Proceedings of the American Control Conference
San Diego, California June 1999
●
A Real-timeRolloverThreatIndexforSportsUtilityVehicles
Bo-Chiuan Chenl
Huei Peng2
Department of Mechanical Engineering and Applied Mechanics
University of Michigan
Ann Arbor, MI 48109-2125
TEL: (734) 936-0352
E-mail: hpeng@umich.edu
Abstract
In this paper, the methodology
capable
of
computing a Time-To-Rollover
(TT’R) index in real-time
[1] is verified by using test data of two Sports-Utility
Vehicles (S W)-a
1988 Suzuki Samurai and a 1997 Jeep
Cherokee.
First, simple yaw-roll models are constructed
based on the test data. The TTR is computed from the
simple model and then corrected by using an Artificial
Neural Network. The TT’R generated
by the Neural
Network is then verified against the data for the two test
vehicles.
1. Introduction
The safety performance of passenger cars has been
an important factor influencing consumers’ purchasing
decisions and government regulations. Currently, a New
Car Assessment Program (NCAP) exists under the US
Department of Transportation which assess the crash
worthiness of new cars. The results are published as a wellknown one-to-five star rating system indicating the crash
protection for passengers within the same weight class. The
final results are easy to understand and have been a
valuable index for the general public. Recently, the
National Highway Traftlc Safety Administration (NHTSA)
announced its plan to provide information about the
rollover stability of vehicles in its future safety rating. One
major driving force behind this new initiative is the wellpublicized rollover incidents of several Sports-Utility
Vehicles (SW) and passenger cars (Suzuki Samurai, Isuzu
Trooper, Mercedes
A-class and the Mercedes/Swatch
SmartCar).
It seems fair to say that rollover stability is
becoming an important element in the vehicle safety
performance.
Along another line, active counter-measures
to
prevent rollover crashes are being pursued by research
supported by NHTSA. To prevent rollover, one of the most
important enabling
techniques
is the development
of
accurate rollover threat indices. A rollover warning/control
algorithm will work well only if the impending vehicle
rollover threat can be accurately represented.
Most of the
existing rollover warning algorithms [2,3,4,5,6,7] are based
on acceleration or roll angle threshold values. Rakheja and
Pich~ proposed an early warning safety monitor in 1990 [2].
1 Graduate Student
2 Assistant Professor, corresponding
0-7803-4990-6/99
$ 10.00@
author
1999 AACC
1233
Under static cornering maneuvers, a rollover acceleration
threshold is defined when the inside tire deflection reaches
zero, i.e. the tire normal force becomes zero and the tire is
litling off the ground. In implementation,
a warning is
issued whenever the measured lateral acceleration exceeds
this threshold. They also applied the same idea to define a
threshold for the semi-trailer roll angle. Preston-Thomas
and Woodrooffe [3] used the lateral load transfer ratio
(LTR) to indicate rollover threat. LTR can vary from O
when the loads carried by the tires are equal, to 1 when the
tires lift off and the lateral acceleration reaches the rollover
acceleration threshold. Freedman et al. [4] proposed a
rollover advisory sign on highway exit ramps. When the
velocity of the truck is higher than the advised speed limit,
which is determined based on the curvature of the exit
ramp, a warning light is triggered. In their research, the
warning signal did not use additional information
to
identify each truck’s rollover acceleration threshold. McGee
et al. [5,6] proposed a warning system which is similar to
that of [4]. However, their system detects the type, speed,
weight, and height of the truck to identify the rollover
acceleration threshold of each vehicle tkom a look-up table.
Winkler et al. [7] recently proposed a rollover stability
advisor (RSA) system. RSA determines
the rollover
acceleration threshold based on real-time measurements of
the status of the vehicle.
The measurement
they used
includes force and moment at the fifth wheel and many roll
motion variables.
Three acceleration threshold values are
calculated dynamically
to determine rollover threshold
level.
Nalecz et al. proposed an energy based function
named Rollover Prevention Energy Reserve (PRER) in
1987 [8,9, 10]. PRER is defined as the difference between
the energy needed to bring the vehicle to its tip-over
position and the rotational kinetic energy, which can be
transferred into the gravitational potential energy to lift the
vehicle. PRER remains positive for non-rollover
cases.
When it becomes negative, a rollover will occur if nothing
is done to take energy out of the roll mode.
A special
advantage of PRER is that the same concept can be applied
to both maneuver induced and tripped rollover incidents.
The above-mentioned
concepts were based on
acceleration, roll angle or energy threshold values which
are estimated from the information at a fixed time. In
analogy, it is like taking a still picture of a dynamic system
and uses the information (frozen in time) to determine the
rollover threat. Apparently, a method that covers factors
over a longer time horizon, particularly into the future,
could give us a better perspective. Furthermore, the
“distance” away from these threshold levels is not an
intuitive measure. Therefore, we proposes a “Time-T&
Rollover” (T’TR) metric [1], which was proposed as the
basis to assess rollover threat since it “counts-down” toward
rollover and is an intuitive threat index. To realize this
simple threat index, however, we need to construct a lTR
calculation unit that cart accurately predict TI’R under all
vehicle speed, load and steering pattern. Implementing ‘Ill?
in real-time seems to involve design trade-off. On the one
hand a faster-than-real-time
vehicle model is needed. For
example, in order to predict a Ill? of (up to) 3 seconds, one
needs to predict vehicle response in the next 3 seconds
repeatedly. If TTR is to be updated every 50msec, the
vehicle model needs to be 60 times faster than real-time.
On the other hand, the lTR predicted by this model needs
to be accurate enough under all driving scenarios so that
good warning decision can be made. An innovative
approach was proposed in [1] to solve this dilemma.
Although the focus in [1] was on heavy trucks, it
is obvious that the design procedure proposed in [1] can
also be applied to SUV as well. We will maintain the basic
architecture of the algorithm shown in [1] and will apply it
to two SUVS. The major challenges of this seemingly
straight-forward extension lie in the fact that in the previous
work, the design and verification were based on simulation
data, while in this work, real vehicle test data will be used.
llte remainder
of this paper is organized
as
follows: the simplified yaw-roll models identified from the
vehicle test data are presented in Section 2. In Section 3, the
T’f’R index is defined and presented. The Neural Network
trained to produce
“desired”
TTR response
is also
dkcussed. The results of real-time rollover treat index are
then presented in Section 4. Finally, conclusions are made
in Section 5.
2.
Modeling
A vehicle yaw-roll model needs to be constructed
to compute vehicle roll motion under steering excitations.
In this paper, a yaw-roll model will be obtained by fitting
test data obtained on two SUVS: a 1988 Suzuki Samurai
and a 1997 Jeep Cherokee. In the remainder of this paper,
these two vehicles will be referred to as “Samurai” and
“Cherokee”,
respective y.
The vehicle test data was
obtained from the Vehicle Research and Test Center
(VRTC) of NHTSA. The test data will be used first for the
vehicle model construction (see Figure 1) and later for
rollover algorithm verification.
,
Yaw
Roll
Model
Model
Lateral
Roll angle
angle
acceleration
Roll rate
Figure
1 Structure
of
the
simplified
model
1234
As can be seen from Figure 1, the proposed yawroll model is separated into a yaw and a roll part. This
series arrangement may propose less-than accurate results
compared with an integrated yaw-roll model.
However,
this simplified structure was found to be superior in two
aspects:
ease
of model
construction,
and
faster
computations.
2.1 Yaw Model
The vehicle yaw model (see Figure 2) was
assumed to be described by a linear bicycle model and the
vehicle speed is assumed to be constant. It was found that
when the vehicle speed varies significantly due to excessive
steering/brrtking, a constant-speed
linear model no longer
matches the test data closely.
c \
J
Longitudinal
Yaw rate,r
‘
. . .
=Ieral
=-
----Y“
-.
tire steering
angle, ~tire
,
I
velocity, v ~
acceleration,ay
Figure
2 Yaw model
(Bicycle
model)
For a linear bicycle model, the discrete-time
transfer function
from the steering angle to lateral
acceleration is known to have the following form:
~
#oz2+lqz+lq
yaw
Z2 +alz+q
=fi
(1)
6
where ay is the lateral acceleration
and 6 is the steering
wheel angle which is related to the tire steering angle by a
steering gear ratio. After factorization,
we can get the
following form.
Z–zl
z’yaw(z)=~o+~(z-p*)(z-Fl)
(2)
We applied standard system identification
techniques to
crdculate k, Z1, PI and p2 and then take the average of
their values for multiple files of the same maneuver. ‘l?tese
transfer fimctions have been found to work well under
constant speed cases.
When the vehicle speed varies
significantly due to large steeringhraking,
we found it
necessary
to use a speed-dependent
(gain-scheduled)
transfer function to predict vehicle lateral acceleration
accurately.
A simple interpolation technique is used. As
can be seen from Table 1, the lateral acceleration prediction
error can be reduced significantly
for steering+brakirtg
maneuvers.
Table
1
prediction
RMS values
errors
of
lateral
acceleration
Samurai
Cherokee
Const.
Braking
Braking
Const.
w/o gain-scheduling
0.3972
2.2749
0.2194
2.8197
w/ gain-scheduling
0.3660
0.6004
0.1868
0.5667
Note: Const. denotes for constant speed maneuvers.
Ay (m/Sz)
2.2 Roll Model
The roll model was found to be well behaved and a
2 degree-of-tkedom
(sprung mass roll and unsprung mass
roll, refer to Figure 3) model fits the test data well
independent
of longitudinal
vehicle speeds and lateral
acceleration
levels. The structure of the discrete-time
transfer fiction
from lateral acceleration to sprung mass
roll angle is shown as follows.
boz3 +b1z2 +b2z+b3
Trou (z)=
=1
‘Y
~
z
+alz3+a2z2+a3z+a4
(3)
Where # is the roll angle of the sprung mass.
4
Figure
0.4 g left
Roll
turn
G-1
8-IL-5
0ml
.-
f.lo~
0
5
Figure
Alternatively,
3 Roll
—-.I+..---W
------
2
4
6
time (sac)
Figure
5 Roll
response
of
O. 6 g left
turn maneuver)
model
3.
(.2-21)(2-22)(2-%!)
(4)
(Z- P1)(Z-F1)(Z- P2)(Z-F2)
k= (l-p~)(l-pl)(l-
pz)(l-pz)
O-Z:) O- Z2)0-Z’2)
52 are zeros , and
pl,
Z* , 22 ~d
-
j?j, p2 and ~2 are poles of
Troll(z). Again, the averaged values of k,
Z1, Z2, 22,
PI> R * P2 ~d % me ~ken aCrOSSmultiPle files tO
obtain the final roll model.
As can be seen from Table 2, the roll angle
prediction error for the Samurai is much larger than the
Cherokee. A representative roll response of these two
vehicles are shown in Figures 4 and 5, respectively. Under a
step steering input, the roll angle of Samurai didn’t reach
steady-state value quickly. The roll angle keeps drifting for
several seconds after the step input. This drifting may arise
from either a its suspension damping or corrupted sensor.
The roll angle shown in Figure 5 is derived from two
optical sensor measurements installed on the right and letl
sides of the vehicle front bumper. It reaches a steady-state
response for a step steering input quickly even under a
high-g maneuver. Due to its poor roll angle measurement,
the roll model identified for Samurai does not perform very
well.
Table
errors
of
Samurai
(50
--’1
1
8
10
Cherokee
(50
mph
Eq. (3) can be shown in the following form:
Ti-o[[(z)= k
where
response
maneuver)
2 RMS values
Roll sngle (deg)
of
roll
Samurai
0.5926
angle
prediction
Cherokee
0.1391
1235
Time-To-Rollover
(’ITR) Metrics
When the vehicle roll angle exceeds a certain
threshold level, wheel lift-off will occur.
Since most
existing passenger vehicle models were developed based on
4-wheel assumptions, the vehicle response under wheel liftoff conditions can longer be predicted accurately.
In this
study, we define wheel lift-off as an unacceptable rollover
incident. In other words, while we use the term “rollover
threat index”, the index was in fact issued based on “TimeTo-Wheel-lift-off’.
This decision will not result in any
major change in the overall algorithm development.
A
more aggressive roll angle threshold level can be used and
the overall design process to be described below will stay
the same. The roll angle threshold for these two vehicles
are selected to be 3 degrees (for Samurai) and 3.5 degrees
(for Cherokee), respectively.
3.1 Model based TI’R
From the “rollover incidents” (more precisely,
wheel-lift-off incidents) detected in the test data files, a true
TIR will be computed off-line. In other words, whenever
the roll angle exceeds the defined threshold value, we can
roll back the clock and define a point 0.2 seconds before
this wheel-lift-off incident to have a “T’f’R” of 0.2 seconds.
Ideally, if we can calculate this TIR index in real-time, the
rollover threat can be accurately represented.
A model based TTR is defined as following:
assuming the input (steering angle) stays fixed at its current
level in the foreseeable future, the time it takes for the
vehicle sprung mass to reach its critical roll angle is defined
as TTR. Under normal driving conditions, ‘lTR is usually
quite large. For implementation considerations, we saturate
lTR at 1 second. In other words, we will integrate the
speed-dependent
yaw-roll model (shown in Section 2) for
up to 1 second (see Figure 6). If it is found that the vehicle
does not rollover, the model-based TI’R will be defined as 1
second.
Steering
Future
angle
Vehicle mll ~gle
,.,
;, $~~ ~xe
“;,:q~$
‘tkkfis!!okl
Model
.—
Yes
T1-R
and brake, right/left turn (no braking), double lane change,
pulse steering rightieft, and ramp steering right/left. ‘Ihese
maneuvers are repeated at (nominally) constant speed of 25
and 50 mph. Usually each test condition is repeated for 10
runs.
Some of the test maneuvers were done at lateral
acceleration as high g as 0.6 g. There are 29 and 31
measurement
channels for the Samurai and Cherokee,
respective y. And more than 200 test files for each vehicle
were used. As can be seen from Figure 8, some signals are
noisy and contain large offsets. Therefore, these signals
were processed before they are used in the verifications.
%
2
0
Q19
54 62
41
Figure
6
Flow
chart
for
the
TTR calculation
so
2
3.2 Neural Network Tf’R
From a previous study [1], it was found that the
model-based lTR might not be accurate enough. Ideally,
the predicted lTR should have a straight line of slope = -1
(on a time-lTR plot) under all driving conditions. In other
words, we hope the model-based ‘ITR gives an accurate
countdown
toward rollover incidents. In [1], it was
proposed to use a Neural Network (NN) to correct the error
between the model-predicted TTR and the true TT’R. The
structure of the NN is shown in Figure 7. In additional to
the model predicted ~
the NN also uses vehicle roll
angle and change of roll angle to produce a corrected NNTTR. The desired NN-’ITR is the straight line (of slope = 1) as described above. In the non-rollover cases, the desired
TI’R will be a straight line of lsecond.
. .2
~
-0.4
Y
I
‘,
\
-+
——
-----
\\
i!
\{
,!t
&o
m]
Steering input
Roll angle
Lateral
1234567
a
to
time
Figure
8
calculations
Required
measurements
for
TTR
4.2 ‘ITR verification
The dettil system implemented
to compute a
model based ‘ITR is shown in Figure 9. The measurements
required to implement this system include steering wheel
angle, lateral acceleration and roll angle of the sprung mass.
Initial Conditions at the K%unptiog Time
LateralAccel.
...........
Steerin ~
angle :
k+
~:
—:
Yaw
Model
Latemt
Acceleration
Roll
Motion
ted for up to 1 Second ..........
implementation
for
the TTR
I
I
Layer 1
Figure
Network
7
Input
and
output
of
the
Neural
The NN used is quite simple. It has two hidden
layers with 5 and 1 neurons, respectively. The functions of
those neurons are standard tansig functions.
The NN is
trained for each vehicle by using test data files both with
and without rollover incidents.
In other words, the same
NN will be used to generated a corrwted T1’R across all
vehicle speeds and all steering and braking maneuvers.
4. System Verifications
4.1 Maneuver description for the test data
The test data from VRTC contains the following
maneuvers suitable for TTR verifications: right/left steer
1236
Figure
9 Detail
calculation
We have verified the TTR for both rollover and
non-rollover cases. Since the TTR for the non-rollover
cases, we only show the rollover cases here (see Figures 10
and 11). The l’TR for Samurai sudden] y reduces to near
zero. For the Cherokee, we can get about 0.3 seconds of
“warning” before a rollover actually starts. From these two
figures, it seems a lot more challenging to design a rollover
warning (or control) algorithm for the Samurai since the
roll motion occurs suddenly. We would like to emphasize,
however, that the test data of the Samurai is not in as good a
condition as those of Cherokee.
Therefore, the problem
exhibited in Figure 10 may not be completely due to the
vehicle design.
Since we are more confident with the
model identified from the test data of the Cherokee, we will
focus the future TTR discussion on the Cherokee data.
vs. 3 seconds). It seems to suggest theneed
prevention controls on SW.
5.
“o
0 .5
Figure
10 Model
0.6g right
turn
1
1 .5
t im e (Se e )
based
TTR of
maneuver)
Samurai
data were fwst processed to remove noise and DC offset.
The steering, lateral acceleration and roll angle data were
then used to construct simple yaw-roll models. By utilizing
standard gain-scheduling technique, a speed-dependent yaw
model was obtained which provides a more accurate lateral
It was found that the TIT?
acceleration prediction.
computed from the simple yaw-roll model is accurate
enough due to the fact the vehicle roll angles for SW
under wheel lift-off are generally very small.
mph
(50
‘~
“o
0 .5
25
2
1
ti:(sec)
Figure
11 Model
based
TTR of
mph 0.6g right
turn maneuver)
3
Cherokee
Acknowledgement
(50
After training the INN(Figure 7), we found that no
significant improvement was obtained over the modelbased TTR (refer to Figure 12). This isin contrast to the
results forheavy trucks reported in [1]. In [l], the NN was
found to be necessary to obtain an accurate ‘ITR, as can be
seen from Figure 13.
“o
0.5
2.5
2
1
3
tim155(sec)
Figure
12 Model
based
TTR of
mph 0.6g right
turn maneuver)
Cherokee
(50
3
‘x
%
y
E
+3
“,,
,
\
1
0
n
0
Figure
[1].
10
5
Time (see)
13 TTR of an M916A1/M870A2 army truck
(“+”: Model based TTR, “x”: NN TTR).
12 that the NN
Due to their
lower e.g. height, the critical roll angle to denote
unacceptable roll incidents (e.g., wheel lift-ofo for SW is
generally much lower than that for heavy trucks.
The
lowered level of roll angle produces much lower ‘fTR, i.e.,
the advance warning one can get to prepare for proper
action is much shorter (see Figures 11 and 13, 0.3 seconds
It can be seen
from
Conclusion
A Time-To-Rollover
(’ITR) based rollover threat
index is developed and verified by using the test data of two
SWs—a
Suzuki Samurai and a Jeep Cherokee. The test
25
2
Figure
correction may not be necessary for SWS.
This research is supported by the U.S. Army
TARDEC under the contract DAAE07-98-C-R-LO08. The
authors also wish to thank Dr. Riley Garrett and Mr. Paul
Greiger of VRTC for supplying the test data.
References
[1] Chen, B. and Peng, H., “Rollover Warning For Articulated
Vehicles Based on A Time-To-Rollover Metric; To appear
in the Proceedings of the 1999 International Mechanical
Engineering Congress and Exposition, Nov. 1999.
[2] Rakhej~ S. and Pich&, A., “Development of Directional
Stability Criteria for an Early Warning Safety Device,” SAE
Paper No. 902265, 1990.
[3] Preston-Thomas, J. and Woodrooffe, J. H. F., “A Feasibility
Study of a Rollover Warning Device for Heavy Trucks,”
Transport Canada Publication No. TP 1061OE, September
1990.
[4] Freedman, M., Olson, P. L. and Z.ador, P. L., “Speed
Actuated Rollover Advisory Signs for Trucks on Highway
Exit Ramps,” Insurance Jnstitute for Highway Safety and
University of Michigan Transportation Research Institute,
December 1992.
[5] McGee, H. et al., “Feasibility of An Automatic Truck
Warning System,” FHWA-RD-93-039, June1993.
[6] Strickland, R. and McGee, H., “Evaluation of Prototype
Automatic Truck Rollover Warning Systems,” FHWA-RD97-124, 1997.
[7] Winkler, C. et al. “Cooperative Agreement to Foster the
Deployment of a Heavy Vehicle htelligent Dynamic
Stability Enhancement System,” University of Michigan
Transportation Research Institute, Interim repor~ NHTSA U.S. DOT Contract No. DTNH22-95-H-07002, January
1998.
[8]. Nalecz, A. G. and Bindemarm, A. C., “Sensitivity Analysis of
Vehicle Design Attributes that Affect Vehicle Response in
Cnticaf Accident Situations – Part I: User’s Manual,” DOTHS-807-229, December 1987.
[9]. Nalecz, A. G., “Intermediate Maneuver Induced Rollover
Simulation (JIvIIRS) a Sensitivity Analysis,” DOT-HS-807672, February 1991.
[10].Nalecz, A. G. et al., “An investigation into Dynamic Measure
of Vehicle Rollover Propensity, ” SAE Paper No. 930831,
1993.
1237
.
for active roll-