MULTIBODY DYNAMICS 2007, ECCOMAS Thematic Conference

C.L. Bottasso, P. Masarati, L. Trainelli (eds.)

Milano, Italy, 2528 June 2007

REAL TIME SIMULATION OF LARGE VEHICLE SYSTEMS

Rill Georg? , Chucholowski Cornelius
? Department

of Mechanical Engineering
Regensburg University of Applied Sciences, Galgenbergstr. 30, 93053 Regensburg, Germany
e-mail: [email protected], web page:
http://homepages.fh-regensburg.de/rig39165/
TESIS

DYNAWare GmbH
Baierbrunnerstr. 15, 81379 Munchen, Germany
e-mail: [email protected], web-page: http://www.tesis.de

Keywords: Vehicle Dynamics, Multibody Systems, Numerical Integration, Differential Algebraic Equations, Real-Time Application
Abstract. This paper shows that real time performance for complex vehicle systems can be
achieved by a sophisticated modeling technique and a tailored integration algorithm. By taking
advantage of the specific structure different types of vehicles can be described by one generic
model. Besides the steering system and complex axle suspension systems including compliances
the critical model part tire and wheel is discussed in detail. It is shown that a sophisticated first
order dynamic tire model allows the simulation of hard braking and stand still manoeuvres.
The coupling of two generic vehicle models to a vehicle combinations is performed by differential algebraic equations. In this paper the modified implicit Euler integration algorithm used
successfully within the software package ve-DYNA is extended to index-3 differential algebraic
equations. The modeling concept and the numerical solution are illustrated on a simple example. It allows to model complex systems very easily just by combining existing models. Real
time applications even for large vehicle systems like tractor trailer combinations will be possible because the additional computation effort is very little.

Rill Georg, Chucholowski Cornelius

1
1.1

INTRODUCTION
Overview

Driving simulators and hardware in the loop test rigs have become very popular in vehicle dynamics. Vehicle modeling is usually done by multibody systems, Ref. [6]. At present several
providers offer full passenger car models for real time applications. Real-time performance
is achieved by rather simple model approaches where details in subsystems are omitted and
the wheel/axle kinematics is approximated by nonlinear lookup tables. An automatic model
reduction procedure is presented in Ref. [7]. To achieve real time performance even for complex vehicle systems a sophisticated modeling technique and a tailored integration algorithm is
needed, Ref. [1]. But the simulation of large vehicle systems like a passenger car with trailer,
multi-axled trucks or tractor trailer combinations in real time is still a demanding task.
1.2

Model Structure

Different vehicles like agricultural tractors, passenger cars, buses and trucks can be handled
by one generic vehicle model, Fig .1. The vehicle framework represents the kernel of such

G. R i

l l 10.

03.

03

Driver'
Cab

Engine

Driver

Passenger
Seat

Load

Chassis
Steering
system

Axle 1

Axle 2

Wheel
Tire

Wheel
Tire

Drive
train

Axle n-1 Axle n

Wheel
Tire

Wheel
Tire

Road
Figure 1: Different types of vehicles and generic model structure

a model. It includes at least the module chassis and modules for the wheel/axle suspension
systems. The vehicle framework may be supplemented by modules for the load, passenger/seat
models, an elastically suspended engine, and a drivers cab. A simple load module just takes the
mass and inertia properties of the load into account. Using Jourdains principle of virtual power
in combination with generalized speeds results in two sets of nonlinear first order differential
equations for the vehicle framework
KV (yV ) y V

= zV ,

MV (yV ) zV

= qV (yV , zV , yS , zS , sF ) ,

(1)

where the vector yV contains the generalized coordinates of the vehicle framework, the kinematic matrix KV defines the generalized speeds which are arranged in the vector zV , MV names
the mass matrix and qV is the vector of generalized forces and torques applied to the vehicle
framework. This vector depends also on the states of the subsystems yS , zS and the internal
states sF of dynamic force elements. For standard vehicle dynamics analysis the chassis can be
2

Rill Georg, Chucholowski Cornelius

modeled by one rigid body. At least the torsional flexibility of the frame has to be taken into
account on trucks, Ref. [1]. A multibody approach to a flexible frame is presented in Ref. [13].
The subsystems elastically suspended engine, passenger/seat and in heavy truck models a
suspended drivers cab can all be handled by a generic free body model, Ref. [13]. Models
of different complexity for the steering systems can be found in Ref. [10]. A generic model
for a 4x4 drive train is described in Ref. [13]. Similar to the vehicle framework, mechanical
subsystems are typically described by two sets of nonlinear first order differential equations
KS (yS ) y S = zS ,

(2)

MS (yS ) zS = qS (yS , zS , yV , zV , sF ) ,
where KS defines the generalized speeds and MS the mass matrix of the subsystem.
Sophisticated vehicle models also take the dynamics of force elements, like tires, dampers,
hydro or rubber mounts into account. The dynamics of these elements can be described by an
additional set of first order differential equations
s F = fF (sF , yV , zV , yS , zS ) .
1.3

(3)

Numerical Integration

The differential equations Eqs. (1), (2) and (3) may be put together to a huge system of
first order differential equations x = f (x) and solved by common ode solvers. In real time
applications the number of function evaluations is limited by the complexity of the system and
the power of the computer. Hence, explicit solvers will run into stability problems because
in order to achieve real time performance comparatively large step sizes are needed. Implicit
solvers require the Jacobian df /dx which for large vehicle system is very time consuming. In
ve-DYNA the dynamic force elements, the different subsystems and the vehicle framework are
solved sequentially by a modified implicit euler integration algorithm. Applied to Eqs. (3), (2)
and (1) one will get


k
k
k
k
skF = h fF sk+1
sk+1
F
F , y V , z V , y S , zS ,

MS ySk+1





MV yVk+1





KS ySk+1

KV yVk+1

zSk+1 zSk

= h qS ySk+1 , zSk+1 , yVk , zVk , sk+1

,
F

ySk+1 ySk

= h zSk+1 ,

zVk+1 zVk

= h qV yVk+1 , zVk+1 , ySk+1 , zSk+1 , sk+1

,
F

yVk+1 yVk

= h zVk+1 ,

(4)

where h denotes the integration step size and the superscripts k and k+1 denote the states at time
t and t + h respectively. Using the approximations MS (ySk+1 ) MS (ySk ), KS (ySk+1 ) KS (ySk ),
MV (yVk+1 ) MV (yVk ), KV (yVk+1 ) KV (yVk ), and





fF  k+1
k
k
k
k
k
k
k
k
k
k
,
(5)
s

f
s
,
y
,
z
,
y
,
z
+
fF sk+1
,
y
,
z
,
y
,
z
F
F
F
F
V
V
S
S
F
V
V
S
S
sF


qS ySk+1 , zSk+1 , yVk , zVk , sk+1

qS ySk +hKS1 zSk , zSk , yVk , zVk , sk+1
F
F
+


qS  k+1 k 
qS  k+1
yS (ySk +hKS1 zSk ) +
z zS ,
yS |
zS S
{z
}


hKS1 zSk+1 zSk

3

(6)

Rill Georg, Chucholowski Cornelius

qV yVk+1 , zVk+1 , ySk+1 , zSk+1 , sk+1

qV yVk +hKV1 zVk , zVk , ySk+1 , zSk+1 , sk+1
F
F
+


qV  k+1 k 
qV  k+1
yV (yVk +hKV1 zVk ) +
zV zV
yV |
zV
{z
}


hKV1 zVk+1 zVk

(7)
results in the integration algorithm
!1

sk+1
F
zSk+1



fF
+ h I h
=
fF skF , yVk , zVk , ySk , zSk ,
sF
!1


qS
2 1 qS
k
= zS + h MS h
h KS
qS ySk +hKS1 zSk , zSk , yVk , zVk , sk+1
,
F
zS
yS

skF

ySk+1 = ySk + h KS1 zSk+1 ,

(8)

1 

q
q
V
V
zVk+1 = zVk + h MV h
h2 KV1
qV yVk +hKV1 zVk , zVk , ySk+1, zSk+1, sk+1
,
F
zV
yV
|

{z

MVI

yVk+1 = yVk + h KV1 zVk+1 ,

where MVI denotes a new mass matrix modified by implicit correction terms. In this approach,
only the partial derivatives fF /sF , qS /yS , qS /zS , and qV /yV , qV /zV are needed.
By avoiding the computation of the Jacobian for the whole system a very fast but still sufficiently stable solver is generated. However, the subsystem tire and wheel, the steering system,
sophisticated axle suspension systems which include compliances and the coupling between
vehicle combinations are critical in particular.
2
2.1

CRITICAL MODEL PARTS

Tire and Wheel

Different approaches and tire models are available, Ref. [9]. Structural tire models are very
complex and computer time consuming. Handling models are characterized by a useful compromise between user-friendliness, model-complexity and efficiency in computation time on
the one hand, and precision in representation on the other hand. Within handling models, the
6

Longitudinal tire force Fx=Fx(sx) [kN]

6

Lateral tire force Fy=Fy() [kN]

150

100

50

0
1.8 kN
3.2 kN
4.6 kN
5.4 kN

-2
-4
-6
-40

-20

0
s [%]
x

20

0
1.8 kN
3.2 kN
4.6 kN
6.0 kN

-2
-4
40

-6

Self aligning torque Tz=Tz() [Nm]

-20

-10

[o]

10

20

-50

1.8 kN
3.2 kN
4.6 kN
6.0 kN

-100
-150

-20

-10

[o]

Figure 2: Steady state tire forces and self aligning torque modeled by TMeasy, Ref. [5]

10

20

Rill Georg, Chucholowski Cornelius

steady state tire forces and torques are generated as functions of the longitudinal slip sx and the
lateral slip sy or the slip angle = arctan(sy ) respectively, Fig. 2.
To investigate the dynamics of the wheel motion a quite simple quarter car model can be
used where neither the suspension travel nor the tire deflection is taken into account, Fig. 3. The
v

Fx

Figure 3: Simple vehicle wheel model

linear momentum for the vehicle and the angular momentum for the wheel result in
m v = Fx

and = T r Fx ,

(9)

where m is the corresponding vehicle mass, and r name the inertia and radius of the wheel
and T describes a driving or braking torque. If the vehicle is neither driven nor braked, the
wheel will be close to the rolling condition. Then, the velocity of the vehicle and the angular
velocity of the wheel can be described by
v = v0 + 4v

and =

v0
+ 4 ,
r

(10)

where 4v  v0 and 4  v0 /r denote small disturbances of the steady state situation v = v0

and = 0 . Now, the longitudinal slip is small too, sx  1 and the steady state force
characteristics can be approximated by
Fx dF0 sx = dF0

r 4 4v
,
|v0 |

(11)

where dF0 describes the initial inclination of the longitudinal tire characteristics Fx (sx ). Assuming a constant driving velocity v0 = const and a vanishing driving or braking torque T = 0
the equations of motion (9) result in a set of linear differential equations
m 4v =

dF0
(r 4 4v)
v0

dF0
4 = r
(r 4 4v) .
v0

and

(12)

The dynamics of this simple vehicle wheel model is characterized by the eigenvalues
dF0
1 = 0 and 2 =
v0

r2
1
+
m

(13)

Using a steady state approach to tire forces causes problems in drive away or braking to stand
still maneuvers because the wheel dynamics will become more and more stiff if the vehicle
slows down. A slip modification allows to overcome this problem, but it significantly changes
the wheel dynamics especially at low wheel angular velocities and may hence produce results
with poor accuracy, Ref. [16].
5

Rill Georg, Chucholowski Cornelius

However, a simple but effective extension to first order dynamic tire forces and torques allows
a smooth transition from normal driving situations to stand still and keeps the dynamics of the
system finite, Ref. [12]. At first a generalized slip
s =

r


sN
x

2

+ sN
y

2

(14)

is introduced where the normalized slips are described by

sN
x =

(vx rD )
rD || sx + vN

and

sN
y =

vy
.
rD || sy + vN

(15)

The angular velocity of the wheel multiplied by the dynamic rolling radius rD results in the
average transport velocity of the tread particles and vx , vy are the longitudinal and the lateral
component of the contact point velocity. The small but artificial velocity vN > 0 was introduced
here to avoid the singularity at stand still. In normal driving situations where rD || = 0  vN
holds the influence of vN will hardly be noticeable. The tire force characteristics in longitudinal
Fx = Fx (sx ) and lateral direction Fy = Fy (sy ) can be combined to a generalized force characteristics F = F (s), Fig. 4. A nearly equally weighted contribution of the longitudinal and
F(s)

dF0
Fy(sy)

Fx(sx)

sy / ^sy

sx / ^sx
Figure 4: Generalized tire characteristics

lateral tire characteristics to the generalized tire characteristics can be achieved by appropriate
normalization factors sx and sy . In return the longitudinal and the lateral forces follow from the
according projections
sN
F N
sx = f s N
Fx = F cos = F x =
x
s
s
and

(16)

sN
F N
y
=
s = f sN
(17)
y ,
s
s y
where the function f (s) = F (s)/s was introduced to describe the generalized tire characteristics. Note, at vanishing slip values f (s = 0) is well defined by the initial inclination of the
generalized force characteristics, F (0)/0 = dF/ds|s=0 = dF 0 .
The tire forces Fx and Fy acting in the contact patch deflect the tire in longitudinal and lateral
direction, Fig. 5. In a first order approximation the dynamic tire force in longitudinal direction
follows from
Fx
Fx (vx + x e ) Fx (vx ) +
x e ,
(18)
|
{z
}
| {z }
vx
FxD
FxS
Fy = F sin = F

Rill Georg, Chucholowski Cornelius

rim

dx

cx

rim
dy
tire

tire

xe vx - rD

Fx

cy

Fy

ye

vy

Figure 5: Tire deflection in longitudinal and lateral direction

where xe names the longitudinal tire deflection. In steady state the longitudinal tire force FxS
will be provided by Eq. (16) as a function of the normalized longitudinal slip sN
x . Its derivative
with respect to the longitudinal component vx of the contact point velocity can be calculated
from
FxS
FxS sN
FxS
1
x
=
=
(19)
N
N
vx
sx vx
sx rD ||
sx + v N
where the definition of the normalized longitudinal slip in Eq. (15) was used to generate the
derivative of the longitudinal slip with respect to the longitudinal component of the contact
point velocity. Corresponding to the first order approximation in Eq. (18) the partial derivative
of the steady state tire force with respect to the normalized slip is approximated by its global
derivative
FxS
f sN
FxS
x

=
= f.
(20)
N
N
sN
s
s
x
x
x
Then, Eq. (18) reads as
FxD f sN
x + f

1
x e ,
rD ||
sx + v N

(21)

where according to Eq. (16) the steady state longitudinal tire force FxS was replaced by the term
f sN
x . On the other hand, the dynamic tire force can be derived from
FxD = cx xe + dx x e ,

(22)

where cx and dx denote stiffness and damping properties of the tire in the longitudinal direction.
Combining Eqs. (21) and (22) results in a first order differential equation for the longitudinal
tire deflection
!
1
x e = f sN
(23)
dx + f
x cx xe .
rD ||
sx + v N
Inserting the normalized longitudinal slip defined in Eq. (15) and multiplying the differential
equation with the modified transport velocity rD || sx + vN yields
(dx rD || sx + f ) x e = f (vx rD ) cx xe rD || sx ,

(24)

where the artificial velocity vN in the first and last term was omitted because it is no longer
needed in this expression. The first order dynamic tire force model is completely characterized
by the generalized steady state tire characteristics f = f (s), and the stiffness cx and damping dx
properties of the tire.
At stand still the contact point velocity and angular velocity of the wheel will vanish, vx = 0
and = 0. Then, Eq. (24) simply results in x e = 0 which means that an existing longitudinal
deflection and hence an existing longitudinal tire force will be maintained. On hard braking
7

Rill Georg, Chucholowski Cornelius

maneuvers the wheel will lock = 0 while the vehicle is still moving vx 6= 0. In this case
Eq. (24) produces
f x e = f vx or x e = vx
(25)
which means that the tire deflection is increased or decreased as long as the vehicle is moving.
But, Eq. (25) takes no longer into account that the longitudinal tire force is limited to a maximum
value, |Fx | FxM . In order to achieve this Eq. (24) is supplemented by an appropriate if
statement
Fx = FxM sign(xe ) ;

if |cx xe + dx x e | > FxM and xe x e > 0 then

x e = 0 .

(26)

In normal driving situations the if statement will not be active at all. At a locked wheel Eq. (25)
together with the if statement in Eq. (26) describes a simple stick slip model.
The results of a driving and braking maneuver on a grade are shown in Fig. 6 for the simple
vehicle model. The vehicle (m = 400 kg, = 0.8 kgm2 , r = 0.3 m) starts from stand still
Velocities [m/s]

20
15
10
5
0
-5
-10

Vehicle position [m]

30
20

10
0
Driving and braking torque [kNm]

1
0.5

Driving

0
-0.5
-1
-1.5

Braking

8 10 12
Time [s]

14

16

18

20

4
3
2
1
0
-1
-2
-3

Longitudinal tire force [kN]

10 12
Time [s]

14

16

18 20

Figure 6: Driving and braking on a grade

on a grade ( = 15 ). At first it rolls backwards, v = r , then, the vehicle is accelerated

by a driving torque r > v. After a short period a braking torque is applied which causes the
wheel to lock in an instant, r = 0. A lockable braking torque model is presented in Ref. [14].
The maximum tire force is limited here to |Fx | < 3.1 kN . When the vehicle comes to a stand
still r = 0, v = 0 at t 6 s the braking torque model automatically changes the sign in
order to prevent the vehicle from moving downhill again. As the brake is not released yet the
vehicle oscillates some time in longitudinal direction. During this period where the wheel is
locked the system vehicle and tire represent a damped oscillator. The stiffness and damping
properties of the tire in longitudinal direction cx = 180 000 N/m and dx = 1500 N/(m/s)
together with the corresponding vehicle mass of m = 400 kg result in a frequency of f =
q
c/m (d/(2m)2 /(2pi) = 3.4 Hz. Inspecting the time history of the longitudinal force in
Fig. 6 one counts approximately 7 cycles in the interval 6s t 8s which yields f = 3.5 Hz,
and corresponds very well with the predicted result. Finally, the vehicle comes to a complete
stand still, v = 0, = 0. The tire force of Fx 1 kN which is needed to compensate the
8

Rill Georg, Chucholowski Cornelius

downhill force is perfectly maintained as long as the brake is applied. At t = 18 s the brake
is fully released and the vehicle starts to roll downhill again. It is shown in Ref. [14] that the
modified implicit Euler integration algorithm grants stability even at comparatively large step
sizes and produces results with sufficient accuracy. In this particular case the modified implicit
Euler algorithm is stable and sufficiently accurate up to a step size of h = 0.01 s. Hence, real
time simulations including hard braking, stand still and drive away maneuvers are possible with
this approach.
The modeling of the lateral tire force dynamics is straight forward. A dynamic model of the
bore torque which is needed for simulating the parking effort is described in Ref. [12].
In complex vehicle models the drive torque is generated by a separate engine model. It is then
transmitted by the drive train to the wheels. The software package ve-DYNA provides different
engine models and a generic model for a large variety of driving concepts. A sophisticated
drive train model can be found in Ref. [13]. The subsystem drive train interacts on one side
with the engine and on the other side with the wheels. The angular velocities of the wheels and
the engine or respectively the gear output angular velocity are inputs to the drive train model.
An implicit Euler integration generates smooth torques in the half shafts which act as driving
torques to the wheels.
2.2

Steering System

The steering system is coupled with the vehicle framework via the tire forces and torques.
Its modeling costs some effort. In addition to the nonlinear kinematics of the steering linkage,
dry friction and the dynamics of the power assistance have to be taken into account. Modern
steering systems include controlled overriding gears. At present even steer-by-wire systems are
under investigation.
Depending on the type of application the steering system can be modeled in different complexity, Fig. 7. The drag link, a part of the wheel suspension system, is included in the minimum
Drag
FW1

Steering box

link

iSB
uR

Stee

ring a
rm

TW1

Torsion bar

Power
assistance
TPA

SS

Steering shaft

BI

Force
coupling
FS2
FR2

TW2

Power
assistance
TPA

Torsion bar

TW1

2
MR2

Tyre 2
Tyre 1

SS

Wheel 1

SW
Steering Wheel

Steering shaft
Overriding
gear OG

Overriding
gear OG

SW

Tie rod

FW1

2
Wheel 2

Wheel 1

Steering box
iSB

uR

FS1

FW2

BI

1
Tyre 1

Force
coupling

Tie rod

Tyre 2
Wheel 2

Steering Wheel
TSI

TSI

Figure 7: Simple and enhanced model of a rack and pinion steering system

steering system model. Hence, the wheel axle kinematics cannot be modeled separately from
the kinematics of the steering linkage. This restricts the simple model to certain arrangements
of wheel/axle suspensions and steering linkages. The enhanced steering system model focuses
to the steering system itself. As the steering motions of the wheels are no longer included it
is possible to integrate third party models of steering systems into the simulation. Even hardware in the loop (HIL) applications are possible. The enhanced model of the steering system
becomes very stiff, because the coupling between the steering system and the steering motion
of the wheels is provided by the forces FS1 and FS2 generated by the compliance of the the drag

Rill Georg, Chucholowski Cornelius

links. However, real time performance can still be achieved by using a modified implicit Euler
integration algorithm, Ref. [10].
2.3

Suspension Systems

The automotive industry uses different kinds of wheel/axle suspension systems. Important
criteria are costs, space requirements, kinematic properties, and compliance attributes. Besides
the tire which was discussed in detail in a previous section vehicle suspension systems consist
of guiding elements, bushings and force elements.
The momentary position and the orientation of a reference frame R fixed to the wheel body
with respect to a body fixed reference frame B can be described by the position vector
h

rBR,B = x y z

iT

(27)

and the rotation matrix

ABR = A A A

(28)

where x, y, z are the components of the position vector expressed in reference frame B and the
rotation matrix was composed of 3 elementary rotations with the angles , around the x-,
y- and z-axis. Restricting oneself to the purely kinematic motion of a wheel the momentary
position and orientation of a wheel is determined by the hub motion z and in case of a steered
wheel for example by the displacement u of the steering rack. Hence, the kinematics of a wheel
suspension can be described by two-dimensional functions
x = x(z, u) ,
y = y(z, u) ,

= (z, u) ,
= (z, u) ,
= (z, u) .

and

(29)

Approximating this functions by two-dimensional lookup tables results in very fast computer
codes. Modern suspension systems however use compliances to improve the handling properties and the ride comfort of passenger cars, Ref. [15]. Hence, sophisticated suspension models
with several degrees of freedom must be used to take all these effects into account, Fig. 8. Here,
Topmount
zR

Spring

Body fixed
zB
reference
frame

Damper
Wheel body
fixed
reference
frame

yB

xR

xB
B

Drag link

yR

Rack
R

Wheel
and tire

Ball
joint

Wishbone

zS

Bushing

yS
xS

Bushing

S
Rubber
mounts

Subframe

Figure 8: Sophisticated suspension model

each wheel body and the subframe have 6 degrees of freedom. The rigid ball joint permits
10

Rill Georg, Chucholowski Cornelius

only 3 rotations of each wishbone relative to the wheel body. Hence, this axle suspension system consisting of two wheels, two wishbones and one subframe amounts to a total number
of f = 2 (6+3)) + 6 = 24 degrees of freedom. The forces in the drag links provide the coupling of the axle model with the model for the steering system. The software package ve-DYNA
provides the user with suspension models in different complexities. Due to a sophisticated modeling technique and an implicit numerical integration real time performance is possible even for
enhanced suspension models with up to f = 30 degrees of freedoms.
Leaf springs continue to be a popular choice for guiding and suspending solid axles. Though
simple in appearance, a leaf spring suspension causes many problems in modeling. Poor models
approximate guidance and suspension properties of the leaf spring by rigid links and separate
force elements, Ref. [4]. But, for realistic ride and handling simulations the deformation of the
leaf springs must be taken into account. Sophisticated beam-element leaf spring models have
a large number of degrees of freedom and result in extremely stiff differential equations which
will slow down the computing time significantly. For real time applications the leaf spring
must be modeled by a simpler but still accurate model, Fig. 9. The internal states of each leaf
FB1
3
1
zL
yL
2

sh

ac
kl

R
S

xL

Q
3

X
front
leaf eye
bushing

fo
ey rces
e b in
us fro
hin nt
gs

FB2

FS1

sh
for ack
ce le
s

Figure 9: Leaf spring suspension model

spring is calculated quasi-statically by minimizing the potential energy. Both leaf springs act
as generalized force elements which guide and support the axle by the forces acting in the front
eye bushings and in the shackles. Hence, the quasi-static five link leaf spring model can easily
be integrated into overall vehicle models. By suppressing high frequent leaf spring eigenmodes
it is perfectly adopted to real-time application. The model quality in kinematic and compliance
performance is proofed by a comparison to measurements, Ref. [8]. The five link leaf spring
model was integrated into a ve-DYNA Ford Transit vehicle model. Using this model at the rear
axle instead of a poor kinematic approach means only 85% more computer run time. Hence,
real time applications are still possible.
Usually the front and the rear axles are not connected by links. Hence, by splitting the states
of the vehicle frame work into states for the body, the front axles and the rear axles
h

yVT = yB , yF , yR

and zVT = zB , zF , zR

11

(30)

Rill Georg, Chucholowski Cornelius

the kinematic matrix KV and the mass matrix MV show a typical structure

KV

T
T
MBR
MBB MBF

0
= MBF MF F
.
MBR
0
MRR

KBB 0
0

IF F
0
= 0

0
0 IRR

and

MV

(31)

For the vehicle body a particular set of generalized speeds defined by KBB y B = zB can be
found in general which will simplify the equations of motion significantly, Ref. [3]. The trivial
choice of generalized speeds y F = zF and y R = zR is used for the axles which is indicated in
the kinematic matrix KV by matrices of identity IF F , IRR with appropriate dimensions. Due to
its simple structure the kinematic Matrix KV can be inverted analytically.
In addition front and rear axles will not be coupled by force elements. Then, the partial
derivatives of the generalized forces qV /yV and qV /zV show the same structure as the
mass matrix MV which keeps the block zero matrices in the modified mass matrix MVI which is
defined in Eq. (8). Taking these block zero matrices into account can speed up the decomposition of MVI significantly in particular when multi-axled vehicles are investigated or sophisticated
axle models with several degrees of freedom are used.
3
3.1

Tractor Trailer Coupling

Modeling Aspects

Vehicle combinations like passenger car with trailer, tractor semitrailer or tractor trailer may
be described by completely new multibody systems which applying Jourdains principle would
result in a set of first order differential equations of the same type as indicated in Eq. (1). But
then, the dimension n of the nn mass matrix MV which is given by the number of degrees
if freedom of the vehicle combination would be very large. For the numerical simulation the
mass matrix MV or according to the modified Euler formula in Eq. (8) the modified mass matrix
MVI must be decomposed. As the effort of a matrix decomposition increases with n3 real time
simulations will not be possible for large vehicle combinations.
A different modeling concept to vehicle combinations is shown in Ref. [11]. This approach
is based on a generic vehicle model which can easily be adopted to different types of single vehicles. Hence, passenger car and trailer, tractor semitrailer or tractor and trailer my be modeled
independently as vehicle 1 and vehicle 2, Fig. 10. Both vehicles are described by smaller sets

Tractor

Trailer

Vehicle 2

Vehicle 1
Coupling

Figure 10: Different model approaches to a tractor trailer combination

of first order differential equations. But in order to describe the vehicle combination these to
vehicle models must be coupled somehow.
12

Rill Georg, Chucholowski Cornelius

3.2

Force Coupling

If the coupling between vehicle and trailer is implemented by means of spring damper force
elements the equations of motions for the tractor (vehicle framework 1) and the trailer (vehicle
framework 2) will read as
KV 1 (yV 1 ) y V 1 = zV 1 ,
MV 1 (yV 1 ) zV 1 = qV 1 (yV 1 , zV 1 , . . .) + q12 (yV 1 , zV 1 , yV 2 , zV 2 ) ,

(32)
(33)

KV 2 (yV 2 ) y V 2 = zV 2 ,
MV 2 (yV 2 ) zV 2 = qV 2 (yV 2 , zV 2 , . . .) + q21 (yV 1 , zV 1 , yV 2 , zV 2 ) ,

(34)
(35)

where the dots in qV 1 and qV 2 indicate states of additional subsystems and dynamic force elements and q12 , q21 describe the force coupling. In order to avoid too large displacements in
the trailer drawbar force elements with a comparatively large stiffness must be used. Appropriate damping elements are needed in addition to avoid oscillations between the tractor and the
trailer. Due to the stiff force coupling explicit integration algorithms will require very small
step sizes to maintain stability. For implicit integration algorithms the generalized forces qV 1 ,
qV 2 and the coupling forces q12 , q21 must be evaluated at the new states. The first order Taylor
expansion of Eq. (7) can be applied directly to qV 1 and qV 2 because qV 1 does not depend on the
states of vehicle V 2 and qV 2 not on the states of vehicle V 1. However, the Taylor expansion of
the coupling forces results in


q12
q12
k
hKV11 zVk+1

z
1
V1 +
yV 1
zV 1


q12
q12
k
+
hKV12 zVk+1
2 zV 2 +
yV 2
zV 2



q21
q21
k
hKV11 zVk+1
1 zV 1 +
yV 1
zV 1


q21
q21
k
+
hKV12 zVk+1
2 zV 2 +
yV 2
zV 2

k+1
k
+
q12
q12

k+1
k
+
q21
q21

k
zVk+1
1 zV 1

(36)


k
zVk+1
2 zV 2 ,

k
zVk+1
1 zV 1

k
zVk+1
2 zV 2

.


(37)

k+1
Hence, the calculation or the approximation of the implicit coupling force q12
applied to vehik+1
cle V 1 will require the new velocity states zV 2 of vehicle V 2 and vice versa. That is why, the
implicit formula defined by Eq. (8) applied to both vehicles will couple the dynamic differential
equations (33) and (33) to one system (V = V 1 + V 2) where the implicit mass matrix is given
by

q12

zV 1

MVI

MV 1
=

0
MV 2

q21

zV 1

q12 1 q12 1
KV 1
K

yV 2 V 2
yV 1
h2

q22 1

q21
KV11
K
yV 1
yV 2 V 2

q12
zV 2

q22
zV 2

(38)

Certainly stability can be achieved for a very stiff force coupling with this implicit approach,
but the decomposition of one large matrix MVI instead of two smaller ones MVI 1 and MVI 2 will
be very costly. Using a softer force coupling and omitting the coupling terms is possible but
critical with respect to numerical stability and accuracy.
13

Rill Georg, Chucholowski Cornelius

3.3

Coupling via Constraint Equations

Describing the coupling between vehicle and trailer by constraint equations result in differentialalgebraic equations (DAE). Without losing generality only one constraint equation will be considered in the following. The corresponding DAEs for vehicle V 1 and V 2 are of index-3 and
will read as
KV 1 (yV 1 ) y V 1 = zV 1 ,

(39)
!T

MV 1 (yV 1 ) zV 1 = qV 1 (yV 1 , zV 1 , . . .) +

g
yV 1

KV11 ,

(40)

KV 2 (yV 2 ) y V 2 = zV 2 ,

(41)
!T

MV 2 (yV 2 ) zV 2 = qV 2 (yV 2 , zV 2 , . . .) +

g
yV 2

KV12 ,

(42)

0 = g(yV 1 , yV 2 ) ,

(43)

where the dots indicate the influence of additional states of subsystems or dynamic force elements, is the Lagrange multiplier and g defines the constraint equation. It is a common
practice to differentiate the constraint equation twice and to combine the new constraint on
acceleration level 0 = g with the dynamic differential Eqs. (40) and (42) to an index-1 DAE.
Then, the Lagrange multiplier can be eliminated and the resulting differential equations can
be solved by standard solvers for ordinary differential equations. However by using 0 = g the
original position constraint 0 = g is no longer satisfied exactly which will lead to a drift off
growing in time. In order to avoid this drift off Baumgarte has suggested to eliminate by
using the modified contraint equation
0 = g + 2 g + g ,

(44)

where not only the acceleration constraint, but also the constraints on velocity and position level
are considered via the stabilizing parameter > 0 and > 0. In Ref. [11] vehicle combinations were modeled by DAEs where two vehicles are coupled via algebraic constraints and the
Baumgarte stabilization technique is used to achieve a stable and sufficiently accurate solution.
The Baumgarte stabilization requires little computational effort and can be integrated very well
in existing simulation environments. But, the choice of suitable stabilization parameters and
is not straight forward and the error in the position constraint can not be predicted.
That is why an attempt was made to solve the index-3 DAE via the implicit Euler algorithm
directly. First results are very promising. Applying the implicit Euler formula defined in Eq. (8)
to Eqs. (39) to (42) will result in
zVk+1
= zVk 1 + h (zVu 1 + 1 ) ,
1

yVk+1
= yVk 1 + h KV11 zVk+1
1
1 ,

zVk+1
2

+ 2 ) ,

yVk+1
2

1
MVI j

zVk 2

+ h

(zVu 2

yVk 2

+ h

KV12

zVk+1
2

(45)

where the abbreviations

zVu j

1
MVI j qV j

j =

14

!T

g
KV1j ,
yV j

j = 1, 2

(46)

Rill Georg, Chucholowski Cornelius

were used and the mass matrices modified by implicit correction terms MVI 1 and MVI 2 are dek+1
fined according to Eq. (8). The new generalized speeds zVk+1
1 , zV 2 and in consequence the
k+1
position vectors yVk+1
1 , yV 2 depend on the yet unknown Lagrange multiplier . But of course
k+1
k+1
k+1
the new position vectors yVk+1
1 = yV 1 () and yV 2 = yV 2 () have to satisfy the constraint
equation (43) which in general will yield a nonlinear equation for


k+1
g yVk+1
1 () , yV 2 () = g() = 0 .

(47)

Applying the Newton algorithm will result in




g   i+1
g(i ) +
 i = 0 ,
i

i = 0, 1, . . .

(48)

The derivative of the constraint g with respect to the Lagrange multiplier is at first determined
by
g
g yV 1
g yV 2
=
+
,
(49)

yV 1
yV 2
where the partial derivatives g/yV 1 and g/yV 2 were already calculated in the dynamic
equations (40) and (42) and Eq. (45) yields finally
zVk+1
yV j
j
= h KV1j
= h KV1j h j = h2 KV1j j ,

j = 1, 2 .

(50)

The iteration in Eq. (48) may be started at t = 0 with 0 = 0 then, the value of the previous
time step 0 = (t h) can be used in the following.
3.4

Example

The concept of combining two separately modeled dynamic systems to a DAE can be demonstrated by an example. In Fig. 11 two simple pendulums are connected by a rigid rod of length
s. Following the modeling concept in section 3.3 the dynamics of two coupled pendulums may

b
2
a1
1

a2

s
m2

m1

Figure 11: Two pendulums connected by a rigid rod

be described by an index-3 DAE consisting of two times two first order differential equations
and one constraint equation
KV 1 y V 1 = zV 1

a1 1 = v1 ,

m1 v 1 = m1 g sin 1 +

!T

MV 1 zV 1 = qV 1 +

g
KV11
yV 1

15

g 1
,
1 a1

(51)

Rill Georg, Chucholowski Cornelius

KV 2 y V 2 = zV 2

a2 2 = v2 ,

m2 v 2 = m2 g sin 2 +

!T

MV 2 zV 2 = qV 2 +

g
KV12
yV 2

0 = g(yV 1 , yV 2 ) 0 =

g 1
,
2 a2

(52)

(b+a2 sin 2 a1 sin 1 )2 + (a2 cos 2 a1 cos 1 )2 s . (53)

{z

Here, the Lagrange multiplier represents the constraint force acting in the connecting rod.
The partial derivatives of the constraint function are given by

g
a1 
=
(a2 cos 2 a1 cos 1 ) sin 1 (b+a2 sin 2 a1 sin 1 ) cos 1 ,
1
`

(54)


g
a2 
=
(b+a2 sin 2 a1 sin 1 ) cos 2 (a2 cos 2 a1 cos 1 ) sin 2 .
(55)
2
`
The following Matlab-Skript solves the index-3 DAE for two connected pendulums with the
modified implicit Euler algorithm described by Eqs. (45) to (50).

% data pendulum 1, pendulum 2

datp1.grav = 9.81;
datp2.grav = 9.81; % Gravity [m/s]
datp1.mass = 15;
datp2.mass = 4;
% Masses [kg]
datp1.a
= 0.25;
datp2.a
= 0.30; % Lengths [m]
% data constraint
datc.b = 0.15;
datc.s
= 0.50; % Distances [m]
% Integration time and step size [s]
tE = 4.25;
h = 0.01;
% simple consistent initial values
time = 0;
% time
phi1 = -pi/2; phi2 = 0;
% Amplitudes pendulum 1, 2 [rad]
v1
= 0;
v2
= 0;
% Velocity pendulum 1, 2 [m/s]
la
= 0;
% Lagrange multiplier [N]
% Time loop
while time+h <= tE
time = time + h;
% constraint derivatives
[g,dgdy] = Constraint(2,[phi1,phi2],datp1.a,datp2.a,datc.b,datc.s);
% pendulum 1
[zp1_u, mu1_la, k1_inv] = p_dyn(phi1,v1,datp1,dgdy(1));
% pendulum 2
[zp2_u, mu2_la, k2_inv] = p_dyn(phi2,v2,datp2,dgdy(2));
% realize coupling via appropriate Lagrange multiplier
[la,g,iter] = Hook( phi1,v1,phi2,v2, la,dgdy(1),dgdy(2) ...
, zp1_u,mu1_la,k1_inv, zp2_u,mu2_la,k2_inv ...
, h, 10, 1.e-4 ,datp1, datp2, datc );
% update states
[phi1,v1] = p_step(phi1,v1, zp1_u, mu1_la, k1_inv, la, h );
[phi2,v2] = p_step(phi2,v2, zp2_u, mu2_la, k2_inv, la, h );
% plot amplitudes
plot(time/tE,phi1*180/pi,ob,time/tE,phi2*180/pi,xb)

16

Rill Georg, Chucholowski Cornelius

hold on, grid on

end

Here, the generalized forces applied to the unconstrained systems consist of non-stiff gravity
forces only. That is why no implicit correction terms in the mass matrices were needed. The
Matlab function
function ...
[g,dgdy]=Constraint(ijob,y,a1,a2,b,s)
phi1 = y(1); sphi1 = sin(phi1); cphi1 = cos(phi1);
phi2 = y(2); sphi2 = sin(phi2); cphi2 = cos(phi2);
x = b + a2*sphi2 - a1*sphi1 ; z = a2*cphi2 - a1*cphi1 ;
ell = sqrt( x2 + z2 );
g = ell - s;
if ijob ==2
dx1 = -a1*cphi1; dz1 = a1*sphi1;
dx2 = a2*cphi2; dz2 = -a1*sphi2;
dgdy = [ ( x*dx1 + z*dz1 ), ( x*dx2 + z*dz2 ) ] / ell ;
end

provides the constraint equation and optionally its partial derivatives. The Matlab functions
function ...
[zp_u,mu_la,k_inv] = p_dyn(y,z,dat,dgdy)
k_inv = 1/dat.a;
m_inv = 1/dat.mass;
q = -dat.mass*dat.grav*sin(y);
zp_u = m_inv*q;
mu_la = m_inv*dgdy*k_inv;

and
function ...
[yk1,zk1] = p_step( yk,zk, zp_u,mu_la,k_inv, la, h )
zk1 = zk + h * ( zp_u + la*mu_la ) ;
yk1 = yk + h * zk1*k_inv ;

define the pendulum dynamics and update the pendulum state by an Euler step. To judge the
results of this index-3 DAE solver the rod was also modeled by a spring damper element. In
order to avoid too large deformations and vibrations in the connecting rod a stiff spring with a
rate of c = 50 000 000 N/m and a damping constant of d = 15 000 N/(m/s) have to be used.
The resulting stiff ordinary differential equations (ODE) were then integrated with ode15s an
appropriate Matlab solver. Although a very nonlinear behavior is initiated, differences between
the simple index-3 DAE solver and the sophisticated ODE solver are hardly noticeable in the
time histories of the pendulum amplitudes and the force in the connecting rod. It turns out
that the value from the last time step serves as quite a good starting value for the Newton
iteration. At most two Newton steps were needed to limit the deformation of the connecting rod
to |4`| = |` s| = |g| < 0.1 mm. Here, this simple index-3 DAE solver is stable for step sizes
h 0.09 s but of course less accurate then. A larger step size increases the number of Newton
steps too. Up to 6 Iterations were needed for h = 0.09 s.
This concept allows to model complex systems very easily just by combining existing models. The additional modeling and programming effort is basically reduced two the constraint
equation. Hence, real time applications even for large vehicle systems like tractor trailer combinations will be possible. After providing the partial derivatives of the constraint function the two
17

Rill Georg, Chucholowski Cornelius

Amplitudes of pendulums

300

Number of Newton steps

[o]

200

100

3
2

0
-100

1
0.2

0.4

0.6

0.8

0
0

t/tE 1

Force in connecting rod

500

0.2

[N]

0.2

0.4

0.6

0.8

t/tE

Deformation of connecting rod

[mm]

400

0.1
300
200

100
-0.1
0
-100
0

0.2

0.4

0.6

0.8

-0.2
0

t/tE 1

0.2

0.4

0.6

0.8

Figure 12: Simulation results h = 0.01 s: DAE, stiff ODE

systems which are combined by constraint equations can be evaluated independently. Hence,
parallel processing is possible which will speed up the computation time significantly. A first
application of this method to a passenger car coupled by a tow bar with a second passenger car
was very satisfactory. No tricky stabilizing parameters are needed. Using an integration step
size of h = 1ms, which is very common in vehicle dynamics, the number of iterations needed
to calculate an appropriate Lagrange multiplier could even be restricted to one without loosing
significant accuracy.
4

CONCLUSIONS

The combination of different modeling techniques and a numerical integration formalism which
is adopted to the specific problems in vehicle dynamics make it possible to simulate large and
complex vehicle systems in real time. Vehicle combinations can be modeled quite easily by two
or more generic vehicle models joined by constraint equations. The resulting index-3 differential algebraic equations can be solved directly and very effectively by the modified implicit
Euler formalism.
REFERENCES
[1] G. Rill. Vehicle Dynamics in Real-Time Simulation. In: The Dynamics of Vehicles on
Road and on Tracks. Ed.: Apetaur, M., Lisse: Swets-Zeitlinger 1987
[2] G. Rill. Modeling and Dynamic Optimization of Heavy Agricultural Tractors. In: 26th
International Symposium on Automotive Technology and Automation (ISATA), Croydon:
Automotive Automation Limited 1993
[3] G. Rill. Simulation von Kraftfahrzeugen. Vieweg, Braunschweig 1994
18

Rill Georg, Chucholowski Cornelius

[4] W. Matschinsky. Radfuhrungen der Straenfahrzeuge. Springer: Berlin 1998.

[5] W. Hirschberg, G. Rill, H. Weinfurter. User-Appropriate Tyre-Modeling for Vehicle Dynamics in Standard and Limit Situations. Vehicle System Dynamics Vol. 38, No. 2, pp.
103-125, 2002
[6] J. Rauh. Virtual Development of Ride and Handling Characteristics for Advanced Passenger Cars. Vehicle System Dynamics Vol. 40, Nos. 1-3, pp. 135-155, 2003
[7] E. Pankiewicz and W. Rulka. From Off-Line to Real Time Simulations by Model Reduction and Modular Vehicle Modeling. In: Proceedings of the 19th Biennial Conference on
Mechanical Vibration and Noise Chicago, Illinois, 2003
[8] G. Rill, N. Kessing, O. Lange, J. Meier. Leaf Spring Modeling for Real Time Applications.
In: The Dynamics of Vehicles on Road and on Tracks - Extensive Summaries, IAVSD 03,
Atsugi, Kanagawa, Japan 2003
[9] P. Lugner, H. Pacejka and M. Plochl. Recent advances in tyre models and testing procedures. Vehicle System Dynamics, Vol. 43, No. 67, pp. 413-436, 2005
[10] G. Rill, C. Chucholowski. Modeling Concepts for Modern Steering Systems. ECCOMAS
Multibody Dynamics, Madrid, Spain, 2005
[11] B. Esterl, T. Butz, B. Simeon, B. Burgermeister. Real-Time Capable Vehicle-Trailer Coupling by Algorithms for Differential-Algebraic Equations Vehicle System Dynamics Vol.
00, No. 00, pp. 1-21, 2005
[12] G. Rill. First Order Tire Dynamics. In: Proceedings of the III European Conference on
Computational Mechanics Solids, Structures and Coupled Problems in Engineering. Lisbon, Portugal, 5-8 June, 2006
[13] G. Rill. Vehicle Modeling by Subsystems. J. of the Braz. Soc. of Mech. Sci.& Eng. - ABCM,
Vol. XXVIII, No. 4, pp. 431-443, 2006
[14] G. Rill. A Modified Implicit Euler Algorithm for Solving Vehicle Dynamic Equations.
Multibody System Dynamics, Vol. 15, Issue 1, pp. 1-24, 2006
[15] G. Rill. Simulation von Fahrwerksystemen. In: Wallentowitz, H.; Reif, K. (Ed.): Handbuch Kraftfahrzeugelektronik, Vieweg-Verlag, pp. 167-177, 2006
[16] G. Rill. Wheel Dynamics. Proceedings of the XII International Symposium on Dynamic
Problems of Mechanics (DINAME 2007), P.S.Varoto and M.A.Trindade (editors), ABCM,
2007

19