16th IFAC Symposium on System Identification

The International Federation of Automatic Control

Brussels, Belgium. July 11-13, 2012

On the Periodic Noise Affecting Wheel

Speed Measurement ?
Giulio Panzani, Matteo Corno and Sergio M. Savaresi ∗
∗
Politecnico di Milano, Dipartimento di Elettronica e Informazione.
Piazza L. da Vinci, 31, 20133. Milano, Italy (e-mail:
{panzani,corno,savaresi}@elet.polimi.it).

Abstract: This paper addresses the measurement of the wheel angular velocity of wheeled
vehicles. The wheel angular velocity measurement is often affected by large periodic distur-
bances. The paper proposes a model of the origin of that disturbance: even a small offset in the
positioning of the encoder can cause large, velocity dependent noise. After having shown that
the periodic term is indeed a noise and not part of the actual signal, an adaptive notch filter
is discussed. Lyapunov theory is employed to prove the quadratic stability of the time-variant
filter. An analysis of both simulation and experimental tests supports the analysis and methods.

Keywords: Speed measurement, Automotive control, Measurement noise, Adaptive filters,

Lyapunov stability

1. INTRODUCTION

In the past decades the automotive industry has wit-

nessed the rise of electronics. From being the epitome
of mechanical engineering, the automotive industry be-
came a perfect example of the potential of mechatronics:
bringing together mechanical, electronics, control system
and software engineering. The electronic systems onboard
commercial vehicles are numerous and diverse; active ve-
hicle dynamics systems are critical to improve safety and
save lifes. Active vehicle dynamics systems (such as ABS
(Corno et al. [2009]), traction control (Tanelli et al. [2009]) Fig. 1. Rear wheel encoder mounted on a motorbike.
and semi-active suspensions (Savaresi et al. [2010])) em-
ploy a set of sensors to measure the state of the vehicle,
actuators to a influence the state of the vehicle and micro- All those techniques that assume the availability of a
processors to decide the action to take given the current model of the system whose velocity is under estimation
and past measurement. fall in the model-based approach. In the literature, the
problem has been addressed with techniques such as model
The sensor lay out strongly varies according to the control reference adaptive systems (Peng and Fukao [1994], Levi
system; however most control systems rely on the mea- and Wang [2002]), sliding mode observers (Proca et al.
surement of the rotational speed of the wheels as the basic [2003]), neural networks (Ben-Brahim et al. [1999], Kim
building block. et al. [2001]), Kalman filters (Belanger et al. [1998])
and extended Kalman filters (Atkinson et al. [1991], Shi
The most commonly used technology to measure wheel
et al. [2002]). These techniques have had many successful
velocity is based on discrete encoders. Different types of
applications for electrical motors for which simple and
discrete position encoders exist. They are mainly made of
fairly accurate models exist. Conversely, in the automotive
two elements: a rotating part with teeth (or lines) and a
field accurate wheel models are rare, since the dynamics is
fixed sensor which detects the passage of a line, outputting
affected by the highly nonlinear road friction dynamic; this
a pulse for each line. Figure 1 depicts the rear wheel
reason makes the non model based approach more suited
encoder of a motorbike. Digital encoders do not directly
for automotive applications.
measure the rotational velocity; it must be reconstructed
from the received pulses. The velocity estimation problem Signal-based approaches are based on signal processing
is usually addressed with two approaches; either model- techniques; the signal processing community has produced
based or signal-based techniques can be applied. plenty of algorithms. There are solutions based on first
order approximation of the velocity (Kavanagh [2000,
2001], Ohmae et al. [1982]), higher order approximations
? This work has been partially supported by MIUR project ’New (Brown and Schneider [1987]), exact polynomial interpo-
Methods for Identification and Adaptive Control for Industrial lation (Janabi-Sharifi et al. [2000]), least square interpo-
Systems’. lation of the position trajectory (Lemkin et al. [1995]),

978-3-902823-06-9/12/$20.00 © 2012 IFAC 1695 10.3182/20120711-3-BE-2027.00260

16th IFAC Symposium on System Identification
Brussels, Belgium. July 11-13, 2012

and on linear or nonlinear numerical integration (Su et al.

[2005, 2006]). Each of the mentioned algorithms has its
advantages and disadvantages that depend on velocity and
acceleration; in fact, it was shown by Phillips and Branicky
[2003], Lemkin et al. [1995] through simulations that for
a system with a wide range of speeds a best velocity
estimator does not exist. The same result is achieved in
Bascetta et al. [2007] with the help of analytical tools.
It should be noted that, despite the numerous available
signal-based methods, they all can be seen as refinements,
specializations or improvements of two basic techniques: 200

the lines per period algorithm and the fixed position

150
algorithm.

Speed [km/h]
As shown by the brief outline of the available literature, 100

the problem of rotational velocity estimation is well ad-

50
dressed; however to the best of the authors knowledge only
Corno and Savaresi [2010] refers to the problem of the peri- 0
odic noise that affects all the above estimation algorithms. 20 40 60 80 100 120
Time [s]
140 160 180 200 220

The problem is graphically depicted in Figure 2. The figure

Fig. 3. Wheel speed spectrogram (top) and speed signal

200
(bottom).
180

160 The paper is structured as follows: in Section 2 different

possible causes for the disturbance/noise are analyzed; an
140
accurate modeling of the encoder and wheel system allows
Speed [km/h]

120 to conclude that the observed phenomenon is indeed a

100 measurement noise and not a disturbance. In Section 3 the
80 already cited solution is re-looked at the stability of the
filter is proven. The paper is closed by some conclusions.
60

40

20
2. ENCODER-WHEEL MODELING
0
0 50 100 150 200 250 Different causes for periodic behavior are taken into con-
Time [s]
sideration. Usually, periodic disturbances are induced by
unbalanced axis, imperfections or cracks (see e.g. Mc-
Fig. 2. Example of periodic noise on wheel speed signal. Fadden and Smith [1984]). This is not the case of the
considered motorcycle wheels, since such vibrations are
shows the logged front wheel velocity (multiplied by the not present in any other signals such as the suspension
radius to get the equivalent vehicle velocity) during a compression or wheel vertical acceleration.
lap on a test track. In this figure and in the rest of the Another source of a periodic noise can be the wheel speed
paper we will refer to the fixed position algorithm as the estimation algorithm, itself. In particular, it is well known
estimation algorithm of choice. From figure a clear periodic that the line per period algorithms are subject to periodic
component is observed. Further analysis reveals that the oscillations; however the fixed positions (FP) algorithm
frequency of that component is locked with the rotational (employed herein) is not subject to these phenomena
frequency of the wheel. This can be easily observed in
Figure 3. The figure plots the measured velocity spectro- To understand another possible cause, consider Figure
gram during the same lap. It shows the main frequency 4 where a sketch of a wheel along with the encoder is
component of the signal as function of the time; the lower depicted. The encoder is represented as a series of rays
panel plots the wheel speed. A clear linear correlation is of length D that originate form the encoder center E.
observed between the velocity and the main component of Each ray should be considered as the edge that divides
the disturbance/noise term. It is clearly visible how the an empty from a full part of the encoder: in the FP
most important harmonic component in the considered algorithm every time that the hall-effect sensor encounters
signal varies its frequency following the same path of the such discontinuity a timer starts to measure the time gap
speed signal. until the next transition. The FP algorithm assumes that
the encoder is fixed with the wheel, rotates around the
These phenomena was already notices by Corno and same axis and has equidistant lines.
Savaresi [2010] and an ad-hoc solution based on a notch
filter was proposed. Being outside of the main scope of that Now consider the fact that the encoder may be mounted
paper, no attempt at identifying the cause nor a rigorous with an offset e, thus E does not coincide with the wheel
proof of the performance of the provided solution were rotation axis C. The hall-effect sensor H is fixed with the
provided. This paper aims at addressing the above issues. motorcycle frame and it is located at distance R form the

1696
16th IFAC Symposium on System Identification
Brussels, Belgium. July 11-13, 2012

These oscillations can be better understood by recalling

the system equations (1) that describe the generic point
that passes the hall-effect sensor. Under the assumption of
e << R, an is almost constant and equal to R. The first
equation of the algebraic system (1) can be thus written
as:
0 = e cos (θn ) + R cos (θn − n∆) .
Using the Euler formula, after some algebraic manipula-
tions, the following expression arises:
 
0 = M cos θ̂ − φ (∆, n) (5)
where p
M = R2 + 2Re cos n∆ + e2
R sin n∆
φ (∆, n) = arctan
Fig. 4. Schematic representation of the encoder, wheel, hall e + R cos n∆
sensor assemblyl. which is solved by
π
wheel center. The nth line of the encoder is described in θ̂ = φ (∆, n) + .
2
the absolute non-rotating reference centered in C by:
If the above expression is plugged in (4) taking the limit
x = e cos (θ) + a cos (θ − n∆) for ∆ → 0, the estimated speed becomes:
y = e sin (θ) + a sin (θ − n∆) 1
with 0 < a < D, where θ is the wheel angular position lim ω̂ = θ̇ .
∆→0 ∂φ
with respect to the reference frame and ∆ is the encoder
angular resolution that depends on the number of teeth ( ∂∆
N ) of the encoder, ∆ = 2π/N . The FP timer starts when By trivial calculation, the above expression can be written
a line passes in front of the hall sensor. As the position of as:
e e + R cos (n∆)
the hall sensor in the non-rotating reference is (0, R), this lim ω̂ = θ̇ +1
happens for θn and an such that: ∆→0 R R + e cos (n∆)
0 = e cos (θn ) + an cos (θn − n∆) that under the assumption e << R becomes:
(1) e e + R cos (n∆)
R = e sin (θn ) + an sin (θn − n∆)
lim ω̂ = θ̇ + 1.
with 0 < an < D. The FP algorithm estimates the wheel ∆→0 R R
velocity as: By noting that
∆ e + R cos (n∆) ≈ R cos φ
ω̂ = (2)
tn+1 − tn the final estimated angular speed expression is:
where tn+1 − tn is the time between two subsequent e
 
passages of encoder line. At a constant angular rate ω ω̂ = ω cos φ + 1
R
θn = ω tn + θ(0) θn+1 = ω tn+1 + θ(0) (3) π
Hence the algorithm (2) can be expressed as function of which plugging in ωt = φ + returns
2
the angular values at which each encoder tooth passes the 
e

π
 
hall-effect sensor: ω̂ = ω cos ωt − +1 (6)
∆ R 2
ω̂ = ω . (4)
θn+1 − θn The expression in equation (6) is based on the assumption
Equation (4) shows that the measured speedω̂ is the actual that the encoder offset is small with respect to its radius.
To appreciate the effectiveness of such approximation,

wheel speed ω if and only if the ratio ∆/ θ̂n+1 − θ̂n is
Figure 6 shows a comparison between the estimated speed
equal to 1. The ratio is constant only if the encoder offset according to the complete expression (4) and the simplified
e is null. Figure 5 shows the difference θ̂n+1 − θ̂n and the (6) for two different values of ratio e/R (e/R = 0.02 for
amplitude an (computed numerically solving equations (1) the left plot and e/R = 0.2 for the right plot). It should be
along with the estimated speed (according to (4)) for the noticed that the approximated formula perfectly describes
following parameters: the measured speed, when the e/R ratio is small. Notice
ω = 62.8319 [rad/s] that in the left panel the encoder has a radius of 5cm and
R = 5 [cm] an offset of 0.1cm; despite the very small machining error
e = 0.1 [cm] a significant error on the measured vehicle speed occurs:
N = 48 in fact, considering a wheel radius of 0.29m, the measured
speed sets in the range 65.5 ± 0.4 km/h.
The figure clearly exemplifies that a small encoder offset The analysis of the approximated FP algorithm speed
generates considerable oscillations in the velocity mea- estimate supports to the following considerations:
surement at the wheel revolution frequency. Note that, as
expected, the distance from the encoder center of the hall (1) The measured speed mean value is equal to the real
sensor has a fluctuation of amplitude e. wheel speed and the period of oscillations is equal

1697
16th IFAC Symposium on System Identification
Brussels, Belgium. July 11-13, 2012

5.2 0.135 64.5

ω̂
5.15 0.134 64 ω

[r ad]

angular speed [rad/s]

5.1 0.133
63.5

n
0.132
a [cm]

5.05

θ
63

n + 1−
n

5 0.131
62.5

θ
4.95 0.13

4.9 0.129 62

4.85 0.128 61.5

0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5
Wheel revolutions Wheel revolutions Wheel Revolutions

Fig. 5. Distance from encoder center of the hall sensor (upper-left), effective angular displacement between two
consecutive triggers (upper right), estimated and actual angular velocity (lower).

real 80 its derivatives may suffer from the component at 1X; this
65 measured complete
measured complete 75
disturbance is remarkably large, and may become critical if
Angular speed [rad/s]

it falls within the closed-loop bandwidth. For motorcycles,

Angular speed [rad/s]

64 real
70 measured complete
63 measured simplified the 1X frequency corresponds to 10 Hz when the vehicle
62
65
speed is about 60 km/h. Since the disturbance can be
60 critical, in Corno and Savaresi [2010] it was proposed to
61
55 filter all the wheel-speed signals using the adaptive notch-
60
0.1 0.2 0.3 0.4 0.5 50
0.1 0.2 0.3 0.4 0.5
filtering scheme illustrated in Figure 8.
Time [s] Time [s]

Fig. 6. Comparison between real, estimated and approx- N

imate formula angular speed for different values of
e/R: e/R = 0.02 for the left and e/R = 0.2 for the
right plot.
to the wheel revolution frequency; thus fitting the
observed periodic component.
(2) The oscillations amplitude increases proportionally to
the angular speed. This is behavior is experienced in Fig. 8. Adaptive notch-filtering scheme.
real sampled data, as shown in Figure 7.
300 255

After a first low pass filtering stage (1st-order low-pass

250 filter at 20Hz), the 1X component is removed by a second
250 order adaptive notch filter Diniz [2008]. The wheel rolling
200 frequency is estimated using a low pass filter with 1 Hz
cut-off frequency. Once the rolling frequency is identified
Speed [km/h]

~3km/h

150 245
203 203.05 203.1 its main harmonic is removed via the adaptive filter:
70

100 68 s2 + ωn2
N (s) = (7)
66 s2 + 2ξωn s + ωn2
50 64
~0.8km/h
62
where ωn corresponds to the rolling frequency. Figure 9 ex-
0 60
emplifies the effect of the filtering scheme in the frequency
50 100 Time [s] 150 200 72 72.25 72.5
domain. It is also shown that the other parameter that
appears is the pole damping ratio ξ: according to depicted
Fig. 7. Experimental example of once-per-revolution noise: transfer function, this parameter can be used to modify
the increase of amplitude with speed is highlighted. the selectiveness of the notch filter.
If the notch frequency is constant, the stability of the
Finally, it can be noticed that with the mentioned values proposed filter is trivial: 0 < ξ < 1 guarantees the
for encoder radius and offset, the model in equation (6) asymptotic stability of the filter. Conversely, when the
predicts an amplitude oscillation similar to the experimen- notch frequency ωn varies, the system is not more linear
tal ones. This confirms the proposed once-per-revolution time invariant anymore and 0 < ξ < 1 may not be enough
modeling. to guarantee stability (Khalil [1996]); luckily the following
can be proven:
3. ADAPTIVE FILTERING Theorem 1. The adaptive notch filter in equation (7) is
asymptotically stable for any notch frequency trajectory
The above analysis has shown the main noise affecting ωn = ωn (t) > 0.
the wheel rotational speed. From the control perspective,
any control loop closed on the wheel velocity or any of Let us consider a state-space realization of the notch filter:

1698
16th IFAC Symposium on System Identification
Brussels, Belgium. July 11-13, 2012

of many wheeled vehicles. In the first part of the paper a

10 model was proposed to explain the origin of the periodic
Magnitude [dB]

0 term. The analysis indicates that a small offset in the ma-

−10 chining of the encoder is sufficient to explain the measured
ξ = 0.1
−20 ξ = 0.3
noise. This analysis concludes that it is a measurement
noise and not an actual signal component. This conclusions
−30
is the employed in the second part of the paper where an
−40 −1
10 10
0 1
10
2
10 adaptive notch filtering technique is recalled and applied
to the problem. The stability properties of the notch filter
100 are discussed and commented upon.
50 Throughout the paper experimental tests have been em-
Phase [°]

0
ployed to support the analysis and test the proposed
methods.
−50

−100 −1
10 10
0 1
10
2
10
REFERENCES
Frequency [Hz]
D. Atkinson, P. Acarnley, and J. Finch. Observer for In-
duction Motor State and Parameter Estimation. IEEE
Fig. 9. Adaptive notch filter scheme for once-per- Transactions on Industry Applications, 27:1119–1127,
revolution noise suppression. 1991.

−2ξωn ωn
 
2ξωn
 L. Bascetta, G. Magnani, and P. Rocco. Velocity Estima-
ẋ = x+ u tion: Assessing the Performance of Non Model-Based
−ωn 0 0
Techniques. Advanced intelligent mechatronics, 2007
y = [ −1 0 ] x + [1] u ieee/asme international conference on, pages 1–6, Sept.
2007.
Since ωn = ωn (t) the system can be written as a linear P. Belanger, P. Dobovolny, A. Helmy, and X. Zhang. Esti-
time variant (LTV) system with equations mation of Angular Velocity and Acceleration from Shaft-
ẋ = A (ωn ) x + B (ωn ) u Encoder Measurements. The International Journal of
Robotics Research, 17(11):1225–1230, 1998.
y = Cx + D u L. Ben-Brahim, S. Tadakum, and A. Akdag. Speed Control
The above system is said to be quadratically stable if there of Induction Motor Without Rotational Trasducers.
exists a time invariant matrix P > 0 such that: IEEE Transactions on Industry Applications, 35(4):844–
T 850, 1999.
A (t) P + P A (t) < 0 R.H. Brown and S. Schneider. Velocity Observations from
at all time t. In such a case, V (x) = xT P x is a Lyapunov Discrete Position Encoders. In IEEE Conference on
function for the above system. The notch filter state matrix Decision and Control, volume 6, pages 1111–1118, 1987.
can be written  as follows:  M. Corno and S.M. Savaresi. Experimental identification
of engine-to-slip dynamics for traction control applica-
 
−2ξωn ωn −2ξ 1
A (ωn ) = = ωn = ωn à tions. European Journal of Control, 16(1):88–108, 2010.
−ωn 0 −1 0
M. Corno, Savaresi S.M., and G.J. Balas. On linear
and it can be easily noticed that à is Hurwitz. Thus the parameter varying (LPV) slip-controller design for two-
Lyapunov equation can be written as: wheeled vehicles. International Journal of Robust and
Nonlinear Control, 19(12):1313–1336, 2009.
h i
ωn (t) ÃT P + P Ã < 0
P.S.R. Diniz. Adaptive filtering: algorithms and practical
Since ωn (t) is a positive scalar and à is Hurwitz, matrix implementation, volume 694. Springer Verlag, 2008.
P > 0 satisfying the above equation exists (see Khalil F. Janabi-Sharifi, V. Hayward, and C. Chen. Discrete-time
[1996]). The stability of the notch filter is thus proved. Adaptive Windowing for Velocity Estimation. IEEE
It is of interest to notice that the stability is guaranteed Transaction of Control Systems Technology, 8(6):1003–
for any positive value of ωn , regardless of the speed with 1009, 2000.
which the frequency varies. R.C. Kavanagh. Shaft Encoder Characterization via The-
oretical Model of Differentiator with Both Differential
In Figure 10 the power spectra of the measured signal and and Integral Nonlinearities. IEEE Transactions on In-
the adaptively filtered one are shown. For comparison a strumentation and Measurements, 49(4):795–801, 2000.
linear low pass filter of the second order is shown. In order R.C. Kavanagh. Performance Analysis and Compensation
to reach the same 1X attenuation level of the adaptive of M/T-Type Digital Tachometers. IEEE Transactions
structure with a low pass filter, a cut off frequency of 4 Hz on Instrumentation and Measurements, 50(4):965–970,
is required. Such a low pass filter filters out many of the 2001.
relevant features of the signal. H.K. Khalil. Nonlinear Systems. 2nd Edition, Prentice
Hall, Upper Saddle River, New Jersey, 1996.
4. CONCLUSION S.H. Kim, T.S. Park, J.Y. Yoo, and G.T. Park. Speed-
sensorless Vector Control of an Induction Motor Using
This paper proposed a detailed discussion of the periodic Neural Network Speed Estimation. IEEE Transactions
component often measured on the wheel angular velocity on Industrial Electronics, 48(3):609–614, 2001.

1699
16th IFAC Symposium on System Identification
Brussels, Belgium. July 11-13, 2012

measured
filtered apative notch
filtered low pass

power spectrum

1X
frequency [Hz]

155

150
velocity [km/h]

145

140
measured
filtered adaptive notch
135
filtered low pass

87.9 88 88.1 88.2 88.3 88.4 88.5 88.6

time [s]

Fig. 10. Speed measurement spectra and time domain, before and after notch filtering.

M. Lemkin, P.H. Yang, A.C. Wuang, J. Jones, and D.M. Y. Su, C. Zheng, D. Sun, and B. Duan. A Simple Non-
Auslander. Velocity Estimation From Widely Spaced linear Velocity Estimator for High Performance Motion
Encoder Pulses. In American Control Conference, pages Control. IEEE Transactions on Industrial Electronics,
998–1002, Seattle,Washington, USA, 1995. 52(4):1161–1169, 2005.
E. Levi and M. Wang. A Speed Estimator for High Y. Su, C. Zheng, P. Mueller, and B. Duan. A Simple
Performance Sensorless Control of Induction Motor in Improved Velocity Estimation for Low-Speed Regions
the Field Weakening Region. IEEE Transactions on Based on Position Measurements Only. IEEE Transac-
Power Electronics, 17:365–378, 2002. tions on Control Systems Technologies, 14(5):937–942,
PD McFadden and JD Smith. Vibration monitoring of 2006.
rolling element bearings by the high-frequency reso- M. Tanelli, C. Vecchio, M. Corno, A. Ferrara, and S.M.
nance technique–a review. Tribology International, 17 Savaresi. Traction control for ride-by-wire sport mo-
(1):3–10, 1984. torcycles: A second-order sliding mode approach. In-
T. Ohmae, T. Matsuda, K. Makiyama, and M. Tachikawa. dustrial Electronics, IEEE Transactions on, 56(9):3347–
A Microprocessor Controlled High Accuracy Wide 3356, 2009.
Range Speed Regulator for Motor Drives. IEEE Trans-
actions on Industrial Electronics, IE-29:207–211, 1982.
F.-Z. Peng and T. Fukao. Robust Speed Identification for
Speed Sensorless Vector Control of Induction Motors.
IEEE Transactions on Industry Applications, 30(5):
1234–1240, 1994.
S. Phillips and M. Branicky. Velocity Estimation Using
Quantized Measurements. In IEEE Conference on
Decision and Control, pages 4847–4852, 2003.
A. Proca, A. Keyhani, and A. Akdag. Sensorless Sliding-
mode Control of Induction Motors Using Operating
Condition Dependent Models. IEEE Transactions on
Energy Conversion, 18(2):205–212, 2003.
S.M. Savaresi, C. Poussot-Vassal, C. Spelta, O. Sename,
and L. Dugard. Semi-active Suspension Control Design
for Vehicles. Elsevier (Butterworth-Heinemann), Ox-
ford, UK, 2010.
K. Shi, T. Chan, Y. Wong, and S. Ho. Speed Estimation
on an Induction Motor Drive Using an Optimized Ex-
tended Kalman Filter. IEEE Transactions on Industrial
Electronics, 49:124–133, 2002.

1700