A methodology for the design of robust rollover prevention
controllers for automotive vehicles using differential braking
March 15, 2009
Abstract
In this paper we apply recent results from robust control to the problem of rollover prevention in automotive vehicles. Specifically, we exploit the results of [6] which provide controllers to robustly guarantee that the peak values
of the performance outputs of an uncertain system do not exceed certain values. As a measure of performance for
rollover prevention, we use the Load Transfer Ratio LT Rd introduced in [21], and design differential-braking-based
rollover controllers to keep the value of this quantity below a certain level; we also obtain controllers which yield
robustness to variations in vehicle speed. We present numerical simulations using a nonlinear multi-body vehicle
simulation model to demonstrate the effectiveness of our controllers in preventing rollover.
Keywords: vehicle rollover, vehicle active safety, vehicle active rollover prevention, robust control.
1 Introduction
It is well known that vehicles with a high center of gravity such as vans, pickups, and the highly popular SUVs (Sport
Utility Vehicles) are more prone to rollover accidents. According to the 2004 data [1], light trucks (pickups, vans and
SUV’s) were involved in nearly 70% of all the rollover accidents in the USA, with SUV’s alone responsible for almost
35% of this total. The fact that the composition of the current automotive fleet in the U.S. consists of nearly 36%
pickups, vans and SUV’s [2], along with the recent increase in the popularity of SUV’s worldwide, makes rollover an
important safety problem.
There are two distinct types of vehicle rollover: tripped and un-tripped. A tripped rollover commonly occurs
when a vehicle slides sideways and digs its tires into soft soil or strikes an object such as a curb or guardrail. Driver
induced un-tripped rollover can occur during typical driving situations and poses a real threat for top-heavy vehicles.
Examples are excessive speed during cornering, obstacle avoidance and severe lane change maneuvers, where rollover
occurs as a direct result of the lateral wheel forces induced during these maneuvers. In recent years, rollover has
been the subject of intensive research, especially by the major automobile manufacturers; see, for example, [3, 4].
That research is geared towards the development of rollover prediction schemes and occupant protection devices. It
is however, possible to prevent such a rollover incident by monitoring the car dynamics and applying proper control
effort ahead of time. Therefore there is a need to develop driver assistance technologies which would be transparent
to the driver during normal driving conditions, while acting in emergency situations to recover handling of the vehicle
until the driver recovers control of the vehicle [5].
We present in this paper a robust rollover prevention controller design methodology based on differential braking.
The proposed control design is an application of recent results on the design of control systems which guarantee
that the peak values of the performance outputs of a plant do not exceed certain thresholds when subject to bounded
disturbance inputs [6, 7]. The main selected performance output for the rollover problem is the Load Transfer Ratio
LT Rd . This measure of performance is related to tire lift-off and it can be considered as an early indicator of impending
vehicle rollover. We also include the braking force as a performance output to take into account limitations on the
maximum braking force. The aim of our control strategy is to maximize the magnitude of the allowable disturbance
inputs which do not drive the performance outputs outside their prespecified limits; in this case the disturbance input
is the driver steering input. We also want to guarantee robustness with respect to the parameter uncertainty that arises
from changing vehicle speed. We indicate how our design can be extended to account for other sources of uncertainty
such as unknown vehicle center of gravity and tire stiffness parameters.
2 Related work
Rollover prevention is a topical area of research in the automotive industry (see, for example, the rollover section at
http://www.safercar.gov/ for a good introduction to the problem) and several studies have recently been published.
Relevant publications include that of Palkovics et al. [8], where they proposed the ROP (Roll-Over Prevention) system
for use in commercial trucks making use of the wheel slip difference on the two sides of the axles to estimate the
tire lift-off prior to rollover. Wielenga [9] suggested the ARB (Anti Roll Braking) system utilizing braking of the
individual front wheel outside the turn or the full front axle instead of the full braking action. The suggested control
system is based on lateral acceleration thresholds and/or tire lift-off sensors in the form of simple contact switches.
Chen et al. [10] suggested using an estimated TTR (Time To Rollover) metric as an early indicator for the rollover
threat. When TTR is less than a certain preset threshold value for the particular vehicle under interest, they utilized
differential braking to prevent rollover. Ackermann et al. and Odenthal et al. [11], [12] proposed a robust active
steering controller, as well as a combination of active steering and emergency braking controllers. They utilized a
continuous-time active steering controller based on roll rate measurement. They also suggested the use of a static
Load Transfer Ratio (LT R s ) which is based on lateral acceleration measurement; this was utilized as a criterion to
activate the emergency steering and braking controllers.
3 Vehicle modeling and LT Rd
In this section we introduce the model that we use for controller design. We define the rollover detection criterion
LT Rd and present the assumptions on the sensors and actuators used in the design. We also present the higher fidelity
nonlinear multi-body simulation model to which the controllers will be applied.
3.1 Vehicle model for control design
We use a linearized vehicle model for control design. Specifically, we consider the well known single-track bicycle
model with a roll degree of freedom as illustrated in Figure 1. In this model the steering angle δ , the roll angle φ and
the vehicle sideslip angle β are all assumed to be small. We further assume that all the vehicle mass is sprung, which
implies insignificant unsprung mass.
x
Įf
į
Fy,f
a
y
v
z
ȕ
mg
ĭ
ƺ
6
h
Įr
b
c
y
k
Fy,r
Fz
Figure 1: Linear bicycle model with roll degree of freedom.
The lateral forces on the front and rear tires, denoted by F y, f and Fy,r , respectively, are represented as linear
2
functions of the tire slip angles α f and αr , that is, Fy, f = Cα , f α f and Fy,r = Cα ,r αr , where Cα , f and Cα ,r are the
front and rear tire stiffness parameters, respectively. In order to simplify the model description, we further define the
following auxiliary variables
σ
Cα , f + Cα ,r
ρ
Cα ,r b − Cα , f a
2
Cα , f a + Cα ,r b
κ
(1)
2
where the lengths a and b are defined in Figure 1. For simplicity, it is assumed that, relative to the unsprung mass, the
sprung mass rolls about a horizontal roll axis which is along the centerline of the unsprung mass and at ground level.
Using the parallel axis theorem of mechanics, J xeq , the moment of inertia of the vehicle about the assumed roll axis, is
given by
(2)
Jxeq = Jxx + mh2 ,
where h is the distance between the vehicle center of gravity (CG) and the assumed roll axis and J xx is the moment
T
of inertia of the vehicle about the roll axis through the CG. Introducing the state x = β ψ̇ φ̇ φ , where ψ̇ is the
yaw rate of the unsprung mass, the motion of this model can be described by
ẋ = Ax + Bδs δs + Bu u
(3)
where
⎡
σ Jx
− mJxxeqv
⎢
ρ
⎢
A = ⎢ Jzz
⎣ − hσ
Jxx
0
π
Bδs =
180λs
Bu
=
ρ Jxeq
−1
mJxx v2
− Jzzκ v
hρ
Jxx v
0
Cα , f Jxeq
mJxx v
0 − 2JTzz
Cα , f a
Jzz
0 0
− Jhc
xx v
0
− Jcxx
1
hCα , f
Jxx
h(mgh−k)
Jxx v
0
⎥
⎥
⎥,
⎦
mgh−k
Jxx
0
T
0
⎤
,
(4)
T
where the steering wheel angle δ s is the steering input applied to the steering wheel (in degrees) and λ s is the steering
ratio between the steering wheel input and the steering angle δ of the front wheels.
Control input u represents the differential braking force on the wheels; it is positive if braking is on the right wheels
and negative if braking is on the left wheels. Note that we can brake either front, rear or both of the wheels on each
side of the vehicle depending on the maneuver and u is the total effective braking force acting on either side.
Further definitions for all the parameters in (4) are given in Table 1. See [13] for a detailed derivation of this
vehicle model.
3.2 The Load Transfer Ratio, LT Rd
Traditionally, as discussed in the related work section, some estimate of the vehicle load transfer ratio (LTR) has been
used as a basis for the design of rollover prevention systems. The quantity LTR [12, 14] can be simply defined as the
load (i.e., vertical force) difference between the left and right wheels of the vehicle, normalized by the total load; these
forces are illustrated in Figure 2. In other words
LT R =
Load on Right Tires-Load on Left Tires
.
Total Load
(5)
Clearly, LT R varies within [−1, 1], and for a perfectly symmetric car that is driving straight, it is zero. The extrema
are reached in the case of a wheel lift-off of one side of the vehicle, in which case LT R becomes 1 or −1 depending
on the side that lifts off. If roll dynamics are ignored, it is easily shown [12] that the corresponding LTR (which we
denote by LT R s ) is approximated by
LT Rs
2ay h
,
gT
3
(6)
Table 1: Model parameters and their definitions
Parameter
m
v
Jxx
Jzz
a
b
T
h
c
k
Cα , f
Cα ,r
δ
δs
λs
Description
vehicle mass
vehicle speed
roll moment of inertia at CG
yaw moment of inertia at CG
longitudinal CG position w.r.t. front axle
longitudinal CG position w.r.t. rear axle
vehicle track width
distance of CG from roll axis
suspension damping coefficient
suspension spring stiffness
linear tire stiffness for front tire
linear tire stiffness for rear tire
steering angle
steering wheel angle
steering ratio
Unit
[kg]
[m/s]
[kg · m 2 ]
[kg · m 2 ]
[m]
[m]
[m]
[m]
[N · m · s/rad]
[N · m/rad]
[N/rad]
[N/rad]
[deg]
[deg]
z
mg
ĭ
d
c
k
Fz,L
Fz,R
T
Figure 2: Combined vertical forces for each side of vehicle
where ay is the lateral acceleration of the CG.
The rollover estimation based upon (6) is not sufficient to detect the transient phase of rollover (due to the fact that
it is derived ignoring roll dynamics.) Consequently, we follow [21] and obtain an expression for LTR which does not
ignore roll dynamics. We denote this by LT R d . In order to derive LT R d we write a torque balance equation. Recall
that we assumed the unsprung mass is insignificant and the main body of the vehicle rolls about an axis along the
centerline of the unsprung mass at the ground level. We can write a torque balance for the unsprung mass about the
assumed roll axis in terms of the suspension torques and the vertical wheel forces as follows:
−Fz,R
T
T
+ Fz,L − kφ − cφ̇ = 0 .
2
2
4
(7)
Now substituting the definition of LT R from (5) and approximating the total load by the vehicle weight, yields the
following expression for LT R d :
LT Rd
= −
2(cφ̇ + kφ )
.
mgT
(8)
In terms of the state x, LT R d can be described by
LT Rd = C1 x ,
(9)
where
C1 =
2c
0 0 − mgT
2k
− mgT
.
(10)
3.3 Actuators, sensors and parameters
We are interested in control design based on differential braking. Active braking actuators are already available in
many modern production cars that are equipped with systems such as ABS (Anti-lock Braking System) and EBS
(Electronic Brake System) or similar systems, which are capable of selectively braking each of the wheels. The fact
that control designs using these actuators can be commissioned without much financial overhead makes them the
preferred actuator candidates in the literature.
We also assume full state feedback information for the design of the controllers and that all the model parameters
are known. This is an unrealistic assumption; however, our control design is easily extended to account for uncertainty
in these parameters. As a side note, although we assumed all the vehicle model parameters to be known, it is possible
to estimate some of these that are fixed (but unknown) using the sensor information available for the control design
suggested here; this however is outside the scope of this work [15].
3.4 A high-fidelity nonlinear simulation model
Although we base control design on the linear bicycle model, controller evaluation is carried out on a higher fidelity nonlinear simulation model of a vehicle which we call the SimMechanics Vehicle Simulation Model (SM).
This model is created using the multi-body simulation package SimMechanics which is integrated into Mathworks’s
Matlab/Simulink. This is convenient for using various analysis tools in Matlab.
In general, a Simechanics model consists of bodies connected together by various joints and subject to various
forces. Our SM model consists of six unsprung bodies (four wheels and two axles) of negligible mass and one sprung
body as shown in Figure 3. Between the sprung mass and each of the axles, there is a joint which permits a roll degreeof-freedom (DOF); the location of these two joints defines the body roll axis. At each of these roll joints there is a
torsional spring and damper between the sprung mass and the corresponding axle; this models the vehicle suspension.
Connected to each axle are left and right wheels, with the front wheels having a yaw DOF relative to the front axle to
allow for a steering angle, whereas the rear wheels are rigidly fixed to the rear axles. Each wheel has a contact point
where longitudinal, lateral and vertical tire forces are applied. These contact points can leave the ground allowing the
vehicle to roll over. This is a nonlinear model in which any or all wheels can leave the ground and can be used to
simulate rollover.
The tire force model used here is based on the Magic Formula model developed by Pacekja [16]. This model
captures the nonlinear characteristics of the tire forces at large sideslip angles and the effect of the vertical tire force
Fz on the lateral tire force Fy ; see Figure 4 for an definition of the tire force components. With zero longitudinal force
Fx , the vertical force depends on the vehicle loading and motion. the lateral force is a function of the vertical force,
the tire slip angle α , tire lateral stiffness and the friction coefficient between the tire and ground, as described by the
Magic Formula model [16]. Figure 5 shows the lateral force as a function of slip angle for various vertical forces (
with fixed tire stiffness and friction coeffcient), demonstrating the non-linearity of the Magic Formula function. Note
that the peak lateral force and the saturation slip angle are functions of the vertical force.
The introduction of a longitudinal force Fx acting on the tire due to braking and acceleration makes it necessary
to consider the limitations of the combination of lateral and longitudinal forces acting on the tire at the same time.
The physical limitation on the forces applied to the tire are determined by the ground/tire longitudinal and lateral
friction coefficients µ x and µy , resulting in a friction ellipse of longitudinal and lateral forces. In the case where the
5
TopView
RearView
T/2
d
T
a
hrr
T/2
b
T/2
T
z
T
x
y
y
x
z
Figure 3: SimMechanics vehicle model layout
Fy
Fx
Fz
Figure 4: Tire force components
3000
2000
Fy [N]
1000
0
−1000
Fz = 500N
Fz = 1500N
−2000
−3000
−20
Fz = 2500N
−15
−10
−5
0
5
10
15
20
Slip angle α [deg]
Figure 5: Lateral force as a function of the slip angle for various vertical loads
longitudinal and lateral friction coefficients are equal (µ x = µy = µ ), the friction limit results in a friction circle which
bounds the available lateral and longitudinal forces:
6
Table 2: Model parameters
parameter
m
Jxx
Jzz
a
b
T
h
c
k
Cα , f
Cα ,r
λs
value
2800
2275
16088
1.58
1.97
1.6252
0.79
12160
221060
153540
123650
18
unit
[kg]
[kg · m 2 ]
[kg · m 2 ]
[m]
[m]
[m]
[m]
[N · m · s/rad]
[N · m/rad]
[N/rad]
[N/rad]
Fx2 + Fy2 ≤ µ 2 Fz2
(11)
Our model takes this constraint into account. The effects of camber on the tire result in an equivalent slip angle (α eq )
instead of the true tire slip angle (α ), reducing the lateral forces that can be generated by the tire. These corrections
need to be taken into consideration as the tire camber varies as with the wheel roll angle during rollover events.
The linear tire lateral stiffnesses used in the bicycle model are equivalent to the ‘effective axle cornering stiffness’
of the SM model defined by Pacejka [16]. The effective axle stiffnesses are defined as the ratio of the lateral force to
virtual slip angle for each axle i in a steady state turn, where i is either the front or rear axle.
Cα ,i =
Fy,i
.
αa,i
(12)
To ensure that the linear tire characteristics are captured, the effective axle stiffnesses have to be calculated from the
SM model in a steady state turn at low speeds to ensure that linear characteristics are captured.
Since the SM model is sprung at both front and rear axles, the overall suspension roll stiffness and damping is split
between front and rear by the roll stiffness ratio KF and the roll damping ratios CF. This distribution highly affects
the handling behavior of the vehicle from understeer to oversteer. In order to achieve maximum cornering ability by
reducing understeer without inducing oversteer (and thus maximizing rollover propensity), the stiffness and damping
ratios are set to 60% front bias.
Furthermore the limitations of the friction circle on the tire forces necessitates the use of a brake bias between front
and rear braking forces to minimize the case of wheel lockup without utilizing the maximum braking force available.
Although longitudinal load transfer during braking would vary the sizes of the friction circles, we consider the case of
a static brake bias tuned to achieve maximum braking force initially for a range of constant lateral accelerations, and
this is set to 55% front bias.
3.5 Simulation of the SM model
For comparison purposes, the sprung mass of the SM model is set equal to that of the bicycle model, while the unsprung
masses set are very small; they are non-zero to prevent numerical singularities in the simulation. Furthermore, the
bicycle model assumes that the roll center is on the ground plane, thus the roll center height of the SM model is set
to be on the ground. The primary set of simulation parameter values used are as defined in Table 2, as obtained from
[18] and [19], with parameters representative of a typical commercial van.
The vehicle is subjected to an ‘elk-test’ maneuver with steering wheel input profile as illustrated in Figure 6. The
simulated ‘elk-test’ response of the linear bicycle model is compared with the response of both the nonlinear SM
model and a linearization of the SM model generated via MATLAB’s model linearization command ‘linearize.m’.
7
At a speed of 20 m/s (72 kph) with a peak steering wheel angle of δ s,max = 60 deg, Figure 7 demonstrates that the
linearized dynamics of the SM model agree that of the bicycle model. However this also demonstrates the limitation
of the linear bicycle model as a tool for prediction of rollover as it does not consider the saturation of lateral tire forces.
As the maximum steering input is increased to δ s,max = 80 deg (Figure 8) the differences between the linear bicycle
model and the nonlinear SM model becomes even more apparent. The nonlinear SM model exhibits a significantly
lower peak LTR compared to the linear models when large steering inputs are applied or when the initial velocity of
the vehicle is large.
60
+ δs,max
Steering wheel input δs [deg]
40
20
0
−20
Δ T = 3 sec
−40
−δ
s,max
−60
0
2
4
6
8
10
12
Simulation time [s]
Figure 6: Steering wheel input history for simulated ‘Elk-Test’
1
Bicycle
SM
Linearized SM
0.8
Load Transfer Ratio
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
0
2
4
6
8
10
12
Simulation Time [s]
Figure 7: LTR comparison for bicycle, SM and linearized SM models at v=20 m/s and δ s,max =60 deg
Further simulations with the SM model show that with large steering inputs and speeds the vehicle does roll
over, which confirms the necessity for control systems to prevent such phenomenon. The SM model can be used to
determine the conditions of steering input and speed which result in untripped rollover for a vehicle with parameters
given in Table 2. One of the distinctions that the SM model is able to provide over the linear bicycle model is that
wheel liftoff and rollover are not equivalent. In Table 3 the vehicle is tested at initial longitudinal velocities of 20, 30
and 40m/s. The steering input to lift a single wheel off the ground is denoted by δ s,li f t , whereas the input to drive the
magnitude of LT R d to one is given by δ s,maxLT R . The third column of results shows δ s,roll , or the steering input needed
to induce untripped rollover.
Full details of the SimMechanics vehicle simulation model development and verification can be found in [17].
8
1
Bicycle
SM
Linearized SM
0.8
Load Transfer Ratio
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
0
2
4
6
8
10
12
Simulation Time [s]
Figure 8: LTR comparison for bicycle, SM and linearized SM models at v=20 m/s and δ s,max =80 deg
Table 3: Conditions for wheel lift, |LT R d | = 1 and vehicle rollover
Speed v [m/s]
20
30
40
δs,li f t [deg]
137
67
45
δs,maxLT R [deg]
147
72
49
δs,roll [deg]
167
80
52
4 State feedback controllers for robust disturbance attenuation
We are interested in designing a controller to prevent rollover that is robust with respect to parameter uncertainty. Our
starting point is in results obtained by Pancake, Corless and Brockman [6] for uncertain systems of the form
ẋ = A(θ )x + B(θ )w + Bu(θ )u
zi
= Ci (θ )x + Diu(θ )u ,
i = 1, . . . , r
(13)
(14)
where x(t) ∈ Rn is the state at time t ∈ [0, ∞), w(t) ∈ Rm is a bounded disturbance input, u(t) ∈ R mu is the control
input, and zi (t) ∈ R pi , i = 1, 2, ..., r are performance outputs. All the uncertainty and nonlinearities in the system are
captured in the parameter vector θ which can depend on t, x, w and u. We wish to synthesize a stabilizing controller
which prevents the peak values of the performance outputs exceeding certain values. In other words, we want to design
a feedback controller, which guarantees a bounded performance output given a bounded uncertain disturbance, that is,
||w(t)|| ≤ wmax . We consider linear state feedback controllers of the form
u = Kx ,
(15)
where K is a constant state feedback gain matrix. This results in a closed loop system described by
ẋ =
zi
=
[A(θ ) + Bu(θ )K]x + B(θ )w
(16)
[Ci (θ ) + Diu (θ )K]x ,
(17)
i = 1, . . . , r .
We require the following assumptions.
Assumption 1 There are matrices
A j , B j , Bu j ,
j = 1, . . . , N
so that for each θ , the matrix [A(θ ) B(θ ) B u (θ )] can be written as a convex combination of [A 1 B1 Bu1 ],. . ., [AN BN BuN ].
9
Assumption 2 For each i = i, . . . , r, there are matrices
Cik , Diuk ,
k = 1, . . . , Mi
so that for each θ , the matrix [Ci (θ ) Diu (θ )] can be written as a convex combination of [C i1 Diu1 ], . . . , [CiMi DiuMi ].
Remark 1 Suppose that each of the matrices A(θ ), B(θ ), B u (θ ) depends in a multi-affine fashion on the components
of an L̄-vector θ and each element of θ is bounded; specifically,
θ l ≤ θl ≤ θ l
for
l = 1, . . . , L̄ .
(18)
Then, for all θ , the matrix [A(θ ) B(θ ) B u (θ )] can be expressed as a convex combination of the 2 L̄ vertex matrices
corresponding to the extreme values of the components of θ , that is, θ l = θ l or θ l for l = 1, . . . , L̄.
We have now the following result which is useful for control design.
Theorem 1 Consider a nonlinear/uncertain system described by (13)-(14) and satisfying Assumptions 1 and 2. Suppose that there exist matrices S = S T > 0 and L along with scalars α 1 , . . . , αN > 0 and γ 1 , . . . , γ r ≥ 0 such that the
following matrix inequalities hold:
A j S+Bu j L+SATj +LT BTu j +α j S
BTj
Bj
−α j I
≤ 0,
(19)
for j = 1, . . . , N and
−S
SCikT +LT DTiuk
Cik S+Diuk L
−γ 2i I
≤ 0,
(20)
for i = 1, . . . , r and k = 1, . . . , Mi . Then the controller
u = Kx
K = LS−1
with
(21)
results in a closed loop nonlinear/uncertain system which has the following properties.
(a) The undisturbed system (w = 0) is globally exponentially stable, that is, all state trajectories decay exponentially.
(b) If the disturbance input is bounded, that is, w(t) ≤ w max for all t then, for zero initial state, the performance
outputs z1 , . . . , zr of the closed loop system are bounded and satisfy
zi (t) ≤ γ i wmax .
(22)
The scalars γ 1 , . . . , γ r are called performance levels and can be regarded as measures of the ability of the closed
loop system to attenuate the effect of the disturbance input on the performance outputs; a smaller γ i means better
performance in the sense of increased attenuation. For a proof of the theorem, see [7].
5 Rollover prevention controllers and simulation results
Here we use the results of the previous section to obtain rollover prevention controllers using differential braking as
the control input. We consider the steering wheel angle δ s (in degrees) as the disturbance input.
For reasons discussed earlier, we choose z 1 = LT Rd given by (8) as one performance output; we want to keep
z1 ≤ 1 for the largest possible steering inputs. We consider the magnitude of the braking force u to be limited by
the weight mg of the vehicle; so we choose z 2 = u as a second performance output. The resulting system with two
performance outputs can be described by
ẋ = Ax + Bδs δs + Buu
z1 = C1 x
z2
=
(23)
u,
First we obtain a control design which is based on the above model with a fixed speed using the vehicle parameters
in Table 2; we call this the fixed model controller. We then consider the effect of varying speed in our control design
and we obtain a control design assuming that the speed varies over some prespecified range; we call this the robust
controller.
10
5.1 Controller based on fixed speed
Here we base controller design on model (23) in which all matrices are constant and correspond to a fixed vehicle
speed of v = 40m/s. To obtain a state feedback controller, we applied Theorem 1. Since we desire that z 1 ≤ 1 and
z2 ≤ mg for the largest possible steering inputs, we considered γ 2 = mgγ 1 . By performing a line search with respect
to the scalar α we obtained a minimum value of 0.0096 for γ 1 . The corresponding control gain matrix is
Kuconst = mg · [ −12.7651 5.1246 0.0854 −3.6968 ] .
Remark 2 Consider the constant speed linear bicycle model subject to the above control gain matrix. According to
(22), the constraints on the outputs will not be violated for this constant speed closed loop system if the maximum
magnitude δs,max of the steering wheel disturbance input satisfies δ s,max ≤ 1/γ 1 ≈ 104.69◦. However application of
steering inputs and the braking controller both reduce vehicle speed in the SM model. As the vehicle speed reduces,
its tendency to rollover decreases and the vehicle can actually tolerate disturbance inputs with magnitude considerably larger than 1/γ 1 . In addition, as a consequence of the friction circle, application of braking forces reduces the
maximum allowable lateral force at each wheel. In numerical simulations the above controller gain matrix was able
to maintain |LT Rd | ≤ 1 for steering input magnitudes up to δ s,max = 165◦. However it is necessary to note that with
the SM model the condition of |LT R d | = 1 does not correspond to the limit for rollover, only the condition that both
inside wheels lift off the ground.
For numerical simulations we chose a driver steering input corresponding to an ‘elk-test’ (Figure 6); we chose
an initial speed of v = 40m/s and first simulate the system with a steering input δ s,max = 35 deg, below the liftoff
requirement as stated in Table 3. The LTR response of both the controlled and uncontrolled cases are shown in Figure
9. Although the design intention of the controller is to prevent rollover, it also reduces the load transfer ratio of the
vehicle significantly.
Uncontrolled
Fixed speed controller
1
0.8
0.6
0.4
LTRd
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
1
2
3
4
5
6
7
8
9
10
11
12
Simulation Time [s]
Figure 9: LT Rd response for controlled and uncontrolled vehicle simulated at below liftoff conditions.
Next we simulate the vehicle subject to an ‘elk-test’ maneuver with a peak steering magnitude of δ s,max = 104.69◦.
The vehicle with the proposed controller satisfies |LT R d | < 1, achieving the intended design goal whereas the uncontrolled vehicle rolls over by t = 4sec.The corresponding normalized control history u/mg for the controller is shown
in Figure 11, where we observe that the maximum input magnitude during the maneuver was 80% of the weight of the
vehicle. The speed history of the vehicle is shown in Figure 11. Notice that the dramatic speed drop of the controlled
vehicle is a combination of both steering inputs and braking action, suggesting there may be potential performance
gains if the controller design takes into account the deceleration of the vehicle.
In the following subsection we demonstrate how our control design method can be extended to account for varying
parameter uncertainties such as a variable velocity.
11
Uncontrolled
Fixed speed controller
1
0.8
0.6
0.4
LTR
d
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
1
2
3
4
5
6
7
8
9
10
11
12
Simulation Time [s]
Figure 10: LT Rd response for controlled and uncontrolled vehicle
1
40
Uncontrolled
Fixed speed controller
0.8
35
0.6
Speed v [m/s]
Control Input/Weight
30
0.4
0.2
0
25
20
−0.2
15
−0.4
10
−0.6
−0.8
0
2
4
6
8
10
5
0
12
2
Simulation Time [s]
4
6
8
10
12
Simulation Time [s]
Figure 11: Normalized control history and speed history
5.2 Controller based on variable speed model
In this section, we present a rollover controller design which takes into account varying vehicle speed; we use the
constant model parameters given in Table 2.
We assume that the speed is bounded above and below by v and v,
respectively, that is, v ≤ v ≤ v. In order to represent typical freeway driving conditions for a compact passenger vehicle
we chose v = 25m/s, and v = 40m/s as the extremum design speeds. Again, we used the model (23) for controller
design, where the matrices A, B δs , Bu and C1 are given in (4) and (10). System matrices B u and C1 are independent of
speed. The matrices A and B δs can be expressed as affine linear functions of the time-varying parameters θ 1 := 1/v
and θ2 := 1/v2 . These parameters are bounded as follows:
θ 1 ≤ θ1 ≤ θ 1 ,
where
θ1 =
1
,
v
θ1 =
1
,
v
θ 2 ≤ θ2 ≤ θ 2
θ2 =
12
1
,
v2
θ2 =
(24)
1
.
v2
Hence our system description satisfies Assumptions 1 and 2 with the following vertex matrices
A1
A3
Bδs ,1
Bδs ,3
A2 = θ 1Y1 + θ 2Y2 + Y3 ,
= θ 1Y1 + θ 2Y2 + Y3 ,
A4 = θ 1Y1 + θ 2Y2 + Y3 ,
= θ 1Y1 + θ 2Y2 + Y3 ,
π
Cα , f Jxeq
Cα , f a
hCα , f
= Bδs ,2 =
0
mJxx θ 1
Jzz
Jxx
180λs
π
Cα , f Jxeq
hCα , f
Cα , f a
= Bδs ,4 =
0
mJxx θ 1
Jzz
Jxx
180λs
T
T
,
,
where
Y1
Y2
Y3
=
=
=
⎡
⎢
⎢
⎢
⎣
σ Jx
− mJxxeq
0
0
0
⎢
⎢
⎣
hρ
Jxx
0
ρ Jxeq
mJxx
⎡
0
⎢ 0
⎢
⎣ 0
0
⎡
0
− Jκzz
0
0
0
0
ρ
Jzz
− Jhxxσ
0
0
0
0
0
h(mgh−k)
Jxx
− Jhcxx
0
0
0
⎤
0
0
0
⎤
⎥
⎥
⎥,
⎦
0
0 ⎥
⎥,
0 ⎦
0
−1
0
0
0
0 − Jcxx
0
1
0
0
mgh−k
Jxx
0
⎤
⎥
⎥.
⎦
We used Theorem 1 to design a controller which guarantees performance levels γ 1 and γ2 = mgγ1 , in presence of the
any variations in speed satisfying v ≤ v ≤ v. We achieved γ 1 = 0.0097. Also the corresponding control gain matrix is
Krobust = mg · [ −14.8499 5.4573 0.1372 −4.6320 ] .
Note that, according to (22) the maximum theoretical driver steering disturbance input permitted is, δ s,max = 1/γ 1 ≈
102.60◦. In our simulations however, for the reasons explained in Remark 2, the robust controller was able to keep
|LT Rd | ≤ 1 for driver steering inputs with magnitudes up to δ s,max = 165◦.
For numerical simulations, we used the same obstacle avoidance (‘elk-test’) scenario as before (see Figure 6), however with a peak driver steering input of magnitude δ s,max = 102.60◦ and an initial speed of v = 40m/s. Comparison of
the LT Rd response with with a steering input of δ s,max = 102.60deg in Figure 12 shows that both controllers result in
similar LT Rd responses. However inclusion of the vehicle speed as a varying parameter uncertainty in the controller
design does not improve the rollover prevention capability of the vehicle.
The steering profile corresponding to this maneuver and a comparison of speed histories for the uncontrolled
vehicle as well as the controlled vehicles with the two suggested control designs are shown in Figure 13. Both
controllers show similar control histories and deceleration profiles, with the fixed velocity controller ending with a
final velocity of 6.98m/s and the robust controller with 7.07m/s.
It is of particular interest for us to see how the suggested controllers affect the vehicle path, shown in Figure 14.
The trajectory of the vehicle CG with and without controllers shows where the rollover for the uncontrolled vehicle
occurs and a comparison of the response from both controllers relative to the uncontrolled case. The fixed velocity
controller shows a tighter turn radius than the robust controller through the initial steering input, resulting in a slightly
higher lateral translation of 20.1m to the robust controller’s 19.9m from the initial lateral position. Also, because of
the changing velocities and effects of tire force saturation, the symmetric input as shown in Figure 6 does not result
in the vehicle heading returning to its initial direction. However in a real driving situation a driver would be able to
react to the difference in heading and control the steering wheel such that the final trajectory was purely longitudinal if
necessary. Furthermore, the yaw angles (measured inertially) at each longitudinal position for each controller is shown
below to illustrate that the vehicle does maintain the correct orientation throughout the trajectory (i.e. the vehicle does
not spin out at any point in the maneuver).
Comment : From the simulation results for the fixed model and the robust controllers, we observe that both
controllers are effective in reducing the vehicle load transfer ratio LT R d , and thus preventing rollover, however no
notable improvement is achieved using the ‘robust’ controller.
13
Fixed speed controller
Robust controller
1
0.8
0.6
0.4
LTR
d
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
1
2
3
4
5
6
7
8
9
10
11
12
Simulation Time [s]
Figure 12: LT Rd response for comparison for different controllers K uconst and Krobust
1
40
Fixed speed controller
Robust controller
Fixed speed controller
Robust controller
0.8
35
0.6
Speed v [m/s]
Control Input/Weight
30
0.4
0.2
25
20
0
15
−0.2
10
−0.4
−0.6
0
2
4
6
8
10
5
0
12
Simulation Time [s]
2
4
6
8
10
12
Simulation Time [s]
Figure 13: Comparison of control histories and longitudinal velocities for both controllers
Comment : Our design is easily extended to incorporate other sources of parameter uncertainty such as the vehicle
parameters, mass and center of gravity height.
Since the wheel can only deliver as much lateral and braking force as the friction circle permits, it is of interest
to compare the commanded braking force and the actual braking force on each tire; this is shown in Figure 15. Both
the fixed speed and robust controllers have near identical control input histories and responses, so although the values
shown in Figure 15 apply to the robust controller, there is little deviation if plotted using simulation data corresponding
to the fixed speed controller. The commanded braking force from the differential braking controller (‘Cmd’), the actual
braking force applied to the tire (‘Act’) and the maximum available braking force as determined by the vertical load
and the ground friction coefficient (‘Avail’) for each wheel is shown for the maneuver. With the way the tire model is
set up, at no point is the actual applied braking force (‘Act’) allowed to exceed the maximum available force (‘Avail’),
however between t=2s and t=5s on the right rear tire the commanded braking force far exceeds the available force and
thus the braking force saturates. Similarly when the steering input switches directions at t = 5 s, it can be seen that the
commanded braking force on the rear left tire exceeds the available force.
Due to the friction circle constraint, if most of the available friction force on a tire is being used to apply a
braking effort then there is little lateral force acting on that tire and load transfer is reduced. Of course, this reduces
lateral acceleration and the vehicle effectively slides out of the initiated turn. The above remarks suggest that that the
controller can be designed for better performance if longitudinal load transfer is taken into consideration.
14
Lateral distance [m]
30
20
10
No controller
Fixed speed controller
Robust controller
0
−10
0
50
100
150
200
250
300
Longitudinal distance [m]
Yaw Angle [deg]
60
No controller
Fixed speed controller
Robust controller
40
20
0
−20
0
50
100
150
200
250
300
Longitudinal distance [m]
Figure 14: Comparison of CG trajectories and yaw angles
FL
FR
15000
15000
Cmd
Act
Avail
10000
Force [N]
Force [N]
Cmd
Act
Avail
5000
0
0
2
4
6
8
10
10000
5000
0
12
0
2
Simulation time [s]
4
8
10
12
RR
RL
8000
12000
Cmd
Act
Avail
7000
Cmd
Act
Avail
10000
Force [N]
6000
Force [N]
6
Simulation time [s]
5000
4000
3000
8000
6000
4000
2000
2000
1000
0
0
2
4
6
8
10
0
12
Simulation time [s]
0
2
4
6
8
10
12
Simulation time [s]
Figure 15: Commanded and actuated braking forces of each wheel
6 Conclusions
We have presented a methodology for the design of vehicle rollover prevention systems using differential braking. By
introducing the load transfer ratio LT R d , we obtain a system performance output whose value provides an accurate
measure for determining the onset of rollover. Our rollover prevention system is based upon recent results from Pancake, Corless and Brockman, which provide controllers to robustly guarantee that the peak values of the performance
outputs of an uncertain system do not exceed certain values. Simulation of the differential braking controller on a
high-fidelity nonlinear vehicle model demonstrates the benefits of the proposed approach in a real-life problem.
15
References
[1] National Highway Traffic Safety Administration (NHTSA), Traffic Safety Facts 2004: A Compilation of
Motor Vehicle Crash Data from the Fatality Analysis Reporting System and the General Estimates System,
Technical Report, 2006.
[2] Runge, J., Statement to the Committee on Commerce on SUV Safety, 2003, (available at
http://www.nhtsa.dot.gov.)
[3] Chou, C.C., McCoy, R.W. and Le, J., “A literature review of rollover test methodologies,” Int. J. Vehicle
Safety, Vol. 1, pp. 200-237, 2005.
[4] Chou, C.C., Wu, F., Gu, L. and Wu, S.R., “A review of mathematical models for rollover simulations,”
AMD-Vol. 230/BED-Vol. 41, Crashworthiness, Occupant Protection and Biomechanics in Transportation
Systems, ASME, 1998.
[5] Carlson C.R. and Gerdes J.C., “Optimal rollover prevention with steer by wire and differential braking”, Proceedings of the ASME International Mechanical Engineering Congress and Exposition, IMECE’03, Washington, D.C., USA, 2003.
[6] Pancake T., Corless M. and Brockman M., “Analysis and control of polytopic uncertain/nonlinear systems
in the presence of bounded disturbance inputs”, Proceedings of the American Control Conference, Chicago,
IL, USA, 2000.
[7] Pancake T., Corless M. and Brockman M., “Analysis and Control of a class of uncertain/nonlinear systems
in the presence of bounded disturbance inputs”, in preparation.
[8] Palkovics L., Semsey À. and Gerum E., “Roll-over prevention system for commercial vehicles-additional
sensorless function of the electronic brake system”, Vehicle System Dynamics, Vol. 4, pp. 285-297, 1999.
[9] Wielenga T.J., “A method for reducing on-road rollovers: anti-rollover braking”, SAE Paper No. 1999-010123, 1999.
[10] Chen B. and Peng H., “Differential-braking-based rollover prevention for sport utility vehicles with humanin-the-loop evaluations”, Vehicle System Dynamics, Vol. 36, pp. 359-389, 2001.
[11] Ackermann J. and Odenthal D., “Robust steering control for active rollover avoidance of vehicles with
elevated center of gravity”, Proceedings of International Conference on Advances in Vehicle Control and
Safety, Amiens, France, 1998.
[12] Odenthal D., Bünte T. and Ackermann J., “Nonlinear steering and braking control for vehicle rollover avoidance”, Proceedings of European Control Conference, Karlsruhe, Germany, 1999.
[13] Kiencke U. and Nielsen L., Automotive Control Systems for Engine, Driveline and Vehicle, Springer-Verlag
& SAE Int., Berlin, 2000.
[14] Kamnik R., Böttiger F. and Hunt K., “Roll dynamics and lateral load transfer estimation in articulated heavy
freight vehicles: A simulation study”, Proceedings of the Institution of Mechanical Engineers, Part D, 2003.
[15] Shorten R., Solmaz S. and Akar M., Method for determining the centre of gravity for an automotive vehicle,
US Patent no: US 2009/0024269 A1, Publication date: January 22, 2009.
[16] Pacejka, H.B., Tire and Vehicle Dynamics, Second Edition, SAE Int., 2006
[17] Chiu, J., “A Simulink Model For Vehicle Rollover Prediction and Prevention”, MS Thesis, Purdue University, 2008.
[18] Schofield, B., “Vehicle Dynamics Control for Rollover Prevention”, Licentiate Thesis, Lund University,
2006.
16
[19] Sugiyama, H., Shabana, A., Omar, M. and Loh, W., “Development of nonlinear elastic leaf spring model
for multibody vehicle systems”, Computer Methods in Applied Mechanics and Engineering, Vol. 195 No.
50-51, pp. 6925-6941, 2006.
[20] Solmaz S., Corless M. and Shorten R., “A methodology for the design of robust rollover prevention controllers for automotive vehicles: Part 2-Active steering”, HYCON-CEMaCS Joint Workshop on Automotive
Systems & Control, June 1-2, 2006.
[21] Solmaz S., Corless M. and Shorten R., “A methodology for the design of robust rollover prevention controllers for automotive vehicles with active steering”, International Journal of Control, Vol. 80, No. 11, pp.
1763-1779, November 2007.
[22] Schofield, B., Hägglund, T., and Rantzer, A., “Vehicle Dynamics Control And Controller Allocation for
Rollover Prevention”, Proceedings of the IEEE International Conference on Control Applications, Munich,
Germany, 2006.
[23] Schofield, B., and Hägglund, T., “Optimal Control Allocation in Vehicle Dynamics Control for Rollover
Mitigation”, Proceedings of the 2008 American Control Conference, Seattle, WA, USA, 2008.
17