Escolar Documentos
Profissional Documentos
Cultura Documentos
filter pid.sagepub.com
Abstract
This paper presents a new robust road bank angle estimation method that does not require a differential global position-
ing system or any additional expensive sensors. A modified bicycle model, which is less sensitive to model uncertainties
than is the conventional bicycle model, is proposed. The road bank angle estimation algorithm designed using this model
can improve robustness against modelling errors and uncertainties. A proportionalintegral HN filter based on the game
theory approach, which is designed for the worst cases with respect to the sensor noises and disturbances, is used as
the estimator in order to improve further the stability and robustness of the bank estimation. The effectiveness and per-
formance of the proposed estimation algorithm are verified by simulations and tests, and the results are compared with
those of previous road bank angle estimation methods.
Keywords
Road bank angle estimation, proportionalintegral HN filter, modified bicycle model, observer-based disturbance
estimation
angle and modelling errors are defined as uncertain additional expensive sensors. A modified bicycle model,
parameters and they are estimated using a disturbance which is less sensitive to model uncertainties such as
observer10 and the adaptive control theory.14 These the cornering stiffnesses of the tyres than is the conven-
estimation methods require the side-slip angle, which tional bicycle model, is proposed in this paper.
can be estimated using the differential global position- Therefore, the road bank angle estimation algorithm
ing system (DGPS) measurement. Different from the designed using this model can be more robust against
one-antenna global positioning system (GPS), which is modelling errors and uncertainties than using the con-
generally used in automotive navigation systems, the ventional bicycle model. A proportionalintegral HN
DGPS with two antennae is too expensive to be used in filter (PIF) based on the game theory approach, which
passenger cars and is frequently unreliable in urban is designed for the worst cases with respect to the sen-
environments.9 Even though these methods can guar- sor noises and disturbances, is used as the estimator in
antee an acceptable accuracy of estimation, they are order to improve further the stability and robustness of
not practical solutions owing to the cost and reliability the bank estimation. The effectiveness and performance
issues related to the DGPS. A road bank angle estima- of the proposed estimation algorithm are verified by
tion using a vertical accelerometer is proposed.9 simulations and tests, and the results are compared
However, the vertical accelerometer measurement with those of previous road bank angle estimation
also cannot provide an acceptable accuracy of the methods.
road bank angle because the vertical accelerometer is
almost insensitive to the narrow range of vehicle tilts
owing to the non-linearity of the arccosine functional Vehicle dynamics model
relation between the vertical acceleration and the vehi- Figure 1 shows the schematic diagrams of the target
cle tilt.15 system. As shown in Figure 1(a), by assuming the
The road bank angle has been estimated using the bicycle model (i.e. by assuming that the dynamics of
differences between the lateral tyre force estimate and the left and right sides of the vehicle are identical) and
the lateral accelerometer measurement,16 or the differ- no pitch motion, the lateral motion of the vehicle22,23
ences between the lateral acceleration measurement and can be expressed by
the products of the yaw rate and the longitudinal
speed17 in the linear observer framework. However, may = Fyf + Fyr
these methods tend to be inaccurate under transient = Lf Fyf Lr Fyr
Iz c 1
driving conditions because they neglect the derivative
term of the lateral velocity of the vehicle. A road bank The lateral accelerometer measurement consists of three
angle estimation method based on the transfer function components, namely the linear motion term, the lateral
and dynamic filter compensation (DFC), which is motion term and the gravity term, according to22
related to the model uncertainty of the lateral dynamics,
was previously introduced.8,12 This method illustrated ay = v_y + vx c_ + g sin (fb + f) 2
the robustness issues of the road bank angle estimation;
however, the physical meaning of the DFC term was where fb is the road bank angle and f is the roll angle
not explicitly explained in the papers. of the vehicle. By equation (2), it is shown that the lat-
Disturbance observers based on the unknown input eral acceleration measurement includes not only the
observer (UIO), which is a well-known solution for dynamic component v_y + vx r of the vehicle motion but
the state and disturbance estimation of linear systems, also the gravity component g sin (fb + f) of the road
were proposed to estimate the road bank angle.13,18 bank angle and the roll angle. By assuming that the lat-
This method can guarantee the stability and conver- eral forces of the tyres are linearly proportional to the
gence of the estimation error, but estimation by this cornering stiffnesses of the tyres and that the slip angles
method is sensitive to the output changes because of of the tyres are very small, the lateral force of each tyre
the derivative term of the output in the observer. can be expressed as22,23
More complex methods that consider the roll
Fyf Cf af = Cf (bf df )
dynamics of vehicles19,20 were developed in order to
estimate the roll angle of the vehicle and the road Fyr Cr ar = Cr br
bank angle individually. On the other hand, non-lin- _ vy + Lf c_
1 vy + Lf c
ear modelling and table-based estimation methods21 bf = tan
vx vx
were developed in order to improve the accuracy of the
v y Lr
_
c v y Lr c_
state estimation of the lateral dynamics. However, most br = tan1
of the previous road bank angle estimation methods vx vx
explained above did not address the robustness issue of Fyf Fyr
Cf = , Cr =
the estimation due to uncertainties and disturbances, af af = 0 ar ar = 0 3
such as the cornering stiffnesses of the tyres and the
changes in the vehicle mass. For simplification, let u = fb + f. From equations (1),
This paper presents a new robust road bank angle (2) and (3), the lateral and yaw motions of a vehicle are
estimation method that does not require DGPS or any expressed by the state equations
Kim et al. 781
b v_y = ay vx c_ g sin u
= Lf may (Lf + Lr )Fyr
Iz c 5
(b)
From equations (5) and (3), the state equations of the
modified vehicle dynamic model can be obtained as
vx x axis x axis
f r vx x_ = Am x + Bm um + Em w
ym = Cm x 6
where
f
vy
x= , um = ay , w = sin u
c_
Front tyre Rear tyre !
0 vx
Am = (Lf + Lr )Cr (Lf + Lr )Lr Cr
(c) Iz vx Iz vx
!
Figure 1. Schematic diagrams of the target system: (a) bicycle 1
Bm = Lf m
model of a vehicle; (b) rear view of a vehicle; (c) tyre diagram. Iz
CG: centre of gravity.
g
Cm = 0 1 , Em =
x_ = Ao x + Bo uo + Eo w 0
yo = C o x + D o uo 4 As shown in equations (6), the modified model is not
affected by the cornering stiffness of the front tyres.
where
Therefore, it can be said that the modified model is less
sensitive to variations in the cornering stiffnesses of the
vy ay
x= , yo = tyres than is the conventional bicycle model, and the road
c_ c_
bank angle estimation designed using the modified model
uo = df , w = sin u can be more robust against the uncertainties. Similar
0 Lf Cf Lr Cr
1
C +C
fmvx r vx analyses can be conducted for rear-wheel-drive vehicles
mvx
Ao = @ L2f Cf + L2r Cr
A by eliminating the cornering stiffness terms of the rear
Lf Cf Lr Cr
I z vx I z vx
tyres from the equations of motion of the vehicle.
Cf !
m
Bo = Lf C f Robustness analysis of the estimation
Iz
! methods
Cf + Cr Lf Cf Lr Cr
mvx mvx As commented in the previous section, variations in the
Co =
0 1 cornering stiffnesses of the tyres exist under real driving
!
Cf
g conditions and they make the parameters of the vehicle
Do = m , Eo = model uncertain. For this reason, the effects of the
0 0
parameter uncertainties on the estimation errors should
782 Proc IMechE Part D: J Automobile Engineering 226(6)
On the other hand, for the modified vehicle dynamic In equation (11), (Haw + vx Hrw )=g is a form of
model (6), the uncertainties in the system matrices can second-order low-pass filter and therefore DFC can be
be expressed as considered as an estimation of v_y . If the model uncer-
! tainties exist, the transfer functions of equations (4) can
0 0 be expressed as
DAm = (Lf + Lr ) DCr (L + L )L DC
Iz vx f Izrvx r r
Hau
= (Co + DCo )(sI Ao DAo )1 (Bo + DBo )
DBm = 0 H ru
0 0 + Do + DDo
DCm =
0 0 Haw
= (Co + DCo )(sI Ao DAo )1 Eo
0 H rw
DDm = 8
0 12
Then, the uncertainties in the transfer functions are
Dynamic filter compensation method au Hau ,
DHau = H ru Hru
DHru = H
If the derivative of the lateral speed is zero (i.e. v_y = 0), aw Haw ,
DHaw = H rw Hrw
DHrw = H 13
the road bank angle estimation can be derived from
equation (2) as In this case, the DFC can be obtained as
are non-zero even if the system is in a steady state. For order to make equation (18) asymptotically stable (i.e.
this reason, w^dfc has steady state errors if uo or w are limt! (x x^uio ) = 0), the equations that should be
not zero. valid are
It was proposed by Tseng8 that the steady state val-
ues of the transfer functions Haw and vx Hrw are actually Euio = Eo (Co Eo ) + Quio I Co Eo (Co Eo )
implemented for the actual automotive applications to Nuio = (I + Euio Co )Ao Muio Co
mitigate the computational burden, according to Luio = Nuio Euio + Muio
g(Lf + Lr ) Guio = (I + Euio Co )Bo 19
lim Haw =
s!0 (Lf + Lr ) + Kus v2x
Quio is a constant matrix, which consists of design
gKus v2x parameters and (Co Eo ) is the left inverse of Co Eo
lim vx Hrw =
s!0 (Lf + Lr ) + Kus v2x 15 (i.e. (Co Eo )T Co Eo 1 (Co Eo )T ). The estimation of
w based on the UIO was proposed by Imsland et al.18
where as
Lr m Lf m
Kus = ^uio = E
w o Luio yuio Euio y_ uio
(Lf + Lr )Cf (Lf + Lr )Cr
(Luio Co Euio Co Ao )^
xuio + Euio Co Bo uo 20
If the model uncertainties do not exist,
Unknown input observer method
Eo (w w
^uio ) = Eo w Luio Co x + Euio Co (Ao x + Bo uo + Eo w)
The UIO is a state observer designed to decouple the + (Luio Co Euio Co Ao )^xuio Euio Co Bo uo
state estimation error from the disturbance.24 The dis- = (Euio Co Ao Luio Co )(x x^uio ) + (I + Euio Co )Eo w
turbance can be estimated by using the state estima-
= (Euio Co Ao Luio Co )(x x^uio )
tion of the UIO. The form of the UIO13,18 is expressed
by 21
This shows that equation (20) was designed to achieve
z_uio = Nuio zuio + Luio yuio + Guio uo
^uio = w by using limt! (x x^uio ) = 0 but, if the
limt! w
x^uio = zuio Euio yuio 16 model uncertainties exist, the derivative of x^uio is chan-
ged to
where
yuio = yo Do uo x^_ uio = z_uio Euio (Co + DCo )x_ Euio DDo u_ o
Nuio , Luio , Guio and Euio can be designed by the fol- = Nuio zuio + Luio yo + (Guio Luio Do )uo
lowing steps. The derivative of x^uio can be derived from Euio (Co + DCo )(Ao x + DAo x + Bo uo
equations (4) and (16) as + DBo uo + Eo w) Euio DDo u_ o 22
x^_ uio = z_uio Euio Co x_
= Nuio zuio + Luio yo + (Guio Luio Do )uo Then, the error dynamics of the UIO are given by
Euio Co (Ao x + Bo uo + Eo w) 17 x_ x^_ uio = (I + Euio Co + Euio DCo )(Ao x + DAo x
If the model uncertainties do not exist, the dynamics of + Bo uo + DBo uo + Eo w)
the estimation error are given by Nuio zuio Luio yo (Guio Luio Do )uo
+ Euio DDo u_ o
x_ x^_ uio = (I + Euio Co )(Ao x + Bo uo + Eo w) Nuio zuio Luio yo = Nuio (x x^uio ) + Muio Co x + Nuio x^uio
(Guio Luio Do )uo
Nuio zuio Luio yo + Luio Do uo
= (I + Euio Co )Ao Muio Co (x x^uio ) + Muio Co x
+ Euio DCo (Ao x + DAo x + Bo uo + DBo uo + Eo w)
+ (I + Euio Co )Ao Muio Co ^xuio + (I + Euio Co )Bo uo
+ (I + Euio Co )(DAo x + DBo uo ) + Euio DDo u_ o
+ (I + Euio Co )Eo w Nuio zuio Luio yo (Guio Luio Do )uo
= (I + Euio Co )Ao Muio Co (x x^uio )
= Nuio (x x^uio ) + Euio DCo Eo w + Euio DDo u_ o
+ (I + Euio Co )Ao Muio Co Nuio zuio + (I + Euio Co ) DAo Muio DCo
+ (I + Euio Co )Ao Euio + Muio Co Euio + Muio Luio yo + Euio DCo (Ao + DAo )x
+ f(I + Euio Co )Ao Muio Co Euio Do + (Luio Muio )Do + (I + Euio Co ) DBo Muio DDo
+ (I + Euio Co )Bo Guio guo + (I + Euio Co )Eo w + Euio DCo (Bo + DBo )uo
18 23
where Muio is a constant matrix, which should be This means that limt! (x x^uio ) 6 0 if the model
selected to make (I + Euio Co )Ao Muio Co stable. In ^uio is
uncertainties exist. The error of w
784 Proc IMechE Part D: J Automobile Engineering 226(6)
1
Eo (w w
^uio ) = Eo w Luio (Co + DCo )x Luio DDo uo x x^po Ao Kpo1 Co Eo
lim =
+ Euio Co (Ao x + DAo x + Bo uo t! w w ^po Kpo2 Co 0
+ DBo uo + Eo w) Kpo1 I DCo x + DDo uo
3 29
+ (Luio Co Euio Co Ao )^ xuio Euio Co Bo uo Kpo2 0 DAo x + DBo uo
= (Euio Co Ao Luio Co )(x x^uio )
It is possible to select Kpo1 and Kpo2 to minimize equa-
+ (Euio Co DAo Luio DCo )x
tion (29) at the cost of reducing the freedom of the
+ (Euio Co DBo Luio DDo )uo observer design (e.g. pole placement methods should be
24 modified in order to minimize equation (29)). However,
variations in the other vehicle parameters are ignored
Therefore, it is concluded that the model uncertain- in equation (29) and they can amplify the steady state
ties make both x^uio and w ^uio inaccurate and w ^uio has error of the estimation even if equation (29) is mini-
steady state errors if x or uo are not zero. mized by the gain selection.
Because differentiating the output amplifies the
effect of the sensor noise, a low-pass filter is used in this
paper according to PIO of the modified vehicle dynamic
model
s
y^_uio (s) = yuio (s) 25
t uio s + 1 A PIO can be derived from the modified vehicle
dynamic model (6) according to
where t uio is the time constant of the filter.
x^_ pm = Am x^pm + Bm um + Kpm1 (ym y^m ) + Em w
^pm
w ^_ pm = Kpm2 (ym y^ ), y^ = Cm x^pm
Proportionalintegral observer of the original m m
The proportionalintegral observer (PIO) is a state where Kpm1 and Kpm2 are the observer gain matrices. If
observer designed to reduce the steady state error by the model uncertainties exist, the error dynamics of the
using one or more integration terms of the estimation PIO of the modified model are given by
error.25,26 A PIO can be derived from the original vehi-
cle dynamic model (4) as x_ x^_ pm = (Am Kpm1 Cm )(x x^pm ) + Em (w w
^pm ) + DAm x
w_ w _
^pm = Kpm2 Cm (x x^pm ) + w_
x^_ po = Ao x^po + Bo uo + Kpo1 (yo y^o ) + Eo w
^po 31
^_ po = Kpo2 (yo y^o ),
w y^o = Co x^po + Do uo 26
If the time goes to infinity
where Kpo1 and Kpo2 are the observer gain matrices. If 1
the model uncertainties do not exist, the dynamics of Am Kpm1 Cm Em
lim (w w
^pm ) = 0 I
the estimation error are given by t! Kpm2 Cm 0
DAm x
x_ x^_ po = (Ao Kpo1 Co )(x x^pio ) + Eo (w w
^po ) 3 =0 32
0
w_ w _
^po = Kpo2 Co (x x^po ) + w_
Therefore, it is concluded that the PIO of the modi-
27 fied model can eliminate the steady state error of w ^pm
even if DAm exists. It is notable that the steady state
This means that Kpo1 and Kpo2 should be selected
error remains zero even if variations in the other vehicle
to make Ao Kpo1 Co and Kpo2 Co stable. If the
parameters exist owing to the structure of equations
model uncertainties exist, the error dynamics of the
(6).
PIO derived from the original model are changed
to
Proposed road bank angle estimation
x_ x^_ po = (Ao Kpo1 Co )(x x^po ) + Eo (w w
^po ) method
+ (DAo x + DBo uo ) Kpo1 (DCo x + DDo uo )
The results of the previous section show that the PIO
w_ w _
^po = Kpo2 Co (x x^po ) + w_ Kpo2 (DCo x derived from the modified model is the best solution
+ DDo uo ) from the viewpoint of the robust performance of the
28 steady state error. This paper proposes to apply the
PIO algorithm to the bank angle estimation using the
If the time goes to infinity, the error dynamics of the modified vehicle model (6). By assigning w = sin u as a
PIO become new state, equations (6) can be modified as
Kim et al. 785
and v is the noise in the yaw rate measurement such as where P0 , Q, R and S are positive definite matrices that
offset and stiction.27 It is notable that the Luenberger depend on the performance requirements. Because all
observer derived from equations (33) is the same as the the state equation matrices of equations (33) are contin-
PIO in equations (30). uous, the system N-norm kGerr k is finite.31 This means
Because the rank of the observability matrix of the that there exists a positive scalar u such that the optimal
system described by equations (33) is estimation w^ satisfies
T 1
rank Cw Cw Aw Cw A2w =3 34 sup Jw 4 41
e60 u
the system described by equation (34) is observable.
In order to make the observer derived from equa- Conditions of the existence of a bound on the
tions (33) robust against a set of disturbances including system N-norm kGerr k and the lower bound of the
the disturbance term n and the noise term v, this paper N-norm can be derived using theorems given by
proposes to use a continuous-time HN filter based on Burl;31 therefore the upper bound of u can also be
the game theory approach.2830 The error of the road derived.30 The derivation of the upper bound of u is
bank angle estimation is omitted because it is beyond the scope of this paper.
The continuous-time H filter can be derived from
ew = w w
^ 35 equation (41) as
where w
^ is the estimation of the road bank angle term. x^_ w = Aw x^w + Bw um + PCTw R1 (ym Cw x^w ) 42
Because the input um and the output ym are the only
known terms, the estimate of the road bank angle term where
should be derived from them. Let
^ = Lw x^w ,
w x^w (0) = x^w0
w^ = Lw x^w
36 and as
x^w = fest (um , ym , x^w (0))
P_ = Aw P + PATw + Ew QETw
where fest (um , ym , x^w (0)) is the estimation function, which
should be determined. From equations (33), (35) and P(CTw R1 Cw uLTw SLw )P 43
(36), the estimation error at time 0 can be derived as where
ew (0) = w(0) w(0)
^ P(0) = P0
= Lw xw (0) x^w (0) 37
and P is a symmetric positive definite matrix if the
On the basis of equations (33), (35), (36) and (37), it can solution of equation (43) exists 8t 2 0, tf . The
be said that the estimation error ew is a function of n, v detailed derivations of equations (42) and (43) can be
and xw (0) x^w (0). Therefore, ew can be expressed as found in the papers by Banavar and Speyer28 and de
Souza et al.29 It is notable that the solution of the
ew (s) = Gerr (s)e(s) 38 algebraic Riccati form of equation (43) can be used as
786 Proc IMechE Part D: J Automobile Engineering 226(6)
P0 in order to reduce the calculation complexity of Table 2. Parameters of the UIO method.
equation (43) by taking the risk of increasing the esti-
mation errors due to the initial state xw (0).30 In this Symbol Quantity Selected value
case, the steady-state HN filter can be obtained from
Muio Weight matrix for the outputs 10 10
equations (42) and (43) as
10 10
Quio Weight matrix for Euio 1 1
x^_ w = Aw x^w + Bw um + P0 CTw R1 (ym Cw x^w ) 1 1
Aw P0 + P0 ATw + Ew QETw t uio Time constant of the filter 1/9
Simulation results
Table 4. Parameters of the PIF method.
Simulations were performed in order to verify the effec-
tiveness and performance of the proposed estimation Symbol Quantity Selected value
algorithm. A front-wheel-drive sport utility vehicle Q Weight matrix for the 0.1
(SUV) model in CarSim was selected as the vehicle disturbances
model, and the proposed estimation algorithm was R Weight matrix for the noises 0.0035
implemented by MATLAB/Simulink using the para- S Weight matrix for the 1
meters shown in Table 1. The cornering stiffnesses Cf ss estimation errors 0 1
P0 Initial value of the Riccati equation 0 0 0
and Cr ss used for the estimator design were calculated @0 0 0 A
by using the results of steady state cornering simula- 0 0 0:5
tions and experiments. The lateral acceleration, the u Parameter for the cost function 1
yaw rate and the longitudinal speed are assumed to be
known in the simulations. These signals are normally
available through the in-vehicle network of modern
vehicles equipped with an electronic stability program
Straight road with a constant bank angle
(ESP). Figure 2 shows simulation results when the vehicle tra-
In order to assess the performance and robustness of vels on a straight road with a constant bank angle
the proposed algorithm, the DFC method,8 the UIO (30% = 16.7). The vehicle was driven for 15 s with a
method13,18 and the PIO method,25,26 which do not constant longitudinal speed (80 km/h) and maintained a
require DGPS or any additional expensive sensors, straight direction on the banked road. The purpose of
were also implemented and simulated in this paper. this simulation is to compare the performance of the
The estimation parameters of the methods are carefully road bank angle estimation methods during steady state
tuned to produce the best results. The estimation para- cornering on the banked road. Simulations were con-
meters for the UIO, the PIO and the PIF methods are ducted for several different driving conditions with dif-
decided by hand tuning processes based on the results ferent types of modelling error and uncertainty in order
of various simulations and experiments. The para- to examine the robustness of the bank angle estimation
meters identified through the processes are shown in methods. Figure 2(b) shows the simulation results when
Tables 2,3 and 4. the driving condition is the same as the driving
Kim et al. 787
30 10
Throttle Angle Steering Wheel Angle
25
5
20
0
15
-5
10
-10
5
0 -15
0 5 10 15 0 5 10 15
Time (s) Time (s)
(a)
0 0
Reference Reference
PIF Method PIF Method
PIO Method PIO Method
-5 -5
Estimated Angle (deg)
-10 -10
-15 -15
-20 -20
0 5 10 15 0 5 10 15
Time (s) Time (s)
(b) (c)
0 0
Reference Reference
PIF Method PIF Method
PIO Method PIO Method
-5 -5
Estimated Angle (deg)
-10 -10
-15 -15
-20 -20
0 5 10 15 0 5 10 15
Time (s) Time (s)
(d) (e)
Figure 2. Simulation results (constant road bank angle): (a) driver inputs; (b) estimated u (m = 0:85); (c) estimated u (m = 0:3); (d)
estimated u (ay = ay real cos (108 )); (e) estimated u (Cf = 1:2Cf ss ).
PIF: proportionalintegral HN filter; PIO: proportionalintegral observer; DFC: dynamic filter compensation; UIO: unknown input observer.
condition used for the estimation algorithm design. 2(e) shows the simulation results when the actual cor-
Figure 2(c) shows the simulation results when the road nering stiffness Cf is different from the cornering stiff-
surface condition becomes more slippery (i.e. low m). ness Cf ss used for designing the estimation algorithm in
Figure 2(d) shows the simulation results when the lat- Table 1, owing to the changes in the vehicle mass (e.g.
eral acceleration is measured incorrectly owing to the more passengers). Simulation results were obtained by
misaligned installation of the accelerometer and Figure using the PIF method based on the modified vehicle
788 Proc IMechE Part D: J Automobile Engineering 226(6)
30 60
Throttle Angle Steering Wheel Angle
25
40
20
20
15
0
10
-20
5
0 -40
0 5 10 15 0 5 10 15
Time (s) Time (s)
(a)
40 40
Reference Reference
30 PIF Method 30 PIF Method
PIO Method PIO Method
Estimated Angle (deg)
10 10
0 0
-10 -10
-20 -20
0 5 10 15 0 5 10 15
Time (s) Time (s)
(b) (c)
40 40
Reference Reference
30 PIF Method 30 PIF Method
PIO Method PIO Method
Estimated Angle (deg)
10 10
0 0
-10 -10
-20 -20
0 5 10 15 0 5 10 15
Time (s) Time (s)
(d) (e)
Figure 3. Simulation results (double lane change on a flat road): (a) driver inputs; (b) estimated u (m = 0:85); (c) estimated u
(m = 0:3); (d) estimated u (ay = ay real cos (108 )); (e) estimated u (Cf = 1:2Cf ss ).
PIF: proportionalintegral HN filter; PIO: proportionalintegral observer; DFC: dynamic filter compensation; UIO: unknown input observer.
model, and by the PIO method, the DFC method and CarSim. As shown in the figures, because of the integral
the UIO method based on the original model (4). The state in the observers, the PIF and the PIO methods
curves labelled Reference in Figure 2(b),(c),(d) and (e) yield better performances on estimating the bank angle
are calculated as the difference between the absolute without the steady state errors than the DFC and the
heights of the left and the right wheels provided by UIO methods do.
Kim et al. 789
30 400
Throttle Angle Steering Wheel Angle
300
25
20
100
15 0
-100
10
-200
5
-300
0 -400
0 5 10 15 20 25 0 5 10 15 20 25
Time (s) Time (s)
(a)
20 20
Reference Reference
15 PIF Method 15 PIF Method
PIO Method PIO Method
10 10
Estimated Angle (deg)
0 0
-5 -5
-10 -10
-15 -15
-20 -20
0 5 10 15 20 25 0 5 10 15 20 25
Time (s) Time (s)
(b) (c)
20 20
100% Cf 100% Cf
15 150% Cf 15 150% Cf
50% Cf 50% Cf
10 10
Estimated Angle (deg)
5 5
0 0
-5 -5
-10 -10
-15 -15
-20 -20
0 5 10 15 20 25 0 5 10 15 20 25
Time (s) Time (s)
(d) (e)
Figure 4. Simulation results (S-curve manoeuvre on a flat road): (a) driver inputs; (b) estimated u (m = 0:85); (c) estimated u
(m = 0:3); (d) robustness of PIF with the modified vehicle model (m = 0:85); (e) robustness of PIF with the original vehicle model
(m = 0:85).
PIF: proportionalintegral HN filter; PIO: proportionalintegral observer; DFC: dynamic filter compensation; UIO: unknown input observer.
790 Proc IMechE Part D: J Automobile Engineering 226(6)
150
Long. Vehicle Speed
v (kph)
145
x
140
10
Steering Wheel Angle
5
f (deg)
-5
5
ay (m/s 2), (deg/s)
Lateral Acceleration
Yaw Rate
0
-5
0 5 10 15
Time (s)
(a)
70 70
Reference w/ LPF Reference w/ LPF
60 PIF Method 60 PIF Method
PIO Method PIO Method
Estimated Angle (deg)
30 30
20 20
10 10
0 0
0 5 10 15 0 5 10 15
Time (s) Time (s)
(b) (c)
Figure 6. Test results (lane 1): (a) vehicle states from sensors; (b) estimated u (Cf = Cf ss ); (c) estimated u (Cf = 1:2Cf ss ).
Long.: longitudinal; w/ LPF: with a low-pass filter; PIF: proportionalintegral HN filter; PIO: proportionalintegral observer; DFC: dynamic filter
compensation; UIO: unknown input observer.
Conclusion
on the game theory approach, which is designed for
This paper presents a new robust road bank angle the worst cases with respect to the sensor noises and
estimation method that is based on a PIF and does disturbances, is used as the estimator in order to
not require expensive sensors such as the DGPS. In improve further the stability and robustness of the
this work, the robustness of the estimation was bank estimation. Simulations and actual vehicle tests
enforced by the use of a more robust system model are conducted for various road and vehicle driving
and a robust estimation algorithm. A modified bicycle conditions. The simulation and vehicle test results
model, which reduces the model uncertainty by elimi- showed that the proposed PIF method provides the
nating the lateral force term of the front tyre from the best accuracy and robustness against the model uncer-
system equation, was derived and used to design a tainties compared with the previous methods, such as
road bank angle estimation algorithm. A PIF based the DFC and the UIO methods.
792 Proc IMechE Part D: J Automobile Engineering 226(6)
160
Long. Vehicle Speed
150
v (kph)
140
x
130
120
-100
40
ay (m/s2), (deg/s)
Lateral Acceleration
20 Yaw Rate
-20
0 5 10 15 20 25
Time (s)
(a)
60 60
Reference w/ LPF Reference w/ LPF
50 PIF Method 50 PIF Method
PIO Method PIO Method
40 40
Estimated Angle (deg)
20 20
10 10
0 0
-10 -10
-20 -20
0 5 10 15 20 25 15 20 25
Time (s) Time (s)
(b)
60 60
100% Cf 100% Cf
50 150% Cf 50 150% Cf
50% Cf 50% Cf
40 40
Estimated Angle (deg)
30 30
20 20
10 10
0 0
-10 -10
-20 -20
15 20 25 15 20 25
Time (s) Time (s)
(c) (d)
Figure 7. Test results (lane 2 $ lane 3): (a) vehicle states from sensors; (b) estimated u; (c) robustness of PIF with the modified
vehicle model; (d) robustness of PIF with the original vehicle model.
Long.: longitudinal; w/ LPF: with a low-pass filter; PIF: proportionalintegral HN filter; PIO: proportionalintegral observer; DFC: dynamic filter
compensation; UIO: unknown input observer.
Kim et al. 793