See discussions, stats, and author profiles for this publication at: [Link]

net/publication/254220960

Robust estimation of road friction coefficient using lateral and longitudinal

vehicle dynamics

Article in Vehicle System Dynamics · June 2012

DOI: 10.1080/00423114.2012.659740

CITATIONS READS
88 3,502

3 authors, including:

Huei Peng Eric Tseng

University of Michigan uni
397 PUBLICATIONS 27,586 CITATIONS 58 PUBLICATIONS 5,350 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Huei Peng on 07 September 2014.

The user has requested enhancement of the downloaded file.

Robust Estimation of Road Friction Coefficient Using Lateral and
Longitudinal Vehicle Dynamics

Changsun Ahn*, Huei Peng*, and H. Eric Tseng**

* Department of Mechanical Engineering, University of Michigan, Ann Arbor, MI,

USA
**
Powertrain Controls Research & Advanced Engineering, Ford Motor Company,
Dearborn, MI USA

Corresponding author: Changsun Ahn, sunahn@[Link]

(Received 31.12.2010; final version received XX Month Year)

Vehicle active safety systems stabilize the vehicle by controlling tire forces. They work
well only when the tire forces commanded by the safety systems are within the friction
limit. Therefore, knowledge of the tire/road friction coefficient can improve the
performance of vehicle active safety systems. This paper presents four methods to
estimate the friction coefficient based on four different excitation conditions: medium
lateral excitation, large lateral excitation, small longitudinal excitation, and large
longitudinal excitation. For the lateral excitation cases, the estimation is based on vehicle
lateral/yaw dynamics and Brush tire model, whereas, for the longitudinal excitation
cases, the estimation basis is the relationship between the tire longitudinal slip and
traction force. These four methods are then integrated to increase the working range of
the estimator and to improve robustness. The performance of the integrated estimation
algorithm is verified through experimental data collected on several surface conditions.

Keywords: friction coefficient estimation; nonlinear observer; recursive least squares

1. Introduction

The safety of ground vehicles relies heavily on tire-road friction. The risk of traffic

accidents increases significantly on slippery surfaces [1]. Tire-road friction

influences the ability of tires to generate steering, traction, and braking forces, which

in turn affect vehicle motions. Knowledge of the road friction coefficient is thus

important for the design and analysis of active safety systems [2]. Examples of

vehicle control systems that can benefit from the knowledge of tire-road friction

include anti-lock braking systems (ABS), electronic stability control (ESC), adaptive

cruise control, and collision warning or collision avoidance systems. The quality of

traffic management and road maintenance work (e.g., salt application and snow
ploughing) can also be improved if the estimated friction value is communicated to

the traffic and highway authorities.

Figure 1. Classification of friction coefficient estimation algorithms

Many approaches to estimate tire-road friction have been proposed based on different

vehicle dynamics and phenomena. The estimation methods can be categorized into

“cause-based” and “effect-based” approaches according to the fundamental

phenomena, which are depicted in Figure 1. The cause-based methods detect

exogenous materials covering road surfaces, such as water, ice, and snow, by using

vision, temperature, or other sensors [3-5]. These methods can estimate the friction

coefficient of the road ahead which can be beneficial but they may not manifest other

factors affecting the friction coefficient, such as tire conditions. Effect-based methods

utilize vehicle and tire dynamic behaviours directly, e.g., tire tread deformation [6-8],

the relationship between tire slip ratio and longitudinal force [9-11], wheel speed

frequency content [12], vehicle lateral dynamics [13-16], and front tire aligning

moment [17, 18]. Each method has pros and cons, such as cost, robust performance,

and required excitations.

The main objective of this research is to develop a robust and cost effective road

friction estimation algorithm with a wide operation range. Having adequate and rich

excitations is crucial for estimation algorithms. In ground vehicle applications, we do

not have the luxury of choosing the level and type of excitations; the nature of

excitations is typically dependent on the discretion of the driver, in response to road

and traffic conditions. For example, daily driving involves a lot of driving on straight

roads, with only longitudinal excitations, and at very low excitation levels most of the

time. In such situations, a road friction estimation method based on longitudinal

dynamics and wheel dynamics [19] may be a viable choice, whereas a lateral dynamic

based method will not be very useful. On the other hand, when adequate lateral

excitations exist, a method based on vehicle lateral and steering system dynamics may

become appropriate. Furthermore, under conditions involving extreme manoeuvring

of the vehicle (ploughing out or spinning out), the algorithms based on normal

manoeuvres may not work appropriately. Therefore, it is necessary to develop and

combine different approaches to form a comprehensive estimation strategy that works

under a wide range of manoeuvres.

In this paper, we develop friction coefficient estimation methods using both

longitudinal and lateral excitations. For the lateral excitation methods, we chose

vehicle lateral dynamics and front tire aligning moment as the basis of estimation

because the required sensors are readily available for existing vehicle control systems,

such as ESC and Electric Power Steering (EPS). Therefore, incremental hardware cost

is low. Furthermore, the dynamics are less sensitive to high frequency disturbances

from uneven road surfaces because of the vehicle inertia. In the authors’ previous

papers [14, 15], an algebraic and a dynamic estimator were developed, which

achieved good performance under nominal conditions. In this paper, we propose an

observer based on a parameter/state estimation methodology for improved robust

performance. For the cases with longitudinal excitations, we chose the relationship

between longitudinal slip and tire force as the basis of friction estimation. In many

studies [9-11], the friction coefficient is estimated based on the observation that tire

longitudinal stiffness depends on surface conditions. This estimation approach is

useful because it works even with small excitations, e.g., less than 2% of tire slip

ratio, and few sensors are required.

The main contributions of this study are the following: first, we propose a design

methodology for parameter/state estimation of a class of nonlinear systems. This

design methodology is then applied to determine the observer gains for simultaneous

estimation of slip angle/friction coefficient; and second, estimation algorithms for

friction coefficient will be formulated and divided into two groups based on the types

of excitations they use. In addition, an integrated estimator will be formulated,

combining the estimators from both types. This integrated estimator can work in

combined slip cases as well as pure slip cases.

The remainder of the paper is organized as follows: the lateral and longitudinal

excitation based methods are described in Section 2 and Section 3, respectively; the

integration of the developed algorithms is discussed in Section 4; the developed

methods are evaluated via experimental data in Section 5; and finally conclusions are

given in Section 6.

2. Lateral Excitation Based Methods

2.1 System Models for Lateral Dynamics

2.1.1 Vehicle Model

This chapter presents the dynamic models used for the lateral and yawing based

methods.
Figure 2. Vehicle bicycle model

The centre piece of our lateral estimation method is the vehicle bicycle model, which

describes the vehicle lateral and yaw dynamics of a two-axle, single-rigid body

ground vehicle, presented in Figure 2. The equations of motion are derived from

force and moment balance:

m(v y  vx r )  Fyf  Fyr ,

(1)
I z r  aFyf  bFyr ,

where vx is the vehicle forward speed, vy is the vehicle lateral speed, r is the yaw rate,

m is the vehicle mass, and Iz is the yaw moment of inertia. Fyf and Fyr are the lateral

forces at the front and rear axles, respectively, and a and b are the distances from the

vehicle’s centre of gravity to front and rear axles. Using small angle approximation,

the slip angles αf and αr of the front and rear tires are as follows:

 f   v y  ar  vx   ,
(2)
 r   v y  br  vx ,

where δ is the front wheel steering angle. This vehicle model is simple but effective in

representing relevant vehicle input/output relationships. Ignoring suspension roll and

road bank, the sum of tire lateral force can be calculated by using the vehicle lateral

acceleration measured at the centre of mass as follows:

Fyf  Fyr  ma y , (3)

where ay is the measured lateral acceleration.

2.1.2 Tire Model

The Brush tire model [20] is selected because it uses few parameters to compute the

tire/road forces. The lateral tire force and self-aligning moment equations are

 
 f y  3 Fz  y 1   y  13  y 2

 , for    sl ,
 
3
 a   Fz l  y 1   y
 (4)
 f y    Fz sgn( )
 , for    sl ,
 a  0

where ρy=θyσy, αsl=tan-1(1/θy), θy =2cpl2/(3µFz), σy = tan(α), l is half of the length of

the tire contact patch, α is the tire slip angle, µ is the friction coefficient, Fz is the tire

normal force, and cp is the stiffness coefficient of the tire tread in unit length.

Figure 3. Force and self-aligning moment curve of the Brush tire model

These equations and Figure 3 show the basic characteristics of the tires and the

inherent challenge for friction coefficient estimation. For example, if we know tire

lateral force or aligning moment together with the tire slip angle, then we can infer the

friction coefficient. When accurate slip angle information is not available, which is

the situation we are facing, the tire slip angle and friction coefficient need to be

estimated simultaneously from the two curves. However, when slip angle is small, the

friction coefficient is hard to differentiate because the curves are close to each other,
but the slip angle is still differentiable. When the slip angle is large, on the other hand,

slip angle cannot be differentiated but the friction coefficient is differentiable.

2.1.3 Steering System Model

Although the tire aligning moment is typically not measured in production vehicles,

electrically power steering systems provide signals enabling us to estimate the

aligning moment of the front tires. The aligning moment of front tires are dynamically

balanced by the driver’s steering moment, the assist-motor’s moment, and the

frictional moment of the steering system. These moments and related signals are

measured through the power steering assist system. Therefore, aligning moment of the

front tires can be estimated using steering system dynamics and the measured signals.

Figure 4. Steering system model

The steering system shown in Figure 4 is modelled as a simple inertia-damping

system:

J eff   beff   k   a   s   m   fx   f , (5)

where δ is the front wheel steer angle, Jeff is the effective moment of inertia, beff is the

effective damping of the steering system with respect to the road wheels, and k is the

jack up moment coefficient. τa, τs, τm, τfx and τf represent the self-aligning moment,

steering wheel torque, motor torque, torque due to longitudinal force, and friction
torque with respect to the king pin axis, respectively. τs is measured with a torque

sensor installed on the steering column, τm can be computed using motor current, and

τf can be predetermined from a Coulomb friction model.

2.2 Medium Lateral Excitation Based Estimation

The two physical quantities lateral acceleration and front tire aligning moment are

characterized by two independent variables, slip angle and friction coefficients. In

other words, we can estimate friction coefficient with the slip angle as a side product.

In this chapter, we will introduce a parameter estimation algorithm based on nonlinear

observer technique. Figure 5 shows the diagram of an estimator that uses the two

measurements.

Figure 5. Diagram of the estimator with medium lateral excitations

2.2.1 Observer Design Synthesis

The observer to simultaneously estimate road friction and slip angle is based on a

nonlinear observer framework [21]. Consider a general dynamic system:

x  f ( x, u, ),
(6)
y  h( x, u, ),
where x is the state vector of the system, y is the measurement, u is the input, and θ is

a vector of unknown parameters. To estimate the parameter vector and the states

simultaneously, we suggest the following formulation:

xˆ  f ( xˆ, u,ˆ)  L1 ( xˆ, u,ˆ)  y  yˆ  ,

 (7)
ˆ  L2 ( xˆ, u,ˆ)  y  yˆ  ,

which can be rewritten as:

zˆ  F ( zˆ, u)  L( zˆ, u)  H ( z, u)  H ( zˆ, u)  , (8)

where,

z  [ x,  ]T , F ( zˆ, u )  [ f ( xˆ, u,ˆ) 0]T ,

L( zˆ, u )  [ L ( xˆ, u,ˆ) L ( xˆ, u,ˆ)]T , H ( z, u )  h( x, u, ).
1 2

In Eq. (8), the unknown parameters become part of the augmented state and will be

estimated together with the states.

The error dynamics of the observer are shown in Eq. (9), which consist of a linear

term and a residual term.

e   A( z, u )  L( zˆ, u )C ( z, u ) e  r1 ( z, e, u )  L( zˆ, u )  r2 ( z, e, u ), (9)

where

F ( z, u ) H ( z, u )
e  z  zˆ, A( z, u )  , C ( z, u )  ,
z z

r1 ( z, e, u )  F ( z, u )  F ( zˆ, u )  A( zˆ, u )e, r2 ( z, e, u )  H ( z, u )  H ( zˆ, u )  C ( zˆ, u )e.

Consider a Lyapunov candidate function, V(z, e, u) = eTPe, where P=PT is a positive

definite matrix, we have

V (e, z, u)  eT Qe  (r1  Lr2 )T Pe  eT P(r1  Lr2 ), (10)

where Q = ‒ (A‒LC)TP‒P(A‒LC). A sufficient condition for Eq. (9) to be

asymptotically stable is that Eq. (10) is negative definite for all z and u. If F(z, u) and

H(z, u) are continuous and differentiable in z, then the following condition holds:
 r1 ( z, u , e)  Lr2 ( z, u , e)
lim  0, (11)
e 0 e

because r1 and r2 are residual terms of the first order Taylor series approximation.

Consequently,

 (r  Lr2 )  2
lim V (e)  lim  min (Q)  2 1 Pe
e 0 e 0  e 
  (12)
 lim  min (Q)  e .
2

e 0

Therefore, local stability of the observer can be determined using standard linear

system theories, for example by checking the positive definiteness of matrix Q.

There may or may not exist P and L that satisfy the stability conditions for all z and u.

An important requirement for such P and L is that they must result in a positive

definite Q, which is a necessary condition for local stability. Therefore, choosing P

and L that satisfy Q>0 is a good starting point for gain selection.

In this paper, robust stability against plant uncertainties is a key design objective of

the observer. If the system has uncertainties ΔF and ΔH, i.e.,

z  F ( z, u)  F ( z, u),
(13)
y  H ( z, u)  H ( z, u),

and if the observer is stable for all ΔF and ΔH, then the observer is said to be robust-

stable. We define robust stability region in the state space as follows:

RS ( d , 1 ,  2 )
(14)
 {( z , u ) | V ( z , e, u )   d , for e  D ( 1 ,  2 ), F   , H  },

where D(ε1, ε2) = {e|ε12< eTPe <ε22}, a region of attraction allowing a steady state

error with a radius ε1, d is a design parameter for convergence rate, Ψ is the union of

ΔF, and Θ is the union of ΔH. The robust stability region defined above quantifies the

region inside which the observer is stable, under a given region of attraction and

bounded uncertainties.
The size of the robust stability region RS(d,ε1,ε2) depends on the selection of P and L.

Therefore, the optimal P* and L* can be numerically computed, as follows:

( P* , L* )  arg max P, L size of RS (d , 1 ,  2 ) , (15)

where

size of RS (d , 1 ,  2 )  
RS ( d , 1 ,  2 )
1 dzdu.

In the case where x is one-dimensional, y is two-dimensional, and θ is scalar, we

suggest a family of P and L in the following, which is clearly not the only possible

choice to ensure local stability:

p 0 l l2 
PI  1 , L1 , (16)
0

1 l3 l4 

where

 h f h1   h 
l1   k1 1   , l2  k2  2  ,
 x x x  x  xˆ , ˆ  x  x  xˆ , ˆ
 h   h 
l3  k3  1  , l4  k 4  2  ,
   x  xˆ , ˆ    x  xˆ , ˆ

when p1= k1=k2=k3=k4=1, Q is positive definite. If we search the observer for positive

p1, k1, k2, k3 and k4, larger robust stable region can be found through the optimization

problem expressed in Eq. (15).

The observer design process is summarized below:

(1) Construct the augmented form of observer, as shown in Eq. (8).

(2) Derive partial derivatives of the nonlinear system equations to build P and L,
as shown in Eq. (16).
(3) Determine the observer design parameters d, ε1 and ε2, where d determines the
convergence rate, ε1 is related to steady state error and ε2 is related to the
maximum allowable initial estimation error.
(4) Determine the range of possible plant uncertainties, Ψ, and Θ.
(5) Perform the optimization problem by numerically computing the size of RS(d,
ε1, ε2) in the state space and determine optimal p1, k1, k2, k3, and k4.

2.2.2 Slip Angle and Friction Coefficient Observer

In general, a tire curve may have positive or negative slope depending on its operating

region. A vehicle with this tire behaviour may experience limit cycle during active

controls. The range of validity of an observer depends on the underlying model and

whether the model to be parameterized can capture all possible tire/road

characteristics or not. In this slip angle/friction coefficient estimation, since the Brush

tire model is used, which may not accurately represent the behaviors in large slip

angle/slip cases, the mismatch between the actual and modelled behaviour is lumped

in the model error term. The robust observer design methodology developed in the

previous section is applied to the friction estimation problem.

Assuming that vx is constant and taking time derivate of the first equation in Eq. (2),

αf = (vy+ar)/vx  δ, the dynamics of front tire slip angle are derived as follows:

 f  (v y  ar) / vx  . (17)

The time derivative of vy and r can be obtained from Eq. (1):

vy  ( Fyf  Fyr ) / m  vx r ,
(18)
r  (aFyf  bFyr ) / I z .

By plugging Eq. (18) into Eq. (17), the dynamics of front tire slip angle are derived.
 1 a2   1 ab 
 f     Fyf     Fyr  r  . (19)
mv
 x I v
z x  mv
 x I v
z x 

where Fyf = 2fy(αf, μ)cosδ, Fyr = 2fy(αr, μ), and αr = αf  (a+b)r/vx + δ. The system

state and parameter to be estimated are αf and μ. The augmented system matrices for

the nominal system are:

 1 a2   1 ab  
   Fyf     Fyr  r     f 
F0 ( z, u )   mv I z vx   mv I z vx   :  0  ,
 0    (20)
 
 F  Fyr   ma y   h1 
H 0 ( z, u )   yf   :   .
 2 a   2 a   h2 

Then, the observer model is as follows:

  1 a2  ˆ  1 ab  ˆ
ˆ f     Fyf     Fyr  r  
mv
 x I v
z x  mv
 x I v
z x 

1 
 l ma  Fˆ  Fˆ
y  yf yr 
 l   ˆ  , 2 a a (21)

 
ˆ  l3 ma y  Fˆyf  Fˆyr   l  4 a  ˆa  .

where l1~l4 are defined in Eq. (16). To design an observer that is robust against plant

uncertainties, we need to optimize the observer gains through the process suggested in

Section 2.2.1. We need to first define the set of plant uncertainties, Ψ and Θ. These

uncertainties are defined assuming that the most important plant uncertainties

originate from variations in vehicle mass and tire characteristics. When driving a

vehicle, the vehicle mass changes slowly, the effect of which can be estimated using

vehicle longitudinal dynamics. There are many vehicle mass estimation algorithms

reported in the literature [22-25]. We can also lump the effect of mass uncertainty

into the tire uncertainties, in which case, tire uncertainty becomes the single most

important plant uncertainty. The tire uncertainties used in this evaluation are defined

by the upper and lower bound curves, as shown in Figure 6.

Figure 6 Tire uncertainties expressed in tire force curves

In the optimization, we only consider the range, |αf | ≤ αfsl and 0 < μ ≤ μmax =1 and for

the same reason, we restrict the range of error to |e1| ≤ αfsl and |e2| ≤ μmax. We also
define emax=[αfsl, μmax]T and εmax=(emaxTPemax)1/2. In the optimization search, the range

of ε1 and ε2 are selected to be 0.15εmax and 0.5εmax, and d is set at 0.01||P||. The

parameters of the target vehicle model are shown in Table 1. The optimal observer

gains for the target vehicle are obtained by optimizing Eq. (15) and were found to be:

p1  52.4,
(22)
k1  2.5  10 9 , k 2  2.8  10 6 , k3  1.8  10 8 , k 4  1.9  10 4.

Table 1. Model parameters of the test vehicle

Parameter Value Unit Description

m 2050 kg Vehicle mass

Iz 3344 kg·m2 Yaw moment of inertia of the vehicle

a 1.432 m Distance from front axle to centre of mass

b 1.472 m Distance from rear axle to centre of mass

w 1.52 m Vehicle track width

l 0.1 m Half of the tire contact length

cp 1.65 ×106 N/m2 Tread stiffness of unit length

Jeff 4 kg·m2 Effective rotational inertia

beff 88 Nm/(rad/s) Effective damping coefficient

k 362 Nm/rad Jack-up moment coefficient

The optimization was performed using a Sequential Quadratic Programming

technique implemented in commercial software iSIGHT of Engineous. Figure 7 (a)

shows the robust stability region achieved by the optimized gains. The stability region

shown in a bright color indicates the estimator should be activated when the

normalized front tire slip angle and the normalized rear tire slip angle are in the bright

region. The normalized slip angles are defined as follows:

 f 
f  , r  r , (23)
 f ,sl  r , sl

where

 3 Fzf   3 Fzr 
 f , sl  tan 1   ,  r ,sl  tan 1   .
 2c l 2  2c l 2
 p   p 

These angles are critical angles when the tire forces get saturated, and 24.9 and 24.3

degrees for front and rear tires when μ=1.0, respectively. Figure 7 (b) also shows the

robust stability region but in the lateral force-yaw rate space when steer angle, δ, is 0.

Non-zero steer angles shift the overall graph up or down. This plot shows the

quantitative criteria of the lateral acceleration and the yaw rate for a robust

estimation.

Figure 7. Robust stability region with optimized observer gains

2.3 Large Lateral Excitation Based Estimation

When the tire-road friction coefficient changes, the self-aligning moment

characteristics, as well as its peak value varies. These variations can be used as the

basis for friction estimation. The relationship between the friction coefficient (µ) and
the peak value is found from the extremum of the aligning moment curve. Assuming

that the Brush tire model is a good representation of tire aligning moment behaviour,

the maximum value of the aligning moment can be found through the following

equation:

 a
  Fz l y (1  4 y y )(1   y y ) 2  0, (24)
 y

where τa is defined in Eq. (4). Eq. (24) is satisfied when σy =1/(4θy) or σy =1/ θy. The

aligning moment curve reaches its peak value in the first case and returns to the zero

in the second case. At the extremum point, we have two equations:

 1  27 256  a , peak
 a , peak   a   y     Fz l , or   . (25)
 4 y
  256 27 Fz l

This indicates that if we know the peak aligning moment, then we can calculate the

friction coefficient. In addition, the equation also shows that the friction coefficient is

linearly related to the peak aligning moment. For implementation purposes, we can

use the maximum value of the measured aligning moment within a past time window

instead of true peak value. The problem with this method is that the estimation result

is accurate only when (i) lateral excitation is strong enough, and (ii) the window of

data observation is large—which has the drawback of resulting in a slow response to

sudden friction change. This method is still valuable, however, in that a lower bound

of the friction coefficient can be calculated from the maximum aligning moment data

in the window:

256  a ,max
 a ,max  max  a ( ) , lower bound  . (26)
 [ t th ,t ] 27 Fz c

3. Longitudinal Excitation Based Method

The relationship between longitudinal slip and longitudinal tire force is frequently

used as a basis of friction estimation. Two main benefits are that longitudinal

excitations are more commonly available and that the related signals are readily

available. The relationship involves the traction force generation characteristic of the

tire, and many tire model models have been developed to describe these

characteristics. the small slip range, linear models [26-28] can be used because one

can observe the proportional relationship between the longitudinal tire forces and slip

ratio from experiments, but the linear relationship is valid only in the small slip range

(0~2%) because the tire force generation becomes nonlinear as the slip ratio increases.

For the larger slip range, extensive nonlinear longitudinal force models [29, 30] are

used but they have more parameters than the linear models. These multiple

parameters may cause a convergence issue and sometimes do not have physical

meaning. Some studies used a nonlinear model that has fewer numbers of parameters

or they reduced the number of parameters to estimate by using additional

assumptions. For example, Liu [9] used the Brush tire model with an adaptive

estimator to estimate the friction coefficient. Once the tire properties are given, the

longitudinal force curves are determined only by a parameter, . Canudas-de-Wit [31]

used the LuGre model [32] that has two parameters, S and C, which can be used to

adjust the force curve shape, once tire properties are given. In order to avoid multiple

parameter estimation, he assumed that the two parameters are proportional to each

other and introduced a scaling parameter  to set the two parameters S and C as

constants.

By using a force generation model that has a single parameter, estimation of the

friction coefficient can be handled over a large range of slip ratio with less concern

for convergence issues until the model is valid. However, the model uncertainties
caused by the factors not being considered, such as the vehicle dynamic states, tire

wear of the tire, temperature, and the inflation pressure, cannot be represented by the

single parameter. Also, the simple nonlinear models do not have differentiability

between different force curves with different friction coefficients at a small slip range

(0-2%), as shown in Figure 8. Test results [19, 27, 33, 34], though, shows that the

differentiation at the small slip range is possible and many papers on friction

estimation [26-28] have used this property. Furthermore, when the vehicle was

slowly accelerating from standing still, clear separations of tire force-slip ratio curves

on different surfaces is observed in Figure 9. The data shown in Figure 9 were

collected while the vehicle was slowly accelerating from the standing still, therefore,

the slip ratio was also slowly increasing. The traction force was computed using the

longitudinal acceleration and the aerodynamic drag force. Physics based tire models

do not demonstrate this behaviour under low slip on low friction road. Some papers

[33, 35] explained the discrepancy between this observed phenomenon and the

physical models and they asserted that the discrepancy is due to the intermediate

material such as snow or water between the tire and the road. Currently, we are not

aware of any physics-based tire model that demonstrates the observed behavior. This

is why many papers [19, 27, 35, 36] use heuristic linear tire models to estimate tire-

road surface friction coefficient.

Figure 8. Longitudinal tire force generation in three regions

Figure 9. Longitudinal tire force in the small-slip region under different road
conditions

As shown in Figure 8, there are three distinct regions of force generation: linear,

transient, and saturation regions. In the linear region, the force generation curves can

be modeled as a linear function and the gradient of the linear function is believed to

be dependent on the friction coefficient, as shown in Figure 8 and Figure 9. In the

saturation region, the force generation becomes insensitive to the slip ratio and is

mostly dependent on the friction coefficient and the tire normal force. In the transient

range, the force curves are affected by not only the friction coefficient but also vehicle

dynamic states and road surface materials which are not represented only by the

friction coefficient. Therefore, a single nonlinear model having a single parameter is

not accurate enough. To represent the force generation in the transient region, we may

need a multiple-parameter model, which requires the simultaneous estimation of

multiple parameters. Furthermore, in the transient region, electric slip control systems

such as ABS and TCS may be activated. Therefore, the force/slip relationship is

affected by the operation of these control systems. In this region, approaches using

information internal of the electric slip control systems can work satisfactorily. For

example, Capra [37] utilized a look-up table to relate braking, wheel deceleration, and

brake pressure to estimate the friction coefficient. The estimation algorithm is

activated only when the brake pedal is pressed and the ABS activated. Sui [38]

modeled wheel force and torque models for ABS and designed an estimator for the

friction coefficient. Estimation of road friction during ABS operation requires internal

ABS signals and possibly modification of the ABS control algorithm. Due to these

requirements, we will focus on the linear region (0~2%) and the saturation region

(30~100%).

3.1 Small Longitudinal Excitation Based Method

3.1.1 System Models for Longitudinal Dynamics

When a vehicle is under only longitudinal forces such as straight driving, the vehicle

dynamics are different from those explained in the previous chapter. The equations of

motion can be derived from force equilibrium, as shown in Figure 10 and the resulting

equation is

max  Fxrl  Fxrr  Rxf  Rxr  DA , (27)

where m is the vehicle weight, ax is the longitudinal acceleration, Fxrl and Fxrr are the

tractive forces on the rear-left and rear-right tires respectively, Rxf and Rxr are the

rolling resistance of the front and rear tires, and DA is the aerodynamic drag force. The
rolling resistance of a tire is usually determined by coast down tests. The aerodynamic

drag force is proportional to velocity squared as follows:

Cd A vx 2
DA  , (28)
2

where Cd is the drag coefficient and can be determined experimentally, A is the cross-

sectional area of the vehicle, and ρ is the air density. If the rolling resistance is known

and the vehicle is moving on a level surface, we can compute the total tractive force

of the vehicle by measuring vehicle longitudinal speed and acceleration.

Figure 10. Forces related to the vehicle longitudinal dynamics

The longitudinal tire force distribution extracted from our test results is plotted in

Figure 10. In the small-slip region, the longitudinal force is proportional to the slip

and the gradient of the force-slip curve is defined as longitudinal force stiffness. Thus,

the longitudinal force model in the small-slip range can be expressed as follows:

Fx ( )  K (  )   , for   0.02, (29)

where, κ is tire slip ratio. The longitudinal slip stiffness K(μ) depends on road friction

μ and the tire characteristic. The friction coefficients and longitudinal stiffness

between a tire (Pirelli 255/50R-17 installed on the Jaguar S-type test vehicle used in

this study) and three surfaces are listed in Table 2. The stiffness value is different

enough for different road surfaces and can be used as an effective means to reflect

road friction.
Table 2. Friction coefficient and longitudinal stiffness on several surfaces
Surfaces Friction Coefficient Longitudinal Stiffness

Concrete 0.85~1.0 16.0×104

Snow 0.35~0.4 6.6×104

Ice 0.15~2 1.8×104

3.1.2 Stiffness and Friction Coefficient Estimation

Using the linear tire force model shown in Eq. (29) and the longitudinal vehicle

dynamics, Eq. (27) can be rewritten as follows:

Rxf  Rxr  DA  rl   rr
ax   K (  ). (30)
m m

where κrl and κrr are slip ratio of rear left and rear right wheels, respectively. Each

term on the left hand side can be measured and calculated. Therefore, we can identify

the stiffness K(μ) using standard least square methods.

3.2 Large Longitudinal Excitation Based Method

The friction estimation algorithm based on tire stiffness identification does not work if

the tire slip is larger than a few percent. When a wheel is spinning, the tire

longitudinal force saturates and it becomes insensitive to the slip ratio and wheel

speed. The tire longitudinal force is dependent on the friction coefficient and the tire

vertical force, as shown in Figure 8. In this case, the saturating tire forces provide

information about the friction level. For a rear-wheel-drive vehicle (which is the case

for our test vehicle), the traction force is computed from:

a
Fxr   Fzr   mg. (31)
ab
By combining Eq. (31) and Eq. (27), the friction coefficient can be computed:

a  b  ax Rxf  Rxr  DA 
    . (32)
a g mg 

To use RLS for this formulation, Eq. (32) is modified as follows:

a  b  ax Rxf  Rxr  DA 
    1  , (33)
a g mg 

then μ can be identified using RLS. Note that this estimation algorithm is valid only

when the slip ratio is higher than about 30%.

4. Integrated Algorithm

The algorithms presented in Sections 2 and 3 were developed based on certain

assumptions of the type and level of excitations. These assumptions need to be kept

in mind because the algorithms are likely to behave satisfactorily only under these

excitations. For example, the large lateral excitation based method assumes that large

lateral excitations exist, and the nonlinear observer requires medium level of lateral

excitations, whereas, the longitudinal dynamics based method requires less than 2%

of longitudinal slip for the linear region and 30%~100% of longitudinal slip for the

saturation region.
Figure 11. Coverage of the presented estimation methods in the friction circle

In Figure 11, the two longitudinal excitation based methods only cover a small range

of vehicle operations, whereas, the two lateral excitation based methods cover almost

operations with pure lateral excitations. However, all four methods were developed

based on pure-slip assumption and might not handle combined-slip cases well. To

deal with combined excitation cases, we use a combined slip Brush model and a four-

wheel vehicle model. The combined slip Brush model is as follows[20]:

x y
Fx  F , Fy  F , M z  t ( )  Fy , (34)
 

where

  Fz (1   3 ) for    sl
F ( ,  ,  )   ,
  Fz sgn( ) for    sl

  1   ,   2c p l 2 / (3 Fz ),    x2   y2 ,
 x   / (  1),  y  (tan  ) / (  1),  sl  1/  ,
t ( )  l 1   
3

/ 3  3   
2
.
To differentiate the longitudinal tire forces on left and right sides, the modified

vehicle model (shown in Figure 12) is as follows:

m  v y  vx r   Fyf  Fyr ,
w w (35)
I z r  aFyf  bFyr  Fxr  Fxl .
2 2

Figure 12. Four wheel bicycle model

Based on the four-wheel bicycle model, an observer for combined excitations is:

 1 a2  ˆ  1 ab  ˆ w
ˆ f     Fyf     Fyr  ( Fˆxr  Fˆxl )  r  
 mvx I z vx  mv
 x I v
z x  2 I v
z x

  
 l1 ma y  Fˆyf  Fˆyr  l2  a  ˆa  , (36)

 
ˆ  l3 ma y  Fˆyf  Fˆyr   l 4 a  ˆa  ,

where Fˆyf and Fˆyr are calculated by Eq. (34) using the estimated slip angle and

measured slip ratio.

This observer can handle combined slip cases but still cannot manage small lateral

excitation cases because sufficient excitations are required for robust and stable

estimation. Therefore, we use an open-loop observer for the cases that the four

methods cannot handle. The open-loop observer is as follows:

  1 a2  ˆ  1 ab  ˆ w
ˆ f     Fyf     Fyr  ( Fˆxr  Fˆxl )  r  ,
 mv x I z v x   mv x I z v x  2 I z v x (37)
ˆ  0,

The conditions for selecting the five estimators are summarized in Table 3. The

nonlinear observer for medium lateral excitations is modified to cover combined slip

cases and the open loop observer is used to fill the gap of the four developed

estimators. The last remaining step is to define logic to switch among the five

estimators.

Table 3. Friction coefficient estimators for different excitation conditions

Excitation Estimators for integration Remarks
 1 a2  ˆ  1 ab  ˆ
ˆ f     Fyf     Fyr
 mvx I z vx  mv
 x I z vx 
w
Medium lateral  ( Fˆxr  Fˆxl )  r   Using combined slip tire
2 I zVx
excitation model
 
 l1 ma y  Fˆyf  Fˆyr    l 
2 a  ˆa  ,

 
ˆ  l3 ma y  Fˆyf  Fˆyr   l 
4 a  ˆa  .

Valid only when the

 a ,max  max  a ( )  , longitudinal excitation is
 [ t  th , t ]
Large lateral
256  a ,max small. It provides a lower
excitation lower bound  .
27 Fz c bound for other
estimators.
y (t )   T (t ) (t ),
Small
Rxf  Rxr  DA
longitudinal y (t )  ax  g sin   ,
m
excitation
 rl   rr
(0-2%)  (t )  ,  (t )  K (  ).
m
Large y (t )   T (t ) (t ),
longitudinal a  b  ax Rxf  Rxr  DA 
y (t )    sin   ,
excitation a  g mg 
(30-100%)  (t )  1,  (t )   .
 1 a2  ˆ  1 ab  ˆ
ˆ f     Fyf     Fyr
mv
 x I v
z x  mv
 x I z vx 
All other w
conditions  ( Fˆxr  Fˆxl )  r  
2 I zVx
ˆ  0.
Figure 13. Estimation flow of the integrated estimator

The five methods described in Table 3 are integrated by a switching logic shown in

Figure 13. The estimators are activated based on the magnitude of excitation signals,

and the initial values for the selected estimator are the latest estimated results of the

previously selected estimator. The method used to handle the combined slip cases

increases the coverage of lateral dynamics based method, as shown in Figure 14.
Figure 14. Increased coverage of the estimators by using the combined slip tire model

5. Experimental Validation

5.1 Experimental vehicle

Validation of the developed algorithms was performed on the winter test track of Ford

Motor Company in Sault Ste. Marie, Michigan, USA. The test vehicle is a rear-wheel-

drive Jaguar S-type, which was modified for the development of vehicle dynamics

control algorithms. The vehicle has standard ESC sensors, including yaw rate and

lateral acceleration sensors, four wheel speed sensors, a steering wheel angle sensor,

and a steering torque sensor. To measure the rack force, two strain gauges were

installed on the steering racks. Figure 15 shows the test vehicle and the GPS/INS

system. To measure reference signals, a RT-2500 system of Oxford Technology was

installed, which has two antenna GPS integrated with INS and measures three

dimensional vehicle position and orientation, as well as three dimensional linear and

angular velocities of the vehicle.

Figure 15. Test vehicle and GPS/INS system

5.2 Vehicle and Tire Parameter Identification

While vehicle parameters are easily obtained from the vehicle design specifications,

tire parameters are usually not readily available and therefore need to be identified

through vehicle tests. We performed steady state turning manoeuvres to identify the

key tire parameters: tread stiffness and the contact patch length. Tire lateral forces at

front and rear axle were computed using following equations and post-processed

signals:

b  m  a y  I z r a  m  a y  I z r
Ff  , Fr  , (38)
(a  b) cos  ( a  b)
where the time derivate of yaw rate signal was computed using an ideal filter. The two

tire parameters were identified through minimization of the difference between the

model and the measurement force trajectory in α-F domain, as shown in Figure 16 (a).

In contrast, the steering system parameters, such as the rotational inertia and the

damping coefficient, are identified through fast transient manoeuvres. The steering

system parameters, Jeff, beff, and k, which are shown in Eq. (5), were computed using

the measured kingpin moment when the angular acceleration and velocity components

were rich in the steering angle signal. Finally, additional transient manoeuvre was

performed for the purpose of verifying the fidelity of the combined system model.
Figure 16 shows the validation of the tire and steering models and Figure 17 shows

the validation of the vehicle model integrated with the identified tire and steering

system models. The identified parameters are listed in Table 1.

Figure 16. Model vs. measured tire forces

Figure 17. Validation of the combined system model

5.3 Estimation Results

The test vehicle travelled on four different surfaces: concrete, ice, snow and slippery

concrete surfaces, as shown in Figure 18. Several tests were conducted to verify the

developed algorithms. We present two of them as representative examples. As

reference signals, three dimensional vehicle position and orientation, as well as three

dimensional linear and angular velocities of the vehicle were measured through a GPS

based sensor (RT-2500), the friction coefficients of test surfaces were identified by a

maximum braking acceleration tests.

Figure 18. The four test surfaces

The first test was designed to evaluate the integrated algorithm, where both lateral and

longitudinal excitations exist but are not sufficient for either to work all the time. The

driver was asked to emulate a daily driving pattern, such as slight single lane changes.

The measured signals are plotted in Figures 19 and 20, which compared the three

algorithms. As the longitudinal dynamics based algorithm does not estimate slip

angle, only two algorithms are compared for the slip angle estimate. The longitudinal

dynamics based algorithm is able to estimate the friction coefficient accurately when

pure longitudinal excitations exist. When only lateral excitations exist (between t=20

and t=30), the longitudinal dynamics based algorithm does not update the estimated
value. In this case, however, the lateral dynamics based algorithm has adequate

excitation and its estimation is accurate. The integrated algorithm utilizes both

excitations and thus has more chance to update.

Figure 19. Measured signals of the first test

 Figure 20. Comparison of three estimators using the data of the first test

The second test was designed to have sufficient lateral excitation and the measured

signals are plotted in Figure 21. In that test, we can see the benefit of integration more

clearly. The longitudinal dynamics based algorithm exhibits a very low level of

accuracy of estimation due to the sparse occurrence of longitudinal excitations. The

lateral dynamics based algorithm generally tracks the actual friction coefficient well,

except during abrupt surface changes. The combination of algorithms improves the

estimation performance because the integrated algorithm works with both excitations.
Figure 21. Measured signals of the second test
Figure 22. Comparison of three estimators using the data of the second test

In Figure 23, Graph (a) shows comparisons between the integrated algorithm and the

uni-directional excitation based algorithms, which are evaluated by using the same

test data sets. The tracking delays are calculated during surface changes and the

steady state errors are identified on even surfaces. For the integrated algorithm, the

steady state estimation errors are within 0.2~0.3 and delays are within 1~5 seconds.

For the other two algorithms, the steady state estimation errors and delays are

distributed within 0.1~0.5 and 0~15 seconds, respectively. Graphs (b) and (c) of

Figure 23 show the steady state error and tracking delays obtained from the literature

[2, 4, 6, 12, 14, 16, 18, 19, 36, 39, 40]. Those results may not be directly comparable

to the performance of our estimators because the surface conditions, tires, vehicles,

and excitation conditions are not the same. Considering that our test data emulating

daily driving patterns while results in other papers might be using more favorable
excitations, we conclude that the integrated algorithm we developed performs well

compared to the algorithms found in the literature.

Figure 23. Performance of the integrated algorithm (a) compared to existing

algorithms in literature, (b) and (c)

6. Conclusion

In this paper, we presented friction coefficient estimation algorithms which work with

sensors which are mostly available on typical passenger cars. They are developed

based on either lateral dynamics or longitudinal dynamics but later extended to cover

combined slip cases. With respect to the lateral dynamics based methods, which are

more robust to high frequency noises, a nonlinear observer was proposed.

Furthermore, a process to design robust nonlinear observers was developed using

numerical implementation of Lyapunov stability conditions. Regarding the

longitudinal dynamics based methods, tire longitudinal stiffness is used to infer road

friction. The original lateral and longitudinal dynamics based methods work well only

under pure-slip excitation conditions. To increase the working range, an integration

logic was developed to switch among the developed algorithms based on the nature

and level of excitations.

The algorithms are verified through experimental results. It was found that the road

friction can be estimated accurately. Depending on the nature and level of

excitations, however, the delay in estimation convergence can be as long as five

seconds, when adequate excitation is not present. Both estimation accuracy and

latency compare favourably to the results reported in the literature. Reducing the

estimation delay is a possible topic for further improvement.

Acknowledgements

This research work is supported by Ford Motor Company through the UM-Ford Alliance Program.

References

[1] C.G. Wallman, and H. Åström, Friction measurement methods and the
correlation between road friction and traffic safety. A literature review,
Swedish National Road and Transport Research Institute, Linköping, Sweden,
2001.
[2] M. Andersson, et al., Road friction estimation, Intelligent Vehicle Safety
Systems, Sweden, 2007.
[3] F. Holzmann, et al., Predictive estimation of the road-tire friction coefficient,
in IEEE International Conference on Control Applications, Munich, Germany,
2006, pp. 885-890.
[4] Y. Sato, et al., Study on recognition method for road friction condition, in
JSAE Transaction, 2007, pp. 51-56.
[5] M. Yamada, et al., Road surface condition detection technique based on image
taken by camera attached to vehicle rearview mirror, in Review of Automotive
Engineering, 2005, pp. 163-168.
[6] U. Eichhorn, and J. Roth, Prediction and monitoring of tyre/road friction, in
FISITA, London, U.K., 1992, pp. 67-74.
[7] A. Tuononen, Vehicle lateral state estimation based on measured tyre forces,
in International Symposium on Dynamics of Vehicles on Roads and Tracks,
Stockholm, Sweden, 2009.
[8] A. Tuononen, On-board estimation of dynamic tyre forces from optically
measured tyre carcass deflections, in International Journal of Heavy Vehicle
Systems, 2009, pp. 362-378.
[9] C.-S. Liu, and H. Peng, Road friction coefficient estimation for vehicle path
prediction, in Vehicle System Dynamics, 1996, pp. 413 - 425.
[10] M. Ito, et al., Estimation of road surface conditions using wheel speed
behavior, in International Symposium on Advanced Vehicle Control, Tsukuba,
Japan, 1994, pp. 533-538.
[11] F. Gustafsson, Monitoring tire-road friction using the wheel slip, in IEEE
Control Systems Magazine, 1998, pp. 42-49.
[12] T. Umeno, et al., Estimation of tire-road friction using tire vibration model, in
SAE 2002, Detroit, Michigan, USA, 2002.
[13] C. Sierra, et al., Cornering stiffness estimation based on vehicle lateral
dynamics, in Vehicle System Dynamics, 2006, pp. 24 - 38.
[14] J.-O. Hahn, et al., GPS-based real-time identification of tire-road friction
coefficient, in IEEE Transactions on Control Systems Technology, 2002, pp.
331-343.
[15] W.R. Pasterkamp, and H.B. Pacejka, The tyre as a sensor to estimate friction,
in Vehicle System Dynamics, 1997, pp. 409 - 422.
[16] L.R. Ray, Nonlinear tire force estimation and road friction identification:
simulation and experiments, in Automatica, 1997, pp. 1819-1833.
[17] Y. Yasui, et al., Estimation of lateral grip margin based on self-aligning
torque for vehicle dynamics enhancement, in SAE 2004, Detroit, Michigan,
USA, 2004.
[18] Y.-H.J. Hsu, and J.C. Gerdes, A feel for the road: A method to estimate tire
parameters using steering torque, in International Symposium on Advanced
Vehicle Control, Taipei, Taiwan, 2006.
[19] K. Yi, et al., Estimation of tire-road friction using observer based identifiers,
in Vehicle System Dynamics, 1999, pp. 233 - 261.
[20] H.B. Pacejka, Tyre brush model, in Tyre and Vehicle Dynamics, Elsevier,
Oxford, U.K., 2005, pp. 93-134.
[21] C. Ahn, Robust Estimation of Road Friction Coefficient for Vehicle Active
Safety Systems, in Mechanical Engineering, The University of Michigan, Ann
Arbor, MI, 2011.
[22] H.S. Bae, et al., Road grade and vehicle parameter estimation for longitudinal
control using GPS, in IEEE Intelligent Transportation Systems Conference,
Oakland, California, USA, 2001.
[23] H.K. Fathy, et al., Online vehicle mass estimation using recursive least
squares and supervisory data extraction, in American Control Conference,
Seattle, Washington, USA, 2008.
[24] V. Winstead, and I.V. Kolmanovsky, Estimation of road grade and vehicle
mass via model predictive control, in IEEE Conference on Control
Applications, Toronto, Canada, 2005, pp. 1588-1593.
[25] A. Vahidi, et al., Recursive least squares with forgetting for online estimation
of vehicle mass and road grade: theory and experiments, in Vehicle System
Dynamics, 2005, pp. 31 - 55.
[26] C.R. Carlson, and J.C. Gerdes, Consistent nonlinear estimation of longitudinal
tire stiffness and effective radius, in Control Systems Technology, IEEE
Transactions on, 2005, pp. 1010-1020.
[27] S.L. Miller, et al., Calculating longitudinal wheel slip and tire parameters
using GPS velocity, in American Control Conference, Arlington, Virgina,
USA, 2001, pp. 1800-1805.
[28] J. Wang, et al., Friction estimation on highway vehicles using longitudinal
measurements, in Journal of Dynamic Systems, Measurement, and Control,
2004, pp. 265-275.
[29] H.B. Pacejka, Semi-empirical tyre models, in Tyre and Vehicle Dynamics,
Elsevier, Oxford, U.K., 2005, pp. 156-215.
[30] C. Canudas de Wit, and P. Tsiotras, Dynamic tire friction models for vehicle
traction control, in IEEE International Conference on Decision and Control,
Phoenix, Arizona, USA, 1999, pp. 3746-3751.
[31] C. Canudas de Wit, and R. Horowitz, Observers for tire/road contact friction
using only wheel angular velocity information, in IEEE International
Conference on Decision and Control, Phoenix, Arizona, USA, 1999, pp. 3932-
3937.
[32] C. Canudas de Wit, et al., A new model for control of systems with friction, in
IEEE Transactions on Automatic Control, 1995, pp. 419-425.
[33] J. Deur, et al., On low-slip tire friction behavior and modeling for different
road conditions, in IAVSD, 2005.
[34] Y. Nakao, and H. Kawasaki, Estimation of friction levels between tire and
road, in SAE 2002 World Congress, Detroit, Michigan, USA, 2002.
[35] T. Dieckmann, Assessment of road grip by way of measured wheel variables,
in FISITA, London, 1992.
[36] F. Gustafsson, Slip-based tire-road friction estimation, in Automatica, 1997,
pp. 1087-1099.
[37] D. Capra, et al., Experimental test of vehicle longitudinal velocity and road
friction estimation for ABS system, in SAE 2009 World Congress, Detroit,
Michigan, USA, 2009.
[38] D. Sui, and T.A. Hohansen, Moving horizon estimation for tire-road friction
during braking, in IEEE Multi-conference of Systems and Control, Yokohana,
Japan, 2010.
[39] C. Lee, et al., Real time slip based estimation of maximum tire road friction
coefficient, in IEEE/ASME Transactions on Mechatronics, 2004, pp. 454-458.
[40] Y.-H.J. Hsu, et al., A method to estimate the friction coefficient and tire slip
angle using steering torque, in International Mechanical Engineering
Congress and Exposition, American Society of Mechanical Engineers, New
York, NY 10016-5990, United States, Chicago, IL, USA, 2006.
 Table 1. Model parameters of the test vehicle
Parameter Value Unit Description

m 2050 kg Vehicle mass

Iz 3344 kg·m2 Yaw moment of inertia of the vehicle

a 1.432 m Distance from front axle to centre of mass

b 1.472 m Distance from rear axle to centre of mass

w 1.52 m Vehicle track width

l 0.1 m Half of the tire contact length

cp 1.65 ×106 N/m2 Tread stiffness of unit length

Jeff 4 kg·m2 Effective rotational inertia

beff 88 Nm/(rad/s) Effective damping coefficient

k 362 Nm/rad Jack-up moment coefficient

Table 2. Friction coefficient and longitudinal stiffness on several surfaces
Surfaces Friction Coefficient Longitudinal Stiffness

Concrete 0.85~1.0 16.0×104

Snow 0.35~0.4 6.6×104

Ice 0.15~2 1.8×104

Table 3. Friction coefficient estimators for different excitation conditions
Excitation Estimators for integration Remarks
 1 a2  ˆ  1 ab  ˆ
ˆ f     Fyf     Fyr
 mvx I z vx   mvx I z vx 
w
Medium lateral  ( Fˆxr  Fˆxl )  r   Using combined slip tire
2 I zVx
excitation model
 
 l1 ma y  Fˆyf  Fˆyr    l 
2 a  ˆa  ,

 
ˆ  l3 ma y  Fˆyf  Fˆyr   l 
4 a  ˆa  .

Valid only when the

 a ,max  max  a ( )  , longitudinal excitation is
 [ t  th , t ]
Large lateral
256  a ,max small. It provides a lower
excitation lower bound  .
27 Fz c bound for other
estimators.
y (t )   T (t ) (t ),
Small
Rxf  Rxr  DA
longitudinal y (t )  ax  g sin   ,
m
excitation
 rl   rr
(0-2%)  (t )  ,  (t )  K (  ).
m
Large y (t )   T (t ) (t ),
longitudinal a  b  ax Rxf  Rxr  DA 
y (t )    sin   ,
excitation a  g mg 
(30-100%)  (t )  1,  (t )   .
 1 a2  ˆ  1 ab  ˆ
ˆ f     Fyf     Fyr
 mvx I z vx   mvx I z vx 
All other w
conditions  ( Fˆxr  Fˆxl )  r  
2 I zVx
ˆ  0.
 Figure 1. Classification of friction coefficient estimation algorithms

Figure 2. Vehicle bicycle model

Figure 3. Force and self-aligning moment curve of the Brush tire model

Figure 4. Steering system model

Figure 5. Diagram of the estimator with medium lateral excitations

Figure 6. Tire uncertainties expressed in tire force curves

Figure 7. Robust stability region with optimized observer gains

Figure 8. Longitudinal tire force generation in three regions

Figure 9. Longitudinal tire force in the small-slip region under different road

conditions

Figure 10. Forces related to the vehicle longitudinal dynamics

Figure 11. Coverage of the presented estimation methods in the friction circle

Figure 12. Four wheel bicycle model

Figure 13. Estimation flow of the integrated estimator

Figure 14. Increased coverage of the estimators by using the combined slip tire model

Figure 15. Test vehicle and GPS/INS system

Figure 16. Model vs. measured tire forces

Figure 17. Validation of the combined system model

Figure 18. The four test surfaces

Figure 19. Measured signals of the first test

Figure 20. Comparison of three estimators using the data of the first test

Figure 21. Measured signals of the second test

Figure 22. Comparison of three estimators using the data of the second test

Figure 23. Performance of the integrated algorithm (a) compared to existing

algorithms in literature, (b) and (c)

View publication stats