Institutionen för systemteknik Department of Electrical Engineering Examensarbete

Institutionen för systemteknik Department of Electrical Engineering Examensarbete
Institutionen för systemteknik
Department of Electrical Engineering
Examensarbete
Steering System Modelling for Heavy Duty Vehicles
Examensarbete utfört i Reglerteknik
vid Tekniska högskolan vid Linköpings universitet
av
Rickard Sjölund and Nicklas Vedin
LiTH-ISY-EX--15/4852--SE
Linköping 2015
Department of Electrical Engineering
Linköpings universitet
SE-581 83 Linköping, Sweden
Linköpings tekniska högskola
Linköpings universitet
581 83 Linköping
Steering System Modelling for Heavy Duty Vehicles
Examensarbete utfört i Reglerteknik
vid Tekniska högskolan vid Linköpings universitet
av
Rickard Sjölund and Nicklas Vedin
LiTH-ISY-EX--15/4852--SE
Handledare:
Gustav Lindmark
isy, Linköping University
Jonny Andersson
Scania
Examinator:
Martin Enqvist
isy, Linköping University
Linköping, 26 June 2015
Avdelning, Institution
Division, Department
Datum
Date
Avdelningen för Reglerteknik
Department of Electrical Engineering
SE-581 83 Linköping
2015-06-26
Språk
Language
Rapporttyp
Report category
ISBN
Svenska/Swedish
Licentiatavhandling
ISRN
Engelska/English
Examensarbete
C-uppsats
D-uppsats
—
LiTH-ISY-EX--15/4852--SE
Serietitel och serienummer
Title of series, numbering
Övrig rapport
ISSN
—
URL för elektronisk version
http://urn.kb.se/resolve?urn=urn:nbn:se:liu:diva-119770
Titel
Title
Modellering av styrsystem för tunga fordon
Författare
Author
Rickard Sjölund and Nicklas Vedin
Steering System Modelling for Heavy Duty Vehicles
Sammanfattning
Abstract
Framtidens tunga fordon kommer att designas och tillverkas med förbättrade förarstödsfunktioner, adas. Vid utveckling av adas förenklas utvecklingsprocessen av en modell över
fordonets dynamik. Ett system med stor påverkan på fordonets laterala dynamik och utvecklingen av adas är styrsystemet.
Syftet med detta examensarbete är att utveckla en tillförlitlig och simuleringseffektiv modell
av ett tungt fordons styrsystem. Insignalen i systemet är ett moment på ratten och den
resulterande hjulvinkeln betraktas som utsignal.
Med hjälp av bindningsgrafer och kända samband utvecklas fysikaliska modeller av systemets komponenter. Vissa komponenter modelleras med olinjära effektivitetsförluster och
friktion av olika komplexitet. Verktyg som linjär regression och olinjär sökning i rutnät
används för att skatta okända parametrar och funktioner utifrån mätdata. De olika delsystemen identifieras separat i den utsträckning det anses möjligt.
Olika modelleringsalternativ övervägs, valideras, och jämförs. Fördelar och nackdelar med
olika modelleringsval diskuteras. Slutligen väljs en olinjär tillståndsmodell för dess höga
precision och beräkningseffektivitet. Eftersom den slutgiltiga modellen kan användas för att
simulera ett tungt fordons styrsystem på en skrivbordsdator snabbare än realtid uppfyller
den sitt syfte.
Nyckelord
Keywords
Advanced Driver Assistance System, Steering System, Lane Keep Assist, Power Steering Gear,
Physical Modelling, Hydraulic Power Assist, Bond Graph
Sammanfattning
Framtidens tunga fordon kommer att designas och tillverkas med förbättrade förarstödsfunktioner, adas. Vid utveckling av adas förenklas utvecklingsprocessen av en modell över fordonets dynamik. Ett system med stor påverkan på fordonets laterala dynamik och utvecklingen av adas är styrsystemet.
Syftet med detta examensarbete är att utveckla en tillförlitlig och simuleringseffektiv modell av ett tungt fordons styrsystem. Insignalen i systemet är ett moment på ratten och den resulterande hjulvinkeln betraktas som utsignal.
Med hjälp av bindningsgrafer och kända samband utvecklas fysikaliska modeller
av systemets komponenter. Vissa komponenter modelleras med olinjära effektivitetsförluster och friktion av olika komplexitet. Verktyg som linjär regression och
olinjär sökning i rutnät används för att skatta okända parametrar och funktioner
utifrån mätdata. De olika delsystemen identifieras separat i den utsträckning det
anses möjligt.
Olika modelleringsalternativ övervägs, valideras, och jämförs. Fördelar och nackdelar med olika modelleringsval diskuteras. Slutligen väljs en olinjär tillståndsmodell för dess höga precision och beräkningseffektivitet. Eftersom den slutgiltiga
modellen kan användas för att simulera ett tungt fordons styrsystem på en skrivbordsdator snabbare än realtid uppfyller den sitt syfte.
iii
Abstract
Future heavy duty vehicles will be designed and manufactured with improved
Advanced Driver Assistance Systems, adas. When developing adas, an accurate
model of the vehicle dynamics greatly simplifies the development process. One
element integral to the vehicle lateral dynamics and development of adas is the
steering system.
This thesis aims to develop an accurate model of a heavy duty vehicle steering
system suitable for simulations. The input to the system is an input torque at the
steering wheel and the output is the wheel angle.
Physical models of the system components are developed using bond graphs and
known relations. Some components are modelled with non-linear inefficiencies
and friction of different complexity. Unknown parameters and functions are identified from measurement data using system identification tools such as, for example, linear regression and non-linear grid search. The different subsystems are
identified separately to the extent deemed possible.
Different model designs are considered, validated, and compared. The advantages and disadvantages of different model choices are discussed. Finally, a nonlinear state space model is selected for its high accuracy and efficiency. As this
final model can be used to simulate a heavy duty vehicle steering system on a
desktop computer faster than real time, it fulfills its purpose.
v
Acknowledgments
This work has been performed at the department Driver Assistance Controls at
Scania, where we want to thank Jonny and Daniel for their interest, support, and
commitment. Thanks to Christoffer, Joseph, Linus, and Markus for your guidance towards a solution. Our work would not have been possible without the
support of Emil, Jolle, and Krystof, thank you for your help providing us with
measurement data. Additionally we want to thank all other staff involved in
our work, as well as our department at Scania for an enjoyable experience. Many
thanks to Gustav for your hard work interpreting unclear drafts of the report and
contributing with input on how to improve the quality of our work. We would
also like to thank Martin for his inputs throughout the process, not least the insights he provided early in the planning phase leading to a more structured and
specific goal. Finally, we want to thank our friends and family for your support
throughout our university studies and an unforgettable experience.
Södertälje, June 2015
Rickard Sjölund and Nicklas Vedin
vii
Contents
Notation
xiii
1 Introduction
1.1 Background . . . . . . . . . . . . . . . . . . . . . .
1.2 Problem Formulation . . . . . . . . . . . . . . . . .
1.3 Aim of the Thesis . . . . . . . . . . . . . . . . . . .
1.4 Method . . . . . . . . . . . . . . . . . . . . . . . . .
1.5 Delimitations . . . . . . . . . . . . . . . . . . . . . .
1.6 Previous Work . . . . . . . . . . . . . . . . . . . . .
1.6.1 Previous Models of Mechanics . . . . . . . .
1.6.2 Previous Models of Hydraulics . . . . . . .
1.6.3 Previous Models of Friction and Hysteresis
1.6.4 Previous Models of External forces . . . . .
1.6.5 Practical Basis for the Modelling . . . . . .
1.7 Outline of the Report . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
1
2
2
3
3
3
4
4
5
5
5
2 System Description
2.1 Mechanics . . . . . . . . . .
2.1.1 Steering Column . .
2.1.2 Power Steering Gear
2.1.3 Steering Linkage . .
2.2 Hydraulics . . . . . . . . . .
2.3 External Behavior . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
7
7
7
8
9
10
10
3 General Modelling and Simulation Techniques
3.1 Model Types . . . . . . . . . . . . . . . . . .
3.2 Physical Modelling . . . . . . . . . . . . . .
3.2.1 Bond Graphs . . . . . . . . . . . . . .
3.3 System Identification . . . . . . . . . . . . .
3.4 Parameter Estimation . . . . . . . . . . . . .
3.5 Simulation and Solvers . . . . . . . . . . . .
3.5.1 Discretization of State Space Models
3.5.2 Differential Algebraic Equations . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
13
14
14
14
16
16
17
17
18
.
.
.
.
.
.
.
.
.
.
.
.
ix
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
x
Contents
3.5.3
Solvers in Simulink . . . . . . . . . . . . . . . . . . . . . . .
4 Steering System Mechanics
4.1 Fundamental Mechanical Principles . . . . . . . . . . . .
4.1.1 Moment of Inertia Model . . . . . . . . . . . . . . .
4.2 Steering Column . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1 Previously Developed Steering Column Models . .
4.2.2 Proposed Steering Column Model . . . . . . . . . .
4.3 Power Steering Gear . . . . . . . . . . . . . . . . . . . . . .
4.3.1 Previously Developed Power Steering Gear Models
4.3.2 Proposed Power Steering Gear Model . . . . . . . .
4.4 Steering Linkage . . . . . . . . . . . . . . . . . . . . . . . .
4.4.1 Previously Developed Steering Linkage Models . .
4.4.2 Proposed Steering Linkage Model . . . . . . . . . .
4.5 Combined State Space Model . . . . . . . . . . . . . . . .
4.6 Wheel Dynamics . . . . . . . . . . . . . . . . . . . . . . . .
19
.
.
.
.
.
.
.
.
.
.
.
.
.
21
21
22
22
23
23
25
25
25
29
29
30
31
33
5 Steering System Hydraulics
5.1 Hydraulic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Valve Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
37
37
41
6 Data Analysis and Parameter Estimation
6.1 Measurement Data . . . . . . . . . . . . . . . . . . .
6.1.1 Data Set 1 – Mechanical Efficiency . . . . . .
6.1.2 Data Set 2 – Valve Characteristics . . . . . . .
6.1.3 Data Set 3 – Hydraulic Efficiency . . . . . . .
6.1.4 Data Set 4 – Input Analysis . . . . . . . . . .
6.1.5 Data Set 5 – Reversible Torque . . . . . . . . .
6.1.6 Data Set 6 – Stationary Vehicle Measurements
6.2 Parameters and Functions . . . . . . . . . . . . . . .
6.2.1 Steering Column . . . . . . . . . . . . . . . .
6.3 Valve Characteristics . . . . . . . . . . . . . . . . . .
6.3.1 Valve Offset . . . . . . . . . . . . . . . . . . .
6.3.2 Valve Function . . . . . . . . . . . . . . . . . .
6.4 Hysteresis . . . . . . . . . . . . . . . . . . . . . . . .
6.4.1 Efficiency Constants . . . . . . . . . . . . . .
6.4.2 Adaptation of Hydraulic Valve Coefficients .
6.4.3 Hammerstein-Wiener Models . . . . . . . . .
6.5 Friction . . . . . . . . . . . . . . . . . . . . . . . . . .
6.5.1 Linear Friction Model . . . . . . . . . . . . . .
6.5.2 Non-linear Friction Models . . . . . . . . . .
6.5.3 Proposed Friction Model . . . . . . . . . . . .
6.6 External Influence . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
43
43
44
44
46
46
46
46
49
49
49
51
51
55
57
60
61
62
62
64
66
67
7 Results
7.1 Static Behaviour and Hysteresis . . . . . . . . . . . . . . . . . . . .
7.1.1 Static Gain . . . . . . . . . . . . . . . . . . . . . . . . . . . .
69
70
70
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
xi
Contents
7.1.2 Efficiency Constants . . . . . . . . . . . . . . . . . . .
7.1.3 Hydraulic Adaptation . . . . . . . . . . . . . . . . . .
7.1.4 Wiener Model . . . . . . . . . . . . . . . . . . . . . . .
7.1.5 Comparison of Hysteresis Models for Static Data . . .
7.2 Dynamic Behaviour and Full Vehicle Steering System Model
7.2.1 Comparison of Hysteresis Models for Dynamic Data .
7.2.2 Comparison of Models for the Jacking Torque . . . . .
7.2.3 Comparison of Friction Models . . . . . . . . . . . . .
7.2.4 Comparison of Solvers . . . . . . . . . . . . . . . . . .
7.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8 Summary
8.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . .
8.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . .
8.2.1 Suggestions for Model Accuracy Improvement
8.2.2 Model Expansion . . . . . . . . . . . . . . . . .
Bibliography
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
70
73
73
76
76
77
77
81
85
85
.
.
.
.
.
.
.
.
.
.
.
.
89
89
90
90
91
93
Notation
Acronyms
Acronym
adas
lka
ode
dae
hpas
mse
nrmse
dof
Meaning
Advanced Driver Assistance Systems
Lane Keep Assist
Ordinary differential equation
Differential Algebraic Equation
Hydraulic Power Assisted Steering
Mean Square Error
Normalized Root Mean Square Error
Degrees Of Freedom
xiii
xiv
Notation
Nomenclature Mechanics
Notation
bsc
eof f
Fdl
Fdl,j
Fg,p
Ftf ,n
Ftf ,r
Fz
fhyd
fl,p
fµ,p
g
hµ,g
hµ,kp
hµ,sc
hµ,wh
hsus
Jda
Meaning
Unit
Steering column damping coefficient
Lateral offset at the hub
Drag link force
Draglink force induced by jacking torque
Gravitational force acting on the piston
Force acting on the piston from the worm
Force acting on the piston from the sector
shaft
Front axle normal force
Hydraulic force
Piston loss force
Piston friction
Gravitational acceleration constant
Wheel to ground friction
King pin friction
Steering column friction
Wheel friction
Suspension height
Drop arm moment of inertia
[N ms/rad]
[m]
[N ]
[N ]
[N ]
[N ]
[N ]
Jlsc
Lower steering column moment of inertia
Jss
Sector shaft moment of inertia
Jsw
Steering wheel moment of inertia
Jusc
Upper steering column moment of inertia
Jwh
Wheel moment of inertia
Jwo
kdl
ksc
ktb
lda
ldla
mdl
meq
Worm moment of inertia
Drag link spring stiffness
Steering column spring stiffness
Torsion bar spring stiffness
Drop arm length
Drag link arm length
Drag link mass
Equivalent mass of the sector shaft, piston and
worm
Piston mass
Track rod mass
Worm to piston gear ratio
Sector shaft radius
Total aligning torque
Drop arm torque
Steering column string stiffness torque
Torque acting on the wheels from the drag
link
mp
mtr
n
r
Ta
Tda
Tk,sc
Tin,wh
[N ]
[N ]
[N ]
[N
h ] i
m/s2
[N m]
[N m]
[N m]
[N m]
[m]
h
i
kgm2 /rad
h
i
kgm2 /rad
h
i
kgm2 /rad
h
i
kgm2 /rad
h
i
kgm2 /rad
h
i
kgm2 /rad
h
i
kgm2 /rad
[N m/rad]
[N m/rad]
[N m/rad]
[m]
[m]
[kg]
[kg]
[kg]
[kg]
h
i
m−1
[m]
[N m]
[N m]
[N m]
[N m]
xv
Notation
Notation
Tis
Tj
TJwh
Tsa
Tsc
Ttf ,n
Ttf ,r
Tsw
Ttb
Tdla
vp
xp
δ
δwh,l
θK P I
θpsg
τl,is
ωda
ωlsc
ωss
ωusc
ωwh
ωwo
Meaning
Unit
Input shaft torque
Jacking torque
Torque acting on the moment of inertia of the
wheels
Self aligning torque
Steering column torque
Torque acting on the worm from the piston
Torque acting on the sector shaft from the piston
Steering wheel torque
Torsion bar torque
Drag link arm torque
Piston velocity
Piston position
Vehicle steering angle
Left wheel angle
Kingpin inclination angle
Mounted angle of the power steering gear
Input shaft loss torque
Drop arm angular velocity
Lower steering column angular velocity
Sector shaft angular velocity
Upper steering column angular velocity
Wheel angular velocity
Worm angular velocity
[N m]
[N m]
[N m]
[N m]
[N m]
[N m]
[N m]
[N m]
[N m]
[N m]
[m/s]
[m]
[rad]
[rad]
[rad]
[rad]
[N m]
[rad/s]
[rad/s]
[rad/s]
[rad/s]
[rad/s]
[rad/s]
xvi
Notation
Nomenclature Hydraulics
Notation
Ap
Meaning
Exposed piston area
Ap,L
Left side exposed piston area
Amin
Minimal valve area opening
Ap,R
Right side exposed piston area
Av
Valve area
A0
cq
k1
k3
Valve area opening with 0 input
Flow coefficient
Valve area parameter
Valve area parameter
pl
Load pressure
pp
Pump pressure
pp,min
Minimum pump pressure
pp,max
Maximum pump pressure
pL
Left side load pressure
pR
Right side load pressure
p0
Pump return pressure
Ttb0
Qf
Torsion bar torque that delivers the minimum
pump pressure
Flow of fluid
Ql
Qp
Unit
h i
m2
h i
m2
h i
m2
h i
m2
h i
m2
h i
m2
[−]
h[1/N m] i
1/(N m)3
i
h
N /m2
i
h
N /m2
i
h
N /m2
h
i
N /m2
h
i
N /m2
h
i
N /m2
i
h
N /m2
h
i
m3 /s
h
m3 /s
i
Flow of fluid over the piston
h
m3 /s
i
Pump flow
h
m3 /s
i
VL
Left side chamber volume
h
m3 /s
i
VR
Right side chamber volume
h
m3 /s
i
V0
φ
ψ
ρ
Chamber volume at valve equilibrium
Valve angle
Pressure drop over an orifice
Oil density
h
i
m3 /s
[rad]
[P
h a] i
kg/m3
1
Introduction
1.1
Background
Vehicle manufacturers and researchers across the world are developing autonomous
driving to improve traffic safety and increase fuel efficiency. Autonomous driving
requires Advanced Driver Assistance Systems (adas) that integrate both lateral
and longitudinal control of the vehicle. To develop adas with increased functionality and reliability, good information is needed about the vehicle dynamics
and how the vehicle is affected by different operations such as steering, applied
throttle, or brakes.
Scania is a manufacturer of heavy vehicles such as buses and trucks. For Scania to
be competitive in this market it is becoming increasingly important with driver
assistance systems, leading to low fuel consumption and traffic safety. Considering the importance of adas, contributing to this development is an interesting
challenge.
An example of an adas is Lane Keep Assist, lka. This is a system that helps
the driver stay in the intended lane. Vehicles can be equipped with sensors that
allow control units to detect where the lane is headed. By using this information
and a controller, the system can adjust the torque being applied to the steering
system to help the driver follow the lane and thereby increase comfort and avoid
collisions.
1.2
Problem Formulation
A good model of the vehicle’s lateral motion is a prerequisite for multiple adas
and fully autonomous vehicles. The problem treated in this thesis is designing an
1
2
1
Introduction
accurate vehicle model for the lateral dynamics of heavy duty vehicles. The focus
is on the physical steering system from an input torque on the steering wheel to
the output angle of the wheels.
1.3
Aim of the Thesis
As noted previously, development of driver assistance functionality requires extensive knowledge of vehicle dynamics; lka specifically requires insight into the
steering and lateral dynamics. This project aims to investigate the steering system of a heavy duty vehicle and to design mathematical models representing the
system. To maximize the usage of the knowledge, the modelling is performed
using varying methods and complexity. The resulting models are evaluated for
accuracy and computability. The final objective of the modelling process is developing an accurate model that can be simulated efficiently on a desktop computer.
1.4
Method
As the first part of the work process, previous work in the area is studied in order
to provide useful information and methods, as well as to learn about the type
of models previously developed. Both mathematical models as well as models
developed in the programs Simulink and Simscape are studied.
The different components that make up the steering system are modelled and
validated individually before they are merged to an integrated model. By first
developing individual subsystem models of different complexity, the importance
of each subsystem is assessed. A visual representation of the modelling methodology can be seen in Figure 1.1.
Throughout the development process, physical models are studied to gain a good
understanding of the equations that constitute the system. After establishing a
theoretical basis for the modelling, the system is divided into appropriate subsystems that can be modelled individually with suitable estimation and validation data. The model development is iterative for each subsystem; the subsystem
model is validated and evaluated before selecting one or several acceptable models that can be used in an integrated model. Finally, the result of the model
validation is used to compare the models.
Figure 1.1: The flow of the method used when designing models.
1.5
Delimitations
1.5
3
Delimitations
Due to the size and complexity of the system some delimitations are necessary.
All components cannot be modelled in full detail and complex components are
replaced with simplified representations – e.g. the dependence of the hydraulic
pump on engine speed is omitted.
Considering that the main objective of the thesis is modelling the steering system,
from the steering wheel to the wheels, little effort is put into modelling the load
from the truck and the road. To make use of measurement data from the total
truck system either a new simple simulation model or an existing simulation
model is used. These simulation models for the external dynamics provide the
steering model with a suitable input and use the modelled steering system to
simulate the vehicle’s motion.
As steering systems vary between truck models, one tractor configuration is chosen for this modelling assignment. The model design is performed with a generalizable and parameterizable implementation to as large extent as possible. It
is furthermore assumed that the lateral and the longitudinal motions are decoupled.
1.6
Previous Work
Obviously this is not the first model of a steering system developed, and this
thesis builds on previous work in both academic and corporate contexts. This
section briefly presents previous models that are referenced later in the text or
have provided inspiration. As this thesis aims to model the steering system of a
heavy duty vehicle, previous work in this area is the most relevant. Both Tagesson
[2014] and Dell’Amico [2013] give extensive descriptions of steering system for
heavy duty vehicles. Models for the same system are also developed in Mårtensson [2005], Dell’Amico and Krus [2015], and Tai et al. [2004].
Many steering system models have also been developed for light automotive vehicles. Even though light vehicle steering systems differ from heavy vehicle ditto,
the main functionality is the same. Consequently, much can be learned regarding,
for example, the steering column and friction by studying literature regarding
light vehicle steering systems. Light vehicle steering system studies used for this
purpose are Badawy et al. [1999], Chul Shin et al. [2013], Rösth [2007], Ancha
et al. [2007], Qiu et al. [2001], Pfeffer et al. [2008], Wu et al. [2015], and Viana
et al. [2010]
1.6.1
Previous Models of Mechanics
In Dell’Amico [2013] a review of different approaches to model the steering system mechanics is given. As stated in the review, the previously developed models
have between 2 and 5 degrees of freedom – i.e. between 2 and 5 moments of inertia or masses. For example Badawy et al. [1999] uses a complex 4 dof model
4
1
Introduction
while Tai et al. [2004], Dell’Amico and Krus [2015] and Chul Shin et al. [2013]
prefer simpler two moment of inertia models. Rösth [2007] develops one 2 dof
model for the steering system of a light automotive vehicle but chooses to add a
third dof to a model for the steering system of a heavy duty vehicle due to the
latter’s greater flexibility in the steering linkage.
1.6.2
Previous Models of Hydraulics
In addition to the mechanical connection of the steering system, heavy duty vehicles generally provide Hydraulic Power Assisted Steering, hpas, to reduce the
steering torque required from the driver. Manufacturers of hpas strive to construct power assist systems with a good trade-off between assistance and road feel
for the driver. Furthermore, the fuel consumption is affected by the hydraulics,
especially since the pump flow is commonly designed to be close to constant. To
further complicate the modelling of power assist systems, the valve characteristics are difficult to model and are kept secret by the manufacturers.
The active safety features as well as energy efficiency of the hpas and how these
can be improved are discussed in Rösth [2007]. The power steering is investigated
using both linear and non-linear modelling techniques. In order to model these
systems, the valve design and corresponding characteristics are explained. A
slightly more complex model is given in Dell’Amico and Krus [2015] and Dell’Amico
[2013] which considers the compressibility of the hydraulic fluid. The compressibility of the hydraulic fluid is also taken into account in Nilsson [2014] where
an electrohydraulic system at Scania is studied and modelled. Even though Nilsson [2014] models an open circuit hydraulic system, the methods can be used to
model closed circuit electrohydraulic actuator systems such as the hydraulic assist. A model of a hpas is also modelled using bond graphs in Viana et al. [2010].
Furthermore, a model for the hpas system of Scania trucks and suggestions for
future improvements is given in Mårtensson [2005].
1.6.3
Previous Models of Friction and Hysteresis
Another important phenomenon to model in the steering system is the friction.
In the 2 dof model developed in Dell’Amico and Krus [2015], viscous friction
is added to both the upstream mass and downstream mass. Dell’Amico and
Krus [2015] also adds static friction to the upstream mass and pressure dependent coulomb friction to the downstream mass. However, Dell’Amico and Krus
[2015] later neglects the coulomb friction when linearizing the model. In Qiu
et al. [2001] the downstream friction is modelled as a combination of viscous
and coulomb friction while Wu et al. [2015] approximates all friction in the steering system with one constant coulomb friction. Like Dell’Amico and Krus [2015],
Tagesson [2014] models viscous friction, however Tagesson [2014] argues that the
simple coulomb friction should be replaced by a more complex friction model developed in Pfeffer et al. [2008]. Just as the pressure dependent coulomb friction
suggested by Dell’Amico and Krus [2015], this model is developed to take the
hysteresis into account.
1.7
Outline of the Report
1.6.4
5
Previous Models of External forces
In addition to the driver input torque, the steering system is affected by external
forces from the tire-road interaction via the wheels. This self aligning torque is
decribed in Section 2.3 but models for these torques are developed in: Bastow
et al. [2004], Hsu [2009], Tagesson [2014], Wong [2008], and Wu et al. [2015].
1.6.5
Practical Basis for the Modelling
The models proposed in this thesis are influenced in parts by models previously
used at Scania. Primarily two models have constituted the foundation and source
of inspiration; the ADAMS model and the RT model. At Scania there is a model
covering the entire truck; the steering part of this model is henceforth referred
to as the ADAMS model. The ADAMS model is a physical model in three dimensions implemented using the multibody modelling language ADAMS. This
model is presented in Ikjema [1999].
Another simulation model that is currently used at Scania is the RT model. This
model is implemented in Simulink and includes dynamics for the entire truck.
However, the steering system in this model is represented by a static relationship including a look-up table. Although the RT model only models the steering
system schematically it provides some inspiration when modelling the load and
wheel dynamics.
1.7
Outline of the Report
This introductory chapter is followed by Chapter 2, where the reader is given
a basic understanding of the system that is modelled and how the components
interact. The reader is then introduced to the modelling theory in Chapter 3 that
is used throughout the thesis. A special focus is given to the distinction between
system identification and physical modelling.
After providing a general introduction, Chapter 4 contains modelling of mechanical components. In this chapter the mechanical connection in the steering system
is modelled based on physical principles.
To further explain the behavior of the power steering gear, Chapter 5 contains
modelling of the hydraulic system. The hydraulic power assist in the steering
system is modelled based on physical principles.
Before presenting the system identification methods used in Chapter 6, the measurement data analyzed in the thesis is presented. In the chapter, the measurement data is used to identify unknown parameters and relations in the physical
models from Chapter 4 and Chapter 5. Furthermore, model structures are suggested for unknown functions and the measurement data is used when estimating the parameters in the model structures.
Previously developed models are compared and validated in Chapter 7. Models
are compared with respect to accuracy and computability. The final section of
6
1
Introduction
the thesis, Chapter 8, contains discussions concerning the results and proposals
of future work.
2
System Description
This section provides a description of the steering system that is being modelled.
The components that make up the system transfer the input torque on the steering wheel to the wheels and provide power assist. An overview of the system can
be seen in Figure 2.1.
2.1
Mechanics
In this section the mechanical connection between steering wheel and wheels
is discussed. The mechanical connection consists of the steering column, the
mechanical connection of the power steering gear and the steering linkage.
2.1.1
Steering Column
The steering wheel is attached to the top of the steering column. The mechanism
of the steering column allows a mechanical torque to be transferred from the
steering wheel to the power steering gear.
Upper Steering Column
The upper steering column is directly attached to the steering wheel and is also
attached to the cab. The lower end is attached to a joint that allows the applied
torque to be transferred to the lower steering column. The upper steering column
has a certain degree of flexibility.
Lower Steering Column
Torque is transferred to the power steering gear through the lower steering column and hence the output torque of the lower steering column is the input torque
to the power steering gear. As the upper steering column is attached to the cab
7
8
2
System Description
Upper steering column
Lower steering
column
Power steering gear
Drop arm
Drag link
Drag link arm
Track rod
Figure 2.1: Steering system being modelled. Source: Scania
and the lower steering column is attached to the chassis some translational flexibility is needed; the cab and the chassis are not rigidly connected. This is accomplished by allowing the lower steering column to be elongated or compressed.
2.1.2
Power Steering Gear
In the power steering gear the torque is transferred from the steering column to
the steering linkage. As the name implies, this system also amplifies the torque
using a hydraulic system powered by the engine. Another important function of
the power steering gear is the gear ratio which is approximately 1:20; the gear
ratio depends non-linearly on the driver torque and steering angle. An overview
of the power steering gear is presented in Figure 2.2.
Following the steering column, the applied torque is transferred to the input
shaft of the power steering gear. The input shaft operates on a torsion bar that
transfers the torque to a worm; the torsion bar basically acts as a spring. The
worm is connected via a recirculating ball mechanism to a piston. The piston in
turn is connected to the sector shaft and drop arm of the steering linkage.
In addition to transferring the torque, the torsion bar acts as a rotor valve that
controls the support from the hydraulic servo. The hydraulic force is exerted on
2.1
9
Mechanics
the piston and is described more in Section 2.2. It is important that the steering
mechanism works even if the hydraulics fail. To fulfill this criteria, the steering
box can also be controlled mechanically without additional support from the hydraulics.
Oil feed
Valve
Oil
return
Worm
Torsion
bar
Input
shaft
Piston
Sector shaft
Figure 2.2: Power steering gear. Source: Robert Bosch Automotive Steering
GmbH
2.1.3
Steering Linkage
The steering linkage transfers the output torque from the power steering gear to
the steering knuckle rotating around the king pin of the wheels. The sector shaft
of the power steering gear is directly connected to the drop arm, which means
that these bodies can be modelled as a single body with a joint moment of inertia.
The torque that acts on these bodies is then transferred to a translational force on
the drag link.
Since the drag link is attached at the kingpin to the wheel through the drag link
arm which is attached to the steering knuckle, these bodies can be modelled as a
single body. The purpose of the drag link arm and steering knuckle is to convert
the translational force of the drag link to a rotational motion of the stub axle that
is connected to the wheel. The steering knuckle is also connected to the track rod
and the suspension. The track rod connects the two wheels by using ball joints
10
2
System Description
that are attached to their respective steering knuckles. As a result, the left and
right wheels have similar steering angles; there is an offset which increases with
the steering angle and is related to Ackermann Steering Geometry but this will
not be modelled in this thesis.
2.2
Hydraulics
In this section the hydraulic system in the power steering gear is described. The
system is depicted in Figure 2.2. The main purpose of the hydraulics is to apply
an additional force to the piston of the power steering gear to ease the lateral
handling of the vehicle. To achieve this, the power steering gear is required to
sense when a torque is applied. When a torque is applied, a lateral motion is
conducted. This can be sensed and controlled with a flow device, a valve located
between the torsion bar and the input shaft.
The hydraulics of the power steering gear consist of an oil reservoir with a filter,
a pump to move the oil at an approximately constant flow rate, as well as an
oil flow regulator. The studied hydraulic system is described in ZF Lenksysteme
[2015] and its components are modelled accordingly.
When a torque is applied at the input shaft of the power steering gear, the result
is an angular displacement of the valve. An enlarged image of the valve can be
seen in the upper left corner of Figure 2.2. This leads to the connection of ports
in the valve housing to channel high or low pressured oil. Inside the housing
of the steering box, this pressure difference that comes from the pump and is
controlled by the valve and the oil flow regulator affects the piston that moves
translationally. The pressure creates an additional force on the piston, in addition
to the mechanical force applied from the steering column, by the simple relation
that a pressure that acts on an area creates a force. There is a chamber on each
side of the piston to allow assisted pressure in both directions. When the torque
on the steering wheel ceases, the twisted torsion bar makes the control groves
of the valve return to their neutral position. The neutral position of the control
groves leads to an equal pressure being applied in both chambers on the side of
the piston.
The high and low pressures in the hydraulic system are the results of a pump
powered by the engine. To limit the influence of varying motor engine speeds the
hydraulic system uses the oil flow regulator to acquire a suitable flow rate. The
hydraulics are also built with a pressure release valve to prevent excessively high
pressures applied on the piston and a hydraulic steering limiter.
2.3
External Behavior
The lateral motion of a vehicle is influenced by several external parameters such
as wind, road bank, and self-aligning torque. In this thesis these influences are
not modelled in detail. However, in order to understand how these external
2.3
External Behavior
11
forces affect the expected steering output, some understanding is helpful. In
this section a rudimentary explanation is given for how the self-aligning torque
affects the vehicle.
Whenever there is a difference between the direction of travel of the vehicle and
the wheel plane, denominated a slip angle, a self-aligning torque acts on the
wheels and the steering system. This torque originates from the tire-road interaction and acts to decrease the slip angle by rotating the wheel plane towards the
direction of travel. The torque depends on a number of parameters such as the
weight of the load, the lateral stiffness of the tire, the tire contact length and the
velocity of the vehicle.
For example, assuming that a vehicle is on a track that has a constant turning
radius as well as a constant and high velocity. The vehicle is then affected by a
side force. The tires of the vehicle need to produce an opposing cornering force
to stay on the track. This cornering force results in a torque acting to rotate
the wheel plane towards the direction of travel. Furthermore, this self-aligning
torque is what makes the steering wheel return to its original position when driving straight at high velocity.
There is also a self-aligning torque resulting from the vertical displacement of
the truck at non-zero steering angles. This torque is a function of the caster angle
and king pin inclination. The caster angle is the forward rotation of the king pin,
the camber angle is the angle between a vertical plane and the wheel plane, and
the king pin inclination is the angle of the king pin relative to the vertical plane.
These three angles are shown in Figure 2.3.
To distinguish between the self-aligning torque caused by slip angles and selfaligning torques caused by the vertical displacements, the former is denominated
total aligning torque and the latter is denominated jacking torque.
12
2
Caster angle
System Description
King pin inclination
Camber angle
Vertical plane
Vertical plane
Vertical plane
Figure 2.3: Caster angle, camber angle and king pin inclination. The figure
is inspired by Bastow et al. [2004].
3
General Modelling and Simulation
Techniques
The input to the system described in Chapter 2, is a torque applied on the steering
wheel. The output is the angle of the wheels that results in a change of heading.
There are disturbances that affect the system output. Examples of such disturbances are wind, track geometry and component deviations.
In Ljung [1999] models are described as assumed relationships between different
variables. However, this thesis is mainly concerned with mathematical models
where relations between variables are stated using mathematical expressions. As
stated in Ljung [1999] models can fundamentally be developed in two different
ways: either by physical modelling or by system identification.
Physical modelling implies splitting the complete system into subsystems and
using previous experience to establish relationships between the variables. Previous experience may consist of physical laws and physical principles among other
established relationships. The model of the complete system is obtained by connecting the subsystem models using mathematical relations. Models developed
by modelling using laws of nature are called white box models.
System identification is another approach to building a model. In this case the
model is obtained by analyzing system inputs and outputs without considering
their physical significance. Using system identification, models can be developed
of systems that are not well-understood; these are called black box models.
Physical modelling and system identification are combined in grey box modelling. This is accomplished by first modelling using the laws of nature and
subsequently using system identification to identify parameters and model dynamics that are not well-understood. The resulting models are based on physical
underpinnings but some parameters and dynamics have no physical interpreta13
14
3
General Modelling and Simulation Techniques
tion.
3.1
Model Types
Mathematical models can be structured in different ways with different advantages and disadvantages. In this section state space models and more generally
differential algebraic equation models are described.
A differential algebraic equation (dae) is, as defined by Ljung and Glad [2004],
any equation on the general form
F(ż, z, u) = 0
y = h(z, u),
(3.1)
where F is a vector valued function. Please note that the equation y = h(z, u)
could have been included in F but is written separately for notational convenience. An important special case of dae:s is the state space model given by
ẋ = f (x, u)
.
y = h(x, u).
(3.2)
and is explained in Ljung and Glad [2004]. This model structure is well studied
and having equations on this form can greatly facilitate analysis and simulation.
For a general dae the index is given by the number of times the dae has to be
differentiated to arrive at a state space model.
3.2
Physical Modelling
As stated above, modelling is often performed using the physical principles that
govern the system. This approach is viable whenever the physical properties of a
particular subsystem are known. As even the physical laws are models and some
simplifications have to be made, the physical models are not correct representations of the real system(Ljung and Glad [2004]).
An advantage of physical modelling compared to system identification is that
a white box model can be adapted to new similar systems by adapting the parameters accordingly. For example, a physical model of a steering system can be
adapted to a new steering system with a heavier steering column by just changing
the mass parameter whereas a new black box model would have to be estimated.
3.2.1
Bond Graphs
Bond graphs are useful tools when performing systematic modelling. Bond graphs
describe the flow of energy between components by visualizing the efforts and
flow streams of the system.
3.2
15
Physical Modelling
A flow of energy is represented by an arrow that indicates the direction of the
energy flow. In this direction, the product of the effort, e, and flow, f , is positive.
The side with the causal stroke (the right side in Figure 3.1) gives the direction for
the numerical calculation for an effort variable. If all causal strokes except one
are facing a serial junction, the intensity for the last one can also be calculated.
The reversed relation applies to a parallel junction, if all causal strokes except one
are facing away from the junction, the flow of the last one can be calculated(Ljung
and Glad [2004]).
e
f
Figure 3.1: The energy flow representation in a bond graph.
The bonds of the system can be connected to elements; there are capacitive, inductive, and resistive elements as well as gyrators and transformers. A system may
exchange energy with the surroundings. The surrounding energy is then referred
to as a source, S. A source can be either based on the effort or flow depending on
which unit is specified.
A system can consist of several connection points,which are either serial or parallel. A serial point has the property that the sum of all ingoing efforts subtracted
by the sum of all outgoing efforts into the connection point equals zero, which
can be written
ein,1 + ein,2 + · · · + ein,i − eout,1 − eout,2 − · · · − eout,j = 0.
(3.3)
Furthermore, all contiguous flows are equal. This can be written
f1 = f2 = · · · = fn .
(3.4)
Analagously, a parallel connection point uses the reversed definitions, where all
attached efforts are equal
e1 = e2 = · · · = en
(3.5)
and the sum of all ingoing flows subtracted by the sum of all outgoing flows equal
zero,
fin,1 + fin,2 + · · · + fin,i − fout,1 − fout,2 − · · · − fout,j = 0.
(3.6)
Causality markings are used to specify the order of calculations. To obtain a suitable and solvable calculation, a serial connection point should have one outgoing
effort and the remaining efforts should be incoming. The reversed computational
order should be used for parallel connection points. For a system with consistent
16
3
General Modelling and Simulation Techniques
causality markings, a state space model is easily found by choosing the flow variables of the inductive elements and effort variables of the capacitive elements as
states.
The steering system model is based on hydraulic and mechanical representations.
In the rotational mechanical domain the effort is a torque and the flow is an angular velocity. In the translational mechanical domain the efforts are forces and
the flows are velocities. The hydraulics are described with the pressure as effort
and the hydraulic flow as flow.
3.3
System Identification
According to Ljung [1999], three components are needed to perform system identification. The first part is analyzable data recorded with a suitable experimentation setup to capture the system behavior. The second component is a set of
candidate model structures to choose from. Given a model structure, a candidate
model is obtained by estimating its parameters from the data. The last component of the system identification process are rules by which candidate models are
evaluated. Following the development of a model using the three components
described above, the model should be validated to measure how well it explains
the measured signals.
3.4
Parameter Estimation
When a model structure is selected, either by physical modelling or inspecting
the data, the unknown parameters, θ, need to be determined. Ljung and Glad
[2004] suggests a criteria that is related to the prediction errors. At the time t − 1
the model provides a prediction
ŷ(t|t − 1, θ)
(3.7)
of the output at time t. Using this notation the prediction error, e is
e(t, θ) = y(t) − ŷ(t|t − 1, θ).
(3.8)
A natural intent of the modelling can be to minimize the mean square error,
VN (θ) =
N
1 X 2
e (t, θ).
N
(3.9)
t=1
Using this criteria, the optimal cost is given by the parameter vector θ that best
matches the measurement is obtained by minimizing VN , i.e.
3.5
17
Simulation and Solvers
θ̂N = arg min VN (θ).
θ
(3.10)
For a linear regression model ŷ(t|t − 1, θ) = θ T φ(t), where φ(t) is an arbitrary
function independent of θ, minimizing the loss function, VN has an analytical
solution given by the least square solution to


y(t1 ) = θ T φ(t1 )



T


 y(t2 ) = θ φ(t1 )
.

..



.


 y(t ) = θ T φ(t )
N
N
(3.11)
For many non-linear model structures (ŷ(t|t − 1, θ) = h(t, θ)) where the loss function is a complex function of θ, it is much more difficult to find an analytical
solution. In these circumstances the parameter estimation has to be performed
numerically. A general numerical approach is described in Gustafsson [2012] as
θ (i+1) = θ (i) + α (i) f (i) ,
(3.12)
where α is a step length and f is the search direction. Basing the search direction
on a first order Taylor approximation gives the Gauss-Newton algorithm defined
by the recursion
θ (i+1) = θ (i) + α (i) (J(θ (i) )J T (θ (i) ))−1 J(θ (i) )(ȳ − h̄(θ (i) )).
(3.13)
where J(x) is the Jacobian of the prediction errors, ē, with respect to θ and ȳ
is a vector containing all N measurements. For a linear regression model this
recursion converges in the first iteration to the least squares estimate (with α = 1).
3.5
Simulation and Solvers
One important usage of models is simulation, where outputs and states of the
model are determined based on initial conditions and known inputs. The method
used to obtain an approximate solution depends on the type of system. In this
section a brief description of solver types is given for state space and dae models
based on Ljung and Glad [2004].
3.5.1
Discretization of State Space Models
To simulate a continuous state space model at f different time points t1 , ..., tf ,
the time-derivative of the state ẋ(t) has to be numerically approximated using
18
3
General Modelling and Simulation Techniques
previous values of x(t). The simplest difference approximation is Euler’s method,
in which
ẋ(tn ) ≈
xn+1 − xn
,
h
h = tn+1 − tn .
(3.14)
Euler’s method is generally not the most efficient one. For example, like Euler’s
method, Runge-Kutta methods use more sophisticated approximations to arrive
at the next state, improving accuracy at the expense of more calculations. More
generally, approximations of the next state can depend on any number of previous states according to
xn+1 = G(t, xn−k+1 , xn−k+2 , . . . , xn , xn+1 ),
(3.15)
where k is the number of states used.
An important distinction is between explicit and implicit methods. In the former,
the right hand side of (3.15) does not depend on xn+1 . This implies that xn+1 is
found directly by evaluating the expression. In the latter, a system of equations
must be solved to obtain xn+1 . Although implicit methods require more computations, they have a greater area of stability than explicit methods with the same
k. Examples of multi-step methods (k > 1) are the explicit and implicit Adams
methods.
In general it is inefficient to keep the step length h = tn+1 − tn constant during simulations. In many cases, the dynamics shift between fast and slow. Methods determining step length automatically usually depend on the local error estimate.
Stiff differential equations are characterized by widely spread time constants in
the same system. To efficiently simulate stiff systems with both slow and fast dynamics, both large and small step lengths must to be managed by the solver. This
is because the maximum step length is limited by stability requirements from
the fast dynamics. Consequently, simulating stiff differential equations requires a
trade-off between accuracy and simulation speed. Implicit methods are preferred
for solvers used for stiff differential equations as they have greater stability areas.
3.5.2
Differential Algebraic Equations
To simulate a general dae, one method is to approximate the derivative by using
Euler’s method. This gives
F( 1h (zn − zn−1 ), zn , u(tn )) = 0,
.
h = tn − tn−1
(3.16)
Subsequently the equation can be solved for zn with the knowledge of zn−1 ; this
process is then iterated for each zk , k ∈ {1, ..., N }.
Obviously the accuracy of this simulation method is limited by the accuracy of
3.5
Simulation and Solvers
19
the difference approximation. To improve the difference approximation, more
previous values zn−k can be used. This is called using a backward differentiation
formula. Using these methods daes with index one and lower can be solved.
Solving higher index daes requires other more sophisticated methods.
3.5.3
Solvers in Simulink
Several solvers are implemented in the MATLAB & Simulink software. Almost
arbitrary accuracy can be achieved with any solver by reducing the step size at
the expense of more computations. Therefore these solvers are useful for different
purposes and model types. The initial choice when choosing solver is deciding
between fixed and variable step solvers. The latter type of solver is often faster
than the former in accordance with the reasoning in Section 3.5.1, particularly
for stiff problems. However, code generation from a MATLAB model is more
complicated when using a variable step solver. The fixed and variable step solvers
in Simulink are summarized in Table 3.1 and 3.2, respectively.
For non-stiff problems the ode45 is generally the fastest and most accurate solver.
However, for stiff problems implicit step solvers are recommended in accordance
with Section 3.5.1. Primarily, ode15s is recommended for normal stiff problems.
For dae problems with index < 2, a methodology similar to the backward differentiation formula described in Section 3.5.2 is implemented in both ode15s
and ode23s. The numerical differentiation formulas used in ode15s are similar
to but more efficient than the backward differentiation formula described in Section 3.5.2.
20
3
General Modelling and Simulation Techniques
Table 3.1: Fixed step solvers in Simulink.
Solver
Description
Order
Explicit solvers
ode1
ode2
ode3
ode4
ode5
ode8
Euler’s Method
Heun’s Method
Bogacki-Shampine Formula
Fourth-Order Runge-Kutta (RK4) Formula
Dormand-Prince (RK5) Formula
Dormand-Prince RK8(7) Formula
First
Second
Third
Fourth
Fifth
Eighth
Implicit solvers
ode14x
Newton’s method and extrapolation
Variable
Table 3.2: Variable step solvers in Simulink.
Solver
Description
Order of accuracy
Explicit solvers
ode45
ode23
ode113
Runge-Kutta, Dormand-Prince (4,5)
pair
Runge-Kutta (2,3) pair of Bogacki &
Shampine
PECE Implementation of AdamsBashforth-Moutlon
Medium
Low
Variable, Low to High
Implicit solvers
ode15s
ode23s
ode23t
ode23tb
Numerical Differentiation Formulas
Fourth
(NDFs)
Second-order, modified Rosenbrock forFifth
mula
Trapezoidal rule using a "free" interEighth
polant
TR-BDF2
Variable, Low to Medium
4
Steering System Mechanics
The mechanical connection between steering wheel and wheels is discussed in
this section. First the physical laws governing the mechanics are summarized.
Then a one-dimensional modelling approach is described where the subsystems
are derived separately before an integrated model of the steering system mechanics is presented.
4.1
Fundamental Mechanical Principles
According to Newton’s Second Law of physics, a force F on a body is given by the
mass m multiplied by its acceleration a, i.e.,
F = ma.
(4.1)
The rotational equivalent of (4.1) is a torque, T , affecting a body given by the
moment of inertia, J, multiplied by the angular acceleration, α, resulting in
T = J α.
(4.2)
Since the angular acceleration is the derivative of the angular velocity, α = ω̇,
(4.2) gives
21
22
4
1
ω(t) =
J
Steering System Mechanics
Zt
T (s)ds.
(4.3)
0
For a torsional spring, the torsion is proportional, with constant k to the torque,
Tspring . The derivative of this relationship results in
ω(t) =
1
Ṫ
(t).
k spring
(4.4)
A rotational friction torque, Tf riction acting on the body is typically a nonlinear
function of ω,
Tf riction (t) = h(ω(t)).
4.1.1
(4.5)
Moment of Inertia Model
The steering system in a truck moves in three dimensions with a moment of inertia that can be written
J3D

Jxx

= Jxy
J
xz
Jxy
Jyy
Jyz

Jxz 

Jyz  .
Jzz 
(4.6)
However, in attempt to reduce the complexity of the model, a one dimensional
representation of the dynamics is considered. This representation only allows
each body to rotate around a given axis, commonly the z-axis, simplifying the
moment of inertia to
J1D = Jzz .
(4.7)
As the lateral dynamics being modelled concern a one-dimensional input torque
to a wheel rotation in one dimension this is a reasonable simplification.
4.2
Steering Column
As described in Section 2, the steering column consists of a steering wheel, an
upper steering column and a lower steering column that are attached with joints.
4.2
23
Steering Column
4.2.1
Previously Developed Steering Column Models
One way of modelling the upper steering column is to consider the three bodies
and their corresponding moments of inertia. The torque applied to the steering wheel acts on the inertia through the steering column and is then applied
to the power steering gear. The three-dimensional model developed at Scania
described in Ikjema [1999] follows this approach. However, the same model constrains the relative motion of the bodies so that most torque from the driver is
transferred directly to the input shaft of the power steering gear. Consequently
the one-dimensional equivalent is one inertia directly connected to the torsion
bar. In Tai et al. [2004] this one inertia model is used.
The bodies constituting the steering column are not completely rigid but these dynamics are ignored in the model proposed by Tai et al. [2004]. Ancha et al. [2007]
and Mårtensson [2005] use an approach where the steering column model is split
into two bodies with separated moments of inertia, with springs and dampers inserted between the bodies.
4.2.2
Proposed Steering Column Model
The moment of inertia for the steering wheel, Jsw , upper steering column, Jusc and
lower steering column, Jlsc can separately be found by making the approximation
that the shape of the bodies is close to that of a cylinder. The three-dimensional
moment of inertia of a cylinder is
J3D
 m(3r 2 +h2 )
 12

= 
0

0
0
m(3r 2 +h2 )
12
0

0 

0  ,

mr 2 
(4.8)
2
where m is the mass of the cylinder, r is the radius and the height is given by h.
For the one-dimensional case considered here, (4.8) simplifies to
J1D =
mr 2
.
2
(4.9)
There is friction between the steering column and its connection to the surrounding truck. This is modelled with a damper connecting one of the masses to
ground. The rotational friction is typically non-linear. The full model of the
three connected moments of inertia is shown in Figure 4.1.
Mathematical Description
The mathematical model of the steering column is developed for the full system
consisting of two bodies, i.e. upper and lower steering column. As this is the most
complex model of the steering column, it is easy to also find the one inertia model
by eliminating the spring between the upper and lower steering column. Figure
4.2 contains a bond graph of the full steering column model. The causality of the
24
4
Steering System Mechanics
Tsw
Jsw + Jusc
Jlsc
Figure 4.1: The steering column representation.
bond graph is consistent so the states can be selected as the flow variables for the
moments of inertia and the effort variable for the springs.
R : bsc
C : ksc
Se
Tsw
I : Jsw + Jusc
s
R : hµ,sc
I : Jlsc
s
ωusc
p
s
ωlsc
Tis
Tsc
Se
Se
τl,is
Figure 4.2: Bond graph of the steering column dynamics from the input
torque of the steering wheel to the output torque at the output interface of
the lower steering column.
Consequently, the first state is the angular velocity of the steering wheel and
upper steering column, ωusc . This state is calculated from the total moment of
inertia of the steering wheel and the upper steering column, Jsw + Jusc , the input
torque at the steering wheel, Tsw and the torque at the joint with the lower part
of the steering column, Tsc ,
ω̇usc =
1
Tsw − Tsc .
Jsw + Jusc
(4.10)
The second state is the torque of the spring in the joint, Tk,sc . The state equation
4.3
25
Power Steering Gear
is
Ṫk,sc = ksc ωusc − ωlsc ,
(4.11)
where ksc is the torsional stiffness of the spring and ωlsc is the angular velocity of
the lower steering column. The steering column torque, Tsc is equal to the torque
of the spring in the joint and the friction torque given by the viscous constant bsc .
Consequently (4.10) can be rewritten as
ω̇usc =
1
Tsw − Tk,sc − bsc (ωusc − ωlsc ) .
Jsw + Jusc
(4.12)
The last state in the steering column model is the angular velocity of the lower
steering column, ωlsc , determined by
ω̇lsc =
1 Jlsc
Tk,sc + bsc (ωusc − ωlsc ) − hµ,sc (ωlsc ) − Tis − τl,is .
(4.13)
The state equation is dependent on the torque acting on the input shaft of the
power steering gear, Tis , its associated loss torque, τl,is and a non-linear friction,
hµ,sc dependent on the angular velocity of the lower steering column. The loss
torque will be described in Chapter 6 and emanates from mechanical deformations and inefficiencies.
4.3
4.3.1
Power Steering Gear
Previously Developed Power Steering Gear Models
In the three-dimensional model developed in Ikjema [1999] the four inductive elements of the power steering gear are modelled separately; the torsion bar, worm,
piston and sector shaft are modelled as four inertias or masses. However, in the
one-dimensional case the torque bar and the lower steering column can be modelled as one body. Furthermore, the motions of the worm, the piston and the sector shaft are constrained which means they can be replaced with an equivalent
inertia and a gear constraint. This simplification is done in Mårtensson [2005]. In
Tai et al. [2004] the dynamics of the power steering gear are neglected completely;
the rotational torque entering the power steering gear is directly transformed via
a gear constraint to the sector shaft.
4.3.2
Proposed Power Steering Gear Model
The angular displacement, between the steering column and the worm, caused
by the torque bar deformation is modelled as a spring between the two bodies.
This modelling alternative is used in both Ancha et al. [2007] and Mårtensson
[2005]. Friction in the power steering gear is modeled with a friction force acting
on the piston. In accordance with the reasoning regarding the inductive elements,
26
4
Steering System Mechanics
resistive elements spread out over the piston, worm, and sector shaft can be replaced by one equivalent resistive element at the piston. There are considerable
efficiency losses in the power steering gear that are modelled as asymmetries and
hysteresis.
Mathematical Description
In this section a model including four inertias is developed. The model is then
simplified to achieve causality throughout the bond graph. The efforts and flows
of the mechanical forces are visualized by the bond graph in Figure 4.3. The input
of the bond graph is a torque at the input shaft, transferred to the torsion bar. The
flexibility of the torsion bar is modelled as a spring constant, ktb . Following
Ṫtb = ktb (ωlsc − ωwo ),
(4.14)
the derivative of the torque that acts on the torsion bar, Ttb depends on the angular velocity of the lower steering column, ωlsc and the angular velocity of the
worm, ωwo . The torsion bar torque is thus selected as a state.
Se
Tis
I : Jtb
C : ktb
I : Jwo
s
p
s
ωwo
Ttb
I : mp
`n−
TF
s
Se
Se
fl,p
fhyd − Fg,p
I : Jss
−ra
TF
s
ωss
Se
Tda
R : fµ,p
Figure 4.3: Bond graph of the power steering gear from the input torque of
the valve slide to the output torque at the sector shaft.
The angular velocity of the worm is modelled according to
ω̇wo =
1
(T − Ttf ,n ),
Jwo tb
(4.15)
where Jwo is the moment of inertia of the worm and Ttf ,n is the torque of the
transformer. The angular velocity of the worm is selected as a candidate state.
This angular velocity is transformed in the recirculating ball mechanism to the
piston, causing it to move in a translational motion with the piston velocity, vp .
This is modelled as a transformer with n as the gear ratio in
vp = nωwo .
(4.16)
4.3
27
Power Steering Gear
The corresponding translational force that causes the piston to move, Ftf ,n , can
also be found through the torque of the worm using
Ttf ,n = nFtf ,n .
(4.17)
The piston velocity is another direct candidate to use as a state since its value
can be calculated from the piston mass, mp , the friction force, fµ,p , the applied
hydraulics force, fhyd and the forces applied by the adjacent transformers, Ftf ,n
and Ftf ,r . An extra force fl,p is subtracted from the force at the piston to account
for the inefficiency of the power steering gear, described in Section 6. The power
steering gear is mounted with the piston moving in an inclined plane with angle
θpsg , and it is thus affected by a gravitational force, Fg,p . The piston acceleration
is
v̇p =
1
(F
+ fhyd − Fg,p − Ftf ,r − fµ,p − fl,p ).
mp tf ,n
(4.18)
The interface between the piston and the sector shaft is modelled as a rack and
pinion with radius r and acts as a transformer,
Ttf ,r = rFtf ,r .
(4.19)
The effort is transferred from a force connected to the piston, Ftf ,r to a torque on
the sector shaft, Ttf ,r . The corresponding transfer of flow,
vp = rwss
(4.20)
allows the velocity of the piston to be transferred to the angular velocity of the
sector shaft, wss .
The last body of the power steering gear to be modelled is the sector shaft. Its
angular velocity is calculated from its moment of inertia, Jss and the difference in
torque between the incoming torque from the transformer and the torque at the
droparm, Tda according to
ω̇ss =
1
(T
− Tda ).
Jss tf ,r
(4.21)
Causality Conflicts and Model Adjustments
It can be observed in Figure 4.3 that the bond graph of the power steering gear
has causality conflicts at the piston and at the sector shaft. The serial junction at
28
4
Steering System Mechanics
the piston has four causal strokes facing the junction and two causal strokes facing away from the junction. In this case, the conflict originates from the junction
having two bonds trying to determine the effort. Likewise, the serial junction
at the sector shaft has two bonds trying to determine the effort. These conflicts
imply that a state space model cannot be derived directly from the graph. In
addition to the state space equations, the static relationships from the transformers need to be included in the model. Consequently, the mathematical equations
from the bond graph results in a general dae instead of a state space model. This
problem can be remedied by eliminating moments of inertia and replacing them
by equivalent masses. Thereby the piston, worm, and sector shaft is replaced by
one equivalent mass.
In order to merge the inductive elements, the moment of inertia of the sector
shaft must be rescaled according to the transformation (4.19). Rewriting (4.18)
as
Ftf ,r = Ftf ,n + fhyd − Fg,p − fµ,p − fl,p − mp v̇p
(4.22)
in combination with rewriting (4.21) using (4.20) as
Ttf ,r = Jss ω̇ss + Tda = Jss
v̇p
r
+ Tda
(4.23)
results in
v̇
v̇p ( Jrss2
Jss rp + Tda
+ mp ) + Trda
= r(Ftf ,n + fhyd − Fg,p − fµ,p − fl,p − mp v̇p ) ⇐⇒
= Ftf ,n + fhyd − Fg,p − fµ,p − fl,p .
(4.24)
The relation has been produced by partially removing the dae through the removal of one equation. This method is possible since the two bodies, the piston
and the sector shaft are assumed to have their motions constrained one with respect to the other.
Figure 4.4 shows the bond graph of the power steering gear when the piston and
sector shaft are modelled as one equivalent body through the implementation
of (4.24). There is still a causality conflict between the transformers. It can be
solved by once again eliminating an equation and transforming the general dae
into a state space model. In this case the worm is moved through the remaining
transformer using (4.17) as a base and the input torque and output force of the
equation need to be found.
The input torque can be found with the relation that the sum of the torques at
the worm node equals zero and by using (4.16),
4.4
29
Steering Linkage
Se
I : Jtb
C : ktb
I : Jwo
s
p
s
ωwo
Ttb
Tis
I : mp +
`n−
TF
Jss
r2
s
−r a
TF
Se
Tda
Se
Se
fl,p
fhyd − Fg,p
R : fµ,p
Figure 4.4: Simplified model of the power steering gear from the input
torque at the input shaft to the output torque at the drop arm.
Ttf ,n = Ttb − Two = Ttb − Jwo ω̇wo ⇐⇒
v̇p
Ttf ,n = Ttb − Jwo .
n
(4.25)
Equation (4.24) and (4.25) can now be inserted in (4.17) to eliminate the last
equation that causes the causality conflict, resulting in
T
v̇p 1
⇐⇒
+ mp + da + fµ,p + fl,p − fhyd + Fg,p = Ttb − Jwo
r
n
n
J
J Tda
T
v̇p ss2 + mp + wo
+
= fhyd − mp gcos(θpsg ) − fµ,p − fl,p + tb .
(4.26)
r
n
r
n2
v̇p
Jss
r2
Using (4.26) results in the causal bond graph in Figure 4.5, where the worm, piston and sector shaft are modelled as one equivalent body. The equivalent mass,
meq , is described by
meq =
4.4
4.4.1
Jss
J
+ mp + wo
.
r2
n2
(4.27)
Steering Linkage
Previously Developed Steering Linkage Models
In Tai et al. [2004] the entire steering system after the power steering gear is modelled as one moment of inertia with a damper affected by a non-linear friction. In
Mårtensson [2005], a more complex model is developed by modelling the drop
arm one inertia, followed by a transformer and then the wheels.
30
4
I:
I : Jtb
Se
Tis
Ttb
p
s
ωlsc
`n−
TF
Jwo
n2
+ mp +
Steering System Mechanics
Jss
r2
s
−r a
TF
Se
Tda
C : ktb
Se
Se
floss
fhyd − Fg,p
R : fµ,p
Figure 4.5: Causal representation of the power steering gear from the input
torque of the valve slide to the output torque at the sector shaft.
4.4.2
Proposed Steering Linkage Model
Since the sector shaft of the power steering gear is rigidly connected to the drop
arm, they can be modelled as a single body with one moment of inertia. The
torque that acts as an effort on these bodies is transferred to a translational force
on the drag link. The proposed model is inspired by the model developed in
Ikjema [1999] where the drag link is modelled as a spring and its mass is rescaled
and combined with the inductive element of the drop arm.
Mathematical Description
A representation of the steering linkage is visualized in Figure 4.6. The input
to the subsystem is the drop arm torque that originates from the power steering
gear.
The drop arm and drag link are modelled with an equivalent moment of inertia,
2
Jda + mdl lda
, that is affected by the input torque and the force on the drag link,
Fdl . Its angular velocity, ωda , is calculated according to
ω̇da =
1
Tda − lda Fdl .
2
Jda + mdl lda
(4.28)
In this equation lda is the length of the drop arm and consequently the ratio of
the torque to force transformation.
Large forces act on the drag link, causing flexibility in the body [Eriksson, 1998].
The flexibility can be captured by modelling the drag link as a spring with the
spring stiffness constant kdl . The force in the spring is selected as a state. The
derivative of this force depends on the angular velocity of the drop arm and the
angular velocity of the wheels, ωwh according to
Ḟdl = kdl lda ωda − ldla ωwh .
(4.29)
4.5
31
Combined State Space Model
In this relation, ldla is the length of the drag link arm that transfers the translational velocity of the drag link to the rotation of the wheels.
Finally, the angular velocity of the wheels is given by
ω̇wh =
1 Jwh
Fdl ldla − Tsa − hµ,wh ,
(4.30)
where Jwh is the moment of inertia of the wheels including the mass of the track
rod as the track rod connecting the wheels is modelled as stiff, constraining the
steering angle of the left and right wheel to an equal angle. The non-linear friction of the king pin and wheels is given by hµ,wh .
2
I : Jda + mdl lda
` lda −
Se
Tda
s
ωda
TF
Fdl
p
I : Jwh
R : hµ,wh
s
ωwh
Se
− ldla a
TF
Tsa
C : kdl
Figure 4.6: Causal representation of the steering linkage from the input
torque of the drop arm to the output angle of the wheel.
4.5
Combined State Space Model
Combining the models of the subsystems yields a complete model of the steering
system mechanics. The bond graph of the complete model is shown in Figure 4.7.
As the complete bond graph has integrating causality, a state space model of the
system is easily derived with the flow variables of the inductive elements and the
effort variables of the capacitive elements as states. For this system the states are
selected as
  

x1  ωusc 
x   T 
 2   k,sc 
  

x3   ωlsc 
  

x = x4  =  Ttb  .
x   v 
 5   p 
  

x6   Fdl 
  

x7
ωwh
(4.31)
Using these states and the equations described above results a non-linear state
space model,
32
4

 0

 ksc

 bsc
 J
 lsc

ẋ =  0

 0


 0


0
1
− J +J
sw usc
0

 0
 0

 1
− Jlsc

+  0
 0


 0

 0
0
0
0
0
1
meq
0
0
1
Jlsc
0
−ksc
b
− J sc
0
0
0
0
0
0
0
0
0
0
0
ktb
0
0
− ntb
0
k
0
0
0
0
0
lsc
n
meq
0
0
l
kd l da
r
0
Steering System Mechanics


 1

 Jsw +Jusc



 0



 0



0  x +  0


 0
0 


 0



kdl ldla 
0

0
0
0
0
l
− rmda
eq
0
ldla
Jwh





 "
 T #
 sw

 Tsa




− J1 
0
0
0
0
0
0
0
0
0
0
0




 


hµ,sc (x1 ) + τl,is (x3 )
 

 

 fhyd (x4 , x5 ) − mp gcos(θpsg ) − fµ (x5 ) − floss (x4 ) .
 


hµ,wh (x7 )


0 
1 
wh
(4.32)
Jwh
As can be seen in the equations, this model is a 4 dof model, where 4 moments
of inertia or masses move individually.
By considering the steering column rigid, as proposed above, the model from
(4.32) can be simplified and two states removed. The new state vector is given by

  
x1   ωsc 
x2   Ttb 
  

x = x3  =  vp 
  

x4   Fdl 
x5
ωwh
(4.33)
(4.34)
and the new model by


1

0
0
0
− J1

 0
sc
0 
 Jsc


ktb



k


0
− n
0
0
#
 0

 tb
0  "



 Tsw
lda
1



0
−
0



0
0
x
+
ẋ =  0


 T
nmeq
rmeq
 0


l
0  sa


 0
0
kdl da
0
−k
l



dl dla 
r
 0 − 1 


ldla
Jwh
0
0
0
0
Jwh
 1

0
0 
− Jsc



 0
0
0  
hµ,sc (x1 ) + τl,is (x1 )


 

1
 0
0  fhyd (x2 , x3 ) − mp gcos(θpsg ) − fµ,p (x3 ) − fl,p (x2 , x3 ) .
+ 
m
eq



 
 0
hµ,wh (x5 )
0
0 



0
0
− J1
wh
(4.35)
4.6
33
Wheel Dynamics
This implies removing one dof and arriving at the three dof model suggested by
for example Rösth [2007] for the steering system of a truck.
4.6
Wheel Dynamics
As mentioned in Section 2.3, there are important forces from wheel and tire dynamics that the steering system needs to overcome to turn the wheels. The incoming torque, Tin,wh that attempts to turn the wheel equals the sum of the torque
that moves the moment of inertia, TJwh , the non-linear friction that acts between
the tire and the ground, hµ,g , the non-linear between the king pin and steering
knuckle, hµ,kp , and the self-aligning torque, Tsa , that acts to reduce the steering
angle, i.e.,
Tin,wh = TJwh + hµ,g + hµ,wh + Tsa .
(4.36)
Hsu [2009] uses a simple but efficient way to separate the self aligning torque,
separating it into a jacking torque, Tj and the total aligning torque, Ta , according
to
Tsa = Ta + Tj .
(4.37)
In this thesis, neither the total aligning torque Ta , which depends greatly on the
vehicle slip angle, nor the friction between the ground and tire, hµ,g , is modelled.
The jacking torque is a function of suspension geometry and is described in Hsu
[2009] as
Tj (δ) =Fz
dhsus
(δ).
dδ
(4.38)
The vertical load on the tire is given by Fz and the change in suspension height
due to steering is given by hsus . A simplified model of the derivative of the suspension height is given in Bastow et al. [2004] as
eof f
dhsus
(δ) =
sin(2θK P I )sin(δ),
dδ
2
(4.39)
where eof f is the lateral offset at the hub and θK P I is the kingpin inclination angle.
Using (4.38) together with (4.39) gives one of the proposed models of the jacking
torque that will be used later and is henceforth denoted the Sinus Model. This
simple equation neglects the influence of caster and camber angle on the jacking
torque. This influence is described in Wu et al. [2015] and Tagesson [2014] but
not modelled in this thesis.
Tsw
Se
s
p
s
ωusc
Se
τl,is
R : hµ,sc
s
I : Jsw + Jusc
Tsc
R : bsc
C : ksc
`n−
TF
Ttb
p
C : ktb
Se
I : Jlsc
s
vp
I : meq
fhyd − Fg,p − fl,p
−
TF
r
lda
a
R : fµ,p
C : kdl
Fdl
p
TF
− ldla a
s
ωwh
I : Jwh
Tsa
Se
R : hµ,wh
34
4
Steering System Mechanics
Figure 4.7: Bond graph of the complete steering system.
4.6
35
Wheel Dynamics
An alternative way to describe the jacking torque is found in Bastow et al. [2004].
The force in the draglink induced by the jacking torque, Fdl,j is given by
Fdl,j =
Fz dhsus
.
ldla δ dδ
(4.40)
For small to medium angles this expression is strictly larger than 0 and it denotes
the magnitude of the force. In the model, (4.40) is multiplied with sgn(δ) to find
the force. Deriving the jacking torque from (4.40) by inserting (4.39) results in
Tj (δ) = Fz
eof f
2
sin(2θK P I )
sin(δ)
sgn(δ)
δ
(4.41)
which differs from the Sinus Model. This will be the second modelling alternative
for the jacking torque used here and is denoted the Sinc Model.
The difference between the Sinus Model and the Sinc Model is that the latter
model applies a higher torque for lower steering angles, in comparison to a sine
function that has low torques for small angles and increases the torque as the
angle increases. However, the model (4.41) causes a discontinuity close to δ =
0. For computational purposes it is advisable to eliminate the discontinuity by
creating a linear transition band. Notably, the two models are contradictory with
very different characteristics. The reason for this disparity is not known but the
models in Wu et al. [2015] and Tagesson [2014] are more similar to the Sinus
model than the Sinc model.
5
Steering System Hydraulics
In this section, models are developed for the hydraulic system that generates
steering assistance. First the model for the hydraulics is presented and then the
valve model is described.
General modelling of hydraulic actuator systems for heavy vehicle applications
has previously been studied in Nilsson [2014]. The modelling of the power chambers of the hydraulic system in this section is based on this thesis. Hydraulic
power steering in particular has been discussed at length in Rösth [2007] and the
proposed model of the valve in this thesis is inspired by Rösth [2007] and Merritt
[1967]. The modelling of the hydraulics is also influenced by a previous model
developed at Scania and described in Ikjema [1999]. The modelling approach
used in Ikjema [1999] is applied when developing the power steering gear model
in Mårtensson [2005].
5.1
Hydraulic Model
The system that provides the hydraulic assist in the power steering gear is suitably modelled as a hydraulic actuator system. In Figure 5.1 the proposed modelling principle is shown. The hydraulic force, fhyd , is generated by the pressure
difference in the chambers on the two sides of the piston. Let pL and pR be the
pressure in the chambers and let Ap,L = Ap,R = Ap be the area of the piston
exposed to each chamber; the symmetry follows from the symmetry of the hydraulic force. Using these variables the hydraulic force can be expressed as
fhyd = pL Ap,L − pR Ap,R = (pL − pR )Ap = pl Ap .
37
(5.1)
38
5
Steering System Hydraulics
𝑄2
𝑄1
𝑃𝑝
𝜓1
Pump
𝑄3
𝜓3
𝜓2
𝑄𝑝
𝑄4
𝜓4
𝑃0
𝑄𝑙
𝑃𝐿
Piston
𝑃𝑅
Figure 5.1: Respresentation of the hydraulics.
When omitting the compressibility of the hydraulic fluid, the following three
flows are equal: from the piston motion Ql , entering the left chamber and leaving
the right chamber. In this case the chamber volume will depend on the position
of the piston, xp which is bounded according to
xp,min ≤ xp ≤ xp,max .
(5.2)
If the piston is modelled as a brick, this results in
V L = V0 + A p x p
(5.3)
V L = V0 + A p x p
(5.4)
and
In these equations V0 is the chamber volume when the valve angle is 0 and xp
is the piston displacement from the equilibrium. Differentiating (5.3) and (5.4)
results in
5.1
39
Hydraulic Model
dVL
= Ap ẋp = Ap vp
dt
(5.5)
dVR
= −Ap ẋp = −Ap vp .
dt
(5.6)
and
L
Finally Ql = dV
dt , which implies the load flow depends on the piston velocity, vp
and the piston area, Ap according to
Ql = vp Ap .
(5.7)
Each of the two power cylinder chambers are connected via the valve to the feed
and return side of the hydraulic pump system. In Merritt [1967], the pressure
drop over the valve and to the actuator chamber via any connection in particular
can be modelled as an orifice resulting in
X
∆p = ψ φ, Qf ,
ρQf2
2cq2 A(φ)2
(5.8)
and
√
Qf = ψ
−1
φ,
X
2cq A(φ)qX
∆p.
∆p ,
√
ρ
(5.9)
In these equation φ is the valve angle, ρ is the oil density, cq is a constant, and Av
is the orifice area. This pressure drop function is henceforth denominated Ψ .
The following calculations are based on Merritt [1967] and Rösth [2007]. The
model for the hydraulic assist is derived by considering the hydraulic fluid incompressible and assuming constant pump flow passing through the valve; statically
the pump flow and the flow passing through the valve are equal but they can vary
dynamically. Furthermore the valve is modelled as symmetric, meaning that
Av [φ] = A1 [φ] = A4 [φ] = A2 [−φ] = A3 [−φ].
(5.10)
is fulfilled. In (5.10), the different valve areas, Av , that redirect the flow of oil are
presented, these correspond to the area openings in Figure 5.1. Considering the
symmetry of the hydraulic assist it is reasonable to assume that (5.10) holds.
40
5
Steering System Hydraulics
This model of the system is illustrated as a bond graph in Figure 5.2. The bond
graph contains a causal loop, illustrated with red lines in Figure 5.2. To avoid
this issue, the bond graph can be reduced to the bond graph in Figure 5.3. The
removal of the causal loop corresponds to the assumption that the pressure before
the pump equals 0. This assumption seems reasonable as the pressure before the
pump is small in comparison to the pressure after the pump. The modification
in the bond graph does not change the value of the external variables such as the
load pressure according to Ljung and Glad [2004].
From the bond graph it can be seen that


Q


 p
Q1



 Q2
= Q1 + Q2
= Q3 + Q l .
= Q4 − Q l
(5.11)
= Q p − Q l − Q2
= Q p + Q l − Q1
(5.12)
This can be re-written as
(
Q3
Q4
Due to the symmetry of the valve and the fact that the inflow equals the outflow,
Q1 = Q4 and Q2 = Q3 . This can be further transformed into
(
2Q1
2Q2
= Qp + Ql
.
= Qp − Ql
(5.13)
Using (5.8) and (5.13), Rösth [2007] derives


ρ  (Qp − Ql )2 (Qp + Ql )2 
−
pl = p1 + p2 = 2 

8cq
A2v [φ]
A2v [−φ]
(5.14)
for the pressure over the piston and


ρ  (Qp − Ql )2 (Qp + Ql )2 
pp = p1 + p2 = 2 
+

8cq
A2v [φ]
A2v [−φ]
(5.15)
for the pump pressure. As no dynamics are modelled in this physical model
the result is a static relationship between the piston velocity, valve angle, and
hydraulic force. These calculations will not be given in this thesis.
5.2
41
Valve Model
R : ψ3
R : ψ1
p1
p
Q1 s
pp p
Sf
R : ψ2
Sf
s
Q2
p
p2
s Q3
R : ψ4
s Ql
s
Q4
p
p0
s
Qp
Figure 5.2: Bond graph of the simple hydraulic model with causal loop.
5.2
Valve Model
According to Rösth [2007] there are two alternatives to valve modelling, either
modelling the valve area or modelling the valve geometry. Rösth [2007] argues
the importance of studying the geometry when designing a new valve but states
that the area function might be more interesting when modelling an existing
valve. Since this work focuses on modelling an existing valve, the approach used
in this thesis is based on modelling the valve area openings.
The valve is controlled by the torque in the torsion bar. When a deformation of
the torsion bar occurs, this means that the angle of the input shaft is not the same
as the angle of the output shaft of the power steering gear. The hydraulic force
should thus increase as the deformation of the torsion bar increases in order to
help steer the vehicle according to the input torque at the input shaft.
The torsion at the torsion bar results in a torque equal to the spring constant
multiplied by the deformation, i.e. the integral of the angular velocity according
to
Ṫtb = ktb ω =⇒ Ttb = ktb φ.
(5.16)
Ikjema [1999] presents the valve area model
3
Av [φ] = Amin + A0 e−k̃1 φ−k̃3 φ .
(5.17)
The area function consists of a constant and an exponential part. The design by
Ikjema [1999] is motivated in part by the physical system being modelled and in
part by stability concerns. As e x > 0, the valve area in (5.17) is always greater
42
5
Steering System Hydraulics
R : ψ3
R : ψ1
p1
p
Q1 s
R : ψ2
pp p
Sf
s
Q2
s Q3
Ql s
Sf
p
p2
s
Q4
R : ψ4
s
Qp
Figure 5.3: Bond graph of the simple hydraulic model with causal loop removed.
than 0 when A0 and Amin are positive constants. Physically this means the valve
is always slightly open. As the valve area function is always positive, the division
in (5.14) does not cause any numerical issues. Numerical stability can be assured
by adjusting Amin . The non-linear behavior of the hydraulics can be modelled by
selecting A0 , k̃1 , and k̃3 .
Since the torsion bar torque and deformation angle are proportional according to
(5.16), the torque in the torsion bar can be used as a control variable instead of
the angular deformation. This is convenient as the torque of the torsion bar is a
state in the model shown in (3.2) and (??). Using the torque as input, the valve
area function can be designed with different constants, k1 and k3 suitable to the
control variable Ttb .
As stated above Ikjema [1999] and Rösth [2007] model the valve as symmetrical.
However Rösth [2007] states that small deviations in geometry when mounting
the valve will result in the pressure curve becoming unbalanced. This asymmetry
results in four different area functions, A1 [φ], A2 [φ], A3 [φ] and A4 [φ]. In production valves, such deviations are likely to occur. In (5.14), the valve area openings
are modelled as symmetric. The small deviations and their resulting asymmetric
pressure curves can be modelled by assigning different area functions to positive
and negative torques. This solution however would only be valid for one particular power steering gear. Another method is to model a torsion bar offset, Ttb0 ,
to capture the unbalanced behavior in the pressure curve. The latter solution is
implemented in
3
Av [Ttb ] = Amin + A0 e−k1 (Ttb −Ttb0 )−k3 (Ttb −Ttb0 ) .
(5.18)
6
Data Analysis and Parameter
Estimation
In this chapter measurement data is used to identify parameters in the grey box
physical models developed in Section 4 and 5. The data is also used to assign suitable model structures to the unknown functions. Furthermore black box models
are estimated for particular subsystems.
6.1
Measurement Data
The measurement data used to estimate models of different subsystems should be
divided into estimation and validation data. Ideally the estimation and validation
data are recorded under identical or similar circumstances. The input signal
should ideally be dynamic and excite as many frequencies as possible, which
means that the energy of the input should be evenly distributed across a large
range of frequencies. The measurement data used to identify the system will be
presented in this section.
In total, six different data sets have been used in this thesis and information regarding the data sets is summarized in Table 6.1. Data Sets 1-5 are measured in a
rig, under the circumstances described below. These data sets focus on capturing
the behavior of the power steering gear and its hydraulics. Data Sets 1-3 are all
measured statically – that is measurements are made in steady state. Data Set 6
contains measurements of a stationary vehicle.
Out of these data sets, Data Set 6 is the only experiment made specifically for
this thesis. Consequently, the other data sets do not necessarily have the optimal
input signals or output signals for this estimation purpose. Instead the input
signal to Data Sets 1-5 only consists of one single frequency; the zero frequency
for Data Set 1-3 and a low non-zero frequency for Data Set 4 and 5. Furthermore,
43
44
6
Data Analysis and Parameter Estimation
the components of the steering systems are not identical in all measurements,
which can be a source of error. Finally, an added complication is that for Data
Sets 1-5 there is not separate validation data available.
Table 6.1: Measurement data.
Data set
Sample Time
Samples
No. of Series
Data Set 1
0.05s
2622
1
Data Set 2
0.1s
3173
1
Data Set 3
1s
1086
1
Data Set 4
0.05s
12360
1
Data Set 5
0.1s
7479
1
Data Set 6
0.01s
5992-15212
6
6.1.1
Signals
Input shaft torque and
output shaft torque
Input shaft torque and
pump pressure
Input shaft torque,
pump pressure, and
output shaft torque
Input shaft torque and
input shaft angle
Output shaft torque and
output shaft angle
Input shaft torque, input shaft angle, and
steering angle at the
wheels
Data Set 1 – Mechanical Efficiency
The first data set consists of static measurements from the power steering gear.
The input is a torque applied at the input shaft of the power steering gear and
the output is a torque measured at the drop arm. To better isolate the mechanical
efficiency, the hydraulic assist is deactivated. However, there is still oil in the
cylinder of the power steering gear.
In this scenario, the drop arm can be fixed at different angles. Figure 6.1 shows
a measurement when the drop arm is fixed at zero degrees. The data set shows
that the torques needed to perform a right turn and a left turn are not identical
in the power steering gear mechanics. This is due to a non-linear inefficiency that
needs to be modelled.
6.1.2
Data Set 2 – Valve Characteristics
The second data set provides a basis for studying the valve characteristics. The
input for the data set is a torque measured at the input shaft. The output is the
pump pressure in the hydraulic power assist system, which in this scenario has a
constant pump flow. To build up the pump pressure, the drop arm is fixed and
the measurements are made under static conditions. Figure 6.2 shows the signals
of the data set when the drop arm is fixed at zero degrees.
6.1
45
Measurement Data
Input torque at the input shaft
Normalized torque
1.5
1
0.5
0
-0.5
-1
-1.5
0
20
40
60
80
100
120
100
120
Time [s]
Output torque at the droparm
Normalized torque
20
10
0
-10
-20
0
20
40
60
80
Time [s]
Figure 6.1: Data Set 1 with the drop arm fixed at 0 degrees and hydraulics
disabled. The top plot shows the torque applied to the input shaft and the
bottom plot shows the torque necessary to keep the output shaft fixed. The
differences between the torques show that there are static friction and inefficiency.
Input torque at the input shaft
Normalized torque
1
0.5
0
-0.5
-1
0
20
40
60
80
100
120
140
160
180
200
220
240
260
280
300
220
240
260
280
300
Time [s]
Normalized pressure
Hydraulic pump pressure
1
0.5
0
0
20
40
60
80
100
120
140
160
180
200
Time [s]
Figure 6.2: Data Set 2. The top plot shows the torque applied to the input shaft and the bottom plot shows the pump pressure resulting from the
deformation of the torsion bar.
46
6.1.3
6
Data Analysis and Parameter Estimation
Data Set 3 – Hydraulic Efficiency
To further study the hydraulic power assist, a data set is available for the same
scenario as Data Set 1 but with the hydraulics activated. This third data set contains static measurements of the relation between the pump pressure and the
torque measured at the drop arm when the drop arm is fixed. This data set can
be used to estimate inefficiencies that are not only related to the power steering
gear mechanics, but the hydraulic inefficiencies as well. One plot from this data
set is shown in Figure 6.3.
6.1.4
Data Set 4 – Input Analysis
The fourth data set contains measurements of the input shaft torque and input
shaft angle when the drop arm is disconnected from the drag link and the hydraulics are disconnected. During the experiment, the magnitude of the velocity
at the input shaft is constant and low. This data set may be used to estimate one
value of the viscous friction in the power steering gear and on the input shaft.
The data from this data set can be seen in Figure 6.4.
6.1.5
Data Set 5 – Reversible Torque
In the fifth next data set the input shaft is disconnected from the power steering
gear and the drop arm is forced to move with constant magnitude of angular
velocity as shown in Figure 6.5. For this data set the necessary torque and the
sector arm angle are measured. Just as for the input analysis, these measurements
are made with low constant velocities.
6.1.6
Data Set 6 – Stationary Vehicle Measurements
In order to estimate the parameters of the steering linkage and the wheels, full
system measurements are needed. However, as measurements from driving on
the road are noisy and require a model of the total self-aligning torque, measurements have been made on a stationary vehicle.
During the experiments, the wheels of the test vehicle are placed on low-friction
turn tables that allows the wheels to turn with next to no friction from the ground.
The angles of the wheels and steering column are measured, as well as the input
torque applied to the input shaft. However, there are no measurements of the
steering wheel torque and therefore a correct steering column model is difficult
to obtain from this data set. For small angles, the left and right wheel angles are
similar. However for large angles, a small deviation can be observed. A major
advantage of this data set is that the self aligning torque is limited to the jacking
torque and disturbances are essentially eliminated.
The input signals used as input torques during these experiments are varied and
separate estimation and validation data sets are available for this data set. One
measurement series from this set is shown in Figure 6.6.
47
Measurement Data
Normalized torque
6.1
Input torque at the input shaft
1
0.5
0
-0.5
-1
Normalized torque
Normalized pressure
0
100
200
300
400
500
600
700
800
900
1000
700
800
900
1000
700
800
900
1000
Time [s]
Hydraulic pump pressure
1
0.5
0
0
100
200
300
400
500
600
Time [s]
Sector arm torque
1
0.5
0
-0.5
-1
0
100
200
300
400
500
600
Time [s]
Figure 6.3: Data Set 3 with sector arm fixed at 0 degrees. The top plot shows
the torque at the input shaft, the middle plot shows the pump pressure, and
the bottom plot shows the torque necessary to keep the output shaft fixed.
Input shaft torque
Normalized torque
1
0.5
0
-0.5
-1
0
100
200
300
400
500
600
400
500
600
Time [s]
Sector arm angle
Normalized angle
1
0.5
0
-0.5
-1
0
100
200
300
Time [s]
Figure 6.4: Data Set 4. The top plot shows the torque applied to the input
shaft and the bottom plot shows the angle of the input shaft.
48
6
Data Analysis and Parameter Estimation
Sector arm torque
Normalized torque
1
0.5
0
-0.5
-1
0
100
200
300
400
500
600
700
500
600
700
Time [s]
Sector arm angle
Normalized angle
1
0.5
0
-0.5
-1
0
100
200
300
400
Time [s]
Figure 6.5: Data Set 5. The top plot shows the torque applied to the output
shaft and the bottom plot shows the resulting angle of the output shaft.
Input torque at the input shaft
Normalized torque
1
0.5
0
-0.5
-1
0
2
4
6
8
10
12
14
16
18
20
14
16
18
20
Time [s]
Left wheel angle
Normalized angle
1
0.5
0
-0.5
-1
0
2
4
6
8
10
12
Time [s]
Figure 6.6: Data Set 6. Measurements from a stationary vehicle where the
top plot shows input shaft torque and the bottom shows the steering angle
of the left wheel.
6.2
49
Parameters and Functions
6.2
Parameters and Functions
In this introductory section the parameters and functions are summarized.
6.2.1
Steering Column
As shown in Table 6.2 almost all parameters in the model of the steering column
from Section 4 are known. The notable exception is the friction and damping
coefficients.
Most parameters in the power steering gear are known; this is shown in Table 6.3.
Just as in the steering column, no information is available on the damping and
other friction forces. Even though most rig data are static, both Data Set 4 and
Data Set 5 have some dynamics and can be used to estimate the friction. All parameters in the steering linkage are known and summarized in Table 6.4. As can
be seen in Table 6.5, the wheel inertia and friction are unknown. Consequently
these need to be estimated.
Table 6.6 summarizes the parameters in the model given in Chapter 5. The density ρ in (5.14) can be found in a table. The function Av that describes the relationship between valve opening and torsion bar displacement depends on the valve
design. This function will be found using system identification. The flow coefficient, cq , is unknown and has to be identified from measurement data. However,
as can be seen in (5.14), cq is multiplied directly with Av and will thus not be observable as more than a scaling of the area function. Consequently the constant
can be given a fixed value without affecting the solution. After this simplification, the only unknown parts of the hydraulic subsystem is the area function.
This function in turn is parametrized according to (5.18) where all constants are
unknowns.
6.3
Valve Characteristics
As Data Set 2 shows the relationships between the torsion bar torque and pump
pressure, it is used to estimate the parameters in the valve model (5.18). In Figure
6.7, the relationship between the input shaft torque and pump pressure for Data
Set 2 is shown. The pump pressure increases with increasing torque magnitudes
but the pump pressure is limited within an interval due to a pressure release
valve,
pp,min ≤ pp ≤ pp,max .
(6.1)
The pump pressure is given by the sum of the pressure applied on both sides of
the piston as in
50
6
Data Analysis and Parameter Estimation
Table 6.2: Steering column parameters.
Parameter
Description
Estimation status
Jsw
Jusc
Jlsc
ksc
Steering wheel inertia
Upper steering column inertia
Lower steering column inertia
Steering column stiffness
Calculated from geometry
Calculated from geometry
Calculated from geometry
Known from measurements
Function
Description
Estimation status
hµ,sc
Steering column friction
Unknown
Table 6.3: Power steering gear parameters.
Parameter
Description
Estimation status
Jwo
mp
n
r
Worm inertia
Piston mass
Worm to piston ratio
Sector shaft radius
Calculated from geometry
Known
Known
Known
Function
Description
Estimation status
fµ,p
Piston friction
Unknown
Table 6.4: Steering linkage parameters.
Parameter
Description
Estimation status
Jda
mdl
Jdla
mtr
kdl
lda
ldla
Drop arm inertia
Drag link mass
Drag link arm inertia
Track rod mass
Drag link flexibility
Drop arm length
Drag link arm length
Calculated from geometry
Known
Calculated from geometry
Calculated from geometry
Known
Known
Known
Table 6.5: Wheels and external influence parameters.
Parameter
Description
Estimation status
Jwh
eof f
θK P I
Fz
Wheel inertia
Lateral offset at hub
Kingpin inclination
Normal force on front axle
Unknown
Known
Known
Known
Function
Description
Estimation status
hµ,wh
Wheel friction
Unknown
6.3
51
Valve Characteristics


ρ  (Qp − Ql )2 (Qp + Ql )2 
pp = pL + pR = 2 
+ 2
 .
8cq
A2v [Ttb ]
Av [−Ttb ]
(6.2)
Figure 6.8 shows how the valve area function from (5.18) depends on the parameters A0 and Amin . These parameters can be determined from the minimal applied
pump pressure in the boost curve when the piston is fixed. Logically the pump
pressure will be at its lowest when the valve is in its equilibrium position, i.e.
T̃tb = Ttb − Ttb0 = 0. The described scenario can be observed by simplifying (5.15)
to
pp,min = pp (vp = 0, Ttb


2Qp2

ρ 
 .
= 0) = 2 

2
8cq (A0 + Amin )
(6.3)
The parameters A0 and Amin can be determined since all the parameters except
pp,min in the right hand side of
r
Qp
ρ
A0 + Amin =
(6.4)
2cq pp,min
are known. The minimal pressure can be approximated by the minimal measured
pump pressure in the boost curve.
The minimal area, Amin represents the minimal area opening of a valve. If the
valve redirecting oil to the right side of the piston is completely open, the opening
that redirects oil to the left hand side of the piston will be represented by Amin .
As the primary purpose of Amin is guaranteeing computational stability, it can be
fixed at a small value. Theoretically, Amin could be set to zero and the parameters
k1 and k3 in (5.18) could be estimated accordingly.
6.3.1
Valve Offset
As can be seen in Figure 6.7, there is a valve offset skewing the boost curve to
the left. As noted previously, this valve offset will be different for any given
power steering gear. To estimate the valve offset the measurements are sorted
and low-pass filtered with a fifth order zero-phase Butterworth filter with cut-off
frequency fc = 0.025H z. These filtered measurements are shown as a red line
in Figure 6.9. Following the filter, a second degree polynomial is fitted to the
line and shown as a black line in the same figure. Finally the input torque of the
minimum of the second degree polynomial is chosen as a valid approximation of
the valve offset.
6.3.2
Valve Function
In Data Set 2, there is no load flow as the piston is fixed. In this case (5.14)
simplifies to
52
6
Data Analysis and Parameter Estimation
Table 6.6: Hydraulic parameters.
Parameter
Description
Estimation status
ρ
Qp
pp,min
pp,max
cq
Ap
Hydraulic fluid density
Pump flow
Minimum pump pressure
Maximum pump pressure
Flow coefficient
Piston area
Known
Known
Known
Known
Unknown
Known
Function
Description
Estimation status
fhyd
Hydraulic assist force
Unknown
1.2
Normalized pump pressure
1
0.8
0.6
0.4
0.2
0
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Normalized input torque
Figure 6.7: In this figure, the relationship between the input shaft torque
and pump pressure for Data Set 2 is shown. The pump pressure increases
with increasing torque magnitudes.
6.3
53
Valve Characteristics
Valve area opening
1
A v [-Ttb ]
A v [Ttb ]
A 0 + Amin
Normalized valve area opening
A min
0.5
0
-5
-4
-3
-2
-1
0
1
2
3
4
5
Torsion bar torque [Nm]
Figure 6.8: This figure shows the valve openings area as a function of the
torsion bar torque.
1.4
Measured pump pressure
Filtered measurement
Fitted polynomial
Estimated valve offset
1.2
Normalized pump pressure
1
0.8
0.6
0.4
0.2
0
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Normalized input torque
Figure 6.9: In this figure, the relationship between the input shaft torque
and pump pressure for Data Set 2 is shown. The red line is a sorted and
filtered version of the measurements and the black line is a second degree
polynomial fitted to the red line.
54
6
pp
= pL + pR =
ρ
8cq2
Data Analysis and Parameter Estimation
(Qp )2
A2v (Ttb )
+
(Qp )2
(6.5)
A2v [−Ttb ]
which in turn implies that
8cq2 pp
ρQp2
=
=
1
+ 21
2
A2v (Ttb ) 2Av (−Ttb )
Av (Ttb )+Av (−Ttb )
.
A2v (Ttb )A2v (−Ttb )
=
(6.6)
For any given measurement, the pump pressure is known and this means that
the only unknown in (6.6) is the area function. After inserting the valve function
given by (5.18), this expression can be further transformed into
0 = Ax4 + Bx3 + Cx2 + Bx + A.
(6.7)
where























x
= e−k1 (Ttb −Ttb0 )−k3 (Ttb −Ttb0 )
A = A20 −
8cq2 pp
ρQp2
(A2min A20 )
= 2A2min −
8cq2 pp
(2A3min A0 + 2Amin A30 )
ρQp2
8cq2 pp
(A4min + A40 + 4A2min A20 )
ρQp2
B = 2Amin A0 −
C
3
.
(6.8)
As this expression is a fourth degree polynomial it can be solved – e.g. by using
the roots function in MATLAB.
Obviously the equation in (6.8) will have four solutions for every measurement.
To find the right solution first the complex solutions are discarded. The remaining solutions are compared with a reference solution; this reference for x is found
ad hoc by tuning k1 and k3 so that they mimic the main behaviour of the valve
function. The solution that is closest to the reference solution is then chosen for
further manipulation and estimation. Finally, the logarithm of the chosen solution is considered a pseudo-measurement of the polynomial in the exponential
of x. This measurement equation can be written
y = ln(x) = −k1 (Ttb − Ttb0 ) − k3 (Ttb − Ttb0 )3 + e,
(6.9)
where y is the pseudo-measurements and e is a measurement error. These pseudomeasurements are shown in Figure 6.10.
A first estimate for k1 and k3 is found by omitting the hysteresis and approximating Ttb by Tis . From this model k1 and k3 are found directly by linear regression.
6.4
55
Hysteresis
Normalized pseudo-measurements
1
0.5
0
-0.5
-1
-1
-0.5
0
0.5
1
Normalized input torque
Figure 6.10: On the y-axis this shows pseudo-measurements of the valve
function calculated from the pump pressure in Data Set 2. The pseudomeasurements are plotted against the input shaft torque on the x-axis.
This estimate is revised for some hysteresis models in Section 6.4 by developing
other estimates of Ttb and using the estimate in the linear regression.
6.4
Hysteresis
As noted in Section 6.1 the data from both the Mechanical Efficiency and Hydraulic Efficiency experiments are static. According to the ideal model detailed
in Sections 4.5 and 5.1, this implies one input torque should always result in the
same output torque, due to the lack of transients. However, as can be seen in Figure 6.11, the input-output relationship in the Mechanical Efficiency tests gives
two output torques for any given input torque. This hysteresis phenomenon is
also apparent in the Hydraulic Efficiency tests as is shown in Figures 6.12 and
6.13.
In accordance with the definition of hysteresis phenomena, the system behaviour
depends on both the current torque and its previous values. In Figure 6.14, the
valve hysteresis is shown together with the time plot for the same measurement
series. From this plot it can be seen that for a given input torque magnitude there
are two different pump pressure curves, one when the absolute value of the input
torque is increasing and a second value for decreasing torques.
Upon closer analysis this conclusion is logical. When the torsion bar is being
deformed, the input torque has to overwin the resistance inherit in the system –
including but not limited to static friction. Analogously, when the torsion bar is
returning to its natural state, the torque from the torsion bar has to overwin the
56
6
Data Analysis and Parameter Estimation
1
Normalized output torque
0.5
0
-0.5
-1
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Normalized input torque
Figure 6.11: This figure shows the relationship between the input shaft
torque(x-axis) and output shaft torque(y-axis) for Data Set 1. As the measurements are static, the two different output shaft torques for any given
input shaft torque show that there is hysteresis.
1.2
Normalized pump pressure
1
0.8
0.6
0.4
0.2
0
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Normalized input torque
Figure 6.12: This figure shows the relationship between the input shaft
torque(x-axis) and pump pressure (y-axis) for Data Set 2. As the measurements are static, the two different pump pressures for any given input shaft
torque show that there is hysteresis.
6.4
57
Hysteresis
same resistance. By performing the same analysis with respect to the other data
sets, the same conclusion can be drawn.
6.4.1
Efficiency Constants
It is natural to assume there are hysteresis phenomena in three places in the
power steering gear: from input shaft to torsion bar, from torsion bar to piston, and from piston to output shaft. In for example Figure 6.11 it can be noted
that for a given hysteresis state (e.g. increasing and positive torque) the inputoutput relationship seems approximately linear. Following this line of reasoning
a reasonable model for the hysteresis is a hysteresis state dependent efficiency
constant, Cef f , as shown in
Tout = Cef f Tin
(6.10)
with
Cef f







=





C1
C2
C3
C4
ωin ≥ 0,
ωin < 0,
ωin ≥ 0,
ωin < 0,
Tout
Tout
Tout
Tout
≥0
≥0
.
<0
<0
(6.11)
Ideally, the efficiency constants would be identical for increasing torque magnitudes, i.e. C1 = C4 , and for decreasing, C2 = C3 . However, as the measurement
data demonstrates asymmetrical behaviour for positive and negative torques, all
four constants are necessary.
Using the efficiency constants in the three places suggested above, the loss torque
and loss force in (4.32) and (4.35) can be re-written as



 Tl,is


 fl,p
= (1 − Cef f ,in )Tsw
=
1−Cef f ,worm
Ttb
n
+
1
1
r 1−Cef f ,out Fdl
.
(6.12)
With these functions and for static data – in for example Data Set 3 – (4.32) and
(4.35) can be simplified to
Tout
= Cef f ,out r(fhyd + Cef f ,worm n1 Cef f ,in Tsw ).
(6.13)
Consider for example the hysteresis between the input shaft and the torsion bar
and equation (6.10). Let the torque and angular velocity of the input shaft be
Tin = Tis and ωin = ωis , respectively, and let Tout = Ttb be the torque in the
torsion bar. If ωis and Ttb are both positive this implies the input shaft acts
to increase the torque in the torsion bar. This implies that Tis has to overcome
58
6
Data Analysis and Parameter Estimation
1
Normalized output torque
0.5
0
-0.5
-1
0
0.2
0.4
0.6
0.8
1
Normalized pump pressure
Input torque
Figure 6.13: This figure shows the relationship between the input shaft
torque(x-axis) and output shaft torque(y-axis) for Data Set 3. As the measurements are static, the two different output shaft torques for any given
input shaft torque show that there is hysteresis.
1
0
-1
Pump pressure
0
50
100
150
200
250
300
200
250
300
Time (s)
1
0.5
0
0
50
100
150
Time (s)
1.2
Increasing torque
Decreasing torque
Pump pressure
1
0.8
0.6
0.4
0.2
0
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Input torque
Figure 6.14: Data Set 2. The first plot shows the input shaft torque over time,
the second plot shows the pump pressure over time, and the third plot shows
the relationship between the input shaft torque(x-axis) and pump pressure
(y-axis). The figure shows that a given input shaft torque gives one pump
pressure when the torque magnitude is increasing and another when it is
decreasing.
6.4
59
Hysteresis
the resistance in the system and Cef f will therefore be smaller than 1. Similar
reasoning can be made for each case.
The hysteresis model in (6.10) and (6.11) has a discontinuity for ωin = 0. To
prevent that high frequency oscillation of the angular velocity cause overall oscillatory behaviour, it is advisable to filter the measurements used to calculate the
efficiency constants. To remove the discontinuity and improve simulation speed
it is also possible to filter the efficiency constants. Combining (6.10) and (6.11)
with these two filters give
Tout
Ċef f
= Cef f Tin
= Ĉef f − Cef f Kef f
(6.14)
with
Ĉef f







=





C1
C2
C3
C4
ωin ≥ 0,
ωin < 0,
ωin ≥ 0,
ωin < 0,
Tout
Tout
Tout
Tout
≥0
≥0
<0
<0
(6.15)
and
ω̇in = ω̂in − ωin Kω
(6.16)
where Kef f and Kω are filter constants and ω̂in is the angular velocity state in the
overall model. In this extended hysteresis model ωin can be considered a state
describing the torque direction rather than a direct measurement of the angular
velocity.
Total Efficiency
As Data Set 1 is both static and without hydraulic assist, (6.13) can be simplified
into
Tout
= Cef f ,out Cef f ,worm Cef f ,in nr Tin .
(6.17)
Furthermore as the data is static the three hysteresis states are found by assuming that sgn(ωi ) ≈ sgn(Ṫin ) for all hysteresis states. Under this assumption the
hysteresis states are found by differentiating and then filtering the input torque.
The differentiation is done by the Euler approximation given in (3.14) and the filter is a low-pass, fifth order, zero-phase, Butterworth filter with cut-off frequency
fc = 1H z. This same differentiation approach is used in remainder of this section.
When the hysteresis states are found the values of the joint efficiency constant
Cef f ,tot = Cef f ,out Cef f ,worm Cef f ,in are found by linear regression.
Input Efficiency Constants and Valve Parameters
Using the efficiency constants, (6.9) can be re-written as
60
6
Data Analysis and Parameter Estimation
y = −k1 (Cef f ,in Tin − Ttb0 ) − k3 (Cef f ,in Tin − Ttb0 )3 + e
(6.18)
using the hysteresis model (6.13). The hysteresis states can be found by differentiating and smoothing Tin (as the piston is fixed sgn(ωin ) = sgn(Ṫtb ) ≈ sgn(Ṫin )).
The numerical differentiation and smoothing is done as above. When the hysteresis states are found linear regression can be done on (6.18) to find k1 , k3 and the
four states of Cef f ,in .
Output and Worm Efficiency Constants
In the Hydraulic Efficiency data set there are measurements of both input torque
and output torque. For these measurements (6.13), can be written as
Tout
= Cef f ,tot nr Tin + Cef f ,out fhyd =
= Cef f ,out Cef f ,worm Cef f ,in nr Tin + Cef f ,out fhyd .
(6.19)
As the total efficiency constant is known from Section (6.4.1) and the hydraulic
force is known from Section 6.3, the only unknown in the equation is Cef f ,out .
Analogously to the methodology used for finding the efficiency constants previously, the hysteresis constants are found by assuming that sgn(ωi ) ≈ sgn(Ṫout )
and differentiating the output torque (using the same approximation and filter
as above). Finally, linear regression is used to find values for Cef f ,out conditioned
on the hysteresis states.
As there is now knowledge of Cef f ,out , Cef f ,in and Cef f ,tot , the final efficiency
constant Cef f ,worm can now be found by using
Cef f ,worm
6.4.2
=
Cef f ,tot
Cef f ,out Cef f ,in .
(6.20)
Adaptation of Hydraulic Valve Coefficients
An alternative to the modelling alternative described in Section 6.4.1 is to only
adapt the valve function in the hydraulics. As there is hysteresis in Data Set 1,
where there are no hydraulics, this model will capture less of the system behaviour. However, as the hydraulics is already a non-linear part of the model, this
modelling approach can prove advantageous from a simulation time perspective.
Furthermore, it is worth noting that the hydraulic force on the piston is generally
much greater than the force from the torsion bar. Considering the importance
of the hydraulics, it is possible to achieve good accuracy with this model as well.
One possible implementation of a valve function that compensates for the hysteresis is modelling the constants in the valve function similarly to the efficiency
constants in (6.14). For both constants k1 and k3 in (5.18) this give
k̇i
= k̂i − ki Ki
(6.21)
6.4
61
Hysteresis
where
k̂i







=





ki,1
ki,2
ki,3
ki,4
Ṫtb
Ṫtb
Ṫtb
Ṫtb
≥ 0,
< 0,
≥ 0,
< 0,
Ttb
Ttb
Ttb
Ttb
≥0
≥0
<0
<0
(6.22)
and
d
Ṫtb = Ṫˆtb − Ṫtb Kdot .
dt
(6.23)
Similarly to the efficiency constants, the constants ki can be found using (6.9) by
linear regression.
6.4.3
Hammerstein-Wiener Models
A third approach for modelling the hysteresis in the power steering gear is to
estimate a Hammerstein-Wiener model. As this black box modelling approach
combines a linear filter with an input and output non-linearity it might capture the system behaviour. Models with only an input non-linearity or output
non-linearity belong to the subset of the Hammerstein-Wiener models known as
Hammerstein models and Wiener models respectively.
Although the Hammerstein-Wiener models are dynamical models, the static data
from Data Sets 1-3 are used to estimate the models. This is far from optimal and
can lead to erroneous conclusions. However, as only static data for the power
steering gear is available, this methodology is used under the assumption that
the dynamic behaviour resembles the static behaviour. To ensure the validity of
the model, it is essential that the final model is validated using dynamic measurement data. The estimation of the models is performed using the System Identification Toolbox in MATLAB.
Experimentation shows that using piecewise linear functions as non-linearity is
a suitable model. Empirically the Wiener model
Tout = hhw (Tin ) = h G(q)Tin
(6.24)
where
h(x)










=








a0 x + b 0
a1 x + b 1
..
.
an−1 x + bn−1
an x + bn
x ≤ x0
x1 < x ≤ x2
..
.
,
(6.25)
xn−1 < x ≤ xn
xn < x
G(q) is an LTI filter, and n = 4 is a good choice. A simplified explanation of why
this model works is that the function works by estimating the influence of previous inputs by using the linear filter and then compensating for the asymmetrical
62
6
Data Analysis and Parameter Estimation
behaviour using the non-linearity.
To capture the system behaviour, two Wiener models are implemented – one on
the input shaft of the power steering gear and one on the output shaft. Under the
assumption that the valve constant k3 = 0 (empirically shown to be an insignificant simplification), the input shaft function, hhw,in can be estimated from the
valve characteristics data set using the pseudo-measurements from (6.9). Finally,
the output shaft function, hhw,out can be found using the relationship
(
Ttb = hhw,in (Tin )
(6.26)
Tout = hhw,out (rFout )
leading to
Tout
.
= hhw,out (r(fhyd + n1 hhw,in (Tin )))
(6.27)
for the measurements from Data Set 3 – Hydraulic Efficiency. For simulation in
continuous time the estimated linear filters are converted from discrete to continuous time using the MATLAB function d2c.
6.5
Friction
As the friction depends on angular velocity, it obviously cannot be identified using static data. Consequently, the Input Analysis and Reversible torque data sets
are the only rig data used to estimate the friction coefficients. In addition to
these data sets, the Stationary Vehicle Measurements can be used to estimate the
friction.
Some previous studies have been made on the friction in the steering system of
Scania trucks. In Rothhämel [2013] the steering system of a truck was disassembled and then the friction was measured for varying degrees of re-assembly.
Figure 6.15 shows the friction of the input shaft and the incremental friction resulting from the addition of the power steering gear and front axle respectively.
The measurement of the friction is done in steady state at the steering wheel and,
consequently, the friction in Figure 6.15 is not necessarily identical to the one
actually acting on the piston and king pin. As can be seen in Figure 6.15, the
friction on the input shaft and king pin is almost constant, i.e. Coulomb friction.
On the other hand, the friction affecting the piston demonstrates a clearly viscous
behaviour.
6.5.1
Linear Friction Model
Considering that there is no hydraulic assist in the Input Analysis data set, the
system is approximately linear. If the hysteresis is neglected, the gravitational
force is considered an input, and the friction is modelled as viscous, the state
space model (4.35) can be simplified into the linear representation of the system
63
Friction
Normalized force
6.5
Input Friction
1
0.5
0
Normalized force
0
0.2
0.6
0.8
1
0.8
1
0.8
1
Normalized velocity
Piston Friction
1
0.5
0
0
Normalized force
0.4
0.2
0.4
0.6
Normalized velocity
King Pin Friction
1
0.5
0
0
0.2
0.4
0.6
Normalized velocity
Figure 6.15: The top plot shows the friction friction force at different angular
velocities when the steering column is isolated. The middle and bottom plot
shows the incremental friction resulting from adding the power steering gear
and wheels respectively.
given in
ẋ =
 Binput
− J
 input

 ktb


 0

1
−J
1
input
0
− kntb
1
nmeq
− mp
eq
0
0

 1
0
 Jinput


 0

0

 x +  0

0


 0
0
0
B


 "
 T #
 in

 mp g
cos(θpsg ) 

−
0
0
0
(6.28)
meq
where
   
x1  ωsc 
x   T 
x =  2  =  tb  .
x3   vp 
   
x4
θsc
(6.29)
In this model the only unknowns are the gravitational force scaling – resulting
from the mounting angle of the power steering gear – and the damping coefficients of the input shaft and piston. In this linear grey box model, the parameter
values can be estimated numerically by e.g. the Gauss-Newton algorithm using
the measurement function
y=
h
0
0
0
i
1 x .
This algorithm is implemented in the MATLAB function greyest.
(6.30)
64
6.5.2
6
Data Analysis and Parameter Estimation
Non-linear Friction Models
To improve the friction model, a non-linear model can be implemented. In the
measurements from Data Set 6, the steering angle does not return to zero when
it is released from a non-zero initial value with no steering torque. This implies
there is non-negligible static friction in the wheels. In MATLAB’s Simscape the
non-linear friction model described in Figure 6.16 is included. This model includes static friction with a breakaway force, Coulomb friction (constant friction),
and viscous friction. The Stribeck friction is the initial decrease of the friction
that can be seen in the figure. For computational reasons, the function is implemented without a discontinuity as shown in Figure 6.17 where the threshold
velocity is a small number. MATLAB’s implementation of this friction model is
given in
fµ


(FC + (Fbrk − FC ) · e−cv |v| ) · sgn(v) + Bv




=

−cv |v| )


 v (Bvth +(FC +(Fbrk −FC )e
v
th
|v| ≥ vth
(6.31)
|v| < vth
where B is the viscous friction coefficient, cv is a coefficient affecting the transition
velocity FC is the Coulomb friction, Fbrk is the breakaway friction, and vth is
the threshold velocity. The rotational friction block is implemented analogously.
Using this friction model the Coulomb and viscous friction models suggested by
Qiu et al. [2001] and Wu et al. [2015] can be implemented.
F
Stribeck friction
Breakaway friction
Viscous friction
Coulomb friction
v
Figure 6.16: General friction model used by MATLAB. The friction force
(y-axis) is shown as a function of velocity (x-axis).
In Ikjema [1999] the friction on the steering column of the truck is modelled as
non-linear with a constant friction torque for large velocities and viscous friction
6.5
65
Friction
F
v
Threshold velocity
Figure 6.17: General friction model used by MATLAB with linear region to
remove the discontinuity. The friction force (y-axis) is shown as a function
of velocity (x-axis).
torque for small velocities. One possible implementation of this non-linearity, is
a havsin-function that depends on the angular velocity as in
τµ,sc = havsin(ω, −ω̄, ω̄, −c̄, c̄)
(6.32)
where

c̄1 ,
ω > ω̄1 ,







 c̄0 +c̄1
havsin(ω, ω̄0 , ω̄1 , c̄0 , c̄1 ) = 
( 2 ) + ( c̄1 −2 c̄0 ) · sin π ω̄ω−−ω̄ω̄0 − π2 , −ω̄ < ω < ω̄, .

1
0






−c̄0 ,
ω < ω̄0
(6.33)
The maximum amplitude of the output is given by c̄. The transition velocity
between linear and constant output is given by ω̄.
Ikjema [1999] also develops a specific friction model for the friction on the output
shaft of the power steering gear. This friction is scaled so that the friction is
largest when the sector arm angle is small and decreases for larger angles. This
model is given in
τµ,da = A · cos(arctan(B · θda )) · havsin(ωda ).
(6.34)
Furthermore, Ikjema [1999] develops a model of the king pin friction. In this
model the friction is modelled as Coulomb friction but, analogously with the
66
6
Data Analysis and Parameter Estimation
Simscape implementation, eliminates the discontinuity by creating a linear region around zero angular velocity. The model is described in
τµ,w


τ̄µ,w




=

τ̄


 v vµ,w
th
6.5.3
|v| ≥ vth
(6.35)
.
|v| < vth
Proposed Friction Model
Normalized force
In this thesis several different friction models are tested and compared. The proposed friction models are presented in Table 6.7 and their general appearance in
Figure 6.18. Firstly, viscous models are tried for all friction functions. Secondly,
MATLAB’s non-linear implementation, combining Coulomb friction and viscous
friction, is tried to mimic the behaviour in Figure 6.15. Thirdly, the models developed in Ikjema [1999] are tried.
Input friction
2
Viscous
Coulomb
Coulomb-Viscous
Havsin
1
0
Normalized force
0
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Normalized velocity
Piston friction
2
Viscous
Coulomb-Viscous
Havsin
1
0
0
Normalized force
0.1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Normalized velocity
King pin friction
2
Viscous
Coulomb
Coulomb-Viscous
1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Normalized velocity
Figure 6.18: The three plots show the proposed friction models for the input
shaft, piston, and king pin respectively. The figure details how the friction
force (y-axis) depends on the velocity (x-axis).
The havsin-function can be used to implement both Coulomb friction and Viscous friction. This is accomplished by tuning of of ω̄0 and c̄. For Coulomb friction, the transition velocity, determined by ω̄0 should ideally by zero, but for
computational reasons it should be chosen as a non-zero small number. Similarly,
for semi-viscous friction the transitional velocity should be chosen large.
Identification of parameters in non-linear models is done by a grid search aiming
to minimize the mse cost function. The initial parameter estimates are based on
the linear damping coefficients previously found.
6.6
67
External Influence
Table 6.7: Proposed friction models.
Model Name
Description
Input friction
Viscous
Coulomb
Coulomb-Viscous
Havsin
Viscous friction model.
Model implemented using the non-linear friction in MATLAB. Coulomb friction without discontinuity.
Model implemented using the non-linear friction in MATLAB. Combines Coulomb and viscous friction.
Model based on Ikjema [1999]. Viscous for low
ωlsc and constant for large ωlsc .
Piston friction
Viscous
Coulomb-Viscous
Havsin
Viscous friction model.
Model implemented using the non-linear friction in MATLAB. Combines Coulomb and viscous friction.
Model based on Ikjema [1999]. Viscous for low
ωlsc and constant for large ωlsc . Friction decreases with increasing drop arm angles.
King pin friction
Viscous
Coulomb
Coulomb-Viscous
6.6
Viscous friction model.
Model implemented using the non-linear friction in MATLAB similar to the model suggested
by Ikjema [1999]. Coulomb friction without discontinuity.
Model implemented using the non-linear friction in MATLAB. Combines Coulomb and viscous friction.
External Influence
In this section the influence from the self aligning torque is discussed. As the
vehicle is stationary in the measurements of Data Set 6, the total aligning moment
from the lateral tire forces is zero. Consequently Equation (4.36) simplifies into
Tdla = J δ̈ + hµ,wh + Tj .
(6.36)
In this equation, only the steering angle is known from measurements. Since
the steering torque in the drag link arm, Tdla , depends both on the steering
system and the external forces, it cannot be directly calculated from the driver
torque. However, to find initial estimates of the unknown parameters and func-
68
6
Data Analysis and Parameter Estimation
tions, the steering torque can be approximated by its static relationship with the
input torque when the steering angle is fixed. Furthermore, the steering angle
derivatives can be found by numerical differentiation. Just as previously, the
differentiation is done by the Euler approximation given in (3.14) and the filter
is a low-pass, fifth order, zero-phase, Butterworth filter with cut-off frequency
fc = 5H z.
Using this static approximation to find the steering torque from the input torque,
initial parameter estimates can be found by linear regression using either
Tin = J δ̈ + beq δ̇ + Tc sgn(δ̇) + Cj sin(δ)
(6.37)
or
Tin = J δ̈ + beq δ̇ + Tc sgn(δ̇) + Cj
sin(δ)
sgn(δ)
δ
(6.38)
depending on jacking torque model. In these equations, beq and Tc is a total
viscous and coulomb friction, respectively, and Cj is given by
Cj = Fz
eof f
2
sin(2θK P I )
(6.39)
and is almost constant.
Following this initial estimate, the model is continuously refined – by a grid
search with a mse cost function – together with the friction models where beq
and Tc are used as guidance for the friction models. As there is no measurement
data where any of these constants can be separated they are all estimated at the
same time.
7
Results
In this chapter the models previously developed are validated and compared using measurement data. As the measurement data presented in Chapter 6 are
either for the power steering gear in isolation or for the total steering system, this
chapter will be structured according to these two categories.
In accordance with the purpose of this thesis, models are compared for accuracy
and computability. These concepts are measured using the normalized root mean
square error, nrmse,
nrmse =
kŷ − yk
ky − ȳk
(7.1)
as a proxy for accuracy and simulation time as a proxy for computability. The
computability is only calculated for the complete vehicle simulations as the simulation time for static measurements does not correspond well to the times for
dynamic simulations.
As it is the relative simulation time and relative nrmse that is interesting for
each model combination, the nrmse and simulation time will be normalized for
each comparison table. In practice this implies that the fastest simulation will be
given relative simulation time 1, the most accurate model will be given relative
nrmse 1, and the others are scaled accordingly.
Although normalizing the nrmse and simulation time facilitates the comparison
of model combinations, it also makes it impossible to determine whether the
model is good in absolute terms. In order to show the absolute accuracy and
efficiency, the absolute nrmse and simulation time is given for the final model
and when nrmse is discussed in the text.
69
70
7.1
7
Results
Static Behaviour and Hysteresis
In this section the result from the estimation and validation using Data Set 1-3
is presented. Unfortunately, there is no validation data for these data series and
comparisons are therefore made using estimation data.
7.1.1
Static Gain
As stated previously, the gear ratio of the power steering gear nr , is approximately
constant and known for small to medium angles. Ideally this gear ratio, as a
static gain, would describe the input-output relationship for the static data of
Data Set 1. However, as stated in Section 6.4 this model has poor predictive ability due to torque losses in the power steering gear. Similarly, the least squares
estimate of the static gain has low predictive ability due to the hysteresis phenomena. The measurements are compared to simulations using these two choices of
static gains in Figure 7.1. The simulations using ideal and estimated gain have
nrmse of 0.464 and 0.19 respectively.
Simulation of Data Set 3, for the same ideal case with no torque loss, results in
Figure 7.2. As can be seen in the figure, there are considerable torque losses in
the system and the nrmse for this data is 0.492 for the pressure and 0.448 for the
output torque.
7.1.2
Efficiency Constants
Following the estimation approach suggested in Section 6.4.1, efficiency constants
are estimated for (6.14), (6.15), and (6.16) and subsequently used for simulation.
Simulation using the identified efficiency constants, without taking any system
dynamics into account, results in the result showed in Figure 7.3. As the figure
shows, there is generally good correspondence (nrmse 0.049) between the measurement data and simulation. The notable exception is given by the maximum
value for the torque where the efficiency constant changes value due to a change
of state. Considering that the efficiency constants transition will be smoother
when the steering dynamics change gradually, this is natural.
To assess the models true fit to measurement data, the system is simulated dynamically where the input is kept constant between samples. In this simulation the
steering dynamics results in a natural low pass filter and the simulation showed
in Figure 7.4 with a slightly improved nrmse of 0.043.
In accordance with the method described in Section 6.4.1, efficiency constants
are estimated for the pseudo measurements of the torsion bar torque. As can be
seen in Figure 7.5, the efficiency constants capture the hysteresis behaviour in the
pseudo measurements with nrmse 0.06. Furthermore, the simulated pump pressure closely resembles the measurements but with considerable noise for large
pump pressures resulting in a nrmse of 0.268.
Following identification of all three sets of efficiency constants, the model is simulated for Data Set 3. The result of this simulation is shown in Figure 7.6. As
7.1
71
Static Behaviour and Hysteresis
1.5
Measurement
True gear ratio
Estimated gear ratio
1
Normalized torque
0.5
0
-0.5
-1
-1.5
0
20
40
60
80
100
120
Time [s]
Normalized pressure
Figure 7.1: Simulations of Data Set 1 with static gains. Simulation with the
true gear ratio and an estimated gear ratio is compared to the estimation
data. The y-axis shows the output torque, normalized with the maximum
absolute value of the measurement data, as a function of time.
1
0.5
0
-0.5
Measurement
Simulation
-1
0
200
400
600
800
1000
1200
800
1000
1200
Time [s]
Normalized output torque
1.5
1
0.5
0
-0.5
-1
-1.5
0
200
400
600
Time [s]
Figure 7.2: Simulation of Data Set 3 with the ideal static model. The simulated output pressure and output torque is compared to the estimation data
in the top and bottom plot, respectively. The y-axis is normalized with the
maximum absolute value of the measurement data.
72
7
Results
Normalized output torque
1
0.5
0
-0.5
-1
Measurement
Simulation
-1.5
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Normalized input torque
Normalized output torque
1
0.5
0
-0.5
-1
-1.5
0
20
40
60
80
100
120
Time [s]
Figure 7.3: Simulations of Data Set 1 with efficiency constants but no system
dynamics. The simulated output torque (y-axis) is compared to the estimation data with input torque on the top x-axis and time on the bottom x-axis.
The y-axis and top x-axis are normalized with the maximum absolute value
of the measurement data.
Normalized output torque
1
0.5
0
-0.5
-1
Measurement
Simulation
-1.5
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Normalized input torque
Normalized output torque
1
0.5
0
-0.5
-1
-1.5
0
20
40
60
80
100
120
Time [s]
Figure 7.4: Simulations of Data Set 1 with efficiency constants and with system dynamics. The simulated output torque (y-axis) is compared to the estimation data, with input torque on the top x-axis and time on the bottom
x-axis. The y-axis and top x-axis is normalized with the maximum absolute
value of the measurement data.
7.1
Static Behaviour and Hysteresis
73
the figure shows, the simulation has reasonable fit to the measurement data with
pressure nrmse 0.263 and output torque nrmse 0.204. However, considering
that there is no separate validation data for these data sets, it is possible that the
model suffers from over-adaptation to this measurement set or this type of input
and spectral content. This risk is partially mitigated by the fact that this model
has a physical interpretation; the efficiency constants correspond to the resistance
in the steering system.
7.1.3
Hydraulic Adaptation
As an alternative to modelling the non-linear inefficiencies at the input shaft
and at the sector shaft of the power steering gear, the valve can be modelled
to compensate for the inefficiencies. This approach leads to different constants
estimated in (6.21), (6.22), and (6.23).
Simulations with the hydraulic adaptation results in Figure 7.7 with good results,
0.249 nrmse to measurement data of the hydraulic pressure and 0.154 nrmse
of the output torque. However due to variations in the valve, there is a risk of
over-adaptation to a specific power steering gear or to the data set itself.
7.1.4
Wiener Model
Analogously to the efficiency constant approach, first the hysteresis of the input
is estimated using the pseudo-measurements in (6.24) and (6.25). Simulating the
estimated model gives Figure 7.8. The simulation has good correspondence with
the estimation data but there is a slight offset leading to too low pump pressure
for positive torques and too high for negative torques. This gives a nrmse compared to the pseudo measurements of 0.113 and nrmse 0.293 compared to the
pressure data.
Using the input function from Figure 7.8 and the data from Data Set 3, the output
hysteresis is estimated and then simulated. The simulation results can be found
in Figure 7.9. Obviously the offset from the input function is still apparent in this
simulation with nrmse 0.293. However, it is interesting that the simulation has
very good nrmse 0.048 compared to the output torque. This implies the output
function compensates for the bias introduced by the input function.
Due to the lack of validation data, it is more probable that the Wiener approach
suffers from over-adaptation to the measurement data than, for example, the efficiency constants; the functions derived from the Wiener model lack the clear
physical interpretation of the efficiency constants. In order to diminish this risk,
each linear filter is selected with fewer than 4 poles and 4 zeros. It is also problematic that the dynamic Wiener models are estimated using static data. As there
is no other data available, this method is still used which increases the risk for
over-adaptation.
74
7
Results
Pseudo-measurements
3
Measurement
Simulation
2
1
0
-1
-2
-3
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Normalized input torque
Normalized pump pressure
1.2
1
0.8
0.6
0.4
0.2
0
0
200
400
600
800
1000
1200
Time [s]
Normalized pressure
Figure 7.5: Simulation of Data Set 2 with efficiency constants. The simulated
pump pressure (y-axis) is compared to the estimation data in the bottom plot
and the simulated pseudo-measurements (y-axis) are compared to the estimation data in the top plot with input torque on the x-axis. The y-axis and
top x-axis are normalized with the maximum absolute value of the measurement data.
1
0.5
0
Measurement
Simulation
-0.5
0
200
400
600
800
1000
1200
800
1000
1200
Normalized output torque
Time [s]
1
0.5
0
-0.5
-1
0
200
400
600
Time [s]
Figure 7.6: Simulation of Data Set 3 with efficiency constants. The simulated
output pressure and output torque is compared to the estimation data in
the top and bottom plot, respectively. The y-axis is normalized with the
maximum absolute value of the measurement data.
7.1
75
Static Behaviour and Hysteresis
Normalized pressure
1
0.8
0.6
0.4
0.2
0
0
200
400
600
800
1000
1200
Time [s]
Normalized output torque
1
0.5
0
-0.5
Measurement
Simulation
-1
0
200
400
600
800
1000
1200
Time [s]
Figure 7.7: Simulation of Data Set 3 with the Hydraulic Adaptation model.
The simulated output pressure and output torque is compared to the estimation data in the top and bottom plot, respectively. The y-axis is normalized
with the maximum absolute value of the measurement data.
Pseudo-measurements
3
Measurement
Simulation
2
1
0
-1
-2
-3
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Normalized input torque
Normalized pump pressure
1.2
1
0.8
0.6
0.4
0.2
0
0
200
400
600
800
1000
1200
Time [s]
Figure 7.8: Simulation of Data Set 2 with the Wiener Model. The simulated
pump pressure (y-axis) is compared to the estimation data in the bottom plot
and the simulated pseudo-measurements (y-axis) are compared to the estimation data in the top plot with input torque on the x-axis. The y-axis and
top x-axis are normalized with the maximum absolute value of the measurement data.
76
7.1.5
7
Results
Comparison of Hysteresis Models for Static Data
Table 7.1 summarizes the result from the simulations of Data Set 3. As can be
seen from the simulation results, there is truly non-negligible torque losses in
the power steering gear that need to be modelled. Best fit to the pressure data
is obtained through the hydraulic adaptation, followed by the efficiency constant
and the Wiener approach. Regarding the output torque, the best fit is obtained
from the Wiener approach even though both the efficiency constants and the hydraulic adaptation give acceptable results.
For all these approaches there is a risk for over-adaptation to static data in general
and these measurement series in particular. As the Wiener models is a black
box approach and estimated using static data, this is extra vulnerable to overadaptation.
Table 7.1: Comparison of Hysteresis Models for static data from Data Set 3.
The nrmse is normalized.
7.2
Modelling Approach
Pressure nrmse
Torque nrmse
Ideal Model
Efficiency Constants
Hydraulic Adaptation
Wiener
1.98
1.06
1.00
1.18
9.31
4.29
3.20
1.00
Dynamic Behaviour and Full Vehicle Steering
System Model
For the full vehicle data from Data Set 6, both simulation time and fit is calculated. However, as there is only measurement of input torque and steering angle,
is is not possible to isolate the friction, self-aligning torque, or hysteresis model.
To avoid an unnecessary amount of figures, all possible modelling alternatives
are not shown. Generally all modelling options but one are kept constant for any
given comparison. For example, the hysteresis models are compared using one
model for the self-aligning torque – not every self-aligning torque model. However, where considerable interdependence between modelling options is noticed,
the best combination for each alternative is shown. An example is the hysteresis
models and input friction models. Both these modelling options aim to model
the resistance on the input shaft and from the grid search it can be seen that the
hysteresis models work best with different input friction models.
Since Data Set 6 is designed specifically for this thesis, the input signals are varied
and excite a large range of frequencies. To ensure that the models developed are
not good only for one specific input signal, all comparisons are made using the
two distinct measurement series Data Set 6.1 and Data Set 6.2 (belonging to Data
Set 6). As these measurement series, shown in Figure 7.10 and 7.11 respectively,
7.2
77
Dynamic Behaviour and Full Vehicle Steering System Model
contain fundamentally different frequencies, this decreases the risk that models
are over-adapted. To further decrease this risk, the measurement series are split
into estimation and validation data.
7.2.1
Comparison of Hysteresis Models for Dynamic Data
In Table 7.2, the results of the simulations with different hysteresis models are
summarized and the results are plotted in Figure 7.12 and 7.13. As stated above
there is considerable interdependence between the hysteresis model and input
friction model and therefore the input friction has been adapted to each hysteresis model individually.
The best fit is obtained from the ideal model but the efficiency constants and
the Hydraulic Adaptation model also has reasonable fit. The Wiener model in
no way manages to follow the dynamic data for neither data set and it must be
concluded that the model is both over-adapted to static data in general and the
specific estimation data in particular.
Regarding the simulation time, the ideal model, which contain the fewest states,
naturally simulates the fastest. However, both the efficiency constant model and
the Hydraulic Adaptation model have acceptable simulation speeds as well. In
addition to having less than satisfying fit to the measurement data, the Wiener
model is the slowest due to a larger number of states.
Table 7.2: Comparison of Hysteresis Models for dynamic data from
Data Set 6. The nrmse is normalized.
Modelling Approach
Ideal Model
Efficiency Constants
Hydraulic Adaptation
Wiener
7.2.2
Data Set 6.1
Relative Time nrmse
1.00
1.97
1.40
5.24
1.00
1.00
1.10
3.30
Data Set 6.2
Relative Time nrmse
1.04
1.00
1.32
2.71
1.00
1.01
1.36
1.87
Comparison of Models for the Jacking Torque
In Section 4.6, two models for the jacking torque were presented, (4.38) and
(4.41). As there is no total aligning moment for these measurements this is the
complete models. In Table 7.3 the results are presented. As can be seen in the
table and Figure 7.14, the Sinc Model, which applies a stronger torque at low
steering angles, achieves slightly better fit for one of the data sets and the Sinus
Model for the other. The simulation time for the two models are the same. However, to achieve this fit with the Sinus Model, the constants have to be selected
larger than they should be physically. Likewise, the parameter in the Sinc Model
needs to be adapted to the measurement data but to a lesser extent.
78
7
Results
Normalized pressure
1.2
1
0.8
0.6
0.4
0.2
0
0
200
400
600
800
1000
1200
Time [s]
Normalized torque
1
0.5
0
-0.5
Measurement
Simulation
-1
0
200
400
600
800
1000
1200
Time [s]
Figure 7.9: Simulation of Data Set 3 with the Wiener model.The simulated
output pressure and output torque is compared to the estimation data in
the top and bottom plot, respectively. The y-axis is normalized with the
maximum absolute value of the measurement data.
Input torque at the input shaft
Normalized torque
1
0.5
0
-0.5
-1
0
2
4
6
8
10
12
14
16
18
20
14
16
18
20
Time [s]
Left wheel angle
Normalized angle
1
0.5
0
-0.5
-1
0
2
4
6
8
10
12
Time [s]
Figure 7.10: Data Set 6.1. The measurement is made for dynamic data with
activated hydraulics on a stationary vehicle. The input torque is applied at
the input shaft and is displayed in the top plot. The resulting wheel angle of
the left wheel is displayed in the bottom plot. The y-axis is normalized with
the maximum absolute value of the respective signal.
7.2
79
Dynamic Behaviour and Full Vehicle Steering System Model
Input torque at the input shaft
Normalized torque
1
0.5
0
-0.5
-1
0
2
4
6
8
10
12
14
16
18
20
14
16
18
20
Time [s]
Left wheel angle
Normalized angle
1
0.5
0
-0.5
-1
0
2
4
6
8
10
12
Time [s]
Figure 7.11: Data Set 6.2. The measurement is made for dynamic data with
activated hydraulics on a stationary vehicle. The input torque is applied at
the input shaft and is displayed in the top plot. The resulting wheel angle of
the left wheel is displayed in the bottom plot. The y-axis is normalized with
the maximum absolute value of the respective signal.
Data Set 6.1
Normalized angle
1
0.5
0
-0.5
-1
-1.5
0
2
4
6
8
10
12
14
16
18
20
14
16
18
20
Time [s]
Data Set 6.2
Normalized angle
1
0
-1
Measurement
Ideal model
Efficiency constants
Weiner
-2
-3
0
2
4
6
8
10
12
Time [s]
Figure 7.12: Comparison of Ideal, Efficiency Constant, and Wiener models
for dynamic data. The simulated left wheel angle (y-axis) is compared to
validation data for two different data sets in the two plots. The y-axis is
normalized with the maximum absolute value of the measurement data.
80
7
Data Set 6.1
1
Normalized angle
Results
Measurement
Ideal model
Hydraulic adaptation
0.5
0
-0.5
-1
-1.5
0
2
4
6
8
10
12
14
16
18
20
12
14
16
18
20
Time [s]
Data Set 6.2
Normalized angle
1
0.5
0
-0.5
-1
-1.5
0
2
4
6
8
10
Time [s]
Figure 7.13: Comparison of the Hydraulic Adaptation model and Ideal
model for dynamic data. The simulated left wheel angle (y-axis) is compared
to validation data for two different data sets in the two plots. The y-axis is
normalized with the maximum absolute value of the measurement data.
Data Set 6.1
Normalized angle
1
0.5
0
-0.5
-1
0
2
4
6
8
10
12
14
16
18
20
16
18
20
Time [s]
Data Set 6.2
Normalized angle
1
0.5
0
-0.5
-1
Measurement
Sinus
Sinc
-1.5
-2
0
2
4
6
8
10
12
14
Time [s]
Figure 7.14: Comparison of jacking torque models for dynamic data. The
simulated left wheel angle (y-axis) is compared to validation data for two
different data sets in the two plots. The y-axis is normalized with the maximum absolute value of the measurement data.
7.2
81
Dynamic Behaviour and Full Vehicle Steering System Model
Table 7.3: Comparison of Jacking Torque Models for dynamic data from
Data Set 6. The nrmse is normalized.
Modelling Approach
Sinus Model
Sinc Model
7.2.3
Data Set 6.1
Relative Time nrmse
1.00
1.04
1.22
1.00
Data Set 6.2
Relative Time nrmse
1.00
1.17
1.00
1.14
Comparison of Friction Models
Input Friction
Following iterative use of the grey-box estimation approach for Data Set 4 in
(6.28), described in Section 6.5.1, the estimation converges and results in the
simulation output in Figure 7.15. As is apparent in the figure the identified coefficients give good nrmse 0.094 to estimation data. However, as the friction
depends on the angular velocity which is constant for this measurement set, this
result is only valid for this particular angular velocity. It is worth mentioning
that the friction estimate from this approach also compensates for the hysteresis
that is omitted from this linear model.
When the estimated Viscous model is used to simulate Data Set 6 that has faster
dynamics, the estimated friction is too high. Similarly, iterative estimation of viscous models for Data Set 6.1 and Data Set 6.2 shows that Data Set 6.1 with faster
dynamics requires less damping than Data Set 6.2. Consequently, the friction
over angular velocity ratio has to decrease as the angular velocity increases. This
behaviour can be simulated both using the Havsin model, (6.32) suggested by
Ikjema [1999] and a Coulomb-Viscous friction according to (6.31).
Iterative estimation of Havsin Input Friction Models and Coulomb-Viscous Models result in the model output shown in Figure 7.16 and summarized in Table 7.4.
The Coulomb-Viscous model demonstrates the same, but less expressed, tradeoff between the fast and slow dynamics as the Viscous model. To a certain extent
this is also true for the Havsin Model but by fine-tuning the parameters much
of this behaviour can be eliminated. The Coulomb Model, as indicated by Rothhämel [2013], causes the opposite trade off; more friction is required for the less
dynamic Data Set 6.2.
The parameter values in the Coulomb-Viscous and Havsin models depend heavily on the hysteresis model. In comparison to the ideal model, the efficiency constant model requires less Coulomb friction. This is natural as the torque losses
modelled by the efficiency constants are probably caused, to a large extent, by
friction.
Piston Friction
Table 7.5 presents a comparison of friction models that are estimated for dynamic
data. As can be seen in Table 7.5 and as Rothhämel [2013] suggested, a com-
82
7
1.5
Results
Measurement
Simulation
Normalized angle
1
0.5
0
-0.5
-1
0
100
200
300
400
500
600
700
Time [s]
Figure 7.15: Simulation of Data Set 4 with a Linear Friction model. The
simulated input shaft angle is compared to estimation data. The y-axis is
normalized with the maximum absolute value of the measurement data.
Data Set 6.1
Normalized angle
1
Measurement
Viscous
Coulomb
Coulomb-Viscous
Havsin
0.5
0
-0.5
-1
0
2
4
6
8
10
12
14
16
18
20
12
14
16
18
20
Time [s]
Data Set 6.2
Normalized angle
1
0.5
0
-0.5
-1
0
2
4
6
8
10
Time [s]
Figure 7.16: Simulation of Data Set 6 with different input friction models.
The simulated left wheel angle (y-axis) is compared to validation data for
two different data sets in the two plots. The y-axis is normalized with the
maximum absolute value of the measurement data.
7.2
83
Dynamic Behaviour and Full Vehicle Steering System Model
Table 7.4: Comparison of Input Friction Models for dynamic data from
Data Set 6. The nrmse is normalized.
Modelling Approach
Viscous Model
Coulomb Model
Coulomb-Viscous Model
Havsin Model
Data Set 6.1
Relative Time nrmse
1.00
1.17
1.16
1.01
1.85
1.49
1.70
1.00
Data Set 6.2
Relative Time nrmse
1.00
1.40
1.88
1.82
1.26
1.00
1.46
1.23
pletely viscous model gives good fit to validation data. Adding Coulomb friction
implemented according to the Matlab function, does not improve the accuracy
of the simulation and increases simulation time. Similarly, adapting the Havsin
model for sector arm friction developed in Ikjema [1999] to the piston does not
improve the simulation either. The results from these experiments are shown in
Figure 7.17.
Table 7.5: Comparison of piston friction models for dynamic data from
Data Set 6. The nrmse is normalized.
Modelling Approach
Viscous Model
Coloumb Model
Coloumb-Viscous Model
Havsin Model
Data Set 6.1
Relative Time nrmse
1.00
1.59
1.12
1.17
1.00
1.08
1.00
1.46
Data Set 6.2
Relative Time nrmse
1.00
1.59
1.10
1.03
1.10
1.00
1.10
1.22
King Pin Friction
When adapting the viscous wheel friction model to the estimation data it is clear
that the wheel friction has a rather small impact on the model accuracy. Consequently, exchanging the friction model for a Coulomb friction model does not
greatly affect the simulation – except for increasing the simulation time. The results for the different friction models can be found in Table 7.6 and Figure 7.18.
Table 7.6: Comparison of king pin friction models for dynamic data from
Data Set 6. The nrmse is normalized.
Modelling Approach
Viscous Model
Coulomb Model
Data Set 6.1
Relative Time nrmse
1.00
1.39
1.02
1.00
Data Set 6.2
Relative Time nrmse
1.00
1.57
1.00
1.10
84
7
Data Set 6.1
1.5
Normalized angle
Results
Measurement
Viscous
Coulomb-Viscous
Havsin
1
0.5
0
-0.5
-1
0
2
4
6
8
10
12
14
16
18
20
12
14
16
18
20
Time [s]
Data Set 6.2
Normalized angle
1
0.5
0
-0.5
-1
0
2
4
6
8
10
Time [s]
Figure 7.17: Simulation and comparison to Data Set 6 with different piston friction models. The simulated left wheel angle (y-axis) is compared to
validation data for two different data sets in the two plots. The y-axis is
normalized with the maximum absolute value of the measurement data.
Data Set 6.1
Normalized angle
1.5
Measurement
Viscous
Coulomb
1
0.5
0
-0.5
-1
0
2
4
6
8
10
12
14
16
18
20
12
14
16
18
20
Time [s]
Data Set 6.2
Normalized angle
1
0.5
0
-0.5
-1
0
2
4
6
8
10
Time [s]
Figure 7.18: Simulation and comparison to Data Set 6 with different king
pin friction models. The simulated left wheel angle (y-axis) is compared to
validation data for two different data sets in the two plots. The y-axis is
normalized with the maximum absolute value of the measurement data.
7.3
85
Conclusions
7.2.4
Comparison of Solvers
In order to investigate how the choice of solver affects the accuracy and computability, the model was simulated using three different solvers, ode5, ode45,
and ode15s. As can be seen in Table 7.7 these tests unanimously suggest ode15s
is the most suitable solver. The simulation using ode45 achieved the same simulation result but required much longer time. When using ode5, the simulation
either reached maximum simulation time without finishing or failed, depending
on step size. This result implies that the state space model is stiff and requires an
implicit variable solver.
Table 7.7: Comparison of Solvers for dynamic data from Data Set 6. The
nrmse is normalized.
Modelling Approach
ode5
ode45
ode15s
7.3
Data Set 6.1
Relative Time nrmse
NaN
62.8
1.00
NaN
1.00
1.00
Data Set 6.2
Relative Time nrmse
NaN
127
1.00
NaN
1.00
1.00
Conclusions
Throughout this chapter there have been both accurate and less accurate models.
In this section the necessary choices and trade-offs are made that maximize the
accuracy without requiring too long simulation time.
In Section 4.5 two steering column models are developed with one and two moments of inertia respectively. As the torque for all measurement sets is measured
below the split between the upper and lower steering column, no comparison of
accuracy can be made using the measurement data available. In order to make
such a comparison it would be necessary to measure the torque applied at the
steering wheel in addition to the torque applied at the input shaft. As the added
flexibility is believed to have negligible impact on the model accuracy and some
impact on the computability, the simpler one moment of inertia steering column
model is selected. This decision is also supported by the review of mechanical
models in Dell’Amico [2013].
Regarding the hysteresis modelling, the best choice for dynamic data is considered the ideal model. Although at a clear disadvantage compared to other models
for static data, the ideal model has the best fit and simulation time for dynamic
data. The hysteresis that the ideal model fails to capture for static simulations is
compensated using the friction model for dynamic data. An additional benefit
of the ideal model is that it has fewer parameters and is less likely to be overadapted than the hysteresis models.
Although the jacking torque models have fundamental differences they both man-
86
7
Results
age to capture significant model behaviour. However, even though the Sinus
model has better theoretical support, its parameters need to be one order of magnitude larger than they should be physically to achieve this. On the other hand,
the Sinc Model has less support in the literature but its parameters are in the right
order of magnitude and it performs slightly better experimentally. Consequently,
as neither of the alternatives completely constitutes a physical representation of
the system, the Sinc model is chosen for its superior performance empirically.
Considering that the input friction model presented in Ikjema [1999] has better
fit than the alternatives and shorter simulation time, it is clearly the best choice
to use the Havsin friction model for the input friction. However, for the piston
and the wheels, the viscous model is relatively accurate and significantly faster
than the alternatives. Consequently, this is the modelling alternative of choice.
Finally, the choice of solver is facilitated greatly by the superiority of ode15s.
The final simulation result is shown in Figure 7.19. This model is rather accurate
for both validation series with nrmse 0.307 and 0.499, respectively. Furthermore,
the model can be simulated with acceptable speed as 100 seconds of data require
a simulation time of approximately 2 seconds.
As the model to a large extent is developed by physical modelling – as opposed
to system identification – the model can be adapted for some limited variations
of the steering system and truck configuration without having to estimate a completely new model. On the other hand, other modelled parts are specific to this
steering system. An example is the valve offset which will be distinct for every
single power steering gear. It is also possible that the estimated efficiency constants are only valid for this particular power steering gear.
7.3
87
Conclusions
Data Set 6.1
Normalized angle
1
0.5
0
-0.5
Measurement
Simulation
-1
0
2
4
6
8
10
12
14
16
18
20
12
14
16
18
20
Time [s]
Data Set 6.2
Normalized angle
0.4
0.2
0
-0.2
-0.4
-0.6
0
2
4
6
8
10
Time [s]
Figure 7.19: Simulation of the final model and comparison to Data Set 6.
The simulated left wheel angle (y-axis) is compared to validation data for
two different data sets in the two plots. The y-axis is normalized with the
maximum absolute value of the measurement data.
8
Summary
In this chapter the results from Chapter 7 are discussed and compared with the
purpose of the thesis. Additionally, future work is proposed and discussed. The
rest of this chapter is structured along these two topics.
8.1
Discussion
The underlying objective of this thesis is deepening the understanding of steering
systems in general and for heavy duty vehicles in particular. More specifically
this objective leads to the purpose of this thesis which is developing an accurate
and efficient physical model of the steering system. In pursuit of this purpose,
the modelling effort has included the basic mechanical connection and various
non-linear phenomena such as hysteresis, friction, and hydraulic assist.
As a result of the modelling, there is a simulation model achieving the necessary
accuracy. By also identifying the most important model constituents and simplifying the model, efficiency in the form of simulation time is achieved.
However, although a relatively accurate and efficient model of the steering system
has been identified, this is not enough to use for development of e.g. lka. While
driving, the steering system is merely a subsystem of the total truck which in turn
can be considered a subsystem of the truck-and-environment system. Considering this, the model describes the steering system but requires interaction from
other models or sensors to serve its primary purpose.
Regarding the model itself, it meets the requirements of accuracy and computability. However, the model is only better than the rest of the identified models – it
can in no way be considered optimal. It is possible that better results could be
89
90
8
Summary
obtained from the same modelling approach with better conditions, including
but not limited to a consistent measurement set-up, more dynamic data from
isolated subsystems, and more validation data. There are also other modelling
approaches that could be attempted. This would result in other models with
their own advantages and drawbacks.
Finally, it is worth discussing the generalizability of the results obtained. The
data used for estimation is from a few selected steering systems and the extent to
which these results and methodologies can be transferred to other trucks of the
same and different types is uncertain. It is probable that the model developed
in this thesis is best adapted to the particular truck studied. However, it is also
probable that the model would work rather well for a similar truck and that the
same methodology would be directly applicable. For different steering systems,
it is likely that some adaptations are necessary to find a well-working simulation model, but it is also likely this thesis would greatly facilitate that modelling
effort.
In summary, this thesis has resulted in a well-functioning simulation model and
thus fulfilled its purpose. In doing so it has also granted insight into the mechanics and function of steering systems in general. Consequently, this thesis has
made its own contribution to the emergence of autonomous driving.
8.2
8.2.1
Future Work
Suggestions for Model Accuracy Improvement
Some of the work done in this thesis has been developed based on measurement
data from isolated power steering gears where the power steering gears are not
identical in all measurements. This is a source of error that could be eliminated
by using rig data from the same power steering gear as the one from which full
vehicle data is recorded. The developed model would be further improved by
studying different power steering gears and their corresponding full vehicle data.
The development process would also be simplified and model accuracy would be
increased by studying more subsystems separately before identifying the fully assembled system. In the developed model, the power steering gear has been identified separately. Using the same methodology, the steering column and steering
linkage could also be isolated and identified. The drag link can advantageously
be assigned sensors that measure its position. As a result the force in the drag
link can be derived. This would also mean that the steering system dynamics
could be measured when decoupling all external effects and applying incremental modelling.
The used rig data does not contain a wide frequency spectrum on the input signal
since the rig data sets are not primarily designed for this purpose. Measuring
separate estimation and validation data is also suggested to design models that
are not over-adapted to the specific data set.
8.2
Future Work
8.2.2
91
Model Expansion
Throughout the development of the model, little emphasis has been put on finding a good vehicle model that models the self aligning torque, weight distribution
during turns and steering angle to yaw rate conversion. A complete model of the
lateral dynamics needs to consider such external influences.
The developed model in this thesis is a good model for a stationary vehicle without friction acting on the tires. How the wheel-tire friction affects a stationary
vehicle as well as vehicles at higher velocities need to be studied. The authors
recommend that additional measurements are recorded and used for such estimations. These measurements should advantageously include the input torque,
the wheel angle, the drag link position and yaw rate for a vehicle driving on a flat
surface at varying velocities.
Finally, the model can be further generalized and expanded by studying the steering system of different vehicle configurations. For example, the king pin friction
likely depends on the weight on the front axle – the importance of which has not
been studied in this thesis.
92
8
Summary
Bibliography
Santosh Ancha, Abhijit Baviskar, John R. Wagner, and Darren M. Dawson.
Ground vehicle steering systems: modelling, control, and analysis of hydraulic,
electric and steer-by-wire configurations. Int. J. Vehicle Design, 44(1):188–208,
2007. Cited on pages 3, 23, and 25.
Aly Badawy, Jeff Zuraski, Farhad Bolourchi, and Ashok Chandy. Modeling and
analysis of an electric power steering system. SAE Technical Paper 1999-010399, 1999. Cited on page 3.
Donald Bastow, Geoffrey Howard, and John P. Whitehead. Car Suspension and
Handling. SAE International, fourth edition, 2004. Cited on pages 5, 12, 33,
and 35.
Min Chul Shin, Seong Han Kim, Gwang Hee Cho, and Chong Nam Chu. Development of a steering-system model considering viscous friction and its verification. Journal of Automobile Engineering, 228(2):144–163, 2013. Cited on
pages 3 and 4.
Alessandro Dell’Amico. Modelling, Simulation and Experimental Investigation
of an Electrohydraulic Closed-Centre Power Steering System. Licentiate thesis,
Linköping University, 2013. Cited on pages 3, 4, and 85.
Alessandro Dell’Amico and Petter Krus. Modelling, simulation and experimental investigation of an electrohydraulic closed-centre power steering system.
IEEE/ASME Transaction on Mechatronics, 2015. Cited on pages 3 and 4.
Joel Eriksson. Technical report C108/309 (Internal Scania document). Scania,
Södertälje, 1998. Cited on page 30.
Fredrik Gustafsson. Statistical Sensor Fusion. Studentlitteratur, second edition,
2012. Cited on page 17.
Yung-Hsiang Judy Hsu. Estimation and Control of Lateral Tire Forces using Steering Torque. PhD thesis, Stanford University, 2009. Cited on pages 5 and 33.
Jolle Ikjema. Technical report C70/163 (Internal Scania document). Scania,
93
94
Bibliography
Södertälje, 1999. Cited on pages 5, 23, 25, 30, 37, 41, 42, 64, 65, 66, 67, 81,
83, and 86.
Lennart Ljung. System Identification: Theory for the User. PTR Prentice Hall,
second edition, 1999. Cited on pages 13 and 16.
Lennart Ljung and Torkel Glad. Modellbygge och Simulering. Studentlitteratur,
second edition, 2004. Cited on pages 14, 15, 16, 17, and 40.
Herbert E. Merritt. Hydraulic Control Systems. Wiley, first edition, 1967. Cited
on pages 37 and 39.
Kim Mårtensson. Alternative concepts for power steering in heavy vehicles. Master’s thesis, Royal Institute of Technology, Sweden, 2005. Cited on pages 3, 4,
23, 25, 29, and 37.
Oskar Nilsson. Modeling and validation of an electro hydraulic actuator system
for heavy vehicle applications. Master’s thesis, Linköping University, Sweden,
2014. Cited on pages 4 and 37.
P. E. Pfeffer, M. Harrer, and D. N. Johnston. Interaction of vehicle and steering
system regarding on-centre handling. Vehicle System Dynamics: International
Journal of Vehicle Mechanics and Mobility, 46(5):413–428, 2008. Cited on
pages 3 and 4.
H.C. Qiu, Q. Zhang, J.F. Reid, and D. Wu. Modelling and simulation of an electrohydraulic steering system. Int. J. of Vehicle Design, 26(2):161–174, 2001. Cited
on pages 3, 4, and 64.
Marcus Rösth. Hydraulic Power Steering System Design in Road Vehicles. PhD
thesis, Linköping University, 2007. Cited on pages 3, 4, 33, 37, 39, 40, 41,
and 42.
Malte Rothhämel. Technical report 7022190 (Internal Scania document). Scania,
Södertälje, 2013. Cited on pages 62 and 81.
Kristoffer Tagesson. Truck Steering System and Driver Interaction. Licentiate
thesis, Chalmers University of Technology, 2014. Cited on pages 3, 4, 5, 33,
and 35.
M. Tai, P. Hingwe, and M. Tomizuka. Modeling and control of steering system of
heavy vehicles for automated highway systems. Transactions on Mechatronics,
9(4):609–618, 2004. Cited on pages 3, 4, 23, 25, and 29.
L. G. Viana, G. Romero, J. Maroto, and J. Felez. Modelling and simulation of a
hydraulic power assisted steering system through bond graphs. Proceedings
of the International Workshop on Applied Modelling and Simulation, WAMS
2010, 2010. Cited on pages 3 and 4.
J.Y. Wong. Theory of Ground Vehicles. Wiley, fourth edition, 2008. Cited on page
5.
Bibliography
95
Liying Wu, Tongqiang Xu, and Tian Luo. Research and realization of simulation
for automobile steering resisting torque. Applied Mechanics and Materials,
733(1):745–748, 2015. Cited on pages 3, 4, 5, 33, 35, and 64.
ZF Lenksysteme. ZF Servocom (Technical brochure). 2015. Cited on page 10.
96
Bibliography
Upphovsrätt
Detta dokument hålls tillgängligt på Internet — eller dess framtida ersättare —
under 25 år från publiceringsdatum under förutsättning att inga extraordinära
omständigheter uppstår.
Tillgång till dokumentet innebär tillstånd för var och en att läsa, ladda ner,
skriva ut enstaka kopior för enskilt bruk och att använda det oförändrat för ickekommersiell forskning och för undervisning. Överföring av upphovsrätten vid
en senare tidpunkt kan inte upphäva detta tillstånd. All annan användning av
dokumentet kräver upphovsmannens medgivande. För att garantera äktheten,
säkerheten och tillgängligheten finns det lösningar av teknisk och administrativ
art.
Upphovsmannens ideella rätt innefattar rätt att bli nämnd som upphovsman
i den omfattning som god sed kräver vid användning av dokumentet på ovan
beskrivna sätt samt skydd mot att dokumentet ändras eller presenteras i sådan
form eller i sådant sammanhang som är kränkande för upphovsmannens litterära
eller konstnärliga anseende eller egenart.
För ytterligare information om Linköping University Electronic Press se förlagets hemsida http://www.ep.liu.se/
Copyright
The publishers will keep this document online on the Internet — or its possible replacement — for a period of 25 years from the date of publication barring
exceptional circumstances.
The online availability of the document implies a permanent permission for
anyone to read, to download, to print out single copies for his/her own use and
to use it unchanged for any non-commercial research and educational purpose.
Subsequent transfers of copyright cannot revoke this permission. All other uses
of the document are conditional on the consent of the copyright owner. The
publisher has taken technical and administrative measures to assure authenticity,
security and accessibility.
According to intellectual property law the author has the right to be mentioned when his/her work is accessed as described above and to be protected
against infringement.
For additional information about the Linköping University Electronic Press
and its procedures for publication and for assurance of document integrity, please
refer to its www home page: http://www.ep.liu.se/
© Rickard Sjölund and Nicklas Vedin
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Related manuals

Download PDF

advertisement