Tire Modeling and Friction Estimation: Svendenius, Jacob
Tire Modeling and Friction Estimation: Svendenius, Jacob
Tire Modeling and Friction Estimation: Svendenius, Jacob
Svendenius, Jacob
2007
Document Version:
Publisher's PDF, also known as Version of record
Link to publication
General rights
Unless other specific re-use rights are stated the following general rights apply:
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors
and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the
legal requirements associated with these rights.
• Users may download and print one copy of any publication from the public portal for the purpose of private study
or research.
• You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal
LUNDUNI
VERSI
TY
PO Box117
22100L und
+4646-2220000
Tire Modeling and Friction
Estimation
Jacob Svendenius
5
Abstract
Acknowledgments
First of all I would like to express my gratitude to the persons that made
the decisions, a huge amount of years ago, necessary to establish my sit-
uation as industrial PhD student, the former technical manager, Kent
Jörgensen and the former personal manager, Lena-Maria Lundberg at
Haldex Brake Products AB. I will also thank my supervisor, Björn Wit-
tenmark, at the Department of Automatic Control for the administrative
work he made to establish this collaboration. His advice, encouraging sup-
port and guiding corrections has all the time lead the projects and the
thesis writing in a rewarding direction. Thanks to my assistant supervi-
sor Per Hagander for proof reading and valuable comments on the thesis.
I would also like thank my supervisor at Haldex, Per-Axel Roth, for good
listening and finding nice parallels from his earlier experiences.
I’m very grateful to my group manager at Haldex, Anders Lindqvist,
for giving great support in all kind of matters and being the only one
at Haldex to always participate on the reference group meetings. It is a
pleasure to be a part of a great working group, having Anders Nilsson,
Ola Nockhammar, Lisette Pedersen, Sabine Marksell and Ola Ahlström
as colleagues. Many thanks to Peter Nilsson for his wise opinions and
skilled advices. I am particularly thankful Sverker Stomrud for his great
commitments in all technical and patent related issues and strong belief
in new ideas.
The work in the Road Friction Estimation project has been very inter-
esting and rewarding and I will gratefully acknowledge Johan Hultén at
Volvo Cars for good project leadership and Fredrik Bruzelius for fruitful
and inspiring cooperation and nice fellowship and proof reading of the
thesis. Both Johan and Fredrik are deeply involved the chapter concern-
ing the road friction estimator. Special thanks to Magnus Gäfvert, now at
Modelon AB, but former at both the Department of Automatic Control and
at Haldex, for his enormous working spirit. I am glad for our rewarding
discussions and nice collaboration. Magnus has contributed in both the
6 7
Acknowledgments
8 9
Contents
5.6
Implementation and Application in Multi-Body Simula-
tion of Vehicle Handling . . . . . . . . . . . . . . . . . .
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . .
95
97
1
6. Measurement and Processing of the Wheel Speed Signal 100
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.2 Effect of Tire Force Transients on the Wheel Speed Mea-
surements . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Introduction
6.3 Method for Elimination of Rotation Dependent Noise . 111
6.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7. Road Friction Estimation . . . . . . . . . . . . . . . . . . . . 120
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 120
7.2 Experimental Validation of the Brush Model . . . . . . 124
7.3 Tire Parameter Estimation . . . . . . . . . . . . . . . . . 143
7.4 Algorithm Implementation . . . . . . . . . . . . . . . . . 155 1.1 Background and Motivation
7.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
7.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 166 This thesis is a result of cooperation between the R&D department at
Haldex Brake Products in Landskrona and the Department of Automatic
8. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
Control at Lund University. The work is closely related to the develop-
8.1 Ideas of Further Research . . . . . . . . . . . . . . . . . 172
ment of brake-by-wire systems using electrically actuated brakes for heavy
9. Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 trucks and buses performed at Haldex. One objective of the collaborative
A. Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183 work has been to further investigate the new field for improvements of
B. Scale Factors and Slip Conversion . . . . . . . . . . . . . . 189 vehicle stabilizing and safety functions, opened by the enhanced control-
lability of an electrically actuated brake compared to the conventional
C. Extension of Optimization Algorithm to Include Hori-
pneumatic brake. For a well-working brake system, the physical proper-
zontal Shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
ties of the brakes are not limiting, their capability to produce braking
torque is very rarely used fully. Instead the conditions of the tire and the
road surface set the functional limits for safe retardation of the vehicle.
When the Anti-lock Braking System (ABS) was introduced in the luxury
vehicles it greatly improved the surface friction utilization of the brake
system in critical situations. Nowadays, it is standard equipment in most
cars. The motivation for developing ABS relies on the observation that
a locked wheel only produces a force in the sliding direction of the tire.
Furthermore, a higher tire force can be obtained from a rolling wheel.
Preventing the wheel from locking while still generating a high braking
force is shown to reduce the braking distance. But, maybe more impor-
tant is that it enables the possibility to the driver to control the vehicle
direction by steering maneuvers.
There is an optimal wheel rotation velocity, depending on the tire prop-
erties, surface condition, and vehicle motion, for which the tire develops its
maximal friction force. Finding and keeping this value under the current
condition in critical maneuvers, while providing stable and safe motion
10 11
Chapter 1. Introduction 1.1 Background and Motivation
of the vehicle, is a great issue for improvement of today’s systems. The and the close collaborative work with both Volvo Cars and VTI (Swedish
new quicker brake actuators increase the possibilities for stable control National Road and transport Research Institute) a multitude of different
of the wheel velocity in such situations. When increasing the bandwidth test scenarios have been performed. There have been many possibilities
of the braking control loop the dynamic effects of the mechanical systems to participate in the planning of the tests and most of them have been
supporting the brake and participating in the transmission of the force directly specified to serve the method development described in the the-
to the road become crucial. Deflections in suspension, bushings, and the sis. There is a great benefit in participating during the testing, since it
tire delay the generated tire force acting on the vehicle and may cause gives a better feeling for how the surrounding conditions and other un-
displacement of the wheel speed sensor. This disturbs the wheel speed expected factors might influence the result. All tests have be performed
measurements. The noise filters have to be faster to catch rapid signal on real tires in real conditions. This has made it impossible to completely
variations, resulting in less attenuation of the disturbances. The inertia separate the measurement of, for example, the tire behavior from distur-
in the wheel rim might cause unwanted oscillations. Hence, utilization bances from the road and the wheel suspension, etc. On the other hand
of the advantages in control of faster brakes requires the algorithms to it gives the most realistic picture of the problems and difficulties that
further predict and estimate the dynamics of the surrounding system. have to be dealt with. Although good equipment has been available, some
Enhanced sensor quality and larger computational capacity increases the of the issues addressed in the thesis, relating to very quick events, have
possibilities to achieve such estimations. not been validated completely. The reason is that the sensors and mea-
The new and improved technology emphasizes attempts to also detect surement systems that have been used have not been capable of such fast
the conditions surrounding the entire vehicle using existing sensors. The acquisition. Although there are weeks of testing behind the results in the
idea is attractive since the functionality of many other vehicle systems thesis, further testing is still one of the most urgent issues for the future
can be improved by such knowledge. Automatic cruise control (ACC), col- work.
lision avoidance warning and, electronic stability program (ESP) are a Three different, but closely related topics, are treated in the thesis.
few examples. Another important feature is to help the driver to adapt
the driving style towards the current conditions. Within the EU-project
Intelligent Roads [Vägverket, 2006a] research is performed about how the Semi-empirical tire-modeling
road surface limitations should be presented to the driver. It might be as Simultaneous cornering affects the behavior of the tire during a braking
estimated braking distance, or as a recommended maximal speed. For an phase. For some vehicle stability systems it is more important how the
on board road-condition estimator it seems quite impossible to determine vehicle cornering performance is affected when applying a brake torque
the circumstances in front of the vehicle. It could however, be possible while turning. If a vehicle exhibits a large lateral acceleration such that
by using a preview sensor or having an infrastructure that receives and a roll-over may be inevitable a concurrent controlled braking can reduce
distributes frictional information about the roads. The distribution can be the lateral force stabilizing the vehicle, see for example [Schofield et al.,
both to a common central or to share between the vehicles in a close neigh- 2006]. Further, a system may not demand a higher brake force in a corner
borhood. Research regarding infrastructures that receives and distributes than that the vehicle can keep its course not running off the road.
estimates of the surface friction along the road is performed within the The thesis includes a description of a newly developed semi-empirical
national program Safety Road Information Structure [Vägverket, 2006c]. tire-model that describes this interaction. A main focus in the model
A major aim of this work is to explain and model the general behavior derivation is the possibility to extract as much information as possible
of tires and to find properties that are common for the majority of tire from the pure slip data. This may simplify and reduce the number of cal-
types. Simplicity in modeling is a keyword throughout the thesis. Often, ibration procedures needed to map the specific tire behavior. The longitu-
the final solutions still tend to be sufficiently comprehensive. Starting dinal and lateral forces are described by scalings of the pure-slip forces.
with complex models might limit the usability in the end. Therefore, the The used scale factors are calculated from physical relations between com-
basic form of the brush-model theory to describe the tire behavior has bined and pure slip.
been fundamental in this work. Transient maneuvers in one direction have a dynamic effect on the tire
The developed theories are based on physical assumptions and most behavior in both directions. The thesis derives the cross-coupling terms
of them are validated from available data or by tests performed partic- and studies the transient properties of a tire for low and high slip situa-
ularly for this purpose. Due to good experimental equipment at Haldex tions.
12 13
Chapter 1. Introduction 1.2 Main Results
Improvement of wheel speed measurement force and slip axis. This will maximize the persistence of excitation of the
The wheel-speed signal is one of the most essential signals for the brake data feeding the estimator and it is a compact form to store and use old
system control. Errors and disturbances on the measurements often re- data, important for characterizing the tire behavior. The tire friction and
quire hard filtering that slows down the brake performance. The wheel tire stiffness parameters are derived by minimizing the error between the
speed is measured as the relative rotational speed between the rim and adopted tire model and the stored data. The model has a practical focus
the axle. Transient force applications may disturb the sensor position and intended for implementation in production cars.
cause disturbances on the measurements. The effect became particularly
evident during a test sequence where the ABS-function of the vehicle was
triggered due to a response on the wheel speed from the quick brake 1.2 Main Results
torque application. The wheel was not even close to locking at that time.
Also tolerances and unevenesses in sensor arrangement may cause extra The thesis mainly contributes in a few related areas. The first is the
jitter on the wheel speed signal. The thesis discusses methods to reduce derivation of a new physically based semi-empirical tire model that pre-
such disturbances. dicts the tire behavior for simultaneous braking, cornering, and cambering
given the tire characteristics for pure braking, cornering, and cambering.
The method incorporates the following features:
Road friction estimation
Friction is a prerequisite for the modern life of human beings. Technical • The tire forces at combined slip are derived using any longitudinal,
solutions for transport and manufacturing would be completely different, lateral, and camber pure-slip model or data set.
if ever possible, if materials did not bond to achieve frictional forces. In • The scale factors for the combined tire forces are computed in a
many normal duties the anticipation of the friction has become automatic unique way from the relation between the combined tire force and
and is nothing that we really reflect over. Everything from small incidents the pure tire force. Any theoretical model describing the tire behavior
to fatal accidents may occur if the friction differs from the expectations at combined slip can therefore be used to form the scale factors.
or changes suddenly. A pedestrian that slips on a slippery spot or a roller
skater whose wheel bearing abruptly cuts are two examples. • The model is, in its presented form, based on the most basic brush-
A motivation for trying to estimate the friction is that the number of model theory. The structure makes it straight forward to incorporate
accidents related to slippery road conditions might be reduced if the driver effects from more variables and more sophisticated modeling.
is correctly informed about the limitations of the vehicle in the current
• Velocity dependence is included in the model, even though this is
condition. Safety can also be increased, since the performance of many ap- not explicitly present in the pure-slip models.
plications in the vehicle can be enhanced by knowing the current frictional
condition. The brake control can be made more efficient by knowing the • The model has the possibility to convert between driving and braking
available friction, for example the gain-scheduled slip controller proposed data.
in [Solyom et al., 2004] bases it switching on whether the actual slip is
• All included parameters can automatically be computed from the
lower or higher than a the slip corresponding to the friction peak. This
shape of the pure-slip models.
slip value is an important parameter for obtaining the optimal tire force
and can be derived from an assumption on the estimated actual friction The wheel speed is an important signal for the applications covered by the
and the tire characteristics. thesis. A few ways to improve the quality of the wheel speed measurements
The last part of the thesis is results from the work within a cooperative are discussed. Particularly, the contribution here is:
project between Haldex, the Department of Automatic Control, and Volvo
• The construction of a feedforward filter on the brake torque that pre-
Cars within the Intelligent Vehicle Safety Systems (IVSS) programme. This
dicts the the disturbances on the wheel speed due to the application
part deals with on board model-based friction estimation based on the
of the brakes.
local measurements on the vehicle.
The proposed algorithm includes a memory-efficient way to store fil- A method for friction estimation is another contribution of this thesis. Its
tered values of the sampled data into storage bins, distributed over the major characteristics are:
14 15
Chapter 1. Introduction 1.4 Related Publications
• A filter that stores representatives of the sampled data in a compact Gäfvert, M. and J. Svendenius (2005): “A novel semi-empirical tire model
form to maximize the possibilities to extract information from the for combined slip.” Vehicle System Dynamics, 43:5.
force-slip relation.
Gäfvert, M., J. Svendenius, and J. Andreasson (2006): “Implementation
• Estimations of the braking stiffness of the tire and the road-tire and application of a semi-empirical tire-model in multi-body simula-
friction using non-linear optimization techniques. tion of vehicle handling.” In Proceedings of the 8th International Sym-
posium on Advanced Vehicle Control. Taipei, Taiwan.
• The model-based estimation relies on brush-model mechanics
Svendenius, J. (2003): “Tire models for use in braking applications.”
• Promising results are shown in validation tests at different driving
Licentiate thesis ISRN LUTFD2/TFRT--3232--SE. Department of
conditions.
Automatic Control, Lund University, Sweden.
Svendenius, J. (2006): “Examples of scale factors for a semi-empirical
1.3 Outline tire-model.” Technical Report ISRN LUTFD2/TFRT--7614--SE. De-
partment of Automatic Control, Lund University, Sweden.
The topics covered by the thesis are related to the wheel-tire-road in-
Svendenius, J. and M. Gäfvert (2004a): “A brush-model based semi-
teraction. Chapter2 explains the fundamental properties and definitions
empirical tire-model for combined slips.” SAE Technical Paper 2004-
related to the characteristics of tires. In Chapter 3 there is a brief review
01-1064.
of existing tire models. A basic concept for tire modeling is the brush model
developed in the early age of the tire-modeling era. The theory behind the Svendenius, J. and M. Gäfvert (2004b): “A semi-empirical tire-model for
brush model and some extensions are described in Chapter 4. This model transient combined-slip forces.” In AVEC ’04.
is fundamental, both in the derivation of the semi-empirical tire-model
described in Chapter 5 and for the friction estimator developed in Chap- Svendenius, J. and M. Gäfvert (2005): “A semi-empirical tire model for
ter 7. Chapter 6 contains material describing a few ways to improve the combined slips including the effects of cambering.” In Lugner and
quality of the wheel speed signal, which may enhance the performance of M.Plöchl, Eds., Tyre Models For Vehicle Dynamics Analysis, vol. 43,
pp. 317–328. Taylor & Francis Group, Glasgow. Supplement to Vehicle
vehicle systems relying on this signal, e.g. the proposed friction estimator.
Proposals for further work and some concluding remarks are mentioned System Dynamics Volume 43.
in the final chapter of the thesis. Svendenius, J. and M. Gäfvert (2006): “A semi-empirical dynamic tire
model for combined-slip forces.” Vehicle System Dynamics, 44:2,
pp. 189 – 208. Special Issue: AVEC ’04: 7th International Symposium
1.4 Related Publications on Advanced VEhicle Control, 23-27 August 2004 HAN University,
Arnhem, The Netherlands.
The thesis is based on research documented in several publications where
The conference paper [Svendenius and Gäfvert, 2004b] was chosen to-
the author of the thesis is a coauthor or the sole author.
gether with another six submissions from all contributions published at
The following publications and reports contain theory, implementation,
the AVEC-04 conference to be included in the special issue of the Journal
and validating results of the semi-empirical tire model:
of Vehicle System Dynamics, reflecting the highlights of the conference,
Gäfvert, M. and J. Svendenius (2003): “Construction of semi-empirical tire see [Svendenius and Gäfvert, 2006].
models for combined slip.” Technical Report ISRN LUTFD2/TFRT-- The following publications and reports are related to friction estima-
7606--SE. Department of Automatic Control, Lund University, Sweden. tion project:
Gäfvert, M. and J. Svendenius (2004): “A semi-empirical tire- Svendenius, J. (2007a): “Validation of the brush model towards VTI-
model including the effects of camber.” Technical Report ISRN measurement data recorded at Hällered 2005.” Technical Report ISRN
LUTFD2/TFRT—7611–SE. Department of Automatic Control, Lund LUTFD2/TFRT—7616–SE. Department of Automatic Control, Lund
University, Sweden. University, Sweden. RFE-project.
16 17
Chapter 1. Introduction
18 19
Chapter 2. Tire Fundamentals 2.4 Kinematics
2.2 Design and Construction the coupling between the tire model and the multi-body vehicle system.
However, the multitude of different standards in this area can be confus-
The pneumatic tire is a flexible structure that together with the rim holds ing. In 1974 SAE published a vehicle dynamics standard specifying the
the pressure of the inflated air. The most important functions of the tire choice of coordinate system and notations [SAE Recommended Practice
are to reduce vibrations from unevennesses in the road and to achieve J670e, 1976]. Fifteen years later the European version ISO 8855 was re-
a high friction coefficient in the interaction with the road surface. Its leased [ISO 8855, 1991]. One of the greatest pioneers in tire modelling, H.
main structural component is the carcass, which consists of layers of stiff B. Pacejka used an adapted version of the SAE standard [Pacejka, 1988]
cords to hold the shape of the tire and the tension from the compressed in his works. In 2002, TNO together with TU Delft promoted a version
air. High tensile steel wires, called beads, keep the carcass to the rim of the ISO 8855 in a standard specialized for acquisition of measurement
and when a load is applied on the wheel, the rim primarily hangs on data for calibration of tire model parameters called TYDEX (Tyre Data
the sidewall cords and the beads. The entire structure is covered with a Exchange Format) [TYDEX-Working group, 1997]. In close connection to
wear resistant rubber compound, often styrene-butadiene, to protect the TYDEX is the Standard Tire Interface (STI) that prescribes the interface
carcass and to build up the friction to the road. There are mainly two ways for the interaction between the tire model and the Multi Body Simulation-
to design the carcass, radial-ply and bias-ply. Radial-ply implies that the system. All of these four standards deviate regarding sign conventions of
sidewall cords are oriented radially and the wear surface cords are laid in slips and forces, which makes it important to state which standard that
layers tangentially with small angles between the cords in each layer. The is followed or, alternatively, clearly define the signification of the used
bias-ply tires have the cord layers diagonally positioned over the entire tire notations.
surface. For a rolling bias-ply tire the deformation of the cords gives rise to The choice of reference system in this work follows the SAE stan-
a wiping motion of the rubber tread, which causes higher wear and power dard [SAE Recommended Practice J670e, 1976], with the longitudinal x-
dissipation. Therefore, radial tires are nowadays mostly used for cars and axis aligned with the wheel heading, the lateral y-axis perpendicular to
trucks, even though their manufacturing process is more complex and the the wheel, and the vertical z-axis pointing downwards. The ISO-standard,
expense is about 50% higher than for a bias-ply tire [Thomson, 2003]. instead prescribes the direction of the z-axis to be upwards. Hence, the
direction of both the y- and z-axis differ between the two standards, since
their reference systems are orthogonal. In Sections 2.4–2.5 the used no-
2.3 Standards tations and sign conventions are explained, as from [Gäfvert, 2003]. The
table in Appendix A lists the nomenclature and its deviation from ISO
Many issues related to the tire are controlled by regulations and stan- 8855:1991.
dards. One of those that might be interesting for the consumer is the
standard for the tire marking. In Europe the ECE (Economic Commis-
sion for Europe) Regulation and EEC (European Economic Community)
Directives set standards for the tire marking, which differ from the Amer-
2.4 Kinematics
ican way of marking the tires. However, the size of the tire is always
printed on the sidewall as a combination of four designations, for instance
This section describes the relevant tire kinematics and introduces defi-
P 205/75R15. The first letter (sometimes left out) signifies the vehicle for
nitions which are used in the thesis. The entities are illustrated in Fig-
which the tire is intended. The following number denotes the tire width
ure 2.1. Vectors have two components and are denoted by a bar as in
in millimeter. A slash separate the width from the height of the tire side,
v̄. The corresponding components and magnitude are denoted by vx , vy,
which is expressed as the fraction in percent of the width. After the letter q
R, the rim radius is given in inches. The size marking is often followed and v = v2x + v2y . The x-axis is often refereed to as the longitudinal di-
by markings denoting load and speed classes. rection and the y-axis as the lateral direction. The wheel-travel velocity
There are also standards for the nomenclature regarding the kinemat- v̄ = ( vx , vy ) deviates from the wheel orientation by the slip angle α
ics and the mechanics of the tire. The movements and forces developed by
the tire are important signals for the control or simulation of the dynam- vy
ics of the vehicle. Many errors can therefore be avoided by standardizing tan(α ) = (2.1)
vx
20 21
Chapter 2. Tire Fundamentals 2.5 Tire Mechanics
v̄ This convention is followed also in this work. Implicitly, this assumes that
vx α
the forces do not depend on the magnitude of the slip velocity, vs . In
general, at least the sliding friction is velocity dependent. The ISO and
SAE standards [ISO 8855, 1991; SAE Recommended Practice J670e, 1976]
v̄s use −100κ x [%] to represent longitudinal slip, and α [deg] for lateral slip.
vsx Here, the slips are defined such that signs are consistent for the different
vy
slip definitions, and such that a generated tire force has opposite sign to
vx
− Fy vsy
the slip. This means that braking or left cornering will result in positive
slip and negative force. For convenience the slip ratio, λ , will be used
Mz ω
R to denote longitudinal slip as: λ = κ x . It is straightforward to translate
F̄ − Fx between the different slip representations
( λ , tan(α ) ) κ̄ s̄
x
− Fx x σ̄ = = =q (2.6a)
(1 − λ ) 1 − κx 1 − s2y − s x
y z σ̄ s̄
κ̄ = ( λ , tan(α ) ) = =q (2.6b)
1 +σx 1 − s2y
Figure 2.1 Kinematics of an isotropic tire during braking and cornering. Force
vectors are also included. (Left: top view; Right: side view) σ̄ κ̄
s̄ = ( λ cos(α ), sin(α ) ) = q = q (2.6c)
(1 + σ x )2 + σ 2y 1 + κ y2
The circumferential velocity of the wheel is
22 23
Chapter 2. Tire Fundamentals 2.6 Tire Deformation
Aligning Torque Mz up to a certain slip value (σ ∗x ) and then diminishes slightly when σ x goes
to infinity or λ reaches unity, see the measurements in Figure 5.1.
Camber γ
Combined slip Combined slip signifies simultaneous longitudinal and
lateral slip, i.e. braking and cornering actions performed at the same time.
A steering maneuver during braking, generally, decreases the braking
stiffness, the longitudinal peak force, and its corresponding slip value. A
Wheel Torque model for the interaction between the slip in both directions is therefore
Direction of Wheel Heading
Rolling Resistance Torque My
Tractive Force Fx
inevitable for more advanced vehicle simulations. In Section 3.2 and in
x Chapter 5 this topic will be discussed in detail.
Slip Angle α
the slip of the tire. The velocity difference between the carcass and the Rolling resistance
road is a result from continuous deformation of the rubber treads and
The deformation of the tire due to the vertical load is often clearly vis-
sliding between the tire and the road surfaces. At low slip the relation
ible. The wheel rotation for a moving vehicle continuously changes this
is approximately linear and the forces can be described as Fx = − Cxσ x ,
deformation and power dissipates from the system due to the visco-elastic
longitudinally and Fy = − Cyσ y, laterally. The braking stiffness, Cx and the
properties of the carcass, see Section 2.6. The effect is called rolling resis-
cornering stiffness, Cy are correspondingly defined as the linearization of
tance and is, in general, assumed to depend linearly on the tire load [Wong,
the force-slip relation at σ x = 0 and σ y = 0 [Wong, 2001]
2001] and radius, as
My = q0 Rp Fz p (2.8)
dFx
Cx = −
dσ x σ x =0,σ y=0
(2.7) For more accurate results the dependency on the velocity has to be in-
dFy cluded in the tire specific factor, q0 . In [Pacejka, 2002] an addition to the
Cy = −
dσ y σ x =0,σ y=0 resistance torque due to the extra energy loss when a longitudinal force
is developed, is proposed, as My,add = ( R e − R) Fx . The energy loss comes
At higher slip the relation is strongly nonlinear and a more complex func- from the difference between the effective rolling radius and the tire ra-
tion is necessary to express the relation, see Sections 3.1 and 4.1. The dius that works as a lever for the torque on the rim to the transmitted
normal behavior for a tire on asphalt is that the force increases with slip tire force to the ground.
24 25
Chapter 2. Tire Fundamentals 2.6 Tire Deformation
y x
y
z
z
Figure 2.3 Cambered wheel. Left: Rear view; Right: Top view with contact patch
(dashed rectangle). Note that the contact patch is greatly exaggerated in size. Figure 2.4 Illustration of the schematic tire structure used in this thesis. The
carcass is visualized as a stiff string attached to the rim by visco-elastic springs.
The springs are only showed in the radial direction but work also tangentially and
2.6 Tire Deformation laterally (out of the wheel plane). The same holds for the elastic tread springs.
One of the main tasks for the tire is to reduce the vibrations caused by
stretched string method gives, however, a good visualization of the physics
the unevennesses in the road. The elastic and damping properties are
behind the tire behavior and in many cases the resulting equations can
therefore of great importance and large vertical shape deformations are
be solved analytically. Restrictions can be introduced to obtain results in
required for efficient damping. The price is slower handling and reac-
fairly simple and sufficiently accurate expressions that easily can be used
tion on maneuvers from the driver, since a softer tire, generally, allows
for estimation and control issues.
larger horizontal deformation. The deformation can schematically be di-
vided between the carcass and the tread. The carcass flexibility is the
major source for the dynamic behavior and the rolling resistance of the
tire. In the longitudinal direction the minor deflection of the rubber treads
determines the force-slip characteristics, while both the tread and carcass
deflections affects it in the lateral direction. A good way to illustrate the
carcass behavior is to use the stretched string carcass model that origi-
nates from the work of von Schlippe in 1941, described in [Pacejka, 2002],
which is still used for tire modeling, see for instance [Thorvald, 1998]. The
carcass is then, as the name intends, approximated as a stretched string
attached to the rim by visco-elastic springs, see Figure 2.4. Finite element
methods, allowing more sophisticated models, are usually used to derive
more accurate results for the tire deformation, since the shape is complex
and the deflections are large [Pauwelussen et al., 1999; Gipser, 2005]. The
26 27
3.1 Steady-State Tire Models
extensively presented in Section 4.1 and therefore left out here, but the
3 different assumptions on the brush model often results in a polynomial
including at least two parameters. One example is the proposal for the
lateral force by Smiley and Horne in 1958
2
3µ Fz
C α 1 − α
α ≤ α○ =
α
Review of Existing F0y =
µ Fz
3(α ○ )2
otherwise
2Cα (3.1)
Tire-Models where Cα is the cornering stiffness and µ the friction coefficient. The
definition of the slip angle α is given by (2.1). The friction is assumed to
be constant in the model, which disables its ability to also characterize the
longitudinal force accurately for longitudinal slips, since there, generally,
is a pronounced peak of the force in this direction. Interpolation between
the data points was then mostly used to model the longitudinal tire force.
An extensive amount of work has been done in the field of tire modeling. By use of velocity dependent friction, a relation for combined slip that
The research covers everything from derivation of simplistic expressions better expressed the behavior of the brake force for large slip was derived
describing the tire behavior to advanced finite-element methods that pre- in [Dugoff et al., 1969], given on the form
dicts the important states at any point of the tire. The tire modeling was
Cx λ
initiated by the vehicle and air craft industry back in the 1940’s. The first Fx = f (θ ) (3.2)
descriptions of the tire characteristics were derived from physical model- 1−λ
ing. Later on the interest for finding empirical equation structures, that Cy tan(α )
Fy = f (θ ) (3.3)
easily could be adjusted to fit the measurements from the tire tests in- 1−λ
creased due to the complex nature of the tire. The early tire models only
(
θ (2 − θ ) θ ≤1
covered the static force-slip relation, but the development of fast control f (θ ) = (3.4)
1 θ >1
systems has now increased the focus on the dynamical aspects of the tire
behavior. The larger amount of computer power has enabled the possibility where q
to solve the physical differential equations related to the tire deformation. µ 0 Fz((1 − ǫv λ 2 + tan2 (α ))(1 − λ ))
Therefore, physical modeling of both dynamical and frictional aspects has θ= q (3.5)
become popular again. This chapter gives a short review of the tire models 2Cx2 λ 2 + Cy2 tan2 (α )
that are relevant for the continuation of the thesis. and µ 0 is the nominal friction coefficient and ǫ is the velocity dependency
factor. This model is one of the three models usually referred to as the
HSRI-models developed at the Highway Safety Research Institute. The
3.1 Steady-State Tire Models other two models are described in [Fancher et al., 1972] and [Tielking and
Mital, 1974] and are based on different assumptions on the brush model
The aim of a tire model is often to obtain a structure that can fit mea- regarding the vertical pressure distribution and sliding properties of the
surement data well by optimal choice of included parameters. A detailed rubber.
description of the early tire models can be found in [Nguyen and Case, An empirical way to describe the lateral force and account for com-
1975], from where also the following examples in this section are gath- bined braking and cornering by using the friction circle criterion, has
ered, if not otherwise is stated. been proposed by Chiesa (1965) as
The first tire models did only concern the steady-state relation between s n
the slip and the developed force. They were physically derived from vari-
Fx
Fy = 1 − Fz (a1 + a2 Fz)α + (a3 + a4 Fz )α 2 + . . . (3.6)
ants of the brush model. The most common form of the brush model is 2µ N
28 29
Chapter 3. Review of Existing Tire-Models 3.1 Steady-State Tire Models
The method requires that the longitudinal force is known instead of the where ȳ(⋅) is the region-separated function-approximation. The adjusted
longitudinal slip, compared to the previous methods. To further increase combined slips σ x , σ y are shaped from combinations of the pure slips
the flexibility in the modeling Holmes (1969) proposed an empirical struc- λ , α and γ and do not correspond to the physical slips from (2.5), see
ture, quite different to the others, to use for curve fitting the original paper for a detailed description. The frictional values at the
peak, µ x,max and µ y,max and corresponding slips s xm ( Fz ) and s ym ( Fz) are
F0y = a0 + a1 vx + a2 v2x + a3α + a4α 2 + a5α 3 + a6 R + a7 P (3.7) then used to set the difference in the corresponding direction. The factors
fst1 and fst2 are used to adjust between different surfaces. Similar ideas
where P is a tire-pattern constant and R is a tire-tread constant. The are proposed for deriving the self-aligning torque at combined slip. The
coefficients a1...7 have no physical interpretation. The idea to treat the method is claimed to have a short on-line calculation time, and to rep-
dependence of the vehicle velocity as an additional contribution was not resent the tire behavior in areas without measurements in a reasonable
really accepted by the other researchers in the field. Further references way.
regarding these early tire models can also be found in [Bernard et al., The paper [Bakker et al., 1987] presents the “Magic Formula”, which
1975]. quickly became the most predominating model. The model expresses the
In [Kiencke and Nielsen, 2000] another empirical model, developed by lateral and longitudinal tire forces, as well as aligning torque, on the form
Burckhardt is used, which models the real tire behavior more accurately. h i
The input sRes denotes the resultant slip (s2L + s2S )1/2 and an additional slip y( x) = D sin C arctan (1 − E) x + ( E/ B ) arctan( B x) (3.11)
30 31
Chapter 3. Review of Existing Tire-Models 3.2 Combined-Slip Semi-Empirical Tire-Models
3.2 Combined-Slip Semi-Empirical Tire-Models Some early efforts to model tire forces under combined-slip conditions
are described and compared in [Nguyen and Case, 1975]. One of the most
The term semi-empirical is used for models that relies on physical model- well-known is presented in [Nicholas and Comstock, 1972]:
ing to transform empirical data collected at a certain condition to be valid
for another situation at different conditions. There is always a trade-off Fx (λ ) Fy(α )λ
Fx (λ , α ) = q
since the semi-empirical model is more or less accurate. The measure- λ 2 Fy2 (α ) + tan2 (α ) Fx2 (λ )
ment data on the other hand includes errors and disturbances. The testing (3.19)
circumstances also mostly differ from the driving or simulation circum- Fx (λ ) Fy(α ) tan(α )
Fy(λ , α ) = q
stances. In some cases collection of data can be difficult, unreliable, or λ 2 Fy2 (α ) + tan2 (α ) Fx2 (λ )
very expensive and a theoretical model is inevitable.
Semi-empirical methods are often used to derive the tire forces at com- In [Brach and Brach, 2000] this model is shown to give incorrect result
bined slip situations, when the pure-condition tire-forces often are avail- for small slips and a modified version is presented.
able from test-bench experiments. In [Bakker et al., 1987], a procedure for computing combined forces
The most simplistic model of combined forces is based on the friction for the Magic Formula is presented. It is essentially a refinement of the
ellipse concept, see for instance [Wong, 2001; Ellis, 1994; Nielsen and Kamm Circle for non-isotropic tire characteristics and a normalization of
Eriksson, 1999]. While the friction ellipse is the envelope of the maximum the slips to guarantee simultaneous sliding. The normalized slip
achievable forces, the ellipse is here used also for modeling intermediate
forces. It is used to compute a combined lateral force Fy(α , λ ) at a given
s 2 2
σx σy
longitudinal force Fx , and is based on the assumption σN = + (3.20)
σ ∗x σ ∗y
2 2
Fy(α , λ ) is an entity that is less than one for non-sliding conditions. It is based on
Fx
+ =1 (3.16)
∗
F0x F0y(α ) an elliptic assumption where σ ∗x and σ ∗y are the longitudinal and lateral
slips that corresponds to full sliding for pure slips, normally taken as
where F0x∗ is the maximum achievable longitudinal force, and F (α ) de-
0y
the slips at the peak values F0x ∗ , and F ∗ . Now the combined forces are
0y
scribe the lateral force at pure slip. The combined lateral force becomes computed as
(3.22)
sx sy J 2
Fx = F (s) and Fy = ks F (s) (3.18) with ϑ = π arctan( q1σ 2N ) and
s s
J
′
F0x = F0x (σ N ) − sat(σ N ) ( F0x (σ N ) − F0y(σ N )) sin2 (β ○ )
A drawback with this model is that longitudinal and lateral characteristics (3.23)
J
are assumed to be the same, modulo the corrective factor. ′
F0y = F0y(σ N ) + sat(σ N ) ( F0x (σ N ) − F0y(σ N )) cos2 (β ○ )
32 33
Chapter 3. Review of Existing Tire-Models 3.3 Transient Tire-Models
34 35
Chapter 3. Review of Existing Tire-Models 3.4 Finite Element Models
road, the patch is described by cams, see Figure 3.1. The contact between
the road and the cam decides the vertical position of the contact patch.
σ 0i λ (vr ) To be able to manage inclination of the surface, two cams (tandem) are
z̄˙i = vri − z̄i − κ i (t)pω rp z̄ (3.30)
µ 2ki necessary. Four cams (two rows of tandem cams) are needed to derive
Fi = − Fn (σ 0i zi (t) + σ 1i żi (t) + σ 2i vri )), i = x, y (3.31) camber angles due to road inclination in the lateral direction.
where
pp Mk2 vs pp
λ (vs ) = (3.32)
(vs )
and the friction function is redefined as
pp Mk2 vs pp
2
p Mk vs p p Mk2 vs p pvs p γ
(vs ) = + − exp − (3.33)
pp Mk vs pp p Mk vs p p Mk vs p vst
" # " #
µ kx 0 µ sx 0
Mk = Ms = (3.34)
0 µ ky 0 µ sy
and the parameters from the distributed steady-state description can then
be used in the lumped formulation. The derivation of κ , from (3.30) has to Figure 3.1 Illustration of the method of the SWIFT-model to derive the vertical
position for the contact point between the tire and the road. To the left single cam
be done to cover for the difference between the distributed and the lumped
and to the right tandem cams allowing determination of patch inclination.
formulation. More details on the model are given in [Velenis et al., 2002].
A drawback with the method is that the description of the bristle dy-
namics is complicated, resulting in comprehensive equations, while other
phenomenon are roughly simplified. The major source of the tire dynamics
still comes from the carcass behavior. 3.4 Finite Element Models
The Short Wavelength Intermediate Frequency Tyre (SWIFT) model
described, for instance, in [Pacejka, 2002] developed at TU-Delft and TNO- The increase of available computation power has provided a new platform
Helmond is a collection of methods employed to extend the MF-tire concept for use of finite element methods to solve the complex partial differen-
with dynamic properties and ability to handle uneven roads. The model tial equations describing the behavior of the tire. Briefly, this means that
can handle frequencies up to 60 Hz and wavelengths larger than 0.2 m the tire structure is divided into small elements. Each element has a set
and acts on longitudinal and lateral slip, camber, and turn slip. The tire of differential equations describing its physical behavior. The elements
belt is modelled as a rigid ring with inertia. The ring has a flexible at- are connected to each other through boundary surfaces. All elements con-
tachment to the rim, with stiffnesses in all directions, compare to the tains points or nodes, with constraints on the externally applied force, or
stretched string model in Section 2.6. To improve the transient properties deflection. The constitutive conditions and laws of equilibrium of forces,
further, the stiffness in the contact patch rubber is modelled as a first or- torques, and energy flow are used to derive matrix equations from which
der differential equation with a time constant corresponding to the size of the states in each node are solved. The finer meshing (small elements)
the adhesion area in the patch (see Section 4.1 for explanation of adhesive the larger matrices and memory consumption. The F-tire model [Gipser,
area). The pneumatic trail, used for the self-aligning torque is modelled 2005] uses between 80 and 200 sector elements to describe the carcass
as a second order differential function from the lateral slip, which means deformation. Each element has five degrees of freedom denoting, longi-
that self-aligning torque builds up slower than the lateral force. The turn- tudinal, lateral and vertical displacement, rotation, and bending. Above
slip torque-contribution to Mz on the rigid ring is modelled as a fourth that, between 5 and 10 tread elements on each carcass element are used to
order polynomial function relative the turn slip in the contact patch. To model the rubber treads. Each tread element contains information about
calculate the position for the contact patch when running on an uneven wear, temperature, and deflection. Since the tread elements outside the
36 37
Chapter 3. Review of Existing Tire-Models
contact area are unloaded they will not be included in the calculations
of the deflections. For a normal PC (2005) the calculations takes about
2 to 10 times real time. The model properties are determined by sets of
parameters, which can be calibrated from measurement data. F-tire is the
4
middle-complexity part in a family of three models. The simpler R-tire is
restricted by having a rigid carcass which is similar to the SWIFT-model
in Section 3.3. The most advanced model FE-tire is a coarse mesh finite
element tire model allowing multiple elements in the lateral direction.
Tire Modeling According to
The Rmod-K [Oertel and Fandre, 1999] is another tire model that
similarly to the FE-tire is based on coarse-mesh finite-elements. the Brush Model Theory
3.5 Conclusions
38 39
Chapter 4. Tire Modeling According to the Brush Model Theory 4.1 Basics of the Brush Model
Z t c ( x)
xr ( x ) = a − vx dt
Road Slide Adhesion 0
Z t c ( x)
(4.1)
x
−a xs 0 a yr ( x) = − vy dt
0
where tc ( x) is the time elapsed since the bristle entered the contact re-
gion. The velocities vc , vx and vy are assumed to be constant as a bristle
travels through the adhesive region of the contact patch, i.e. during the
vsy integration interval [0, tc ( x)]. Hence, the bristle position is x = a − vc tc ( x),
δ xb ( x) and tc ( x) = (a − x)/vc . The deformation becomes
δ xs ( x) = xr ( x) − x
(4.2)
δ yb ( x) δ ys ( x) = yr ( x)
y
where the slip definition from (2.5) is used in the last equality. As the
region into small brush elements. Each element stretches laterally ( y-
carcass is assumed to be stiff the the bristle deformations are accordingly,
direction) over the entire contact region, but their length is infinitesimal
δ xb = δ xs and δ yb = δ ys . With the assumption of linear elasticity, the
in the longitudinal, x, direction. The elements are regarded as elastic rect-
deformation force working on the bristles becomes
angular blades, or bristles, see Figure 4.1. Even though rubber, in general,
is not linearly elastic, this assumption is made in the brush model. Posi-
tions in the contact region are expressed in a reference system attached dFax ( x) = cpx dx δ xb ( x)
(4.4)
to the carcass, with the origin located in the center of the contact region. dFay( x) = cpy dx δ yb ( x)
The length of the contact region is 2a. Each bristle is assumed to deform
independently in the longitudinal and lateral directions. In the adhesive where cpx and cpy are the longitudinal and lateral bristle stiffnesses per
region the bristles adhere to the road surface and the deformation force is unit length. The assumption of constant vc , vx , vy in the interval [0, tc ( x)]
carried by static friction. In the sliding region the bristles slide on the road is relaxed to the assumption of slow variations in σ x and σ y with respect
surface under influence of sliding friction. Hence, in the sliding region the to the duration 2a/vc , which is the maximum time for a bristle to travel
resulting force is independent of the bristle deformations. through the adhesion region. The total adhesive tire force is computed by
40 41
Chapter 4. Tire Modeling According to the Brush Model Theory 4.1 Basics of the Brush Model
µ sy dFz( x)
dFy( x)
dF
yr ( x) x dFx ( x)
µ sx dFz( x)
xr ( x)
x
y
Figure 4.3 Illustration of the elliptic static friction constraint at anisotropic fric-
tion and rubber characteristics. Note that the direction of dF̄ ( x) and σ̄ is equal only
Figure 4.2 The deformation of a bristle element in the contact patch. Compare if c px /c py = µ sx /µ sy.
with Figure 4.1.
region and start to slide. Introduce the pressure distribution qz( x), with
integration of (4.4) over the adhesive region. With (4.3) this gives dFz ( x) = qz( x) dx. By combining (4.3) and (4.4) with (4.6) the static fric-
tion constraint may be written as
Z a Z a
Fax = dFax ( x) = −cpxσ x (a − x) dx
s 2 2
cpxσ x cpyσ y
x
Z sa
x
Z as (4.5) + (a − x) ≤ qz ( x) (4.7)
µ sx µ sy
Fay = dFay ( x) = −cpyσ y (a − x) dx
xs xs
The position xs in the contact area is the break-away point where the static
friction limit is reached and the bristles starts to slide. If the pressure
where xs is the position in the contact patch which divides the adhesive
distribution qz( x) is known then xs can be calculated by setting equality
and sliding regions. To compute the total adhesive force it is necessary to
in (4.7) with x = xs .
know xs .
A common assumption is to describe the pressure distribution in the
contact patch as a symmetric parabolic function:
The size of the adhesion region The size of the adhesive region is
determined by the available static friction. The deformation will be limited 3Fz x 2
by the largest force that can be carried by the static friction between the qz ( x) = 1− (4.8)
4a a
tire and the road. The static friction is assumed to be anisotropic (i.e.
depending on the sliding direction) with the friction coefficients µ sx and This is proposed, for example, in [Tielking and Mital, 1974] and has shown
µ sy, respectively. With a normal force dFz ( x) acting on the infinitesimal to give a good agreement with experimental longitudinal force-slip curves
bristle at position x, the available static friction force is described by the for real tires. In [Svendenius, 2003] the influence of the pressure distribu-
elliptic constraint tion is further examined. A more realistic pressure distribution is obtained
by moving the pressure-peak slightly forward in the contact patch. The
2 2
gain in accuracy is, however, small compared to increase in complexity of
dFax ( x) dFay( x)
+ ≤1 (4.6) the resulting formulas and it can be concluded that the proposed parabolic
dFz ( x)µ sx dFz ( x)µ sy
assumption is sufficient for the applications related to the brush model in
As a result, the magnitude of the available static friction force is depen- this work. Inserting (4.8) in (4.7) with equality gives
dent of the direction of the deformation force dF̄a ( x), defined by (4.4). s 2 2
The static friction constraint is illustrated in Figure 4.3. When dF̄a ( x) cpxσ x cpyσ y 3Fz
+ ( a − xs ) = ( a − xs ) ( a + xs ) (4.9)
exceeds the static friction constraint the bristle will leave the adhesive µ sx µ sy 4a3
42 43
Chapter 4. Tire Modeling According to the Brush Model Theory 4.1 Basics of the Brush Model
J 3Fz µ sx J 3Fz µ sy
σ ○x = ; σ ○y = (4.11) Fax (σ x , 0)
2a2 cpx 2a2 cpy
Introduction of normalized slips with respect to the limit slips will simplify
the notation in the following. The normalized slip is defined as −a xs (σ x , 0) 0 a x
s 2 2
J σx σy
ψ (σ x , σ y) = + (4.12) Figure 4.4 Illustration of the adhesive tire-force for pure longitudinal slip. The
σ ○x σ ○y elastic deformation force for an element at x in the adhesive region depends linearly
on x as c pxσ x ( x − a), where the slope is proportional to the slip σ x . The transition
Equation (4.10) may now be rewritten as from adhesion to slide occurs at the intersection of the lines at the break-away point
xs . For slips σ x > σ x○ full sliding occur in the contact area since there is then no
intersection.
xs (σ x , σ y) = (2ψ (σ x , σ y) − 1)a (4.13)
It is clear that partial sliding occurs when ψ (σ x , σ y) < 1. At full sliding Note that it follows from (4.4) and (4.3) that the produced adhesive force
then ψ (σ x , σ y) ≥ 1 and Fax (σ x , σ y) = Fay(σ x , σ y) = 0. In the following per unit length in the adhesion region is not affected by combined slips:
the construction of adhesive and sliding forces at partial sliding will be
determined.
dFax (σ x , x)
= −cpxσ x (a − x)
Knowing the size of the adhesive region, xs (σ x , σ y), dx
Adhesion force (4.16)
given by (4.13), the adhesive forces are obtained by solving the integrals dFay(σ y , x)
= −cpyσ y (a − x)
(4.5) yielding dx
44 45
Chapter 4. Tire Modeling According to the Brush Model Theory 4.1 Basics of the Brush Model
Combined-slip slide forces The normal force acting on the sliding a way that the mechanical work W = −v̄s ⋅ F̄sMDR is maximized under the
region at partial sliding may be computed from (4.8) and (4.13) as constraint !2
MDR 2 MDR
Fsx Fsy
Z xs (σ x ,σ y ) + ≤1 (4.21)
Fsz(σ x , σ y) = qz ( x) dx = Fzψ 2 (σ x , σ y) (3 − 2ψ (σ x , σ y)) (4.17) Fsz µ kx Fsz µ ky
−a
This results in the sliding forces
In case of isotropic sliding friction with the friction coefficient µ k ,
the friction force is collinear with the slip velocity with the magnitude MDR µ 2kx vsx
Fsx (σ x , σ y) = − p Fsz (σ x , σ y) = −µ kx cos (β SPM ) Fsz
Fsz (σ x , σ y)µ k . Its components are given by (µ kx vsx )2 + (µ kyvsy )2
MDR
µ 2kyvsy
Fsx (σ x , σ y) = − cos (β )µ k Fsz(σ x , σ y) Fsy (σ x , σ y) = − p Fsz (σ x , σ y) = −µ ky sin (β SPM ) Fsz
(4.18) (µ kx vsx )2 + (µ kyvsy )2
Fsy(σ x , σ y) = − sin (β ) µ k Fsz (σ x , σ y)
(4.22)
where β is defined by (2.4). Assumptions on isotropic sliding friction are
where β SPM is defined as
common in tire modeling, see for instance [Schuring et al., 1996].
If the sliding-friction is anisotropic with the different friction coeffi- J µ ky vsy
tan(β SPM ) = (4.23)
cients µ kx and µ ky, there are several ways to calculate the magnitude and µ kx vsx
the direction of the resulting force. Three different methods are presented
in the following and which one to choose depends on the assumptions The angle of the resulting force F̄sMDR is denoted by β MDR and is given by
made on the friction behavior for the actual case.
2
µ ky vsy
Collinear slide forces This method should be used if the friction be- tan(β MDR ) = (4.24)
µ kx vsx
tween two surfaces is supposed to be isotropic, but the values of µ kx and
µ ky are unequal. A reason for that could, for instance, be errors in the
measurements of the pure-slip forces. The friction forces are given by
Slip-projection method An intermediate approach to model aniso-
Col tropic sliding friction is to simply replace µ k in (4.18) with µ kx and µ ky in
Fsx (σ x , σ y) = − cos (β )µ kx Fsx (σ x , σ y)
′
(4.19) the corresponding directions:
Col
Fsy (σ x , σ y) = − sin (β ′ )µ ky Fsy(σ x , σ y)
SPM
Fsx (σ x , σ y) = − cos (β ) µ kx Fsz(σ x , σ y)
where β is defined as
′
SPM
(4.25)
Fsy (σ x , σ y) = − sin (β )µ ky Fsz(σ x , σ y)
−1
J µ ky vsy This means a projection of the pure-slip sliding-forces on the slip vector.
tan(β ′ ) = (4.20)
µ kx vsx The angle of the resulting force is then equal to β SPM . From the definitions
of β (2.4), β SPM (4.23), and β MDR (4.24), it is clear that the direction of
The choice of β ′ ensures that F̄s acts in the opposite direction to the F̄sSPM is between the directions of F̄sCol and F̄sMDR , see Figure 4.5.
sliding motion, with a friction coefficient that is somewhere in the interval To summarize: The sliding forces are described by
[µ kx , µ ky] depending on the sliding angle β .
Fsx (σ x , σ y) = − cos(β f )µ kx Fsz (σ x , σ y)
(4.26)
Maximum dissipation rate The correct way to treat anisotropic fric- Fsy(σ x , σ y) = − sin(β f )µ ky Fsz (σ x , σ y)
tion according to the literature is to apply the Maximum Dissipation Rate
(MDR) principle. The theory which is further presented in [Goyal, 1989] with
claims that the resulting sliding-friction force F̄sMDR is generated in such Fsz(σ x , σ y) = Fzψ 2 (σ x , σ y) (3 − 2ψ (σ x , σ y)) (4.27)
46 47
Chapter 4. Tire Modeling According to the Brush Model Theory 4.1 Basics of the Brush Model
µ sx qz( x)
µ kx qz( x)
FsSPM
FsCol FsMDR
σ x c px
β SPM
β MDR
β
µ ky Fsz
µ kx Fsz
Fax (σ x , 0)
Fsx (σ x , 0)
σ̄
−a xs (σ x , 0) 0 a x
Figure 4.5 Illustration of methods to describe kinetic friction in case of different
longitudinal and lateral friction coefficients.
Figure 4.6 Illustration of partition of the contact area into a sliding and an ad-
hesive region for the case of pure longitudinal slip. The slide force for an element at
x is determined by the pressure distribution µ kx qz( x) dx. The horizontally striped
and β f is any of β ′ (collinear), β (slip-projection) or β SPM (MDR) depend- area is the total slide force.
ing on choice of friction model:
!−1
J µ ky vsy J vsy J µ ky vsy
tan(β ) =
′
; tan(β ) = ; tan(β SPM ) = (4.28)
µ kx vsx vsx µ kx vsx
To illustrate the effect of combined slips Figure 4.7 shows the produc-
In the special case of pure-slip the sliding-forces are tion of longitudinal force in the case of combined longitudinal and lateral
slip (σ x , σ y) with σ x ,= 0, σ y ,= 0. Equation (4.13) shows that the ad-
F0sx (σ x ) = −µ kx Fsz(σ x , 0) sgn(σ x ) hering region shrinks compared to the case with pure slip (σ x , 0). The
(4.29) sliding region grows accordingly. From (4.16) it is clear that the adhesive
F0sy(σ y) = −µ ky Fsz(0, σ y) sgn(σ y)
force per unit length is the same for the combined slip (σ x , σ y) as for the
pure-slip (σ x , 0). Hence, the slope is the same, but the area corresponding
In Figure 4.6 the case of pure longitudinal slip is again regarded, now with to the force is smaller since the adhering region is smaller. The corre-
also the sliding force introduced. Since qz( x) is the normal force per unit sponding adhesive-force slope derived from (4.16) is cpxσ ○xψ (σ x , σ y). The
length, the sliding force per unit length is simply µ kx qz ( x), as marked in corresponding expression applies for the lateral force. It is therefore clear
the figure. The horizontally striped area corresponds to the total sliding
that sliding will occur simultaneously in both directions as ψ (σ x , σ y) ap-
force. proaches unity. It is important to note that the indicated area under the
pressure distribution no longer corresponds to the resulting sliding force.
Effects of combined slips The total tire force is given by adding the Instead it describes µ kx Fsz (σ x , σ y), which is the force that would result
adhesive forces of (4.14) and the sliding forces of (4.26): for pure longitudinal sliding with the sliding region xs (σ x , σ y). This force
must be limited by a friction constraint according to Section 4.1.
Fx (σ x , σ y) = Fax (σ x , σ y) + Fsx (σ x , σ y) The braking and cornering stiffnesses are the linearizations of the
(4.30)
Fy(σ x , σ y) = Fay(σ x , σ y) + Fsy (σ x , σ y) pure-slip friction curves at small slips and may be computed by derivation
48 49
Chapter 4. Tire Modeling According to the Brush Model Theory 4.1 Basics of the Brush Model
µ kx qz( x) Z a
′
Maz (σ x , σ y) = −cpyσ y x (a − x) dx
xs (σ x ,σ y )
σ x c px σ x○ψ (σ x , σ y)c px σ x○ c px 2
= −cpy a3σ y (1 − ψ (σ x , σ y))2 (4ψ (σ x , σ y) − 1) (4.33)
3
Z xs (σ x ,σ y )
µ kx Fsz(σ x , σ y) Fax (σ x , σ y) ′
Msz (σ x , σ y) = −µ ky sin (β ) x qz( x) dx
−a
= −3µ kx sin (β )aFzψ 2 (σ x , σ y)(1 − ψ (σ x , σ y))2 (4.34)
−a xs (σ x , 0) xs (σ x , σ y) 0 a x
Mz′ (σ x , σ y) = Maz
′
(σ x , σ y) + Msz
′
(σ x , σ y) (4.35)
Figure 4.7 Illustration of the effect of combined slip. The combined-slip has the
effect of decreasing the size of the adhesive region, compare with Figure 4.6.
When there is a lateral slip the tire deflects laterally and the point
of action for the longitudinal force will have an offset from the central
plane of the wheel. This produces an additional deformation torque in the
z-direction. A longitudinal deflection together with a lateral force has the
of (4.30):
same effect. Since it is assumed that the carcass is stiff the deformation is
here described by bristle deflections, see also Section 4.3. The deformation
torque developed at position x in the contact region is described by
Fx (σ x , 0)
Cx = − = 2cpx a2
σ x
σ =0
x (4.31) dMz′′ ( x) = dFy ( x)δ xb ( x) − dFx ( x)δ yb ( x) (4.36)
Fy(0, σ y)
Cy = − = 2cpy a2
σ y
σ y =0 In the same way as above, integration over the adhesive and the sliding
regions is performed separately. The deformation δ xb ( x) is computed from
Self-aligning torque The self-aligning torque consists of two parts. (4.3) in the adhesive region and from (4.4) using the infinitesimal sliding
The main part is Mz′ , which is the torque developed by the non symmetric force in the sliding region. Hence
distribution of the lateral force Fy. The additional part Mz′′ is due to the
deformation of the tire. a
Z
The torque dMz′ developed at position x in the contact region is
′′
Maz (σ x ,σ y) = cpyσ y(a − x)σ x (a − x) dx
xs (σ x ,σ y)
Z a
dMz′ ( x) = dFy ( x) x (4.32) − cpxσ x (a − x)σ y(a − x) dx
xs (σ x ,σ y )
(4.37)
4
= ( Cy − Cx )aσ xσ y(1 − ψ (σ x , σ y))3
In the adhesive part of the contact region the expression for dFy ( x) is 3
given by (4.4) together with (4.3). In the sliding zone it is given by differ- 4 1 1
a
entiating (4.26) using dFz ( x) = qz ( x) dx from (4.8). Integration over the = − Fax (σ x , σ y) Fay(σ x , σ y)
3 Cx Cy (1 − ψ (σ x , σ y))
50 51
Chapter 4. Tire Modeling According to the Brush Model Theory 4.1 Basics of the Brush Model
Z xs (σ x ,σ y)
1
Msz (σ x , σ y) =
′′
µ ky sin (β f ) qz ( x)µ 2kx cos (β f ) qz( x) dx 40
−a c px
−Fx [kN]
Z xs (σ x ,σ y)
1 20
− µ kx cos (β f ) qz( x)µ ky sin (β f ) qz ( x) dx
−a c py
0
6 1 1
2 0 20 40 60 80 100
= − µ kx µ ky a sin (β f ) cos (β f ) Fz (4.38)
5 Cx Cy 40
−Fy [kN]
20
6 1 1 a(10 − 15ψ (σ x , σ y) + 6ψ 2 (σ x , σ y))
= −
5 Cx Cy ψ (σ x , σ y)(3 − 2ψ (σ x , σ y))2 0
0 20 40 60 80 100
⋅ Fsx (σ x , σ y) Fsy (σ x , σ y) 1
where (4.14) and (4.26) have been used in the last step. The total addi-
M [Nm]
0.5
tional torque is
z
0
Mz′′ (σ x , σ y) = ′′
Maz (σ x , σ y) + ′′
Msz (σ x , σ y) (4.39) −0.5
0 20 40 60 80 100
λ [%]
and finally the total self-aligning torque is
Figure 4.8 Tire forces as function of λ with α = [0, 5, 10, 20] deg. The dotted line
shows the force from the adhesive region, the dashed line shows it from the sliding
Mz (σ x , σ y) = Mz′ (σ x , σ y) + Mz′′ (σ x , σ y) (4.40) region. The solid line is the total force and the dashed-dotted line is the reference
curve generated from a Magic Formula approximation of real tire data [Gäfvert
A commonly used parameter is the pneumatic trail, which denotes the and Svendenius, 2003]. For the self aligning torque the dotted line denotes the
distance between the center of the tire and point of action for the lateral deformation torque Mz′′ and the dashed line M ′ . The slip definition λ is used since it
is most common when data is visualized. Refer to Equation (2.6) for transformation
force. It is defined as t(σ x , σ y) = Mz(σ x , σ y)/ Fy(σ x , σ y). The coordinate between λ and σ x .
for the point of action for the adhesive force is denoted by ta (σ x , σ y) and
for the sliding force by ts (σ x , σ y). By using (4.33) and (4.14) respectively
(4.34) and (4.26) the contributions from Mz′ (σ x , σ y) to the pneumatic trail, Properties of the brush-model
t′a (σ x , σ y) and t′s (σ x , σ y), are given by
Figures 4.8 and 4.9 show the behavior of the brush model at pure and
combined slip. The pure slip curves are compared to a Magic-Formula
′
Maz (σ x , σ y) a
t′a (σ x , σ y) = = (4ψ (σ x , σ y) − 1) parametrisation of a classified truck tire. The brush-model parameters are
Fay(σ x , σ y) 3 chosen so that the pure-slip curves have the same braking and cornering
(4.41)
′
Msz (σ x , σ y) (1 − ψ (σ x , σ y))2 stiffnesses and the same peak force as of the Magic-Formula reference-
t′s (σ x , σ y) = = −3a
Fsy (σ x , σ y) (3 − 2ψ (σ x , σ y)) curve. The length of the contact-patch is chosen such that the aligning-
stiffness agrees with the self-aligning torque reference-curve. For small
The contributions from Mz′′ (σ x , σ y) can be read directly from (4.37) and pure longitudinal slip the coherence between the brush model and the
(4.38). reference curve is good. For slips larger than the value corresponding to
In the same way as for the braking and cornering stiffness, the aligning the peak force, the curves do not agree since no velocity-dependency is
stiffness is defined as modelled for the friction. This have the effect that the modelled tire force
is constant for higher slip, while here and most commonly the tire force
decreases in the region. Different ways to introduce velocity dependence
Mz 2 a
Cz = = cpy a3 = Cy (4.42) in the friction is, for example, discussed in [Svendenius, 2003]. At partial
σ y 3 3
σ x ,σ y=0 sliding (ψ < 1) this approximation normally has small effects.
52 53
Chapter 4. Tire Modeling According to the Brush Model Theory 4.2 The Effect of Camber
40
xs
−F [kN]
20
x
0 i
0 10 20 30 40 50
δ xs
−F [kN] 40
20
y
y δ y,cam δ ys
0
0 10 20 30 40 50
1 −a 0 a x
M [Nm]
0.5
Figure 4.10 Schematic illustration of the contact patch showing the bristle defor-
z
0
mation due to longitudinal and lateral slip and cambering.
−0.5
0 10 20 30 40 50
α [deg]
For pure lateral slip there are disparities in the lateral force and in
Bristle deflection
the self-aligning torque compared to the Magic-Formula reference, even
at lower slips. Probably, the main reason for this is the assumption of a The developed tire force due to tilting of the tire, can with some approxi-
stiff carcass, which is reasonably accurate in the longitudinal direction. mations, be explained by the brush model. In Figure 2.3 a cambered tire
Laterally, where the carcass is weaker, the effects of this simplification is is shown together with the orbit, an ellipse, that describes the projection
noticeable. on the road surface of a point on the carcass during rolling motion. The
There are several ways to include carcass flexibility based on assump- deviation from the straight contact patch from a non-cambered tire is
tions on stretched string or beam behavior in the brush model. Some of p √
them are further discussed in Section 4.3. Also the self-aligning torque y = − sin(γ ) R2 − x2 − R2 − a2 (4.43)
depends on the flexibility of the carcass which might explain some of its
disagreement to the reference. A factor that can explain the deviation of where R is the average wheel radius and a is half the contact length.
the self-aligning torque, particularly at high lateral slip is the deviation Figure 4.10 also shows the deformation of bristle element due to cam-
of the actual pressure distribution from the parabolic assumption in (4.8). bering and lateral slip according to the brush model. Relation (4.43) will
If the position of the center of the vertical load differs from the hub cen- result in difficult expressions when the standard parabolic pressure distri-
ter, the pneumatic trail will not become zero at higher slips. Hence, the bution is employed. In, for example, [Gim and Nikravesh, 1991] and [Pace-
lateral force then gives a torque contribution. This have not been further jka, 2002] the deviation due to cambering is approximated as a parabolic
examined in this thesis, but the effects on the longitudinal tire force, due function similar to the assumed pressure distribution as
to non-uniform pressure distribution is penetrated in [Svendenius and
Wittenmark, 2003]. δ y,cam ( x) = −γ k a2 − x2 (4.44)
54 55
Chapter 4. Tire Modeling According to the Brush Model Theory 4.2 The Effect of Camber
Bristle Force
−a −a 1
0
1
0
1111111111
0000000000 1
0
1
0
0000000000
1111111111
0000000000000000000000000
1111111111111111111111111
c pyδ yb 1
0
1
0
The approximation simplifies the calculations considerably. The deforma- 0000000000
1111111111
0000000000000000000000000
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
1
0
0000000000
1111111111
0000000000000000000000000
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
1
0
tion of a bristle that adheres to the road in the contact patch is then 0000000000
1111111111
0000000000000000000000000
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
1
0
−δ y,cam and the total bristle deformation including the effects caused by 0000000000000000000000000
1111111111111111111111111
c pyδ ys 1
0
0000000000000000000000000
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
1
0
slip, under the assumption of a stiff carcass, can be written as 0000000000000000000000000
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
1
0
2aψ 0000000000000000000000000
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
1
0
δ xb ( x) = −σ x (a − x) (4.46) 0000000000000000000000000
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
c pyδ y,cam 1
0
0000000000000000000000000
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
1
0
2 2 0000000000000000000000000
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
1
0
δ yb( x) = −σ y(a − x) + γ k(a − x ) (4.47) 0000000000000000000000000a
1111111111111111111111111
1
0
0000000000000000000000000
1111111111111111111111111
1
0
1
0
1
0
−a xs 0
Contact Patch
Size of adhesive region
The calculations in Section 4.1 can be used to determine the size of the
Figure 4.11 Force generation in the contact patch due to slip and camber. The
adhesion region. By using (4.5) in (4.6) with the addition from camber on
shaded area is the force generated in the sliding region. The lower area with diagonal
δ yb from (4.44), the expression for deriving the break-away point is given lines denotes the adhesion force generated by slip and the upper the contribution
as due to camber. The dashed straight line indicates the sliding region in case of zero
camber.
s 2 2
cpxσ x cpy (σ y + γ k(a + xs )) has the same shape as the pressure distribution along the contact patch,
+ (a − xs ) ≤ qz ( xs ) (4.48)
µ sx µ sy ψ will be zero as long as σ x = σ y = 0. At the camber angle γ ○ the whole
contact patch starts to slide against the road and ψ can not be computed
with the assumption of the pressure distribution from (4.8) and the defi- since the denominator reaches zero. In this case the model is not valid
nition of the normalized slip, ψ , compared to (4.12), is extended to and if accurate results are to be obtained a different tire model should be
used. Such large camber angles are, however, not relevant for studies on
trucks and cars.
s
σx 2
2
γ ○2 σ yγ σ xγ 2
J σy
ψ (σ x , σ y, γ ) = ○2 + + −
γ − γ 2 σ ○yγ ○ σ ○x σ ○y σ ○xγ ○
Forces and torque
(4.49) The forces generated in the contact patch are in Section 4.1 derived sepa-
where the limit slips, σ ○x , σ ○y are defined in (4.11), and the camber limit rately for the adhesive and sliding regions. As mentioned, camber affects
angle γ ○ are defined as the natural path of the carcass in the contact region, which affects the
J 3Fz µ sy
γ○ = (4.50) force contribution from the adhesive region. In the sliding area the de-
2Cy ka formation, δ x , δ y, changes, since the vertical force on the bristle varies
The break-away point can be written as in (4.13) according to the parabolic pressure distribution. The rate of change of the
deformation is small and the effect on the sliding velocity is neglected. The
xs (σ x , σ y, γ ) = (2ψ (σ x , σ y, γ ) − 1) a (4.51) curvature of the bristle path resulting from cambering is small for small
camber angles and the effect on the sliding friction forces is neglected. In
The partitioning of the contact patch and the significance of ψ and xs is other words, the only effect cambering has on the sliding force is that it
visualized in Figure 4.11. It is clear that values of ψ larger than unity lack changes the size of the sliding region, which already is included by use of
physical interpretation. Since the bristle deformation due to cambering ψ from (4.49)
56 57
Chapter 4. Tire Modeling According to the Brush Model Theory 4.3 Effects of a Flexible Carcass
Adhesion The forces from the adhesive region derived in (4.14) with use 4.3 Effects of a Flexible Carcass
of ψ from (4.49) are still valid. In the lateral direction the force addition
due to the camber is the sum of the force contribution from each bristle
given by (4.44) with (4.4) as Lateral deformation of the carcass
As mentioned previously the brush model described in Section 4.1 is based
a
on the assumption of a stiff carcass. In reality the carcass is flexible, as
Z
Fay,cam (σ x , σ y, γ ) = cpyγ k(a2 − x2 )dx mentioned in Section 2.6 and exhibits significant deformation. Figure 4.12
xs (σ x ,σ y ,γ )
illustrates how the carcass deformation affects the brush model laterally.
2
= γ kaCy 2ψ 3 (σ x , σ y, γ ) − 3ψ 2 (σ x , σ y, γ ) + 1 (4.52) Mainly, the break-away point will move backwards and the deflection of
3
the bristles increases differently in the contact patch. The bristles and
the carcass can be seen as two spring elements connected serially, with
The camber stiffness, Cγ is defined as
the distinction that each bristle deflects individually, when the carcass is
a coherent unit. The total carcass deflection is denoted by δ yctot , and δ yc
Fy(0, 0, γ ) 2kaCy
Fz µ sy
Cγ = = = (4.53) is the relative deflection, compared to the deflection in the leading point
γ
γ =0 3 γ○ of the contact patch, ( x = a), as δ yc ( x) = δ yctot ( x) − δ yctot (a). The total
deformation relative the carcass deformation in same point is the sum of
In case of pure cambering for γ < γ ○ the lateral force is the bristle and carcass deflections δ yb ( x) and δ yc ( x)
In Section 5.4 results on how the combined slip forces and torque are Figure 4.12 Lateral tire deformation in the contact patch according to the brush-
affected by camber are shown. model. Left: Stiff carcass. Right: Flexible carcass.
58 59
Chapter 4. Tire Modeling According to the Brush Model Theory 4.3 Effects of a Flexible Carcass
Using (4.6) at pure lateral slip together with a parabolic pressure distri-
bution (4.8) the position for the break-away point can be solved from
δ yb
Fy′ (σ y) 3µ ay Fz
cpy σy − = ( a + xs ) (4.65)
Cc 4a3
−a a α x
x′s δ yc y
The total lateral force can be derived from the following equation
Figure 4.13 Lateral tire deformation in the contact patch with carcass deforma- Z xs ( Fy′ (σ y ))
a Fy′ (σ y )
Z
tion according to the assumption in Section 4.3.
Fy′ (σ y) = cpy σy − (a − x)dx + µ qz dx (4.66)
xs ( Fy′ (σ y )) Cc −a
60 61
Chapter 4. Tire Modeling According to the Brush Model Theory 4.3 Effects of a Flexible Carcass
of the points of action for the tire forces due to the carcass deformation.
20
where Ccx and Ccy are the carcass-stiffness coefficients, which may be
5 calculated from (4.82). The self-aligning torque contribution Mcz
′′
should
be added to the result in Equation (4.40).
0
0 5 10
α [deg]
15 20 25
Tire dynamics
In a tire with rigid carcass the motion in tire-road interface may be directly
Figure 4.14 Comparison of the brush-model with (dashed) and without (dashed described by the motion of the wheel rim. With a flexible carcass there is
dotted) compensation for a flexible carcass. Asterisks denotes points for total sliding
a dynamic relation between these motions. When deriving the relation a
α ○′ and α ○ . Solid line is a Magic-Formula parametrisation of empirical data further
presented in [Gäfvert and Svendenius, 2003]. few entities have to be redefined, compare to Figure 2.1 and Section 2.4.
The slip speed in the contact patch of a tire with rigid carcass is given by
The divergence between the longitudinal and lateral stiffnesses may now v′sx = vx − ω R e ; v′sy = vy (4.73)
be explained by the carcass stiffness, which can be calculated from (4.69)
as where the motion of the rim is described by the velocities v′x and v′y, the
Cx Cy′ wheel rotational velocity Ω , and the effective rolling radius R e . To be able
Cc = (4.70) to distinguish between motions related to the carcass and the motions
Cx − Cy′
related to the rim, the latter entities are marked by an apostrophe. The
where 2a2 cpx = Cx according to (4.31). Using (4.70) then the limit-slip corresponding slip velocities with a flexible carcass are
adjusted for carcass deformation of (4.67), σ ○y ′ , can be written as
vsx = v′sx + δ˙x ; vsy = v′sy + δ˙y (4.74)
2 1
σ ○y ′ = Fz µ y + ′ (4.71) where δ x and δ y are the deformations of the carcass as illustrated in
Cx Cy
Figure 4.15. The slip velocities may be normalized to form the tire slips. In
practical models they are best normalized with the longitudinal velocity
vx . The slips corresponding to tires with rigid (primed variables) and
Discussion Introducing flexibility in the carcass as described above flexible carcasses can be written as
improves the accuracy of the brush model in the lateral direction signif-
icantly, as can be seen in Figure 4.14. The approach can be used even v′sx δ˙x
λ′ = ; λ = λ′ +
for combined slip. The reliability of the brush model is still better in the vx vx
(4.75)
longitudinal direction, but assuming more realistic lateral deflections will v′sy δ˙y
complicate the expressions considerably. The major aim of this section is tan (α ′ ) = ; tan (α ) = tan (α ) +′
vx vx
62 63
Chapter 4. Tire Modeling According to the Brush Model Theory 4.3 Effects of a Flexible Carcass
The tire forces that are generated in the contact patch, Fx (λ , α , γ ) and Then, by expansion of the time derivative of the forces as
Fy(λ , α , γ ), are transmitted to the wheel rim by the flexible carcass. The
contact patch forces depend on the motion in the friction interface, which Fx Fx Fx
Ḟx = λ̇ + α̇ + γ˙
may be described by the tire slips, and the camber angle γ . In this sec- λ α γ
tion (λ , α ) are chosen to represent the slip instead of (σ x , σ y) as used (4.78)
Fy Fy Fy
previously. The reason is that (λ , α ) is most commonly used when visu- Ḟy = λ̇ + α̇ + γ˙
λ α γ
alizing data and results and the conversion between the slips are often
straight forward by using (2.6). Here, those slips are used to show the and by differentiation of the right column of (4.75) under the assumption
extra step that is necessary for the differentiation of tan(α ), see (4.79). that the vehicle velocity varies slowly, the carcass dynamics are formulated
Use of (σ x , σ y) simplifies the calculations in that λ , tan(α ) and vx are as
replaced by σ x , σ y and vc in (4.75).
1
Dx ˙′ Fx Fx Fx
λ −λ =′
λ + − D x vx λ̇ + α̇ + γ˙ (4.79a)
Ccx Ccx vx λ α γ
Dy 1
Ω tan(α ) − tan(α ′ ) = (1 + tan2 (α ′ ))α˙ ′ +
Ccy Ccyvx
Fy Fy Fy
y x
⋅ λ̇ + − D yvx (1 + tan2 (α )) α̇ + γ˙ (4.79b)
z λ α γ
z
which relates the contact patch slips to the wheel-rim slips. Note that the
−δ y −δ x dynamics include cross-couplings.
vsy vsx
Without damping the system becomes unstable in slip regions where
− Fy − Fx
the slip-force characteristics have positive slope. To avoid this, the damp-
v′sy v′sx ing coefficients should be chosen so that D x vx,min > sup( Fx /λ ) and
D yvx,min > sup( Fy/α ) for an arbitrary choice of slowest vx,min > 0. The
Figure 4.15 Carcass deformation of a tire. To the left: Lateral deformation. To system is then well posed for all vx ≥ vx,min . In [Pacejka, 2002] it is sug-
the right: Longitudinal deformation for a braked wheel. gested to solve this problem by preventing the diagonal terms, Fx /λ ,
Fy/α from being positive by limitation. The introduction of damping
may be a more physically motivated modification.
The tire carcass is assumed to behave like a linear spring and damper
The carcass stiffnesses are properties that might not be available for
such that
the specific tire, but quite often the relaxation length, σ a of a tire is mea-
Fx = Ccxδ x + D xδ˙x ; Fy = Ccyδ y + D yδ˙y (4.76) sured. The relaxation length is usually defined for small slips [Pacejka,
2002] and relate the dynamics between the rim and the carcass as
where Ccx and Ccy are the carcass stiffnesses in respective directions.
Using (4.75) together with the time derivative of (4.76) gives σ ax σ ay
λ̇ + λ = λ ′ ; α̇ + α = α ′ (4.80)
vx vx
Ḟx D xδ¨x Note that vx is time varying, and that (4.80) therefore is not a linear
λ − λ′ = −
Ccx vx Ccx vx time-invariant system. Instead, the differential equations are linear space-
(4.77)
Ḟy D yδ¨y invariant, since the independent time variable t can be exchanged for the
tan(α ) − tan(α ′ ) = − space variable s (rolling distance) as λ̇ = dλ / ds ⋅ ds/ dt and ds/ dt = vx ,
Ccyvx Ccyvx
64 65
Chapter 4. Tire Modeling According to the Brush Model Theory
and correspondingly for the lateral slip. In the case of small pure slip, zero
camber, and neglecting the damping effects (4.79) might be simplified to
1 Fx 1 Fy
5
λ − λ′ = λ̇ ; α −α′ = α̇ (4.81)
Ccx vx λ Ccyvx α
It is then clear that the relaxation lengths are related to the stiffnesses
as
A Dynamic Semi-Empirical
Cx Cy
; (4.82)
σ ax =
Ccx
σ ay =
Ccy Tire-Model for Combined
4.4 Summary Slips and Camber
In this chapter the well-known brush model has been presented in detail.
The brush model in its ordinary form renders simplistic expressions for
the tire-force and slip relations. There are, however, many shortcomings,
which can be overcome by more extensive modeling. The trade-off prob-
This chapter presents a method to derive the tire forces during com-
lem between simplicity and accuracy is obvious and it is important to be
bined slip and camber, given the empirical tire-force models at pure slip.
aware about the demands of the application when adopting a model for a
The method has previously been presented in a number of journal arti-
particular purpose. Each extra feature increases the complexity consider-
cles [Gäfvert and Svendenius, 2005; Svendenius and Gäfvert, 2005; Sven-
ably. An example shown here is the introduction of camber, which still is
denius and Gäfvert, 2006] and conference proceedings [Svendenius and
possible with reasonable complexity.
Gäfvert, 2004a; Svendenius and Gäfvert, 2004b; Gäfvert et al., 2006]. A de-
In the presented form, the brush model gives a good description of
tailed presentation of the theory behind the model can be found in [Gäfvert
the longitudinal tire characteristic at low slip up to the peak force. If
and Svendenius, 2003], while the first publication was in [Gäfvert, 2003].
good accuracy for higher slip is necessary the velocity dependence has
The steady-state part of the model, described in Sections 5.1–5.3, is
to be included in friction coefficient, which is discussed in Section 5.2
based on the observation that the pure-slip characteristics contains all
and more generally in e. g. [Svendenius and Wittenmark, 2003; Pacejka,
necessary information to reproduce combined-slip behavior with reason-
1988]. An extensive evaluation of the accuracy of the longitudinal proper-
able accuracy. The separation of the empirical pure-slip forces into com-
ties of the brush model can be found in Section 7.2. In the lateral direction
ponents of adhesion and sliding and the different treatment of respective
also the behavior of the carcass effects the tire characteristics. An accu-
contribution are fundamental parts of the method. The model is “semi-
rate modeling of the carcass require more sophisticated software, such as
empirical”, since the pure-slip models may be entirely empirical, while
Finite-Element tools, see Section 3.4.
the method to construct the combined-slip characteristics strictly follows
The most simple corrections available in the literature have been dis-
from theory on brush-model mechanics, explained in Section 4.1. This
cussed here. A new touch is given by the practical way of calculating the
distinguishes the presented model from similar previous methods that
lateral slip corresponding to the peak force in Equation (4.71). Most of
are purely empirical, for example [Bakker et al., 1989; Schuring et al.,
the contents in this section are gathered from earlier presented material
1996; Lugner and Mittermayr, 1991; Sharp, 2004]. Due to the physical
by other authors.
assumptions adopted, it is possible to include velocity dependency in the
model, even if this is not explicitly present in the pure-slip models.
A validation and a presentation of the results of the model are given
in Section 5.4 using measurement data from a Jeep Cherokee. The Magic
Formula is used to approximate the pure-slip data. An implementation
66 67
Chapter 5. A Dynamic Semi-Empirical Tire-Model . . . 5.1 Introduction
and a simulation of the dynamic model derived in Section 4.3 are de- 6. The resulting force magnitudes should stay within the friction con-
scribed, here. The dynamics behavior is explained by the deflection of the straints.
tire carcass, that causes differences between the forces and movements
7. The combined force should be F̄ = − Fz µ v̄s /vs at full sliding for tires
of the rim and the forces and movements in the contact patch. This is of
with isotropic friction characteristics, i.e. equal friction properties in
practical significance in, for example, systems for control of the longitudi-
all directions.
nal tire slip where motion sensors are, in general, attached to the wheel
rim and the force are generated by the motions in the contact patch. These criteria are used when developing the semi-empirical tire-model
for combined slip presented in the following.
68 69
Chapter 5. A Dynamic Semi-Empirical Tire-Model . . . 5.2 Scale Factors
70 71
Chapter 5. A Dynamic Semi-Empirical Tire-Model . . . 5.3 Parameters
and where
G mz = p sin(β ′ )p (5.20)
2 2 2 2 2 2
12
1 ○
v ( + σ ) + σ − v (σ + σ ) Λ σ
0
x y x y y
(4ψ − 1)(1 − ψ )2 p sin(β )p
sin ′
(5.21)
Gfz = a − p ( β )p
ψ (σ x , σ y, γ )(3 − 2ψ (σ x , σ y, γ )) ϒ σ ○y Λ z
J
Γy = vel (5.14b)
⋅ vel
if ψ (0, σ 0y , 0) < 1
vϒ(0, σ 0y , 0) and
Gcamz = 3aψ 2 (σ x , σ y, γ ) (1 − ψ (σ x , σ y, γ ))2 (5.22)
2 vel
ψ (σ x , σ y, γ )(3 − 2ψ (σ x , σ y, γ )) if ψ (0, σ 0y , 0) ≥ 1
and with
v
○2
sin cos
2 2 !
sin
2 γ ○2
u
J γ ( β ) γ u ( β ) γ ( β ) J
Λ z = ○2
Λ = ○2 +t 1− + γ −γ 2
γ −γ 2
σ ○y γ ○ σ ○x γ○ σ ○y
s 2 2 2
(5.15) sin ( β ) γ cos ( β ) sin ( β ) cos ( β )γ
⋅ + + − sgn(βσ y)
if ψ (σ x , σ y, γ ) < 1, otherwise σ ○y γ ○ σ ○x σ ○y σ ○xγ ○
(5.23)
( −1 vel
J
ψ (σ 0x , 0, 0)ϒ −1 (σ 0x
vel
, 0, 0) vel
if ψ (σ 0x , 0, 0) < 1
Γx = (5.16a)
vel
1 if ψ (σ 0x , 0, 0) ≥ 1 Note that the brush model states that the contact patch length can be de-
rived by the relation between the cornering stiffness and the self-aligning
and stiffness, see (4.42), as a = 3Cz/ Cy. When using the region-invariant slips
( −1 vel the friction angle is computed from
J
ψ (0, σ 0y , 0)ϒ −1 (0, σ 0y
vel
, 0) vel
if ψ (0, σ 0y , 0) < 1
Γy = (5.16b)
1 vel
if ψ (0, σ 0y , 0) ≥ 1 σ y F̂0x (σ xreg )
tan(β ′ ) = (5.24)
σ x F̂0y(σ yreg )
The friction-constraint angle β can be derived from
′
vel
σ y Γ x F̂0x (σ 0x )
tan(β ′ ) = vel
(5.17)
σ x Γ y F̂0y(σ 0y ) 5.3 Parameters
see (4.20). Four parameters are needed in the steady-state part of the model and two
if the transient properties are to be included. The parameters
Self-aligning torque
For simplicity, another pure-slip definition is used for the self-aligning • σ ○x , longitudinal limit slip
reg reg
torque. The pure slips (σ 0x , 0, 0) and (0, σ 0y , 0), with
• σ ○y , lateral limit slip
reg reg
σ 0x = σ ○xψ (σ x , σ y, γ ) sgn(σ x ) ; σ 0y = σ ○yψ (σ x , σ y, γ ) sgn(σ y ) (5.18) • γ ○ , adhesive camber limit
result in adhesion and sliding regions of the same size as the combined • v0 , velocity for the pure-slip model
reg
slip (σ x , σ y, γ ). It is therefore obvious to choose σ 0z = σ 0y and the self-
aligning torque is derived as • Ccx , longitudinal carcass stiffness
Mz (σ x , σ y, γ ) = G f z F0y(σ yreg ) + G mz M0z(σ yreg ) + Gcamz F0cam (γ ) (5.19) • Ccy, lateral carcass stiffness
72 73
Chapter 5. A Dynamic Semi-Empirical Tire-Model . . . 5.4 Validation and Results of the Semi-Empirical Model
scribe the pure slips where transition from partial to full sliding occur. 8
assumption is that these transitions occur when the tire forces obtain 2 4
their maximum. Some road foundations, such as gravel and snow might
−F [kN]
2
Fx [kN]
not provide a clear maximum point. In any case, the parameters may sim- 0
y
0
ply be set to the slip values corresponding to the maximum of the tire −2
−2
∗ and F ∗ . From the brush model it can be shown that
forces, F0x 0y −4
−4
−6
3 F̂0x
∗ 2 F̂0y
∗ ∗
F̂0y
σ ○x ( ; σ ○y ( + (5.25) −6
150 200 250 300 350 400 450 500 550
−8
−15 −10 −5 0 5 10 15
Ĉx Ĉx Ĉy Wheel rotational speed [RPM] Slip Angle [deg]
which is a result from (4.11) and (4.71). Figure 5.1 Available measurement data. Left: Longitudinal force for different
The parameter v0 denotes the wheel-travel velocity at which the empir- wheel rotational speeds; Right: Lateral force for sweeps of α .
ical pure-slip model is valid. The actual wheel-travel velocity v is assumed
to be a model input. If v0 is not known then v/v0 = 1 may be used, which
will neglect any velocity dependence. This is the common assumption in at combined slip at α = [±2, ±4] deg for sweeps of λ and forces at some
most other models. camber angles were also included. The different vertical loads for the mea-
The camber parameter, γ ○ , is outside the cambering range that is nor- surements of the pure longitudinal behavior are 1.1, 2.2, 3.3, 4.4, 5.3 or
mally measured on truck and car wheels. Therefore, it is not possible to ob- 5.5 kN. It is unclear from the data whether the highest load for sweeps
tain it directly from measurement data. Instead it can be used that (4.53) of λ and the only load for the combined slip is 5.3 kN or 5.5 kN. For the
is well approximated by pure lateral and cambering behavior 2.4, 4.2, 6.0, 7.8, 9.6 kN. The avail-
γ ○ ( F̂0y
∗
/ Ĉγ (5.26) able steady-state measurement data are shown in Figures 5.1, 5.2 and 5.3.
The camber stiffness can either be given from the empirical camber model Measurements of the transient properties of the tire are available as si-
or be computed from tire parameters according to (4.53) using (4.45) nusoidal excitations of the slip angle at nine different frequencies. The
and (4.42). data are shown in Figure 5.17 and further discussed in Section 5.4.
The carcass stiffnesses can be derived from (4.82) as
Validation of tire-forces at combined slip
Ĉx Ĉy The measurements of the pure lateral properties were performed at differ-
Ccx = ; Ccy = (5.27)
σ ax σ ay ent conditions than the longitudinal and combined properties. The vertical
loads do not correspond between the different setups. The longitudinal
if the relaxation lengths are known. This is further discussed in the last pure slip model and the combined validation data are taken at highest
part of Subsection 5.4 available load and the lateral pure slip model at 6.0 kN. A linear correc-
tion is applied to adjust for the different loads. The pure slip models are
Magic Formula parametrisation of the raw data, but for the torque model
5.4 Validation and Results of the Semi-Empirical Model interpolation of the tabular data was used. The empirical pure slip mod-
els are shown in Figure 5.4. In Figure 5.5 the forces and torque from the
Empirical data, provided by NHTSA, US, for P225/75R15 Goodyear Wran- model are compared with the measured values for sweeps of λ . In Figure
gler RT/S, OWL, all terrain steel belted radials [Salaani et al., 1999], are 5.6 the pure slip measurement for sweep of α is shown together with the
used for validation. The reference speed used was 48 km/h and inflation pure slip Magic Formula parametrisation and the computed characteris-
pressure 2.4 bar. The data consist of pure-slip forces and aligning mo- tics for combined slip. Note that (λ , α ) is used to signify the slip instead
ments in the slip ranges λ = [−50, 50]% and α = [−15, 15] deg. Forces of (σ x , σ y), since it is the most common slip definition when dealing with
74 75
Chapter 5. A Dynamic Semi-Empirical Tire-Model . . . 5.4 Validation and Results of the Semi-Empirical Model
10
−F [kN]
5
10
x
0
5
−5
F [kN]
0 5 10 15 20 25 30 35 40 45 50
0 Longitudinal Slip [%]
x
10
−5
−F [kN]
0
−10
y
200 250 300 350 400 450 500 550
−10
5 −20 −15 −10 −5 0 5 10 15 20
Lateral Slip [deg]
0.2
M [kNm]
F [kN]
0 0
y
z
−0.2
−20 −15 −10 −5 0 5 10 15 20
−5 Lateral Slip [deg]
200 250 300 350 400 450 500 550
Wheel rotational speed [RPM]
Figure 5.4 Pure-slip measurement data (dots) with Magic Formula parametrisa-
Figure 5.2 Available measurement data for combined slip. Longitudinal and lat- tion (solid lines) for the forces and interpolation of raw data for the aligning moment.
eral force at Fz=5.3 or 5.5 kN, see comment in text. The limit slips calculated from (5.25) are marked by crosses in the two upper plots.
measurement data and may look more familiar to the reader. Conversions
0.4
Camber Force for Fz=2.4, 4.2, 6.0, 7.8, 9.6 [kN]
between the two slips can be found in Appendix B. The overall behavior
of the model is very similar to the real measurements. Particularly good
0.3 agreement is shown at small and large slips. The largest deviations can
be found at slips around the force peak value. Since the tire behavior is
0.2
sensitive to many factors it is very important that the pure slip empir-
0.1 ical models are generated at the same conditions as the combined slip
Lateral Force [kN]
76 77
Chapter 5. A Dynamic Semi-Empirical Tire-Model . . . 5.4 Validation and Results of the Semi-Empirical Model
0
3
0 5 10 15 20 25 30 35
5 2
−F [kN]
0 1
y
−F [kN]
−5
0
y
0 5 10 15 20 25 30 35
0.2
−1
−M [kNm]
0
−2
z
−0.2
0 5 10 15 20 25 30 35 −3
Longitudinal Slip [%]
−4
Figure 5.5 Combined measurement data compared to the presented model (solid
−5
line) for sweeps of λ at different slip angles. −1 0 1 2 3 4 5 6
−F [kN]
x
Combined Slip Forces at λ=0, 2, 4, 10 [%] Figure 5.7 Combined-slip forces with the proposed model (solid line) compared
6
to the COMBINATOR model (dashed-dotted line) and the BPL model (dashed-line),
for fixed slip angles α and varying λ ranging from 0 to 35%. for comparison.
−F [kN]
4
x
The resulting force is always collinear with the slip vector, which is an
0
−30 −20 −10 0 10 20 30
5
assumption with weak physical motivation. In the COMBINATOR model
the lateral force initially increases, as a longitudinal slip is applied. This is
−F [kN]
0 structures may appear different. The BPL model, see (3.2), uses convex
z
78 79
Chapter 5. A Dynamic Semi-Empirical Tire-Model . . . 5.4 Validation and Results of the Semi-Empirical Model
. Velocity dependency for v=[0.5 1 1.5 2] ⋅ v0 Velocity dependency at combined slip for v=[0.5 1 1.5 2] ⋅ v
0
6 6
6
5 5
5
4 4
4
−Fx [kN]
−F [kN]
3 3
Fy [kN]
3
2 2
1 1 2
0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 1
λ [%] Lateral Slip α [deg]
Figure 5.8 Velocity dependence at pure slips using the models shown in Fig- 0
ure 5.4. Velocities: v = 0.5v0 , (dotted) v0 (solid), 1.5v0 (dashed), 2v0 (dash-dotted). −6 −4 −2 0
Fx [kN]
2 4 6
Figure 5.9 Velocity dependence at combined slips using the pure-slip models
fined in (5.12) used for the sliding contribution behaves more similar to shown in Figure 5.4. Sweeps of λ for α = −2, −4 [deg] and sweeps of α for λ = ±2, ±4
the region-invariant slips, when v = v0 . The resultant magnitude for the [%]. Velocities: v = 0.5v0 , (dotted) v0 (solid), 1.5v0 (dashed), 2v0 (dash-dotted).
BPL model at full sliding is
80 81
Chapter 5. A Dynamic Semi-Empirical Tire-Model . . . 5.4 Validation and Results of the Semi-Empirical Model
° °
Velocity dependency of extracted friction coefficient λ =[8.1 10.6 12.9] % Parameter dependency at combined slip for α =[6.69 8.38 10.1] deg
1.4 6
1.2
1 5
Long. Friction
0.8
0.6 4
0.4
Fy [kN]
0.2 3
0
0 2 4 6 8 10 12 14 16 18 20
Sliding velocity [km/h]
2
1
1.2
1
Lat. Friction
0.8 0
−6 −4 −2 0 2 4 6
Fx [kN]
0.6
0.4
Figure 5.11 Illustration of the parameter sensitivity of the force envelopes at
0.2 combined slip with sweeps of λ for α =-2, -4 [deg] and sweeps of α for λ = ±2, ±4.
0 The different limit slip, in increasing order, are denoted by dashed, solid and dash-
0 2 4 6 8 10 12 14 16 18 20
Sliding velocity [km/h] dotted lines, respectively.
Figure 5.10 Calculated friction coefficient for different choices of limit slip. Re-
spective limit slip, in increasing order, are denoted by dashed, solid and dash-dotted Formula with default shape factors has decent extrapolation properties
line. and may be used with slip ratios λ 0vel > 100%, but α 0vel is limited to 90
deg. In general, extrapolation of empirical data must be carried out with
great care.
recommendations for the choice of limit slip in this aspect. At actual speeds, v, lower than v0 the friction properties are extracted
If it is necessary to elaborate with the lateral limit slip to get proper from lower slip at v0 to the ensure that the sliding velocity is invariant
velocity dependence it is also important to study the effect on the combined (obvious from (5.12)). At low slip the sliding contributions are small and
slip behavior. Figure 5.11 shows the parameter sensitivity of the force they have to be magnified to match the proportion of sliding friction at v.
envelopes at combined slip for variation of α ○ . The figure shows that the Normally neglectable uncertainties of the model are then also magnified
resulting tire force gets larger for higher values on α ○ . A higher limit slip which might cause strange and non-accurate results.
increases the proportion of the adhesion contribution for lower slip than
this limit. The force limitation due to the friction constraints, therefore, Utilizing braking data to generate driving data In the proposed
starts to act later. This will give higher forces before the limit slip is model it is assumed that the given longitudinal pure-slip model, F̂0x (λ ),
reached. is valid both at braking (λ > 0) and driving (λ < 0). The Magic Formula
The expressions for pure slips can be solved for λ 0vel < 100% (σ 0x vel
< 1) is an odd function and therefore F̂0x (λ ) = − F̂0x (−λ ). Generally, this is
and α 0 < 90 deg (σ 0y < 1), which is when (v/v0 ) λ cos (α ) + sin2 (α )
vel vel 2 2 2
not in accordance with empirical observations. If the pure-slip model is
≤ 1 is fulfilled, see (5.12) or (B.20). This restricts the interval of v, λ , and assumed to be valid for braking, then the following procedure suggests
α . If λ 0vel or α 0vel are outside the valid range of the pure-slip model, then how to modify the argument to the pure-slip model at driving, so that a
extrapolation is necessary. This is actually when the model tries to extract more accurate force is obtained.
the friction properties from speeds not available in the empirical model. A The brush model states that the adhesion force developed at braking
straightforward method is to use the end points of the models. The Magic with a slip σ x > 0 will have the same size and the opposite sign, as the
82 83
Chapter 5. A Dynamic Semi-Empirical Tire-Model . . . 5.4 Validation and Results of the Semi-Empirical Model
−F [kN]
0
x
4
−5
−15 −10 −5 0 5 10 15
2
2
−Fy [kN]
F [kN]
0 1
x
0
−15 −10 −5 0 5 10 15
−2
0.08
0.06
M [kNm]
−4 0.04
0.02
z
0
−0.02
−6 −15 −10 −5 0 5 10 15
−15 −10 −5 0 5 10 15
Longitudinal Slip [%]
Longitudinal Slip [%]
Figure 5.12 Illustration of the difference between driving and braking. Solid line Figure 5.13 The effect of camber on the tire forces. The dashed dotted lines are
is the proposed method to convert between braking and driving, with sliding contri- the adhesion force and the dashed line the sliding forces. The camber effect when
bution as the dashed dotted line and the adhesive part as the dotted line. Dashed included as a shift on the lateral slip is shown by the dotted line for γ = 0, −5, −10,
line shows the Magic Formula tire-model, F0x (λ ) = − F0x (−λ ). and −15 deg.
are best used for driving and braking, respectively. For the sliding case it Results of cambering
is more natural to let the force depend on the relative velocity vs = λ vx . The information about the camber properties for the tire in the avail-
Then, −vs will simply correspond to −λ . able data was restricted to a few measurement points at different loads,
84 85
Chapter 5. A Dynamic Semi-Empirical Tire-Model . . . 5.4 Validation and Results of the Semi-Empirical Model
Tire Forces at γ=0, −5, −10 [deg] Tire Forces at γ=0, −5, −10 [deg] and α=0, 3, 6 [deg]
6
5
−Fy [kN]
5
0
−5 4
−15 −10 −5 0 5 10 15
−F [kN]
3
y
0.15
0.1
2
0.05
M [kNm]
0
z
−0.05 1
−0.1
−15 −10 −5 0 5 10 15 0
Lateral Slip [deg] −6 −4 −2 0 2 4 6
−Fx [kN]
Figure 5.14 Illustration of the effects of camber on the tire forces. Solid lines
are the total force and aligning torque. The dashed dotted lines are the camber Figure 5.15 Illustration of the effects of camber on the lateral force as a function
forces. For the torque the dashed dotted lines denote the additional torque from the of the longitudinal tire force at combined slip. The solid lines shows zero camber,
cambering forces. The camber effect when included as a shift on the lateral slip is dashed lines γ = −5 deg and dashed-dotted lines γ = −10 deg. The transitions to
shown by the dotted lines for γ = 0, −5, and −10 deg. full sliding are marked with asterisks.
camber is plotted in the figures and the results for pure slip cases are
see Figure 5.3. A validation of the effect of camber on combined slip was similar. For the combined-slip cases the results differ significantly, since
therefore, not possible. Instead the properties and results are illustrated the offset has an undesired influence on the sliding velocity and also on
in the following. The parameters of interest for cambering are Cy = 1.4 the sliding-force component. The self-aligning torque also differs between
kN/deg, Cz = 54 Nm/deg, and Cγ = 26 N/deg. It can be noticed that the two methods. Figure 5.15 shows the camber effect on the lateral force
the influence of camber on the lateral tire force is very low for the tire, as a function of the longitudinal tire force at combined slip. Note that
i.e. Cγ is small, and can be neglected for any practical case. Interestingly, camber has a strong influence on the point of transition to full sliding at
the camber stiffness calculated from the aligning stiffness, as described larger lateral slip.
in Section 4.2 by combining (4.53), (4.45) and (4.42), Ĉγ′ = 2kĈz , is sig-
nificantly larger, Ĉγ′ = 120 N/deg. To be able to visualize the effects of Discussion The difference between the theoretical camber stiffness, Ĉγ′
camber, the computed larger stiffness was chosen. In Figure 5.13 the ef- and the measured camber stiffness, Ĉγ for six different tires is shown in
fect of camber is shown for sweeps of λ and in Figure 5.14 for sweeps of Table 5.1. It is noticed that the deviation between the calculated and the
α. measured camber stiffness is smaller for tires for personal cars and motor-
It can be seen from the result that camber hardly affects the longitu- cycles than for tires for heavier vehicles as trucks or the Jeep, used in the
dinal force. It affects the partitioning between sliding and adhesion, but example above. A theory that is discussed in [Pacejka, 2002] is that the
the lower adhesion force is compensated by a higher sliding force. In other assumed bristle deformation due to the camber is most accurate for tires
tire models, for example presented in [Hirschberg et al., 2002] the camber with a rounded shoulder. For wider and flatter tires the camber effect
effect is included as a shift on the lateral slip, i.e. the lateral slip gets might be better explained by deformation in the tire walls and changes in
an additional term of Cγ γ / Cy . This much simpler approach to handle the the contact pressure distribution instead of bristle deflections and carcass
86 87
Chapter 5. A Dynamic Semi-Empirical Tire-Model . . . 5.4 Validation and Results of the Semi-Empirical Model
γ γ tot
Table 5.1 Comparison between measured and computed camber stiffness for dif- γ
ferent tires and loads. Note that â is calculated as 3 Ĉy / Ĉz for all tires except 185/60
R14, where it is measured. The calculation is an approximation which gives unre-
alistic values at higher vertical loads.
Fz [kN] 2.4 4.2 6.0 7.8 9.6 2.5 4.5 6.3 8.1
R [mm] 350 350 350 350 350 317 317 317 317
Ĉy [kN/deg] 0.75 1.2 1.4 1.4 1.4 0.7 0.8 0.9 1
Ĉz [Nm/deg] 16 33 54 75 94 20 30 50 70
â [mm] 64 83 116 161 201 86 112 167 210
Cy
Fy = α′ (5.36)
σ ay
deformations, see Figure 5.16. Through the difference between Ĉγ and Ĉγ′ s+1
it may be possible to calculate the stiffness of the carcass during camber, vx
since γ tot = γ + Ccγ Mx , Ccγ is the carcass stiffness in the actual direction
The parameters in (5.36) were identified from the measurement data as
and Mx the overturning torque. It would also hold that Ĉγ′ γ = Ĉγ γ tot for Cy = 1.4 kN/deg and σ ay = 0.57 m. In Figure 5.18 the identified transfer
small angles, which gives Ccγ = γ ( Ĉγ′ − Ĉγ )/ Mx . A further validation has function is compared with the data in a Bode-plot. The misfit of the phase
not been performed and the reasoning should be seen as a possible ex- was supposed to be the effect of a time delay between the collection of the
planation to the distinction between the measured and calculated camber slip-angle signal and the force signal. The fit is greatly improved with the
stiffnesses. addition of a delay of a half sampling period. The identified stiffness also
corresponds well to the static measurements in Figure 5.4 where Cy =1.3
The dynamic properties of the tire model kN/deg and Cx = 1.4 kN/%. Since no measurements are available in the
The available measurements from NHTSA included data for evaluation of longitudinal direction σ ax = 0.5 m is used. Relation (4.82) now gives the
the lateral dynamic properties at one velocity. The transient data is shown carcass stiffnesses as Ccy =2.3 kN/deg⋅m and Ccx =2.8 kN/%⋅m. Carcass
in Figure 5.17 and generated by a sinusoidal excitation of the slip angle. damping coefficients were chosen to D x = 4 s/%⋅m and D y = 7 s/deg⋅m.
The amplitude of the oscillations is small and the tire can be assumed The aligning stiffness was identified from Figure 5.4 to Cz =60 Nm/deg.
88 89
Chapter 5. A Dynamic Semi-Empirical Tire-Model . . . 5.4 Validation and Results of the Semi-Empirical Model
Dynamic measurements in lateral direction, v=48 km/h, Fz=6 kN Comparison between estimated system and measurements, v=48 km/h, Fz=6 kN
1 1
10
Amplitude
0
−0.5 −1
10
−1
0 10 20 30 40 50 60 70
−2
10 0 1 2 3
10 10 10 10
2
1 0
Fy [kN]
Phase (degrees)
0 −20
−1 −40
−2 −60
0 10 20 30 40 50 60 70
Time [s]
−80
Figure 5.17 Measurement of the force response for sinusoidal slip-angle inputs at 10
0
10
1 2
10 10
3
Figure 5.18 Bode plot showing the fit of a linear relaxation-length model (solid)
with the measurement data (stars). Addition of a time delay (10 ms) results in
The test velocity was given as v0 = 48 km/h. All results below is for v = v0 .
better fit for the phase (dashed).
Result for transient input The dynamic system (4.79) was simulated response for larger steps. The figure also illustrates the influence of com-
on the form bined slip on the dynamics. Figure 5.20 illustrates the dynamic cross-
coupling between λ and α at step inputs in λ ′ and α ′ . The cross-coupling
Fx Fx is larger and the dynamics are faster at larger slips.
" #" #
λ
− D x vx α λ̇
To visualize the effect of the dynamics of the tire in a more realis-
“ ”
Fy Fy
cos2 (α ) cos(α ′ ) cos2 (α ) − D yvx cos(α ′ ) α̇
λ α
" # tic setup the following results consider the dynamic of a whole wheel. In
−1
Cxσ ax v x (λ − λ ′ ) automotive applications the control of λ ′ using the brake torque is a del-
= (5.37)
−1
Cyσ ay vx cos(α ) sin(α − α ′ ) icate problem. Regard the wheel in Figure 4.15. Torque equilibrium can
be stated as
where the contributions from λ˙′ and α˙ ′ are neglected. Note the trigono- Jω̇ = T − Fx (λ ) R e (5.38)
metric re-writings to avoid numerical problems at slips angles of 90 deg. where R e is the wheel radius and T is the applied torque. Assuming
Partial derivatives of tire forces with respect to slips are required for sim- constant vehicle velocity vx the equilibrium can be rewritten as
ulation of the model. In general, these are not analytically available and
in the following they are computed by finite differences in each simulation J λ̇ ′ vx = ( Fx (λ ) R e − T ) R e (5.39)
step.
Figure 5.19 shows the response of the contact slip, λ , and longitudinal Note that the tire is not a rigid body and that the assumption of a fixed
force for rim-slip steps, λ ′ , of different magnitude at pure longitudinal inertia J is an approximation. The inertia for the tested wheel was not
slip and constant lateral rim slip α ′ = 5 deg. The results agree qualita- known and was set to J = 5 kg⋅m2 . It is also pointed out that the effec-
tively well with reports in the literature [Pacejka, 2002]. Note the quicker tive rolling radius R e might not be identical to the moment arm by which
90 91
Chapter 5. A Dynamic Semi-Empirical Tire-Model . . . 5.4 Validation and Results of the Semi-Empirical Model
14 6 6
Long.
12 Lat.
10 4 4
Slip [Percent]
−Fx [kN]
−Fy [kN]
8
6
2 2
4
α′=0 [deg]
2
α′=5 [deg]
0 0
0 0 0.05 0.1 0.15 0.2 0.25 0.3
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
8 8
6
4 4
3
2 2 2
1
0 0
0 0.05 0.1 0.15 0.2 0.25 0.3
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Time [s]
Distance [m]
Figure 5.21 Simulation of step responses in the brake torque (500, 1000, 1500
Figure 5.19 Illustration of the responses in λ and Fx for 10%, 20% and 30% steps Nm) while keeping the rim-slip angle constant at 2 deg. Upper plot: contact forces;
in λ ′ at α ′ = 0 deg and α ′ = 10 deg. Lower plot: contact slips.
6 6
the tire force applies a moment on the wheel. However, the difference
Long. Slip [Percent]
3
dynamics at a stationary operating point with constant brake torque is
2
92 93
Chapter 5. A Dynamic Semi-Empirical Tire-Model . . . 5.5 Implementation and Application in MBS Systems
60 and for ϕ → 0 or vx → ∞ to
−∞
Ccx
40
s s+ (5.42)
Dx
φ
Future work will address quantitative validation with measured tran-
20 sient data. A problem in this context is the possibilities to separate the
φ φ contribution of dynamic behavior from the wheel and tire from deflections,
0
0 back-lash, and hysteresis of the wheel suspension in a real application.
Section 6.2 treats these questions in a more practically way.
Im
−60
−400 −350 −300 −250 −200 −150 −100 −50 0 50
There are many different software environments available for modeling
Re and simulation of multi-body mechanics, both for general mechanics and
applications specific for vehicle dynamics. The VehicleDynamics LibraryTM -
Figure 5.22 Root locus of the characteristic polynomial (5.40). [Modelon, 2006] for DymolaTM [Dynasim, 2006] is chosen, due to the close
connection to the developers, to extend the usability and the possibility
of validation of the model, see [Gäfvert et al., 2006]. Dymola is a model-
sponds to the braking stiffness at small slip the poles are complex and ing and simulation environment for general multi-domain modeling and
poorly damped. When ϕ increases the poles become real. This explains simulation. It is based on the open high-level ModelicaF R language that
the increased damping at larger brake torques. As ϕ approaches zero one is maintained by the Modelica Association [Mod, 2005]. The VehicleDy-
pole goes to a stable limit and the other approaches zero. The introduc- namics Library is an add-on to Dymola that offers complete and detailed
tion of the damping D x lets ϕ grow to the positive value specified by the multi-body vehicle models for vehicle dynamics analysis.
conditions in Section 4.3 while maintaining stability for the carcass dy- The multi-body model describes the vehicle as a set of rigid bodies con-
namics. The wheel dynamics, however, becomes unstable as the slope of nected by joints (constraints) or force elements. The description is based
the force-slip curve gets positive. The root locus clearly shows that wheel on orthonormal reference frames that are attached to the bodies. The ref-
and tire dynamics can not be separated and that any control application erence frames follows the ISO 8855 standard with x pointing forward, y
that involves the wheel rotation benefits from taking the tire dynamics to the left, and z upwards.
into account. It is well known that the vehicle velocity vx greatly affects The wheels are treated as bodies with force elements that describe
the wheel dynamics and that slip control is harder at low velocities. In- the interactions between the chassis and the track surface. The rim of
spection of the characteristic polynomial at limit conditions reveals the each wheel is rigidly connected to the suspension wheel carrier by the
interesting fact that the dynamics change equally for variations in the hub, and may be kinematically defined by the chassis and wheel-spin
slip operating point and the vehicle velocity. For ϕ → −∞ or vx → 0 the states. The tire is a vertically deformable body that is attached to the
characteristic polynomial goes to rim and interacts with the ground in a contact patch of varying size and
location. In principle, the contact patch is represented by a contact point
at which the forces and moments are assumed to apply, located where a
R2e D x R2 Ccx line through the wheel center, aligned with the wheel plane, intersects
s2 + s+ e (5.41) the track surface. Moreover, the deformation of the tire is normally only
J J
94 95
Chapter 5. A Dynamic Semi-Empirical Tire-Model . . . 5.6 Conclusions
Figure 5.23 To the left: Simulated test rig for analyzing the transient and sta-
tionary behaviour of tyre models. The test rig can apply longitudinal and lateral
slips as well as camber at desired loads and velocities. To the right: Visualization
of simulated NHTSA-Fishhook handling maneuver.
96 97
Chapter 5. A Dynamic Semi-Empirical Tire-Model . . . 5.6 Conclusions
based on understanding of the physical sources of tire forces, as given BPL model [Bakker et al., 1989] performs similarly. The COMBINATOR
by the theoretical rigid-carcass brush-model. Based on the brush model, model [Schuring et al., 1996] behaves slightly differently at small com-
the combined-slip forces may be described by a scaling of corresponding bined slips. None of these previous models include velocity dependence.
empirical pure-slip forces. There is a freedom in choosing the pure-slip The tire model is successfully implemented in Matlab code and in the
forces, which may be used to emphasize different physical effects. This multi body simulation environment, Dymola. It works well and is compu-
made it easy to include velocity dependence in the model, in a unique tationally sound on both platforms.
way. The generation of the camber force has been described in a simple, The proposed dynamic extension of the proposed combined-slip model
but physically motivated manner. Its influence on the partitioning of the exhibits a qualitatively realistic tire relaxation behaviour. The model in-
sliding and adhesive regions in the tire contact patch, affects the scale cludes cross-couplings between longitudinal and lateral dynamics, that
factors for the tire-forces. For pure slips it is common to introduce camber are not commonly found in other models of similar complexity. The issue
by an offset on the lateral slip. However, for the self-aligning torque and of stability is handled by including a physically motivated damping. Ex-
the combined slip forces the proposed model gives significantly different amination of the combined dynamics of a wheel and the tire model gives
result. It has also been discussed that the effects of camber, with varying insights that might be relevant for brake-control applications.
accuracy, can be approximated from the lateral properties of the tire. The
model is simple to use since it does not introduce any new parameters.
All necessary information is given by the used pure-slip model.
Even though the model in its present state is useful for a number of
applications, the addition of effects of turn-slip and flexible carcass would
further enhance the applicability. The physical foundation of the model is
expected to make this possible with reasonable effort.
Relations to similar, previously published combined-slip models are
analyzed. Previous models are partly heuristically based, while the pro-
posed model is entirely based on physical principles. An advantage with
the presented model is that it combines the data-fitting accuracy of em-
pirical pure slip models with the stringency of physical modeling. The
calibration of physically-based models to measurements is often limited
by the small set of parameters and the approximations and idealized or
neglected effects in such models. Therefore, more flexible heuristic mod-
els such as the Magic Formula [Pacejka, 2002] are more commonly used
nowadays both for pure and combined slip situations. Despite good fit at
calibration with measurement data, deviations from real behaviour will
still arise as the normal operating conditions for a tire usually differ sig-
nificantly from those at calibration-measurement experiments. Empirical
models are normally valid only within the input intervals where calibra-
tion measurements exists and extrapolation properties are often very un-
reliable. Moreover, the large amount of factors that affects the tire behav-
ior makes it practically impossible to collect and manage data for all kind
of influencing factors. These observations raise the question whether it is
better to use physically based models with less perfect fit to measurement
data, but with better extrapolation properties outside the measurements
ranges.
Good results are obtained in validation with the available empirical
data. Experimental validation indicates that the proposed model and the
98 99
6.1 Introduction
100 101
Chapter 6. Measurement and Processing of the Wheel Speed Signal 6.2 Effect of Tire Force Transients on the Wheel Speed Measurements
A simple model of the axle behavior and a proposal on how the influence of
the transient brake force on the measured wheel speed can be accounted
for is described in the following section. The tire dynamics is previously
Figure 6.1 Picture of a rear truck-axle showing axle, rim, brake unit, tooth wheel treated in Section 2.6 and 4.3. Section 6.3 reviews and describes some
(TW) and the placement of the speed sensor (SP) (Modified, by courtesy of Haldex methods to eliminate the rotation dependent noise from the speed signal.
Brake Products AB). Processing the noise from road irregularities is a large area for further
work. In, for example, [Gustafsson et al., 2006a] the dependency between
the front and the rear wheel speed covariances are used for determination
with great care, not to delay the detection of a beginning wheel-lock situ- of the absolute velocity of the car. This field in not further treated in this
ation. In some conditions, particularly at low speed, the wheel might lock thesis. Additional noise sources are, e. g. mentioned and discussed in
quickly and any delay on the signal might be hazardous for the stabiliza- [Schwarz, 1999]
tion system. However, the signal processing has to be efficient enough not
to unnecessarily invoke e. g. the anti-lock system.
The sensors used nowadays often measure accurately, but it is impor-
tant to regard that they measure the magnetic field due to the passage of 6.2 Effect of Tire Force Transients on the Wheel Speed
a tooth in front of the sensor pick-up. This do not always coincide with the Measurements
tire speed that is useful for the vehicle control system. There are three
main noise sources, among others, that affect the measurement of the This section will mainly discuss the disturbance on the wheel speed mea-
wheel speed: surements caused by movements of the sensor position. The effect is par-
ticularly obvious for front axles, since these, generally, are relatively weak
• Brake force transients There is a dynamic response from the and flexible mechanical constructions. The brake that creates and carries
mechanical system around the wheel axle when a braking torque the reacting braking torque is mounted on the axle end, which deflects
is applied onto the wheel. Two different physical effects are then when supporting the load. Forces acting on the tire due to, for example,
present. First, the movement of the wheel-axle end, where the brake the engine torque are also subjected to a dynamic behavior, but do not, in
and wheel speed sensor is mounted, when the axle is subjected to the same way, affect the wheel speed measurement. Rear axles are often
a reaction torque from the brake affects the wheel speed measure- of more robust design and deflects less, while the dynamic behavior of the
ment. Since the relative speed between the wheel and the sensor is tire still is present.
102 103
Chapter 6. Measurement and Processing of the Wheel Speed Signal 6.2 Effect of Tire Force Transients on the Wheel Speed Measurements
30 ϕa
25
48.4 48.6 48.8 49 49.2 49.4
10
MB
5
0
48.4 48.6 48.8 49 49.2 49.4
5
2
0
ω
−5
48.4 48.6 48.8 49 49.2 49.4
0.2
0.1
Slip
0
−0.1
48.4 48.6 48.8 49 49.2 49.4
Time [s] Figure 6.3 Schematic figure of the axle and wheel hub. The axle rotation are
here supposed to carry the effects from all kind of deflections of the suspension and
Figure 6.2 Plot showing the effect of a front-axle brake-force transient on the tire. The movement of the sensor causes the disturbances on the wheel rim speed
wheel-speed measurement. A step command on the brake pedal causes an apparent measurements.
decrease of the wheel speed. The ABS reacts and lowers the brake force. After an
overshoot the wheel speed returns to a stable level and the brake system can come
back to normal mode. The rear axle is totally unbraked. The solid lines denotes the Dynamic model
front left, dashed lines the front right, dashed dotted the rear left and dotted line
describes the rear right wheel. The vehicle retardation (third plot) are derived from
The reaction of the wheel speed measurements, due to transients in the
accelerometer measurements. brake forces, is explained by many different effects. One part is the move-
ment of the wheel hub and wheel speed sensor due to deflection of the
axle and suspension. Another part is the dynamics of the tire, caused by
the carcass flexibility and the inertia of the wheel, discussed in detail in
Sections 4.3 and 5.4.
An example of the effect can be viewed in Figure 6.2, that shows mea- In the following it is assumed that the movement of the sensor relative
surements collected from a Scania test truck at Haldex. The transient the tooth wheel is caused by axle torsion, see Figure 6.3. In reality, the
of the brake torque, in the interval t=48.4–48.6 s, causes a dynamic be- movements of the axle are much more complex, see e. g. [Matschinsky,
havior of the wheel speed signal, which is clearly seen from the top and 1997], and the sensor together with the wheel axle will additionally exhibit
bottom subplots. The overshoot in slip during the application phase largely linear motions. The deflections are here lumped together in one equation,
exceeds the settled value at the end of the data series even though the since the major part of the respective dynamics can be described as a
brake torque is as largest at the end. mechanical system, characterized as
In [Schwarz et al., 1997b] the effect is modelled as longitudinal suspen-
sion stiffness and accounted for in the slip determination. In [Sugai et al., 1
ϕa = MB (6.1)
2003] a method is presented to on-line estimate a coefficient denoting the J a s2 + D a s + K a
entire influence of the longitudinal suspension motion and tire carcass
stiffness. Here, the behavior will be modelled and discussed. Based on the It is therefore emphasised that, the inertia, Ja , the damping, Da , and,
available measurements, the requirements on a model that predicts the the stiffness, K a are generic parameters gathered from the entire system.
tire-wheel-axle dynamics are studied in an empirical way. They do not directly correspond to the physical, measurable quantities of
104 105
Chapter 6. Measurement and Processing of the Wheel Speed Signal 6.2 Effect of Tire Force Transients on the Wheel Speed Measurements
the axle, since many parallel systems are incorporated in this equation. 34
The deflection angle, ϕ a , denotes the movements of the wheel speed sensor
and the brake torque is denoted by MB . The actual speed of the rim ω
33
can be calculated from
ω = ω̂ − ϕ̇ a (6.2)
where ω̂ is the measured wheel speed. For the wheel speed measurement 32
Time [s]
sensor and the wheel axle together gets a linear deflection. In the first case 31
the speed measurement is actually wrong, since the sensor moves with
the wheel. In the other case the speed is correct, but the wheel reference
speed, ω actually deviates from the vehicle reference. It is assumed that 30
lumping of these motions into one rotational motion is a good engineering
approximation.
29
Inserting (6.1) in (6.2) gives
s
ω = ω̂ − MB (6.3) 28
J a s2 + D a s + K a 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Included in ω there is another dynamic response of the force transient due Figure 6.4 Simulated wheel speed response (solid line) compared to measure-
ments (dashed line) for the scenario shown in Figure 6.2.
to the tire deformation. The response is velocity dependent and should be
considered when the speed of the tire carcass in the contact with the road
is to be derived, which normally is the case when modeling tire forces identified model dynamics is assumed to behave as any first or second-
from the slip. The relation between the wheel rim speed and tire carcass order transfer function, G (s). The signals necessary for the identification
speed is in detail discussed in Section 5.4. In Figure 6.4 the result from are derived from Equation (6.3), rewritten as
a simulation of the axle, wheel, and tire including these effects is shown.
For the axle simulation Equation (6.1) is used and the tire and wheel is ω − ω̂ = G (s) MB (6.4)
modelled according to (5.37) and (5.38). The brake torque sequence visi-
ble in Figure 6.2 is used as input signal to the system. It can be seen that The real wheel velocity ω can not be measured. Here, it is approximated
major part of the dynamics is caught by the model. Proper parameters for from wheel speed measurements of the unaffected rear axle, but compen-
the system is, however, difficult to find since the tire and axle parameters sated for the slip (due to the applied brake force), such that there is no
are not entirely known. The unknown parameters used in the simulation stationary gain in the identified system. The slip is caused by the tire be-
example are more or less manually chosen to give acceptable result and havior and should not be incorporated in the axle behavior. In that way the
more effort can be done to further increase the accuracy. Tire characteris- transient effects of the torque application becomes clear, see Figure 6.5.
tics given by the manufacturer is used, even though the reliability is low The velocity dependency of the dynamics caused by the tire carcass
since the conditions at the testing facilities certainly, are different from is then completely neglected, since it is incorporated in the velocity inde-
those at the measurement occasion. pendent system for identification. In practise it is difficult to distinguish
between the different effects. Tests at different speed are required for a
Identification of the dynamics
further evaluation.
The System Identification toolbox in Matlab is employed to increase the Using the Matlab command arx, the discrete equivalence H ( q) to G (s)
agreement of the model. The aim is to find an accurate relation between is identified on the form
the brake torque and the wheel speed during transients. Finding a sim-
ple transfer function with good accuracy is of more interest than finding b0 + b1 q−1 + b2 q−2
the optimal parameters to the physical model from the prior section. The H ( q) = (6.5)
a0 + a1 q−1 + a2 q−2
106 107
Chapter 6. Measurement and Processing of the Wheel Speed Signal 6.2 Effect of Tire Force Transients on the Wheel Speed Measurements
8 8
35 Input (Br Tq) Input (Br Tq)
Output Output
1st ord. sys 6 1st ord. sys
6
33 2
2
0
Wheel Speed [kph]
32 0
−2
−2
31 −4
−4 −6
48.4 48.45 48.5 48.55 48.6 48.65 48.4 48.45 48.5 48.55 48.6 48.65
30 Time [s] Time [s]
29 Figure 6.6 Comparison between the output from the identified system and the
measurements ω −ω̂ for left (left plot) and right (right plot) front wheels. The dotted
line denotes the result from the full-model simulation described in Section 6.2
28
34 39
Wheel Speed Wheel Speed
27 1st ord. corr
38
1st ord. corr
0 0.05 0.1 0.15 0.2 0.25 0.3 33
2nd ord. corr 2nd ord. corr
Time [s] 37
32 36
30 33
32
The measured data set contains two occasions where the effects of the 29
31
torque transients are clearly visible. The first is shown in Figure 6.2 from 28 30
where the time slot [48.4, 48.65] is used for identification. The second
48.3 48.4 48.5 48.6 48.7 48.8 137 137.1 137.2 137.3 137.4 137.5
Time [s] Time [s]
occasion is used only for validation. The identified parameter values are
Figure 6.7 Comparison between the measured wheel velocity and the corrections
presented in Table 6.1 and the obtained dynamics are graphically shown
applied with the two different filters according to (6.6). To the left: The first brake
in Figure 6.6. The result is almost similar for the first and second order application where the disturbance is seen, which also is used for the identification.
system and no accuracy is gained by higher order than one. The right To right: The second brake application for validation. Data from the left-front wheel.
wheel shows better accuracy than the left-side from where there seems to
be a timing error of approximately 0.01 s between the simulation result
and the measurements. Reasons for this are discussed later on. removed. In the other plots the disturbances still are present, but reduced
The result from the identification is used to reduce the disturbance of by at least 50% and the first order system performs slightly better than
the brake force application on the wheel speed. The corrected rim speed the second order system.
is then calculated as
Conclusion The aim of this section has been to find a way to com-
ω (t) = ω̂ (t) − H ( q) MB (t) (6.6) pensate for disturbances on the wheel speed signal created by axle and
suspension deflections during the brake application. From the validation
In Figures 6.7 and 6.8 the correction is validated towards the measure- it can be seen that the disturbance is not entirely eliminated, but its
ment data for the two available test occasions. It is clear that the dis- amplitude is greatly reduced by very simple means. The correction on the
turbance on the wheel speed is greatly reduced. The best result is seen wheel speed signal can be seen as a feed-forward filter of the brake torque,
in the first transient on the right side, where the disturbance is almost predicting the disturbance. A first order filter is sufficient and higher or-
108 109
Chapter 6. Measurement and Processing of the Wheel Speed Signal 6.3 Method for Elimination of Rotation Dependent Noise
35
Wheel Speed
39
Wheel Speed axles have less impact on the wheel speed and the dynamics then might
1st ord. corr
be completely disregarded. In practise, it is difficult to distinguish the tire
1st ord. corr
34 38
2nd ord. corr 2nd ord. corr
33
37
dynamics from the axle dynamics. A result from Section 5.4 is that the im-
36
pact of the carcass deflections are less obvious at higher slip. Comparing
Wheel Speed [kph]
31
35 the dynamics at high and low slip might be an accessible way to separate
34 between wheel and axle dynamics. Performing experiment at different ve-
hicle velocities is another. The method of approximating the suspension
30
33
29
32 and tire stiffnesses in one coefficient and identify it on-line as described
28 31 in [Sugai et al., 2003] seems to be an realistic approach, since the major
27
48.3 48.4 48.5 48.6 48.7 48.8
30
137 137.1 137.2 137.3 137.4 137.5
part of the dynamics shows to be described by this coefficient. A problem
Time [s] Time [s]
with on-line estimation in this case may be that quick and substantial
excitations necessary for the estimation might occur rarely during normal
Figure 6.8 Data from the right front wheel, compare to Figure 6.7.
run. When they suddenly occurs it might be in a limited condition and
optimal performance of the system is a demand.
Table 6.1 Result of the parameter identification of (6.5).
On the other hand there are many ways to further improve the param-
eter identification. The system is very fast and the sample time of 0.01 s,
Side a0 a1 a2 b0 b1 b2 used in the collected measurements, is quite slow compared to the reac-
tion of the wheel speed. It is therefore assumed that shorter sampling time
Left 1 0.283 0 -0.00249 0.00255 0 would be advantageous for the estimation. For such fast system dynamics
Right 1 0.411 0 -0.00301 0.00299 0 delays and jitter on the signals caused by the CAN-bus transmission and
Left 1 -0.131 -0.217 -0.00293 0.00497 -0.00206 different sensor characteristics can be crucial. In Figure 6.6 there is a
Right 1 -0.15 -0.348 -0.00319 0.00531 -0.00213 delay of 0.01 s between the measurements and the identified systems for
the left side, while the right axle shows better agreement. The setup is
identical, but with the hard timing constraints the error may be explained
by varying latency on the signals.
der filters do not improve the results. A remarkable result is that poles
of the identified discrete system are on the negative real axis. For the
second order system one pole stays on the negative and the other on the 6.3 Method for Elimination of Rotation Dependent Noise
positive stable axis. This makes it impossible do a physical interpretation
of the systems. Comparing the pole to the slightly unstable zero in the There is, in general, noise on the wheel speed signal that is repeated ev-
first order system shows that behavior due to the pole is overridden by ery revolution, with a scheme that is depending on the rotation angle of
the effect of the zero. The behavior can almost be seen as a pure differ- the wheel. This type of noise might be hard to discern since its pattern
entiation, which also will be the result if the the denominator is set a is distorted after sampling and filtering. To clearly see this kind of dis-
constant while re-doing the identification. turbance, the signal has to be unfiltered and should be sampled at each
The disturbances are reduced by at least 50% in the available valida- passage of a tooth flank or faster. This is not commonly done even on test
tion data and this is probably sufficient to increase the performance of vehicles. The sources for the disturbances come, for example, from irregu-
vehicle systems relying on the wheel-speed signal during the brake-force larities of the teeth, varying air-gap between sensor, and a distorted tooth
transients. A drawback, so far, is the limited amount of measurement wheel, and uneven worn or oval tires. To be able to eliminate such a noise
data. More test occasions are necessary for a more comprehensive valida- source, the position of the rim has to be known during the wheel rotation.
tion. Tests on different types of axles are another necessary requirement. It is not necessary to determine the absolute angle, if there are sufficient
A question mark is if the neglection of the tire characteristics still is ac- time for calibration of the algorithm after each restart of the system. The
ceptable in the identification of a stiffer axle or at a lower vehicle velocity. important thing is to be able to compare a pulse signal at a certain po-
Since the relative contribution of the tire then is larger. Also, the stiffer sition to the pulses from the same position at the prior revolutions. The
110 111
Chapter 6. Measurement and Processing of the Wheel Speed Signal 6.3 Method for Elimination of Rotation Dependent Noise
R eθ i−1 R eθ i R eθ i+1
∆ t k−1 ∆ t k ∆ t k+1
Figure 6.10 Illustration of the pulses from the wheel speed sensor relative the
rotation of the wheel. The space between the pulses relates to the distance travelled
Figure 6.9 Illustration of a part of a tooth wheel, with angle θ between the teeth,
by the wheel, when it is not subjected to a slip. Any irregularities in the tooth wheel
ideally.
and the dynamic tire radius are accounted for in δ θ . Therefore, R e might be treated
as constant.
general idea for the noise elimination is to derive a vector containing one
element per pulse generated by the sensor during one revolution of the varying system on state-space form as
tooth wheel. The vector values are used for correction of the actual wheel
speed signal and calibrated in such a way that the signal irregularities x(tk+1 ) = Φ x(tk ) + e p (6.9)
are cancelled. Similar methods have been presented in [Schwarz et al., y(tk ) = C (ik , tk ) x(tk ) + em (6.10)
1997a] and [Persson, 2002].
where the state vector x is
The measurement system
Let N denote the number of pulses from the sensor during one revolution.
T
1
Each pulse is indexed by k. Hence, tk is the time at pulse k, which is x( t k ) = δ θ1 δ θ2 .. δ θ N (6.11)
ω (tk )
generated by the tooth flank ik (when i is used as index, k will be left
out). The rotation angle of the rim between each pulse is ideally θ = 2π / N, The system matrix, Φ is the identity matrix with N + 1 rows and columns.
see Figure 6.9. Due to the circumstances mentioned above the pulses are The mean velocity during the time between tk−1 and tk is denoted by ω (tk ).
not evenly distributed along the travelled distance for one revolution of The process error, e p relates to the assumed variation of the inverted
the wheel. The angle corresponding to the rolled distance between two wheel speed and the irregularities of the teeth. The irregularities are
pulses at the tooth i, see Figure 6.10, is assumed not to change, hence
where δ θ i is a measure for the distortions in the tooth wheel and uneven The observation y(tk ) = [∆ tk ] = tk − tk−1 is the time elapsed between the
rolling radius around the peripheral of the tire and belongs to the vector two latest pulses. The observation vector becomes
δθ,
1
N
C (i k , tk ) = θ 0 .. 0 0 .. 0 (6.13)
X ω (t k )
δθ = [ δ θ 1 δ θ 2 .. δ θ N ], δθi = 0 (6.8)
i=1 where the factor 1/ω (tk ) is placed corresponding to the element δ θ i , i. e.
The measurement of the rotational velocity can be expressed as a time on instance ik + 1. The measurement error, em is assumed to be white noise
112 113
Chapter 6. Measurement and Processing of the Wheel Speed Signal 6.3 Method for Elimination of Rotation Dependent Noise
due to road irregularities and disturbances from the signal conversion in where rω is the variance relating to the change rate of the wheel speed,
the sensor. Hence, the observation of the inverse velocity is or more correctly, the inverse of the wheel speed. The variance of non-
repeating measurement errors due to, i.e road irregularities are incor-
θ δθi porated in r t . It should, however, be a balance between r t and rω such
y(tk ) = ∆ tk = + + e(tk ) (6.14)
ω (tk ) ω (tk ) that quick wheel speed changes not are misinterpreted as temporal dis-
turbances. To be on the safe side, then r t are set very low. The covariance
From the observations of ∆ t the aim is to identify the correct velocity, or of the irregularities, rθ , might not be known at the beginning, but after a
in this case, its inverse. By estimation of the irregularities of the teeth, successful state estimation, it can be computed as cov(δθ). The vector δθ
their influence on the noise on the speed can be eliminated. does not change during normal run. However, there are occasions where
the counter can miss a tooth. Therefore, both at start up and when a
Wheel speed filtering by use of a minimal-variance technique missed tooth is suspected the filter has to be reinitialized. Earlier com-
If the disturbances on the system are Gaussian and respective covariances putations of cov(δθ) can then be used for rω . An alternative is to set the
are known sufficiently accurately, than a Kalman filter is an optimal way diagonal elements in the process covariance matrix, R1 in the same way
to estimate the state vector even for time-variant system [Åström, 1970]. as those of R0 and in that way introduce a forgetting factor in the estima-
The recursive Kalman filter and the state estimation scheme is formulated tions. A drawback is that a trade off problem is encountered in choosing
as between a quick re-adaptation and minimum variance on the estimates.
Different ways to supervise the algorithm, using different update modes,
x̂(tk+1 ) = x̂(tk ) + K ( y( k) − C (ik , tk ) x̂(tk )) (6.15) can be more effective and give and better output in this case.
T T
K (tk ) = P(tk ) C (ik , tk )( R2 + C (ik , tk ) P(tk ) C (ik , tk )) −1
(6.16) As an illustration of the method, a tooth wheel and a sensor that
outputs 10 pulses per revolution are simulated at constant speed. A vector
P(tk+1 ) = P(tk ) + R1
of tooth asymmetries is created, but there are no other disturbances on the
−( P(tk) C (ik , tk )T )( R2 + C (ik , tk ) P(tk ) C (ik , tk )T )−1 ( C (ik , tk ) P(tk )) signals, so that the function of the state estimator can be easily illustrated.
(6.17) Signal disturbances would increase the convergence time. The difference
P(0) = R0 (6.18) of the measured output (time between the pulses, ∆ tk ) and the corrected
output (1/ω (tk )) is shown in Figure 6.11 using the proposed method. It
The condition of having the sum of δθ equal to zero is not yet included in can clearly be seen that the variance of the corrected signal decreases
the formulations. while the values in δθ-vector adjust. The result of the δθ estimation is
P To fulfill this specification, the observer, y, is extended
to also observe (δθ). Hence, shown in Figure 6.12. Note, that all values in the vector changes at each
update such that the sum of the entire vector is zero.
1 The presented observer is regarded as an optimal observer, predict-
∆ tk ) 0 .. 0 0 .. 0 et
θ ing the states with minimal variance relating the to available measure-
yE (tk ) = ; C E (i, k) = ω (tk ) ; em =
0 0 ments. Setting r t low, the major tuning parameter is rω . It is important to
0 1 .. 1 1 1 .. 1
(6.19) have a sufficiently high value on rω such that the velocity estimate reacts
which means that an error-less observation that the sum is zero is per- fast and is calculated with minimum delay. If the actual wheel velocity
formed at each instant and the estimated states will adapt to this condi- changes rapidly, for example, in a wheel-lock situation and the estimated
tion. The covariance matrices can be chosen as velocity can not adapt quick enough, there will be a disturbance on the
δθ-estimates. On the other hand, if the velocity reacts to fast the δθ-
rω 0 ⋅ ⋅ ⋅ 0 values will not be calculated correctly, since the tooth-irregularities can
0 r ⋅⋅⋅ 0
θ
rω 0(1, N )
rt 0
not be separated from velocity changes. In the method illustrated above
; ; the variances are set as rθ = 0.04, rω = 0.0005, and r t = 0. In a realistic
R0 =
.. .. . . . R1 = R2 =
. ..
. .
0( N, 1) 0( N, N ) 0 0 approach the variances much be set more careful, concerning the temporal
0 0 0 rθ disturbances on the wheel speed signal and the desired ability of the al-
(6.20) gorithm to catch velocity changes of the wheel. A separate algorithm may
114 115
Chapter 6. Measurement and Processing of the Wheel Speed Signal 6.3 Method for Elimination of Rotation Dependent Noise
Figure 6.11 Results from minimum variance estimation of simulated tooth irreg- Figure 6.12 Estimated normalized angle errors compared to the true error
ularities. The upper plot shows the “measured” pulse time and the lower plot shows (dashed-dotted lines)
the pulse time corrected for the estimated δθ. The algorithm starts at 0.5 s.
116 117
Chapter 6. Measurement and Processing of the Wheel Speed Signal 6.4 Conclusion
for indicating the tire pressure. Therefore the frequency spectra for the 6.4 Conclusion
wheel speed signal is derived during the travel of the vehicle. It has been
showed that the tooth irregularities creates disturbing frequency peaks in The functionality of many vehicle system relies on the wheel speed signals.
the spectrum plot at the frequencies f K (m) = mω k /(2π ) and the proposed In some important applications traditional noise filtering is difficult since
method do only attempt to suppress the harmonics in the frequency win- the timing constraints on the signal are high. This section has discussed
dow of interest. A Fourier series is employed to describe the disturbance a few methods to improve the quality of the wheel speed signal, without
pattern around the tooth wheel peripheral causing any phase shift or time delay.
On, particularly, front axle wheels force transients caused by quick
N /2 −1)/2
( NX changes of the brake command can give large disturbances on the wheel
2π 2π speed signal. The disturbances arise due to wheel axle deflection and a
X
δ θ k = a0 + am cos m k + bm sin m k (6.23)
N N corresponding movement of the speed sensor position, since the measured
m=1 m=1
speed is the relative speed between the wheel and the sensor. The result
shows that the effect of such a disturbance can by simple means be reduced
and the included parameters a and b are estimated from the recursive by over 50% by appling a first-order feed forward filter from the brake
scheme torque signal.
Further, the wheel speed signal has a rotation-dependent noise, due to,
tolerance errors on the tooth wheel that together with a sensor generate
θ = [ am bm ]T (6.24)
the pulses to the frequency counter that calculates the wheel speed. The
1 2π 2π
ϕk = cos m k sin m k (6.25) noise pattern can be removed by creating a vector that adaptively keeps
∆ tk N N track on the error related to each tooth. The expectation is that noise of
ε k = yk − ϕ kT θ k (6.26) 1.5% on the speed signal could be reduced to around 0.5% for free rolling
θ k+1 = θ k + µ kϕ k ε k (6.27) on smooth asphalt roads.
Conclusive remarks
Measurements performed on cars have shown a rotational dependent noise
level up to 1.5% percent of the wheel speed. The expectation is that the
disturbances be heavily reduced leaving noise levels of maximum 0.5%
on smooth asphalt roads during free rolling. In [Pavkovi et al., 2006] even
better results are shown. The variance of the remaining noise in a general
case is difficult to determine since it depends on many factors, for exam-
ple, the actual surface conditions and the sensitivity towards the force
transients.
118 119
7.1 Introduction
The aim of the RFE project is to estimate the friction between tire and
7 road and to evaluate and optimize the reliability as well as the delay of the
estimation. Three different concepts are analysed within the RFE-project.
This chapter cover work done within the sub project, Model-based road Surface detection by preview techniques Volvo Technology and Luleå
friction estimation in the IVSS-programme, Road Friction Estimation, Technical University study different concepts to estimate the surface con-
RFE. dition in front of the vehicle by optical means. A proposal under evaluation
IVSS, Intelligent Vehicle Safety Systems, is a unique joint venture by is a sensor called RoadEye [Optical Sensors, 2006] that transmits a couple
public-sector agencies, private-sector companies, and industry organiza- of laser frequencies on to the road and measures the reflections. With this
tions, that was set up to stimulate research and development for the road technique it is possible to detect whether there is a coating layer on the
safety of the future and aims at moving the emphasis from passive solu- road surface, such as water, snow, or ice, see [Casselgren et al., 2007].
tions to active systems. In other words, preventing problems from arising
in the first place [Vägverket, 2006b]. Model-based road friction estimation The thesis presents work per-
formed within this sub project. The main aim is to develop algorithms
for friction estimation during longitudinal tire force excitations. The re-
sults are derived in close cooperation with Johan Hultén at Volvo Cars,
7.1 Introduction Fredrik Bruzelius at Volvo Technology, and Magnus Gäfvert at Modelon
AB. Besides IVSS Haldex Brake Products AB and Volvo Cars finance the
work.
The Road Friction Estimation project
Road vehicles rely strongly on friction. Their large masses that often move Model-based road friction estimation
at high speeds may cause fatal damage if they loose steerability. The A popular and in many circumstances successful approach to assess the
controlling tire forces are generated by and dependent on a sometimes friction is to estimate the braking stiffness, i.e. the incline of the tire force
abruptly changing friction. A large safety margin in the traffic should be relative the slip at low slips and from this value distinguishing between
compulsory, but is often not sufficiently regarded by the drivers. Modern different surface conditions. Note that the term “braking” in this context
vehicle control systems can, to some extent, correct for incautious actions refers to longitudinal slips and forces and is also used to denote the force-
from the driver, but a correct appraisal of the driving circumstances is slip inclination at low slips during driving. A linear tire model is hereby
mandatory for safe driving. used with an assumption on the relation between the slope and the fric-
Many investigations show a correlation between the road condition and tion condition, see e.g. [Gustafsson, 1997]. The braking stiffness depends
the accident risk, see e.g. [Wallman and Åström, 2001]. The output from on many factors and a generic relation between the slope and the exact
a road friction estimator might be used as a detecting device that warns friction coefficient is therefore not possible to obtain. Even the surface
the driver about a bad or suddenly changed road condition. Information detection is not always reliable due to the large variation of the braking
about the friction can also be used to enhance the functionality of ac- stiffness caused by other reasons, see for instance [Pavkovi et al., 2006].
tive and adaptive control systems within the vehicle or sent to a global This work aims at estimating the value of the friction coefficient by
infrastructure that receives and transmits information about the roads. using a well-defined accurate model for the effect of the friction on the
120 121
Chapter 7. Road Friction Estimation 7.1 Introduction
tire behavior. The model adopted for this purpose is the physically based tion estimation by a validation towards measurement data. Section 7.3
brush-model, that describes the effects of the rubber deformation and fric- presents the idea of storing sampled data in data bins, which is a way
tional limitations in the contact patch and specifies the curvature of the to lump measurements in separate regions distributed over the force and
force-slip relation. The information of the friction is derived from the de- slip axles together. This is a way to remember and filter old data and
viation from linearity in the force-slip slope, which is clearly detectable at to decrease the sensitivity to disturbances. The states in the data bin,
larger slips. The estimation of the linearity at small slips is a similarity representing the normalized longitudinal tire force and the tire slip are
to the previous method, but its slope is not used as indicator of the fric- used in the cost function from which the optimal parameters denoting the
tion. In the proposed method the detection of the incline at low slip is a braking stiffness and friction coefficient are calculated. The least squares
necessity for deriving the curvature. method and the Gauss-Newton algorithm are proposed for this purpose.
One advantage of the method is that the algorithm is based on a clear, Implementation details and the additional vehicle estimators neces-
verified relation between the measurements and the friction, which makes sary for deriving the tire force and slip signal are described in Section 7.4
the estimation more reliable. The drawback is it that two tire-road related followed by Section 7.5 presenting the results from tests in reality. The
parameters have to be estimated and the method requires higher force and chapter concludes with a discussion about the estimator properties and
slip excitations compared to stiffness-based methods. the requirements on further work.
Another problem that arises when the curvature of a slope is to be
determined is that the measurement data need to be sufficiently spread Friction estimation — Prior art
in the force and slip plane for an accurate estimate. If the data is clus- Much effort is world-widely spent on developing algorithms for friction
tered around one point the solution become ambiguous and any parameter estimation. A literature study in this area is quite difficult, since pub-
couple that results in a curve that crosses this point may be a possible lished articles often describe academic nice solutions, but these are often
solution. The robustness of the estimation decreases and noise, model er- restricted to certain circumstances with few real implementations. The
ror, and other temporal disturbances will have a very large impact on the companies, having powerful test equipment, mostly do not publish articles
result. To overcome this, a particular data storage system is developed to, due to the hard competition. Instead the ideas are protected by patents,
as good as possible, utilize the spread of the available measurements. that are difficult to penetrate and understand. One exception is work done
by NIRA-dynamics, which is been presented in many articles, for exam-
Outline ple [Gustafsson, 1997] and also developed into a product. The product
A main prerequisite for the choice of a model-based estimator for the is an algorithm that uses information from the wheel speed signal, the
road friction is a clear formulation of the model, which describes how the vehicle velocity, and the engine torque to distinguish between the road
friction coefficient influences the behavior of the tire in the interaction foundations: gravel, ice, snow, and asphalt. The idea is to evaluate the
with the road. Here, it is assumed that the friction affects the relation braking stiffness and correlate its value to the surface friction. Using the
between the slip and the developed tire force and that these signals are braking stiffness as a friction indicator is not always reliable and the al-
available for measurement or estimation. The model also has to reflect the gorithm is limited to certain conditions. The variance in the wheel speed
variation rate of the included parameters and should specify the expected signal is used to separate between rough and smooth surfaces. Further
model and measurement noise distributions. developments and discussion about this method can be found in [Müller
The function of this kind of estimator can be divided into three sub and Uchanski, 2001] and [Müller et al., 2003]. Another approach has been
problems: published in [Ray, 1997] where the vehicle longitudinal and lateral accel-
erations, the vehicle yaw rate, and the speed of the wheels are measured.
• Choice of the tire model The corresponding tire forces are estimated by an extended Kalman-Busy
• Determination of signals to the tire model. In this case the normal- filter and among several tire models the most probable is chosen through
ized tire force and the slip. Bayesian learning. An expectation is that such an algorithm is depending
on extensive calibration and that the choice of tire models has to take the
• Construction of the estimator algorithm behavior of the actual tire into account.
In this work the brush model, see Section 4.1 is used to describe the One method based on the brush model is published in [Pasterkamp
tire behavior. Section 7.2 contains an evaluation of its reliability for fric- and Pacejka, 1997], where the relation between the self-aligning torque
122 123
Chapter 7. Road Friction Estimation 7.2 Experimental Validation of the Brush Model
and the lateral tire force is used to estimate the friction while turning. agreement of the tire model towards real conditions and to investigate the
The approximations in the brush model becomes apparent in the real sensitivity of the included parameters toward various factors that may
application and neural networking is employed to learn the estimator to change during normal run of the vehicle. It is important to validate the
determine the friction. A simpler method for using the relation between model to measurement data collected at real driving. The environmental
the self-aligning torque and the lateral tire force can be found in [Yasui disturbances and the limited measurement conditions in a production car
et al., 2004]. have to be regarded when evaluating the fit of the data to the brush model.
Several different friction-estimation techniques are discussed in [Eich- Otherwise, the results may not be valid and the estimator will not work
horn and Roth, 1992]. One unique approach that is presented is to mount as expected in the real implementation. However, in the first verification
a microphone close to the tire-road contact and by the acoustics separate of the usability of the brush model, it is desirable to minimize the un-
the different road surfaces. It is concluded that the recorded sound con- certainties from the signal estimations, necessary for the limited sensor
tains information about the friction, but distinguishing it from the other situation in a normal car. Therefore, the measurements are performed by
possible noise sources is difficult. The paper also includes an evaluation of a special tire test-truck from VTI, but the tires and road conditions are
a tread strain sensor, which shows promising result at least for detecting normal. Additional available tests results, not collected within the project
low friction. This kind of sensor is further described in [Pohl et al., 1999]. will also be referenced in this section.
Due to the varying pressure distribution in the contact patch there is of-
ten a region where the tread slides on the road. The friction is derived by Test equipment The test vehicle, denoted BV12, is a Scania truck
evaluating the vertical and horizontal tread forces in that region. LB80, see Figure 7.1, owned and run by VTI and equipped with a fifth
In [Umeno et al., 2002] the frequency contents of the wheel speed signal wheel for various measurements and slip and force excitations of tires
is analysed. From the frequency and magnitude of the resonance peak, for personal cars. The test wheel is pressed against the road by the pres-
the tire stiffness is estimated. The method can not predict the friction sure from a hydraulic cylinder, which can achieve different vertical loads
coefficient, but might work as a detector of a change of the road surface. in the range 1–6 kN. A varying brake torque can be applied on to the
The number of patents dealing with this kind of problem is voluminous. wheel through a disc brake controlled by a hydraulic brake system. The
Searching on friction estimation of tires give almost 500 hits at this date. angle between the vehicle travel and the wheel rotational direction can
In this project, a share of the found patents were selected for further also be changed during a test sequence. There are sensors on the wheel
examination, but a suitable approach for friction estimation was not found suspension for measurements of the vertical and horizontal forces and the
among those. Probably, related innovations are hidden from being found by vertical torque working on the wheel hub. There is also a sensor to mea-
forming the title without mentioning words as “friction” and “estimation”. sure the rotational velocity of the test wheel. The vehicle reference speed
One such example is [Levy and Fangeat, 2004] which describes a device is obtained from the left-front wheel of the truck. Further information is
that measures the secant G (λ ) = f x (λ )/λ in the force-slip plane. From the available from [Nordström and Åström, 2001].
presumed tire model, G (λ ) = a − bλ , predicting the secant behavior, a and
b are estimated. By knowing a and b, further conclusions about µ and λ ○ , Test facilities To be able to collect information about a number of
substantiated by measurement data, can be drawn. Similar expectations tires on several different road foundations with rich variation, tests were
can be derived from the brush model where σ ○ ( b/ a, λ ○ = 1/(σ ○ + 1) and performed at two test grounds.
µ ( a2 /(3b) even if these differ slightly from the conclusions in the patent.
The proposed method resembles this approach in that two parameters • Hällered is a proving ground close to Borås owned by Volvo Com-
are estimated from the tire characteristics. The estimation routines are, pany. The facility has an area of 700 hectares and includes tracks
however, completely different. for high-speed, endurance, off-road, and comfort testing tracks [Volvo
AB, 2006]. For the testing in this project mainly the brake and hand-
ling area was used, where high and low-friction tracks are available
7.2 Experimental Validation of the Brush Model with the possibility of water spraying. The high friction surface is
normal asphalt, having the same properties as a well treated high-
This section presents an experimental validation of the brush-tire model way. The low friction area is built up by basalt bricks providing low
used in the Road Friction Estimation project. The aim is to validate the friction when sprayed with water.
124 125
Chapter 7. Road Friction Estimation 7.2 Experimental Validation of the Brush Model
which were:
Test results
The result from the measurements are shown in figures with the slip, λ ,
see definition in (2.5), on the x-axis and the normalized longitudinal brake
force − Fx / Fz on the y-axis. The different plot-styles of the data points
denote which of the repeated sequences in the test file the measurements
belong to. The solid line is the optimal-fitting brush-model adaptation to
Figure 7.1 The test truck BV12 in Arjeplog.
the data. The brush model is given by
2 3
− C σ + 1 Cx σ x pσ x p − 1 ( Cxσ x )
• Arjeplog is a famous place in northern Sweden for winter test-
if pσ x p < σ ○x
x x
ing of vehicles. Many companies provide test facilities for the au- Fx = 3 µ Fz 27 (µ Fz) 2 (7.1)
tomotive industry. The tests in the project were performed at the −µ Fzsign(σ x ) otherwise
proving ground run by Colmis AB, see [Colmis AB, 2006], since both
Haldex and GM have contracts with them. The available surfaces are where σ ○x = 3µ Fz/ Cx and the optimal values on C0x = Cx / Fz and µ are
ice, snow and asphalt. The asphalt patch is held free from snow by derived by minimizing the squared error along the y-axis. The optimiza-
heat from elements in the ground, but when it is snowing the lanes tion algorithm is a Gauss-Newton method similar to the one that will be
quickly fills with water and melting snow. The frictional properties presented in Section 7.3. For the optimization all visible data points in
of both snow and ice are temperature sensitive and the temperature the plots, with slip lower than approximately 20% are used and equally
can change in a large interval from day to day. weighted in the cost function. The shown data points belongs to the first
part of the test sequences where the braking torque is increasing.
Test procedure The longitudinal tire data was generated by applying The measured signals contain biases which effect the fit of the brush
a braking torque on the wheel as a ramp function from free rolling to model. The bias on the longitudinal force sensor is mostly small and need
complete lock-up of the wheel. When the wheel got locked the brake pres- no further correction. The bias on the longitudinal slip may be explained
sure was released and this sequence was repeated for a number of times by error in the estimated free rolling wheel-radius, R e , defined as the
depending on the available road length. Between 2–5 sequences where col- rolling radius of the wheel when no force is transmitted by the tire, see
lected in one file. For each test setup 1–3 files were recorded. In Hällered Section 2.4. Before start up of the tests R e is identified. There are, how-
the vehicle speed was 70 km/h during the braking phases and in the Ar- ever, many factors that affect R e and changes in the radius may alter the
jeplog it was 50 km/h. The winter tire with 4 kN vertical load was chosen slip bias. Therefore, R e has to be identified and accounted for in each
as reference setup and was subjected to all available test conditions. The test sequence. The slip bias, often called horizontal shift, sh , is related to
load dependence was measured by varying the vertical load between 2, 4, changes in the wheel radius as
and 6 kN. Dry asphalt was tested both for summer and winter tires. Re-
sult from wet basalt giving a low-friction firm surface were only obtained ∆ Re
for the winter tires. Snow and ice was tested with all three tire types, sh ( (7.2)
Re
126 127
Chapter 7. Road Friction Estimation 7.2 Experimental Validation of the Brush Model
Only a few of the available measurements are shown here, but more 1.2
Braking −Winter tire −Wet asphalt −H0511_107.dat
Braking −Winter tire −Wet asphalt −H0511_107.dat
denius, 2007b].
1
0
0.8 −5
Winter tire Winter tires, generally, have relative soft rubber compound 6
mer tire, which can be seen on lower values on C0x . The rubber texture is 0.4 4
3
rough to provide good grip in loose foundations such as snow. 0.2 1
2
Measurements together with the adaptations of the brush model to the C =28.3 µ=1.02 s =1.72
x h
2
3 1
winter tire on wet asphalt are shown in the force-slip-plane to the left in
Slip
0 4
5 0
Figure 7.2. The right plot shows the test sequences generating the data.
brush
−0.2 −1
−5 0 5 10 15 20 0 2 4 6 8 10
The force excitation is similar to the other tests, but the number of appli- Slip % Time [s]
cations varies. The results from dry asphalt and from basalt are shown Figure 7.2 Measurement data from Hällered showing results from the winter
in Figure 7.3. On both wet and dry asphalt the agreement of the brush tire on wet asphalt with a vertical load of 4 kN. To the left: Adaptation of the brush
model is very good. The measurement deviations are symmetrical around model. The normalized brake force is shown as a function of the slip, λ . To the right:
the model realisation and can be explained by temporary disturbances. Recorded signals showing the test sequences in the measurement file.
The cluster of points in the low-force region (p f x p < 0.3) are explained
by the vibrations in the measurement rig when the braking torque is 1.4
Braking −Winter tire −Dry asphalt −H0511_149.dat
0.4
Braking −Winter tire −Low friction −H0511_124.dat
released. See the force measurements in the upper plot to the right in 1.2 0.35
Figure 7.2 between, for instance, 2.5 and 3 sec. For the low-friction foun- 0.3
dation the temporal disturbances largely affects the measurements. The 1
0.25
splices between the basalt bricks introduce extra road unevennesses and 0.8
0.2
the disturbances are slightly larger than for asphalt, but their relatively 0.6 0.15
impact become larger due to the lower excitation force needed to lock the 0.4 0.1 Cx=14.6 µ=0.312 sh=0.807
wheel. It is difficult to verify the model from these measurements, but no 1 0.05
for snow and ice. The snow measurements are noisier due to the uneven-
brush brush
−0.2 −0.1
−5 0 5 10 15 20 −5 0 5 10 15
nesses and traces in the randomly packed snow. In spite of this, it can Slip % Slip %
be seen from the data that the performance on snow diverge from the
Figure 7.3 Measurements of the winter tire at Hällered together with a brush
brush model particularly at normalized forces above 0.2. The tire curve model adaptation at the vertical load 4 kN. To the left: On wet asphalt. To the right:
has more substantial curvature and no pronounced force peak. The tire On basalt. Note, that scaling on the axes differs between the figures.
force increases with the slip, at least up to λ = 40%. This is probably an
effect of the snow deformation.
Due to the large spread in the data it is difficult to draw any clear evacuating water from the contact patch to prevent friction losses due to
conclusions of the ice measurements. a water film when driving on a wet road. The texture on the tested sum-
mer tire seems to efficiently cope with the water, since the difference in
Summer tire Summer tires are designed to achieve good grip in sum- performance between the dry and wet asphalt is small, which can be seen
mer condition. Compared to the winter tire, the somewhat simpler task in Figure 7.5. A difference is that the tire force for the wet asphalt has a
increases the possibilities to improve properties such as rolling resistance slightly less curvature up to the peak, which comes at an lower slip then
and wear resistance to provide a more cost-effective driving. Slick tires predicted by the brush model. The force peak is more pronounced in the
without texture is said to give best grip in good weather conditions on wet condition and the friction force reduces quicker for higher slip. The
asphalt [Haney, 2003]. Grooves in the tire are, however, inevitable for poor frictional properties of the summer tire on snow and ice, can be seen
128 129
Chapter 7. Road Friction Estimation 7.2 Experimental Validation of the Brush Model
rak broms −Vinterdäck −Snö −Arj0603_100.dat rak broms −Vinterdäck −is −Arj0603_106.dat rak broms −Sommardäck −Snö −Arj0603_157.dat rak broms −Sommardäck −is −Arj0603_167.dat
0.6 0.12 0.6 0.14
0.08 0.1
0.4 0.4
0.06 0.08
0.3 C =7.16 µ=0.0723 s =0.0996
x h
0.3
0.04 0.06
0.2 0.2 C =4.14 µ=0.0854 s =1.32
x h
0.02 0.04
0.1 1 1 0.1 1
Cx=14.7 µ=0.382 sh=0.571 2 0 2 Cx=33.3 µ=0.29 sh=0.874 2 0.02
3 3 3 1
0 4 4 0 4 2
−0.02 0
5 5 5 3
BM BM BM BM
−0.1 −0.04 −0.1 −0.02
−5 0 5 10 15 20 25 30 35 40 −5 0 5 10 15 −5 0 5 10 15 −5 0 5 10 15
Slip % Slip % Slip % Slip %
Figure 7.4 Adaptation of the brush model towards measurements data from Ar- Figure 7.6 Adaptation of the brush model towards measurement data from Arje-
jeplog with winter tire and with the vertical load 4 kN. To the left: On snow. To the plog with summer tire and with a vertical load of 4 kN. To the left: On snow. To the
right: On Ice. Note, that scaling on the axes differs between the figures. right: On Ice. Note that scaling on the axes differs between the figures.
Braking −Summer Tire −Wet asphalt −H0511_154.dat Braking −Summer Tire −Dry asphalt −H0511_146.dat
1.2 1.4
in the environment to a larger extent than when using normal tires. The
attention to this kind of pollution problem have gained a larger interest
1 1.2
0.8
1
recently. Studded tires have a positive effect on the road friction in that
0.8 they tear away the polished surface caused by normal tires. On the other
hand, the traces from the tires in the road get deeper and collect more
0.6
0.6
0.4 water, which require more extensive road maintenance [Gustafsson et al.,
0.4
0.2 1 1
2006b]. The studded tire is, however, outstanding in friction on ice, which
C =41 µ=1.1 s =1.17
x h
2
3
0.2
Cx=37.2 µ=1.2 sh=1.02
2
3
can be seen from the results in Figure 7.7. Particularly on ice, the mea-
0 4
5
0 4
5
surement signals contain more noise than the other tires. The grip of the
−0.2
−5 0 5 10 15
brush
20
−0.2
−5 0 5 10 15
brush
20
spikes seems to be varying and unpredictable. The measured behavior on
Slip % Slip %
snow is very similar to the winter tire, but with a larger noise level. The
latter is most probably a result of the snow packing when running back
Figure 7.5 Adaptation of the brush model towards measurement data from
Hällered with the summer tire and with a vertical load of 4 kN. To the left: On
and forth on the test lane or of shifting of the snow properties by other
wet asphalt. To the right: On dry asphalt. Note that scaling on the axes differs reasons.
between the figures. In Figure 7.8 the footprints from the studded tire on ice in a braking
phase can be seen. The increasing length on grooves from the spikes is a
consequence of the increasing tire slip caused by ramping up the braking
in Figure 7.6. The tire do not get a good grip in the snow and the charac- torque. Since a bristle, according to the brush-model theory from Sec-
teristics clearly differs from both the winter and studded tire, in that the tion 4.1 both enters and leaves the road undeformed, the sliding distance
friction is lower and do not increase for higher slip. The ice properties do of a bristle onto the ground is ls ( vsx tc (−a) where vsx is the relative
not significantly diverge from the winter tire. velocity between the tire carcass and the road and tc (−a) is the time
that a bristle spends in the contact patch. The time for the road contact
Studded tire Studded tires are gripping the road foundation with can also be expressed as tc (−a) = 2a/vc . The groove length can then be
spikes. The tire force is build up with approximately 10% from the rubber calculated as ls ( vsx 2a/vc = σ x 2a. Hence, the slip could theoretically
friction and 90% from the spikes. A drawback with studded tires is that be derived from measurements of the groove length if the length of the
the abrade on the roads is extensive, and that micro particles are spread contact patch, 2a, is known.
130 131
Chapter 7. Road Friction Estimation 7.2 Experimental Validation of the Brush Model
0.7
rak broms −Dubbdäck −Snö −Arj0603_171.dat
0.3
rak broms −Dubbdäck −is −Arj0603_175.dat • The braking stiffness for the summer tire is larger than for the win-
ter tire and the studded tire.
0.6 0.25
0.5 0.2 • The difference between the friction of the summer and the winter
0.4
tire on asphalt is small.
0.15
0.3 0.1 • The normalized braking stiffness seems to increase for increasing
0.2 0.05 Cx=4.34 µ=0.158 sh=0.68 vertical load for the winter tire on asphalt and snow, but not on
1 1 ice. This topic is important for the friction estimator and further
0.1 2 0
discussed later on.
2
Cx=10.6 µ=0.458 sh=0.663 3 3
4 4
0 −0.05
5 5
−0.1
BM
−0.1
BM • The studded tire has better friction on ice, that the other tires.
−5 0 5 10 15 20 25 30 35 40 −5 0 5 10 15 20 25 30 35 40
• The summer tire has poor grip on snow, when studded tire and the
Slip % Slip %
Figure 7.7 Adaptation of the brush model towards measurements data from Ar- winter tire performs similar.
jeplog with studded tire and with the vertical load 4 kN. To the left: On snow. To
the right: On Ice. Note that scaling on the axes differs between the figures.
• There seems to be a relation between the tire stiffness and friction
coefficient, see Figure 7.9.
• The braking stiffness is larger for wet asphalt than for dry. This is
further discussed later on.
An interesting observation is that the winter tire and the studded tire
behaves similar on snow. The force-slip curve has no obvious maximal
point and the tire force seems to be an increasing function of the slip.
The curve also bends off more than the brush model which might be an
effect of that the tires grip into the snow which yields at a particular
strain. The summer tire differs and behaves more as it would slide on the
packed snow surface. There are models on how to deal with deformable
foundations, e.g [Meschke et al., 1996] and [Saino, 2001] and they tend to
Figure 7.8 Footprints of the studded tire on ice during brake application. The become very complex. One reason for not incorporating the snow behavior
brake torque is applied as a ramp and the increasing grooves are a result of the
in the brush model is the reluctance in introducing more parameters in
increasing tire force. the wheel is rolling from the left to the right.
the model in this work.
The higher noise levels on the measurements on snow for both the
Conclusions From the shown figures it can be seen that the brush summer tire and the studded tire compared the winter tire might be con-
model mostly fits the measurement data well. The optimized tire model fusing. These are most probably results of the snow packing when running
is well within the spread of the data points. The spread is, however, rela- back and forth on the test lane or of shifting of the snow properties by
tively large on the low friction surfaces, which depends on that the force other reasons. It is most unlikely that the different noise levels entirely
disturbances on the measurement rig becomes more apparent when the depends on the tire properties.
force excitation is low. The disturbances mainly come from low frequency
vibrations in the mechanical structure invoked by unevenesses in the road Brush model verification
and the force excitation on the wheel. The major question about the us- The most relevant issue when evaluating the measurement data is the
ability of the brush model for friction estimation for the tested tires on ability to estimate the friction coefficient from the brush model realisa-
the available surfaces are treated further on. tion. It is obvious that the friction can be accurately estimated when data
The derived tire parameters for the different test setups are presented from the full slip curve 0 < λ < 1 is available since the friction is com-
in Table 7.1 and visualized in Figure 7.9. From the results it is stated that pletely given by the normalized tire force at high slip. At low slip the entire
132 133
Chapter 7. Road Friction Estimation 7.2 Experimental Validation of the Brush Model
45 1.4
Table 7.1 Results of the parameter optimization 40
Winter tire 2 kN
Winter tire 4 kN
Winter tire 2 kN
Winter tire 4 kN
Winter tire 6 kN 1.2 Winter tire 6 kN
35 Summer tire Summer tire
Friction coefficient
Winter Wet asphalt 2 kN 23.8 0.98 1.2 25 0.8
134 135
Chapter 7. Road Friction Estimation 7.2 Experimental Validation of the Brush Model
rak broms −Vinterdäck −Snö −Arj0603_099.dat Braking −Winter tire −Wet asphalt −H0511_108.dat
0.6 30 1.4 42
1.3 39
1.2 36
0.5 25
x
Longitudinal Tire Stiffness
Figure 7.10 Parameter estimates as functions of the highest used slip value for Figure 7.11 Parameter estimates as functions of the highest used slip value for
the winter tire on snow. The estimated tire stiffness (dashed line with circles) and the winter tire on asphalt.
friction coefficient (full line with asterisks) by using only the data points to the left
of the slip limits [0.5,1,1.5 .. 8] (dashed-dotted vertical lines) are plotted. Also the
tire characteristics for the estimated parameters at each slip limit are plotted.
measurements. The optimization algorithm used to derive the examples
is not fully developed and further enhancements, such as filter and cor-
Table 7.2 Required utilization of friction (in percent) to achieve a friction estimate rect data weighting may improve the result significantly. Overall, from
within an accuracy of ±0.15. Only tests at 4 kN is used. the experiences from the brush model verification, the conclusion is that
an estimation algorithm based on the brush model will work, but high
Road foundation Winter Summer Studded tire friction utilization is required for an accurate friction estimate.
Dry asphalt 75 69 -
Wet asphalt 74 95 - Parameter sensitivity
Basalt 87 - - This section will discuss the sensitivity of the tire behavior to a few con-
ditional factors. It is also important to know in what ranges and circum-
Snow 66 81 74
stances a parameter might be expected to vary when designing an estima-
Ice 100 100 68 tor. An aim with the investigation is to find out whether further signals,
to cover for parameter changes, have to be regarded within the algorithm.
The friction coefficient depends on the tire-road interaction, while the
model around the force peak which is mentioned earlier. braking stiffness is assumed to depend on more tire related properties as
The investigation shows on the possibilities and restrictions for an fric- the tread stiffness and the size of the contact patch between the tire and
tion estimator relying on the brush model. The study has to be considered road, see (4.31). As been noticed previously and in, for example, [Gustafs-
as very approximate, since some of the disturbances that are present will son, 1997] and [Müller and Uchanski, 2001] there is a co-variation between
not affect the signals in a car. One example is the sensitivity towards the friction coefficient and the braking stiffness. The relation is physically
road unevenesses. The hydraulic piston generating the pressure between difficult to explain and does not always hold. Tests have confirmed that
the tire and the road has stiff and slow suspension. The vertical load on very rough asphalt can provide high friction, but a low stiffness, due to the
the tire therefore fluctuates more than what is expected in a passenger smaller effective contact patch compared to a smooth surface. The braking
car. The sensors in a car do, on the other hand, not deliver that precise stiffness is sensitive to many other factors and a change of the stiffness
136 137
Chapter 7. Road Friction Estimation 7.2 Experimental Validation of the Brush Model
30 1.4
1 1 25 1.2
Friction coefficient
Snow 0.8 Snow
0.6 0.6 15 Ice Ice
Wet Asphalt 0.6 Asphalt
0.4 0.4 10
0.4
0.2 0.2
5
0.2
0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0 0
Slip Slip 2 4 6 2 4 6
Vertical load on tire [kN] Vertical load on tire [kN]
Figure 7.13 Diagrams showing the load dependence in the previously presented
Figure 7.12 Plot of the force-slip relation of the original brush model. To the left:
measurements, see Table 7.1. To the left: The normalized braking stiffness, C0x . To
Varying tire stiffness Cx = [6, 8, 10, 12]. To the right: Varying road friction coefficient,
the right: The friction coefficient, µ .
µ = [0.6, 0.8, 1, 1.2].
138 139
Chapter 7. Road Friction Estimation 7.2 Experimental Validation of the Brush Model
0
1 2 3 4 5 6
Vertical load [kN]
First, the lower share of load on the frictional carrying contact surface and
then the lubricating effect of the water in the effective area in the contact.
Figure 7.14 Comparison of the load dependence between the braking stiffness and
the normalized braking stiffness. Longitudinal data from vehicle test, described in
In [Gothie et al., 2001] the friction for different tires, asphalt surfaces, and
Section 5.4 and [Salaani et al., 1999]. water depths is further examined.
140 141
Chapter 7. Road Friction Estimation 7.3 Tire Parameter Estimation
2003] where shaved tires are tested. The friction properties are mainly ef-
Table 7.4 Summary of parameter dependence. A, ’+’, signifies that the factor
fected on wet road or on loose foundation such as snow and gravel, where increases if the factor increases and vice versa. – means the opposite. A small and
the tread pattern is important. normally neglectable relation is denoted by ∼. No sign means that no results are
available.
Velocity dependence The velocity mainly affects the friction in the
meaning that the sliding velocity between the tire and the road is higher Parameter Pressure load Wear Water Temp. Vel.
for a higher vehicle speed at a particular slip. The friction is often refereed Re + – – ∼+ ∼+
to as sliding speed dependent, see e.g. [Guo et al., 2006]. Different road
Cx – + + –
foundations and different tires certainly exhibits different interactions. A
water film on the road increases the velocity dependence and increasing µ – – –
velocity drastically increases the risk for aquaplaning. An example of the
velocity dependence of the friction on a 3 mm thick water layer on the road
can be found in [Do and Delanne, 2003], where the locked wheel friction best case 66%, i.e if µ = 0.5 a normalized tire force of 0.5 ⋅ 0.66 = 0.33
drops from 0.7 at stand still to around 0.05 at 100 km/h. is required. In the worst case the utilization has to be 95%. The used
In [Pirelli, 2005] the velocity dependence on the rolling radius was parameter optimization is a pure minimization of the squared error be-
tested for a few type of tires. The result showed that the rolling radius tween the data and the model. It is emphasized that a more advanced
increases for higher speed, but first above a (in Sweden) non-legally speed estimation algorithm might reduce the need of friction utilization. The
limit it became noticeable. measurements performed by VTI incorporate discrepancies not present
in a personal car. High demands on the signal processing that provides
Temperature sensitivity Rubber is normally very sensitive to tem- the required signals, Fx , Fz , λ from the existing sensors, might further
perature changes. Therefore, also the tire parameters will change with improve the estimation results. It is also concluded that the parameters
the temperature. In [Carlson and Gerdes, 2003] a decrease of 20% on the included in the brush model varies with many factors. Particularly impor-
braking stiffness between the first and the fifth braking in a number tant for a friction estimator is the load dependence of the rolling radius
of braking sequences is reported which is explained by the temperature and the braking stiffness. Since the application of an acceleration or re-
increase. For the same condition the increase in rolling radius was less tardation force causes a load transfer on the vehicle, the vertical force on
0.1% (0.3 mm). In [Mizuno et al., 2005] the braking stiffness was mod- the tires will change. Changes in the rolling radius will then affect the
elled, based on measurement data, to decrease 5% within a temperature slip calculation and cause a distorted slip-force curve. The variation of the
change from 40 to 60 [deg] C. braking stiffness for varying loads differs between the tires. Generally, the
normalized braking stiffness is less sensitive to load fluctuations than the
Summary The dependence of the parameters in the brush model and braking stiffness. Another important factor is the temperature sensitiv-
the rolling radius towards the examined factors are summed up in Ta- ity of the braking stiffness. The braking stiffness can change around 20%
ble 7.4 within a couple of brake applications, due to warming, without any change
in friction.
Conclusions
This section has treated the issue, whether it is possible to describe any
tire at any common road surface with the brush model containing only two 7.3 Tire Parameter Estimation
parameters. The answer is of course, that it depends on the requirements
on the accuracy of the model. In this context the accuracy of the model The literature contains many different methods to identify a relation that
towards the data affects the possibilities to estimate the friction coeffi- describes the behavior of an observed system, see for instance [Johansson,
cient. A conclusion is that there are deviations between the data and the 2002]. In some cases the identification aims at finding the set of equations
model, but friction estimation is possible, even though rather high utiliza- that best describes the relation between the input and output signal of an
tion of the friction is required for an accurate estimate. The utilization of unknown, so called black-box system. In the actual case the system is
the friction for estimating the friction coefficient within ±15% is in the partly known and better explained as a grey-box model, where the behav-
142 143
Chapter 7. Road Friction Estimation 7.3 Tire Parameter Estimation
ior is described by a particular equation structure, which includes known point. The measurements around this point will not contain information
and unknown parameters. The latter are to be estimated. Normally, the sufficient to estimate two parameters and the parameter observability and
procedures for parameter estimation aim at finding the set of parameters accuracy decreases drastically. As long as the estimated curve intersects
that minimizes the error between the adopted model and the measured the stationary point the parameters will change in between according to
output for a known input signal. The calculation of the error differs be- the present noise conditions and lose the relation to the actual friction
tween the methods. In off-line methods the accumulated error for a set condition and braking stiffness. Proposals on how to supervise adaptive
of data is used for the minimization. In the parameter computation all controllers in related problems at limited excitation situations are given
values in the data set, which can be measurement during a certain time in, e.g [Hägglund and Åström, 2000]. A different approach is proposed in
period or observations collected by other means, are regarded without any this work.
demands on causality. In real-time systems often recursive methods are In the first phase, invariant surface conditions are assumed, such that
used to on-line update the estimation at each sampling instant. Mostly, there is no need to distinguish between old and new data. Information
the estimates at each time instant are based on a weighted sum of the from the force and slip measurements are stored in data bins distributed
previous estimate and information that can be extracted from the present over the force and slip-axis. The optimal tire parameters are then calcu-
measurements, see [Åström and Wittenmark, 1994]. Previous input data lated using the stored data. Techniques to treat detection of a new surface
are only remembered implicitly in the estimated parameters and are for- and adaptation to slow changes of the braking stiffness and friction can
gotten as time passes. be implemented together with the data storage and will be discussed af-
A major problem in the present application is the issue of persistence terwards.
of excitation. For the further description, the excitation is separated into
space and time domains. Having persistence of excitation in the space Formulation of the estimation problem
domain requires that measurements are available in a sufficiently large The tire characteristic according to the normalized brush model, compare
region in the force-slip plane, providing an accurate stiffness and friction to Equation (7.1) used for the estimation is given as
estimation. The aim of the method is to, as good as possible, estimate the
friction coefficient at a low force excitation. The lower friction utilization, 2 3
− C σ + 1 C0xσ x pσ x p − 1 ( C0xσ x )
the smaller amount of information about the friction is contained in the if pσ x p < σ ○x
0x x 2
Ψ(σ x , C0x , µ ) = 3 µ 27 µ
data. Hence, the sensitivity towards measurement errors and model de- −µ sign(σ ) otherwise
viations toward real conditions increases. Further, it is required that the ⋅ x
data is spread over sufficiently large region to be able to get an accurate (7.3)
and robust estimate. If the data points are clustered the cost function will where the slip, σ x and the normalized tire force, f x = Fx / Fz = Ψ(σ x , C0x , µ )
not have a distinct minimum and the quality of the results decrease. are the measured signals. The included parameters θ = [ C0x , µ ] are to be
Having persistence of excitation in the time domain requires that the estimated.
inputs vary sufficiently irregularly and often. When the conditions have
changed, the data information should be fresh, to provide an accurate Data filtering by use of storage-bins
estimation that reflects the actual surface and tire condition. The motivation for data-filtering based on storage-bins is the problem of
Traditional estimation techniques are time depending which makes estimating the parameters of a scalar nonlinear function.
it hard to formulate the space-distributional requirement on the data. A
typical situation where those methods fail is a brake application starting y = ( x, θ) (7.4)
with a tire-force ramp which provides evenly spread data in the space
domain. If the final tire-force value is large enough, this excitation is The specific application at hand is the estimation of the normalized brak-
sufficient for a proper parameter estimation. When the ramp has settled ing stiffness, C0x , and the friction coefficient, µ , from measurements of tire
the tire force is kept constant for the major part of the event. A recursive slip, σ x , and normalized tire force, f x , using the brush tire-model (7.3).
algorithm calibrated to be fast enough to catch the parameters during The problem is divided into the construction of a point-wise approximation
the ramp phase will soon forget the information collected in the low slip of Ψ described by pairs (σ ′xi , fi′ ) and (σ ′′xi , f xi
′′
) constructed from the mea-
region and adapt to the measurements around the reached stationary surements, and the estimation of C0x and µ is performed by minimizing
144 145
Chapter 7. Road Friction Estimation 7.3 Tire Parameter Estimation
the cost:
X 2 X ′′ ′′ 2
w′i f xi
′
− Ψ(σ ′xi , Ĉ0x , µ̂ ) + w j f x j − Ψ(σ ′′x j , Ĉ0x , µ̂ ) (7.5)
i j
146 147
Chapter 7. Road Friction Estimation 7.3 Tire Parameter Estimation
the estimate xi′ will be aligned to the center of the bin X i , and the esti- The least-squares method The least-squares method is a linear-re-
mate yRi′ will go toward the mean value of function in the actual bin, i.e. gression analysis method based on the model formulation
pX i p−1 x∈X i ( x) dx, under the assumption of equal and uniform distribu-
tion of the samples x̂. Note that the local estimate only fulfills yi′ = ( xi′ ) yi = φT θ + ei (7.11)
if the length on the interval pX i p is infinite. The error depends on the de-
viation from linearity of in the bin range. The variance of the estimates where the output, yi and the input, xi are measured, calculated, or, in
decrease with the rate 1/ ni . In practice, it is unlikely that the Central other way derived signals. The elements in the regressor vector φ corre-
Limit Theorem convergence will have any significant visible effects, since spond to the parameters for estimation in the vector θ. The linearised
the sample-count limit Ni will truncate the averaging sum. approximation, φT θ should not largely deviate from the real parameter
relation for proper work of the least squares. The method in this form is a
typical offline-method, which means that a batch of data, is necessary for
Confidence assessment The data-bin tuple can also be extended with
the calculation of the parameters. For on-line use, the method is rewrit-
measures of the content quality. In particular, a recursive variance es-
ten, as the recursive least-squares, see for example [Ljung and Söderström,
timate of the samples can be introduced. Likewise, a freshness measure
1983]. The result of the least-squares, derived for example, in [Gelb et al.,
can be introduced to keep track of which data bins that contain the most
1977], is the estimate that minimizes the sum of the square of the errors
recent information. Measures like this can be used to form weights wi to
(7.9) and can be calculated as
be used in the cost function (7.5) for parameter estimation.
θˆ = (ΦT W Φ)−1 ΦT W y (7.12)
Optimization routine
The estimator is an optimization S algorithm that from a set of data points, The regressor matrix, Φ contains the corresponding regressor vectors as
Ω = {Ω i = [ xi , yi ], [ x, y] = [ x′ , y′ ] [ x′′ , y′′ ]}, calculates the parameters, θ, [φ1 φ2 .. φ N ]T and the weight matrix is formed as W = Iw.
included in a function, , such that the cost function Using the least-squares method together with the brush model, the
demand on linearity in the parameters, restricts the working range. For
N low slip, the tire stiffness can be calculated by choosing θ = C0x , with
X wi e2 (Ω i , θ) φi = − xi . For estimation of the friction when a curvature of the force-slip
V (Ω , θ) = (7.9)
2 relation is detected, the parameters are, θ = [ C0x θ 2 ]T and the regressors
i=1
φi = [− xi xi p xi p]T . The friction coefficient can be derived from the result
2
is minimized. The number of coupled observations in Ω is N. The weight- as µ = C0x /(3θ 2 ). The neglection of the third-order factor of the slip from
ing of each data point for the estimation is specified in the vector w = the brush model will give an error at higher slip and the method is not
{wi ∈ [0, 1]}. The error is derived as reliable if the data set includes slip values above σ ○x . The input and output
signals, x, y are given by the σ x and f x values derived and stored using
the storage-bin algorithm.
ei = e(Ω i , θ) = yi − ( xi , θ) (7.10)
Gauss-Newton optimization In the area of convex optimization the
Two different optimization algorithms are found suitable in the actual Gauss-Newton method is used as a substitute to the least-squares method
case. The least-squares method, a robust method, limited to linear param- when non-linear problems have to be solved. The general problem is to
eter relations and therefore used to detect the slope curvature at low slips minimize the sum of errors (7.9). A non-linear problem can in most cases
where the third order term, from Equation (7.3) is neglectable. The it- not be solved directly and the solution has to be found by iteration. The
erative and more computer demanding Gauss-Newton algorithm catches amount of iterations depends, e.g. on the properties of the function, the ini-
non-linear behavior and is used to reach a more accurate and reliable es- tial values of the parameters and the required accuracy on the estimates.
timation when data points closer to the friction limit are available. The A common way is to update the parameters in the direction of steepest
methods and their implementation using the brush model are described descent of the cost, V (Ω , θ), hence θk+1 = θ k − ∇ V (Ω , θk ). In Newton’s
in the following. method the quadratic term, corresponding to the change of the gradient,
148 149
Chapter 7. Road Friction Estimation 7.3 Tire Parameter Estimation
is used to enhance the step length and direction of the search. The up- with σ ○x = 3µ / C0x . The method works on a batch of data just as the method
date becomes θk+1 = θk − H −1 (Ω , θk )∇ V (Ω , θ k ) where H is the Hessian, of least-squares. The benefit with Gauss-Newton is that it catches the non-
a quadratic matrix, whose entries are the second order derivatives of V linearity in the brush model, which makes the result more accurate at slip
with respect to θ, defined as values close to and above the limit slip. A drawback is that the method
is more computer demanding, due to the many operations necessary to
2 V derive J and the risk of numerous iterations for an accurate result.
H (Ω , θ) = (Ω , θ) (7.13)
θ i θ j
Properties and further functionalities of the estimator
In most cases the calculation of the Hessian is not convenient and in the Combining the storage-bin concept with one of the presented optimization
Gauss-Newton method it is approximated as H (Ω , θ) ( 2J (Ω , θ)W J T (Ω , θ). algorithms gives a parameter estimator that copes with space-domain per-
The Jacobian matrix of e is defined as, J (Ω , θ) = ( ei /θ j ) with i = 1...N sistence in a consistent manner. The bin-filtering will have consequences
denoting the row number and j the column number. The update scheme on the signal properties, which may affect the reliability of the resulting
becomes estimates. This will be discussed in the following. The low-pass filtering
of the data in the bin-storage will only partly cover the ability of the algo-
θk+1 = θk − ( J (Ω , θk )T W J (Ω , θk))−1 J (Ω , θk)W e)) (7.14) rithm to adapt for changing parameters. Proposals on how to enhance such
behavior and how to minimize the impact of outliers are also discussed
For the Gauss-Newton method to converge it is required that the initial below.
parameters are sufficiently close to the optimal values and that the ap-
proximation of H is valid, which is the case if the residual errors e are Effects on the optimization due to the storage-bin classification
small enough [Böiers, 2004]. Methods derived in the area of parameter optimization mainly aims at
To adapt the method to the brush model, θ will denote the parameters minimizing the measurement error of a model output, hence the plant is
[ C0x µ ]. The error vector, e is calculated from (7.10) with = Ψ as the assumed to be on the form y = G (θ , x) + e, where the input, x, is cal-
brush model from (7.3). The model output, y, and the input, x, are replaced culated by the controller before fed into the system and therefore known
by the contents in the data bins, the tire force f x and the tire slip σ x with good accuracy. In the present case, the system is given an excitation
respective. The Jacobian matrix becomes from where it is possible to obtain two signals from measurements on the
vehicle. Due to the relation between these signals information can be ex-
e1 e2 eN T
tracted such that estimation of the friction coefficient is possible. Hence,
C0x ⋅⋅⋅ there are measurement errors on both included signals. In the process of
C0x C0x
J (Ω , θ ) =
e
(7.15) computation and estimation of the particular signals used for the param-
1 e2 eN
⋅⋅⋅ eter optimization further errors and uncertainties are encountered. Both
µ µ µ the input and output are corrupted by noise.
where respective partial derivative is given by There are estimation and optimization methods that cope with sys-
tems having noise on both the input and the output signals. One example
2 3 is the total least-squares-method described in [de Groen, 1996], where the
−σ + 2 C0xσ x pσ x p − 1 C0xσ x
if pσ x p < σ ○x orthogonal distance between the data points and the adopted function is
e x 2
= 3 µ 9 (µ ) (7.16)
C0x minimized. In [Carlson and Gerdes, 2003] this method is applied estimat-
0 otherwise
ing the braking stiffness and rolling radius of a tire. Image-processing
methods treat similar problems when locating obstacles in a picture. The
and most probable incline, shape or curvature of an edge are to be identified
2 3
out of a number of points with the correct color. Methods for this have to
− 1 C0xσ x pσ x p + 2 ( C0xσ x )
e
if pσ x p < σ ○x deal with noise in both directions.
= 3 µ2 27 µ3 (7.17) An obvious effect of the use of storage-bins is that the noise in the two
µ −sign(σ ) otherwise signals are differently treated. The signal deciding which bin to activate,
x
150 151
Chapter 7. Road Friction Estimation 7.3 Tire Parameter Estimation
normalized tire force are created with realistic assumptions on magnitude Figure 7.17 that the Y-bins consisting of the [ xi′′ , yi′′ ]-values, with y as
and noise variance. Since the effect of the bin-storage is dependent on the leading signal, are averaged in the x-direction and gives a better estimate,
shape of the tire curve, two situations, that reflect the tire behavior at but the X -bins ([ xi′ , yi′ ]), with x as leading signal, reinforces the mismatch
low respective high slips are analysed. In the low slip region the linear and shows an even more biased result. Since σ ex > σ ey a better estimate
slope of the curve is to be estimated. The high slip region is in the interval of the slope can be reached by using the relation xm = ym / k̂ + ex for the
around the peak force where the tire force is saturated and the level of a least-squares estimation, see right-most P 2 plot 2in Figure
P 7.17. The result
horizontal line will be estimated. is a less biased estimate, k̂ = ( k2 xi + Nσ ey)/( k xi2 ) using all data
In the first case, shown in Figure 7.17, the y-signal is obtained as points. Minimizing the error in the x-direction for the bin-values do not
ym = kx + ey and the x-signal as xm = x + ex . The deviations are set largely change the result for the Y-bins, since they already are derived as
to σ ex = 0.25 and σ ey = 0.0125. The value of σ ex depends on the wheel an average in the x-direction. The estimation from X -bins are, however,
speed magnitude and quality and the choice of realistic variance of ey enhanced but still not as good as using the Y-bins.
depends on how y is derived. In this case the variance roughly corresponds The estimation of the friction level at high slip is illustrated in Figure
to obtaining y from a filtered acceleration signal. If the engine torque is 7.18. Here, it is only possible to minimise the error in the y-direction. Since
used, as described in Section 7.4 σ ey decreases further. Using a force sensor the brush-model is not strictly monotone in this area the Y-bins can not
similar to the equipment used by VTI, see Section 7.2 would give a larger be placed such that they resemble the tire model. Therefore the X -bins
noise level. have to be used in this region.
The x-values for generating the data are 300 points evenly distributed As a result of this discussion the weight in the cost-function in the real
152 153
Chapter 7. Road Friction Estimation 7.4 Algorithm Implementation
0 if
n < Nlow
n− N
low
wi (n) = if n < Nhigh (7.18)
N high − Nlow
1 otherwise
Figure 7.18 Illustration of the bin classification at high slip. The bin-values are The variance measure can be used to eliminate outliers, for example
here used to estimate the level of a straight horizontal line. The left plot shows the by allowing only new samples that lies within a variance-dependent range
storage of [ x′i , y′i ] and the right plot [ x′′i , y′′i ]. from the current estimate. The effect of outlying bin-values can also be
reduced by two steps in the optimization. After a successful optimisation
the residuals of all bin-values (7.10) are analysed and any bin having a
implementation should be reduced on the X -bins (related to the slip) for
residual larger than a factor times the standard deviation of all residuals
low slip.
will have its weight reduced. Care must be taken so that such mechanisms
do not interfere with change detection.
Time varying tire parameters It is assumed to this point that the
estimated function Ψ is time invariant. The tire parameters are assumed
to vary in two different manners. Either the variation is small and slow,
signifying temperature variations during driving, tire wear etc. or they 7.4 Algorithm Implementation
are changing abruptly as when the vehicle rolls in on a different surface.
If Ψ depends on time-varying parameters, θ(t), it is, in general, diffi- The algorithm was implemented in a dSPACE AutoBox using code gen-
cult to maintain the storage-bin contents in a consistent state. But com- erated by Real-Time-Workshop from a Simulink model for use in a front-
bining the low-pass filter in the storage-bins with freshness information wheel driven Volvo S40. The data-bin storage and the parameter opti-
it can be guaranteed that the latest updated bins are most significant in mization were directly coded in C. The dSPACE-box contained a 1 GHz
the optimization. Also old data not necessary for the persistence of the PowerPC with the possibilities of floating-number operations and the pos-
excitation can be disregarded. sibility of data acquisition at 1 kHz with an A/D-conversion of 16 bit res-
If θ(t) changes at a discrete event in time, the changes can be detected olution. The sampling time for the estimator code is 10 ms. This section
by using information on bin variance or from the parameter optimization. shortly describes the implementation and the computations required to
A sudden change of θ(t) at time t1 such that θ(t) = θ0 for t < t1 and derive the signals used by the friction estimator. The algorithm in this
θ(t) = θ1 for t ≥ t1 will lead to a corresponding increase in the estimated implementation is only active during acceleration, since the required ac-
variance v in the storage-bins at the present operating point. This may be curacy on the vehicle speed estimation was not reached by other means
used to trigger operations, e.g. a reset in the oldest updated storage-bins, than using undriven or unbraked wheels. The implementation is only in-
to be performed at a change detection. Suddenly occurring inconsistencies tended for straight driving without any cornering.
in the bin contents may also lead to a corresponding increase of the cost
function (7.5) in the parameter optimization, which may also be used as Implementation structure
a trigger. Other sensors available on the vehicle can also be employed for The model structure is described in Figure 7.19. The input signals to the
this purpose. Starting the windscreen wipers may, for example, be one algorithm, the wheel speeds, the engine torque and rotation speed are
154 155
Chapter 7. Road Friction Estimation 7.4 Algorithm Implementation
already available signals in the car. Hence, no extra sensors are needed.
Stiffness
Friction
The operations required to derive the normalized tire force and the slip
that are the inputs to the estimator are described in the following.
Parameter
Estimation
Algorithm
Input signal conditioning
DataBins
the speed of the undriven rear wheels.
σx
fx
w
Vertical force estimator The vertical wheel load, Fz, on each wheel
Data-Bin
Storage
has to be determined to calculate the normalized tire force and the rolling
wheel-radius. The tire load is calculated from the force and moment equi-
librium of the car, concerning only longitudinal motion changes.
σx
2( Fz f + Fzr ) − m = 0 (7.19a)
Fx
fx
2Fzr l − ma x h − ml f = 0 (7.19b)
Fz
Vertical Force
where l is the length between the wheel axles and l f is the horizontal
Driving force
Wheel Slip
Normalized
Estimator
Estimator
Estimator
Estimator
Force
distance from the front axle to the vehicle center of gravity. The height of
the latter is denoted by h.
Gear estimator The different gear ratios of the gear-box are stored in a
ear
ax
δR
table. By comparing the relation between the wheel speed and the engine
ω
rotation the most appropriate ratio is chosen.
Long itudinal
Acceleration
Correction
Estimator
Estimator
Radius
Wheel
Gear
Driving force estimator The tire force is generated by multiplying the
Teng
engine torque with the actual gear ratio and the approximate effective
radius, given as a vehicle parameter. Energy losses in the power line are
ω eng
accounted for.
ω
Condtioning
Input Signal
Normalized force estimator The force input to the friction estima-
tor is derived by normalizing the tire force on the front wheels with the
corresponding vertical tire load.
ω eng
Teng
ω
that the rolling radius, R e , of each wheel is accurately estimated. The
rolling radius affects the slip calculation and a deviation from its nominal
value causes a horizontal shift on the force-slip relation, such that it do not Figure 7.19 Schematic block diagram of the implementation of the road friction
estimator.
cross the origin. There are many factors that affect R e and the deviation,
δ R, has to be continuously updated by comparing the wheel speeds of
each tire with a reference speed when no forces are transmitted by the
156 157
Chapter 7. Road Friction Estimation 7.5 Results
tire. If the periods of free rolling are not sufficient the estimation can Implementation of the friction estimator
be complemented by adding one extra parameter, denoting the horizontal The bins in the storage algorithm are filled with slip and force values
shift, in the optimization routines. The resulting Jacobian is derived in according to the description in Section 7.3. The x̂-values are replaced by
Appendix C. measurements of the slip σ x and ŷ-values are replaced by the normalized
Since the undriven rear wheels are used as reference for the vehicle tire force. Consequently the X - and Y-intervals are refereed to as S- and
speed, only the differences in-between the radii of each tire are necessary F-bins. The estimator implementation reported in this thesis contains no
estimates for the slip computation. The deviations, δ R, are derived by change-detection. The impact of outliers is reduced by (7.18). From the
comparing rotational speed of each wheel, ω i , with the rotational speed of experiences derived in the discussion concerning the reliability of the S-
the front-left wheel, ω f l and computing correction factors for the wheels and F-bin values, the weights corresponding to S-bins with pσ x p < K s are
as set to zero to enhance the estimation of the tire stiffness at low slip.
δ Ri = δ Ri + K ( R e f l ω f l − ( R ei )ω i ) (7.20) The parameter estimation routine works as described in Section 7.3
with one iteration per sample. Some rules are set to ensure that the bins
where R ei = Rnom (1 + δ Ri + ∆ Ri (⋅)) is the rolling radius for the wheel i. contain enough information for a valid estimate. Attempting to estimate
Known dependencies on other factors, such as the vertical load on the tire the friction based on an insufficient amount of data may produce com-
can be included in ∆ R(⋅). Correction is only performed when no horizontal pletely erroneous results and conclusions about the reliability. The rules
forces are transmitted by the tire. for the estimation scheme are made clear in Algorithm 1. The condition
In many cases the dependency of the vertical load on the rolling radius on Line 1 states that there must be reliable values (at least Nlow mea-
is small and almost neglectable. Investigations performed in the project surement points) in K 1 number of bins for estimation of the tire stiffness.
have shown that incorporating the assumed load effects on the radius, Before trying to estimate the friction, there must be even more reliable
reduces the braking stiffness estimate on asphalt about 25%, during ac- bins ( K 2 > K 1 ) together with a minimum demand on the magnitude of
celeration. Measurements performed within the project, see Figure 7.15, the lowest available slip and force value in the bins. Implicitly, this also is
and tests in the Pirelli laboratory [Pirelli, 2005] show an approximate a check of the ability to invert the matrices H T H and Φ T Φ , even if each
relation between the increment of rolling radius and the vertical load as determinant has to checked before any inversion procedure.
158 159
Chapter 7. Road Friction Estimation 7.5 Results
Algorithm 1 Pseudo-code implementation of the rules of the friction es- between asphalt and snow at that period. It should, however, be noted that
timator. the slip values in the bins are very low and the high sensitivity towards
Old
Require: Ω ∈ {Ω i = [σ xi , f xi , wi ]}, Ĉ0x , µ̂ Old wheel speed noise at low slip is one factor for the discrepancy. By a more
Produces: Ĉ0x , µ̂ conservative tuning, for example using Kσ > 0.005 (see Algorithm 1) no
1: if Ω (wi > 0) ≥ K 1 then estimation would be available at the actual time slot and the misjudge-
P
L1
2: Calculate C0x using the Least-Squares method as described in (7.12) with ment is avoided. Also the disturbance on the braking stiffness estimation
φi = −σ xi in this period, invoked by the start of the Gauss-Newton iteration would
3: Compute the corresponding cost function, Jcx from (7.9) then be prevented. The plots of the bin-contents in Figure 7.21 explains
4: if ( Ω (wi > 0) ≥ K 2 ) & (max(p f xi p) > K f ) & (max(pσ xi p) > Kσ ) then
P
L2
the reason for the erroneous estimation at t=6.7 s. The group of bins sur-
5: Estimate C0x , µ L2 using the Least-Squares method as described in (7.12)
rounded by the dashed circle have a disturbance that gives an apparent
with φi = [−σ xi σ xi pσ xi p]T
6: Compute the corresponding cost function, Jcxµ from (7.9)
curvature in the tire-force relation. It is also clear from the right-most
7: if µ L2 ∈ [0, K µ ] then plot that this group deviates from the others. The linearity of the bin-
8: if Jcxµ < K j Jcx then values in the right-most plot also shows the difficulty of estimating the
9: Estimate C0x GN
, µ GN by one iteration of the Gauss-Newton algorithm friction at the low tire-force utilization. The remarkable difference of the
(7.14). tire stiffness compared to previously reported VTI-measurement will be
10: Ĉ0x , µ̂ Z C0x GN
, µ GN discussed later.
11: else
L2
12: Ĉ0x , µ̂ Z C0x , µ̂ L2 Estimation results on snow
13: end if In the test on snow, see Figure 7.22, the estimator stabilizes at t (7.7 s.
14: else
L1 The maximal f x -value in any bin at that time is below 0.2, approximately
15: Ĉ0x , µ̂ Z C0x , µ̂ Old
50% of the available friction. The precision of the delivered friction esti-
16: end if
17: else mate is ±0.2 at that time, but improves with the increase of the available
18: Ĉ0x , µ̂ Z C0x L1
, µ̂ Old force measurements. The underestimation of the friction, explained by
19: end if the error of the brush model on snow, also discussed in Section 7.2, is
20: else an accuracy problem, limiting the precision of the method. The mismatch
21: Ĉ0x , µ̂ Z C0x Old
, µ̂ Old (Algorithm inactive) between the model and the real tire behavior is clear from the plots in
22: end if Figure 7.23. It is, however, obvious that the difference in tire behavior is
clearly distinguished from asphalt at low tire force excitation.
160 161
Chapter 7. Road Friction Estimation 7.5 Results
25
20
0.4
0.2
x
f
0
Slip [%]
−0.5
Figure 7.21 Illustration of the bin contents from the winter tire test on asphalt,
shown together with the estimated tire characteristics for the right wheel during
−1
the time instants 6.7 s (left plot) and 10 s (right plot). The size of the marker is
60 relative to the weight of the bin in the cost function. Only F-bins are present in the
plots, since the weight of the S-bins are reduced for slip below K s = 0.02. The group
C0x
40 of points surrounded by the dashed circle contains a disturbance that enforces the
bad estimation during the time period t =6.5–7.1 s.
20
1.5
1
The estimations performed by the least-squares method are indepen-
µ
0.5
dent of the prior estimations. The choice of initial value on the estimate,
0
2 3 4 5 6 7 8 9 10 Cxinit and µ init , at the algorithm initialization is therefore small. But be-
Time [s]
fore starting the Gauss-Newton solver it should be secured that the ac-
tual estimates are in a valid range, not to risk an unstable iteration.
Figure 7.20 Plot of measured vehicle velocity, the calculated normalized tire force,
and wheel slip and resulting estimations of the normalized braking stiffness and The choice of K j < 1 guarantees that a curvature is discerned from the
friction coefficient on asphalt. Solid line denotes left wheel and dashed line right available bin-values, before starting Gauss-Newton iterations. The least-
wheel. squares algorithm on Line 5 in Algorithm 1 should then have provided
descent initial estimates. The parameter K j determines how much the
cost function should decrease when using a second order relation instead
Discussions of a first order relation. If the curvature is small the estimation of two
As mentioned previously the algorithm is tuned to be fast and rather parameters in a linear relation, see (7.11) is sufficient. A difficulty with
volatile in the presented results. The tuning parameters are set according the least-squares estimation is that negative values on the friction can be
to Table 7.5. In a real application it is important to tune it slower and more obtained, if a negative curvature is detected. This is viewed as an tempo-
robust towards noise and disturbances, since it is favourable to prevent ral error and only the stiffness from the one-parameter linear estimation
the deliverance of an erroneous estimate. The algorithm should, however, should be delivered in such case.
be fast enough not to unnecessarily delay any reliable prediction, even on The quickness of the algorithm to adapt to a changing condition is
ice. The parameters deciding when the storage-bins contain sufficient data determined by Ni , which should be set large enough such that tempo-
for friction estimation are K 2 , Kσ , and K f . Where K 2 describes the lowest ral disturbances do not have a too large effect the estimation, but small
number of bins containing reliable force and slip values and Kσ and K f enough such that smaller deviations in tire characteristics, i.e. due to tem-
determine the minimum requirement of the magnitude of the available perature changes, can be correctly accounted for. Larger changes should
slip and force values, respectively. These parameters denote fix limits on be detected by a change-detection algorithm, enforcing a complete reini-
the bin-contents, independent on the road foundation. If the data contents tialisation of the bin-memory. The estimation robustness is adventured
are below these limits only the normalized braking stiffness is estimated. if some bins contains measurements from new conditions and other from
162 163
Chapter 7. Road Friction Estimation 7.5 Results
0.5 0.5
15
0.2 0.2
0 0.1 0.1
0
0 0
Slip [%]
−20 Figure 7.23 Bin contents from measurement of the winter tire on snow together
with estimated tire characteristics for the right wheel during the time instants 8 s
40
(left plot) and 12 s (right plot). The size of the marker denote the weight of the bin
in the cost function. The F-bins are marked with ’o’ and the s-bins with ’*’.
0x
20
C
1.5 by the applied acceleration torque. Simulations show, however, that this
1 has a small and neglectable effect. The derivation of the normalized force
µ
0.5 introduces error sources, as efficiency losses, gear ratios, and uncertain-
0
5 6 7 8 9 10 11 12 ties in torque measurements and load transfer etc that might affect the
Time [s] results. A distinct difference between the measurements are that the VTI
measures while braking, when the estimator implemented in the personal
Figure 7.22 Measured signals and result of friction estimation of the winter tire
on snow. Solid line denotes left wheel and dashed line right wheel.
car only uses driving data. The potential error sources will be further pen-
etrated in the future work. Any physical reason for the tire to be stiffer
on the car relative the test vehicle has not yet been found.
It is very important that the tire radius is accurately determined. For
older conditions. foundations such as e.g. asphalt that gives a high stiffness on the tire the
The parameters Nlow and Nhigh are limits for the reliability of the bin- accuracy on δ R/ Rnom must be lower than one per thousand. It is not en-
value, stating that a certain number of data points are required for full tirely verified that it is possible to obtain this high accuracy. As previously
weight of a bin. It is logically to set Nhigh = Ni , since Nhigh points out that mentioned, the radius estimator can be combined with the estimation of
the bin-value has full confidence. the horizontal shift as a third parameter in the Gauss-Newton algorithm.
There is a large difference in the normalized braking stiffness between The expression for the extended Jacobian is described in Appendix C.
the estimation on the car and the VTI measurements on asphalt. The dif- In Figure 7.26 the resulting estimates are shown for the measurement
ference is almost a factor two, 50 compared to 25. One reason for the stiff- case from Figure 7.20 when starting with incorrect wheel radius, but cor-
ness deviation is the load dependency of the wheel radius. The test wheel recting its value by inclusion of the horizontal shift as an estimated pa-
on the VTI-truck is not exposed to any load transfer during the force ramp, rameter in the Gauss-Newton iteration. It is clear from the results that
which is inevitable for the personal car. Without adjustments for the load the estimates stabilize and that the stiffness value is identical to the re-
dependency, the stiffness would be about 70 and a further increase of the sult with correct and fixed wheel radius. This ensures that the horizontal
impact of load dependency, by raising ki , can not be motivated. Another shift is correctly estimated. The period of an uncertain friction estimation
explanation might be an existing difference in the motion of the front and is, however, longer which indicates, not surprisingly, that the estimation
rear axle, e. g. due to longitudinal deflection of the suspension caused of three parameters makes the algorithm more sensitive towards noise
164 165
Chapter 7. Road Friction Estimation 7.6 Conclusions
10 0.16 0.16
0.14 0.14
0.08 0.08
0.1
x
f
0.06 0.06
0.04 0.04
0
0.02 0.02
0 0 0
−3 −2.5 −2 −1.5 −1 −0.5 0 −35 −30 −25 −20 −15 −10 −5 0
Slip [%]
Figure 7.25 Bin contents from measurement of the winter tire on ice together
−40
with the estimated tire characteristics for the right wheel during the time instants
40 5.9 s (left plot) and 9 s (right plot). The size of the marker denote the weight of the
bin in the cost function. The F-bins are marked with ’o’ and the s-bins with ’*’.
C0x
20
100
0
1.5
Cx
50
1
µ
0.5 0
0
4 5 6 7 8 9
Time [s] 1.5
µ
Figure 7.24 Measured signals and result of friction estimation of the winter tire 0.5
on ice. Solid line denotes left wheel and dashed line right wheel.
0
0
and errors, but on the other hand gives a better estimate if the data ex-
S [%]
citation is sufficient and an accurate wheel radius is hard to obtain by −0.5
h
other means.
−1
2 3 4 5 6 7 8 9 10
Time [s]
7.6 Conclusions Figure 7.26 Plot of parameter estimation with the horizontal shift as as third es-
timated factor. The measurement sequence is the same as one shown in Figure 7.20.
This chapter has described a new type of friction estimator based on force
and slip measurements. The tire behavior is assumed to follow the brush
model prescribing the force and slip relation, depending on the braking The algorithm switches between different optimization methods de-
stiffness of the tire and the road-tire friction. From the measurements pending on the contents in the bins. If only low slip and forces are avail-
these two parameters can be derived. A major invention of the proposed able the braking stiffness is estimated as the sole parameter using the
method is the way to collect the sampled measurements into bins, such least-squares method. For slightly higher values of the signals both the
that the available data used for optimization is evenly spread and weighted friction and the stiffness are derived by the same method. When the curva-
along the force and slip axis. ture of the force-slip relation becomes evident the Gauss-Newton method
166 167
Chapter 7. Road Friction Estimation 7.6 Conclusions
which is before the wheel starts its locking phase and before a high ac-
Table 7.5 List of tuning parameter for the actual implementation of road friction
estimator celeration/retardation can be noticed. The accuracy on asphalt has not
been further validated, since larger accelerations, than could be achieved
Parameter Description Nom. value during the measurement occasion, are needed to reach sufficient friction
utilization. The output from the estimator, in that test, only indicates
Cxinit Initial value on Cx 40
that there are much more friction available. One of the task for the fu-
µ init Initial value on µ 0.5
ture work is to optimize the signal quality and estimation performance to
Ks Slip limit for using S-bin 0.02 reduce the need of friction utilization for a sufficiently accurate friction
K1 Required filled bins for C0x - 3 determination. The estimator has been validated through force excitation
estimation as an acceleration ramp. Its structure has been constructed to effectively
K2 Required filled bins for 6 cope with weighting problems when the tire force is kept constant on a
C0x , µ -estimation certain level for a long time or varying in a way such that the data is not
Kj Residual difference to en- 1 well spread in the force-slip plane. The accuracy for estimation during
able Gauss-Newton algo- this kind of unfavourable excitations needs further validation. The esti-
rithm mator implementation is so far restricted to acceleration with two-wheel
Kσ Slip limit for enable µ - 0 driven vehicles. The major reason for this restriction is the insufficient
estimation accuracy of the estimate vehicle reference velocity if not measured by the
Kf Limit on normalized force to 0 undriven wheel-rotations. Further work incorporates enlargement of the
enable µ -estimation usable area of the estimator to work for both braking and acceleration
kl Load dependence on rolling 0.7 ⋅ 10−6 during simultaneous cornering.
radius
Nlow Threshold for measurement 2
points in bin
Nhigh Number of measurement 20
points in bin to maximize
weight
Ns Number of slip bins 150
Smax Maximum slip bin value 0.5
Nf Number of force bins 150
Fmax Maximum force bin value 1.2
Ni Number of points in bin 100
memory
is used for better accuracy. The evaluation of the result has focused on
the Gauss-Newton method.
The estimator works well in the presented situations. From no prior
information it distinguishes between snow and asphalt for an available
normalized tire force below 0.25. The estimate on snow available at that
time is lower than expected, which depends on that the tire behavior
on snow slightly differs from the model. The estimator detects and pre-
scribes the low friction on ice when the tire force reaches its peak level
168 169
The tire model was successfully implemented in Matlab and Modelica
8 code and tested together with the VehicleDynamics Library in the multi-
body simulation environment, Dymola. The implementation was verified
to be well-working and computationally sound on both platforms. Good re-
sults are obtained in validation with the available empirical data. Experi-
mental validation indicates that the proposed model and previous models
Conclusions perform similarly. The efficiency in computational means compared to
other models has not yet been evaluated. The presented model is advan-
tageous in that it combines the data-fitting accuracy of empirical pure slip
models with the stringency of physical modeling. This claims descent be-
havior, according to the physical constraints, in the defined working area
without the need for calibration, at the cost of quite large expressions. The
The thesis has presented research performed mainly in the related areas methodology to derive the scale-factors based on first principles makes it
of tire modeling and friction estimation. The theory behind the brush tire straight forward to include new functionality to the model.
model developed by the pioneers in tire modeling has been a central theme The wheel speed signal is a most important signal in many system
throughout the work. The thesis contains an extensive review of the brush applications. Based on problems observed during measurements, a few
model including the effects of camber. It also presents a validation of the ways to reduce noise from the signal are discussed in the thesis. One
model towards measurements and an evaluation of the variation of the method to compensate the wheel-speed signal for disturbances caused by
included parameters, due changes in conditional factors, such as load and axle and suspension deflections during transients in brake applications is
road surface. proposed. The disturbance is predicted by a first order feedforward filter
A new method to derive the tire forces for simultaneous braking, cor- of the brake torque signal. From the validation it can be seen that the
nering and camber, by combining empirical models for pure braking, cor- signal error can, by this simple means be reduced by at least 50%. This
nering and cambering has been presented. Based on brush-model me- is probably sufficient to significantly increase the performance of vehicle
chanics, the combined-slip forces may be described by a scaling of corre- systems relying on the wheel-speed signal during brake-force transients.
sponding empirical pure-slip forces. The way to derive the scale factors by The thesis also describes the development of a new type of friction
dividing the expression for the combined slip force with the pure slip force estimator based on the assumption that the tire behavior follows the brush
in the appropriate direction is an unique approach. The generation of the model. Since the brush model predicts the tire characteristic, as depending
camber force has been described and included in a simple, but physically on the braking stiffness and friction coefficient, these parameters can be
motivated manner. The pure-slip tire model can be given as a empirical estimated from the force and slip measurements. Experimental data has
model or as raw tabular data, as long as the horizontal and vertical shift been collected and evaluated to ensure the validity of the brush model
are zero, i.e. that there is no tire force for zero slip. during certain conditions. The shortcomings and limitations of the model
A major aim of the semi-empirical modeling has been to extract as accuracy are discussed. A major invention in the proposed method is the
much information from the available pure-slip curve as possible. The phys- way of collecting the measurements into bins, such that the available data
ical approach has made it possible to include velocity dependency and used for optimization is evenly weighted along the force and slip axis. The
conversion between braking and driving data in the model. The proposed least-squares and the Gauss-Newton methods are used for finding the
model is simple to use, since it does not introduce any new parameters most accurate values of the stiffness and friction.
and all necessary information is given by the chosen pure-slip model. The estimator has been implemented and validated on a passenger
A dynamic extension of the steady-state combined-slip model is pro- car. For a sequence in the form of an acceleration ramp it has shown to
posed, based on qualitatively realistic tire relaxation behaviour. The model work well in the tested situations.
includes cross-couplings between longitudinal and lateral dynamics, that As expected, the amount of friction utilization is strongly connected
are not commonly found in other models of similar complexity. The issue to the accuracy of the estimation and the usability of the method will
of stability is handled by including a physically motivated damping. strongly relate to the requirement regarding accuracy and possibilities
for utilization.
170 171
Chapter 8. Conclusions 8.1 Ideas of Further Research
The structure of the estimator is developed to effectively cope with be enlarged to enable friction estimation during cornering, when a
weighting problems when the tire force is kept constant on a certain level lateral slip affects the longitudinal tire properties.
for a long time or varying in a way such that the data is not well spread
in the force-slip plane. The accuracy for estimation during this kind of An important conclusion of a research project is often how the work
unfavourable excitations needs further validation. It can be concluded that should proceed in the development of an idea or a product. Since the pos-
this way of approaching the problem of friction estimation is promising. sible fields of improvements in vehicle related issues have been expanded,
by the newly available techniques and instrumentation possibilities, it is
important to put the efforts where they have the best effect. Hopefully,
8.1 Ideas of Further Research this thesis may serve as a guide in this purpose for the related topics.
All of the presented methods and ideas can be further developed, analysed,
and tested. The particular needs have been described in the respective
sections, from where the most urgent and important issues are concluded
in the following list:
• The wheel speed signal is essential for many vehicle systems and
further work to improve the quality of that signal is important. The
proposals mentioned in the thesis have to be further evaluated and
developed by testing in real implementations.
• The proposed semi-empirical tire-model is, in its present state, use-
ful for a number of applications, but the addition of effects of turn-
slip and flexible carcass would further enhance the applicability. The
physical foundation of the model is expected to make this possible
with reasonable effort. Further validation of, particularly, the effects
of camber are need.
• It has to be emphasised that the proposed friction estimator is de-
veloped in an ongoing project. The presented result should therefore
be seen as the first glimpse of the possibilities to use this type of
methodology for this purpose. A large amount of work is still required
to achieve good response of the estimate when changing surface and
to ensure the ability of catching slow variations of the parameters.
Solutions for change detection and outlier management have to be
implemented and calibrated, such that a maximum of robustness
and accuracy is achieved. Improvement of signal quality and esti-
mator performance to minimize the need of friction utilization are
other needs. Further, a lot of tests and validations are necessary
really evaluate the performance in action and to reveal not yet dis-
covered weak points.
• A demand for developing a slip-based friction estimator for use dur-
ing braking or for “All Wheel Drive” vehicles is to have a good es-
timate of the vehicle velocity or to succeed in removing this depen-
dence from the estimator structure. The area of usability is also to
172 173
Casselgren, J., M. Sjödahl, and M. S. Sara Woxneryd (2007): “Classifica-
9 tion of road conditions – to improve safety.” In 11th International Fo-
rum on Advanced Microsystems for Automotive Applications. Berlin.
To be published.
Clark, S. K., R. N. Dodge, and G. H. Nybakken (1971): “An evaluation of
string theory for the prediction of dynamic tire properties using scale
Bibliography model aircraft tires.” Technical Report. University of Michigan.
Colmis AB (2006): “Colmis proving ground, Arjeplog, Sweden.” Internet.
http://www.colmis.com.
Continental (2003): “History of the tire.” Internet. http://www.conti-
online.com/generator/www/us/en/continental/transport/themes/
Analog Devices (2001): Wheel speed sensor for ABS systems, AD22157. about_continental/tire_history/tire_history_en.html.
Åström, K. J. (1970): Introduction to Stochastic Control Theory. Academic de Groen, P. (1996): “An introduction to total least squares.” Nieuw Archief
Press. voor Wiskunde, pp. 237–253.
Åström, K. J. and B. Wittenmark (1994): Adaptiv Control. Addison-Wesley
Do, M.-T. and Y. Delanne (2003): “Prediction of tire/wet road friction and
Pub Co.
its variation with speed from road macro- and microtexture, and tire-
Bakker, E., L. Nyborg, and H. B. Pacejka (1987): “Tyre modelling for use related properties.” Technical Report. French Public Works Research
in vehicle dynamics studies.” SAE Technical Paper 870421. Laboratory, France.
Bakker, E., H. B. Pacejka, and L. Lidner (1989): “A new tire model with an Dugoff, H., P. S. Fancher, and L. Segel (1969): “Tire performance char-
application in vehicle dynamics studies.” SAE Technical Paper 890087. acteristics affecting vehicle response to steering and braking control
Bayle, P., J. F. Forissier, and S. Lafon (1993): “A new tire model for vehicle inputs. final report.” Technical Report. Highway Safety Research In-
dynamics simulations.” Automotive Technology International, pp. 193– stitute, Ann Arbor, Michigan.
198. Dynasim (2006): “Dymola, dynamic modeling laboratory.” Internet.
Bernard, J. E., P. S. Fancher, R. Gupta, H. Moncarz, and L. Segel (1975): http://www.dynasim.se/.
“Vehicle-in-use limit performance and tire factors - the tire in use. Eichhorn, U. and J. Roth (1992): “Prediction and monitoring of tyre/road
appendix a, b. c.” Technical Report UM-HSRI-PF-75-1-2. Highway friction.” In XXIV FISITA Congress: Safety, the Vehicle, and the Road,
Safety Research Institute, Ann Arbor, Michigan. pp. 67–74. London.
Böhm, F. and H.-P. Willumeit, Eds. (1996): Tyre Models for Vehicle Ellis, J. R. (1994): Vehicle Handling Dynamics. Mechanical Engineering
Dynamics Analysis. Swets & Zeitlinger Publishers. Supplement to Publications Limited, London.
Vehicle System Dynamics Volume 27.
Fancher, P. S., L. Segel, C. C. MacAdam, and H. Pacejka (1972):
Böiers, L. C. (2004): Lectures on Optimization. Center for Mathematical “Tire traction grading procedures as derived from the maneuvering
Science, Lund University. characteristics of a tire-vehicle system. volume i and ii. final report.”
Brach, R. M. and R. M. Brach (2000): “Modeling combined braking and Technical Report HSRI-71-129. Highway Safety Research Institute,
steering tire forces.” SAE Technical Paper 2000-01-0357. Ann Arbor, Michigan.
Carlson, C. and C. Gerdes (2003): “Nonlinear estimation of longitudinal Gäfvert, M. (2003): Topics in Modeling, Control, and Implementa-
tire slip under several driving conditions.” In Proceedings of the tions in Automotive Systems. PhD thesis ISRN LUTFD2/TFRT--
American Control Conference, vol. 6, pp. 4975–4980. LUTFD2/TFRT–1066–SE--SE, Lund University.
174 175
Chapter 9. Bibliography
Gäfvert, M. and J. Svendenius (2003): “Construction of semi-empirical tire Hirschberg, W., G. Rill, and H. Weinfurter (2002): “User-appropriate
models for combined slip.” Technical Report ISRN LUTFD2/TFRT-- tyre-modelling for vehicle dynamics in standard and limit situations.”
7606--SE. Department of Automatic Control, Lund University, Sweden. Vehicle System Dynamics, 38:2, pp. pp. 103–125.
Gäfvert, M. and J. Svendenius (2005): “A novel semi-empirical tire model Isermann, R., P. Scheerer, O. Nelles, and R. Schwarz (2000): “Method for
for combined slip.” Vehicle System Dynamics, 43:5. compensating variations of a wheel speed sensor.” Patent no. 6,446,018.
Assignee: Continental Teves.
Gäfvert, M., J. Svendenius, and J. Andreasson (2006): “Implementation
and application of a semi-empirical tire-model in multi-body simula- ISO 8855 (1991): “Road vehicles — Vehicle dynamics and road-holding
tion of vehicle handling.” In Proceedings of the 8th International Sym- ability — Vocabulary.”
posium on Advanced Vehicle Control. Taipei, Taiwan.
Johansson, R. (2002): System Modeling and Identification. Prentice Hall.
Gelb, A., J. Kasper, R. Nash, C. Price, and A. Sutherland (1977): Applied
Kageyama, I. and S. Kuwahara (2002): “A study on tire modeling for
Optimal Estimation. The MIT press.
camber thrust and camber torque.” JSAE Review Paper No. 20024252.
Gim, G. and P. E. Nikravesh (1991): “An analytical model of pneumatic
Kiencke, U. and L. Nielsen (2000): Automotive Control Systems: For
tires for vehicle dynamics simulations. Part 2: Comprehensive slips.”
Engine, Driveline, and Vehicle. Springer-Verlag.
International Journal of Vehicle Design, 12:1, pp. 19–39.
Kim, S. and A. Savkoor (1996): “The contact problem of in-plane rolling of
Gipser, M. (2005): “FTire homepage.” Internet. http://www.ftire.com.
tires on a flat road.” In Böhm and Willumeit, Eds., Tyre Models for Ve-
Gothie, M., T. Parry, and P. Roe (2001): “The relative influence of hicle Dynamic Analysis, pp. 189–206. Swets & Zeitlinger. Supplement
the parameters affecting road surface friction.” In 2nd International to Vehicle System Dynamics Volume 27.
Colloquium on Vehicle Tyre Road Interaction. Florence, Italy.
Levy, G. and N. Fangeat (2004): “Automatic stability control system for
Goyal, S. (1989): Planar Sliding of a Rigid Body with Dry Friction: Limit a vehicle using an invariant charaterizing any tire.” Patent no US
Surfaces and Dynamics of Motion. PhD thesis, Cornell University, 2004/0032165 A1.
Department of Mechanical Engineering.
Ljung, L. and T. Söderström (1983): Theory and Practice of Recursive
Guo, K., Y. Zhuang, S. Chen, and L. Willlam (2006): “Experimental Identification. The MIT Press.
research on friction of vehicle tire rubber.” Frontiers of Mechanical
Lugner, P. and P. Mittermayr (1991): “A measurement based tyre char-
Engineering in China, 1:1, pp. 14–20. DOI 10.1007/s11465-005-0001-z.
acteristics approximation.” In Pacejka, Ed., Tyre Models for Vehicle
Gustafsson, F. (1997): “Slip-based tire-road friction estimation.” Automat- Dynamics Analysis, pp. 127–144. Swets & Zeitlinger. Supplement to
ica, 33:6, pp. 1087–1099. Vehicle System Dynamics Volume 21.
Gustafsson, F., J. Nilsson, and T. Gustavsson (2006a): “Method and Matschinsky, W. (1997): Road Vehicle Suspension. John Wiley & Sons,
system of determining the absolute velocity of a vehicle.” Patent no. Inc. ISBN 1-86058-202-8.
EP1642140.
Meschke, G., C. Liu, and H. Mang (1996): “Large strain finite-element
Gustafsson, M., C.-M. Berglund, B. Forslund, and I. Forslund (2006b): analysis of snow.” Journal of Engineering Mechanics, 122:7, pp. 591–
“Effekter av vinterdäck – en kunskapsöversikt.” Technical Report. VTI. 602.
Report 543.
Michelin (2004): “Annual report.” Internet. www.michelin.com.
Hägglund, T. and K. J. Åström (2000): “Supervision of adaptive control
Mizuno, M., H. Sakai, and K. Oyama (2005): “Developement of a tyre force
system.” Automatica, 36, pp. 1171–1180.
model incorporating the influence of the tyre surface.” In Lugner and
Haney, P. (2003): Racing & High Performance Tire: Using Tires to Tune Plöchl, Eds., Tyre Models for Vehicle Dynamic Analysis, pp. 395–402.
for Grip and Balance. SAE. Taylor & Francis. Supplement to Vehicle System Dynamics Volume 43.
176 177
Chapter 9. Bibliography
178 179
Chapter 9. Bibliography
SAE Recommended Practice J670e (1976): “Vehicle dynamics terminol- Sugai, M., K. Asano, and T. Umeno (2003): “Physical quantity estimating
ogy.” apparatus and tire state determining aparatus, and estimating method
of same and determination of same.” Patent no. US 2003/0192375 A1.
Saino, P. (2001): “Snow surface model for tyre performance simulation.”
In 2nd International Colloquium on Vehicle Tyre Road Interaction. Svendenius, J. (2003): “Tire models for use in braking applications.”
Florence, Italy. Licentiate thesis ISRN LUTFD2/TFRT--3232--SE. Department of
Salaani, K., D. Guenther, and G. Heydinger (1999): “Vehicle dynamics Automatic Control, Lund University, Sweden.
modeling for the national advanced driving simulator of a 1997 Jeep
Svendenius, J. (2007a): “Validation of the brush model towards VTI-
Cherokee.” SAE Paper No. 1999-01-0121.
measurement data recorded at Hällered 2005.” Technical Report ISRN
Schmeitz, A., I. Besselink, J. de Hoogh, and H. Nijmeijer (2005): “Ex- LUTFD2/TFRT—7616–SE. Department of Automatic Control, Lund
tending the magic formula and swift tyre model for inflationpres- University, Sweden. RFE-project.
sure changes.” In Wies and Wallentowitz, Eds., Reifen, Fahrwerk,
Fahrbahn, pp. 201–225. Hannover, Germany. Svendenius, J. (2007b): “Validation of the brush model towards VTI-
measurement data recorded in Arjeplog 2006.” Technical Report ISRN
Schofield, B., T. Hägglund, and A. Rantzer (2006): “Vehicle dynamics con- LUTFD2/TFRT—7617–SE. Department of Automatic Control, Lund
trol and controller allocation for rollover prevention.” In Proceedings of University, Sweden. RFE-project.
the IEEE International Conference on Control Applications. Munich,
Germany. Svendenius, J. and M. Gäfvert (2004a): “A brush-model based semi-
empirical tire-model for combined slips.” SAE Technical Paper 2004-
Schuring, D. J., W. Pelz, and M. G. Pottinger (1996): “A model for combined
01-1064.
tire cornering and braking forces.” In Investigations and Analysis in
Vehicle Dynamics and Simulation, pp. 61–83. SAE International. SAE Svendenius, J. and M. Gäfvert (2004b): “A semi-empirical tire-model for
Technical Paper 960180. transient combined-slip forces.” In AVEC ’04.
Schwarz, R. (1999): Rekonstruktion der Bremskraft bei Fahrzeugen mit
Svendenius, J. and M. Gäfvert (2005): “A semi-empirical tire model for
Elektromechanisch Betätigten Radbremsen. PhD thesis, Technische
combined slips including the effects of cambering.” In Lugner and
Universität Darmstadt.
M.Plöchl, Eds., Tyre Models For Vehicle Dynamics Analysis, vol. 43,
Schwarz, R., O. Nelles, P. Scheerer, and R. Isermann (1997a): “Increasing pp. 317–328. Taylor & Francis Group, Glasgow. Supplement to Vehicle
signal accuracy of automotive wheel-speed sensors by on-line learning.” System Dynamics Volume 43.
In Procedings of the American Control Conference, pp. 1131–1135. New
Mexico. Svendenius, J. and M. Gäfvert (2006): “A semi-empirical dynamic tire
model for combined-slip forces.” Vehicle System Dynamics, 44:2,
Schwarz, R., M. Willimowski, R. Isermann, and P. Willimowski (1997b): pp. 189 – 208. Special Issue: AVEC ’04: 7th International Symposium
“Improved wheel speed and slip determination considering influences on Advanced VEhicle Control, 23-27 August 2004 HAN University,
of wheel-suspension dynamics and tire dynamics.” SAE-paper No. Arnhem, The Netherlands.
971117.
Svendenius, J. and B. Wittenmark (2003): “Brush tire model with
Sharp, R. (2004): “Testing and improving a tyre shear force computation
increased flexibility.” In ECC. European Union Control Association,
algorithm.” Journal of Vehicle System Dynamics, 41:3, pp. 223–247.
Cambridge, UK.
Sharp, R. and M. Bettella (2003): “Shear force and moment descriptions
by normalisation of parameters and the "magic formula".” Journal of Thomson, J. (2003): “What we should know about tires.” Internet.
Vehicle System Dynamics, 39:1, pp. 27–56. http://www.jags.org/TechInfo/2001/05May/tires.
Solyom, S., A. Rantzer, and J. Lüdemann (2004): “Synthesis of a model- Thorvald, B. (1998): On truck tyre modelling. PhD thesis ISRN
based tire slip controller.” Vehicle System Dynamics, 41:6, pp. 477–511. LUTFD2/TFRT--KTH/FKT/DA–98/29–SE--SE, KTH.
180 181
Chapter 9. Bibliography
182 183
Appendix A. Nomenclature
184 185
Appendix A. Nomenclature
Misc.
ϒ Force expression (5.10)
Friction Estimation
Λ Normalisation of the normalized slip
(5.15) θ Parameters for estimation
Λz Normalisation of the normalized slip V Cost function (7.9)
(5.23) wi Weight of bin i in cost function
Γ x,y Sliding force scale factor without X Bin intervals, S-bin (slip) (7.6)
friction ellipse limitation (5.14) and Y Bin intervals, F-bin (force) (7.6)
(5.16)
J Jacobian matrix (7.15)
H Hessian matrix (7.13)
λi Forgetting factor for bin i
186 187
Appendix A. Nomenclature
Subscripts
Notation
a
Valid on
F, M , G, µ
Description
Contribution from adhesive part of contact
B
patch
s
F, M , G Contribution from sliding part of contact
s
µ
patch
Static friction
Scale Factors and Slip
k
0
µ
F, M
Kinetic friction
Pure-slip entity
Conversion
0
C Normalized entity, C0x = Cx / Fz
x
Vectors Longitudinal direction
y
Vectors Lateral direction
z
Vectors Vertical direction, positive downwards The derivation of the scale factors is only performed in the lateral direction, but
the procedure for the longitudinal scale factors is similar.
cam
F, M , δ y Contribution from camber
v
F, M , δ y Vehicle entity
ij
v̄, λ , α , F̄, M̄ Wheel identifier, where i = f , r (front or Adhesive slip scale factors
rear) or 1.. number of axles on the vehicle
and j = l, r (left or right).
... Measured quantity Fay,slip(σ x , σ y, γ )
m Gay (σ x , σ y, γ ) =
F0y(σ 0ya )
Superscripts − Cyσ y (1 − ψ (σ x , σ y , γ ))2
=
Notation Valid on Description − Cyσ y (1 − ψ (0, σ y, 0))2 − Fz µ y sin(β )ψ 2 (0, σ y, 0) (3 − 2ψ (0, σ y, 0))
ˆ − Cyσ y (1 − ψ (σ x , σ y , γ ))2
F, M , C, ... Measured or estimated quantity = σ
○ − Cyσ y (1 − ψ (0, σ y, 0))2 − Fz µ yσ ○y σ ○yy ψ (0, σ y, 0) (3 − 2ψ (0, σ y, 0))
λ , α , σ̄ Slip where transition to full sliding occurs
2
∗
λ , α , σ̄ , F, M Point for friction maximum. Corresponds to 3 (1 − ψ (σ x , σ y , γ ))
= (B.1)
subscript crit and max in ISO 8855 ϒ(0, σ y , 0)
¯ Vectors Vector ( x, y, z)
′
v, λ , α Value, measured at the rim under the as- when ψ < 1, otherwise Gay (σ x , σ y , γ ) = 0.
sumption of a flexible carcass
′
σ x , f x , x, y S-bin value (the slip signal is leading) J
′′ ϒ( x, y, z) = ψ 2 ( x, y, z) − 3ψ ( x, y, z) + 3 (B.2)
σ x , f x , x, y F-bin value (the force signal is leading)
reg
σ 0x,y Region-based pure slips, Eq (5.18)
vel Camber scale factor
σ 0x,y Velocity-invariant pure slips, Eq (5.12)
Fcamy (σ x , σ y, γ )
Gcamy (σ x , σ y, γ )) = = 2ψ 3 (σ x , σ y , γ ) − 3ψ 2 (σ x , σ y , γ ) + 1 (B.3)
F0cam (γ )
188 189
Appendix B. Scale Factors and Slip Conversion
Sliding scale factors where ta is the pneumatic trail for the adhesive region defined as
Maz(σ x , σ y, γ )
Fsy(σ x , σ y, γ ) ta (σ x , σ y, γ ) = (B.10)
Gsy (σ x , σ y, γ ) = Fay,slip (σ x , σ y, γ )
F0y(σ 0ys )
− Fz µ y sin(β ′ )ψ 2 (σ x , σ y, γ ) (3 − 2ψ (σ x , σ y , γ )) Finally
= Mz,cam (σ x , σ y, γ )
(B.11)
“ ”
vel
− Cyσ 0y vel
(1 − ψ (0, σ 0y , 0))2 − Fz µ yψ 2 (0, σ 0yvel vel
, 0) 3 − 2ψ (0, σ 0y , 0) Gcamz =
Fcam (0, σ reg
y , 0)
190 191
Appendix B. Scale Factors and Slip Conversion
α ○ = atan(σ ○y ) (B.17)
For the adhesive region the deformation-invariant pure-slips are calculated as
C
λ 0a = λ (B.18)
!
tan(α )
α 0a = arctan
1−λ
(B.19)
Extension of Optimization
and for the sliding region the velocity-invariant pure-slips as
Algorithm to Include
v
q
2
(λ cos (α )) + sin2 (α ) sgn (λ )
λ 0s =
v0
v
q
(B.20a)
Horizontal Shift
sin(α 0s ) = (λ cos (α ))2 + sin2 (α ) sgn (α ) (B.20b)
v0
λ ○ψ (σ x , σ y, γ )
λ 0z = sgn(λ ) (B.21a) To include the horizontal shift in the parameter estimations the brush
1 − λ ○ + λ ○ψ (σ x , σ y, γ )
model definition has to slightly redefined
α 0z = arctan(tan(α ○ )ψ (σ x , σ y , γ ) sgn(α )) (B.21b)
sin(α )
tan(β ) = (B.22) 2
λ cos(α ) 1 ( C0x (σ x + sh ))p(σ x + sh )p)
− C0x (σ x + s h ) +
3
µ
3
Ψ(σ x , C0x , µ , sh ) = 1 ( C0x (σ x + sh )) (C.1)
− 2
if pσ x p < σ ○x
27
µ
−µ ⋅ sign(σ x ) otherwise
Using the least-squares method for low-slips the regressor vector can be
set as φ i = [1 xi ] estimating the parameters θ = [θ 0 C0x ] from where
the horizontal shift is derived as sh = θ 0 / C0x . In fact, it is the vertical shift
that is estimated, but in the linear region the relation between horizontal
and vertical shift is proportional. The same relation do not hold when
extending slip area to include detection of the curvature of the force-slip
relation. Using the regressor φ i = [1 xi xi p xi p]T with θ = [θ 0 C0x θ 2 ]
2
and sh = θ 0 / C0x and µ = C0x /(3θ 2 ) might provide erroneous result.
Using the Gauss-Newton algorithm, the Jacobian matrix has to be
extended with the derivative of the shift factor as
192 193
Appendix C. Extension of Optimization Algorithm . . .
e1 e2 eN T
C0x ⋅⋅⋅
C0x C0x
e1 e2 eN
J (Ω , θ ) = ⋅⋅⋅ (C.2)
µ µ µ
e1 e2 e
N
⋅⋅⋅
sh sh sh
with
2 2
e C − 2 C0x pσ x + sh p + 1 ( C0x (σ x + sh ))
if pσ x p < σ ○x
0x
= 3 µ 9 (µ )2 (C.3)
sh
0 otherwise
194