2011 8th International Multi-Conference on Systems, Signals & Devices

Tyre effective radius and vehicle velocity estimation: a variable

structure observer solution
C. E L TANNOURY 1,2 , F. P LESTAN, 1 S. M OUSSAOUI 1 , N. ROMANI 2
1
IRCCyN, UMR CNRS 6597, Ecole Centrale de Nantes, France.
2
RENAULT co, Guyancourt, France

Abstract This paper proposes an application of a tire stiffness. An estimation of this stiffness and the
variable structure observer for wheel effective radius and effective radius is done using a linear regression model
velocity of automotive vehicles. This observer is based or the nonlinear total least squares. But this method is
on high order sliding approach allowing robustness
and finite time convergence. Its originality consists in very demanding of computing complexity and several
assuming a nonlinear relation between the slip ratio details are still remaining for a practical real time
and the friction coefficient and providing an estimation implementation. Also, the relation between the slip
of both variables, wheel radius and vehicle velocity, ratio and the longitudinal force is not perfectly linear.
from measurement of wheel angular velocity and torque. Many works have been made on observers but
These signals being available on major modern vehicle
CAN (Controller Area Network) buses, this system does for other tires variables such as tyre stiffness and
not require additional sensors. A simulation example is tyre-road friction, especially thanks to variable
given to illustrate the relevance of this approach. structure observers[10]. In [16], [12], a strategy using
Keywords. Wheel radius, vehicle velocity, high order second sliding mode observer for the estimation
sliding mode observers. of the effective radius is proposed ; however, the
I. I NTRODUCTION radius is not considered as a state variable. Only
the angular position and the corresponding velocity
The wheel is an essential part of a vehicle since
are considered as state variables. It is supposed
it contributes to directional stability, handling, and
that the vehicle velocity is known (which is not so
comfort. A key-information related to the wheel is its
obvious in real conditions: the velocity is derived
effective radius (also called rolling radius and dynamic
from the angular velocities of the wheels and the
radius) since a strong variation of this radius can
nominal radius). Furthermore, the angular velocity
have an incidence on the vehicle comfort and safety.
dynamics do not depend on the applied torque.
Existing works have shown that an estimated value of
A variable structure observer is also developed in
this radius can be obtained from a joint measurement
[15] in order to estimate the forces and parameters
of the vehicle speed and the wheel angular veloc-
needed for vehicle control, and to ensure a diagnosis
ity [1], [12]. However, this estimation remains an open
preview. In [10], an estimator of tire-road friction is
problem when the vehicles longitudinal velocity is
designed, and based on EKF (Extended Kalman filter).
not available. As a matter of fact, vehicle velocity
can be derived from the angular velocity of the wheel
In this paper, an observer using only the measured
and its a priori known nominal radius. However, if
wheel angular velocity and the torque applied on
the effective radius is decreasing or if its value is
it is proposed: this solution allows to estimate the
lower than the nominal one, the estimated velocity will
vehicle velocity and the wheel radius. The observer
be erroneous. Another way to estimate the velocity
being based on high order sliding mode theory [5],
consists in using the Global Positioning System (GPS)
[6], it implies that it needs to be a third order sliding
system, but it yields a low precision information. In
mode one which is different from the previously
that respect, we propose in this paper an estimator
cited works. The main features of such observers
of the wheel effective radius and the vehicle velocity
are their robustness with respect to uncertainties
using the information provided by the Anti-lock Brak-
and perturbations, and finite time convergence.
ing System (ABS) encoder: engine torque and angular
Furthermore, they are applicable to a very large
velocity.
class of observable systems. This choice of observer
To our best knowledge, only few works have been
strategy has been made because the radius dynamics
done in the design of radius dynamics observers. In
are not known a priori and can be viewed as a
[1], a linear relation between the longitudinal force
bounded uncertainty.
and the slip is considered using a constant longitudinal
Contact author: franck.plestan@irccyn.ec-nantes.fr The rest of this paper is organized as follows. In

978-1-4577-0411-6/11/$26.00 2011 IEEE

Section 2, the wheel model is introduced with the B. State-space representation
view to present the main equations of longitudinal
By denoting the state vector x = [x 1 x2 x3 ]T =
and rotational dynamics of a tyre. Then, the observer
[ vx R]T and u = the control input, one has
design is discussed in Section 3. Finally, some sim-

ulation results are given in Section 4 to illustrate the 1

x3 Fx (x) + Cf x1 1
applicability of the proposed method. J
 
J
II. P ROBLEM STATEMENT x = 1 F (x) F (x) F (x) + 0 u
M x d r
The main problem is how to develop an estimation
0
strategy of the wheel effective radius and vehicle   
velocity using only angular velocity and wheel torque   
g(x)
measurements. Before deriving the proposed estimator f (x) + f
(4)
a good representative model of wheel dynamics is
The force Fx (x) reads as
needed.
 
A. Wheel longitudinal and rotational dynamics x2
0 1
x1 x3
The application of Newtons second law to the
forces acting on the wheel in an acceleration phase
[4], [14] allow to express the main equations of
rotational and longitudinal dynamics
J = R Fx Cf , (1)
M vx = Fx Fd Fr (2)
where is the wheel angular velocity, R is the
effective radius, v x the vehicles linear velocity, C f
the viscous friction coefficient of the wheel, J and M
are respectively the inertia and the quarter-car mass
consisting of the car body and wheel. In addition to the
wheel torque , the main forces acting on the wheel
are the traction force F x , the aerodynamic drag force
Fd and the rolling resistance F r , which are given as
F (v ) = 1 A C v 2 ,
d x 2
d d x
(3)
Fr (vx ) = M g Cr (vx ),
Fx () = M g (),
where Cd is the aerodynamic drag coefficient, is
the mass density of the air and A d the frontal area of
the vehicle. The parameter C r is the rolling resistance
coefficient which is a function of v x , and () the
adhesion coefficient. The adhesion coefficient depends
on the wheel slip ratio [4] defined as
R vx vx
= =1 .
R R
As in [17], [8], the relationship between and is
approximated by the following function
0 generically observable if dim O = n.
() = 20
20 + 2
where 0 is the optimal slip ratio, corresponding to
the peak adhesion value ( 0 ) = 0 . Finally, the
wheel effective radius R is assumed to be unknown
but having a slow variation, that is
R = ,
with || < |0 |.
2
Fx = 20  2 M g (5)
2 x2
0 + 1
x1 x3
III. H IGH ORDER SLIDING MODE OBSERVER
SYNTHESIS
The first part of this section concerns the analysis
of observability. Then, a high order sliding mode
observer (based on differentiation) is developed.
A. Analysis of observability
Consider a nonlinear system
x = f (x) + g(x)u
(6)
y = h(x)
with the state x M, M being an open and dense
subset of IRn and the control input u IR p . y
IRp is the measured output. Let O denote the generic
observability space 1 defined by [9], [13] as
O = X (Y + U) (7)
with X = SpanK {dx}, Y = SpanK {dy (j) , j 0}
and U = SpanK {du(j) , j 0}. SpanK is a space
spanned over the field K of meromorphic functions
of x and a finite number of time derivatives of u.
y (w) (resp. u(w) ) denotes the w th time derivative of
y (resp. u).
Definition 1: The nonlinear system (6) is
This condition is called Rank condition of generic
observability. Obviously, in the nonlinear systems
context, the observability depends on u. In the sequel,
a class of nonlinear systems observable for all u
U IRp is put in the light.
1 If a property is generically satisfied, that means that this property
is locally satisfied around a regular point x0 M.
Consider the nonlinear system (6) with k 1 , k2 , , kp One gets
its observability indices, and the following vector
x1
 T
(k1 1)
y (x) y (x) 1 y 1
1 1 2 = (x) = y = x3 Fx (x)
J
= .
.. = (x)  T
3 y
 T 1
x3
dFx
x
(k 1)
yp (x) yp p (x) J dx
u U (12)
(8)
At this early stage of our work, the transformation
Definition 2: System (6) is U-observable if there
= (x) is evaluated on simulated data (see section
exist T X , U IRp and p integers {k1 , k2 , . . . , kp },
4 for an example). It appears that its jacobian is
constituting the observability indices, such that
in all cases different from 0 along the trajectories.
p
ki = n and k1 k2 kp after Therefore, = (x) seems to be a state coordinates
i=1 transformation. However, an analytical analysis will
renumbering the output components if necessary, be done in forthcoming work to check the validity of
The transformation (x) is a diffeomorphism for this transformation.
x T and u U, which is equivalent to
  Under the state transformation = (x), it yields that
Det (x) = 0 for x T . (9) there exists a function () such that () = y (3) =
x
3 . Then, the nonlinear system (11) is equivalent to

0
= A + 0
Proposition 1: There exist T X , U IR p , and (13)
()
observability indices vector [k 1 k2 . . . kp ]T such that
y = C
system (6) is observable for x T and u X .
with

Application to system (4). In this case, there is 0 1 0
only one measured output. Furthermore, one consider A= 0 0 1 and C = [1 0 0] . (14)
the following assumption: 0 0 0
H1. The additive uncertainty term f does not change
It is possible to establish that the following assumption
the observability property of (4).
is fulfilled over the work domain.
From H1, it yields that the observability analysis can
be done on the ideal system, i.e. system (4) with H2. Function is globally Lipschitzian with respect
f = 0. Furthermore, the term to .

Cf 1
x1 + u Then, the system (11) is locally uniformly observable
J J [3]. The observers developed in the sequel can be
only depends on well-known variables: thanks to written as following
application to the ideal system without the input-
0
output injection [11] given by
= A + 0 + (y, ) (15)

Cf 1 ()
x1 + u
J J
(y, u) = 0 (10) with the function being the correction term of
0 the observer. This term can be obtained by different
methods such as high gain or sliding mode observers
with y = = x1 . (detailed in next section) . . . , and allows to make the
For the observer, is unknown, so the choise obs = 0 observer converging (exponential or finite time conver-
is considered and the observer design is made in the gence) to the real system. Of course, this term only
sequel on the system (with abuse of notations) depends on the measured output y and the estimated
state vector x. From (15), two ways can be used in
1
x3 Fx (x) order to get the estimated vector x:
J

1 
1
When the inverse transformation can be
x = = fr (x).
M Fx (x) Fd (x) Fr (x) analytically derived, x can be deduced from
according to
0
(11) x = 1 ().

3
 Otherwise, from = (x), one gets (21) by using differential inclusion, i.e.
 1
1 = 2 + 2 L1/3 |1 1 |2/3 sign(1 1 )
  
= x x = (16)
x x 1
1/2 1/2
2 = 3 + 1.5 L |1 | sign(1 )
From (11)-(13)-(16), it yields   
2

 1 0 3 [L, L] + 1.1 Lsign(2 )
  
fr (x) = A +
0 3
x () (22)
(17) This inclusion is understood in Filippov sense [2].
Then, from (17)-(15), an observer for system (11) The proof of the finite time convergence now follows
reads as from Lemma 8 in [6]. Then, from (19), a finite
 1 time convergence observer for (4) reads, with 1 =
2 L1/3 |y x1 |2/3 sign(y x1 ) and 2 , 3 defined by
x = fr (x) + (y, x) (18)
x (22), as

 1 1
The application of inverse input-output injection
x = f (x, y) + (y, u) + 2
transformation (10) (y, u) allows to get an x 3
observer for the ideal system (4) (i.e. without
(23)
any uncertainty, f = 0)
 1 IV. S IMULATION AND RESULTS

x = f (x, y) + (y, u) + (y, x) A simulation is performed in order to check the
x applicability of the proposed observer for wheel radius
(19)
and vehicle velocity estimation.
In the next section, an observer solution is provided
in order to have a sufficiently accurate estimation of A. Simulation model
x in spite of the uncertainties term f . The wheel model previously introduced is firstly
used to simulate a vehicle driven at a constant speed
B. High order sliding mode differentiation based ob- (20 m/s). The parameters included in the simulator
server are summarized in table I. It can be noted that, the
nominal value of the radius R is set to 0.3 m and the
In this section, an observer based on high order sliding rolling resistance Fr is supposed to be linearly depen-
mode differentiation is suggested, in the case of the dent on the velocity v x according to F r = M gCr vx .
current problem. Consider the system (13) All the parameter values are taken according to those
proposed in [8], [4]. More realistic values would be
1 = 2 obtained by applying identification tools to a driven
2 = 3 system.
(20)
3 = ()
y = 1 Parameter value unit
J 45 kg m2
with = [1 2 3 ]T and, for all M , () R 0.3 m
L. Then, consider the following observer based on the M 440 kg
third order differentiator [6], [7] Ad 0.65 m2
1.205 kg m3

1 = 2 + 2 L1/3 | 1 1 |2/3 sign(1 1 ) g 9.807 m s2
   Cf 0.08 kg m2 s1
1
Cd 0.25
1/2 1/2
2 = 3 + 1.5 L |1 | sign(1 ) Cr 0.02
  
2 0 0.9
0 0.25
3 = () + 1.1 L sign(2 )
  
3 TABLE I
(21) W HEEL AND CAR MODEL PARAMETERS USED IN THE
The finite time convergence of the estimation error SIMULATION
to 0 can be proved thanks to the rewriting of

4
 320
Before applying the observer, the wheel is simulated Simulated radius
L = 0.02
to be driven at a nearly constant velocity by applying a 310 L = 0.002

very simple controller providing an appropriate wheel 300

torque.
Controller design. To be as close as possible to real 290

R (mm)
situation, it has been supposed that the effective radius 280
and its dynamics are well-known. The vehicle has
270
been simulated to go ahead at a slow time-varying
velocity (with v d a constant value, = 2f and 260

f = 0.05 Hz)
250
0 20 40 60 80 100
vxd d
= v (1 + 0.01 sin(t)). time (s)

Consider the controlled output Fig. 1. Simulation results - Wheel radius (mm) and its estimated
value (mm) versus time (sec).
z = R vxd (t) (24)
The relative degree of z equals 1, which gives provides the best behavior. Concerning the vehicle
R and wheel velocities, Figures 3 and 4 shows that the
z = + [ RFx Cf ] vxd observer provides satisfying reconstructions.
J
R R V. C ONCLUSIONS
= [RFx + Cf ] vxd +
 J   J A high order sliding mode observer has been ap-
(x, t) (x) plied to the problem of effective wheel radius and
vehicle velocity estimation. The choice of high order
= k z.
(25) sliding mode observer has been made for its well-
Therefore, the torque applied to the wheel reads as known features of robustness and accuracy.

 Further research should be conducted to propose an
= 1 (x, t) k R vxd (t) (26) adequate strategy for observer gain determination or
to design an adaptive sliding mode observer. Another
with k > 0 which is set empirically to k = 1
research direction is the application of the proposed
in this experiment. This choice is made to get an
method in real situations where the model parameters
acceptable behavior without any additional constraint
have to be obtained using an identification method.
on the closed-loop system response. The initial values
x(0) read as R EFERENCES

15/0.3 [1] C.-R. Carlson and J.-C. Gerdes. Consistent nonlinear estima-
x(0) = 15
tion of longitudinal tire stiffness and effective radius. IEEE
(27) Transactions on Control Systems Technology, 13(6):1010
0.3 1020, Nov. 2005.
[2] A. F. Filippov. Differential equations with discontinuous right-
Sliding mode observer. The initial values x(0) read hand side. Kluwer, Dordrecht, 1988.
as [3] J.P. Gauthier and C. Bornard. Observability for any u(t) of
15/0.31 a class of bilinear systems. SIAM Journal on Control and
x(0) = 15 . (28) Optimization, 21:721728, 1983.
[4] T. Gillespie. Fundamentals of vehicle dynamics. SAE
0.31 International, 1992.
[5] A. Levant. Sliding order and sliding accuracy in sliding mode
The parameter L is a design parameter whose influ- control. International Journal of Control, 58(6):12471263,
ence will be analyzed hereafter. 1993.
[6] A. Levant. High-order sliding modes: differentiation and
B. Simulation results output-feedback control. International Journal of Control,
76(9/10):924941, 2003.
The real wheel radius is supposed to decrease from [7] A. Levant. Finite-time stability and high relative degrees in
30 cm to 27 cm, between t1 = 40 sec and t2 = 60 sec. sliding-mode control. In International Workshop on Variable
Structure Systems, Mexico-city, Mexico, 2010.
The desired vehicle velocity v d is set to 22 m/s. [8] J.-S. Lin and W.-E Ting. Nonlinear control design of anti-
A zero mean additive random noise is added to the lock braking systems with assistance of active suspension. IET
simulated -measurement. The noise variance is set Control Theory and Applications, 1(1):343348, 2007.
[9] M. Lopez, F. Plestan, and A. Glumineau. An algorithm for
equal to 5. the structural analysis of state space: synthesis of nonlinear
Figure 1 shows that the observer provides an esti- observers. International Journal of Robust and Nonlinear
mation which is not far from the real wheel radius. Control, 11:11451160, 2001.
[10] J. Matusko, I. Petrovic, and N. Peric. Application of extended
However, the convergence time and transient error kalman filter for road condition estimation. Automatica, 44(1-
depends on the value of L. In this example, L = 0.02 2):5965, 2003.

5
 0
diagnosis strategy based on the online estimation of the tire
100
pressure. In Proceedings of the European Control Conference
200 (ECC 2007), Kos, Greece, July 2007.
300 [17] C. Unsal and P. Kachroo. Sliding mode measurement feedback
control for antilock braking systems. IEEE Transactions on
400

Determinant
Control Systems Technology, 7(2):271281, 1999.
500

600

700

800

900

1000
0 20 40 60 80 100
time (s)

Fig. 2. Simulation results - Jacobian determinant.

82

81

80

79

78
(rad/s)

77

76

75

74

73 Simulated wheel speed

Estimated wheel speed
72
0 20 40 60 80 100
time (s)

Fig. 3. Simulation results - Wheel angular velocity (rad. sec1 )

and its estimated value (rad. sec1 ) versus time (sec).
85
Simulated vehicle velocity
Estimated vehicle velocity

80
Vx (km/h)

75

70

65
0 20 40 60 80 100
time (s)

Fig. 4. Simulation results - Vehicle velocity vx (km.h1 ) and its

estimated value (km.h1 ) versus time (sec).

[11] C. Moog, F. Plestan, G. Conte, and A.M. Perdon. On canonical

forms of nonlinear systems. In European Control Conference,
Groningen, The Netherlands., 1993.
[12] N. MSirdi, A. Rabhi, L. Fridman, J. Davila, and Y. Delanne.
Second-order sliding-mode observer for estimation of vehicle
dynamic parameters. International Journal of Vehicle Design,
48(3-4):190207, 2008.
[13] F. Plestan and A. Glumineau. Linearization by general-
ized input-output injection. Systems and Control Letters,
31(31):115128, 1997.
[14] Rajesh Rajamani. Vehicle Dynamics and Control. Mechanical
Engineering Series. Springer-Verlag, 2006.
[15] H. Shraim, B. Ananou, L. Fridman, H. Noura, and M. Oulad-
sine. Sliding mode observers for the estimation of vehicle
parameters forces and states of the center of gravity. In IEEE
Conference on Decision and Control, San Diego, California,
USA., 2006.
[16] H. Shraim, B. Ananou, M. Ouladsine, and L. Fridman. A new