Professional Documents
Culture Documents
net/publication/254559934
Article in Proceedings of the Institution of Mechanical Engineers Part I Journal of Systems and Control Engineering · September 2012
DOI: 10.1177/0959651812451685
CITATIONS READS
0 568
4 authors, including:
Reza Kazemi
Khaje Nasir Toosi University of Technology
241 PUBLICATIONS 1,814 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Squeak and Rattle prediction for robust product development View project
All content following this page was uploaded by Reza Kazemi on 18 January 2014.
Abstract
The use of control systems, especially vehicle dynamic control systems, is increasing at a remarkable rate. A key para-
meter that affects the performance of a vehicle dynamic control system is the interaction forces between the tyre and
the road since these forces determine the acceleration, braking and steering properties of a car. Consequently, the accu-
rate estimate of these forces is highly desired. Technical limitations and cost considerations mean that it is standard prac-
tice to estimate this data instead of performing measurements. Since available tyre models are generally an algebraic
representation of the actual tyre and neglect variations in road conditions and uncertainties, an improved model for vehi-
cle dynamic control applications is required. This paper proposes a tyre model which is formulated based on the least
squares method and variable exponential forgetting estimation. The proposed estimator calculates tyre forces using
information obtained from the sensors in a vehicle dynamic control system. The performance and robustness of the pro-
posed tyre model is evaluated via virtual handling simulations in ADAMS software.
Keywords
Tyre dynamic modelling, estimation, variable exponential forgetting factor, vehicle dynamic control
Segel and Fiala.1–3 Considering that dynamic data for implemented to evaluate the response of a dynamic sys-
a vehicle such as yaw rate, wheel speed, body longitudi- tem and derive controller or estimator rules in control
nal and lateral accelerations and steering angle are applications. In this research, a seven-degree-of-freedom
available from current VDC sensors, we will use this (7DOF) car model (body longitudinal, lateral and yaw
data to estimate the planar forces acting between the motions and spin of the four wheels) is implemented to
road and the tyre. design a tyre force estimator, as is shown in Figure 1.
In 1993, Huang et al.4 performed a study to identify To write the equations of motion of the system we
and estimate vehicle parameters using a gradient consider the free-body diagram of the vehicle as illu-
method. In 1996, Gustafson5 designed an estimator of strated in Figure 1. The equations can be written as
the friction coefficient between tyre and road using
wheel slip data. In that research, which was part of the 1
v_x = rvy + Fxfl + Fxfr cosðdÞ Fyfl + Fyfr
Local Traffic Control System and Automatic Driver m
Assistant Control System project in Sweden, the fric- sinðdÞ + Fxrr + Fxrl
tion coefficient was calculated and estimated via vari- 1
ous methods such as a recursive least square (RLS) v_y = rvx + Fxfl + Fxfr sinðdÞ + Fyfl + Fyfr
m
linear filter, marginalized likelihood ratio exam and a cosðdÞ + Fyrr + Fyrl
bank of parallel RLS filters. In 2001, Kunsoo and
Joonyoung6 estimated the lateral forces acting on a tyre 1
r_ = Fxfl + Fxfr Lf sinðdÞ + Fyfl + Fyfr
using a Kalman filter. In 2003, Carlson and Gerdes7 Izz
published an article in which they used a non-linear esti-
Lf cosðdÞ Fyrl + Fyrr Lr
mator to identify a tyre’s effective radius and its longi-
Tf Tf
tudinal slip coefficient using ABS and GPS sensors. In + Fxfr Fxfl cosðdÞ + Fyfl Fyfr
2006 Tanelli and Savaresi8 estimated the maximum tyre 2 2
friction coefficient obtained in contact with a road sur- Tr
sinðdÞ + ðFxrr Fxrl Þ ð1Þ
face. They evaluated the proposed method under the 2
conditions that the tyre either slips or it purely rotates.
Li et al.9 introduced an algorithm to estimate tyre With reference to Figure 2, equations pertaining to
forces. They simulated the model using the hardware in four degrees of freedom of the wheels can be obtained as
the loop method and evaluated the robustness of the Ji v_ i = Fxi 3 Rwi + Ti Tbi TRolli i = 1, 2, 3, 4
system under different road conditions.
ð2Þ
Modelling and validation where TRoll is the rolling resistant torque, which
appears due to hysteresis losses and is calculated using
Using virtual models in the process of developing vehicle equation (3)
systems is standard practice in the automotive industry.
The virtual and mathematical models are normally TRolli = fr 3 Fzi 3 Rwi i = 1, 2, 3, 4 ð3Þ
DOF Subsystem
containing linear or non-linear parametric uncertain- The logic behind this is that old data is the result of pre-
ties. It is possible to perform parameter estimation via vious values of parameters and in order to estimate the
online or offline approaches. Offline parameter estima- new value of parameters, the weight of old data with
tion methods are useful, for example, for the estimation regard to new data must be reduced. Accordingly, we
of unknown constant parameters prior to the start of a introduce a weighting coefficient that decreases expo-
manoeuvre when sufficient time is available for the esti- nentially over time. The main idea here is to minimize
mation process. On the other hand, to estimate variable the error function with respect to â(t)15
parameters that change during a manoeuvre, online ðt ðt
parameter estimation methods are implemented. 2
J= exp l(r)dr ky(s) W(s)^
a(t)k ds ð8Þ
Noting the importance of tyre force accuracy on the 0 s
VDC systems performance and considering that these
forces can fluctuate instantly during vehicle man- where l(r) is the forgetting coefficient and always has a
oeuvres, it is clear that an online estimation approach non-negative value. In order to estimate the values of
must be taken to identify the instant value of tyre â(t), we set the derivative of equation (8) with respect
forces. to â(t) equal to zero
One of the basic estimation methods is least squares
(LS) estimation. An extensive literature exists on this ∂J d
= 0 ! ½^a = P(t)WT e ð9Þ
topic and in particular there are numerous reported a
∂^ dt
attempts to improve and optimize the performance of
where P(t) is the estimation gain and is defined as
the LS method. One promising route to these objectives
is to use gradient and non-gradient methods.12–14 ðt 1
d 1
Parameter estimation is performed based on calcu- PðtÞ = WT ðrÞWðrÞdr ! ½P
lating the unknown parameters using measurable sys- 0 dt
tem variables. Hence, an estimation mathematical = l(t)P1 + WT (t)W(t) ð10Þ
model is needed to represent the relation between
unknown parameters and available data. A general In practice, it is more convenient to rewrite equation
model to estimate parameters is the linear parameteri- (10) in the form of equation (11) in order to facilitate
zation model which can be written as the formulation of the system in the calculation process
y(t) = W(t)a ð4Þ d
½P = l(t)P PWT (t)W(t)P ð11Þ
where the n-dimensional vector y represents the output dt
of the system, the m-dimensional vector a contains the In online estimation, we need to define the initial val-
unknown parameters of the system and the (n3m) ues of the parameters and the gain matrix. The initial
matrix W consists of the measurable or available sig- value of P must be chosen as large as possible with
nals of the system. In order to form a relation for the respect to noise sensitivity considerations. We choose
estimated parameters, â, equation (5) is introduced, P(0) in the form of a diagonal matrix for the sake of
where ŷ(t) is the predicted output in time t simplicity15
y^(t) = W(t)^
a(t) ð5Þ
Pð0Þ = 1000
The difference between measured and predicted out- 2 3
5 0 0 0 0 0 0
puts is considered to be the prediction error 60 5 0 0 0 0 0 7
6 7
60 0 5 0 0 0 0 7
e(t) = y^(t) y(t) ð6Þ 6 7
3660 0 0 5 0 0 0 7 7
The prediction error is related to the parameter esti- 60 0 0 0 50, 000 0 0 7
6 7
mation error through equation (7) 40 0 0 0 0 50, 000 0 5
0 0 0 0 0 0 10, 000
a Wa = W~
e = W^ a ð7Þ
where ã is the parameter estimation error. Measuring In order to study the convergence and stability of the
the elements of y and W during the process yields a set proposed estimator we solve the equations for the error
of equations in the form of equation (4). To achieve the in the parameters and the gain matrix, which leads to
goal of estimation, these equations must be solved in ðt
term of the unknown variable a. P1 ðtÞ = P1 ð0Þ exp lðrÞdr
0
ðt ðt
LS estimation with variable exponential forgetting + exp lðrÞdr WT ðsÞWðsÞds ð12Þ
0 s
The estimation of the parameters that vary in the pro-
cess time requires that we implement a method to mini- d 1
P a~ = lP1 a~ ð13Þ
mize the effect of old data in the estimation process. dt
Bayani Khaknejad et al. 1197
By solving equation (13) the parameter error is We limit the gain matrix P to avoid fluctuation in the
formed as equation (14) and subsequently, substituting estimation error while guaranteeing the speed of esti-
equation (12) into equation (14) yields equation (15) mation. In fact, equation (18) indicates that the forget-
ðt ting process is conducted using the l0 coefficient when
a~ðtÞ = exp lðrÞdr PðtÞP1 ð0Þ~ að 0Þ ð14Þ the magnitude of P is small. On the contrary, when P is
0 large, forgetting is devalued and if it exceeds the upper
ðt ð s
limit of k0, the forgetting process is cancelled. Hence, in
a~ðtÞ = P1 ð0Þ + exp lðrÞdr order to limit the gain matrix we correct equation (11)
0 0
1 by defining the following rule
WT ðsÞWðsÞds P1 ð0Þ~
að0Þ ð15Þ
d
½P = lP PWT WP kPðtÞk4K0 ð19Þ
dt 0 otherwise
Evidently, the convergence of the answer is guaranteed
if we demonstrate that equation (16) is true.
ðt ð s In order to check the convergence of the estimator
lim exp lðrÞdr WT ðsÞWðsÞds = ‘ ) lim a~ðtÞ = 0 response as regards supplementing equation (19), we
t!‘ 0 t!‘
0 consider the following Lyapunov function
ð16Þ
P1 a~
This condition is spontaneously satisfied when the a, t) = a~T
V(~ ð20Þ
2
signal matrix W is persistently excited. In other words,
a system with persistently excited signals will satisfy Using the derivative of equation (20) we have that
equation (17) which leads to a guaranteed convergence
_ a, t) = 1a~T d P1 a~ + a~T P1 a~
V(~ ð21Þ
ðt 2 dt
lim WT ðrÞWðrÞdr = ‘ ð17Þ
2 3
cos d cos d 1 1 sin d
6 0 7
6 m m m m m
6 sin d sin d cos d 1 7 7
6 0 0 7
6 m m m m 7
6 L sin d T 2 cos d L sin d + T 2 cos d Tr Tr 7
6 f Lf cos d r7
L
f f f
6 Iz Iz 7
6 Iz Iz 2Iz 2Iz 7
6 Rw 7
W(t) = 6
6 0 0 0 0 0 7 7 ð28Þ
6 Iw 7
6 Rw 7
6 0 0 0 0 0 7
6 Iw 7
6 7
6 Rw 7
6 0 0 0 0 0 7
6 Iw 7
4 5
0 0 0 R
I
w 0 0
w
wheel. This will in turn simplify the equations for the Considering that here W(t) is persistently excited, the
estimator and the calculation time will decrease. Hence, stability and convergence of the estimator is guaran-
we rewrite the equations as follows teed. In order to estimate the desired parameters, first,
in each step the signal matrix is calculated as a function
1
of the steering input and the forgetting factor is regu-
ax = Fxfl + Fxfr cosðdÞ Fyf sinðdÞ + Fxrr + Fxrl
m lated using equations (28) and (18). By substituting
1 these values into equation (19) the gain matrix is calcu-
ay = Fxfl + Fxfr sinðdÞ + Fyf cosðdÞ + Fyr
m lated which in turn is used to calculate the parameter
estimation error using equation (15). Eventually, the step steering manoeuvre, the vehicle starts to depart
estimated value of the tyre forces is calculated by sub- from straight line movement in a gradual manner at an
stituting the parameter estimation error vector and the initial velocity of 100 km/h as shown in Figure 6. This
parameters measured by sensors into equation (7). This manoeuvre can be considered as a severe handling man-
procedure is iterated in each step to obtain a recursive oeuvre as the lateral acceleration of the car reaches
calculation which gives the estimated values. about 0.7 g. Figures 7 and 8 indicate the results of the
ADAMS tyre model and the proposed estimator. It is
clear from these results that the tyre model proposed in
Simulations this paper can calculate tyre forces in both transient
A full-vehicle model of the reference car in ADAMS/ and steady state conditions.
Car software was implemented to evaluate the perfor- In order to evaluate the performance of the system
mance of the proposed estimator. In addition, the pro- in rapidly varying and critical conditions two additional
posed estimator was modelled in Matlab/Simulink. For manoeuvres were considered: first, the road surface var-
online estimation, the entire system has to be simulta- ies and second, the tyres slip during braking. In addi-
neously simulated and accordingly, one needs to estab- tion, the considered model vehicle was equipped with a
lish a direct connection between the two programs for VDC system16 to allow the performance of the estima-
concurrent running. The required sensors for the esti- tor in estimating the tyre forces when a disturbed brake
mator are defined in ADAMS and are as follows: torque is applied to the wheels using the VDC system to
be probed. In the second manoeuvre, a vehicle with an
body yaw rate sensor; initial velocity of 100 km/h moving on a dry road is
body longitudinal and lateral accelerations; steered by the driver to change lanes while simultane-
steering wheel angle; ously the brake pedal is applied, as shown in Figure 9.
wheels angular velocity; It is clear from the results presented in Figures 10 and 11
wheel spinning torques. that the proposed tyre model can efficiently and accu-
rately estimate the tyre forces, even when considering the
After providing the required sensors and actuators, fast varying forces applied by the VDC system. In addi-
the input/output channels and ADAMS/Solver system tion, to study the effect of different road surface condi-
files were built and exported into the Matlab/Simulink tions, a single lane change manoeuvre on a road with a
toolbox; the files contained the equations and informa- friction coefficient of 0.8 on the right side and 0.3 on the
tion of the dynamic model. During the simulation, left side was simulated. The steering input for this man-
Matlab called the ADAMS/Solver to solve the dynamic oeuvre is depicted in Figure 12; the initial velocity of vehi-
equations for the vehicle motion. In fact, the ADAMS/ cle was 100 km/h and the driver started to brake after the
Solver acted as the representative of the full-vehicle first second. Evidently, the performance of the estimator
model in Matlab/Simulink that calculated the required in calculating the longitudinal and lateral tyre forces is
outputs given the input data. Table 2 presents the val- precise and fast. As shown in Figures 13 and 14, the esti-
ues of vehicle parameters used in the ADAMS/Car mator can follow the force variations both in dry and wet
model. conditions.
It is clear from the presented results that the para-
meters that significantly influence the obtained results
Results of the simulations are the vehicle mass (m), yaw moment of inertia (Iz ),
The performance of the proposed estimator was veri- centre of gravity position (Lf) and wheel track (Tf,r).
fied using the results of the simulations; this paper con- The robustness of the proposed estimator was investi-
siders three manoeuvres: step steering, braking during a gated by simulating a single lane change (SLC) man-
lane change and a lane change in a m-split road. In the oeuvre while the vehicle mass (m), yaw moment of
1200 Proc IMechE Part I: J Systems and Control Engineering 226(9)
Figure 6. Steer input and car body accelerations during the step steer manoeuvre.
Figure 7. ADAMS tyre model and the estimated longitudinal tyre forces during the step steer manoeuvre.
Bayani Khaknejad et al. 1201
Figure 8. ADAMS tyre model and the estimated lateral tyre forces during the step steer manoeuvre.
Figure 9. Steering input and car body accelerations during braking for the lane change manoeuvre.
1202 Proc IMechE Part I: J Systems and Control Engineering 226(9)
Figure 10. ADAMS tyre model and the estimated longitudinal tyre forces during braking for the lane change manoeuvre.
Figure 11. ADAMS tyre model and the estimated lateral tyre forces during braking for the lane change manoeuvre.
Bayani Khaknejad et al. 1203
Figure 12. Steering input and car body accelerations during the lane change manoeuvre on a m-split road.
Figure 13. ADAMS tyre model and the estimated longitudinal tyre forces during the lane change manoeuvre on a m-split road.
1204 Proc IMechE Part I: J Systems and Control Engineering 226(9)
Figure 14. ADAMS tyre model and the estimated lateral tyre forces during the lane change manoeuvre on a m-split road.
inertia (Iz) and centre of gravity position (Lf) were 3. Wheel speed sensor: output signals up to 1680 Hz
altered by 10% (about 100 kg which is high for this (equal to 200 km/h for the reference car) with 5%
vehicle segment), 7% and 7%, respectively. The results error.18
of the estimation are presented in Figures 15 and 16.
It can be seen that these uncertainties result in a Depending on the working condition of the VDC,
maximum error in the estimated values of 3.5% in the the maximum working frequency of the VDC actuators
tyre longitudinal force and a maximum error of 14% in is 50 Hz.18 Hence, the outputs of the sensors were fil-
the tyre lateral force. Although the estimated results are tered using a low-pass filter tuned to have 50 Hz as the
generally acceptable, it is better to compensate the lack threshold frequency. The results of SLC manoeuvre
of information by implementing an ancillary system. simulations (performed using the ADAMS-Simulink
For this purpose, using the same estimation approach, loop) at an initial velocity of 100 km/h and the data for
the values of the m, Iz and Lf parameters were esti- noisy sensors are presented in Figures 17 and 18. These
mated in the first seconds of car excitation (movement) results indicate the robustness and stability of the pro-
and the results were used in the force estimator for the posed estimator when they experience noisy inputs.
rest of the driving sequence.17 Using the proposed para-
meter identifier ensures that the effects of parametric
uncertainties are minimized. Conclusions
Sensor noise can be considered to be a measurement
In this research a tyre model is introduced that calcu-
error which is common in practice and real-time eva-
lates the tyre forces based on mathematical estimation
luations. In order to evaluate the robustness and flexi-
equations rather than algebraic experimental relations
bility of the system against sensor noise, induced noises
as is the common approach. The proposed model is
by sensors were modelled as a band-limited Gaussian
evaluated by its performance in various handling man-
white noise. The following sensors were considered to
oeuvres and it is compared with one of the most popu-
be sources of noise.
lar tyre models, PAC2000. The results confirm the
accuracy, speed and robustness of the estimator in nor-
1. Body acceleration sensors: output signals up to mal and critical conditions as well as conditions that
64 Hz with 5% error.18 experience disturbances (disturbed brake toques), para-
2. Yaw rate sensor: output signals up to 916 Hz with metric uncertainties (vehicle inertial parameters), envi-
4.3% error.18 ronment perturbations (road surface condition) and
Bayani Khaknejad et al. 1205
Figure 15. ADAMS tyre model and the estimated lateral tyre forces during the SLC manoeuvre with parameter uncertainties.
Figure 16. ADAMS tyre model and the estimated longitudinal tyre forces during the SLC manoeuvre with parameter uncertainties.
noisy sensor data. Implementing this analytical tyre identification tests and can significantly reduce the
model can result in lower costs for tyre modelling effects of uncertainties and simplifications.
1206 Proc IMechE Part I: J Systems and Control Engineering 226(9)
Figure 17. ADAMS tyre model and the estimated lateral tyre forces during the SLC manoeuvre with parameter uncertainties.
Figure 18. ADAMS tyre model and the estimated longitudinal tyre forces during the SLC manoeuvre with parameter uncertainties.
Funding
4. Huang F, Chen JR and Tsai LW. The use of random
This research received no specific grant from any fund- steer test data for vehicle parameter estimation. SAE
ing agency in the public, commercial or not-for-profit paper 930830, 1993.
sectors. 5. Gustafson F. Estimation and change detection of tyre-
road friction using the wheel slip. In: The 1996 IEEE
international symposium on computer-aided control system
References design, Dearborn, MI, 15–18 September 1996, pp.99–104.
1. MSC Inc. ADAMS user manual, 2003. Piscataway, NJ: IEEE Press.
2. Palkovics L, Grindy MEL and Pacejca HB. Modelling of 6. Kunsoo H and Joonyoung K. Active steering control
the cornering characteristics of tyre on an uneven road based on the estimated tyre forces. J Dyn Syst 2001; 123:
surface. Int J Veh Des 1994; 15(1/2): 189–215. 505–511.
3. Pacejca HB, Bakker E and Nybory L. Tyre modelling for 7. Carlson CR and Gerdes JC. Nonlinear estimation of
use in vehicle dynamic studies. SAE paper 870421, 1987. longitudinal tyre slip under several driving conditions.
Bayani Khaknejad et al. 1207
In: The proceedings of American control conference, d front wheels’ steering angle
Denver, CO, 4–6 June 2003, pp.4975–4980. e estimated error vector
8. Tanelli M and Savaresi MS. Friction-curve peak detec- fr rolling resistance coefficient
tion by wheel deceleration measurements. In: The IEEE Fxi tyre longitudinal force
intelligent transportation systems conference, Toronto, Fyi tyre lateral force
17–20 September 2006, pp.1592–1597. Piscataway, NJ: Fzi tyre vertical force
IEEE Press. Izz vehicle yaw moment of inertia
9. Li L, Song J, Wang H and Wu C. Fast estimation and J error function
compensation of the tyre force in real time control for
Ji wheel moment of inertia
vehicle dynamic stability control system. Int J Veh Des
2008; 48: 208–229.
L car wheel base
10. Wong JY. Theory of ground vehicles. USA: John Wiley & Lf distance between the vehicle’s centre of
Sons, 2001. gravity and the front axle
11. Bayani Khaknejad M. Designing vehicle stabilizer control- Lr distance between the vehicle’s centre of
ler estimator. MS Thesis, K. N. Toosi University of Tech- gravity and the rear axle
nology, Tehran, 2007. P estimation gain matrix
12. Marquardt DW. An algorithm for least square estimation r vehicle yaw rate
of nonlinear parameters. J Soc Ind Appl Math 1983; 11: Rw wheel’s effective radius
431–441. Tbi wheel braking torque
13. Lozano R, Dimogianopoulos D and Mahony R. Identifi- Tf car front track
cation of linear time-varying systems using a modified
Ti wheel traction torque
least-squares algorithm. Automatica 1999; 36: 1009–1015.
Tii wheel total torque
14. Song S, Lim JS, Baek SJ, et al. Gauss Newton variable
forgetting factor recursive least square for time varying Tr car rear track
parameter tracking. Electron Lett 2000; 36(11): 988–990. Troll rolling resistance torque
15. Slotine JJ and Li W. Applied nonlinear control. First ed. vx vehicle longitudinal velocity
UK: Prentice Hall, 1992. vy vehicle lateral velocity
16. Bayani Khaknejad M, Kazemi R and Azadi S. Vehicle y system output vector
stabilization via a self-tuning optimal controller. Amirka- ŷ estimated output vector
bir Int J Model Ident Simul Control 2011; 43(2): 33–42.
17. Bayani Khaknejad M, et al. Identification of vehicle para- l forgetting coefficient
meters using modified least square method in ADAMS/ vi rotating angular velocity of the wheel
Car. In: The IEEE international conference on modeling,
identification and control, Shanghai, China, 26–29 June
2011, pp.98–103 paper no. ICMIC-S2-0124. Piscataway,
NJ: IEEE Xplore.
18. Robert Bosch Gmbh. Technical documentation/TCD speci-
fication sheet: ECU EA106/EA1. Iran: Robert Bosch, 2008.
Appendix 1
Notation
a unknown parameters vector
â unknown parameters estimation vector
ã unknown parameters estimation error
vector