Beruflich Dokumente
Kultur Dokumente
Applications
U. Forssell, S. Ahlqvist, N. Persson, F. Gustafsson*
NIRA Dynamics AB
Teknikringen 1F
58330 Linköping, Sweden
Phone: +46/13/329800 Fax: +46/13/329829
Email: urban.forssell@nira.se
stefan.ahlqvist@nira.se
niclas.persson@nira.se
*Dept. of Electrical Engineering
Linköping University
58183 Linköping, Sweden.
Phone: +46/13/282706 Fax: +46/13/282622
Email: fredrik@isy.liu.se
Keywords: adaptive filter, inertial sensors, sensor fusion, signal processing
Abstract
This paper discusses sensor fusion as a means to compute virtual sensor signals
for certain vehicle attitude quantities, in particular vehicle yaw rate. It is shown
how sensor fusion can be used to increase the performance and availability of
standard sensors commonly available in a modern car. Test results from tests
performed with a real vehicle are presented.
1 Introduction
NIRA Dynamics is a research and development company specialising in signal
processing for vehicle dynamics applications. This paper presents some ongoing
activities within our company directed at creating virtual sensors using advanced
sensor fusion software.
The sensor fusion idea is quite general and has many different application areas,
see e.g. [1] and the references therein. As the name indicates, sensor fusion is
about fusing information from several different physical sensors. The goal is to
compute new virtual sensor signals using information from the existing, physical
sensors. The virtual sensors can in principle be of two different types:
Sensor fusion is used in, for example, navigation, target tracking, aircraft attitude
estimation and various other military applications to achieve exactly these goals.
Our primary aim in our research and development efforts is to develop unique
sensor fusion based systems for vehicles in particular for vehicle dynamics
applications – and the challenge is to utilise the potential to both improve
performance and reduce the system cost.
N IR A D y n a m ic s S e n s o r F u s io n R o a d fr ic tio n
H ig h p r e c . y a w r a te
T ir e p r e s s u r e
ABS
A n ti- lo c k in g b ra k e s
V ir tu a l C o n tro l A n ti- s p in
s e n s o rs u n it D y n a m ic s ta b ility
G y ro
S ensor A d a p tiv e c r u is e
in te g r a tio n
u n it
A cc. A u d io /v is u a l w a r n in g s
D ia g n o s is
H M I In fo r m a tio n d is p la y s
u n it
E n g in e
D e g r a d e d /" lim p
h o m e " fu n c tio n a lity
Antilock braking and antispin systems need accurate velocity information to
compute the slip.
Antispin systems for AWD vehicles need absolute vehicle velocity
information for computing the optimal slip.
Dynamic stability systems need an accurate and highbandwidth yaw rate
signal to control the body slip angle (i.e. the difference in angle between the
steering wheel and the wheel's velocity vector).
Adaptive cruise control systems need accurate yaw rate information for its
situation awareness.
Smart airbag systems need accurate velocity and acceleration information to
control the release of the airbag during e.g. vehicle roll over.
A problem in many vehicle projects, however, is that the highprecision sensors
that are needed in order to meet the functional specifications typically have to be
replaced by less accurate sensors due to cost considerations. Hence, there is a huge
potential for using sensor fusion technology to create highprecision virtual
sensors at a very modest cost in this area. To exemplify our ideas we will in this
paper concentrate on the problem of estimating a high precision yaw rate signal
from a standard, lowcost gyro using sensor fusion.
The rest of the paper is organised as follows. Next, in Section 2, we review some
basic algebraic relations, which reveal much of the core ideas of the sensor fusion
approach. Then, we go on to discuss the details for how to estimate a high
precision yaw rate signal in Section 3. Section 4 contains some test results and,
finally, Section 5 summarises the paper.
2 Theoretical Foundation
The core idea behind our sensor fusion solution can be highlighted using the
following example. Consider two different sensors measuring the same varying
physical parameter gives separate measurements yi(t) of a the parameter x, where
each measurement has an offset bi with an offset scaling ci(t) according to a known
function of time. The measurements can be expressed algebraically as the
equations:
y1(t) = x(t) + c1(t)b1
y2(1) = x(t) + c2(t)b2
These two equations have three unknowns and is therefore insoluble, and the
offsets cannot be directly eliminated. When two measurements y1(1),y2(1) and
y1(2),y2(2) are available, there are two more equations and only one more
unknown, i.e. four equations and four unknowns. Thus, the offsets and the
variable parameter values x(1),x(2) can be solved under the condition that there is
no linear dependency in data. In this example, the linear independency condition
is:
c1(1)/c1(2) c2(1)/c2(2)
If, for example, c1 is constant and c2(t) is the velocity vx(t), linear independency
occurs when the velocity has changed between two measurements. This leads to
observability, and under these conditions we can resolve all unknowns and hence
also determine the sought parameter x without error.
In practice there is a measurement noise added to each of the observations. In
order to eliminate the noise, a number of observation samples large enough to
constitute an overdetermined equation system is collected and solved using a least
squares solution. In a realtime application, this should be implemented using a
recursive filter, preferably a Kalman filter, into which the sampled observations
are input. Under certain identifiability assumptions (like persistence of excitation),
the Kalman filter gives consistent estimates of the sought quantity (quantities). As
always in adaptive filtering, there is a tradeoff between noise suppression and
tracking ability, which must be handled with care for optimal performance of the
adaptive filter [1]. Here we will not go into details on this for sake of conciseness.
Next we will see how these basics translate into a more explicit form when
discussing high precision yaw rate signal estimation.
Our ideas for how to compute a high precision yaw rate sensor signal will be
described in this section. As indicated in figure 2, the idea is to utilise available
information from existing sensors in a modern car and to use sensor fusion to
compute an improved yaw rate signal.
The idea of our patent pending [2,3] system is to compute a bias and scale factor
free yaw rate signal using an adaptive filter that estimates the error in the yaw rate
signal from the yaw rate gyro and removes that from the original signal by means
of a simple subtraction operation, cf. figure 2. A feature of this solution is that the
bandwidth of the original signal is preserved and, in case of limited computational
power, the adaptive filter can run at a very moderate rate (it should be high
enough to capture the most important temperature drifts etc. of the sensor,
though).
+ H ig h P r e c is io n Y a w r a te
Y a w ra te
g y ro
_
S ensor
W heel
in te g r a tio n
speed
u n it
E r r o r e s tim a te
L a te ra l
a c c .-
m e te r
The details of our solution are as follows. For the sake of simplicity of the
explanation, we are here assuming that there is no lateral movement of the vehicle.
In the relations:
vx
v x R 1
R
v2
a y x v x2 R 1 v x
R
is the yaw rate from a gyro; vx is the velocity of the vehicle in the xdirection;
ay is the acceleration in the ydirection. The curve radius R is computed according
to the following relation, where R is defined as the distance to the centre of the
rear wheel axle (of length L):
L
vrr Rrr R 2
vrl Rrl R L
2
The angular wheel velocities for each of the respective wheels are received from
an ABS and the inverse R1 of R is solved for in order to avoid numerical problems
when driving straight ahead. The wheel radii ratio is subject to an offset:
rrl
1 ABS
rrr
The influence of the offset on the denominator is negligible, which results in the
following expression for inverse curve radius:
1 2 rl 1 2 rl
R 1 (1 ABS ) 1 Rm1 ABS
L rl 1 rr L rl 1 rr
rr rr
wherein the computable quantity
1 2 rl
Rm1 1
L rl 1 rr
rr
is used for the inverse curve radius. Finally, the velocity at the centre of the rear
wheel axle is
rr
vx rl r
2
where r denotes the nominal wheel radius.
Thus, in a practical implementation of the system depicted in figure 2, the sensor
measurements are:
1. y1(t) from a yaw rate sensor, i.e. gyro signal;
1 1
2. y2(t) = v x Rm , from ABS sensors, R m is computed as above; and
possibly
3. y3(t) from a lateral acceleration sensor.
All these sensor measurements are subject to an offset and measurement noise
given by the relations:
y1 (t ) (t ) YR e1 (t )
1 2 rl
y 2 (t ) v x Rm1 e 2 (t ) (t ) v x ABS e 2 (t )
L rl rr
1
rr
y 3 (t ) v x (t ) ACC e3 (t )
where ABS is an offset that depends on relative tire radius between left and right
wheels. In this model, the offsets may be estimated online using e.g. a recursive
least squares method or a Kalman filter. We prefer the latter due to the Kalman
filter’s advantageous tuning flexibility in multiparameter estimation problems.
Remarks
1. The important question of identifiability, that is, under what conditions are
the offsets possible to estimate, is answered by studying the rank of the matrix
to be inverted in the least squares solution. For the accelerometer sensor, the
matrix is given by:
1 N 1
1
N t 1 v x t
1 N 1 1 N 1
N t 1 v x t N t 1 v x t 2
and, in short, this matrix has full rank if and only if the velocity changes
during the time horizon. Furthermore, the more variation, the better estimate.
Similarly, the offsets are identifiable from yaw rate and ABS sensors if the
velocity or the curve radius changes anytime.
2. Another problem to consider in a practical implementation of the system is
that of wheel spin and other abnormal driving conditions, which must be
considered separately. Due to space limitations, we do not go into the details
of this here. In general, these kinds of problems must be handled by turning
off or slowing down the adaptation for some (or all) of the parameters during
these conditions.
4 Test Results
To demonstrate the performance of the high precision yaw rate function we now
show the results of two different tests:
1. A 105 second drive on public road, with rapid lane changing forth and back
on first part, then an aggressively taken roundabout and then a drive through
the same bend, with hard ABSbraking to complete the stop. The results are
plotted in figures 36.
2. A test drive comprising of four laps in a large roundabout where the
improvement in longterm drift is clearly visible, cf. figure 7.
38
Yaw Rate [deg/sec]
36
34
32
30
46 46.2 46.4 46.6 46.8 47 47.2 47.4 47.6 47.8 48
Time [s]
30
20
Yaw Rate [deg/sec]
10
-10
-20
-30
-40
0 20 40 60 80
Time [s]
4.75
4.65
4.6
4.55
4.5
4.45
4.4
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90
Time [s]
References
[1] F. Gustafsson, Adaptive Filtering and Change Detection, Wiley & Sons, 2000
[2] F. Gustafsson et al., Sensor Fusion for Accurate Computation of Yaw Rate and
Absolute Velocity, SAE paper 2001011064
[3] Swedish patent application SE00013532, Sensor Fusion System