You are on page 1of 8

2400 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO.

6, NOVEMBER 2013
Traction-Control-Oriented State Estimation for Motorcycles
Matteo Corno, Giulio Panzani, and Sergio M. Savaresi
AbstractThis brief addresses two estimation problems
relevant to traction control for motorcycles: longitudinal vehicle
velocity estimation and wheelie (i.e., front wheel lifting off the
ground during acceleration) detection. Two methods to estimate
the vehicle body velocity are discussed and compared: a comple-
mentary lter and a Kalman lter. The Kalman lter reduces the
noise affecting the estimate of the longitudinal vehicle velocity
by an order of magnitude without introducing any phase lag.
Furthermore, a wheelie detection algorithm is developed. The
approach is based on the fault detection paradigm and detects
wheelies in 70 ms. Both methods are computationally efcient
and industrially viable. Track tests on an instrumented sport
motorcycle are employed to illustrate and validate the methods.
Index TermsMotorcycle dynamics, sensor fusion, traction
control (TC), wheelie detection, wheel-slip estimation.
I. INTRODUCTION
T
HE PAST decade has witnessed the success of active
vehicle dynamics control systems for cars. Among them,
antilocking braking (ABS), electronic stability control, trac-
tion control (TC), and semiactive suspensions are the most
common (see [1][3]). These systems increase safety and
performance and are routinely sought by customers.
Active vehicle dynamics control systems have been devel-
oped for motorcycles as well. Initially they did not meet public
favor. The reasons are manyfold. Two-wheeler dynamics are
considerably more complex than those of four-wheelers, and
engineers have struggled to design systems that can improve
safety without being invasive. Their efforts were sometimes
impeded by customer characteristics. Motorcycle riders tend
to be romantic and in general do not accept any help riding
their vehicles, especially if only to improve safety.
The past few years have witnessed a change of approach.
Performance-oriented control systems have encouraged the
acceptance of electronic systems. Racing ABS [4][6], TC [7],
and semiactive suspensions [8] are gaining support, and
research on advanced stability control systems is also being
pursued [9][12]. The design of such systems poses interesting
technical challenges. The vast motorcycle modeling and sim-
ulation literature (see [13] and references therein) indicates
that motorcycle dynamics have several features that require
the development of new solutions.
Manuscript received May 30, 2012; revised May 30, 2012; accepted
October 14, 2012. Manuscript received in nal form January 4, 2013. Date
of publication January 28, 2013; date of current version October 15, 2013.
This work was supported in part by the MIUR Project New Methods for
Identication and Adaptive Control for Industrial Systems. Recommended by
Associate Editor N. K. Kazantzis.
The authors are with the Dipartimento di Elettronica e Informazione,
Politecnico di Milano, Milan 20133, Italy (e-mail: corno@elet.polimi.it;
panzani@elet.polimi.it; savaresi@elet.polimi.it).
Color versions of one or more of the gures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identier 10.1109/TCST.2013.2238539
v

r
R
r

f
a
x
suspension
stroke

Fig. 1. Sport motorcycle diagram with available measurements and symbol


denitions.
Thanks to their direct impact on performance, TC systems
have gained a good reputation. The goal of a TC system is to
avoid excessive rear-wheel slip, dened as

r
=

r
R
r
v
max (v,
r
R
r
)
(1)
where
r
is the rear-wheel angular velocity, R
r
is the rear-
wheel rolling radius, and v is the longitudinal velocity of the
wheel center (see Fig. 1). The importance of avoiding exces-
sive wheel slip is twofold: too high a wheel slip determines
a nonmaximal acceleration, and it leads to loss of stability
(see [14] for more details).
This brief addresses two critical estimation problems for
motorcycle TC: vehicle longitudinal velocity estimation and
detection of front wheel lift-off. An accurate computation
of wheel slip requires the knowledge of the wheel angular
velocity and vehicle longitudinal velocity. The wheel angular
velocity is easily measured; the longitudinal velocity cannot
be directly measured and thus has to be estimated. The need
to estimate the vehicle longitudinal velocity is also present
in four-wheeled vehicles; however, motorcycles tilt and have
only two wheels, and this makes the straightforward applica-
tion of techniques developed for cars impossible. Once both
the longitudinal velocity and the wheel angular velocity are
known, (1) can be employed to compute the wheel slip for TC
purposes. Furthermore, due to the high center of gravity and
height/wheel base ratio, a motorcycle is subject to wheelies
(a phenomenon during which the front wheel lifts off the
ground during acceleration). This phenomenon has to be
explicitly addressed in the implementation of performance-
oriented TC systems for motorcycles.
1063-6536 2013 IEEE
CORNO et al.: TRACTION-CONTROL-ORIENTED STATE ESTIMATION FOR MOTORCYCLES 2401
For the estimation of the longitudinal velocity, a sensor
fusion approach is taken, i.e., the wheel velocity and longitudi-
nal acceleration are fused to overcome each others limitations.
To this end, a complementary lter is compared with a Kalman
lter. Wheelie detection is posed as a fault detection problem
using the wheel velocity and longitudinal accelerometers. To
the best of our knowledge, these topics have never been
addressed in the literature. Particular attention is devoted to
obtaining reliable, easily implementable, and cost-effective
techniques based on off-the-shelf electronics and sensors.
All the proposed methods are discussed from the method-
ology point of view and experimentally validated on an
instrumented sport motorcycle.
The rest of this brief is structured as follows. Section II
presents the experimental setup along with the characteristics
of the available sensors. Section III describes and validates the
velocity estimation algorithms. Section IV details the wheelie
estimation algorithms.
II. SYSTEM DESCRIPTION AND EXPERIMENTAL SETUP
In the remainder of this brief, we will refer to a sport
motorcycle equipped with several sensors (see Fig. 1), as
follows.
1) Front and rear wheel encoders: the discrete position
encoder information is used to estimate the angular
wheel velocity using the 1/T method (see [15]).
2) A one-axis MEMS longitudinal accelerometer is
mounted close to the center of gravity.
3) An optical sensor: this sensor provides an accurate
measurement of the vehicle longitudinal velocity v
(error < 0.1 km/h). It is, however, expensive and not
available for production vehicles. In this context, it
is employed as a reference to validate the proposed
methods.
4) Front suspension stroke sensor: provides a measurement
of the suspension compression, i.e., the distance between
the wheel hub and the motorcycle body. Stroke sensors
are usually not installed on production vehicles.
III. VELOCITY ESTIMATION ALGORITHMS
Currently, no industrially viable vehicle longitudinal veloc-
ity sensor exists, and thus v has to be estimated from indirect
measurements. Estimating the vehicle longitudinal velocity is
a difcult problem, and different approaches are available in
the literature (note that the literature refers entirely to four-
wheeled vehicles). Among the most successful approaches,
one can nd model-based approaches. These methods are
based on dynamic models of the vehicle (see [16]) and have
been shown to be fairly successful in four-wheeled vehicles.
Unfortunately, the difculties in obtaining control-oriented
models of motorcycles have not, up to now, enabled the
extension of these methods to motorcycles.
In current motorcycle TC systems, the vehicle longitudinal
velocity is estimated by assuming that the front wheel is not
affected by wheel slip and thus v =
f
R
f
. In the remainder
of the section, we will show how this basic estimate can be
improved using off-the-shelf cost-effective sensors, namely,
Fig. 2. Block diagram of the frequency separation principle.
wheel encoders and accelerometers. These two sensors have
complementary characteristics. Encoders provide an accurate
measurement of the wheel velocity, but are affected by
high-frequency noise. It is common to measure a peri-
odic high-frequency noise related to the wheel rolling fre-
quency and its harmonics. The acceleration measurement
is, instead, mainly affected by bias (due to misalignment),
making the straightforward numerical integration of the
acceleration impossible. This situation is rather common in
kinematic estimation problems (see [17]), and two solu-
tions are usually employed: the complementary lter or the
Kalman lter.
A. Complementary Filter
The complementary lter (Fig. 2) performs a frequency-
weighted fusion of two sensors. This is useful, as the available
sensors have different noise characteristics in the frequency
domain. The longitudinal velocity estimation is the sum of
two terms: the low-pass ltered wheel velocity V
LF
, and the
high-pass ltered integrated longitudinal acceleration V
HF
.
Low-pass ltering the wheel velocity v
f
=
f
R
f
removes
high-frequency noise, but introduces a phase lag; the phase
lag is compensated for by the integrated longitudinal accelera-
tion. High-pass ltering the longitudinal acceleration removes
the bias before the integration, thus leaving only the high-
frequency components. More details on the complementary
lter approach can be found in [18]. Note that the two lters
ought to be complementary in the frequency-domain sense,
i.e., HP
lter
(s) +LP
lter
(s) = 1. The designer has two degrees
of freedom: the order of the high-pass lter, and its cutoff
frequency. Most of the time, a rst-order high-pass lter is
used
HP
lter
(s) =
s
s +
and LP
lter
(s) =

s +
where is the location of the pole that has to be tuned (most of
the time by trial and error). This approach is efcient and easy
to implement; however, there is no guarantee of optimality, and
tuning is essentially a trial-and-error procedure.
B. Kalman Filter
The Kalman lter is a standard approach to solve model-
based estimation problems (see [18]). To solve the problem
at hand, one could employ a complete dynamic model of
2402 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 6, NOVEMBER 2013
the motorcycle (as is done in [19] for cars); however the
computational complexity of such a model would be very high,
and its accuracy is limited by the unknown riders physical
characteristics. To overcome these disadvantages, a kinematic
model is proposed. The kinematic equations are comple-
mented with a model of the time-varying accelerometer bias.
The resulting model is
v = a
x
+ b +
1

b =
2
v
f
= v +
1
. (2)
The rst equation represents the relation between the vehi-
cle velocity (v) and measured acceleration (a
x
) (considered
an input). The bias term b and a white noise
1
are added
to account for unmodeled dynamics and terms. The bias
dynamics is usually described as a Brownian motion driven
by noise
2
(see [20]). The output of the system is the front
wheel velocity affected by measurement noise
1
. All noises
are assumed to be uncorrelated Gaussian white noise with
variance q
1
and q
2
for the model disturbances, and r
1
for the
measurement noise. The dynamic system can be rewritten in
the matrix form as
x = A x + B u + w y = C x + D u +
1
where
x =
_
v b
_
T
w =
_

1

2
_
T
u = a
x
y = v
f
A =
_
0 1
0 0,
_
B =
_
1 0,
_
T
C =
_
1 0,
_
D = 0.
The steady-state Kalman lter is

x =
_
_
_
_


r
1
1


r
1
0
_

_
x +
_
1
0
_
a
x
+
1
r
1
_

_
v
f
(3)
where
=
_
q
1
r
1
+ 2r
1

r
1
q
2
=

r
1
q
2
, =

r
1
. (4)
Simple algebraic manipulation shows that system (3) can be
recast in an input/output (I/O) form; the I/O representation has
two inputs (wheel velocity v
f
and longitudinal acceleration
a
x
), and the transfer functions can be rearranged in the block
diagram shown in Fig. 2 with
HP
lter
(s) =
s
2
s
2
+

r
1
s

r
1
LP
lter
(s) =

r
1
s

r
1
s
2
+

r
1
s

r
1
. (5)
A complementary second-order lter is readily recognized
as HP
lter
(s) + LP
lter
(s) = 1. The above discussion shows
that the second-order complementary lter tuned according to
(4) is optimal for the model (2).
The literature reports several applications of a rst-order
complementary lter (see [17]); it is therefore interesting to
analyze how it differs from the second-order lter. Let us
consider the following system:
v = a
x
+ v
f
= v + (6)
where the accelerometer time-varying bias has not been mod-
eled and and are Gaussian zero-mean white noises with
variance q and r. Repeating the above computations, the rst-
order estimator becomes

x =

r
x + a
x
+

r
v
f
with =

qr. Also, in this case the complementary lter
structure arises with the following transfer functions:
HP
lter
(s) =
s
s +

r
; LP
lter
(s) =

r
s +

r
. (7)
Note that the rst-order Kalman estimator misses the critical
issue of the accelerometer reliability at low frequency. Fig. 3
compares the bode diagrams of the rst- and second-order esti-
mation schemes. According to the Kalman ltering framework,
the rst-order complementary lter is optimal for a simpler
and less accurate model than the second-order complementary
lter. Therefore, the second-order lter is expected to yield
better performance. The second-order lter outperforms the
rst-order lter because it implements a sharper frequency
separation.
One of the advantages of the above formulation is the lter
tuning indications. Although the rst-order lter depends on
two variances, only the ratio between q and r is relevant. The
pole of both acceleration and front wheel transfer function is
p =

r
=
_
q
r
while the dc gain of the acceleration transfer function is
A
xgain
=
1

r
=
_
r
q
.
As intuition suggests, considering the front wheel noise vari-
ance r xed, when the acceleration noise variance q increases,
the lter pole shifts toward a higher frequency. This means that
the frequency range where the wheel velocity is considered
reliable is enlarged. Accordingly, also the acceleration dc gain
decreases in order to reduce the contribution of acceleration
noisy measurement.
For the second-order Kalman estimator, three noise parame-
ters have to be considered, i.e., q
1
, q
2
, and r. Let us consider
the process noise of (2) to be zero (q
1
= 0). As the most
critical aspect of the accelerometer is the bias, this assumption
is acceptable. In this condition, the natural frequency and
damping coefcient of the lter are

n
=
_

r
=

_
_
q
2
r
, =

r
2
n
=

2
2
.
The complex poles have a xed damping, and the ratio
between the variances of the two noises affects the frequency
of the poles with a square-root dependency. A nonzero q
1
mainly causes an increase of the damping coefcient while
CORNO et al.: TRACTION-CONTROL-ORIENTED STATE ESTIMATION FOR MOTORCYCLES 2403
10
2
10
1
10
0
10
1
10
2
80
60
40
20
0
frequency [Hz]
(a) (b)
m
a
g
n
i
t
u
d
e

[
d
B
]
1st order
2nd order
10
2
10
1
10
0
10
1
10
2
80
60
40
20
0
frequency [Hz]
m
a
g
n
i
t
u
d
e

[
d
B
]
1st order
2nd order
Fig. 3. (a) High-pass and (b) low-pass lters for the rst- and second-order Kalman frequency separation estimators.
116.3 116.4 116.5 116.6 116.7 116.8 116.9 117 117.1 117.2
85
90
95
100
time [s]
v
e
l
o
c
i
t
y

[
k
m
/
h
]
116.3 116.4 116.5 116.6 116.7 116.8 116.9 117 117.1 117.2
1
0
1
2
3
time [s]
v
e
l
o
c
i
t
y

[
k
m
/
h
]
vf
v
V
LF
v
VHF
(a)
(b)
Fig. 4. (a) Vehicle longitudinal velocity v, front wheel velocity v
f
, and
low-frequency component V
LF
, and (b) high-frequency component V
HF
from
a track test.
its effect on the frequency is negligible. In conclusion, the
Kalman observer can be written as two lters that can be
characterized either by their cutoff frequency or by the ratio
between the variances. In the remainder of this brief, we will
thus refer either to the ratio between the variances or the lter
cutoff frequency depending on the context.
C. Algorithm Tuning
To test the effectiveness of the proposed estimator, data
from an instrumented sport motorcycle have been used. Fig. 4
exemplies the method by showing the measured longitudi-
nal velocity v, the front wheel velocity v
f
, the two terms
V
HP
and V
LP
, and the velocity estimate v computed with the
second-order system with f
cutoff
= 0.8 Hz. The gure shows
a sudden acceleration of the motorcycle from around 84 km/h.
The effect of the frequency separation principle is clear:
removing the noise affecting the front wheel measurement
introduces a considerable phase lag, which is compensated
for by V
HF
. The advantages in terms of noise ltering are
57.5 58 58.5 59 59.5
130
135
140
145
150
v
e
l
o
c
i
t
y

[
k
m
/
h
]
57.5 58 58.5 59 59.5
0
2
4
6
time [s]
e
s
t
i
m
a
t
i
o
n

e
r
r
o
r

R
M
S

[
k
m
/
h
]
2
v
f
v
(vf v)
2
( v v )
2
(a)
(b)
Fig. 5. (a) Estimated longitudinal velocity v and front wheel velocity v
f
,
and (b) squared differences
_
v v
_
2
and
_
v
f
v
_
2
.
better appreciated in Fig. 5. The gure compares the esti-
mated longitudinal velocity obtained by using the second-
order Kalman lter and the estimation obtained using only
the front wheel velocity. The gure also plots the squared
estimation error of the two approaches. The gure conrms
that the proposed method considerably reduces the estimation
error without introducing a phase lag. In particular, an almost
10-fold reduction of the squared estimation error is obtained.
Besides providing an accurate longitudinal vehicle velocity
estimate with virtually no phase lag, fusing front wheel
velocity and longitudinal acceleration proves to be of key value
when wheelies occur. When a wheelie occurs, the front wheel
lifts off the ground and slows down. If the TC system uses
only the front wheel to estimate the wheel slip, it will act on
the incorrect assumption of excessive slip. Consider Fig. 6,
where a wheelie occurs at around 18.7 s. Even though the
rear wheel accelerates, the front wheel slows the vehicle down
because of friction. The vehicle velocity estimated with the
second-order Kalman lter is capable of providing a better
velocity estimate after the wheelie has occurred. The top plot
2404 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 6, NOVEMBER 2013
18 19 20 21 22 23 24
0
50
100
150
200
time [s]
v
e
l
o
c
i
t
y

[
k
m
/
h
]
v
{2- 0. 8- 0. 3} Hz
v
vr
v
f
110 110.2 110.4 110.6 110.8 111
0
5
10
15
20
time [s]
s
l
i
p
[
%
]
0.1 s
re ar whe e l sl i p
based on v
re ar whe e l sl i p
based on vf
re ar whe e l sl i p
based on v
(a)
(b)
Fig. 6. (a) Vehicle velocity estimation during a wheelie, for different values
of cutoff frequency Kalman lter and (b) estimated slip during a wheelie,
using the estimated velocity signal with different values of cutoff frequency
Kalman lter.
also shows the effect of three different cutoff frequencies
(2, 0.8, and 0.3 Hz). The lower the cutoff frequency, the
longer the estimate is reliable after the wheelie occurs.
Section III-B provides indications on how to tune the cutoff
depending on the ratio between the noises. This capability
is critical for TC. Consider the rear wheel slip depicted in
Fig. 6, which was computed using a noncausal ltered front
wheel velocity and the second-order Kalman lter. If only
the front wheel measurement is used, the estimated wheel
slip quickly diverges during a wheelie. In the case of a
threshold-based TC, this will cause a premature intervention
of the torque limitation. The proposed sensor fusion method
is, instead, capable of delaying the intervention of the TC
system for at least 100 ms. This does not completely solve
the problem (eventually the estimation will diverge during
prolonged wheelies); however, it has two main advantages.
On one hand, when accelerating, professional riders are able
to keep the front wheel just on the limit of a wheelie; under
these conditions, characterized by subsequent rapid wheelies,
the proposed lter yields an accurate estimation. On the other
hand, the proposed lter can be combined with a wheelie
detection algorithm. In this case, the capability of the velocity
estimation to delay the TC intervention sets the requirement
for the wheelie detection. In order to avoid unwanted TC
activation, the wheelie must be detected before the torque
is cut off because of perceived excessive wheel slip. The
designed lter gives the detection algorithm around 100 ms in
which the detection can be conrmed to avoid false activations.
The following section is devoted to the development of such
an algorithm.
82 82.5 83 83.5 84 84.5 85 85.5 86
60
80
100
120
w
h
e
e
l

s
p
e
e
d

[
k
m
/
h
]
82 82.5 83 83.5 84 84.5 85 85.5 86
0
5
10
time [s]
f
r
o
n
t

s
u
s
p
e
n
s
i
o
n


s
t
r
o
k
e

[
c
m
]
v
r
vf
v
f
vr
whe el i e starts whe el i e ends
Fig. 7. Compression of the front suspension during severe acceleration.
IV. WHEELIE DETECTION
Wheelies are critical phenomena that can be easily detected
by measuring the following as below.
1) The front suspension stroke: During acceleration, the
front suspension extends until it reaches its maximum
length, when the wheel lifts off the ground. Fig. 7
depicts an example of such behavior. Stroke sensors are
expensive and fragile and they are not amenable to large-
scale production.
2) The vehicle pitch angle (see Fig. 1): The front wheel
lifts off at a known

. Unfortunately, the measurement
of pitch angle is not easily obtained unless more sensors
are added (see [21]).
The need for a cost-effective detection algorithm thus arises.
The algorithm should to be easily implementable, robust,
and, when working with the previously presented velocity
estimation, be able to detect a wheelie in approximately 0.1 s.
Fig. 6 shows the discrepancy between the front and rear
wheel velocity during a wheelie: the rear wheel accelerates,
whereas the front wheel decelerates due to friction. The effect
of a wheelie on wheel velocity can be modeled. When the
front wheel is in contact with the ground, the wheel dynamics
is described considering that the front wheel slip is negligible
and the front wheel velocity is equal to the vehicle velocity.
Thus
v
f
= a
x
+
1
y = v
f
+
1
(8)
where the longitudinal acceleration is considered as the input,
and the measured front wheel velocity the output. Process
1
and measurement noise
1
are modeled as zero-mean white
noises with variance q
1
and r
1
, respectively. When the front
wheel lifts off the ground, friction determines the front wheel
deceleration, resulting in the following dynamics:
v
f
= +
2
y = v
f
+
2
(9)
where is the equivalent linear deceleration of the wheel due
to friction, and
2
,
2
are zero-mean white noises with variance
q
2
and r
2
.
Remark 1: The model presented in (8) is analogous to (6).
In this context, the rst-order model is preferred, neglecting
CORNO et al.: TRACTION-CONTROL-ORIENTED STATE ESTIMATION FOR MOTORCYCLES 2405
the presence of longitudinal acceleration offset. The inclusion
of such offset would make the model insensitive to constant
decelerations and the wheelie detection more difcult. This can
be understood by noticing that the main difference between
(8) and (9) is due to a constant acceleration value. Should the
accelerometer offset be consistent, its estimate provided by the
Kalman lter used to estimate the vehicle velocity can correct
the sensor output.
A. Detection Algorithm
Models (8) and (9) describe two possible operating modes.
The wheelie detection algorithm can be posed as a fault-
detection problem (see [22]): given the available measure-
ments, determine in which mode the system is operating.
The core of the algorithm is two separate Kalman estimators,
based on the above dynamic models. They provide two differ-
ent estimates of the front wheel velocity v
nw
(t ) corresponding
to the normal condition, and v
w
(t ) corresponding to the
wheelie condition. The estimator with the lowest residual,
dened as
e
nw
(t ) =
_
v
f
(t ) v
nw
(t )
_
2
, e
w
(t ) =
_
v
f
(t ) v
w
(t )
_
2
indicates the current operating mode.
The Kalman lters are derived as in Section III-B. For
the normal condition, the inputs are the longitudinal acceler-
ation and front wheel velocity, whereas the wheelie condition
observer uses only the front wheel velocity as an input;
is a tuning parameter to be discussed later. Fig. 8 provides
an example of such velocity estimates. The velocity plot
conrms that, when the wheel rolls on the ground, the best
estimation is provided by the normal condition observer.
When a wheelie occurs, the normal condition estimate diverges
from the real measurement and the wheelie condition observer
better ts the measurement. The residuals prove useful for
the practical implementation of wheelie detection: a simple
but effective detection law triggers when the residual of the
wheelie condition observer is lower than the residual of the
no wheelie condition observer. Conversely, when the normal
condition observer residual is lower than the wheelie condition
observer model residual, the front wheel is in contact with the
ground.
The above considerations are valid only if the models on
which they are based are valid. There exist conditions in
which neither of the models is valid (e.g., during braking).
In these conditions, the core algorithm cannot be employed.
To correctly manage all the possible conditions, the detection
algorithm is implemented as a three-state automaton, described
below.
1) The BRAKING mode is active when the motorcycle
is not accelerating. In this condition, the residuals are
disregarded, as neither model is valid and the state is
set to x = 0 (i.e., no wheelie); a wheelie cannot occur
when not accelerating. The mode is activated by either
the brake lever switch or a
x
< 2 m/s
2
.
2) The STANDBY mode corresponds to a fault in the
algorithm. It is activated when both residuals are higher
than 5 m/s and the automaton is not in BRAKING mode.
42.5 43 43.5 44 44.5
60
70
80
90
100
s
p
e
e
d

[
k
m
/
h
]
42.5 43 43.5 44 44.5
0
5
10
time [s]
r
e
s
i
d
u
a
l
s

[
k
m
2
/
h
2
]
normal model
wheelie model
norm al
c ondi ti on
whe el i e det e c ti on
whe el i e
c ondi ti on
vf
v
nw
v
w
Fig. 8. Measured and estimated front wheel velocity for wheelie detection
and model residuals.
In this state, x is not determined. To exit from this state,
both residuals have to drop below the threshold for at
least 5 s.
3) The ACCELERATING mode is active when the
possibility of a wheelie is present, i.e., the motorcycle
is accelerating, and one residual is lower than 5 m
2
/s
2
.
In the ACCELERATING mode
x =
_
0 (normal), if e
nw
(t ) < e
w
(t )
1 (wheelie), if e
nw
(t ) e
w
(t ).
(10)
B. Algorithm Tuning
The proposed method has three tuning parameters: the
parameters of the Kalman lters of both models and . The
friction coefcient models friction and can be identied
by applying linear identication techniques to a lifted wheel
deceleration test. From experiments, we identied = 5 m/s
2
.
The tuning of the observers is more critical. The main
tuning parameters of the Kalman lters are the ratio
between measurement and process noise or, alternatively, their
cutoff frequency f
cut,nw
for the normal condition observer and
f
w
cut,
for the wheelie condition observer. The accuracy and
readiness on the detection depend on the tuning of the two
cutoff frequencies. Consider Figs. 9 and 10. Fig. 9 shows two
different tunings of the Kalman lter for the normal condition
mode. The low q/r ( f
cut,nw
= 0.1 Hz) tuning leads to a
biased velocity estimate. This causes the normal condition
residual to be greater than the wheelie model ones; thus,
according to the wheelie detection rules, a wheelie is wrongly
detected. The comparison with the higher q/r ratio ( f
cut,nw
=
1.5 Hz) shows that no wrong detection occurs in this case.
However, the correct detection occurs with a delay of 30 ms,
showing a tradeoff between false positives and detection
readiness.
Similar considerations hold when analyzing different tun-
ings for the wheelie condition Kalman lter, as shown in
2406 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 6, NOVEMBER 2013
34.8 34.9 35 35.1 35.2 35.3 35.4 35.5 35.6
65
70
75
80
s
p
e
e
d

[
k
m
/
h
]
34.8 34.9 35 35.1 35.2 35.3 35.4 35.5 35.6
0
5
10
15
20
time [s]
r
e
s
i
d
u
a
l

[
k
m
2
/
h
2
]
wheelie model residual
normal residual, q/r low
normal residual, q/r high
correct
detection
false positive
vf
vw
v
nw
,
q / r l ow
v
nw
,
q / r h i gh
Fig. 9. Example of different tunings for the normal condition model, and
their effects on the wheelie detection algorithm.
34.8 34.9 35 35.1 35.2 35.3 35.4 35.5 35.6
65
70
75
s
p
e
e
d

[
k
m
/
h
]
34.8 34.9 35 35.1 35.2 35.3 35.4 35.5 35.6
0
5
10
15
20
time [s]
r
e
s
i
d
u
a
l

[
k
m
2
/
h
2
]
normal model residual
wheelie residual q/r low
wheelie residual q/r high
detection
vf
v
w
, q /r l ow
v
w
, q /r hi gh
v
nw
Fig. 10. Example of different tunings for the wheelie model, and their effects
on the wheelie detection algorithm.
Fig. 10. When the reliability of the model is weighted more
than that of the noise (i.e., q/r low or f
w
cut,
= 1 Hz), the
wheelie detection is more robust (the residuals are farther
apart). Increasing q/r (or increasing the bandwidth, f
w
cut,
=
2.5 Hz) reduces the robustness of the detection but the wheelie
detection is speeded up with respect to the low q/r case
(50 ms), again showing the tradeoff between robustness and
performance.
The above discussion on tuning tradeoff is made quantitative
by the denition of two cost functions. Consider a test run
of duration

T [s] during which N
wheelie
wheelies occur;
0
5
10
0
5
10
0
0.1
0.2
0.3
wheelie condition
model cut off [Hz]
normal condition
model cut off [Hz]
m
i
s
s

i
d
e
n
t
i
f
i
c
a
t
i
o
n

[
s
]
J
false negative
J
false positive
Fig. 11. Wheelie detection tradeoff. cost functions for different values of
the cutoff frequencies of the two lters.
48.4 48.6 48.8 49 49.2 49.4 49.6
70
75
80
85
90
95
s
p
e
e
d

[
k
m
/
h
]
48.4 48.6 48.8 49 49.2 49.4 49.6
0
2
4
6
time [s]
s
u
s
p
.

s
t
o
k
e

[
c
m
]
w
h
e
e
l
i
e

d
e
t
e
c
t
.

[
0

1
]
front suspension
wheelie detection
0.044s
v
r
v
f
v
nw
v
w
Fig. 12. Comparison between direct wheelie detection via front suspension
elongation and the proposed wheelie detection algorithm.
then
J
falsepositive
=
_
T
0

1
(t , f
cut,nw
, f
cut,w
)dt
N
wheelie
J
falsenegative
=
_
T
0

2
(t , f
cut,nw
, f
cut,w
)dt
N
wheelie
where

1
(t , f
cut,nw
, f
cut,w
) =
_
1, if x = 1 and x = 0
0, otherwise

2
(t , f
cut,nw
, f
cut,w
) =
_
1, if x = 0 and x = 1
0, otherwise.
Intuitively, J
falsenegative
represents the average wheelie detec-
tion delay, whereas J
falsepositive
is the average detection antic-
ipation. The objective is minimizing both cost functions.
CORNO et al.: TRACTION-CONTROL-ORIENTED STATE ESTIMATION FOR MOTORCYCLES 2407
10 8 6 4 2 0
0
0.05
0.1
0.15
0.2
0.25
0.30
0.35
[m/s
2
]
J

[
s
]
J
false negative
J
false positive
Fig. 13. Wheelie detection tradeoff. Cost functions for different values .
Fig. 11 plots the cost functions for multiple laps at the
Mugello Racing Circuit. The test is 320 s long and consists
of 70 wheelies. As expected, by decreasing the wheelie
lter cutoff frequency, the estimation is made more reactive
(limiting the detection delay). However, this comes at the cost
of false positives; the converse is true for a high wheelie
lter cutoff frequency and also for the tuning of the normal
condition model. From the previous analysis, the optimal
tuning has been determined to be f
cut,w
= 4.5 Hz and
f
cut,nw
= 2.5 Hz. This tuning yields an average delay of
70 ms and J
falsepositive
= 50 ms, which are well within the slip
estimation algorithms limits.
1
Fig. 12 shows a comparison
between the suspension compression and the binary signal
provided by the proposed tuning.
At the beginning of the subsection, the tuning of was
briey discussed. Being a friction coefcient, is expected to
change during the life of the motorcycle. The same protocol
used to tune the Kalman lters is employed exploring several
values of . The results are shown in Fig. 13. The results
indicate that the method is robust to inaccurate identication
of . In particular, in order to limit the false positive rate, it
is suggested to overestimate rather than underestimate it.
V. CONCLUSION
In this brief, two important estimation problems for motor-
cycle TC applications were discussed: the estimation of the
vehicle longitudinal velocity and wheelie detection. For both
problems, effective and easy to implement algorithms were
proposed.
Two methods have been discussed for the estimation of the
longitudinal velocity: a complementary lter and a Kalman
lter. Experimental tests validated the methods, showing the
benets of the Kalman lter in both reducing the wheel noise
and in estimating the vehicle longitudinal velocity during
wheelies. The proposed estimation algorithm yielded a 100-ms
period useful to avoid unwanted triggering of the TC.
The proposed estimation method could not indenitely
estimate the longitudinal velocity during a wheelie. An effec-
tive and viable wheelie detection algorithm was therefore
developed. The method was developed as a fault detection
algorithm. The tuning of the algorithm was discussed, and
1
With the proposed tuning, no isolate false positive occurs, only anticipa-
tions with respect to the signal from the suspension.
the method experimentally validated, showing that an average
delay of 70 ms could be achieved.
REFERENCES
[1] M. Corno, M. Gerard, M. Verhaegen, and E. Holweg, Hybrid ABS
control using force measurement, IEEE Trans. Control Syst. Technol.,
vol. 20, no. 5, pp. 12231235, Sep. 2012.
[2] W. Pasillas-Lepine, A. Loria, and M. Gerard, Design experimental
validatation nonlinear wheel slip control algorithm, Automatica, vol. 48,
no. 8, pp. 18521859, Aug. 2012.
[3] T. Johansen, I. Petersen, J. Kalkkuhl, and J. Ldemann, Gain-
scheduled wheel slip control in automotive brake systems, IEEE
Trans. Control Syst. Technol., vol. 11, no. 6, pp. 799811,
Nov. 2003.
[4] M. Corno, S. M. Savaresi, and G. Balas, On linear parameter varying
(LPV) slip-controller design for two-wheeled vehicles, Int. J. Robust
Nonlinear Control, vol. 19, no. 12, pp. 13131336, 2009.
[5] M. Corno, S. Savaresi, M. Tanelli, and L. Fabbri, On opti-
mal motorcycle braking. Control Eng. Pract., vol. 16, no. 6,
pp. 644657, 2008.
[6] T. Kazuhiko, N. Yutaka, N. Takehiko, T. Kazuya, and T. Makoto,
Control technology of brake-by-wire system for super-sport
motorcycles, in Proc. SAE World Congr. Exhibit., Apr. 2010,
pp. 1315.
[7] M. Tanelli, C. Vecchio, M. Corno, A. Ferrara, and S. Savaresi, Traction
control for ride-by-wire sport motorcycles: A second-order sliding mode
approach, IEEE Trans. Ind. Electron., vol. 56, no. 9, pp. 33473356,
Sep. 2009.
[8] S. Savaresi, C. Poussot-Vassal, C. Spelta, O. Sename, and L. Dugard,
Semi-Active Suspension Control Design for Vehicles. Amsterdam, The
Netherlands: Elsevier, 2010.
[9] P. De Filippi, M. Tanelli, M. Corno, and S. Savaresi, Enhanc-
ing active safety of two-wheeled vehicles via electronic stability
control, in Proc. 18th IFAC World Congr. Autom. Control, 2011,
pp. 638643.
[10] S. Evangelou, D. Limebeer, R. Sharp, and M. Smith, Control of
motorcycle steering instabilities, IEEE Control Syst. Mag., vol. 26,
no. 5, pp. 7888, Oct. 2006.
[11] P. De Filippi, M. Tanelli, M. Corno, S. M. Savaresi, and L. Fab-
bri, Semi-active steering damper control in two-wheeled vehicles,
IEEE Trans. Control Syst. Technol., vol. 19, no. 5, pp. 10031020,
Sep. 2011.
[12] A. Teerhuis and S. Jansen, Motorcycle state estimation for lateral
dynamics, Veh. Syst. Dynamics, vol. 50, no. 6, pp. 12611276,
2012.
[13] V. Cossalter, R. Lot, and M. Massaro, An advanced multibody code
for handling and stability analysis of motorcycles, Meccanica, vol. 46,
no. 5, pp. 945958, 2011.
[14] S. Savaresi and M. Tanelli, Active Braking Control Systems Design for
Vehicles. New York: Springer-Verlag, 2010.
[15] M. Corno and S. Savaresi, Experimental identication of engine-to-
slip dynamics for traction control applications, Eur. J. Control, vol. 16,
no. 1, pp. 88108, 2010.
[16] F. Jiang and Z. Gao, An adaptive nonlinear lter approach to the vehicle
velocity estimation for ABS, in Proc. IEEE Int. Conf. Control Appl.,
Sep. 2000, pp. 490495.
[17] I. Boniolo and S. Savaresi, Estimate of the Lean Angle of Motorcycles.
New York: VDM Verlag, 2010.
[18] R. G. Brown and P. Y. C. Hwang, Introduction to Random Signals and
Applied Kalman Filtering, vol. 1. New York, USA: Wiley, 1992.
[19] L. Ray, Nonlinear state and tire force estimation for advanced vehicle
control, IEEE Trans. Control Syst. Technol., vol. 3, no. 1, pp. 117124,
Mar. 1995.
[20] E. Lefferts, F. Markley, and M. Shuster, Kalman ltering for space-
craft attitude estimation, J. Guid., Control, Dynamics, vol. 5, no. 5,
pp. 417429, 1982.
[21] I. Boniolo, S. Corbetta, and S. Savaresi, Attitude estimation of a
motorcycle in a Kalman ltering framework, in Proc. 6th IFAC Symp.
Adv. Autom. Control, 2010, pp. 779784.
[22] R. Isermann, Fault-Diagnosis System: An Introduction From
Fault Detection Fault Tolerance. New York: Springer-Verlag,
2006.

You might also like