states in Regions III(a) or III(b) will converge towards
sa = 0, i.e. falling into either Regions II(a), II(b) or I. Then, trajectory
starting in II(a) and II(b) converges towards the switching curve, sa = 0.
Finally, when the error trajectory enters Region I, the asymptotic stability of
a uniform ultimate bound that contains the origin is proved.

Lemma 5.3. Consider the system (5.20) with initial conditions that satisfy
sa e2 > 0 and sa 6= 0. Suppose Assumptions 5.1, 5.2 and 5.3 hold true. Under
the control law (5.26), the trajectories converge to the e1 axis.

Proof. When the states satisfy sa e2 > 0 and sa 6= 0, the control magnitude
is at its maximum limit, that is Tm = −Tm max sgn(sa ). Define a Lyapunov
candidate function V1 = 12 e22 . Its time-derivative is given by

V̇1 = e2 ė2
= e2 (ω̇m − ω̇m,r )
= e2 (−|ω̇m | sgn (sa ) − ω̇m,r )

where use is made of the fact that (5.17b) and there is sufficient control au-
thority to control the sign of ω̇m (from Assumption 5.1). Furthermore, the
bound of the acceleration can be obtained from Assumption 5.1 when the con-
trol magnitude is saturated, that is |ω̇m | ≥ ap + δ. This leads to the following
116 Chapter 5. High performance clamp force tracking control

expressions

V̇1 = −e2 sgn (sa )|ω̇m | − e2 ω̇m,r

≤ −e2 sgn (sa )(ap + δ) − e2 ω̇m,r
≤ −e2 sgn (sa )(ap + δ) + e2 sgn (sa )ar
= −e2 sgn (sa )(ap + δ − ar ),

where use is made of the fact that e2 sgn (sa ) > 0, |ω̇m,r | ≤ ar (from As-
sumption 5.2) and −e2 ω̇m,r ≤ e2 sgn (sa )ar . Since V̇1 < 0 for all e2 6= 0, the
trajectories converge to e2 = 0, i.e. the e1 axis.

Lemma 5.3 indicates that trajectories with initial condition in Region IV(a)
or IV(b) will move into Regions I, III(a) or III(b).

Lemma 5.4. Consider the system (5.20) with initial conditions that satisfy
sa e2 ≤ 0 and |ωm | < ωm max − ∆. Suppose Assumptions 5.1, 5.2 and 5.3 hold
true, and let

N Fcl + T̄f + J ω̇m,r

!
sa e2
 
σ = ks sat + sat − (5.27)
ssat e2 sat Tm max

Under the control law (5.26), |sa (t)| will be strictly decreasing, and the set
{|sa | : |σ| ≤ 1} is positively invariant.

Proof. Firstly, sa = 0 is shown to be contained in {sa : |σ| ≤ 1}. This can be

achieved by substituting sa = 0 into (5.27), where |σ| ≤ 1 is obtained.
Secondly, it is shown that |sa (t)| is strictly decreasing. Given the initial
condition sa (0) ∈ {sa : |σ| > 1}, the input is Tm = −Tm max sgn (sa ). Let the
Lyapunov candidate function be V2 = 21 s2a . The time-derivative is

V̇2 = sa ṡa
∂φ(e2 )
!
= sa ė1 − · ė2
∂e2
 
= sa e2 − ξ(e2 )(ω̇m − ω̇m,r )
 
= sa e2 − sa ξ(e2 ) − |ω̇m | sgn (sa ) − ω̇m,r ,
5.2. Robust near-time-optimal tracking control 117

where use is made of the fact that there is sufficient control authority to control
the sign of ω̇m (from Assumption 5.1). Furthermore,

V̇2 = sa e2 + |sa |ξ(e2 )|ω̇m | + sa ξ(e2 )ω̇m,r

≤ sa e2 + |sa |ξ(e2 )(ap + δ) + sa ξ(e2 )ω̇m,r
≤ sa e2 + |sa |ξ(e2 )(ap + δ − ar )
< 0, ∀sa 6= 0,

where use is made of the fact that sa e2 ≤ 0, ξ(e2 ) < 0 (from Assumption 5.3),
|ω̇m | ≥ ap + δ (from Assumption 5.1 with saturated control) and sa ω̇m,r ≥
−|sa |ar (from Assumption 5.2). Therefore, |sa (t)| is strictly decreasing and
the set {sa : |σ| ≤ 1} is positively invariant.

Lemma 5.4 shows that any trajectories within Regions III(a) or III(b) have
strictly decreasing |sa (t)|. However, the state trajectory may reach Region II
(which contains the state constraint) before reaching its invariant set, which
is Region I. The following lemma shows that |sa (t)| is strictly decreasing if the
control law (5.26) is applied when the state is in Region II.

Lemma 5.5. Consider the system (5.20) with initial conditions that satisfies
sa e2 ≤ 0 and |ωm | ∈ [ωm max − ∆, ωm max ]. Suppose Assumptions 5.1, 5.2 and
5.3 hold true. Under the control law (5.26), |sa (t)| will be strictly decreasing.

Proof. Choosing Lyapunov candidate function as V2 = 12 s2a and taking time

derivative, we have

V̇2 = sa ṡa
 
= sa e2 − ξ(e2 )(ω̇m − ω̇m,r )
 
= sa sgn (e2 ) |e2 | + (ω̇m,r − ω̇m )ξ(e2 ) sgn (e2 )
 
≤ sa sgn (e2 ) |e2 | + (ar + ab )ξ(e2 )
< 0, ∀sa 6= 0

where (SC-4) of Assumption 5.3 is used. Therefore |sa (t)| is strictly decreasing
and reaching the switching curve sa = 0.

Apart from using Lemma 5.5 to show that the |sa (t)| is strictly decreasing in
Region II, the following lemma also shows that the control law (5.26) maintains
the velocity within its constraint in Region II.
118 Chapter 5. High performance clamp force tracking control

Lemma 5.6. Consider the system (5.20) and the control law (5.26), with
initial conditions that satisfy sa e2 ≤ 0 and |ωm | ∈ [ωm max − ∆, ωm max ]. The
region |ωm | ∈ [ωm max − ∆, ωm max ] is uniformly bounded.

Proof. Consider the dynamics of ωm :

 
∆
!
ω̇m = −kc ωm − ωm max − sgn (ωm ) + d(t).
2

Furthermore, consider a new variable ω̂m = ωm − (ωm max − ∆2 ) and a Lyapunov

function V3 = 12 ω̂m
2
. The time-derivative of the Lyapunov function is

V̇3 = ω̂˙ m ω̂m ≤ −kc ω̂m

2
+ δ|ω̂m |.
2
With c > 2k δ
2 , solution starting in the set {V3 (ω̂m ) ≤ c} will remain therein
c
for all future time since V̇3 is negative on the boundary V = c. Hence, ωm is
uniformly bounded in |ωm | ∈ [ωm max − ∆, ωm max ].

Finally, the following lemma shows that the error trajectories in Region I
converge to a uniform ultimate bound that contains the origin.

Lemma 5.7. Consider the perturbed system (5.20), and given that the initial
condition is in the vicinity of the tracking reference, that is sa (0) ∈ {sa :
|σ| ≤ 1}, x ∈ X − {x : |ωm | ∈ [ωm max − ∆, ωm max ], sa ωm < 0}, e2esat 2
≤
   
N Fcl +T̄f +J ω̇m,r N Fcl +T̄f +J ω̇m,r
1+ Tm max
and e2
e2 sat
≥ −1 + Tm max
. Under the control law
(5.26), the trajectories are uniformly ultimately bounded.

Proof. In vicinity of the error state space e1 -e2 origin, the control law (5.26)
reduces to a linear feedback law,

N Fcl + T̄f + J ω̇m,r

!
sa e2
Tm = −Tm max ks + − .
ssat e2 sat Tm max

Therefore the closed-loop system is

ė = Ae + g(t, e), (5.28)

where
   
0 1 0 Tm max
A =  bks , g(t, e) =  bks , b= .
− ssat − e2bsat + sbksats ξ(0) ssat
ζ(e 2 ) + d(t) J
5.2. Robust near-time-optimal tracking control 119

The approximated switching curve is linearised at the origin of the error state,
where φa (e2 ) = ξ(0)e2 + ζ(e2 ). Furthermore, ζ(e2 ) is bounded, such that
ζ(e2 ) ≤ σke2 k2 for all e2 ≤ k2 . Since A is Hurwitz, for any Q = QT > 0, there
exist a unique P = P T > 0 that satisfies the following Lyapunov equation
(Khalil, 2002, Theorem 4.6):

P A + AT P = −Q.

Let V (t, e) = eT P e be a Lyapunov function, where P = [ pp12 p22 ]. Calculat-

11 p12

ing the derivative of V (t, e) along the trajectories of (5.28), we obtain

V̇ (t, e) = −eT Qe + 2eT P g(t, e)

!
bks
≤ −λmin (Q)kek22 + 2(p12 e1 + p22 e2 ) ζ(e2 ) + d(t)
ssat
bks q
≤ −λmin (Q)kek22 + 2|p12 e1 + p22 e2 |ke2 k2 σ + 2δ p212 + p222 kek2
ssat
bk s
q
≤ −λmin (Q)kek22 + 2(p12 + 2p22 ) σ + 2δ p212 + p222 kek2 .
ssat

Suppose (p12 + 2p22 ) sbksats σ ≤ 1 − ν, where 0 < ν < 1. Then,

q
V̇ (t, e) ≤ −νλmin (Q)kek22 + 2δ p212 + p222 kek2
≤ −(1 − θ)νλmin (Q)kek22 ,
√
2δ p212 +p222
for all kek2 ≥ µ = νθλmin (Q)
, where 0 < θ < 1.
Therefore, by applying Lemma 9.2 in Khalil (2002), the solutions are uni-
formly bounded by q v
2δ p212 + p222 u
t λmax (P ) .
u
β= (5.29)
νθλmin (Q) λmin (P )

Remark 5.5. Lemma 5.7 shows that the ultimate bound β is proportional to
the upper bound on the disturbance, δ. This result can be interpreted as a
robustness property of the unperturbed system (5.1) with control law (5.26), as
an arbitrarily small δ will result in a small ultimate bound, deviated from the
origin of the error state-space. Although the ultimate bound can be calculated
qualitatively, it is usually found to be too conservative to be used for tuning
of a practical controller (Khalil, 2002).
120 Chapter 5. High performance clamp force tracking control

Collecting the results from Lemma 5.3 to Lemma 5.7, the convergence
analysis can be summarised as the following theorem.

Theorem 5.8. Consider the perturbed EMB system (5.20). Suppose Assump-
tions 5.1, 5.2 and 5.3 hold true, the near-time-optimal tracking control law
(5.26) asymptotically stabilises the tracking error to a uniform ultimate bound
that contains the origin.

Theorem 5.8 indicates that the error trajectory is bounded if the given
assumptions are satisfied. By verifying these assumptions, the applicability of
the NTOC on a given plant can be determined, and the design limits of the
tracking reference can be decided. Furthermore, the quantitative value of the
ultimate bound (5.29) is found to be too conservative to be used as a control
design criteria. Therefore, another design criteria that considers tracking speed
and control effort is proposed in the next subsection.

5.2.4 Practical implementation and gain tuning

Simplified NTOC

During EMB clamping operation, the motor position reference θm,r (t) is avail-
able online from the driver input, but the velocity ωm,r (t) and acceleration
ω̇m,r (t) commands are not explicitly available. While ωm,r (t) and ω̇m,r (t) can
be numerically approximated utilising the backward Euler differentiation, this
approach introduces noise and phase offset, which are undesirable in applica-
tion. Based on these practical considerations, a special case of the NTOC is
considered, whereby (5.24) is adapted for step reference tracking, given by:
 !
sa ωm N Fcl + T̄f 
 
Tm,step = −Tm max sat ks sat + sat − , (5.30)
ssat ωm sat Tm max

where ks ≥ 2, ssat > 0 and ωm sat > 0 are tunable gains.

Furthermore, if the observed velocity constraint violation using (5.19) is
insignificant in practice, then it can be adopted in place of the more complex
control law (5.25). Therefore, the simplified NTOC is defined by:

Tm,cv

if |ωm | ≥ ωm max and sa ωm ≤ 0,
Tm,ntocs = (5.31)
Tm,step otherwise.

5.2. Robust near-time-optimal tracking control 121

Gain Tuning

The tuning parameters, ks , ssat and ωm sat explicitly represent the size of the
boundary layer (refer to Figure 5.61 ). However, it is unclear how these param-
eters affect the tracking speed and control effort of the controller. Motivated
by the need to address the aforementioned problems and to aid in design, a
systematic gain tuning procedure is proposed. This procedure provides the de-
signer an intuitive tool to choose a set of controller gains that allows a balanced
choice between these conflicting objectives to be made.
Recall that the aims of the NTOC is to provide a chatter free controller with
short rise-time and high-bandwidth. The short rise-time criteria is achieved
by utilising an approximated time-optimal switching curve to determine the
instance to decelerate the motor. To address the high bandwidth criteria, a
local analysis of the closed-loop system is carried out. The local analysis is
sufficient when the trajectory is close to the reference, which is a typical sce-
nario in practise when the reference changes slowly compared to the dynamics
of the system.
Consider the case when the state is in vicinity of zero tracking error and
the control magnitude is unsaturated in the absence of plant disturbance, the
closed-loop dynamics can be calculated using (5.1) and (5.31), governed by
      
θ̇
 m = 
0 1 θ
  m + 
0  
θm,r + ζ(ω m , (5.32)
)
ω̇m − sbksats − x2bsat + sbksats ξωm (0) ωm bks
ssat

where b = Tm max
J
and the linearised switching curve is employed, given by

φa (ωm ) = ξωm (0)ωm + ζ(ωm ), (5.33)

and the gradient of the switching curve is approximated using

∂φa (ωm )
ξωm (ωm ) = . (5.34)
∂ωm
This approximation is shown in Figure 5.7, indicating that ζ(ωm ) , which repre-
sents the higher order terms is small near the origin, and increases nonlinearly
with the magnitude of ωm .
Suitable tools for controller gain tuning are readily available if (5.32) is
rewritten in a linear equation with perturbation, where one choice is the linear
1
The tuning gain ωm sat of the simplified NTOC is analogous to the e2 sat of the tracking
NTOC.
122 Chapter 5. High performance clamp force tracking control

φa(ωm), ξω (0)ωm [rad]

4 φa(ωm)
2 ξω (0)ωm
m

0
m

−2

−4
−300 −200 −100 0 100 200 300
ζ(ωm) = φa(ωm) − ξω (0)ωm [rad]

2
m

−2

−4
−300 −200 −100 0 100 200 300
ωm [rad/s]

Figure 5.7: Comparison of switching curve with its first order linear approxi-
mation.

quadratic regulator (LQR) design. LQR design is based upon optimising an

objective function with respect to tracking speed and control effort. To this
end, (5.32) is adapted to the following form:

        
θ̇
 m = 
0 1  θm   0  0
+ 1 Tm +  , (5.35)
ω̇m 0 0 ωm J
bξω m (ω m )

where the last term is treated as a perturbation2 to the system. To facilitate

the proceeding analysis, the input is compared to the following linear state-
feedback control law:
 
i θ
m
h
Tm,lqr = − kp kd   + kp θm,r (5.36)
ωm

The relationship between the new gains kp , kd and the original tuning param-

2
Can be verified to be small.
5.2. Robust near-time-optimal tracking control 123

eters are
bks
ssat = , (5.37a)
kp
!−1
kd ks
ωm sat = + ξ(0) . (5.37b)
b ssat

The LQR design problem is characterised by the following objective function

consists of penalties on tracking speed zx (t) and control effort, weighted by ρ:
Z ∞
Jlqr = kzx (t)k2 + ρTm (t)2 dt. (5.38)
0

The penalty on tracking speed is a weighted function of the motor position

and velocity tracking, and is given by
 
i θ
m
h
zx = 1 γ   (5.39)
ωm

Since the tuning procedure now translates to the selection of a weighting con-
trasting tracking error and control effort, it is more transparent than direct
alteration of the boundary layer size.
Note that techniques such as loop-shaping can be applied for the tuning
of ρ and γ. Consider the open-loop transfer-matrix of (5.35) from the plant’s
input to the controller’s output, larger ρ results in the decrease of the gain
magnitude and slower response. Additionally, larger γ leads to a larger phase
margin and smaller overshoot in the step response.
Figure 5.8 illustrates the trend of ssat and ωm sat with respect to ρ and γ.
Increased penalty on control effort, ρ leads to a larger ssat and also a larger
boundary layer. Furthermore, the weight on motor velocity tracking error γ
has no apparent effect on ssat as all four γ = {0.001, 0.02, 0.03, 0.05} produce
similar ssat that visually stack on top of each other. A decreased γ results in
the stretching of boundary layer vertically. However, for a small γ = 0.001,
the associated value for ωm sat is negative and does not satisfy the condition
ωm sat > 0 needed for the stability of the controller, therefore this choice of γ
is invalid.
The sizes of unsaturated control region produced from different values of
ρ and γ are shown in Figure 5.9. Due to the negative ωm sat associated with
γ = 0.001, the unsaturated control regions are distinctive to the intended de-
sign shown in Figure 5.6, and the resulted control trajectories may be unstable.
124 Chapter 5. High performance clamp force tracking control

6
ssat [rad]

2
γ=0.001 γ=0.02 γ=0.03 γ=0.05
0
1 2 3 4 5 6 7 8 9 10
−8
x 10
300

200
x2sat [rad/s]

100

−100

1 2 3 4 5 6 7 8 9 10
ρ x 10
−8

Figure 5.8: Controller parameters ssat and ωm sat are plotted against ρ and γ
with ks = 2.

Comparing the three figures where γ = 0.01, larger ρ yields a larger unsat-
urated region. Furthermore, comparing the figures with ρ = 1 × 10−8 and
γ = {0.01, 0.05}, larger value of γ reduces the unsaturated region in vicinity
of the origin.

Implementation Procedure

The NTOC implementation procedures are summarised in Procedure 5.1 and

illustrated in Figure 5.12.

Procedure 5.1 (Implementation of NTOC). The implementation procedure

for the simplified NTOC can be broken up into five steps. The details of these
steps are given in the following:

Step 1: Verify sufficient control authority of the plant. The EMB

motor power rating is chosen such that the available motor torque is
always greater than the torques produced from stiffness and friction.
5.2. Robust near-time-optimal tracking control 125

300 300 300

γ: 0.001 γ: 0.01 γ: 0.05
ρ: 1e−09 ρ: 1e−09 ρ: 1e−09
ωm [rad/s]
0 0 0

−300 −300 −300

−5 0 5 −5 0 5 −5 0 5
300 300 300
γ: 0.001 γ: 0.01 γ: 0.05
ρ: 1e−08 ρ: 1e−08 ρ: 1e−08
ωm [rad/s]

0 0 0

−300 −300 −300

−5 0 5 −5 0 5 −5 0 5
300 300 300
γ: 0.001 γ: 0.01 γ: 0.05
ρ: 5e−08 ρ: 5e−08 ρ: 5e−08
ωm [rad/s]

0 0 0

−300 −300 −300

−5 0 5 −5 0 5 −5 0 5
e1 [rad] e1 [rad] e1 [rad]

Figure 5.9: Comparison of unsaturated control region with different ρ and γ.

Therefore, the motor is able to change the direction of acceleration over

the whole operating range.
The maximum magnitude of N Fcl + T̄f occurs when the clamp force
is at its peak at 30 kN, and when the velocity is at its lower bound,
ωm = −300 rad/s2 . Using the set of parameters provided in Table 4.3,
ap > 0 thus Assumption 5.1 is satisfied.
Step 2: Determine the approximated switching curve. Firstly, the ex-
act switch curve is calculated using the procedure outlined in Sec-
tion 5.1.3. The curves depicted in Figure 5.3 are produced by set-
ting Tm = ±Tm max and integrating (5.1) backwards in time using
xf = (θm,f , 0) as initial condition.
Secondly, by changing the coordinate from x to e, the switching curves
are overlaid and plotted in Figure 5.4. The switching curve corresponds
to xf = (0, 0) is chosen, and approximated using a polynomial fit as
shown in Figure 5.10. Conditions (SC-1), (SC-2) and (SC-3) in Assump-
tion 5.3 can be verified as follows. Since the curve intersects the origin,
(SC-1) is satisfied. Furthermore, (SC-2) and (SC-3) are satisfied because
126 Chapter 5. High performance clamp force tracking control

the curve cuts through the second and fourth quadrants, and has negative
gradient everywhere.
Step 3: Determine the design limits of reference. The class of refer-
ence with with bounded higher derivatives is considered, where vr =
270 rad/s and ar = 4300 rad/s (Assumption 5.2). For the unper-
turbed case, where δ = 0, the bottom plot of Figure 5.10 shows that
|e2 | + ar ξ(e2 ) > 0 for all |e2 | ∈ [e2δ , 650], e2δ = 29 rad/s. Hence, (SC-4)
of Assumption 5.3 is satisfied.
Step 4: Controller synthesis. The EMB clamp force tracking controller is
implemented as shown in Figure 5.11. The clamp force reference and
measurement are first converted to the motor position reference and mea-
surement respectively. Then, the motor position and velocity references,
together with their measurements are fed into the control law (5.31).
Step 5: Gain tuning. The controller gains ks , ssat , ωm sat are tuned using the
procedure proposed in this subsection.

5.2.5 Results and discussions

This subsection details the results from both simulation and experimentation
for the proposed NTOC. The proposed controller is benchmarked against the
time-optimal control presented in Section 5.1, which represents the best pos-
sible performance achievable by the plant limitations under ideal conditions.
Additionally, the proposed controller is also benchmarked against and the base-
line proportional-integral (PI) controller presented in (Line et al., 2008), which
represents the present industry practise. In order to perform fair tests, the PI
gains are tuned such that the available motor torque is fully utilised while
avoiding integral windup due to input and state constraints.

Simulation results

Initially, results obtained via simulation are used to compare the different
control approaches. The investigation begins with the selection of NTOC
gains, followed by the comparison of different controllers on tracking response.
5.2. Robust near-time-optimal tracking control 127

20
5th order polynomial fit
Switching curve
10
φa(e2)

−10

−20
0
ξ(e2)
−0.01
−|e2|/ar
ξ(e2), −|e2|/ar

−0.02

−0.03

−0.04

−0.05
−800 −600 −400 −200 0 200 400 600 800
e2 [rad/s]

Figure 5.10: (Top) Approximation of the switching curve using a 5th order
polynomial. (Bottom) For ar = 4300 rad/s2 , the first order differentiation of
the switching curve, ξ(e2 ) is larger than − |ea2r | for all 29 ≤ |e2 | ≤ 650 rad/s.

Clamp Force Controller

θm,r
EMB
Fcl,r Convert force to θm Simplified NTOC Tm Control-
Fcl
motor position Tm,ntocs Oriented
Model
Fcl ωm

Figure 5.11: Block diagram of the simplified NTOC structure.

128 Chapter 5. High performance clamp force tracking control

Step 1: Verify sufficient control authority in the plant

Determine ap and verify Assumption 5.1.

Step 2: Determine the approximated switching curve, φa

Calculate the exact switching curve, φe using (5.1). Then, approx-
imate it with φa that satisfies (SC-1)∼(SC-3) of Assumption 5.3.

Step 3: Determine the design limits of reference

Choose vr , ar that satisfy Assumption 5.2. Then, determine e2δ
that fulfils (SC-4) of Assumption 5.3.

Step 4: Controller synthesis

Implement the controller as shown in Figure 5.11.

Step 5: Gain tuning

Select the controller gains using the guidelines provided in Sec-
tion 5.2.4.

Figure 5.12: Implementation procedure for NTOC of an EMB.

Finally, the nominal plant model is perturbed to investigate on the robustness

of the NTOC controller.

Gain selection. As proposed in Section 5.2.4, the NTOC gain tuning can
be translated to the selection of weights contrasting control effort and tracking
error, given by ρ and γ. To make an informed selection for the gains, the
control performances using a range of weights are compared.
Figure 5.13(a) compares 2 kN step responses with different weighting on
the control effort, ρ. A smaller value of ρ leads to a faster response, but it is
also more likely to induce control chatter as demonstrated in the case where
ρ = 0.1 × 10−8 . This is because a small value of ρ produces a small ssat ,
therefore leading to a narrower boundary layer. The NTOC with an infinitely
small boundary layer has an abrupt change in control magnitude across the
switching curve, which is also observed in the time-optimal control and can
5.2. Robust near-time-optimal tracking control 129

cause control chatter in practise.

Figure 5.13(b) illustrates the 2 kN step responses with different weighting
on the tracking error for motor velocity, γ. A smaller γ produces response
with increased overshoot. The clamp force response for γ = 0.001 oscillates
around the reference, as the associated controller gain ωm sat is negative which
does not satisfy a stability condition for the controller.
The 0%–95% rise time and energy usage for a 2 kN setpoint tracking us-
ing NTOC with different gains are shown in Figure 5.14. As expected from
previous discussions, a smaller ρ yields a shorter rise time approaching the
time-optimal control performance, albeit requires more energy to accomplish
the setpoint tracking. When ρ is large, it is interesting to observe that a smaller
γ decreases the rise time with slight increase in energy usage, which is led by
the increased overshoot response.
The choice of γ = 0.03 and ρ = 4 × 10−8 seems to give a good balance be-
tween rise time and energy usage and will be used in the following for simula-
tion and experimental studies. These weights yield NTOC gains of ssat = 4 rad
and ωm sat = 50 rad/s.

Closed-loop bandwidth. The sinusoidal tracking performance is further

investigated over a wide range of frequency from 1 Hz to 30 Hz, chosen to
cover both the normal braking scenario and demanding scenarios with anti-lock
braking system (ABS) and electronic stability control (ESC) operating. Fur-
thermore, the means of the references are in the range of {1, 5, 10, 20, 30} kN,
and the amplitudes are 10% of the means.
Although transfer functions and closed-loop bandwidth are undefined on
nonlinear systems, such as the EMB system, it is informative to plot the gain
magnitude and phase-shift of the clamp force output relative to the reference
over a wide range of frequency, as this provides insights on how well the con-
troller performs over different frequencies. Note that this plot approximates
the Bode diagram of the linearised closed-loop EMB system.
Figure 5.15(a) shows the gain and phase-shift of the clamp force output
relative to the reference with the general NTOC (5.26), which is produced
from simulation results. For low frequency (< 10 Hz), it is noticed that the
gain magnitude is close to 0 dB for all clamp force levels. This suggests that
the NTOC has a consistent tracking performance for low frequency references.
130 Chapter 5. High performance clamp force tracking control

γ = 0.05
3000

2000
Fcl [N]

1000 ρ=0.1×10−8
ρ=2.5×10−8
0
−8
0 0.02 0.04 0.06 0.08 ρ=5×10
0.1 0.12
−8
400 ρ=10×10
ωm [rad/s]

200

−200
0 0.02 0.04 0.06 0.08 0.1 0.12
5
Tm [Nm]

−5
0 0.02 0.04 0.06 0.08 0.1 0.12
Time [s]

(a) Responses with ρ = {0.1, 2.5, 5, 10} × 10−8 and γ = 0.05.

ρ = 5 × 10−8
4000
γ = 0.001 γ = 0.02 γ = 0.03 γ = 0.05
Fcl [N]

2000

0
0 0.05 0.1 0.15 0.2
400
ωm [rad/s]

200

−200
0 0.05 0.1 0.15 0.2
5
Tm [Nm]

−5
0 0.05 0.1 0.15 0.2
Time [s]

(b) Responses with γ = {0.001, 0.02, 0.03, 0.05} and ρ = 5 × 10−8 .

Figure 5.13: Comparison of 2 kN setpoint tracking responses with different

NTOC gains.
5.2. Robust near-time-optimal tracking control 131

0.08
Time−optimal control γ = 0.2 γ = 0.3
0.07
Rise Time [s]

0.06

0.05

0.04
1 2 3 4 5 6 7 8 9 10
−8
x 10

25
Energy Usage [J]

20

15

10
1 2 3 4 5 6 7 8 9 10
ρ −8
x 10

Figure 5.14: Rise time and energy usage for a 2 kN step using the time-optimal
control and NTOC with γ = {0.02, 0.03} and ρ ∈ [1, 10] × 10−8 .

However, for high frequency (> 10 Hz), the gain magnitude varies significantly
with clamp force levels. In particular, for a small clamp force level at 1 kN,
the gain magnitude deviates from 0 dB and start decreasing at approximately
25 Hz. For a larger clamp force level at 30 kN, the decrease starts at ap-
proximately 10 Hz. The decrease in gain magnitude starts earlier with higher
clamp force because higher clamp force escalates both the load torque and fric-
tion torque, therefore reduces the achievable acceleration needed to maintain
a small tracking error.
Compared to the general NTOC, the simplified NTOC has smaller gain
magnitudes and larger phase-shifts, as depicted in Figure 5.15(b). This is due
to the removal of motor velocity reference in the control law.
Figure 5.16 shows the gain and phase-shift of the clamp force output rel-
ative to the reference with the industry-tuned baseline PI controller. Unlike
the general NTOC, the gain magnitudes with the PI are inconsistent for for
132 Chapter 5. High performance clamp force tracking control

Bode Diagram with Different Clamp Force Levels

From: Clamp force reference Fcl,r To: Clamp force Fcl
5

0
Magnitude (dB)

−5

−10 1 kN
5 kN
10 kN
−15 20 kN
30 kN
−20
0

−45
Phase (deg)

−90

−135

−180

−225
0 Frequency (Hz) 1
10 10

(a) General NTOC (5.26).

Bode Diagram with Different Clamp Force Levels

From: Clamp force reference Fcl,r To: Clamp force Fcl

0

−5
Magnitude (dB)

−10

−15 1 kN
5 kN
10 kN
−20 20 kN
30 kN
−25
0

−45
Phase (deg)

−90

−135

−180

−225
0 1
10 10
Frequency (Hz)

(b) Simplified NTOC (5.31).

Figure 5.15: Gain and phase-shift of the clamp force output relative to the
reference with NTOCs.
5.2. Robust near-time-optimal tracking control 133

Bode Diagram with Different Clamp Force Levels

From: Clamp force reference Fcl,r To: Clamp force Fcl

0

Magnitude (dB) −10

−20
1 kN
5 kN
−30 10 kN
20 kN
30 kN
−40
0

−45
Phase (deg)

−90

−135

−180
0 1
10 10
Frequency (Hz)

Figure 5.16: Gain and phase-shift of the clamp force output relative to the
reference with baseline PI controller.

all force levels. In particular, the gain magnitudes for smaller clamp force
levels are typically smaller or equal compared to the higher clamp force lev-
els. This reflects a worse tracking performance at small clamp force levels,
which is resulted from not fully utilising the available motor capacity at small
clamp force levels. Additionally, the baseline PI controller produces responses
with larger gain magnitudes and phase-shifts compared to NTOC, therefore
indicating that the NTOC has a better tracking performance.

Robustness against parameter variation. To examine the robustness of

NTOC against parameter variation, simulation responses of scaled models are
compared. It has been found experimentally in (Schwarz et al., 1999) that
the lumped stiffness under different temperatures and pad thicknesses can be
related using a scaled expression:

Fcl,scaled (θm ) = αFcl (θm ).

For the purpose of the robustness study, α ∈ [0.5, 1.5] is randomly sampled
from a uniform distribution. Similarly, the friction parameters are also ran-
134 Chapter 5. High performance clamp force tracking control

domly scaled from 0.5 to 1.5.

To study the control performance for a variety of clamp force levels, the
perturbed models are first simulated with a small reference 2 kN and then
followed by a larger reference at 20 kN. Figure 5.17(a) and Figure 5.17(b)
illustrates the responses for the nominal model and the 25 perturbed models,
where setpoint references are given at 2 kN and 20 kN respectively. While the
input and output trajectories of the perturbed models are different compared
to the nominal model, it is observed that the overshoots and steady state errors
are less than 5%. Furthermore, chatter does not occur albeit the variation in
model parameters, therefore demonstrating the robustness of NTOC.

Experimental results

The experimental results presented here were performed using the static test
rig of the prototype EMB described in Chapter 3. The simplified NTOC (5.31)
was implemented to the onboard 16-bit microcontroller, where the clamp force
control loop operates at 250 Hz. The NTOC gains ks = 2, ssat = 4 and
e2 sat = 50 were used
The performance of the EMB controllers are assessed experimentally by
the rise time and how well they tracks a reference, which are all important
for demanding situations such as panic braking and during ABS operations.
These criteria will be assessed by the execution of tests using standardised
waveforms such as step, sinusoidal and triangle. Furthermore, performance of
the controllers will also be discussed in terms of energy usage.

Rise time. Step tracking performance of the controllers are first assessed
using a large step, and then followed by a smaller step. The large step ma-
noeuvre corresponds to a full brake application in emergency braking while the
small step demand corresponds to a light brake application. A faster step re-
sponse improves reaction time in emergency braking and may reduce stopping
distance. Figure 5.18(a) illustrates the large step response of 20 kN, where
both the proposed NTOC and the baseline PI controller have the same rising
slope and similar rise time, indicating that the PI is well-tuned for a large
manoeuver. However, the PI is slower at bringing the response to 0 kN, as the
input switches from negative to positive too early. The NTOC responses are
close to the ideal case apart from some overshoot during the falling phase, due
5.2. Robust near-time-optimal tracking control 135

4000
Fcl [N] Perturbed Model Nominal Model

2000

0
0 0.05 0.1 0.15 0.2 0.25 0.3
400
ωm [rad/s]

200

−200
0 0.05 0.1 0.15 0.2 0.25 0.3
100
Tm [Nm]

−100
0 0.05 0.1 0.15 0.2 0.25 0.3
Time [s]

(a) Responses for 2 kN step.

4
x 10
4
Perturbed Model Nominal Model
Fcl [N]

0
0 0.05 0.1 0.15 0.2 0.25 0.3
400
ωm [rad/s]

200

−200
0 0.05 0.1 0.15 0.2 0.25 0.3
100
Tm [Nm]

−100
0 0.05 0.1 0.15 0.2 0.25 0.3
Time [s]

(b) Responses for 20 kN step.

Figure 5.17: Responses of the nominal and the perturbed plant models.
136 Chapter 5. High performance clamp force tracking control

to the control switching to the clearance management controller in vicinity of

the pads/rotor contact. Chatter which usually happens in the implementation
of time-optimal control does not occur here.

Although the proposed NTOC performed satisfactorily for a full brake ap-
ply, to ensure the consistency of step tracking performance within the range
of operation, a smaller step manoeuvre of 2 kN is performed, and is shown in
Figure 5.18(b). The rising and falling responses of the NTOC are faster com-
pared to the PI, suggesting that the fixed set of PI gains are too conservative
for a small step reference. As shown in the measured velocity and motor cur-
rent, the NTOC fully utilises the available motor torque to track the setpoint
reference.

Figure 5.19 presents the phase-portrait of the tracking responses of 2 kN

and 20 kN steps. The NTOC demonstrates a longer acceleration period and
achieves higher velocity for the 2 kN tracking. For the 20 kN tracking, the
NTOC maintains velocity at (or slightly above) the state constraint for longer.
The slight violation of the state constraint is due to errors from modelling and
measurement, and could be remedied by imposing the feedback linearising-
stabilising control law (5.25). However, the observed level of constraint vio-
lation is insignificant from an operational perspective. Overall, compared to
the baseline PI controller, the NTOC maintains the acceleration phase longer
and begins the deceleration phase closer to the setpoint. The discrepancies in
initial positions are due to the difficulty of obtaining an accurate 0 kN clamp
force.

To assess how close the NTOC performs relative to the limitation of the
plant, the 0%–95% rise time of the NTOC is compared to the time-optimal
control, as depicted in Figure 5.20. For smaller steps, notably for steps below
2 kN, the rise time of the NTOC is significantly longer compared to the time-
optimal control. This is due to the introduction of the boundary layer in the
NTOC to avoid control chatter. However, the rise times of the NTOC for
larger steps are within 5% of the time-optimal control, indicating that the
NTOC is able to fully utilise the available motor capacity to achieve a rapid
step tracking. Note that the rise time of the baseline PI is always longer
compared to the NTOC over the whole clamp force operating range.
5.2. Robust near-time-optimal tracking control 137

4
x 10
4
Ref. PI (exp.) NTOC (exp.) NTOC (sim.)
Fcl [N]
2

0
0 0.5 1 1.5

300
ωm [rad/s]

−300
0 0.5 1 1.5

3
Tm [Nm]

−3
0 0.5 1 1.5
Time [s]

(a) Responses for 20 kN step.

4000
Ref. PI (exp.) NTOC (exp.) NTOC (sim.)
Fcl [N]

2000

0
0 0.5 1 1.5

300
ωm [rad/s]

−300
0 0.5 1 1.5

3
Tm [Nm]

−3
0 0.5 1 1.5
Time [s]

(b) Responses for 2 kN step.

Figure 5.18: Experimental and simulation setpoint tracking responses.

138 Chapter 5. High performance clamp force tracking control

400

Velocity Constraint
300

200

100
ωm [rad/s]

−100

−200
2 kN: Basline PI Switching Curve
−300 2 kN: NTOC
20 kN: Basline PI Velocity Constraint
20 kN: NTOC
−400
−35 −30 −25 −20 −15 −10 −5 0 5 10 15
e1 [rad]

Figure 5.19: Experimental measurements presented in state plane for 2 kN

and 20 kN setpoint tracking.

0.4
Time−optimal control (simulation)
NTOC (simulation)
0.35 NTOC (experiment)
Baseline PI (experiment)
0.3

0.25
Rise time [s]

0.2

0.15

0.1

0.05

0
0 0.5 1 1.5 2 2.5 3
Fcl,r [N] 4
x 10

Figure 5.20: Rise time for step responses initiated from 0.4 kN.
5.2. Robust near-time-optimal tracking control 139

4
x 10

3 Reference Baseline PI NTOC

Fcl [N] 2

0 0.5 1 1.5 2 2.5 3

4000
Tracking error [N]

2000

−2000

−4000
0 0.5 1 1.5 2 2.5 3
Time [s]

Figure 5.21: Experimentally obtained reference tracking responses using tri-

angular wave with 1 second period and 20 kN magnitude.

Tracking error. Reference tracking of the controllers is assessed using si-

nusoidal and triangle waveforms, and the performance is quantified using the
normalised integral of the absolute error (nIAE), which is defined as
1 Z T
%nIAE = |e(t)| dt × 100%, (5.40)
T × Fcl,r(mean) 0

where e, Fcl,r(mean) and T respectively represent the clamp force tracking error,
the mean and the period of the clamp force reference trajectories.
Figure 5.21 compares the tracking performance of the baseline PI controller
and the NTOC, where a triangular profile from 0 to 20 kN with 1 second period
was commanded. For the 40 kN/s ramp found in the triangular profile, the
NTOC demonstrates an approximately threefold improvement in %nIAE as
listed in Table 5.2.
Figure 5.22 compares the sinusoidal reference tracking performance of both
controllers. It is shown that the simplified NTOC has a larger bandwidth, and
visually appears to track the reference significantly better than the baseline
PI controller. There is a corresponding 65% reduction in %nIAE, as listed in
Table 5.2.
140 Chapter 5. High performance clamp force tracking control

Table 5.2: Tracking performance %nIAE of the baseline PI and NTOC.

Trajectory PI Simplified NTOC
Triangular 1 Hz, 0∼20 kN 15.65% 5.79%
Sinusoidal 1 kN, 1 ± 0.5 kN 28.19% 9.71%

3000
Ref. PI NTOC (5.29) NTOC (5.24)
2000
Fcl [N]

1000

0
0 0.5 1 1.5 2 2.5 3
Tracking error [N]

1000

−1000
0 0.5 1 1.5 2 2.5 3
3
Tm [Nm]

−3
0 0.5 1 1.5 2 2.5 3
Time [s]

Figure 5.22: Experimentally obtained reference tracking responses using sine

wave with 1 second period, 1 kN mean and 0.5 kN amplitude.

Additionally, the tracking response of the general NTOC (5.26) is also plot-
ted in Figure 5.22, where it shows great reduction in tracking error compared to
the PI and the simplified NTOC (5.31), and with the %nIAE of 3.02%. How-
ever, due to the employment of numerical differentiation for state reference,
the torque command is very noisy, and this may be an important consideration
for its implementation. On the other hand, the noise level found in the torque
command from the simplified NTOC is within an acceptable level.

Energy Usage. The energy usage of the baseline PI and NTOC is shown in
Figure 5.23. In order to achieve a shorter rise time, NTOC uses more energy to
achieve higher acceleration and deceleration compared to the PI. Nevertheless,
it is noted that the gain tuning proposed in Section 5.2.4 is flexible enough to
5.3. Explicit model predictive control 141

70
Baseline PI
NTOC
60

50
Energy Usage [J]

40

30

20

10

0
2 5 10 15 20 25 30
Fcl,r [kN]

Figure 5.23: Experimentally obtained energy usage for step responses initiated
from 0 N clamp force.

tradeoff energy usage with tracking speed.

5.3 Explicit model predictive control

Although the NTOC gain-tuning procedure proposed in Section 5.2.4 has

translated the selection of boundary layer size to the selection of two compro-
mising weights for reference tracking speed and energy usage, these weights
are not reflected in the original formulation of the robust time-optimal track-
ing problem (5.22). It is desirable to be able to tune these weights directly
from the problem formulation. To facilitate a holistic design approach that
incorporates convergence speed, control effort and constraints, the clamp force
tracking control problem is investigated under the model predictive control
(MPC) framework in this section.
142 Chapter 5. High performance clamp force tracking control

5.3.1 Problem formulation

The robust time-optimal tracking problem (5.22) can be adapted to incorporate
measures of convergence speed and control effort, and the resultant formulation
is called the finite horizon clamp force tracking problem, which is stated as
the following: Given an initial state x(t0 ) = x0 , a time-varying clamp force
reference Fcl,r (t) and a finite horizon tp , find a torque trajectory Tm∗ (t) such that
the weighted cost of tracking error and control effort is minimised, subjected
to the input and state constraints. This problem can be formulated as the
following:
"
 2
min Fcl (t0 + tp ) − Fcl,r (t0 + tp ) WP,c
Tm (t)
#
Z t0 +tp  2
+ Fcl (t) − Fcl,r (t) WQ,c + Tm (t) WR,c dt,
2
t0

subject to (5.1)–(5.2), (5.41)

Tm (t) ∈ U,
x(t) ∈ X ,
x(t0 ) = x0 .

Compared to the robust time-optimal tracking problem (5.22) that pos-

sesses a variable final time tf that stops whenever the state is close to the
reference, the finite horizon tracking problem (5.41) is ended at the given fixed
period tp . However, the tracking error at the end of the horizon can be pe-
nalised using the weight WP,c . Additionally, the tracking error and the control
magnitude within the horizon are included in the control formulation, where
they are weighted by WQ,c and WR,c respectively.
Since modern controller implementations involve discrete sampling rates,
it is useful to conceive the continuous-time problem formulation (5.41) in
discrete-time. To this end, the model predictive control (MPC) is introduced.

Problem 5.3 (Model predictive clamp force control). Given an initial state,
x(k) = x0 , a prediction horizon Hp , a control horizon Hu ≤ Hp , a clamp
force reference up to the prediction horizon, {Fcl,r (k + i), i = 1, . . . , Hp } and
a sampling period ts , find the optimal change in torque trajectory up to the
control horizon, ∆U ∗ := {∆Tm∗ (k + i), i = 1, . . . , Hu } such that the weighted
cost of tracking error and control deviation is minimised, subjected to the state
5.3. Explicit model predictive control 143

and input constraints. This problem can be formulated as the following:

"
 2
min Fcl (k + HP ) − Fcl,r (k + HP ) WP,d
∆U
Hp −1  2 Hu
#
+ Fcl (k + i) − Fcl,r (k + i) WQ,d + ∆Tm (k + i) WR,d ,
2
X X

i=1 i=1

subject to
  
 xd (k + i + 1) = xd (k + i) + ts f xd (k + i), Tm (k + i), θ ,

  (5.42a)

 yd (k) = h xd (k) , i = 0, . . . , Hp
Tm (k + i) = Tm (k + i − 1) + ∆Tm (k + i), i = 1, . . . , Hp (5.42b)
∆Tm (k + i) = 0, i = Hu + 1, . . . , Hp
Tm (k + i) ∈ U, i = 0, . . . , Hp
xd (k + i) ∈ X , i = 0, . . . , Hp
xd (k) = x0 .

The discretised model (5.42a) has been derived in (3.24)–(3.25), where the
function f (·) and h(·) are given in (3.23). Furthermore, control deviation is
considered here as a means of incorporating integral action within the MPC
framework to ensure offset-free tracking in the face of model uncertainties and
unknown constant disturbances (Maciejowski, 2002). The control prediction
Hp typically needs to be large to ensure stability, although extending Hp has
only minimal effect of the complexity of solving the MPC problem. For good
tracking performance, Hp is chosen large enough to include the transients of the
dominant plant dynamics. On the other hand, extending the control horizon
Hu will increase the complexity significantly. The choice of Hu is a tradeoff
between tracking precision and computation effort. Note the predicted control
input (5.42b) remains unchanged beyond the control horizon.

5.3.2 Explicit model predictive controller design

The MPC problem (5.42) is solved over a moving time horizon. For the horizon
starting at time tk , for time index k, the state of the plant x(tk ) is measured to
be used as the initial condition for (5.42). From the solution of this problem,
the first element of ∆U is injected into the plant. In an ideal condition, this
 
input drives the state of the plant towards x(k) + ts f xd (k), Tm (k) . Note the
144 Chapter 5. High performance clamp force tracking control

Figure 5.24: Motor torque deviation map with fixed reference at 13 kN.

process is repeated recursively at subsequent sampling instances. Since this

recursive strategy involves solving an optimisation problem at every time step,
it is computationally intensive.
To reduce the computational burden, the optimisation process can be per-
formed offline and the control solution is stored as a function of states. This
strategy leads to the explicit model predictive control law, given in the form
of:
 
∆Tm,mpc Fcl,r (k), Fcl (k), ωm (k), Tm (k − 1) , (5.43)

where the change in control action at time step k is a function of the reference
Fcl,r (k), measured outputs Fcl (k), ωm (k) and the previous control input Tm (k −
1).
The optimum torque deviation ∆Tm (k) is depicted in Figure 5.24 as a
function of clamp force Fcl (k), motor speed ωm (k) and motor torque Tm (k −1),
where the clamp force reference Fcl,r (k) is fixed at 13 kN. The control map
∆Tm,mpc (·) is discretised with 31 sections of clamp force reference Fcl,r from
0 kN to 30 kN, 34 sections of clamp force Fcl from 0 kN to 33 kN, 23 sections
of motor speed from -330 rad/s to 330 rad/s, and 26 sections of motor torque
Tm from -3 Nm to 3 Nm. Online implementation of the control law (5.43) is
achieved using a lookup table, where the binary search method is utilised to
locate the optimal deviation in control action ∆Tm (k).
5.3. Explicit model predictive control 145

5.3.3 Results and discussions

The control performance of the MPC is compared with the time-optimal con-
trol and the NTOC proposed in Section 5.1 and Section 5.2 respectively. Ini-
tially, the choices of prediction horizon and control horizon of the MPC are
considered. This is followed by the investigation of MPC weights, and lastly,
the implementation issues of the MPC is discussed.

Prediction horizon and control horizon. The investigation begins with

a setpoint tracking performance comparison between MPCs with different pre-
diction and control horizons. The objective function weights for the MPC are
WP,d = 50, WQ,d = 1 and WR,d = 30, and the update rate for the MPC
is 0.004 seconds. The input and output constraints are |Tm | ≤ Tm max and
|ωm | ≤ ωm max respectively. The reference is a 20 kN constant clamp force
command. In terms of tracking performance, this is a challenging manoeuver
for the EMB as it demands the controller to handle both input and output
constraints. Figure 5.25(b) shows the rise time and maximum overshoot for
each of the MPC configurations. From this graph, it can be seen that for small
prediction horizons Hp < 8, increasing prediction horizon decreases the max-
imum overshoot. Note the length of eight steps prediction horizon matches
the time required for reaching the velocity constraint from stationary with
the maximum acceleration. However, for longer prediction horizons Hp ≥ 8,
there is little to be gained in terms of reducing the maximum overshoot. For
8 ≤ Hp ≤ 10, there is a slight increase in overshoot due to control chatter
during the deceleration phase, where the objective function is insensitive to
the control. This suggests that the MPC needs at least ten steps of predicted
state trajectory ahead in order to foreseen a potential overshoot and to choose
a control deviation accordingly to avoid the overshoot. Also for Hp < 5, the
rise time is less than the time-optimal trajectory, albeit with greater overshoot.
Note that the time-optimal trajectory is the fastest possible trajectory without
overshoot.
Another observation from this graph is that the rise time actually increases
with prediction horizon if the control horizon is Hu = 1. For the single control
horizon case, the control magnitude stays constant over the prediction horizon
and a conservative control have to be chosen to avoid output constraint vio-
lation within the prediction horizon. This suggests a control horizon greater
146 Chapter 5. High performance clamp force tracking control

0.1
Hu = 2
Rise Time [s]

Hu = 3
0.08
Time−optimal control

0.06

0.04
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

80
Maximum Overshoot [%]

60

40

20

−20
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Hp

(a) Rise time and maximum overshoot of 2 kN step.

0.25
Hu = 1
Rise Time [s]

Hu = 2
0.2
Hu = 3
Time−optimal control
0.15

0.1
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

15
Maximum Overshoot [%]

10

−5
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Hp

(b) Rise time and maximum overshoot of 20 kN step.

Figure 5.25: Comparison of simulated setpoint tracking performance between

MPCs with different prediction and control horizons for 2 kN and 20 kN steps.
5.3. Explicit model predictive control 147

than one is needed in practise.

Figure 5.25(a) illustrates the setpoint tracking performance for 2 kN step.
It is observed that as the prediction horizon is long, decreasing control horizon
reduces rise time. This suggests a long control horizon is needed for a long
prediction horizon to avoid conservative control.

Weighting. To make an informed choice of MPC weights selection, control

performance using a range of weights are tested, and the rise time and energy
usage are compared. Figure 5.26 show the 0%–95% rise time and energy usage
for a 2 kN step using MPC with different weights WQ,d , WR,d . The terminal
weight is set to be constant WP,d = 50, and the horizon lengths are given as
Hp = 10 and Hu = 2. As the penalty on the control deviation is increased, as
reflected by ratio WR,d /WQ,d , the rise time typically increases while the energy
usage decreases. However, there is a dip in rise time when WR,d /WQ,d = 21
and WQ,d = 1, where the rise time is shorter than the time-optimal trajectory.
However, the energy usage with this set of weights is greater than the time-
optimal trajectory due to the overshoot in the response. As the weight on the
tracking error WQ,d increases, the rise time is decreased and the energy usage
is increased. Note that this result can be directly compared with the result
using NTOC in Figure 5.14.
Figure 5.27 compares the rise time and energy usage of the time-optimal
control, NTOC and MPC. The NTOC gains used in this graph are γ =
{0.2, 0.3} and ρ ∈ [1, 10] × 10−8 . The horizon lengths of the MPC are Hp = 10
and Hu = 2, and the ratio is WR,d /WQ,d ∈ [1, 104 ], weight on tracking error is
WQ,d = {1, 100} and terminal weight is WP,d = 50. The time-optimal control
has the shortest rise time with the least energy usage. There are several MPC
settings with shorter rise time than the time-optimal control, however with
larger energy usage due to the overshoot response. Furthermore, for the same
rise time, the MPC uses less energy than the NTOC. This is the consequence
of addressing energy usage and tracking error directly from the problem for-
mulation. Nevertheless, both NTOC and MPC show reduction in energy usage
with the expense of longer rise time when the penalties on control effort are
increased.
To compare the control algorithms, both small and large manoeuvres as
chosen as test cases, illustrated in Figure 5.28(a) and Figure 5.28(b) respec-
148 Chapter 5. High performance clamp force tracking control

0.09
Time−optimal control
0.08 WQ,d = 1
Rise Time [s]

0.07 WQ,d = 100

0.06

0.05

0.04
−2 −1 0 1 2 3 4
10 10 10 10 10 10 10

25
Energy Usage [J]

20

15

10

5
−2 −1 0 1 2 3 4
10 10 10 10 10 10 10
WR,d / WQ,d

Figure 5.26: Rise time and energy usage for a 2 kN step using the time-optimal
control and MPC with WQ,d = {1, 100} and WR,d /WQ,d ∈ [10−2 , 102 ].

tively. For both manoeuvres, the time-optimal control is the quickest to reach
the setpoints, but it experiences control chatter when the setpoints are reached.
Both NTOC and MPC are free from chatter. However, the MPC has larger
control deviations compared to the NTOC that can quickly bring the clamp
force trajectory to the setpoint. Although a properly tuned MPC potentially
achieves shorter rise time compared to a NTOC tuned for robustness, the
improvements are visually insignificant in these graphs.

Implementation issues. Since the MPC involves solving an optimisation

problem at every control update instance, the key to a successful implemen-
tation is the controller’s evaluation time. The evaluation time for the online
MPC and the explicit MPC were obtained from simulations and shown in Ta-
ble 5.3. The online MPC evaluated the optimisation problem (5.43) at each
time step during online implementation, while the explicit MPC searched for
5.3. Explicit model predictive control 149

24
MPC WQ,d = 1
22 MPC WQ,d = 100
NTOC γ = 0.2
20 NTOC γ = 0.3
Time−optimal control
18
Energy Usage [J]

16

14

12

10

6
Increasing WR,d / WQ,d (for MPC) or ρ (for NTOC).
4
0.045 0.05 0.055 0.06 0.065 0.07 0.075 0.08 0.085
Rise Time [s]

Figure 5.27: Comparison of rise time and energy usage of the time-optimal
control, NTOC and MPC.

the optimal control deviation from a lookup table, which was prepared offline
for a bounded operating region. The simulation scenario was 0.2 seconds of
13 kN step clamp force response. It was run on a 2.8 GHz Pentium 4 machine
where Simulink were used with a variable-step solver ode45 and the maxi-
mum step of 0.0002 seconds. It is observed that the online MPC requires a
very long simulation time, and is prohibitive for practical implementation on
the EMB microcontroller.
On the other hand, the explicit MPC requires significantly shorter com-
putation time, where majority of the computation time is spent on searching
for the corresponding control input on the lookup table. In order to reduce
the interpolation error of the control input, a finer grid for the control law
domain is desired. However, the size of control law lookup table increases as
the domain is discretised finer, whereby larger controller memory is needed.
Although the explicit MPC alleviates the lengthy computational issue by
150 Chapter 5. High performance clamp force tracking control

4000
Time−optimal control NTOC MPC
Fcl [N]

2000

0
0 0.02 0.04 0.06 0.08 0.1 0.12
400
x2 [rad/s]

200

−200
0 0.02 0.04 0.06 0.08 0.1 0.12
5
Tm [Nm]

−5
0 0.02 0.04 0.06 0.08 0.1 0.12
Time [s]

(a) Responses for 2 kN step.

4
x 10
4
Time−optimal control NTOC MPC
Fcl [N]

0
0 0.05 0.1 0.15 0.2
400
x2 [rad/s]

200

−200
0 0.05 0.1 0.15 0.2
5
Tm [Nm]

−5
0 0.05 0.1 0.15 0.2
Time [s]

(b) Responses for 20 kN step.

Figure 5.28: Responses of 20 kN step using the time-optimal control, NTOC

and MPC.
5.4. Conclusions 151

Table 5.3: Time required for simulating a 0.2 seconds trajectory with a 13 kN
setpoint.
MPC Algorithms Simulation Time [s]
Online MPC 255
Explicit MPC 0.177
Without controller 0.104

performing the optimisation offline and storing the results in a lookup table, it
is found that the storage requirement of the lookup table is prohibitive under
current hardware limits. This suggests the adoption of NTOC with reduced
storage requirements for practical implementation.

5.4 Conclusions
High performance clamp force tracking controllers for the prototype automo-
tive electromechanical brake (EMB) are developed in this chapter. The con-
trollers are designed such that it can track a given reference closely in the
face of nonlinear dynamics and system constraints, where the input and state
constraints are the motor torque saturation and motor torque velocity limits
respectively. A good tracking performance is desirable to meet demanding
requests from higher level controllers, such as the anti-lock braking system
(ABS) and the electronic stability control (ESC). The main contributions of
this chapter are summarised as the following:

• A time-optimal clamp force tracking control that considers both input

and state constraints is developed using Pontryagin’s minimum princi-
ple. It is shown that the time-optimal trajectory consists of maximum
acceleration and/or maximum deceleration when the velocity is within
the limits, and possibly an intermediate interval with constant maximum
speed if one of the limits is reached. The time-optimal trajectory repre-
sents the quickest possible response without overshoot and can be used
as a benchmark for system performance.
• To aid implementation, a feedback implementation of the time-optimal
tracking controller is further developed, where a state-feedback control
law that utilises a switching curve is proposed. The switching curve
is a function of system states that outputs the instance where control
152 Chapter 5. High performance clamp force tracking control

switching should occur. However, the feedback time-optimal control is

very sensitive to modelling and measurement discrepancies, and can lead
to control chatter, which is undesirable from an operational perspective.
Nevertheless, the time-optimal control can be used as a basis for a more
robust implementation.
• The near-time-optimal clamp force tracking control (NTOC) is proposed
to address the control chatter issue found in the time-optimal control.
Note the control chatter is caused by the abrupt change in control mag-
nitude across the switching curve. To mitigate this, a boundary layer
is introduced in the vicinity of the switching curve, where the control
magnitude transits gradually across the boundary layer. Under certain
assumptions, the NTOC is proved to asymptotically stabilise the track-
ing error to a uniform ultimate bounded that contains the origin.
• Practical implementation and gain tuning guidelines are also provided for
the NTOC. Note that the NTOC gains dictate the size of the boundary
layer, which is difficult to be associated with the control performance.
The proposed gain selection procedure translates the tuning effort into
the selection of weighting contrasting tracking error and control effort,
which is more readily related to the control performance.
• The NTOC is experimentally validated on the prototype EMB and
benchmarked against the time-optimal control and an industry-tuned
proportional-integral (PI) control. The PI controller is tuned such that
the available motor torque is fully utilised while avoiding integral windup
due to input and state constraints. The rise time of the NTOC is shown
to be typically within 5% of the time-optimal control, indicating that
the NTOC is able to fully utilise the available motor capacity to achieve
a rapid setpoint tracking. While the rise times of the NTOC are slightly
faster than the PI control for large setpoints, the NTOC is significantly
faster for small setpoints. This also indicates that the fixed set of PI gains
are too conservative for small step reference. Furthermore, the NTOC
also yields smaller tracking error for triangular and sinusoidal tracking
references compared to the PI control, suggesting a larger closed-loop
bandwidth.
• To incorporate tracking error, control effort and system constraints di-
rectly from the start of the control problem formulation, the clamp force
5.4. Conclusions 153

Time-optimal control Tm,to , (5.18) in Sec. 5.1

Near-time-optimal control Tm,ntoc , (5.26) and Tm,ntocs , (5.31) in Sec. 5.2

Explicit model predictive control ∆Tm,mpc , (5.43) in Sec. 5.3

EMB
Clamp Force Tm Control-
Fcl
Fcl,r Controller Oriented
Model
Fcl ωm

Figure 5.29: EMB closed-loop system and clamp force controllers developed
in this chapter.

tracking control problem is examined under the model prediction control

(MPC) framework. This holistic design approach shows greater flexibility
and slight control performance improvements compared to the NTOC.
However, the MPC is computational intensive and is not implementable
in the current hardware as an optimisation problem have to be solved
at every control update. To address this, the explicit MPC is presented,
where the optimisation is solved offline and the control solutions are
stored as a function of states in a lookup table. However, the mem-
ory requirements for the lookup table are still prohibitive under current
hardware limits.

The EMB closed-loop system and the clamp force feedback controllers in-
vestigated in this chapter are illustrated in Figure 5.29. In particular, the
achieved rapid tracking performance and high closed-loop bandwidth of the
NTOC allows additional branching problems to be investigated. In order to
attenuate brake judder, brake torque variation compensators building on the
outer loop of the clamp force controller will be investigated next.
 6

Brake torque variation

compensation with wheel
speed measurements

As alluded in Chapter 1, brake judder is a prevalent braking problem aris-

ing from a repetitive variation in brake torque. The most common type of
brake judder is cold judder, where the judder inducing brake torque variation
(BTV) is caused by geometrical irregularities of a disc rotor, such as thickness
variation and runout. The current approaches to judder are either corrective
measures (applied after the driver has experienced judder), or locally isolating
the vibration (by modifying the vibration transfer path without eradicating the
source of vibration). These approaches have been the only possible solutions
under hydraulic braking system due to the low bandwidth of the closed-loop
system. On the other hand, brake-by-wire (BBW) offers high closed-loop band-
width and fine clamp force control fidelity which provides the opportunity to
eradicate judder at the source.
This chapter proposes a novel method for attenuation of cold judder di-
rectly at the source, utilising an electromechanical brake (EMB) to actively
compensate for the judder inducing BTV. Taking advantage of the high per-
formance clamp force tracking control developed in Chapter 5, an outer-loop
compensator is designed to produce a compensating clamp force command to
cancel the BTV, and therefore reducing the judder phenomenon.

155
156 Chapter 6. BTV compensation with wheel speed measurements

ωd , ω̇d Tbtv (θd )

EMB Closed-Loop System
+
LPV Brake Fcl Brake Disc + Tb
EMB
Tb,r Torque Fcl,r Tm with Nominal
Clamp Force Control-
Variation Thickness
Controller Oriented
Compensator Model

Tb Fcl ωm

Figure 6.1: Block diagram of a brake torque variation compensator scheduled

by wheel speed and acceleration.

The development of the BTV compensator is commenced in two stages.

First, the compensator is designed in continuous-time, where a linear
parameter-varying (LPV) control is scheduled using wheel speed and accel-
eration. An estimator is constructed to estimate the wheel acceleration from
the speed measurements. Furthermore the compensator is discretised to aid
implementation in a microcontroller. Finally, the performance of the BTV
compensator is experimentally validated over fixed and varying BTV frequen-
cies by employing the production-ready prototype EMB and a dedicated brake
dynamometer.

6.1 Problem formulation

A brake torque variation (BTV) compensation algorithm is designed as an
outer-loop controller for the electromechanical brake (EMB) in order to pro-
duce a compensating clamp force that attenuates the BTV caused by geomet-
rical irregularities of a brake rotor. The algorithm proposed in this chapter
is scheduled using wheel speed and acceleration, and the compensating clamp
force command is calculated using the brake torque tracking error. A block
diagram of the EMB system with the proposed BTV compensator is depicted
in Figure 6.1.
The compensator is designed to work under small clamping force level to
address the typical occurrence of judder that is perceived during light braking
(when slow deceleration allows the vibration amplitude to grow to a noticeable
level). Additionally, the BTV magnitude, wheel angular speed and acceleration
are assumed bounded.
6.1. Problem formulation 157

Assumption 6.1 (Bounded operating region of BTV compensator). The

brake torque variation is bounded, that is Tbtv (θd ) ∈ T , where
n o
T := Tbtv : [0, ∞] 7→ R, Tbtv ∈ [−T btv , T btv ], ωd ∈ [ω d , ω d ], ω̇d ∈ [αd , αd ] .

Note that ωd := θ̇d and θd are the disc rotor angular speed and position re-
spectively.

EMB closed-loop system. To track the clamp force command requested by

the proposed compensator closely, a high performance clamp force controller is
desirable. In this work, the near-time-optimal clamp force tracking controller
(NTOC) developed in Section 5.2 is employed. Under Assumption 6.1, the lin-
earised closed-loop model for the EMB derived in (5.32) is used for developing
the proposed compensator, where the closed-loop dynamics are approximated
by
      
θ̇m 0 1 θm 0
 =  +  Fcl,r ,
kp
ω̇m −kp −kd ωm k̄(θ̄m )
  (6.1)
i θ
m
h
Fcl = k̄(θ̄m ) 0   .
ωm

Note that kp and kd can be calculated using the expressions in (5.37), which
are given by
bks b bks
kp = , kd = + ξω (0),
ssat x2 sat ssat m
where ks , ssat , x2 sat are the NTOC gains, ξωm (·) is the gradient of the switching
curve and b = TmJmax . Furthermore, consider the motor position θm can be
approximated by a fixed θ̄m and the clamp force trajectory is in the vicinity of
Fcl (θ̄m ), then a linear approximation of the clamp force can be obtained as

Fcl = k̄(θ̄m )θm , (6.2)

where the linear stiffness coefficient k̄(·) is given by

k̄(θ̄m ) = k1 θ̄m
2
+ k2 θ̄m . (6.3)

Brake torque generation model. The characteristics of brake torque gen-

eration has been investigated in Section 3.4. The brake torque, Tb is produced
158 Chapter 6. BTV compensation with wheel speed measurements

by applying a clamp force on both surfaces of a disc rotor via brake pads, and
is governed by
Tb = 2µrd Fcl + Tbtv (θd ), (6.4)

where µ, rd and Tbtv (·) are the coefficient of friction between pads and rotor,
effective disc rotor radius and BTV respectively. Exact values of the coefficient
of friction, µ is not available during online implementation and it is affected
by many environmental factors, such as temperature, speed and humidity.
However, its variation is slow relative to the system dynamics (6.1), thus the
discrepancy between the estimated friction coefficient, µ̂ and the true friction
coefficient, µ can be assumed constant. This leads to the perturbed form of
(6.4), given by
Tb = 2µ̂rd Fcl + Td + Tbtv (θd ), (6.5)

where Td is the discrepancy in brake torque resulted from the estimated error
in friction coefficient.
The BTV in (6.5) can be approximated using the following grey-box model:
n
Tbtv (θd ) = ab,k cos (kθd + φb,k ), (6.6)
X

k=1

where n is the maximum number of harmonic in the BTV model. In a typical

cold judder, only one to two pulsations occur per wheel revolution, there-
fore n = 2 is sufficient for most applications. Furthermore, the frequency of
cold judder is within the clamp force controller bandwidth. Additionally, the
amplitude ab,k and phase-shift φb,k are unknowns, and are assumed to be time-
invariant in the proceeding compensator design. However in practice, these
quantities may vary slowly relative to the disc rotations.
The grey-box model (6.6) is dependent on the disc rotor position, θd , where
its measurement is not available in production vehicles. On the other hand,
measured disc rotor speed, ωd is available as it corresponds to the wheel speed.
Therefore, it is desirable to represent (6.6) in terms of ωd for the development of
a BTV compensator that can be scheduled by wheel speed. For an appropriate
choice of initial condition, wk (0), the k-th harmonic of BTV can be represented
by the following dynamics (details are referred to Appendix A):

ẇk = Sk (ωd )wk , (6.7a)

h i
ab,k cos (kθd + φb,k ) = 1 0 wk , (6.7b)
6.1. Problem formulation 159

where    
wk,1  0 kωd 
wk =  , Sk (ωd ) =  . (6.8)
wk,2 −kωd 0
The sum of (6.7b) for all harmonic from 1 to n renders an equivalent grey-box
model of the BTV, which is governed by

w̆˙ = S̆(ωd )w̆,

n
(6.9)
h i
Tbtv = wk,1 = 1 0 · · · 1 0 w̆,
X

k=1 | {z }
n pairs.

where
 
w̆ = [w1T , . . . , wnT ]T , S̆(ωd ) = diag S1 (ωd ), . . . , Sn (ωd ) .

Note S̆(ωd ) is a square matrix comprises main diagonal blocks of Sk (ωd ) and
off-diagonal blocks of zeros.
Employing the EMB closed-loop system model (6.1), the brake torque ex-
pression (6.5) and the BTV grey-box model (6.9), the dynamics of the brake
corner can be represented in the following form:
      
ẋ
 =
A 0   x  Bu 
+ Fcl,r (6.10)
ẇ 0 S(ωd ) w 0

where x := [θm , ωm ]T is the system state and w := [Td , w1T , . . . , wnT ]T is the
exogenous system state. The matrices in (6.10) are given by
   
0 1  0
A= , Bu =  kp
,
−kp −kd k̄(θ̄m ) (6.11)
 
S(ωd ) = diag 0, S1 (ωd ), . . . , Sn (ωd ) ,

where Sk (·) is defined in (6.8). Additionally, the brake torque tracking error
can be characterised by

eb = Tb − Tb,r = Cx + Dew w − Tb,r , (6.12)

where Tb,r is the brake torque reference and the matrices are given by
h i
C = 2µ̂rd k̄(θ̄m ) 0 ,
(6.13)
h i
Dew = 1 1 0 ··· 1 0 .
| {z }
n pairs.
160 Chapter 6. BTV compensation with wheel speed measurements

Furthermore, consider process noise and measurement noise, in addition to

the following change of variable:
 
Tb,r
x̄ = x −  d k̄ 
2µ̂r
, (6.14)
0

the expressions in (6.10) can be re-written as

˙
      
x̄
 =
A 0   x̄  Bu 
+ ū + Bv v,



0 S(ωd ) w 0

 ẇ


Σbc :   (6.15)

 h i x̄
eb = C Dew   + Dv v,





w

where ū is the compensating clamp force reference and v ∈ R is a zero-mean

white noise signal with identity power spectral density. Note the total clamp
force reference is Fcl,r = ū + 2µ̂r
1
d
Tb,r .

Compensator structure. The compensator scheduled by wheel speed and

acceleration proposed in this chapter has the following structure:

 ẋc = Ac (ωd , ω̇d )xc + Bc eb
Σbtvc,s : (6.16)

Fcl,r = Cc (ωd , ω̇d )xc + Fcl,f r (Tb,r ),

where xc is the controller state, and Ac , Bc , Cc are chosen to achieve some spe-
cific control objectives that will be defined shortly. Additionally, to facilitate
brake torque tracking, the following feedforward command is defined:
1
Fcl,f r (Tb,r ) = Tb,r . (6.17)
2µ̂rd
The interconnection of the brake corner model (6.15) and the compensator
(6.16) yields the following closed-loop system:

ẋcl = Acl xcl + Bcl w,

(6.18)
eb = Ccl xcl + Dcl w,

where xcl := [x̄T , xTc ]T and the matrices are given by

   
A Bu Cc  0 
Acl =  , Bcl =  ,
Bc C Ac Bc Dew
h i
Ccl = C 0 , Dcl = Dew .
6.1. Problem formulation 161

The control objectives are to internally stabilise the closed-loop system (6.18),
to attenuate the brake torque tracking error, and to minimise a quadratic
performance metric. These objectives are formally stated as the following:

Problem 6.1 (BTV compensation with wheel speed measurements). Given an

operating region that satisfies Assumption 6.1, a fixed brake torque reference
Tb,r and a brake corner model (6.15), design a speed and acceleration scheduled
brake torque variation compensator (6.16) such that the following objectives
are achieved:

(a) Internal stability. The unforced closed-loop system

ẋcl = Acl xcl

is asymptotically stable, where the matrix Acl is Hurwitz.

(b) Output regulation. The forced closed-loop system (6.18) is asymptotically
stable, where
lim eb (t) = 0,
t→∞
n
for all x̄(0) ∈ R2 and w(0) ∈ w := [Td , w1T , . . . , wnT ]T : |Td | ≤ T d , Tbtv ∈
o  
T , where Tbtv = nk=1 [ 1 0 ]wk .
P

(c) Optimal performance. Suppose there exist Π and Γ such that

lim (x̄ − Πw) = 0, lim (ū − Γw) = 0,

t→∞ t→∞

where x̄ is defined in (6.14) and the compensating control is given by

ū = Cc (ωd , ω̇d )xc .

Let x̃ = x̄ − xc be the estimated error of states, then the optimal com-

pensator minimises the following objective function:
Z ∞   n o
Jbtvc,s := x̄T Qx̄ + ūT Rū dt + sup E kx̃(t)k2 , (6.19)
0 S∈PS

where E{·} denotes the expected value, PS := co (S(ω d ), S(ω d )) and

co (·) denotes the convex hull. Additionally ω d and ω d represent the
lower- and upper-bounds of ωd respectively.
 162 Chapter 6. BTV compensation with wheel speed measurements

6.2 Design of a brake torque variation

compensator with wheel speed
measurements
In this section, the speed and acceleration scheduled BTV compensator (6.16)
is developed utilising output regulation theory. The compensator design is
commenced in three stages. First, the compensator structure is developed
using the output regulation theory, where the differential regulator equation
is solved. This is followed by the selection of compensator gains utilising
the solution of the differential regulator equation and linear matrix inequality
(LMI) techniques. Lastly, discretisation of the compensator is performed to
aid implementation in a microcontroller.

6.2.1 Brake torque variation compensation using

output regulation theory
The compensator design begins with the application of the output regula-
tion theory, where it provides the necessary and sufficient conditions for the
existence of a compensator that solves Problem 6.1. To this end, common
assumptions in regulation problems are adapted to the setting of Problem 6.1.

Assumption 6.2. The brake corner model (6.15) has the following properties:

(BC-1) The pair (A, Bu ) is stabilisable.

(BC-2) The pair ([ C Dew ], [ A0 S0 ]) is detectable.
(BC-3) There exist c1 > 0 such that kΦ(t0 , t)k ≤ c1 for any t ≥ t0 , where Φ(t, τ )
 
is the fundamental matrix generated by S ωd (t) .

Remark 6.1. It can be verified using parameters from Table 4.3 and NTOC
gains chosen in Section 5.2 that the system (6.15) satisfies conditions (BC-1)
and (BC-2). Moreover, these conditions guarantee the internal stability of the
closed-loop system stated in Problem 6.1.
Furthermore, a fundamental matrix of S(·) is Φ(t0 , t) = diag (1, Φ1 , . . . , Φn ),
where  
cos (kθd ) sin (kθ )
d 
Φk =  , k = 1, . . . , n.
− sin (kθd ) cos (kθd )
Note that kΦ(t0 , t)k = 1, therefore (BC-3) is satisfied.
6.2. Design of a BTV compensator with wheel speed measurements 163

Solving differential regulator equation. Under Assumptions 6.1 and 6.2,

Theorem 2.1 in Ichikawa and Katayama (2006) states that a compensator solv-
ing Problem 6.1 exists if and only if there exist continuous bounded matrices
Π and Γ which satisfy the following differential regulator equation:

Π̇ + ΠS = AΠ + Bu Γ, (6.20a)
lim (CΠ + Dew ) = 0. (6.20b)
t→∞

As will be shown later, Π is a function of ωd and Γ is a function of ωd and ω̇d .

However, the function arguments are dropped for simplicity in presentation. In
the following, Π and Γ are solved algebraically, first by solving for the first row
of Π after expansion of (6.20b), followed by solving the rest of the unknowns
from the expansion of (6.20a).
Consider Π ∈ R2×(2n+1) and Γ ∈ R1×(2n+1) , where
 
Π1,1 Π1,2 Π1,3 · · · Π1,2n Π1,2n+1 
Π= ,
Π2,1 Π2,2 Π2,3 · · · Π2,2n Π2,2n+1 (6.21)
h i
Γ = Γ1 Γ2 · · · Γ2n+1 .

Substituting (6.11), (6.13) and (6.21) into the steady-state of (6.20b) yields
 
i Π · · · Π1,2n+1  h
1,1
h i
2µ̂rd k̄ 0  + 1 1 0 · · · 1 0 = 01×(2n+1) . (6.22)
Π2,1 · · · Π2,2n+1 | {z }
n pairs.

For clarity, the size of the zero matrix is shown, where 01×(2n+1) denotes a zero
matrix with one row and 2n + 1 columns. Expansion of (6.22) gives
1 1
Π1,1 = − , Π1,2k = − , Π1,2k+1 = 0, (6.23)
2µ̂rd k̄ 2µ̂rd k̄
where k = 1, . . . , n. Note the k-sequence will be assumed in the following steps
without restating. Taking the time-derivative of (6.23) leads to

Π̇1,i = 0, i = 1, . . . , 2n + 1. (6.24)

Substituting (6.11) and (6.21) into (6.20a) yields

01×2 ··· 01×2
 0 
i 02×1 S1 (ωd ) ··· 02×2
 
Π̇1,1 ··· Π̇1,2n+1 Π1,1 Π1,2 Π1,3 ··· Π1,2n
h
Π1,2n+1 
+ Π2,2n+1  .. .. ... ..

Π2,1 Π2,2 Π2,3 ··· Π2,2n
Π̇2,1 ··· Π̇2,2n+1 . . .

02×1 02×2 ··· Sn (ωd )
(6.25)
h ih
Π1,1 ··· Π1,2n+1
i h 0 i
= 0 1
−kp −kd Π2,1 ··· Π2,2n+1 + kp [ Γ1 ··· Γ2n+1 ].
k̄
164 Chapter 6. BTV compensation with wheel speed measurements

The first row of (6.25) gives the following expressions:

Π̇1,1 = Π2,1 ,
Π̇1,2k − Π1,2k+1 (kωd ) = Π2,2k , (6.26)
Π̇1,2k+1 + Π1,2k (kωd ) = Π2,2k+1 .
To obtain the terms for the second row of Π, (6.23) and (6.24) are substituted
into (6.26) to get
kωd
Π2,1 = 0, Π2,2k = 0, Π2,2k+1 = − . (6.27)
2µ̂rd k̄
Differentiating (6.27) provides
k ω̇d
Π̇2,1 = 0, Π̇2,2k = 0, Π̇2,2k+1 = − . (6.28)
2µ̂rd k̄
The results from (6.23) and (6.27) are collected to obtain Π, which is given by
 
1 −1 −1 0 −1 0 · · · −1 0 
Π(ωd ) = . (6.29)
2µ̂rd k̄ 0 0 −ωd 0 −2ωd · · · 0 −nωd
Additionally, the time derivative of Π is given by
 
1 0 0 0 0 0 ··· 0 0 
Π̇ = . (6.30)
2µ̂rd k̄ 0 0 −ω̇d 0 −2ω̇d · · · 0 −nω̇d

To find Γ, the second row of (6.25) is evaluated as the following:

kp
Π̇2,1 = −kp Π1,1 − kd Π2,1 + Γ1 ,
k̄
kp
Π̇2,2k − Π2,2k+1 (kωd ) = −kp Π1,2k − kd Π2,2k + Γ2k , (6.31)
k̄
kp
Π̇2,2k+1 + Π2,2k (kωd ) = −kp Π1,2k+1 − kd Π2,2k+1 + Γ2k+1 .
k̄
Substituting (6.29) and (6.30) into (6.31) lead to
1 (kωd )2 − kp k(ω̇d + kd ωd )
Γ1 = − , Γ2k = , Γ2k+1 = − .
2µ̂rd 2µ̂rd kp 2µ̂rd kp
Hence,

Γ(ωd , ω̇d ) = 1
2µ̂rd kp
[ −kp ωd2 −kp −(ω̇d +kd ωd ) ··· (nωd )2 −kp −n(ω̇d +kd ωd ) ]. (6.32)

Since both ωd and ω̇d are bounded (as stated in Assumption 6.1), the solutions
of Π(ωd ) (given in (6.29)) and Γ(ωd , ω̇d ) (given in (6.32)) are also bounded.
Therefore, a compensator that solves Problem 6.1 exists.
6.2. Design of a BTV compensator with wheel speed measurements 165

Remark 6.2. Note the solution of Π characterises the trajectory of x̄ with

respect to w. Under any compensator that solves Problem 6.1, the trajectories
of the closed-loop system from any initial states in x̄(0) ∈ R2 and w(0) ∈
n o
w := [Td , w1T , . . . , wnT ]T : |Td | ≤ T d , Tbtv ∈ T satisfy

lim (x̄ − Πw) = 0.

t→∞

Correspondingly, the solution of Γ characterises the trajectory of compensating

control input, ū with respect to w, where ū satisfies

lim (ū − Γw) = 0.

t→∞

Observer-based compensator structure. Utilising Π and Γ expressed in

(6.29) and (6.32) respectively, an observer-based controller can be constructed
as the following (Saberi et al., 2000):

x̂˙
           
 =
A 0   x̂  Bu  Lx h i x̂ 
+ Fcl,r +    C Dew   − eb  ,
ŵ˙ 0 S ŵ 0 Lw ŵ
  (6.33)
h i x̂
Fcl,r = Kx (Γ − Kx Π)   + Fcl,f r ,
ŵ

where x̂ ∈ R2 and ŵ ∈ R2n+1 are the estimated values of x̄ and w respectively,

and Fcl,f r is defined in (6.17). Furthermore, Kx ∈ R2 , Lx ∈ R2 and Lw ∈ R2n+1
are vectors chosen such that
 
A + Lx C Lx Dew 
A + Bu K and  (6.34)
Lw C S + Lw Dew

are both Hurwitz. Note the feedforward gain Γ − Kx Π is utilised for com-
pensating the BTV, which enters the brake corner system at the output, as
depicted in Figure 6.1. In the notation of (6.16), xc = [x̂T , ŵT ]T and
  
A + Lx C + Bu Kx Lx Dew + Bu Γ(ωd , ω̇d ) − Kx Π(ωd )
Ac (ωd , ω̇d ) =  ,
Lw C S(ωd ) + Lw Dew
 
Lx h i
Bc = −   , Cc (ωd , ω̇d ) = Kx (Γ(ωd , ω̇d ) − Kx Π(ωd )) .
Lw
166 Chapter 6. BTV compensation with wheel speed measurements

Wheel speed and acceleration estimation. Note both wheel speed and
acceleration are required for implementing the compensator (6.33), however
only the speed measurements are available. Therefore, the acceleration is
estimated from the measured speed, where the following estimator is used:

ω̂˙
    
 d = 
0 1 ω̂d 
+ Lω (ω̂d − ωd ), (6.35)
˙α̂d 0 0 α̂d

where ω̂d and α̂d are the estimated speed and acceleration respectively. The
estimator gain Lω ∈ R2 is chosen such that ([ 00 10 ] + Lw [ 1 0 ]) is Hurwitz stable.

6.2.2 Tuning for compensator gains

The controller gain Kx and observer gain L := [LTx , LTy ]T are tuned such that
the quadratic performance metric (6.19) is minimised, that is
Z ∞   n o
min x̄T Qx̄ + ūT Rū dt + sup E kx̃(t)k2 ,
Kx ,L
|0 {z } S∈PS
| {z }
Jbtvc,s1 Jbtvc,s2
(6.36)
subject to (6.15), (6.33),
PS := co (S(ω d ), S(ω d )),

where co (·) denotes the convex hull. The performance metric (6.36) is devised
based on the separation principle. The first part of the metric, Jbtvc,s1 is used for
optimising the controller gain, Kx with respect to tracking speed and control
effort, whereas the second part of the metric, Jbtvc,s2 is employed to find an
optimal observer gain L that minimises the supremum of the state estimation
error’s expected value over the wheel speed operating range.
Note the system state x̄ is considered, but the exogenous system state w
is excluded from the controller gain Kx selection. This is because the exoge-
nous system ẇ = Sw is decoupled from the input Fcl,r , and hence the pair
([ C Dew ], [ A0 S0 ]) is uncontrollable. Therefore, the controller design cannot be
arranged in the traditional linear quadratic gaussian (LQG) setting, and is
considered using two decoupled objectives, Jbtvc,s1 and Jbtvc,s2 .
The solution for minimising Jbtvc,s1 is provided by the linear quadratic reg-
ulation (LQR) theory, where the following algebraic Ricatti equation is solved
(Dorato et al., 1995):

AT Pare + Pare A − Pare Bu R−1 BuT Pare + Q = 0. (6.37)

6.2. Design of a BTV compensator with wheel speed measurements 167

The controller gain is calculated using

Kx = R−1 B T Pare .

The contradicting objectives of tracking speed and energy usage can be tuned
using the weight matrices, Q and R.
Furthermore, the solution for minimising Jbtvc,s2 is translated into the fol-
lowing LMI (de Souza and Trofino, 2000):

min tr (N ),
N,P,W
 
N BwT P + DwT W T 
subject to  ≥ 0, P > 0,
P Bw + W D w P
ATaug,i P + P Aaug,i + Caug
T
W T + W Caug + LT L < 0, i = 1, 2
(6.38)
where
   
A 0  A 0  h i
Aaug,1 = , Aaug,2 = , Caug = C Dew .
0 S(ωd ) 0 S(ωd )

The observer gain satisfies

L = P −1 W.
Note the two contrasting performance of fast estimation speed and good noise
rejection can be tuned by matrices Bv and Dv .

6.2.3 Discretisation of the compensator

To implement the compensator (6.33) in a microcontroller, emulation of the
design is investigated using two discretisation methods, the forward-Euler and
Tustin bilinear methods. The resulting discrete controller is given in the form
of
xd (k + 1) = Ad (k)xd (k) + Bd (k)eb (k),
(6.39)
Fcl,r (k) = Cd (k)xd (k) + Dd (k)eb (k) + Fcl,f r (Tb,r ),
where eb (k) = Tb (k) − Tb,r (k).
 
The expressions of Ad (k), Bd (k), Cd (k), Dd (k) depend on the discretisa-
tion method, where forward-Euler method yields the following controller ma-
trices:
Ad = I + ts Ac (ωd , ω̇d ), Bd = ts Bc ,
(6.40)
Cd = Cc (ωd , ω̇d ), Dd = 0,
168 Chapter 6. BTV compensation with wheel speed measurements

and Tustin bilinear method provides the following controller matrices:

−1
ts ts
 
Ad = I + Ac (ωd , ω̇d ) I − Ac (ωd , ω̇d ) ,
2 2
√  ts
−1
Bd = ts I − Ac (ωd , ω̇d ) Bc ,
2 (6.41)
√ 
ts
−1
Cd = ts Cc (ωd , ω̇d ) I − Ac (ωd , ω̇d ) ,
2
−1
ts ts

Dd = Cc (ωd , ω̇d ) I − Ac (ωd , ω̇d ) Bc .
2 2
Note the forward-Euler method provides simpler implementation and lower
computational demand, while bilinear method provides a discretised controller
that matches the continuous controller better at the cost of increased compu-
tational burden.

6.3 Results and discussions

6.3.1 Simulation results
Prior to experimental validation of the proposed linear-parameter varying
BTV compensator, simulations are performed to choose the compensator or-
der, gains, discretisation and sampling rate. Simulations are executed on the
nonlinear EMB model (3.6)–(3.11) with parameters provided in Table 4.3. Ad-
ditionally, to facilitate a good tracking of the clamp force reference given by
the compensator, the NTOC (5.24) developed in Section 5.2 is employed as
the clamp force controller, where the gains ks = 2, ssat = 4 and e2 sat = 50
are used. The optimisation of (6.37) and (6.38) are respectively solved using
lqr function and YALMIP Toolbox (Löfberg, 2004) with SDPT3 solver (Toh
et al., 1999) in Matlab.

Order for compensator. To investigate an appropriate order for the com-

pensator, an experimentally obtained BTV profile illustrated in Figure 6.2 is
used in simulations. As can be seen in Figure 3.18, the signal is dominated by
the first three harmonics.
Tracking responses for a 100 Nm brake torque reference with and without
compensator are compared in Figure 6.3, where ωd = 5 Hz. To facilitate
a fair comparison between compensators of the first-order, the second-order
6.3. Results and discussions 169

45

40

35
Brake Torque Variation [Nm]

30

25

20

15

10

−5
0 1 2 3 4 5 6 6.283
Disc Rotor Angular Position [rad]

Figure 6.2: Brake torque variation profile.

and the third-order, a nominal set of the controller weighting is defined as

Q = I and R = 1. Furthermore, the matrices related to noise are set to Bw =
h iT
0 1 1 0 1 · · · 0 1 and Dw = 1. The compensation performance is
| {z }
n pairs.
assessed using the root mean square of tracking error (RMSE) at steady-state.
To assess the attenuation of the sinusoidal variations, the RMSEs without the
DC offsets are calculated. Then, a dimensionless quantity can be obtained
by normalising the RMSEs (without DC bias) of the compensated responses
with respect to the uncompensated trajectories, where the Normalised RMSE
is given by
RMSE w/o DC Offset of Compensated Response
Normalised RMSE = .
RMSE w/o DC Offset of Uncompensated Response
The first-order compensator, which represents the simplest compensator design
exhibits 74% reduction in RMSE compared to the uncompensated response,
as listed in Table 6.1. The second-order compensator shows greater reduction
in BTV, where it is reduced by a further 16%. Although the third-order
compensator demonstrates the greatest reduction in BTV, where it is decreased
by another 3% compared to the second-order compensator, the reduction is
170 Chapter 6. BTV compensation with wheel speed measurements

140
Tb [Nm]

120

100

80
0 0.5 1 1.5 2 2.5 3 3.5 4

1000

800

600
Fcl [N]

400

200
Uncomp. 1st 2nd 3rd
0
0 0.5 1 1.5 2 2.5 3 3.5 4
Time [s]

Figure 6.3: Simulated responses for BTV compensators with different orders
of harmonic.

Table 6.1: Compensation performance and power usage of compensators with

different orders of harmonic.
Compensator Normalised Average Peak
Order RMSE Power (W) Power (W)
1 26.4% 5.1 13.5
2 10.0% 11.1 47.4
3 7.1% 12.4 50.7
No compensation 100% 0 0

insignificant from a practical perspective and does not justify the increased
complexity in implementation. Therefore, compensator that includes up to
the second-order harmonic is anticipated to perform well in practise.
Additionally, the average and peak power requirements of the compen-
sators in steady-state over five wheel revolutions are compared in Table 6.1.
Unsurprisingly, compensators of higher order require more power in order to
achieve a better BTV reduction due to increased motor activity. The peak
power required by the third-order compensator is 50.7 W, which is well within
the 1 kW power rating of the EMB motor.
6.3. Results and discussions 171

140
Tb [Nm]
120

100

80
0 0.5 1 1.5 2 2.5 3 3.5 4

1000

800

600
Fcl [N]

400

200
Q/R = 1 Q/R = 103 Q/R = 106
0
0 0.5 1 1.5 2 2.5 3 3.5 4
Time [s]

Figure 6.4: Simulated brake torque and clamp force trajectories for different
Q/R ratios.

Gain tuning. In order to study the effects of tuning parameters, they are
perturbed one at a time from the nominal set of values, starting from the Q/R
ratio, then the measurement noise matrix Dw , followed by the process noise
matrix Bw . Brake torque and clamp force trajectories for different Q/R ratios
for a second-order compensator are shown in Figure 6.4. A larger Q/R ratio
yields a larger norm on the controller gain Kx . As a result, the feedforward gain
(Γ − Kx Π) decreases and the BTV compensation performance is decreased.
Compensation performance for different magnitudes of measurement noise
Dw are compared in Figure 6.5. Larger measurement noise indicates that the
measurements are unreliable and thus leads to a conservative estimator gain
and a slower compensation response as expected.
To study the effect of process noise on compensation performance, the nom-
inal value of Bw is perturbed in the form of Bw = [0, 1, 1, 0, ν, 0, ν]T , where ν is
the process noise standard deviation for the sinusoidal BTV model. Compen-
sation performance for different magnitudes of ν are compared in Figure 6.6.
A larger observer gain is obtained for a larger process noise, leading to a faster
172 Chapter 6. BTV compensation with wheel speed measurements

140
Tb [Nm]

120

100

80
0 0.5 1 1.5 2 2.5 3 3.5 4

1000

800

600
Fcl [N]

400

200 Dw = 1 Dw = 2 Dw = 10
0
0 0.5 1 1.5 2 2.5 3 3.5 4
Time [s]

Figure 6.5: Simulated brake torque and clamp force trajectories for different
Dw magnitudes.

compensation response. However, overshoot response may be observed if the

observer gain is too large, as demonstrated by the ν = 5 case.

Discretisation method. A discretised compensator is required for prac-

tical implementation on a microcontroller. To this end, discretisation of
the compensator is performed using forward-Euler method and Tustin bi-
linear method, where the respective controller matrices are given in (6.40)
and (6.41). Figure 6.7 compares brake torque responses produced by the
continuous-time and the discretised second-order compensators. For a small
sampling period (ts = 0.0001 seconds), both discrete-time controllers emulate
the continuous-time compensator closely. However, for a larger sampling pe-
riod (ts = 0.001 seconds), the discrete-time controller obtained using forward-
Euler method becomes unstable, but the discrete-time controller derived from
Tustin bilinear method compares favourably with the continuous-time compen-
sator. This suggests that the sampling period can be larger if Tustin bilinear
method is used. Since the continuous-time compensator is parameter-varying,
online discretisation using Tustin’s method involves heavy computational bur-
6.3. Results and discussions 173

180
160
140
Tb [Nm]

120
100
80
60
0 0.5 1 1.5 2 2.5 3 3.5 4

1500

1000
Fcl [N]

500

ν = 0.1 ν=1 ν=5

0
0 0.5 1 1.5 2 2.5 3 3.5 4
Time [s]

Figure 6.6: Simulated brake torque and clamp force trajectories for different
ν magnitudes, where the process noise is defined as Bw = [0, 1, 1, 0, ν, 0, ν]T .

den due to numerous division operations and can be prohibitive in implemen-

tation. Although discretisation using forward-Euler with a sampling period
of ts = 0.0001 seconds has an adequate emulation performance, the sampling
rate is too high for the current hardware setup.
Figure 6.8 compares brake torque responses produced by the continuous-
time and the discretised first-order compensators. Compared to the second-
order compensator, the Euler discretised first-order compensator is stable at
slow wheel speed (ωd = 3 Hz) and large sampling time (ts = 0.001 s), hence
indicating that a lower-order Euler-discretised compensator has less stringent
requirement on sampling rate. However, it remains unstable at higher wheel
speed (ωd = 10 Hz), suggesting that a higher sampling rate is required to
handle a broader range of wheel speed.

6.3.2 Experimental results

Experimental validation of the proposed compensator was conducted on the
dedicated brake dynamometer described in Section 3.1.3. The compensator
174 Chapter 6. BTV compensation with wheel speed measurements

ts = 0.0001 s, ωd = 3 Hz ts = 0.0001 s, ωd = 10 Hz

140 Cont. Time 140

Tustin
Euler Uncompensated
Tb [Nm]

Tb [Nm]
120 120 Envelope

100 100

80 80
0 2 4 6 0 0.5 1 1.5 2
ts = 0.001 s, ωd = 3 Hz
ts = 0.001 s, ωd = 10 Hz

140
140
Tb [Nm]

Tb [Nm]

120 120

100 100

80 80
0 2 4 6 0 0.5 1 1.5 2
Time [s] Time [s]

Figure 6.7: Comparison of discretisation methods and sampling period for a

second-order compensator on different wheel speeds.

was implemented on Vector CANape, which is a software with built-in scripting

capability and facilitates communication with the EMB via a CAN bus. This
setup has a limitation on the lowest applicable sampling period at 0.001 sec-
onds. Additionally, the computational demand of using Tustin bilinear method
to discretise the compensator “on-the-fly” is very heavy and therefore is pro-
hibitive for online implementation. Furthermore, the second-order Euler-
discretised compensator is unstable at 0.001 seconds sampling period, hence
only the first-order discrete compensator was assessed in experiments.

Fixed wheel speed. The first experimental test case corresponds to brak-
ing under fixed wheel speed, which commonly occurs when light braking is
applied to a vehicle descending a ramp. Figure 6.9 shows the BTV compen-
sation for different wheel speeds. Due to the reduction of friction coefficient
over increasing wheel speed, a larger clamp force reference was given (before
compensation) for faster wheel speed. As the compensator was started, the
6.3. Results and discussions 175

ts = 0.0001 s, ωd = 3 Hz ts = 0.0001 s, ωd = 10 Hz

140 Cont. Time 140

Tustin Uncompensated
Envelope
Tb [Nm]

Tb [Nm]
120 Euler 120

100 100

80 80
0 2 4 6 0 0.5 1 1.5 2

ts = 0.001 s, ωd = 3 Hz ts = 0.001 s, ωd = 10 Hz

140 140
Tb [Nm]

120 Tb [Nm] 120

100 100

80 80
0 2 4 6 0 0.5 1 1.5 2

Figure 6.8: Comparison of discretisation methods and sampling period for a

first-order compensator on different wheel speeds.

amplitude of the brake torque responses were attenuated, and the averages
of the responses approached the setpoint at 100 Nm. In Figure 6.9(a), some
fluctuations are seen around 28 seconds as well as that around 6.5 seconds
in Figure 6.9(b). These fluctuations were caused by dropped packets in the
CAN bus. Note that the BTV algorithm was implemented and tested in Vector
CANape, which runs under Windows XP. Since Windows XP is not a real-time
operating system, packet drop may occur during testing.
Due to the integration action included in the compensator, the mean of the
responses were corrected and moved towards the setpoints when the compen-
sator was started. The normalised RMSEs of the compensated responses are
listed in Table 6.2, indicating that the BTV was attenuated for different wheel
speeds. However, the attenuation performance decreases as the wheel speed is
increased.
To investigate the reason behind the reduced attenuation performance as
the wheel speed is increased, the amplitude spectrum of the brake torque tra-
176 Chapter 6. BTV compensation with wheel speed measurements

140
← Compensation Started
120
Tb [Nm]

100

80

60
0 5 10 15 20 25 30

1400
Fcl [N]

1200

1000

800
0 5 10 15 20 25 30
Time [s]

(a) ωd = 1 Hz.

140
← Compensation Started
120
Tb [Nm]

100

80

60
0 2 4 6 8 10

1800

1600
Fcl [N]

1400

1200

0 2 4 6 8 10
Time [s]

(b) ωd = 3 Hz.

Figure 6.9: Experimentally obtained brake torque variation compensation us-

ing the first-order compensator discretised with forward-Euler method. (Part
1 of 2)
6.3. Results and discussions 177

160
← Compensation Started
Tb [Nm] 140

120

100

80

0 1 2 3 4 5 6

1800

1600

1400
Fcl [N]

1200

1000

800
0 1 2 3 4 5 6
Time [s]

(c) ωd = 5 Hz.

160
← Compensation Started
140
120
Tb [Nm]

100
80
60
40
0 1 2 3 4 5 5.7

2000

1500
Fcl [N]

1000

500
0 1 2 3 4 5 5.7
Time [s]

(d) ωd = 7 Hz.

Figure 6.9: Experimentally obtained brake torque variation compensation us-

ing the first-order compensator discretised with forward-Euler method. The
sampling period and brake torque reference are ts = 0.001 s and Tb,r = 100 Nm
respectively. (Part 2 of 2)
178 Chapter 6. BTV compensation with wheel speed measurements

Table 6.2: Normalised RMSE of compensated responses.

Wheel Speed, ωd (Hz) Normalised RMSE
1 54.9%
3 68.9%
5 70.7%
7 94.9%

10
ωd = 1 Hz
5

0
0 1 2 3 4 5 6
Amplittude [Nm]

10
ωd = 3 Hz
5

0
0 2 4 6 8 10 12 14 16 18
20
ωd = 5 Hz
10

0
0 5 10 15 20 25 30
20
ωd = 7 Hz
10

0
0 5 10 15 20 25 30 35 40
Frequency [Hz]

Figure 6.10: Amplitude spectrum of brake torque trajectories for different

wheel speeds, where the peaks before and after compensation are represented
by circles and squares respectively.

jectories before and after attenuation are compared in Figure 6.10. The am-
plitude spectrum shows the amplitude distribution as a function of frequency,
and can be obtained using fast fourier transform (FFT). It is observed that
the amplitudes for the first harmonic (corresponds to the wheel speed) of the
brake torque responses decrease with compensation, indicating that the first-
order compensator is attenuating the first harmonic disturbance as intended.
No changes are expected from the second and third harmonics since the first-
order compensator only addresses the first harmonic of the BTV. However,
the attenuation action of the compensator also introduces low frequency noises
6.3. Results and discussions 179

150
← Compensation Started
Tb [Nm] 100

50
0 2 4 6 8 10 11.6
2000
Uncomp.
Fcl [N]

Comp.
1500

1000
0 2 4 6 8 10 11.6

6
ωd [Hz]

4
0 2 4 6 8 10 11.6
Time [s]

Figure 6.11: Experimentally obtained brake torque responses with and without
compensation under varying wheel speed.

with growing amplitudes as the wheel speed is increased, which leads to the
diminishing attenuation performance. This is due to the decreased emulation
performance of the forward-Euler discretisation method as the wheel speed
is increased. Nevertheless, the attenuated amplitudes outweighed the newly
introduced noise, as the RMSE of the compensated responses are decreased.

Variable wheel speed. The second experimental test case corresponds to

light braking under variable wheel speed, which reflects the general braking
scenario where judder may occur. Figure 6.11 demonstrates the brake torque
compensation under a varying ωd , starting from 6 Hz with 2 rad/s2 decelera-
tion. The brake torque reference was fixed at 100 Nm. Due to the discrepancy
of friction coefficient estimates, the uncompensated response has an offset from
the reference. However, with the compensator activated, the DC offset is sig-
nificantly reduced. Additionally, the BTV is attenuated even under varying
wheel speed, where the normalised RMSE is 57.2%.
Figure 6.12 illustrates another braking scenario with both varying wheel
speed and brake torque reference. The brake torque reference was initially set
180 Chapter 6. BTV compensation with wheel speed measurements

200
← Compensation Started
Tb [Nm]

100

0
0 5 10 15 20
2000
Fcl [N]

1000
Uncomp.
Comp.
0
0 5 10 15 20

5
ωd [Hz]

3
0 5 10 15 20
Time [s]

Figure 6.12: Experimentally obtained brake torque responses with and without
compensation under varying wheel speed and brake torque reference (thick
dash grey line).

at 100 Nm, then reduced to 20 Nm at 9.5 seconds, and gradually increased

to 100 Nm from 12.5 seconds. The ωd was started at 3 Hz and was gradually
increased to 5 Hz from 9.5 seconds. It can be seen that the amplitude of BTV
for the compensated response is attenuated, with the normalised RMSE of
66.5%.

6.4 Conclusions
A novel method for attenuating cold judder directly from the source of vi-
bration is proposed, where an electromechanical brake (EMB) is utilised to
generate a compensating clamp force to counteract the judder causing brake
torque variation (BTV). To this end, a BTV compensator that handles time-
varying disturbance frequencies is developed. Specific outcomes of this chapter
include:

• To facilitate the development of a model-based compensator design, a

6.4. Conclusions 181

linearised brake corner model incorporating the EMB and a disc rotor is
derived. The model assumes a bounded operation region and provides
the trace of brake torque given a clamp force trajectory.
• Utilising the brake corner model and the grey-box BTV model formulated
in Section 3.4, a BTV compensator scheduled using the measured wheel
speed and estimated acceleration is proposed. Output regulation theory
is employed to determine the existence of a compensator and to pro-
vide the feedforward gains for complete rectification of the BTV, where
a differential regulator equation is solved. Furthermore, the observer-
based controller structure is adapted for the proposed compensator by
incorporating the feedforward gains.
• A quadratic objective function is formulated to aid tuning the compen-
sator gains. Contrasting objectives of tracking speed and energy usage
are considered. Furthermore, the expected value of state estimation er-
rors are minimised. Tuning of the gains in the face of varying wheel
speed is handled using linear matrix inequality (LMI) formulations.
• To facilitate implementation of the compensator on a microcontroller,
discretisation using forward-Euler method and Tustin bilinear method
are investigated. Although the bilinear method provides better a emu-
lation performance, the computational load is overwhelming for current
implementation. On the other hand, the forward-Euler method is simple
to implement, albeit requires very small sampling period.
• The compensator is validated experimentally on the EMB and a dedi-
cated brake dynamometer, where the compensator is discretised using
forward-Euler method. Experimental results show that compensator is
able to reduce the root mean square error (RMSE) by up to 55%, and
the attenuation is consistent for both fixed and varying wheel speeds.
Better performance is expected if higher order compensators are tested.
However, this is limited by the stringent sampling rate requirement of
the forward-Euler method.
 7

Brake torque variation

compensation with wheel
position measurements

A novel method for brake judder attenuation using an electromechanical brake

is demonstrated in Chapter 6, whereby the brake torque variation (BTV) that
causes judder is attenuated using a linear parameter-varying (LPV) compen-
sator. The compensator is scheduled using measured wheel speed and es-
timated acceleration. The discretised compensator, however, requires high
sampling rate which is prohibitive in the current setup.

To alleviate the high sampling rate requirement, an adaptive feedforward

brake torque variation compensator is proposed, where the measured wheel an-
gular position is added as an extra scheduling variable. To this end, the brake
judder compensation problem utilising wheel position measurements is defined,
followed by the development of the adaptive compensator and a model-based
gain tuning procedure. Finally, the proposed compensator is experimentally
validated on the production-ready prototype electromechanical brake (EMB)
and a dedicated brake dynamometer.

183
184 Chapter 7. BTV compensation with wheel position measurements

θd , ωd , ω̇d Tbtv (θd )

EMB Closed-Loop System
+
Adaptive Fcl Brake Disc + Tb
Feedforward EMB
Tb,r Fcl,r Tm with Nominal
Brake Torque Clamp Force Control-
Thickness
Variation Controller Oriented
Compensator Model
Fcl ωm
Tb

Figure 7.1: Block diagram of a brake torque variation compensator scheduled

by wheel position, speed and acceleration.

7.1 Problem formulation

The characteristics of brake torque generation has been investigated in Sec-
tion 3.4, and is depicted in Figure 7.1. The brake torque variation (BTV) is
modelled as an output disturbance that is periodic with respect to the wheel
rotations. However, the amplitude and phase-shift of the BTV are unknowns.
This chapter proposes utilising an adaptive feedforward BTV compensator to
estimate the BTV profile, and to utilise the profile to produce a compensating
clamp force that actively attenuates the judder inducing BTV.
Similar to the linear parameter-varying (LPV) compensator developed in
Chapter 6, the adaptive feedforward compensator is designed as an outer-loop
controller for the electromechanical brake (EMB). In addition to the brake
torque tracking error, wheel speed measurements and estimated wheel accel-
eration available to the LPV compensator, the adaptive feedforward compen-
sator also assumes the availability of wheel position measurements. A block
diagram of the EMB system with the adaptive feedforward BTV compensator
is depicted in.

Brake corner model. The operating region of the LPV compensator stated
in Assumption 6.1 is adopted in this chapter for the adaptive feedforward
compensator design. Furthermore, by neglecting the noise in the brake corner
model (6.15), the nominal brake corner model is given by

x̄˙
      
 =
A 0   x̄  Bu 
+ ū,
ẇ 0 S(ωd ) w 0
  (7.1)
h i x̄
eb = C Dew   ,
w
7.1. Problem formulation 185

where the compensating control for the BTV is given by ū and the system
matrices A, Bu , C, Dew , S are given in (6.11) and (6.13). Note the overall clamp
force command is Fcl,r = ū + Fcl,f r (Tb,r ).

Compensator structure. The adaptive feedforward compensator is sched-

uled by wheel position, speed and acceleration. It involves adaptive identifica-
tion of γ̂ ∈ R2n+1 , which is utilised to produce the estimate of w in the form
of ŵ = Φ(θd )γ̂. The dynamics and output of the compensator are governed by
γ̂˙ = g$(θd )eb ,
(7.2)
Fcl,r = Γ(ωd , ω̇d )Φ(θd )γ̂ + Fcl,f r (Tb,r ),
where g ∈ R is the compensator gain, Γ(ωd , ω̇d ) ∈ R1×(2n+1) is the feedforward
gain vector, and $(θd ) and Φ(θd ) are scheduled by the wheel position, given
by
h iT
$(θd ) = 1 cos (θd ) sin (θd ) · · · cos (nθd ) sin (nθd ) ,
| {z }
n pairs.
 
Φ(θd ) = diag 1, Φ1 (θd ), . . . , Φn (θd ) .

The diagonal matrices in Φ(θd ) are given by

 
cos (kθd ) sin (kθd ) 
Φk (θd ) =  , k = 1, . . . , n
− sin (kθd ) cos (kθd )
Additionally, the feedforward clamp force reference is defined as
1
Fcl,f r (Tb,r ) = Tb,r .
2µ̂rd
The control objective is to attenuated the brake torque tracking error, which
stated as the following:

Problem 7.1 (BTV compensation with wheel position and speed measure-
ments). Given an operating region that satisfies Assumption 6.1, a fixed brake
torque reference Tb,r and a brake corner model (7.1), design a position, speed
and acceleration scheduled brake torque variation compensator (7.2) such that
the tracking error is attenuated, that is

lim eb (t) = 0,
t→∞
n o
for all x̄(0) ∈ R2 and w(0) ∈ w := [Td , w1T , . . . , wnT ]T : |Td | ≤ T d , Tbtv ∈ T ,
 
where Tbtv = [ 1 0 ]wk .
Pn
k=1
186 Chapter 7. BTV compensation with wheel position measurements

Remark 7.1. Problem 7.1 corresponds to the output regulation objective stated
in (b) of Problem 6.1. Although criteria for internal stability and optimal per-
formance are not specified in Problem 7.1 (which are analogous to (a) and (c)
in Problem 6.1), they will be addressed shortly in the following by Assump-
tion 7.1 and Proposition 7.1.

7.2 Design of a brake torque variation

compensator with wheel position
measurements
The development of an adaptive feedforward BTV compensator (7.2) is com-
menced in two stages. First, the error dynamics and the feedforward gain
vector, Γ(·) are derived. This is followed by the selection of compensator gains
utilising linear matrix inequality (LMI) techniques.

7.2.1 BTV compensation using adaptive feedforward

cancellation
The compensator design begins with the derivation of the tracking error dy-
namics equation. To this end, the following change of variables are defined:
eb
x̄1 = − Dew w,
2µ̂rd k̄
ėb (7.3)
x̄2 = − Dew Sw.
2µ̂rd k̄
By substitution of (7.3) into the brake corner model (7.1), repeated differen-
tiation and minor algebraic manipulations, the tracking error system can be
shown to be described by

χ̇ = Aχ + B̄u ū − B̄u Γ(ωd , ω̇d )w, (7.4)

where χ := [eb , ėb ]T is the error state vector. The matrices B̄u and Γ(·) are
given by:
 
0 
B̄u =  , (7.5)
2µ̂rd kp
Γ(ωd , ω̇d ) = 1
2µ̂rd kp
[ −kp ωd2 −kp −(ω̇d +kd ωd ) ··· (nωd )2 −kp −n(ω̇d +kd ωd ) ]. (7.6)
7.2. Design of a BTV compensator with wheel position measurements 187

The A matrix can be made Hurwitz stable by proper selection of the near-
time-optimal clamp force controller (NTOC) gains. Therefore, the following
assumption can be made.

Assumption 7.1. The brake corner model (7.1) has a Hurwitz stable A ma-
trix.

Note Assumption 7.1 imposes internal stability on the tracking error system
(7.4). In other words, the tracking error system is stable in the absence of
BTV and compensating clamp force command, where w = 0 and ū = 0. To
enforce regulation of the tracking error eb in the presence of BTV, the adaptive
feedforward compensator (7.2) is adopted, where the compensating clamp force
command is given by
ū = Γ(ωd , ω̇d )ŵ,

where ŵ denotes the estimation of w, given by

ŵ = Φ(θd )γ̂. (7.7)

The γ is estimated using the adaptive control law in (7.2). By defining the
following estimation error:
w̃ = ŵ − w, (7.8)

the tracking error dynamics (7.4) can be rewritten as

χ̇ = Aχ + B̄u Γ(ωd , ω̇d )w̃. (7.9)

To analyse the dynamics of the estimation error, (7.8) is differentiated to

obtain

w̃˙ = ŵ˙ − ŵ
= Φ̇(θd )γ̂ + Φ(θd )γ̂˙ − Sw
= S(ωd )Φ(θd )γ̂ + gΦ(θd )$(θd )eb − Sw
= S(ωd )w̃ + gΦ(θd )$(θd )eb
= S(ωd )w̃ + gDew
T
eb , (7.10)

where (7.2), (7.7), (7.8) and the following expressions:

Φ̇(θd ) = S(ωd )Φ(θd ), Φ(θd )$(θd ) = Dew

T
188 Chapter 7. BTV compensation with wheel position measurements

are used. By combining (7.9) and (7.10), the dynamics for both tracking error
and estimation error are given by
    
χ̇
 =
A B̄u Γ(ωd , ω̇d )  χ 
. (7.11)
w̃˙ gDew [ 1 0 ]
T
S(ωd ) w̃

Under Assumption 6.1, the system matrix

 
A B̄u Γ(ωd , ω̇d )
Ae (ωd , ω̇d ) =  T , (7.12)
gDew [ 1 0 ] S(ωd )

is bounded in a convex hull, that is Ae (·) ∈ PA , where

 
PA := co Ae (ω d , αd ), Ae (ω d , αd ), Ae (ω d , αd ), Ae (ω d , αd ) .

7.2.2 Stability and decay of error dynamics

In the following, sufficient condition for the exponential stability of the er-
ror system (7.11) is established. Additionally, a gain tuning procedure that
maximises the decay of tracking error and estimation error is proposed.

Proposition 7.1. Consider the BTV Compensation Problem 7.1 and suppose
that Assumptions 6.1 and 7.1 hold. The adaptive feedforward compensator
(7.2) with a gain g ∈ R exponentially stabilises the error system (7.11) if there
exists a matrix P such that the following conditions are satisfied:

P > 0, ATe,i P + P Ae,i < 0, i = 1, . . . , 4 (7.13)

where Ae,i are the vertices of the convex hull PA , given by

Ae,1 = Ae (ω d , αd ), Ae,2 = Ae (ω d , αd ),
Ae,3 = Ae (ω d , αd ), Ae,4 = Ae (ω d , αd ).

Furthermore, the lower bound of the decay rate in the error system (7.11)
satisfies the following optimisation problem:

max α (7.14a)
subject to P > 0, ATe,i P + P Ae,i + 2αP ≤ 0, i = 1, . . . , 4 (7.14b)
7.2. Design of a BTV compensator with wheel position measurements 189

Proof. The sufficient condition (7.13) follows directly from the definition of
quadratic stability (Boyd et al., 1994). Note the quadratic stability is a suffi-
cient (but not necessary) condition for the exponential stability of (7.11).
Consider a quadratic Lyapunov function V (ξ) = ξ T P ξ, where ξ =
[χT , w̃T ]T . Suppose dV (ξ)/dt ≤ −2αV (ξ) for all trajectories, then
   
V ξ(t) ≤ V ξ(0) e−2αt . (7.15)

By substituting the following expressions:

λmin (P )kξ(t)k2 ≤ ξ(t)T P ξ(t), λmax (P )kξ(0)k2 ≥ ξ(0)T P ξ(0)

into (7.15), where λmin (P ) and λmax (P ) denote that smallest and largest eigen-
values of P respectively, it can be shown that

λmin (P )kξ(t)k2 ≤ λmax (P )kξ(0)k2 e−2αt . (7.16)

By taking the square root of both sides of (7.16), it then follows that
r
−αt λmax (P )
kξ(t)k ≤ e λmin (P )
kξ(0)k

for all trajectories. Therefore the decay rate of the (7.11) is at least α. Note
the condition that dV (ξ)/dt ≤ −2αV (ξ) for all trajectories is equivalent to the
following LMI:

ATe,i P + P Ae,i + 2αP ≤ 0, i = 1, . . . , 4

Due to the product between α and P , the optimisation problem (7.14) is

nonlinear and cannot be solved directly using a linear semidefinite program-
ming (SDP) solver. Alternative to using a nonlinear SDP solver, a linear SDP
solver can be employed by concurrently using the bisection method, where the
procedure is described as follows.

Procedure 7.1 (Bisection Algorithm). The lower bound of the decay rate, α
in (7.14) can be solved using the following steps:

1. Find a feasible lower bound on α, and let it be αl .

2. Find an upper bound on α by increasing α until (7.14b) becomes infea-
sible, and let it be αu .
190 Chapter 7. BTV compensation with wheel position measurements

3. Perform bisection on α. First, let α = (αl + αu )/2 and check (7.14b)

for feasibility. If feasible, let α = αl and repeat Step 3. If infeasible, let
α = αu and repeat Step 3.
4. Stop when αu − αl < ε, where ε is the tolerance for α.

Remark 7.2. Admittedly with some conservatism, the estimation of the lower
bound of the decay rate in the parameter-dependent error system (7.11) can
be reduced to a LMI problem and hence is numerical tractable. The estimated
value can then be employed for tuning of the observer gain, g.
Note the compensator structure (7.2) can be generalised by consid-
ering a compensator gain matrix with diagonal elements, where g :=
diag (g1 , . . . , g2n+1 ) ∈ R(2n+1)×(2n+1) . The additional flexibility allows fine tun-
ing of the compensation performance.

7.3 Results and discussions

7.3.1 Simulation results

Prior to experimental validation of the proposed adaptive feedforward BTV
compensator, simulations are performed to choose the compensator gain and to
examine the effects of compensator discretisation. Simulations are executed on
the nonlinear EMB model (3.6)–(3.11) with parameters provided in Table 4.3.
Additionally, to facilitate a good tracking of the clamp force reference given
by the compensator, the NTOC (5.24) developed in Section 5.2 is employed
as the clamp force controller, where the gains ks = 2, ssat = 4 and e2 sat = 50
are used. The optimisation problem (7.14) is solved using YALMIP Toolbox
(Löfberg, 2004) with SDPT3 solver (Toh et al., 1999) in Matlab.

Gain tuning. A rapid decay of the error system (7.11) is desirable, as it

corresponds to a fast attenuation of BTV. The lower bound of the decay rate
can be estimated using Procedure 7.1. Consider a first order adaptive feed-
forward compensator operating in slow wheel speed and deceleration, where
(ω d , ω d ) = (0.5, 10) Hz and (αd , αd ) = (−5, 0) rad/s2 , the lower bounds of the
7.3. Results and discussions 191

0.09

0.08

0.07
Lower Bound of Decay Rate, α

0.06

0.05

0.04

0.03

0.02

0.01

0
0 1 2 3 4 5 6 7 8 9 10
Compensator Gain, g

Figure 7.2: The lower bound of the decay rate, α versus the compensator gain,
g.

error decay rate is plotted against compensator gains in the range of [0, 10]
in Figure 7.2. It is shown that the compensator gain g = 5 yields the largest
decay rate.
Furthermore, an experimentally obtained BTV profile illustrated in Fig-
ure 6.2 is employed to simulate the BTV attenuation performance of the pro-
posed compensator. Figure 7.3 illustrates the BTV compensation for first-
order compensators with different gains, where g = {1, 5, 10}. A fixed brake
torque reference, Tb,r = 100 Nm is given. The compensator with g = 5 reaches
steady state in the shortest time as expected. However, it also yields a larger
overshoot compared to the compensator with g = 1. The compensator with
g = 10 has the largest overshoot. Furthermore, the conservativeness of Fig-
ure 7.2 can be seen here, where the α for g = 10 is zero, hence suggesting that
the response may be unstable. However, the simulated trajectory in Figure 7.3
shows that the response is stable.
Figure 7.4 demonstrates the BTV attenuation responses for compensators
with a compensator gain g = 3 and different orders of harmonics. As ex-
192 Chapter 7. BTV compensation with wheel position measurements

200

150
Tm [Nm]

100

50

0
0 0.5 1 1.5 2 2.5 3 3.5 4

1500

1000
Fcl [N]

500

Uncomp. g=1 g=5 g=10

0
0 0.5 1 1.5 2 2.5 3 3.5 4
Time [s]

Figure 7.3: Simulated responses of BTV compensation for compensators with

different gains. The uncompensated response is also shown for comparison.

pected, the third-order compensator achieves the best BTV attenuation at the
expense of higher computational load. Compared to the first-order compen-
sator, a significant improvement in attenuation is observed in the second-order
compensator. However, compared to the second-order compensator, the im-
provement achieved by the third-order compensator is insignificant and does
not justify the increased in implementation complexity.

Compensator discretisation. Before implementing the proposed compen-

sator on a microcontroller, it is discretised using forward-Euler method with a
sampling period of 0.004 seconds to match rate used in the clamp force control
loop. Note the chosen sampling rate is slower compared to the rate used in the
linear-parameter varying BTV compensator proposed in Chapter 6, therefore
reducing the computational requirements.
Figure 7.5 compares the simulated responses of continuous-time and dis-
cretised compensators, where the compensator gain is chosen as g = 3. The
responses of first-order and second-order discretised compensators match their
continuous-time counterparts. This result suggests that the discretised adap-
7.3. Results and discussions 193

200

150
Tm [Nm]

100

50
0 0.5 1 1.5 2 2.5 3 3.5 4

1500

1000
Fcl [N]

500

Uncomp. 1st 2nd 3rd

0
0 0.5 1 1.5 2 2.5 3 3.5 4
Time [s]

Figure 7.4: Simulated responses of BTV compensation for compensators with

different orders of harmonic. The uncompensated response is also shown for
comparison.

tive feedforward BTV compensator can operate at a lower sampling rate com-
pared to the LPV compensator. Additionally, due to the lower sampling rate
requirement, a higher order discretised adaptive feedforward compensator can
be implemented, therefore achieving a better BTV attenuation.

7.3.2 Experimental results

Experimental investigation of the proposed compensator was conducted on a
dynamometer of a production-ready prototype EMB described in Chapter 3.
The proposed compensator was implemented on a PC with Vector CANape,
and a sampling rate of 250 Hz was used. The sampling rate was chosen to
match the clamp force control loop, as a higher rate increases the computa-
tional requirement.
Additionally, a disc rotor with thickness variation (DTV) of approximately
18 µm was employed, in which the DTV is significant enough to cause judder
when included in a vehicle chassis. Furthermore, Fourier analysis of the DTV
194 Chapter 7. BTV compensation with wheel position measurements

ωd = 5 Hz, 1st order ωd = 12 Hz, 1st order

160 160
Cont.
140 Euler 140
Tb [Nm]

120 120

100 100

80 80
0 1 2 3 4 0 0.5 1 1.5

ωd = 5 Hz, 2nd order ωd = 12 Hz, 2nd order

160 160

140 140
Tb [Nm]

120 120

100 100

80 80
0 1 2 3 4 0 0.5 1 1.5
Time [s] Time [s]

Figure 7.5: Simulated responses of continuous-time and discretised compen-

sators.

waveform revealed that the first harmonic has the largest amplitude, followed
by smaller amplitudes of the second and the third harmonics. This suggests
that BTV compensators of up to three orders are sufficient in implementation
for this case. First, experiment results with fixed wheel speed are presented.
This scenario frequently occurs when light braking is applied while a vehicle is
descending a ramp. This is then followed by experiment results with general
braking profiles.

Fixed wheel speed. Figure 7.6 illustrates the BTV compensation for two
different wheel speeds, where ωd = {1, 7} Hz. The test sequence had a fixed
brake torque reference at 150 Nm and began without BTV compensation. The
first-order compensator was started at 4 seconds, followed by the second-order
compensator at 8 seconds, and lastly the third-order compensator was started
at 12 seconds. Due to the variability of friction coefficient, the DC offset of the
uncompensated brake torque trajectories differ from the setpoint at 150 Nm.
As soon as the first-order compensator was switched on, the BTV amplitude
decreased. Additionally, the DC offset reduced and the trajectory was centred
7.3. Results and discussions 195

Table 7.1: Compensation performance of compensators with different orders

of harmonic.
Wheel Speed, Compensator RMSE RMSE w/o Normalised
ωd (Hz) Order DC offset RMSE
1 6.85 6.84 63.1%
1 2 4.06 4.06 37.4%
3 3.66 3.65 33.7%
Uncompensated 13.0 10.8 100%
1 11.0 10.4 66.9%
7 2 5.61 5.52 35.5%
3 5.34 5.33 34.3%
Uncompensated 45.9 15.5 100%

at the setpoint thanks to the integral action within the compensator. Further-
more, additional attenuation was observed when the second-order compensator
was switched on. However, the improvement in BTV attenuation of the third-
order compensator was visually insignificant compared to the second-order
compensator.
To qualitatively assess the BTV attenuation, the root mean square error
(RMSE) of the brake torque tracking trajectories are calculated and listed in
Table 7.1. To remove the DC offsets and isolate the sinusoidal variations in
brake torque trajectories, the RMSEs without the DC offsets are also calcu-
lated. Additionally, a dimensionless quantity can be obtained by normalising
the RMSEs (without DC bias) of the compensated responses with respect to
the uncompensated trajectories, where the Normalised RMSE is given by
RMSE w/o DC Offset of Compensated Response
Normalised RMSE = .
RMSE w/o DC Offset of Uncompensated Response
For the 1 Hz wheel speed scenario, the first-order compensator reduced the
normalised RMSE by 37%, and the second-order compensator achieved another
26% improvement. However, the improvement obtained by the third-order
compensator was only 4%, and is insignificant from a practical perspective.
Similar trend was also observed for the 7 Hz wheel speed scenario, where the
first-order compensator reduced the normalised RMSE by 34%, followed by
another 31% for the second-order compensator, and a 1% improvement for the
third-order compensator.
Compared to the LPV compensator operating at 7 Hz wheel speed (Fig-
ure 6.9), the BTV attenuation attained by adaptive feedforward compensator
196 Chapter 7. BTV compensation with wheel position measurements

250
← Uncomp. → ← 1st order → ← 2nd order → ← 3rd order →

200
Tb [Nm]

150

100
0 2 4 6 8 10 12 14 16

1000

900

800
Fcl [N]

700

600

500
0 2 4 6 8 10 12 14 16
Time [s]

(a) ωd = 1 Hz.

250
← Uncomp. → ← 1st order → ← 2nd order → ← 3rd order →

200
Tb [Nm]

150

100
0 2 4 6 8 10 12 14 16

1500
Fcl [N]

1000

500
0 2 4 6 8 10 12 14 16
Time [s]

(b) ωd = 7 Hz.

Figure 7.6: Experimentally obtained BTV compensation for compensators

with different orders of harmonic. The brake torque reference is fixed at
150 Nm.
7.4. Conclusions 197

is significantly better visually (even though a lower brake torque setpoint was
used in the LPV compensator test case, where the BTV was smaller). Note
a second-order adaptive feedforward compensator can be practically imple-
mented. However, the second-order LPV compensator cannot be practically
implemented in the current setup due to the high sampling rate requirement.
Amplitude spectrum of the uncompensated and compensated responses are
depicted in Figure 7.7. The amplitude spectrum is used to illustrate the ampli-
tude distribution of a signal as a function of frequency. As the BTV is caused
by the DTV, the first harmonic of the uncompensated brake torque response
has the largest amplitude, followed by the second and third harmonics. As
expected, when the first-order compensator is activated, the amplitude of the
first harmonic (of the compensated brake torque response) is reduced. How-
ever, the amplitude of the second harmonic remains unchanged. When the
second-order compensator is activated, the peaks for both the first and the
second harmonics are attenuated (the peak for the second harmonic is below
1 Nm, and is omitted from the plot). Note the first-order and second-order
compensators cannot attenuate amplitudes for the third and higher order har-
monics.

Variable wheel speed. Figure 7.8 shows the brake torque compensation
under a varying wheel speed, starting from 6 Hz while simulating a light brak-
ing event with 2 rad/s2 deceleration. It is demonstrated that the second-order
compensator achieves 44% reduction in normalised RMSE even under varying
speed. Furthermore, the DC bias found in the uncompensated response was
also corrected by the compensator.

7.4 Conclusions
An adaptive feedforward brake torque variation (BTV) compensator is pro-
posed in this chapter, addressing the need of high sampling rate required by
the linear parameter-varying (LPV) BTV compensator developed in Chap-
ter 6. The reduced sampling rate requirement of the adaptive feedforward
compensator allows BTV compensation at higher wheel speed and attenua-
tion of BTV with higher order harmonics. The main contributions of this
chapter are summarised as follows:
198 Chapter 7. BTV compensation with wheel position measurements

15
ωd = 1 Hz
Amplitude [Nm]

10

0
0 1 2 3 4 5 6
Frequency [Hz]

15
ωd = 7 Hz
Amplitude [Nm]

10

0
0 5 10 15 20 25 30 35 40
Frequency [Hz]

Figure 7.7: Amplitude spectrum of brake torque trajectories for different wheel
speeds. The peaks for uncompensated responses are represented by circles,
peaks for compensated responses using the first-order and the second-order
compensators are represented by squares and diamonds respectively.

• Utilising adaptive control techniques, a BTV compensator scheduled us-

ing wheel position, speed and acceleration is proposed. The compensator
is capable of estimating the BTV, and then generates a compensating
clamp force command online.
• A sufficient condition for exponential stability of the tracking error and
BTV estimation error is proposed by employing the concept of quadratic
stability and linear matrix inequality (LMI) techniques. Additionally,
the compensator gain is tuned by minimising the decay rate of the error
system.
• The adaptive feedforward compensator is validated experimentally on
a production-ready prototype EMB and a dedicated brake dynamome-
ter. Experimental investigations demonstrate better BTV attenuation
compared to the LPV compensator.
7.4. Conclusions 199

300
← Compensation Started
Tb [Nm]

200

100

0 2 4 6 8 10 12

2000
Fcl [N]

1000
Uncomp. 2nd order
0
0 2 4 6 8 10 12

6
ωd [Hz]

5
4
3
0 2 4 6 8 10 12
Time [s]

Figure 7.8: Experimentally obtained brake torque responses with and without
compensation under varying wheel speed.
 8

Conclusions

This thesis investigates control techniques for clamp force tracking and brake
torque variation compensation for a production-ready prototype automotive
electromechanical brake (EMB). The review of existing control approaches in
Chapter 2 identified the gap in literature for the active compensation of brake
torque variation, which is the main cause of brake judder. To support control
development for the EMB, a rapid parameter identification for a reduced-
order EMB model was proposed. This was followed by the development of
high performance clamp force controllers, where the time-optimal control, ro-
bust near-time-optimal control (NTOC) and explicit model predictive con-
trol (MPC) were considered. The NTOC was implemented on a production
ready prototype EMB and showed favourable performance compared to the
industry-tuned cascaded proportional integral controller. The achieved high
closed-loop bandwidth of this design was then employed as the inner control
loop in the brake torque compensator design. Two compensation methods, the
linear parameter-varying (LPV) and the adaptive feedforward compensators
were examined. The three specific research aims stated in Section 2.5 were an-
swered throughout Chapters 4 to 7. In the following, the main contributions
are summarised and future research directions are identified.

201
202 Chapter 8. Conclusions

8.1 Contributions
Developed a rapid EMB parameter identification procedure

A rapid parameter identification procedure for the EMB was proposed, which is
intended to support quick model-based controller tunings and post-production
quality control checks. The optimal experimental design problem was formu-
lated and an optimal experimental trajectory was determined with the aim of
minimising the experimental time and maximising parameter estimation ac-
curacy. Additionally, two parameter estimation techniques, the output error
method and prediction error method were investigated in order to infer the
model parameters from experimental measurements. It was observed that the
prediction error method provides a marginally better estimation accuracy com-
pared to the output error method but requires significantly longer evaluation
time. Compared to the incumbent approach, the proposed rapid parameter
identification procedure significantly reduces the time required for experiments
from nine hours to one minute.

Developed fast clamp force tracking controllers

High performance clamp force tracking is desirable to meet demanding re-

quests from higher level controllers, such as the anti-lock braking (ABS) sys-
tem and the electronic stability control (ESC). However, clamp force control
design is complicated by the nonlinear dynamics and system constraints. A
time-optimal clamp force tracking control that considers both input and state
constraints was developed using Pontryagin’s minimum principle. The time-
optimal trajectory represents the quickest possible response without overshoot
and was used as a benchmark for system performance.
However, the time-optimal control is very sensitive to modelling and mea-
surement discrepancies, and can lead to control chatter, which is undesirable
from an operational perspective. Therefore, a more robust near-time-optimal
clamp force tracking control (NTOC) was proposed. Practical implementa-
tion and gain tuning guidelines were also devised. The gain tuning effort was
translated into the selection of weighting contrasting tracking error and control
effort, which is more readily related to the control performance. The NTOC
was experimentally validated on the EMB and benchmarked against the time-
optimal control and an industry-tuned proportional-integral (PI) control. The
8.2. Directions for further work 203

rise time of the NTOC was shown to be typically within 5% of the time-optimal
control, indicating that the NTOC is able to almost fully utilise the available
motor capacity to achieve a rapid setpoint tracking without causing control
chatter. Furthermore, the NTOC better utilises the available motor torque
compared to the PI control, and is significantly faster for small setpoints.
Additionally, a holistic control design approach incorporating tracking er-
ror, control effort and system constraints was examined under the model
prediction control (MPC) framework. Although this design approach shows
greater flexibility and slight control performance improvements compared to
the NTOC, the intensive computational and large memory demands are pro-
hibitive under current hardware limits.

Developed brake torque variation compensators

The achieved rapid tracking performance of the NTOC allows additional

branching problems to be investigated. In particular, a novel method for
attenuating cold judder directly from the source of vibration was proposed,
where the EMB is utilised to generate a compensating clamp force to counter-
act the judder causing brake torque variation (BTV). Two BTV compensation
approaches were investigated. The first approach was the linear parameter-
varying (LPV) BTV compensator, where it is scheduled using the measured
wheel speed and estimated acceleration. A gain tuning procedure that consid-
ers control effort, tracking error and noises was also presented. Experimental
investigations demonstrated favorable results that warrant further investiga-
tion into this concept, whereby the proposed method reduces BTV by up to
55%. However, only the first-order harmonic is attenuated by the discretised
compensator, and this is limited by the stringent sampling rate requirement
of the forward-Euler method.
The second approach considered the adaptive feedforward compensation
scheme, where the compensator is scheduled using the wheel position, speed
and acceleration. Experimental investigations demonstrated improved BTV
attenuation compared to the LPV compensator.

8.2 Directions for further work

There are several opportunities to extend the present work, most notably:
204 Chapter 8. Conclusions

• Develop and extend the output regulation theory for a time-varying dis-
crete system. This result will be beneficial to the development of a
discrete LPV BTV compensator, and possibly reducing the stringent
sampling rate requirement needed for implementation.
• While the adaptive feedforward scheme has been successfully imple-
mented on the EMB system to attenuate the BTV, which can be per-
ceived as an output disturbance, it is unclear whether the current ap-
proach can be generalised to any output regulation problem. A general-
isation of this design approach will be useful for solving a wide range of
vibration attenuation problems.
• In the current work, the proposed BTV compensators have been exper-
imentally validated on the dedicated brake dynamometer. However, it
will be interesting to validate this on a larger brake dynamometer with
higher speed and larger torque capacity. Furthermore, disc rotors with
different DTV waveforms shall be used in the elaborated tests. A stan-
dardised testing procedure should be used, for example, the SAE J3002
“Dynamometer Low-Frequency Brake Noise Test Procedure” (which is
currently a standard work in progress and has been not released to the
public).
• It is noted that the proposed BTV compensator designs utilise brake
torque and/or disc angular position measurements from the dynamome-
ter, which may be lacking in actual implementation in a production vehi-
cle. While it is not uncommon to have brake torque measurement setup
in prototype vehicles, equipping this sensor in a near future production
vehicle is not expected. Nevertheless, brake torque may be estimated us-
ing the available vehicle acceleration measurements. Furthermore, disc
position may be approximated by integrating the measured wheel speed.
• The concept of utilising a brake to actively attenuate brake judder maybe
extended to hydraulic brakes, and the feasibility of this extension war-
rants further investigations.
• Braking action is abrasive in nature. Therefore, it remains a distinct
possibility that the disc thickness variation causing the judder can be
machined away using the brake calliper itself. The research questions
are what type of braking scenarios this approach might be possible, and
the required extensions to the original compensation schemes.
 A

Periodic signals
.
A periodic signal d(·) cyclical in domain X with period p can be charac-
terised as
d(x + p) = d(x), ∀x ∈ X , x + p ∈ X .
It can be decomposed into the sum of a set of simple oscillating functions using
Fourier series (Kreyszig, 2006), which carries the form of
∞  
d(x) = a0 + ak cos (kωx) + bk sin (kωx) ,
X

k=1

whereby the fundamental frequency is denoted by

2π
ω= ,
p
and the Fourier coefficients are given by the Euler formulas
1Z 2
p

a0 = d(x) dx,
p − p2
2Z 2
p

ak = d(x) cos (kωx) dx,

p − p2
2Z 2
p

bk = d(x) sin (kωx) dx, k = 1, 2, . . .

p − p2
Alternatively, the periodic signal can be represented as a linear combination of
harmonically related complex exponentials, or in the form of a complex Fourier
series, given by
∞
d(x) = ck ejkωx , (A.1)
X

k=−∞

205
206 Appendix A. Periodic signals

Periodic signal
Initial function d(t)
−ps
+ e
t t
−p 0 + 0 p 2p 3p

Figure A.1: Generator of periodic signals.

where the complex Fourier coefficients are

1Z 2
p

ck = d(x)e−jkωx dx, k = 0, ±1, ±2, . . .

p −2p

Consider that d(·) = d(t) is defined in time-domain, where t ≥ 0. Then,

(A.1) can be transformed into frequency-domain by the application of Laplace
transform,
∞
ck
D(s) = L{d(t)} = (A.2)
X
,
k=−∞ s − jkω

where the fact that

1
L{e−at } =
s+a
is used. Equation (A.2) indicates that a time-periodic signal has an in-
finite number of poles located along the imaginary axis, s = jkω, where
k = 0, ±1, ±2, · · · .
Furthermore, the periodic signal can be generated by a time-delay system
given by
e−ps
1 − e−ps
with an appropriate initial function, as shown in Figure A.1.

Models for deterministic disturbances. A particular disturbance con-

sidered here are those that can be described as the output of a linear dynamic
system having zero input and certain specific initial conditions, which is given
by the model (Goodwin et al., 2001):

dq d(t) q−1
X dk d(t)
+ γk = 0. (A.3)
dtq k=0 dtk

Note that (A.3) can be arranged into a space-space form:

ẋd = Ad xd , xd (0) given

d = Bd xd .
 207

Utilisation of Laplace transform on (A.3) leads to

N (s)xd (0)
D(s) = ,
Γd (s)
where Γd (s) is the disturbance generating polynomial given by
q−1
Γd (s) := sq + γk sk . (A.4)
X

k=0

For a periodic disturbance, the generating polynomial (A.4) will carry the form
of the denominator in (A.2).

Models for sinusoidal disturbances. A sinusoidal disturbance with a

fixed frequency of oscillation ω can be represented by

d(t) = a cos (ωt) + b sin (ωt). (A.5)

Note (A.5) can also be equivalently described by

d(t) = A cos (ωt + φ),

where
a = A cos (φ), b = −A sin (φ).
If the following new variables are defined:
 
A cos (ωt + φ) 
xd1 = ,
−Aω sin (ωt + φ)
then the following model for the sinusoidal disturbance (A.5) can be obtained:
 
0 1

=  2  xd1 ,

 ẋ


d1
Σ1 : −ω 0 (A.6)

 h i
d = 1 0 xd1 .



Alternatively, if the following new variables are defined:

 
A cos (ωt + φ) 
xd2 = ,
−A sin (ωt + φ)
then an alternative sinusoidal disturbance model is given by
 
0 ω

=

 ẋ xd2 ,


d2
Σ2 :  −ω 0 (A.7)
 h i
d = 1 0 xd2 .


208 Appendix A. Periodic signals

Similarly, for a sinusoidal disturbance with time-varying frequency,

   
dv (t) = a cos ωv (t)t + b sin ωv (t)t ,

the following two state-space realisations can be obtained:

    
0 1 0 ωv (t)


 ẋd1v =   ẋd2v = 
 

  xd1v 
 xd2v
Σ1v : −ωv (t)2 ω̇v (t)
ωv (t) Σ2v :  −ωv (t) 0

 h i  h i
dv = 1 0 xd1v dv = 1 0 xd2v

 


(A.8)
Comparing (A.6), (A.7) and (A.8), Bodson (2004) argued that the state-space
realisation in the form of Σ2 is more versatile as it can be directly extended to
Σ2v for disturbances with time-varying frequency.
 B

Expansion of equation 4.11

.
The terms in (4.11) are expanded in this appendix. Note that the parameter
vector, θ is given by:
h i h i
θ = θ1 θ2 · · · θ7 = k1 k2 TC Ts0 Tv µs ωs ,

where its relationship with the nominal values, θnom,i is

θi
θ̄i = , i = 1, 2, . . . , 7
θnom,i
Additionally, the output function, h(·) is scaled using the following expression:

h̄ = αh h,

where  
10000−1 0 
αh =  .
0 100−1
The Jacobian ∂fd
∂xT
is given by:
d

∂fd h i
= ∂fd ∂fd
,
∂xTd ∂θm ∂ωm

where
 
∂fd 1
=  ts ∂Fcl   ,
∂θm − J ∂θm N + µs tanh (εωm )
1+(ωm /ωs )2
" ts #
∂fd
=
 
.
 
∂ωm − tJs Tv +ε TC +
Ts −TC
1+(ωm /ωs )2
(1−tanh (εωm )2 )− 2ωmω(T2s(−TC ) tanh (εωm )
1+ω 2 /ω 2 )
2 +1
s m s

209
210 Appendix B. Expansion of equation 4.11

Note that the differentiation of the stiffness with respect to the motor position
is given by:
∂Fcl
= 3k1 θm2
+ 2k2 θm .
∂θm
Furthermore, the stiction is governed by:
Ts = Ts0 + µs Fcl .
The Jacobian ∂ h̄
∂xT
is given by:
d
 
∂ h̄ ∂Fcl
0
= α h
 ∂θm .
∂xdT
0 1
The expressions for ∂fd
∂ θ̄i
are procured as:
 
∂fd 0 
= 3 tanh (εω )
µs k1,nom θm

∂ θ̄1 − tJs N k1,nom θm
3
+ 1+ωm 2 /ω 2
m
s
 
∂fd 0 
= 2 tanh (εω )
µs k2,nom θm

∂ θ̄2 − tJs N k2,nom θm
2
+ 1+ωm 2 /ω 2
m
s
 
∂fd 0
=  
∂ θ̄3 − tJs (TC,nom tanh (εωm )) 1 − 2 /ω 2
1+ωm
1
s
 
∂fd 0
=  
∂ θ̄4 − tJs (Ts0,nom tanh (εωm )) 1+ω21 /ω2
m s
 
∂fd 0
= 
∂ θ̄5 − tJs Tv,nom ωm
 
∂fd 0
=  
∂ θ̄6 − tJs (µs,nom tanh (εωm )) 1+ωF2cl/ω2
m s
 
∂fd 0
=  ts .
  
∂ θ̄7 − J (2ωs,nom ωm tanh (εωm )) ω3 (1+ω2 /ω2 )2
2 Ts −TC
s m s

Finally, the expressions for ∂ h̄

∂ θ̄i
are obtained as:
 
3
∂ h̄ k1,nom θm
= αh  
∂ θ̄1 0
 
2
∂ h̄ k2,nom θm
= αh  
∂ θ̄2 0
 
∂ h̄ 0
=  , m = 3, 4, . . . , 7
∂ θ̄m 0
 Bibliography
Abdelhamid, M. K. (1997), Brake judder analysis: Case studies, in ‘SAE Con-
ference Proceedings P’, vol. 3, pp. 1225–1230. SAE Paper No. 972027.

Albassam, B. A. (2002), ‘Optimal near-minimum-time control design for flex-

ible structures’, Journal of guidance, control, and dynamics 25(4), pp. 618–
625.

Arcara, P., Bittanti, S. and Lovera, M. (2000), ‘Periodic control of helicopter

rotors for attenuation of vibrations in forward flight’, IEEE Transactions
On Control Systems Technology 8(6), pp. 883–894.

Armstrong-Hélouvry, B., Dupont, P. and De Wit, C. C. (1994), ‘A survey of

models, analysis tools and compensation methods for the control of machines
with friction’, Automatica 30(7), pp. 1083–1138.

Åström, K. J. (1980), ‘Maximum likelihood and prediction error methods’,

Automatica 16(5), pp. 551–574.

Bartley, G. F. (2000), Boeing B-777: Fly-by-wire flight controls, in C. R.

Spitzer, ed., ‘The Avionics Handbook’, CRC Press, chapter 11.

Baumann, D., Hofmann, D., Vollert, H., Nagel, W., Henke, A., Foitzik, B. and
Goetzelmann, B. (2011), ‘Self boosting electromechanical friction brake’,
United States Patent US8002088B2.

Bodson, M. (2004), Equivalence between adaptive cancellation algorithms and

linear time-varying compensators, in ‘Proceedings of the IEEE Conference
on Decision and Control’, vol. 1, pp. 638–643.

Bodson, M., Sacks, A. and Khosla, P. (1994), ‘Harmonic generation in adaptive

feedforward cancellation schemes’, IEEE Transactions on Automotic Control
39(9), pp. 1939–1944.

211
212 Bibliography

Bosch, R. (2004), Automotive Handbook, 6th edn, Robert Bosch GmbH.

Boyd, S., Ghaoui, L. E., Feron, E. and Balakrishnan, V. (1994), Linear ma-
trix inequalities in system and control theory, vol. 15 of Studies in Applied
Mathematics, Society for Industrial Mathematics.
Boyd, S. and Vandenberghe, L. (2004), Convex optimization, Cambridge Uni-
versity Press.
Bretz, E. A. (2001), ‘By-wire cars turn the corner’, IEEE Spectrum
38(4), pp. 68–73.
Briere, D. and Traverse, P. (1993), Airbus A320/A330/A340 electrical flight
controls - a family of fault-tolerant systems, in ‘The Twenty-Third Interna-
tional Symposium on Fault-Tolerant Computing’, pp. 616–623.
Byrnes, C. I., Priscoli, F. D. and Isidori, A. (1997), Output regulation of un-
certain nonlinear systems, Birkhauser.
Canudas de Wit, C., Olsson, H., Astrom, K. J. and Lischinsky, P. (1995),
‘A new model for control of systems with friction’, IEEE Transactions on
Automatic Control 40(3), pp. 419–425.
Chen, D. and Paden, B. (1990), Nonlinear adaptive torque-ripple cancellation
for step motors, in ‘Proceedings of the IEEE Conference on Decision and
Control’, pp. 3319–3324.
Dang, J. N. (2007), Statistical analysis of the effectiveness of electronic stabil-
ity control (ESC) systems, Technical Report DOT HS 810 794, Department
of Transportation, National Highway Traffic Safety Administration, Wash-
ington, DC.
Danielson, L. (2005), ‘Electric braking debuts in military and commercial ap-
plications’, Aerospace Engineering 2005(9), pp. 42.
de Souza, C. E. and Trofino, A. (2000), A linear matrix inequality approach to
the design of robust h2 filters, in L. El Ghaoui and S. lulian Niculescu, eds,
‘Advances in linear matrix inequality methods in control’, SIAM, chapter 9,
pp. 175–185.
de Varies, A. and Wagner, M. (1992), The brake judder phenomenon, in ‘SAE
International Congress and Exposition’. SAE Paper No. 920554.
Desoer, C. A. and Wing, J. (1961), ‘A minimal time discrete system’, IRE
Transactions on Automatic Control 6(2), pp. 111–125.
Dhanda, A. and Franklin, G. F. (2009), ‘An improved 2-DOF proximate time
optimal servomechanism’, IEEE Transactions on Magnetics 45(5), pp. 2151–
2164.
Bibliography 213

Dietrich, J., Gombert, B. and Grebenstein, M. (2001), ‘Electromechanical

brake with self-energization’, United States Patent US6318513B1.

Doi, K., Mibe, T., Matsui, H., Tamasho, T. and Nakanishi, H. (2000), ‘Brake
judder reduction technology–brake design technique including friction ma-
terial formulation’, JSAE review 21(4), pp. 497–502.

Dorato, P., Abdallah, C. and Cerone, V. (1995), Linear quadratic control: An

introduction, Prentice Hall.

Duan, C. and Singh, R. (2011), ‘Analysis of the vehicle brake judder problem
by employing a simplified source–path–receiver model’, Proceedings of the
Institution of Mechanical Engineers, Part D: Journal of Automobile Engi-
neering 225(2), pp. 141–149.

Engel, H. G., Hassiotis, V. and Tiemann, R. (1994), System approach to brake

judder, in ‘25th FISITA Congress’, pp. 332–339.

Fieldhouse, J. D. and Beveridge, C. (2001), An experimental investigation of

hot judder, in ‘SAE Brake Colloquium and Engineering Display’. SAE Paper
No. 2001-01-3135.

Fleming, A. J. (2010), ‘Nanopositioning system with force feedback for high-

performance tracking and vibration control’, IEEE/ASME Transactions on
Mechatronics 15(3), pp. 433–447.

Francis, B. A. (1977), ‘The multivariable servomechanism problem’, SIAM

Journal on Control and Optimization 15(3), pp. 486–505.

Francis, B. A. and Wonham, W. M. (1975), ‘The internal model principle

for linear multivariable regulators’, Applied Mathematics & Optimization
2(2), pp. 170–194.

Gassmann, S. and Engel, H. G. (1993), Excitation and transfer mechanism

of brake judder, in ‘International Pacific Conference on Automotive Engi-
neering and High Temperature Engineering Conference’. SAE Paper No.
931880.

Goodwin, G. C. (1971), ‘Optimal input signals for nonlinear-system identifica-

tion’, Proceedings of the Institution of Electrical Engineers 118(7), pp. 922–
926.

Goodwin, G. C., Graebe, S. F. and Salgado, M. E. (2001), Control system

design, Prentice Hall, New Jersey.

Gruber, S., Semsch, M., Strothjohann, T. and Breuer, B. (2002), ‘Elements of

a mechatronic vehicle corner’, Mechatronics 12(8), pp. 1069–1080.
214 Bibliography

Guo, X. and Bodson, M. (2010), ‘Equivalence between adaptive feedfor-

ward cancellation and disturbance rejection using the internal model
principle’, International journal of adaptive control and signal processing
24(3), pp. 211–218.

Haigh, M. J., Smales, H. and Abe, M. (1993a), Vehicle judder under dynamic
braking caused by disc thickness variation, in ‘Braking of Road Vehicles’,
pp. 247–258. IMechE paper C444/022/93.

Haigh, M., Smales, B. and Abe, M. (1993b), RTV - A friction material designers
view, in ‘SAE Brake Colloquium and Engineering Display’. SAE Paper No.
933070.

Halasy-Wimmer, G. and Völkel, J. (2005), ‘Electromechanically actuatable

brake’, United States Patent US6889800B2.

Hamilton, Jr, W. (1972), ‘On nonexistence of boundary arcs in control prob-

lems with bounded state variables’, IEEE Transactions on Automatic Con-
trol 17(3), pp. 338–343.

Hara, S., Omata, T. and Nakano, M. (1985), Synthesis of repetitive control

systems and its application, in ‘IEEE Conference on Decision and Control’,
vol. 24, pp. 1387–1392.

Hara, S., Yamamoto, Y., Omata, T. and Nakano, M. (1988), ‘Repetitive con-
trol system: a new type servo system for periodic exogenous signals’, IEEE
Transactions on Automatic Control 33(7), pp. 659–668.

Hartl, R. F., Sethi, S. P. and Vickson, R. G. (1995), ‘A survey of the maximum

principles for optimal control problems with state constraints’, SIAM Review
37(2), pp. 181–218.

Hess, D. P. and Soom, A. (1990), ‘Friction at a lubricated line contact operating

at oscillating sliding velocities’, Journal of Tribology 112(1), pp. 147–152.

Holweg, E. G. M., Klomp, R. L., Klaassens, J. B. and Lomonova, E. A. (2000),

Modeling and inverse model-based control of an electro-mechanical brake ac-
tuator, in ‘1st IFAC Conference on Mechatronic Systems’, vol. 1, Darmstadt,
Germany, pp. 39–44.

Huang, J. (2004), Nonlinear output regulation: Theory and applications, SIAM.

Ichikawa, A. and Katayama, H. (2006), ‘Output regulation of time-varying

systems’, Systems & Control Letters 55(12), pp. 999–1005.

Inoue, T., Nakano, M. and Iwai, S. (1981), High accuracy control of servomech-
anism for repeated contouring, in ‘Annual Symposium on Incremental Mo-
tion Control Systems and Devices’, pp. 258–292.
Bibliography 215

Jacobson, D. H., Lele, M. M. and Speyer, J. L. (1971), ‘New necessary condi-

tions of optimality for control problems with state-variable inequality con-
straints’, Journal of Mathematical Analysis and Applications 35(2), pp. 255–
284.

Jacobsson, H. (1997), Wheel suspension related disc brake judder, in ‘Proceed-

ings of the 1997 ASME Design Engineering Technical Conferences’. ASME
Paper DETC97/VIB-4165.

Jacobsson, H. (2001), Brake Judder, PhD thesis, Chalmers University of Tech-

nology.

Jacobsson, H. (2003a), ‘Aspect of disc brake judder’, Proceedings of the Insti-

tution of Mechanical Engineers, Part D: Journal of Automobile Engineering
217(6), pp. 419–430.

Jacobsson, H. (2003b), ‘Disc brake judder considering instantaneous disc thick-

ness and spatial friction variation’, Proceedings of the Institution of Mechan-
ical Engineers, Part D: Journal of Automobile Engineering 217(5), pp. 325–
342.

Jo, C., Hwang, S. and Kim, H. (2010), ‘Clamping-force control for

electromechanical brake’, IEEE Transactions On Vehicular Technology
59(7), pp. 3205–3212.

Kao, T. K., Richmond, J. W. and Douarre, A. (2000), ‘Brake disc hot spotting
and thermal judder: an experimental and finite element study’, International
Journal of Vehicle Design 23(3/4), pp. 276–296.

Kar, I. N., Seto, K. and Doi, F. (2000), ‘Multimode vibration control of a

flexible structure using h∞ -based robust control’, IEEE/ASME Transactions
on Mechatronics 5(1), pp. 23–31.

Karnopp, D. (1985), ‘Computer simulation of stick-slip friction in mechanical

dynamic systems’, Journal of Dynamic Systems, Measurement, and Control
107(1), pp. 100–103.

Keerthi, S. and Gilbert, E. (1987), ‘Computation of minimum-time feedback

control laws for discrete-time systems with state-control constraints’, IEEE
Transactions on Automatic Control 32(5), pp. 432–435.

Khalil, H. K. (2002), Nonlinear Systems, 3rd edn, Prentice Hall.

Kim, M.-G., Jeong, H.-I. and Yoo, W.-S. (1996), Sensitivity analysis of chassis
system to improve shimmy and brake judder vibration on steering wheel, in
‘SAE International Congress and Exposition’. SAE Paper No. 960734.
216 Bibliography

Kinney, C. E. and de Callafon, R. A. (2011), ‘The internal model principle for

periodic disturbances with rapidly time-varying frequencies’, International
Journal of Adaptive Control and Signal Processing 25(11), pp. 1006–1022.

Kirk, D. E. (2004), Optimal Control Theory: An Introduction, Dover Publica-

tions.

Klier, T. H. and Rubenstein, J. M. (2011), ‘Making cars smarter: The growing

role of electronics in automobiles’, Chicago Fed Letter 291a.

Köroǧlu, H. and Scherer, C. W. (2011a), ‘Robust generalized asymptotic regu-

lation against non-stationary sinusoidal disturbances with uncertain frequen-
cies’, International Journal of Robust and Nonlinear Control 21(8), pp. 883–
903.

Köroǧlu, H. and Scherer, C. W. (2011b), ‘Scheduled control for robust attenua-

tion of non-stationary sinusoidal disturbances with measurable frequencies’,
Automatica 47(3), pp. 504–514.

Kreyszig, E. (2006), Advanced engineering mathematics, 9th edn, John Wiley

and Sons.

Krishnamurthy, P., Lu, W., Khorrami, F. and Keyhani, A. (2005), A robust

force controller for an srm based electromechanical brake system, in ‘Pro-
ceedings of the IEEE Conference on Decision and Control’, pp. 2006–2011.

Krishnamurthy, P., Lu, W., Khorrami, F. and Keyhani, A. (2009), ‘Robust

force control of an SRM-based electromechanical brake and experimental
results’, IEEE Transactions on Control Systems Technology 17(6), pp. 1306–
1317.

Kubota, M., Suenaga, T. and Doi, K. (1998), A study of the mechanism causing
high-speed brake judder, in ‘SAE International Congress and Exposition’.
SAE Paper No. 980594.

Kwak, J. (2005), Modeling and control of an electromechanical brake (brake-

by-wire) system, PhD thesis, Purdue University.

Kwak, J., Yao, B. and Bajaj, A. (2004), Analytical model development

and model reduction for electromechanical brake system, in ‘2004 ASME
International Mechanical Engineering Congress and Exposition’. Paper
IMECE2004-61955.

Lang, H. (2008), ‘Electromechanical braking system’, United States Patent

US2008/0135357A1.

Lee, K. and Brooks, Jr, F. W. (2003), ‘Hot spotting and judder phenomena in
aluminium drum brakes’, Journal of Tribology 125, pp. 44–51.
Bibliography 217

Lee, K. and Dinwiddie, R. B. (1998), Conditions of frictional contact in disc

brakes and their effects on brake judder, in ‘SAE International Congress and
Exposition’. SAE Paper No. 980598.

Leonhard, W. (2001), Control of electrical drives, 3rd edn, Springer Verlag.

Leslie, A. C. (2004), Mathematical model of brake caliper to determine brake

torque variation associated with disc thickness variation (DTV) input, in
‘SAE Brake Colloquium and Exhibition’. SAE Paper No. 2004-01-2777.

Li, C., Zhang, D. and Zhuang, X. (2004), A survey of repetitive control, in

‘Proceedings of IEEE/RSJ International Conference on Intelligent Robots
and Systems’, vol. 2, pp. 1160–1166.

Line, C. (2007), Modelling and control of an automotive electromechanical

brake, PhD thesis, The University of Melbourne.

Line, C., Manzie, C. and Good, M. (2004), Control of an electromechanical

brake for automotive brake-by-wire systems with an adapted motion con-
trol architecture, in ‘SAE 2004 Automotive Dynamics, Stability & Controls
Conference and Exhibition’. SAE Paper No. 2004-01-2050.

Line, C., Manzie, C. and Good, M. (2006), Electromechanical brake control:

limitations of, and improvements to, a cascaded PI control architecture, in
‘Mechatronics 2006: The 10th Mechatronics Forum Biennial International
Conference’, Penn State Great Valley.

Line, C., Manzie, C. and Good, M. (2007), Robust control of an automotive

electromechanical brake, in ‘5th IFAC Symposium on Advances in Automo-
tive Control’, vol. 5.

Line, C., Manzie, C. and Good, M. (2008), ‘Electromechanical brake modelling

and control: from PI to MPC’, IEEE Transactions on Control Systems Tech-
nology 16(3), pp. 446–457.

Little, E., Kao, T.-K., Ferdani, P. and Hodges, T. (1998), A dynamometer

investigation of thermal judder, in ‘SAE Brake Colloquium and Engineering
Display’. SAE Paper No. 982252.

Löfberg, J. (2004), Yalmip : A toolbox for modeling and optimization in MAT-

LAB, in ‘Proceedings of the CACSD Conference’.

Lüdemann, J. (2002), Heterogeneous and hybrid control with application in

automotive systems, PhD thesis, Glasgow University.

Lyshevski, S. E. (1999), Electromechanical systems, electric machines, and

applied mechatronics, CRC Press.
218 Bibliography

Maciejowski, J. M. (2002), Predictive control: with constraints, Pearson Edu-

cation Limited.

Makkar, C., Dixon, W., Sawyer, W. and Hu, G. (2005), A new continuously dif-
ferentiable friction model for control systems design, in ‘2005 IEEE/ASME
International Conference on Advanced Intelligent Mechatronics’, pp. 600–
605.

Manayathara, T. J., Tsao, T.-C., Bentsman, J. and Ross, D. (1996), ‘Rejection

of unknown periodic load disturbances in continuous steel casting process
using learning repetitive control approach’, IEEE Transactions On Control
Systems Technology 4(3), pp. 259–265.

Marconi, L., Isidori, A. and Serrani, A. (2002), ‘Autonomous vertical landing

on an oscillating platform: an internal-model based approach’, Automatica
38(1), pp. 21–32.

Maron, C., Dieckmann, T., Hauck, S. and Prinzler, H. (1997), Electromechan-

ical brake system: Actuator control development system, in ‘SAE Interna-
tional Congress and Exposition’. SAE Paper No. 970814.

McGill, R. (1965), ‘Optimum control, inequality state constraints, and the gen-
eralized newton-raphson algorithm’, SIAM Journal of Control 3(2), pp. 291–
298.

Mehra, R. (1974), ‘Optimal input signals for parameter estimation in dynamic

systems–survey and new results’, IEEE Transactions on Automatic Control
19(6), pp. 753–768.

Meyer, R. (2005), Brake judder–analysis of the excitation and transmission

mechanism within the coupled system brake, chassis and steering system, in
‘SAE Brake Colloquium and Exhibition’. SAE Paper No. 2005-01-3916.

Newman, W. S. (1987), High-speed robot control in complex environments,

PhD thesis, Massachusetts Institute of Technology.

Newman, W. S. (1990), ‘Robust near time-optimal control’, IEEE Transactions

on Automatic Control 35(7), pp. 841–844.

Newman, W. S. and Souccar, K. (1991), ‘Robust, near time-optimal control

of nonlinear second-order systems: theory and experiments’, Journal of Dy-
namic Systems, Measurement, and Control 113(3), pp. 363–370.

Nissan North America (2005), ‘Warranty extension for “brake judder” condi-
tion involving model year 2005 and 2005 Titan and Armada vehicles’, Claims
Bulletin. Reference: WB/05-011.
Bibliography 219

Olsson, H., Åström, K. J., Canudas de Wit, C., Gäfvert, M. and Lischinsky,
P. (1998), ‘Friction models and friction compensation’, European Journal of
Control 4(3), pp. 176–195.
Pao, L. Y. (1994), Characteristics of the time-optimal control of flexible struc-
tures with damping, in ‘Proceedings of the IEEE Conference on Control
Applications’, pp. 1299–1304.
Pao, L. Y. and Singhose, W. E. (1998), ‘Robust minimum time control of
flexible structures’, Automatica 34(2), pp. 229–236.
Pontryagin, L. S., Boltyanskii, V. G., Gamkrelidze, R. V. and Mishchenko,
E. F. (1962), The Mathematical Theory of Optimal Processes, John Wiley
and Sons.
Raol, J. R., Girija, G. and Singh, J. (2004), Modelling and parameter estima-
tion of dynamic systems, Institution of Engineering and Technology.
Reuter, D. F., Lloyd, E. W., Zehnder, J. W. and Elliott, J. A. (2003), Hy-
draulic design considerations for EHB systems, in ‘SAE World Congress
and Exhibition’, Detroit, MI. SAE Paper No. 2003-01-0324.
Roberts, R., Gombert, B., Hartmann, H. and Schautt, M. (2003), Modelling
and validation of the mechatronic wedge brake, in ‘SAE Brake Colloquium
and Exhibition’. SAE Paper No. 2003-01-3331.
Roozen, N. B., Koevoets, A. H. and den Hamer, A. J. (2008), ‘Active vi-
bration control of gradient coils to reduce acoustic noise of MRI systems’,
IEEE/ASME Transactions on Mechatronics 13(3), pp. 325–334.
Rosen, J. B. (1966), ‘Iterative solution of nonlinear optimal control problems’,
SIAM Journal on Control 4(1), pp. 223–244.
Ryan, E. P. (1980), ‘On the sensitivity of a time-optimal switching function’,
IEEE Transactions on Automatic Control 25(2), pp. 275–277.
Saberi, A., Stoorvogel, A. and Sannuti, P. (2000), Control of linear systems
with regulation and input constraints, Springer.
Sacks, A., Bodson, M. and Khosla, P. (1996), ‘Experimental results of adaptive
periodic disturbance cancellation in a high performance magnetic disk drive’,
Journal of Dynamic Systems, Measurement, and Control 118(3), pp. 416–
424.
SAE-Australasia (2007), ‘Special commendation’, Autoengineer 29, pp. 19.
Salton, A. T., Chen, Z. and Fu, M. (2012), ‘Improved control design methods
for proximate time-optimal servomechanisms’, IEEE/ASME Transactions
on Mechatronics 17(6), pp. 1049–1058.
220 Bibliography

Schumann, F. (2001), ‘Electromechanically actuatable brake’, United States

Patent US6257377B1.

Schwarz, R., Isermann, R., Böhm, J., Nell, J. and Rieth, P. (1998), Model-
ing and control of an electromechanical disk brake, in ‘SAE International
Congress and Exposition’, Detroit, MI. SAE Paper No. 980600.

Schwarz, R., Isermann, R., Böhm, J., Nell, J. and Rieth, P. (1999), Clamp-
ing force estimation for a brake-by-wire actuator, in ‘SAE International
Congress and Exposition’. SAE Paper No. 1999-01-0482.

Semba, T. and White, M. T. (2008), ‘Seek control to suppress vibrations

of hard disk drives using adaptive filtering’, IEEE/ASME Transactions on
Mechatronics 13(5), pp. 502–509.

Shim, H., Kim, J.-S., Kim, H. and Back, J. (2010), ‘A note on the differential
regulator equation for non-minimum phase linear systems with time-varying
exosystems’, Automatica 46(3), pp. 605–609.

Shim, H., Lee, J., Kim, J.-S. and Back, J. (2006), Output regulation problem
and solution for LTV minimum phase systems with time-varying exosystem,
in ‘SICE-ICASE International Joint Conference’, pp. 1823–1827.

Siler, E. R. and Drennen, D. B. (2002), ‘Electric caliper having splined ball

screw’, United States Patent US6367593B1.

Smith, H. W. and Davison, E. J. (1972), ‘Design of industrial regulators. in-

tegral feedback and feedforward control’, Proceedings of the Institution of
Electrical Engineers 119(8), pp. 1210–1216.

Sri-Jayantha, M., Sharma, A., Dang, H. and Yamamoto, S. (1997), ‘Robust

servo for disk-shift compensation in rotating storage system’, United States
Patent 5,608,586.

Sri-Jayantha, S. M., Dang, H., Sharma, A., Yoneda, I., Kitazaki, N. and Ya-
mamoto, S. (2001), ‘Truetrack™ servo technology for high tpi disk drives’,
IEEE Transactions on Magnetics 37(2), pp. 871–876.

Stringham, W., Jank, P., Pfeifer, J. and Wang, A. (1993), Brake roughness-
disc brake torque variation, rotor distortion and vehicle response, in ‘SAE
International Congress and Exposition’. SAE Paper No. 930803.

Suryatama, D., Stewart, D. J., Meyland, S. C. and Hou, L. J. (2001), Contact

mechanics simulation for hot spots investigation, in ‘SAE World Congress
and Exhibition’. SAE Paper No. 2001-01-0035.

Sussmann, H. J. (1979), ‘A bang-bang theorem with bounds on the number of

switchings’, SIAM Journal on Control and Optimization 17(5), pp. 629–651.
Bibliography 221

Sussmann, H. J. and Willems, J. C. (1997), ‘300 years of optimal control:

From the brachystochrone to the maximum principle’, IEEE Control Sys-
tems 17(3), pp. 32–44.

Taig, A. G., Grabill, P. J. and Jackson, R. W. (1988), ‘Electrically operated

disc brake’, United States Patent 4793447.

Tamasho, T., Doi, K., Hamabe, T., Koshimizu, N. and Suzuki, S. (2000),
‘Technique for reducing brake drag torque in the non-braking mode’, JSAE
review 21(1), pp. 67–72.

Toh, K., Todd, M. and Tutuncu, R. (1999), ‘Sdpt3 — a matlab software

package for semidefinite programming’, Optimization Methods and Software
11, pp. 545–581.

von Groll, M., Mueller, S., Meister, T. and Tracht, R. (2006), ‘Disturbance
compensation with a torque controllable steering system’, Vehicle System
Dynamics 44(4), pp. 327–338.

Wang, N., Kaganov, A., Code, S. and Knudtzen, A. (2005), ‘Actuating mech-
anism and brake assembly’, International Patent No. WO 2005/124180 A1.

Wang, Y., Gao, F. and Doyle, F. J. (2009), ‘Survey on iterative learning con-
trol, repetitive control, and run-to-run control’, Journal of Process Control
19(10), pp. 1589–1600.

Watanabe, H., Suzuki, T. and Yasukawa, D. (2007), ‘Electro mechanical

brake’, United States Patent US2008/0091326A1.

Whitcomb, R. (2004), Avro Aircraft and Cold War Avaition, Vanwell Publish-
ing.

Winters, F. J., Mielenz, C. and Hellestrand, G. (2004), Design process changes

enabling rapid development, in ‘Convergence International Congress and
Exposition On Transportation Electronics’. SAE Paper No. 2004-21-0085.

Wonham, W. M. (1979), Linear multivariable control: A geometric approach,

Springer-Verlag, New York.

Workman, M. L., Kosut, R. L. and Franklin, G. F. (1987), Adaptive proximate

time-optimal servomechanisms: Continuous time case, in ‘Proceedings of the
American Control Conference’, vol. 1, pp. 589–594.

Wu, D., Guo, G. and Chong, T. C. (2002), ‘Midfrequency disturbance suppres-

sion via micro-actuator in dual-stage hdds’, IEEE Transactions on Magnet-
ics 38(5), pp. 2189–2191.
222 Bibliography

You, K. H. and Kwak, S. W. (2006), ‘Robust mixed time-optimal control

for third-order systems with model uncertainty’, International journal of
systems science 37(3), pp. 173–180.

You, K. H. and Lee, E. B. (2000), Robust, near time-optimal control of non-

linear second order systems with model uncertainty, in ‘Proceedings of the
IEEE International Conference on Control Applications’, pp. 232–236.

Zhang, D. Q. and Guo, G. X. (2000), ‘Discrete-time sliding mode proximate

time optimal seek control of hard disk drives’, IEE Proceedings Control The-
ory and Applications 147(4), pp. 440–446.

Zhang, L., Ning, G. and Yu, Z. (2007), Brake judder induced steering wheel
vibration: Experiment, simulation and analysis, in ‘SAE Brake Colloquium
and Exhibition’. SAE Paper No. 2007-01-3966.

Zhang, Y. (2000), Stability and performance tradeoff with discrete time tri-
angular search minimum seeking, in ‘Proceedings of the American Control
Conference’, vol. 1, pp. 423–427.

Zhang, Z. and Serrani, A. (2006), ‘The linear periodic output regulation prob-
lem’, Systems & control letters 55(7), pp. 518–529.

Zhao, S. and Tan, K. K. (2005), ‘Adaptive feedforward compensation of force

ripples in linear motors’, Control Engineering Practice 13, pp. 1081–1092.

Zheng, J., Guo, G., Wang, Y. and Wong, W. E. (2006), ‘Optimal narrow-band
disturbance filter for pzt-actuated head positioning control on a spinstand’,
IEEE Transactions on Magnetics 42(11), pp. 3745–3751.

Zhou, J., Zhou, R., Wang, Y. and Guo, G. (2001), ‘Improved proximate time-
optimal sliding-mode control of hard disk drives’, IEE Proceedings Control
Theory and Applications 148(6), pp. 516–522.

Zhou, K. and Doyle, J. C. (1998), Essentials of robust control, Prentice Hall.

Minerva Access is the Institutional Repository of The University of Melbourne

Author/s:
Lee, Chih Feng

Title:
Brake force control and judder compensation of an automotive electromechanical brake

Date:
2013

Citation:
Lee, C. F. (2013). Brake force control and judder compensation of an automotive
electromechanical brake. PhD thesis, Department of Mechanical Engineering, The
University of Melbourne.

Publication Status:
Unpublished

Persistent Link:
http://hdl.handle.net/11343/38499

Terms and Conditions:

Terms and Conditions: Copyright in works deposited in Minerva Access is retained by the
copyright owner. The work may not be altered without permission from the copyright owner.
Readers may only download, print and save electronic copies of whole works for their own
personal non-commercial use. Any use that exceeds these limits requires permission from
the copyright owner. Attribution is essential when quoting or paraphrasing from these works.