Sunteți pe pagina 1din 15

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO.

2, MARCH 2007 529

A Genetic Fuzzy Controller for Vehicle


Automatic Steering Control
Lin Cai, A. B. Rad, and Wai-Lok Chan

Abstract—This paper presents the design and experimental consisting of a scaled vehicle running on a simulated road
implementation of a genetic fuzzy controller for automatic steering surface was studied. The vehicle received control commands,
of a small-scaled vehicle. We first derive a dynamic model of the such as steering angle servo input and motor speed, via radio-
vehicle via system identification and show that the model exhibits
similar characteristics to full-sized vehicles. Subsequently, a stable frequency signals while it was put on a treadmill to simulate a
fuzzy proportional-derivative controller is designed and optimized roadway. The parameters of vehicle model were measured by
by genetic algorithms. The control system is transformed into serial experimental testing. In a parallel development, one may
a Luré system, and Lyapunov’s direct method is used to guar- notice a growing volume of research reported on the vehicle
antee the stability of the control system. Experimental studies control, and various control algorithms have been devised and
suggest that the control system is insensitive to parametric uncer-
tainty, load, and disturbances. The performance of the proposed employed, including the incorporation of conventional control
controller is also compared against a conventional proportional and modern control theory [4]–[9].
derivative (PD) controller. Experimental results confirm that it In this paper, we develop a 1/10th scaled vehicle with com-
outperforms the conventional PD controller, particularly in terms puter and sensors onboard. The vehicle is actually a modi-
of robustness. fied radio-controller (RC) toy car. It can independently drive
Index Terms—Automatic steering, genetic fuzzy control, system through any road conditions. The lateral dynamic model of the
identification, vehicle control. vehicle is obtained via a system identification method from
extensive experimental data. We will show that it exhibits
I. I NTRODUCTION similar characteristics with real vehicle models. We suggest
that the controller designed and implemented on this vehicle

A UTOMATIC steering control is an integral part of intelli-


gent vehicle control systems. In recent years, this problem
has been investigated worldwide by many research groups,
should give comparable results if implemented on real vehicles.
Moreover, the controller would not have to be drastically altered
when implemented on an actual vehicle because of similarities
especially in the U.S., Japan, and Europe [1]. The control in the dynamic model.
system is usually designed to detect any difference between the The rest of this paper is organized as follows: The overview
host vehicle and the reference line on the road via on-board and description of the scaled vehicle are presented in Section II.
sensors, which provide the feedback signals. In general, the Then, vehicle lateral dynamic modeling work is described in
references on the road are grouped into look-ahead and look- Section III. Section IV outlines the genetic fuzzy PD controller
down systems according to the point of measurement of vehicle design. Transformation of the vehicle control system into a
lateral displacement from the reference lane [2]. Luré system and stability analysis of the proposed controller
Much of the published research on novel vehicle control are given in Section V. Experiment results are presented in
algorithms has been limited to simulation studies and has not Section VI. Finally, a summary of the results and the conclusion
provided actual performance in practical settings. It can be are given in Section VII.
argued that a simulation environment cannot take into account
numerous parameters, unknown disturbances, and complete
dynamics of the vehicle. In addition, many research groups may II. E XPERIMENTAL V EHICLE S YSTEM O VERVIEW
not be able to experiment on a full-sized vehicle due to safety The prototype vehicle shown in Fig. 1 is a modified commer-
measures, space, and high expenses. Therefore, we decided to cially available 1/10th scaled RC model car, which is driven and
develop our own small-scale vehicle to try out longitudinal steered by two front wheels. We have added hardware and sen-
and lateral control algorithms. In [3], an experimental system sors to the original design while preserving most of its original
hardware such as chassis, motor, drivetrain, and batteries. Two
Manuscript received September 18, 2003; revised February 1, 2005, April infrared sensors and one ultrasonic sensor are used to measure
13, 2005, June 22, 2005, July 19, 2005, September 10, 2005, November 1,
2005, January 14, 2006, and February 22, 2006. This work was supported by
the lateral and longitudinal displacements between the host
Hong Kong Polytechnic University. The review of this paper was coordinated vehicle and objects around it, respectively. Vehicle longitudinal
by Dr. M. S. Ahmed. speed is detected by an incremental encoder sensor installed on
L. Cai is with Beijing Huafeng Company, Beijing, China.
A. B. Rad and W.-L. Chan are with the Department of Electrical Engi- the shaft of the rear wheels. A Pentium III level processor—
neering, Hong Kong Polytechnic University, Kowloon, Hong Kong (e-mail: industrial computer ADVANTECH PCM-9370—provides
eeabrad@polyu.edu.hk). real-time computing and control center. ADVANTECH PCM-
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. 3718HG with 12-bit A/D conversion is used to convert the
Digital Object Identifier 10.1109/TVT.2006.889576 continuous sensor signals to digital form for control design.
0018-9545/$25.00 © 2007 IEEE
530 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 2, MARCH 2007

Fig. 1. Experimental vehicle after modification.

III. V EHICLE S YSTEM M ODELING the displacements yf and yr of the vehicle from the reference
road. A linearized state-space model for lateral dynamics of the
The acquisition of a high-fidelity model of the vehicle is a
front and rear lateral displacement was derived from the general
crucial step toward the successful design of a high-performance
bicycle model as [2]
control system. However, it is a challenging process to identify
vehicle lateral dynamics due to its multi-input multi-output
    
characteristics, nonlinearities, sensors’ noise, and various dis- yf 0 1 0 0 yf
turbances. A considerable amount of work has been published d  ẏf   a21 a22 −a21 a24   ẏf 
 =  
in the field of modeling vehicles since the 1950s. For a small- dt yr 0 0 0 1 yr
scaled vehicle, a dynamic model was obtained by the frequency ẏr a41 a42 −a41 a44 ẏr
 
domain method and verified by step response in [10]. Since 0 0  
that car was driven on a treadmill, its body did not move.  b21 b22  δf
+  (1)
Therefore, it was rather easy to connect the sensors’ output 0 b32 ρref
to a signal analyzer through wires. However, this method was b41 b22
not convenient for our experimental vehicle, which actually µ(cf + cr ) df µ(cr lr − cf lf )
navigated on an experimental road. The so-called bicycle model a21 = −
M (df + dr ) J(df + dr )
is widely adopted for control purposes. In [3], a vehicle model
µ(cr lr − cf lf ) − dr µ(cf + cr )
was derived via measuring parameters of the bicycle model, and a22 =
the measuring methods were described on how to obtain gravity M ν(df + dr )


location, moment of inertia, and cornering stiffness. Although df dr µ(cr lr − cf lf ) − µ cr lr2 + cf lf2
+
some parameters were experimentally measurable, i.e., vehicle Jν(df + dr )
speed, mass, and moment of inertia, other parameters were µ(cr lr − cf lf ) + df µ(cf + cr )
obviously difficult to measure. a24 = −
M ν(df + dr )
This section is intended to provide a dynamic description

of the small-scaled vehicle for front-wheel automatic steering df df µ(cr lr − cf lf ) + µ cr lr2 + cf lf2
+
control through the system identification method. When Jν(df + dr )
modeling, it has been often attempted to borrow a full-size µ(cf + cr ) dr µ(cr lr − cf lf )
vehicle as the template for small-size vehicle dynamics, a41 = +
M (df + dr ) J(df + dr )
assuming that the dynamics for full size and small size are µ(cr lr − cf lf ) − dr µ(cf + cr )
similar. Some prior knowledge has been provided to select the a42 =
M ν(df + dr )
model structure and model order.

We first describe the theory of the bicycle model for a full- dr dr µ(cr lr − cf lf ) − µ cr lr2 + cf lf2

size vehicle. Then, an experiment will be conducted to obtain Jν(df + dr )
input/output data, and subsequently, a system identification µ(cr lr − cf lf ) + df µ(cf + cr )
method would be designed to get a lateral model of the target a44 = −
M ν(df + dr )
small-scaled vehicle.

dr df µ(cr lr − cf lf ) + µ cr lr2 + cf lf2

Jν(df + dr )
A. Vehicle Bicycle Model 
1 d f lf
b21 = µcf +
The bicycle model is widely used for the purpose of vehicle M J
lateral control. Although it is relatively simple, it has been ver- 
1 d r lf
ified to be a good approximation of full-size vehicle dynamics b41 = µcf −
M J
for small steering angle. For the prototype vehicle considered
here, two sensors were mounted at df for the front and dr for b22 = −ν 2
the back with respect to the center of gravity (CG) to measure b32 = ν(df + dr )
CAI et al.: GENETIC FUZZY CONTROLLER FOR VEHICLE AUTOMATIC STEERING CONTROL 531

Fig. 2. Block diagram of the complete steering system.

where the parameters are defined to a full-sized vehicle. Therefore, the control algorithm
used on the prototype vehicle is meaningful and useful
ν vehicle velocity vector with ν = |ν| > 0; speed ν is
for application on real vehicles. On the other hand, since
assumed measurable;
an in-depth study is carried out based on a continuous
δf front wheel steering angle;
bicycle model, it would give us prior knowledge to
ρref reference road curvature;
build an appropriate model structure. From the analysis
M total vehicle mass;
of the model (1), we know that the model is unstable
J total vehicle inertia about vertical axis at CG;
because of two integrations. Therefore, a closed-loop
lf (lr ) distance of front(rear) axle from CG with l =
experiment has to be designed. Moreover, for the multi-
lf + l r ;
output system, the state-space model is more appropriate
cf (cr ) front(rear) tire cornering stiffness;
than other structures. The model is treated as a gray box,
df (dr ) distance of front(rear) sensors from CG;
and parameters are identified by the MATLAB system
µ road adhesion as a factor of effective tire cornering
identification toolbox.
stiffness c∗f = µcf (c∗r = µcr ).
2) Indirect closed-loop identification approach: As in Fig. 2,
Many plant parameters such as mass of the vehicle, inertia, an indirect approach to identify the closed-loop system
and location of the CG are constant and slowly varying. They from reference input r(t) to output [yf (t) yr (t)] was
can be easily measured by physical methods. As for vehicle performed. The open-loop model could be retrieved since
speed, it is experimentally measurable by the encode sensor. the regulator was known. This approach was selected
The cornering stiffness and road adhesion describe the rela- to derive the vehicle model using a simple proportional
tionship between wheel steering force and wheel angle and controller since the open loop was unstable. Its basic
are difficult to measure. They are constant for small steering advantage is that the dynamics model can be correctly
angles. From (1), we obtain a fourth-order model for the estimated without estimating any noise model, even when
vehicle with two sensors. The parameters a21 , a41 , b21 , and it is unstable [11].
b41 do not change with speed. The matrices A and B have a 3) Parameter estimate method: The prediction error method
special structure. All prior knowledge is incorporated in the (PEM) [11] was selected to estimate parameters. The
modeling work. iterative search algorithm is the damped Gauss–Newton
The typical bicycle model for full-sized vehicles often does iterative method.
not include the dynamics related to the steering actuator. The 4) Choice of input and sampling interval: A periodic square
steering command must be directed through a motor, gears, wave input was selected to provide a persistently exciting
and other actuators, whose response may not be immediate or input.
exact, which results in a response that is far different from that 5) Sampling interval: The operating frequency was in the
predicted by the bicycle model. range of 0–3 Hz. Moreover, during the experiment, the
onboard industrial computer on the vehicle fixed the sam-
B. System Modeling Work pling interval for hardware interrupts. Therefore, the sam-
pling time should be selected as integral multiples of
Before starting the modeling experiment, let us discuss the 1/18.2 s. Considering the sensors, computation of control
purpose of the model, its structure, and factors that must be signals, etc., the sampling time was chosen as 1/18.2
considered for identification. (0.055) s.
1) Continuous state-space model: It is better to find a contin- 1) Actuator Modeling: Accurate measurement of the steer-
uous model like (1) for our target vehicle, since the para- ing angle is difficult without an encoder installed on the shaft.
meters have clear physical interpretation. The purpose of We used an indirect method of correlating servo motor poten-
the continuous state-space model lies in two aspects: One tiometer voltage with the steering angle (c). Fig. 3 shows the
is to illustrate that the scaled vehicle has a similar model functional relationship between measured voltage and steering
532 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 2, MARCH 2007

Fig. 3. (a) Static and, (b) dynamic diagram of steering servo potentiometer voltage.

Fig. 4. Block diagram of proportional feedback control system.

command. The static linear equation between steering com- When considering the dynamics of the steering actuator, the
mand and voltage was obtained via the least square method as vehicle model can be transformed into (4) with the assumption
that ρref = 0. The number of state variables is increased to five
δf = 0.00044∗ c (2) by adding δf (t)
   
where c is the steering command. yf 0 1 0 0 0
−a21
d  f  21
In order to find a dynamic model of the actuator, a step ẏ a a22 a24 b1 
   
response was recorded in Fig. 3. A continuous first-order model  yr  =  0 0 0 1 0 
dt    
was fitted based on the experimental data. Combined with (2), ẏr a41 a42 −a41 a44 b2
the resulting linear model was obtained as δf 0 0 0 0 −24
   
yf 0
0.0106  ẏf   0 
δf (s) = c(s). (3)    
s + 24 ×  yr  +  0  0.00044c (4)
   
ẏr 0
Fig. 3 also shows the similarity between the actual response
δf 24
obtained from experiment data and that predicted from the first-
order model. It should be noted that the actuator is nonlinear where b1 = b21 , b2 = b41 .
due to a dead zone. In this paper, we neglect it and use the linear From Fig. 4, we can obtain the following equation:
model (3) to describe the dynamic of the actuator.  
2) Vehicle Modeling: A continuous state-space model of dr df
c = ke r1 − yf + yr +ky [r2 − (yf − yr )]
closed-loop system is obtained using PEM via experimental df + dr df + dr
 
input/output data. The process to obtain the closed-loop system yf
model with proportional controller is presented below.     ẏf 
r1 dr dr  
A detailed block diagram of the closed-loop system is shown = [ke ky ] − ke + kf 0 ke − kf 0 0  yr .
r2 df + dr df + dr  
in Fig. 4. The first reference input (r1 ) is the desired lateral ẏr
distance of the CG, which is a periodic square wave. The second δf
reference input (r2 ) is the yaw angle, which is set to be zero. (5)
CAI et al.: GENETIC FUZZY CONTROLLER FOR VEHICLE AUTOMATIC STEERING CONTROL 533

Substituting (5) into (4), the state-space equation of the in Table I by the least square method. Therefore, the vehicle
closed-loop system is obtained as (6), shown at the bottom of model is
the page.     
Eight parameters are to be estimated as a21 , a22 , a24 , a41 , yf 0 1 0 0 yf
d  ẏf   251.64 −90 50
a42 , a44 , b1 , and b2 . Obviously, it is easier to identify a model −251.64   ẏf 
 = ν ν  
with fewer parameters. From previous knowledge of this model, dt yr 0 0 0 1 yr
−1.92 −40
we know the following. ẏr 239.56 ν −239.56 ν ẏ r
 
1) 0 0  
 63.77 −ν 2  δf
+  (8)
a42 = (a22 + a24 )
a41
− a44 . (7) 0 0.2ν ρref
2
a21 −6.67 −ν

2) a21 , a41 , b1 , and b2 will not change with variance longi- where ν is the longitudinal speed of the vehicle in meters
tudinal speed. per second. Note that the variance of road adhesion is not
3) a22 , a24 , a42 , and a44 are inversely proportional to longi- considered in this model (it is assumed to be unity, i.e., µ = 1).
tudinal speed.
Based on all prior knowledge, we built a gray-box model
IV. G ENETIC F UZZY C ONTROL D ESIGN
for the closed-loop system (in a MATLAB environment) to
estimate the eight parameters. Experimental input/output data The controller presented in this section focuses attention on
at different speeds were collected. The model obtained is listed achieving good tracking for difference road curvature over a
in Table I, and Fig. 5 shows a comparison between measured range of longitudinal speeds of the vehicle. The controller was
output and the simulated model output at different speeds for broken into two parts: a feedback controller and a feed-forward
ke = ky = 30. controller. A fuzzy PD controller was designed as the feedback
From matrices A and B in Table I, we infer that a21 = controller, which considered the input of the current and past
251.64, a4 = 239.56, b1 = 63.77, and b2 = −6.67. Since a22 , conditions of measured deviation and generated as output an
a24 , a42 , and a44 are inversely proportional to longitudinal appropriate steering angle. Feed-forward control was used to
speed, we can obtain the proportional parameters from the value assist the controller when the vehicle engaged in a curved

   
yf yf
 ẏf   ẏf   
d     r1
 yr  = Ac  yr  + Bc
dt     r2
ẏr ẏr
δf δf
 
yf
   ẏf   
yf   r1
= Cc  yr  + Dc (6)
yr   r2
ẏr
δf
 
0 1 0 0 0
 a21 a22 −a21 a24 b1 
 
Ac = 
0 0 0 1 0 

 a41
  a42 −a41  a44 b2 
−0.00044 × 24 ke df d+d
r
r
+ kf 0 −0.00044 × 24 ke df d+d
r
r
− ky 0 −24
 
0 0
 0 0 
 
Bc =  0 0 
 
0 0
0.00044 × 24 ke 0.00044 × 24 ky
 
1 0 0 0 0
Cc =
0 0 1 0 0
 
0 0
Dc =
0 0
534 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 2, MARCH 2007

TABLE I
OPEN-LOOP MODEL FOR EXPERIMENTAL VEHICLE AT DIFFERENT SPEED

section of roadway, assuming that curvature of the road was to design the initial scaling factors of the fuzzy controllers and
known in advance. This control was active only if the system decrease the search space of GA.
was warned that an upcoming curve was nearby (similar to the 2) Fuzzy Logic Controller Structure: The structure of the
road signs that warn drivers of the curved section ahead). two fuzzy controllers is similar: two inputs (error and change of
error), one output (steering angle), five triangular membership
functions for inputs, and seven singleton membership functions
A. Genetic Fuzzy PD Feedback Controller
for output, as shown in Fig. 7, respectively. The primary linguis-
In this section, we propose a fuzzy PD feedback controller tic terms have standard meanings, such as negative big (NB),
to guarantee the stability of the system by using the genetic negative small (NS), zero (ZE), positive small (PS), positive big
algorithm (GA) method to optimize the parameters of the (PB), and so forth. Here, we define the input linguistic variables
fuzzy control to guarantee performance. The proposed control (error, change of error), which fall into the same category of the
structure is shown in Fig. 6. fuzzy sets (NB, NS, ZE, PS, PB). The fuzzy set of the output
1) Why Fuzzy PD Control: The input information of the first linguistic variable is (nl, nm, ns, ze, ps, pm, pl). Mamdani infer-
fuzzy controller is the average of yf and yr , and the other is ence strategy is used, and the initial rule base is shown in Fig. 8.
the difference of yf and yr . The parameters of the two fuzzy In general, to obtain an optimal set of fuzzy scaling factors,
controllers are not the same. The motivation to construct two membership functions, and rules is not an easy task. It requires
such fuzzy controllers is to derive fuzzy rules from human time, experience, and skills for the tuning exercise. Using GA
driving experience. We can easily determine the initial fuzzy methods to optimal fuzzy system parameters have become a
rules of the first controller to keep the center of the vehicle trend for fuzzy logic system development. Suitable scaling
following the desired path and the second controller to make the factors, membership functions, and rules of the fuzzy logic
angle between the centerline of to the vehicle and path tangent controller correspond to high/low fitness value (good system
to zero, which have a similar function with yaw rate angle. performance). We find optimal parameters via GA tuning such
There are two reasons for selecting fuzzy PD structure in- that the trial time greatly decreases under expert’s knowledge.
stead of PID/PI configuration. First, the system steady state 3) GA Optimal Design: Some of the rules can be inferred
is zero, and large overshoot is prohibited. Second, the two from human driving experience. From simulation and practical
fuzzy PD controllers can be seen as state feedback if the fuzzy experience, we know that scaling factors have the best effect on
mapping is linear. Therefore, we can use linear control theory the response performance than individual membership function
CAI et al.: GENETIC FUZZY CONTROLLER FOR VEHICLE AUTOMATIC STEERING CONTROL 535

and parameters of membership functions separately is not a


good method to find the optimal solution. In this paper, the
scaling factors and membership functions are tuning via GA
simultaneously.
A problem that comes to light in the optimization is the use
of mathematical models to evaluate the fitness of a given string.
One of the strong points of fuzzy controllers is the fact that they
do not require mathematical models. However, to obtain fitness
for a given controller, the GA must have a method to evaluate
the controller’s performance; in this sense, we have negated one
of the fuzzy controllers’ advantages in order to use the power
of GAs to optimize this controller.
On the other hand, the vehicle lateral model has been studied
for several decades, and some of them, such as the famous bicy-
cle model, have been applied successfully by control engineers.
With the explicit mathematical model, the traditional control
analysis method can be used to give us a clear understanding
of the plant. Therefore, we choose the GA-based fuzzy control
method, which is model dependent.
a) Decision for coding and decoding: For scaling factors,
as shown in Fig. 6, there are six scaling factors to be tuned. They
are denoted as Kf = [k1 , k2 , k3 , k4 , ku1 , ku2 ]. For membership
functions, a simple and efficient method is proposed to tune
them. The only point to be tuned is the middle point of the
triangle-type fuzzy subset. Considering that the middle point of
the ZE fuzzy subset should be zero, and −1/ + 1 for NB/PB,
respectively, the five triangle-type fuzzy subsets, as shown in
Fig. 7, can be parameterized as [−1, −0.5, 0, 0.5, 1]. Then,
using GA operators to tune the above parameters, we can find an
optimal solution. Also, this method can be used to parameterize
the fuzzy sets of output. Compared with other coding methods,
the obvious benefit of this method is that it is computationally
efficient and less time consuming.
b) Definition of fitness function: The two important
points in GA algorithm design are how to code the chromosome
and definition of the fitness function. To achieve good tracking
performance, the control attempts to minimize a function J that
includes a term penalizing both the front and rear deviation
of the vehicle; the third term in (9) penalizes the feedback
control commands. This will prevent a high rate using control
commands, and the resultant control actions will be smooth

J = w1 (yf )2 + w2 (yr )2 + w3 (δf b )2 (9)

where w1, w2, and w3 are the weight number of three terms.
c) Evolutionary parameters setting: GA parameters are
given as number of generations = 50, population size = 50,
crossover factor = 0.9, mutation factor = 0.05, and binary bits
of parameters = 6.

Fig. 5. Comparison of measured output and simulated model output at 50, 90, B. Feed-Forward Controller
and 130 cm/s, respectively (ke = ky = 30).
The feed-forward controller is used when the vehicle engages
parameters. From the viewpoint of control engineering, the a curved section of roadway. If the curvature of the coming road
function of scaling factors is global when compared with the is available, the performance of the controller will be further
individual membership functions’ parameter, whose functions improved. A steady-state steering angle δff , which corresponds
are local. However, the scaling factors and membership func- to vehicle speed and road curvature, is added to δfb , which is
tions are codependent. Using GA to optimize scaling factors determined by the feedback controller in Fig. 9.
536 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 2, MARCH 2007

Fig. 6. Block diagram of two fuzzy controllers system.

Fig. 7. Initial input and output membership functions.

In this equation, we assume that the mass of the vehicle


lf , lr , cr , and cf is time invariable. Neglecting the change of
µ, δff can be seen as a linear function of ρref and ν 2 .
After adding the feed-forward controller, the steering angle
of the lateral controller is
δf = δfb + δff . (11)

V. S TABILITY A NALYSIS
Fig. 8. Rule table of fuzzy PD controller.
Fuzzy control has been applied successfully to a number of
practical applications. However, an efficient analytical tool to
analyze the stability of fuzzy control systems is still lacking.
Many efforts have been made to improve this problem. Robust
stability of fuzzy control systems has been studied through
the Popov–Lyapunov approach in [12]. First, the fuzzy control
system is transformed into a Luré system with uncertainties.
Then, Lyapunov’s direct method is used to guarantee the sta-
bility of the perturbed Luré system. However, it only dealt
with scalar output and one nonlinear feedback. In this paper,
Fig. 9. Feedback and feed-forward controller block diagram. we have designed two fuzzy controllers, and hence, the overall
control system has multiple nonlinearities. The stability of Luré
Assuming that the radius of a curved section is known, the systems with multiple nonlinearities has been studied in [13],
steering angle can be set constant at curved sections. From the as shown in Fig. 10. We employ the two theorems introduced
vehicle model (1), the steady-state steering angle δff can be therein to show the stability of the proposed control system. The
calculated for a given ρref as nonlinear feedback system is represented by

mν 2 (cr lr − cf lf ) + µcf cr (lf + lr )2 Ẋ = AX + BΦ(σ) + g(X, u)


δff = ρref . (10)
µcf cr (lf + lr ) σ = C TX (12)
CAI et al.: GENETIC FUZZY CONTROLLER FOR VEHICLE AUTOMATIC STEERING CONTROL 537

Lemma 1 [14]: If, for any i = j(i, j = 1, 2, . . . , m),


∂Φi /∂σj (•), is continuous, and the equation

∂Φi ∂Φj
(σ) = (σ) ∀σ ∈ Rm (18)
∂σj ∂σi

holds, then there exists a scalar function V satisfying (17).


The differentiability of Φ is not required when each com-
ponent of Φ is a single-variable function Φi (σ) = Φi (σi ),
i = 1, 2, . . . , m.
According to (12), the state-space equation of the system is
given as
Fig. 10. Block diagram of Luré systems with multiple nonlinearities.
Ẋ = A0 X + B0 u + g0 (X, u)
where X ∈ R and σ ∈ R denote the state and output of
n m
Y = C0 X (19)
the linear part, respectively. A is an n × n stable matrix, and
g(X, u) is the disturbance. where X = [yf ẏf yr ẏr ]T , Y = [yf yr ]T , u = δf . The change of
The transfer function of the linear part of system (12) is road condition can be treated as a disturbance.
given by Assume that the system satisfies the condition

W (s) = C T (sI − A)−1 B. (13) C0 B0 = C0 g0 = 0. (20)

Consider the transfer function Letting the reference input be zero, the following relations
can be obtained:
  1 1
  1 1
  
Z(s) = (n + qs)W (s) (14) ye yf C01
=− 2 2 =− 2 2 C0 X = − X
yy 1 −1 yr 1 −1 C02
where s = −(n/q) is not a pole of W (s). Then, according (21)
to Miyagi and Yamashita [13], we have the following two where ye is the average of yf and yr , and yy is the difference of
theorems. yf and yr .
Theorem 1 [13]: Supposing that (A, B, C) is a minimal
realization of W (s), then if Z(s) in (14) is positive real, there
ye = − C01 X (22)
exist real matrices P , L, and W0 with P as a positive definite
symmetric such that ẏe = − C01 Ẋ = −C01 (A0 X + B0 u + g0 )

AT P + P A = − LLT = − C01 A0 X (23)


P B = nC + qAT C − LW0 σe = σe1 + σe2 = −(k1 C01 + k2 C01 A0 )X (24)
W0T W0 T
= q(C B + B C). T
(15) σy = σy1 + σy2 = −(k3 C02 + k4 C02 A0 )X. (25)
Theorem 2 [13]: If the transfer function Z(s) is positive
Once the structure and parameters of the fuzzy controller are
real, then the null solution of system (12) is stable, under the
determined, the typical relation between fuzzy output Φ0 and
conditions 1), 2), and 3) for Φ(σ).
input σ satisfies the sector condition [12]
1) Φ(σ) is continuous and has the separate nonlineari-
ties such that ΦT (σ) = [Φ1 (σ1 ), Φ2 (σ2 ), . . . , Φm (σm )],
0 ≤ σΦ ≤ Kσ 2 (26)
where Φk (σk ) for k = 1, 2, . . . , m maps R into R.
2) For some nonnegative constant n
where Φ(σ) = (Φ0 − Kmin σ), K = Kmsx − Kmin , Kmin =
nσk Φk (σk ) > 0 (16) inf σ1 ∈[−1,+1] kmin (σ1 ), Kmax = inf σ1 ∈[−1,+1] kmax (σ1 ), and
kmin and kmax are corresponding to a certain value of σ1 . This
for k = 1, 2, . . . , m, and Φk (0) = 0. means that the nonlinearity Φ belongs to the sector [0, K].
3) There is a scalar function V (σ) such that V (σ) > 0 for For our two fuzzy controllers, we denote the sector variable
all σ ∈ Rm and ν(0) = 0, which satisfies as Kmin e , Kmin y , Kmax e , and Kmax y , respectively. Then,
we define
∇V (σ) = qΦ(σ) (17)
Φe (σe ) = (Φe0 − Kmin e σe ) (27)
for all σ ∈ Rm , where q is a positive constant, and
∇ = ∂/∂σ. Φy (σy ) = (Φy0 − Kmin y σy ). (28)
538 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 2, MARCH 2007

TABLE II
VEHICLE PARAMETERS

After these definitions, the following equation can be As shown in Fig. 12, the nonlinear relationships of the two
obtained: fuzzy controllers satisfy properties 1), 2), and 3) of Theorem 2.
We can obtain Kmin e = 0.5, Kmin y = 0.5, Kmax e = 5, and
u = δf = ku1 Φe0 + ku2 Φy0 Kmax y = 2. Using (30), and neglecting disturbance, the Luré
system with two nonlinearities can be obtained as follows:
= ku1 (Φe + Kmin e σe ) + ku2 (Φy + Kmin y σy )  
0 1.0 0 0
= ku1 Φe + ku2 Φy  9.3117 −3.8235 −10.9673 1.1972 
Ẋ =  X
0 0 0 1.0
− ku1 Kmin e (k1 C01 + k2 C01 A0 ) 32.0915 0.0975 −31.7121 −3.3033
 
+ ku2 Kmin y k3 C02 + k4 C02 A0 ) X. (29) 0 0  
 63.9243 52.1863  Φe
+ 
0 0 Φy
Substituting (29) in to (19), we can get
−14.6479 −11.9582
   
Ẋ = AX + BΦ(σ) + g(X, u) σe −0.0259 −0.0116 −0.0259 −0.0116
= X.
σy −0.2618 −0.0536 0.2618 0.0536
σ = C TX (30)
(31)
where A = A0 −ku1 Kmin e (k1 C01 +k2 C01 A0 )−ku2 Kmin y ×
(k3 C02 + k4 C02 A0 ) The pair (A, B) in (30) is controllable for rank
[B AB A2 B A3 B] = 4, and (C, A) is observable for
  rank [C CA CA2 CA3 ] = 4. The denominator polynomial of
T k1 C01 + k2 C01 A0
B = B0 kuq ku2 , C =− . transfer function W (s) is
k3 C02 + k4 C02 A0
D(s) = s4 + 7.1267s3 + 34.9136s2 + 53.1423s + 56.6646.
Once the structure and parameters of fuzzy controller are (32)
determined via GA optimization, the controlled system in Fig. 6
can be transferred to Luré systems with multiple nonlinearities According to (14), let n = 1 and q = 1. It can be shown
in Fig. 10, which is described by (30). Observing the nonlinear that Z(s) satisfies the requirement of a positive real transfer
functions Φ(σ), if it satisfies conditions 1), 2), and 3), the function (see remark 2). Hence, according to Theorems 1 and 2,
stability of system can be analyzed via Theorems 1 and 2. the system is verified to be stable. We can obtain the same
For the vehicle model described by (1), the typical vehicle conclusion for the experimental car using this method.
model is used to analyze the stability via numeric simulation Remark 2: Here, we use a definition of positive real function,
[2]. The parameters are selected as in Table II. All the pa- as suggested in [15, Def. 6.4, p. 237]: A p × p proper rational
rameters are constant to slowly time varying and assumed to transfer function matrix G(s) is called positive real if the
be known, except for the road adhesion factor µ, which may following are satisfied.
change abruptly while driving. 1) Poles of all elements of G(s) are in Re[s] ≤ 0.
The nominal matrices A0 , b0 , and C0 of (19) can be obtained 2) For all real ω for which jω is not a pole of any element of
under the condition of µ = 1 and ν = 40 m/s, and g(X, u) G(s), the matrix G(jω) + GT (−jω) is positive semidef-
stands for parameter uncertainties. inite.
Remark 1: Model (1) with parameters in Table II and model 3) Any pure imaginary pole jω of any element of G(s)
(8) are structurally the same; however, their parameters are is a simple pole, and the residue matrix lims→jω (s −
different. Model (1) is applied with a real car’s parameters jω)G(s) is positive semidefinite Hermitian.
(obtained from the literature), whereas model (8) incorporates
the experimental car’s parameters. We can use the same method
VI. E XPERIMENT R ESULTS
to analyze the stability of the two models and obtain the same
conclusion. An experimental test road was built that consisted of a
We constructed a Genetic-Fuzzy controller for the real car straight section and, a right turn, followed by a left turn, as
model using the method discussed in Section IV. After GA shown in Fig. 11. The turning radii are Rref = 1.31 m and
optimization, the final value of Kf is Kf = [0.0518, 0.0231, Rref = 2.87 m, respectively. We assumed that of road con-
0.2618, 0.0536, 1.1529, 0.9412], and the membership func- ditions were known. As demonstrated earlier in Section III,
tions parameters are [0.1529, 0.3647, 0.8412, 0.8412, 0.5765, the vehicle is very sensitive to velocity. As a consequence,
0.8941] in Fig. 11. our controller should guarantee sufficient performance and
CAI et al.: GENETIC FUZZY CONTROLLER FOR VEHICLE AUTOMATIC STEERING CONTROL 539

Fig. 11. Final fuzzy membership functions of the feedback controller.

Fig. 12. Nonlinear relationships of two fuzzy controllers.

robustness to the variable velocity. To illustrate this condition, the membership functions’ parameters were [0.4706 0.5882
the experiment was conducted at different velocities of 50, 90, 0.2353 0.2941 0.5882 0.7059]. Experiments had been done
and 130 cm/s in the curved sections. using the optimized parameters, and the experimental results
After GA optimization, the final values of the scaling fac- of adjusting the scaling factor and membership function in
tor Kf were [0.2353 0.0588 435 0.6471 0.1521 458], and a small neighbor area around the optimized values were
540 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 2, MARCH 2007

Fig. 13. Genetic fuzzy control performance of genetic fuzzy controller at speed = 50 cm/s.

Fig. 14. Genetic fuzzy control performance of genetic fuzzy controller at speed = 90 cm/s.

compared. We found a little better result under the tuned Kf = accepted. Figs. 13–15 show the system performance with feed-
[0.2 1 400 0.2 0.5 400]. When considering neglected back and feed-forward controllers. The performances of the
dynamics and sensor noise, the modification of Kf can be genetic fuzzy controller and the PD controller are similar.
CAI et al.: GENETIC FUZZY CONTROLLER FOR VEHICLE AUTOMATIC STEERING CONTROL 541

Fig. 15. Genetic fuzzy control performance of genetic fuzzy controller at speed = 130 cm/s.

Therefore, further experiments under the conditions of un-


known disturbance and parameter uncertainty were carried on
to compare genetic fuzzy and PD controller. The system was
also tested for the unknown disturbance that simulated actual
traffic conditions. First, we assumed that there were speed
bumps in the road, which was simulated by a wire affixed in
the road, as shown in Fig. 16(a). From the simulation result
in Fig. 16(b), we see that vehicle speed decreased when it
passed the speed bumps, but the deviation of the CG was not
influenced. The solid line is the response of the genetic fuzzy
controller, and the dotted line is for the PD controller. Both PD
and genetic fuzzy controllers have good performance. Note that
the curved road is not changed, and only a “speed bump” is
added to the road.
Next, the actual mass vehicle was changed when adding
a passenger or load. We added a board on the experimental
vehicle to simulate this condition in Fig. 17. Experimental
results show that the system performance was not affected by
the change of vehicle mass. Fig. 18 shows the system outputs
of PD and genetic fuzzy controllers: a solid line for the genetic
fuzzy controller and a dotted line for the PD controller.
Finally, the most important problem is faulty sensors or
incorrect reflection of walls. This implies that the controller
cannot get reference information to generate the correct control
signals. In order to simulate this condition, we created a gap
in the wall, as shown in Fig. 19. The width of this gap was
about 6 cm, which was considerable when compared with the Fig. 16. (a) Traffic condition of speed bumps. (b) System performance when
vehicle dimensions (20∗ 40 cm). In Fig. 20, the deviation of the passed speed bumps.
542 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 2, MARCH 2007

Fig. 17. Vehicle without load and with load.

Fig. 20. Deviation of the CG.


Fig. 18. System performance with load.
TABLE III
J UNDER DIFFERENT CONDITION OF TWO CONTROLLERS

Beside these figures, we also used the experimental data to


compare the results of the PD controller and genetic fuzzy
control. A function J was used to calculate the sum of the
square of error. The experimental data were collected every
sampling time; therefore, the function J is discrete form:

J= (ygc − 20)2 . (33)

As tabulated in Table III, we can clearly deduce that the


Fig. 19. Six-centimeter gap on the wall. performance of the genetic fuzzy controller is better than the
PD controller, especially under gap conditions.

center of gravity looked like the passing of two gaps. This is


VII. C ONCLUSION
because ygc is calculated by average of yf and yr , which are
the front and rear sensors, which consequently passed the gap. This paper presents the design and experimental implementa-
Therefore, the wave of ygc has two pulses, and the interval of tion of a genetic fuzzy controller for automatic steering control
the two pulses is determined by the vehicle speed. As expected, system of a small-scaled vehicle. In the first step, using the
when comparing PD and genetic fuzzy controller performance standard vehicle model as a reference, the scaled vehicle model
(solid line for genetic fuzzy controller and dotted line for PD obtained through the system identification method shows a
controller), we found that the genetic fuzzy controller copes similar dynamic to the actual vehicle. A second step is to design
with such problems better because it was a nonlinear controller. a stable genetic fuzzy controller. The stability of the system
CAI et al.: GENETIC FUZZY CONTROLLER FOR VEHICLE AUTOMATIC STEERING CONTROL 543

was analyzed through the Luré system and directly through Lin Cai received the B.Eng. degree from the
Lyapunov’s direct method. Experimental results provided in Department of Automation Science and Electrical
Engineering, Beihang University, Beijing, China, in
this paper demonstrate the effectiveness and robustness of the 1998 and the M.Phil. degree from the Department
controller. The system performance has been compared with of Electrical Engineering, Hong Kong Polytechnic
a conventional PD controller in terms of disturbance (speed University, Kowloon, Hong Kong, in 2003.
Currently, she is the Manager with Beijing
bumps and wall gap) and changes in plant dynamics (change Huafeng Company, Beijing.
of vehicle mass). It was demonstrated that the genetic fuzzy
controller is superior to the PD controller.

R EFERENCES
[1] S. E. Shladover, “Review of the state of development of advanced
vehicle control systems (AVCS),” Veh. Syst. Dyn., vol. 24, no. 6–7,
pp. 551–594, 1995.
[2] J. Guldener et al., “Robust automatic steering control for look-down
reference systems with front and rear sensors,” IEEE Trans. Control Syst.
Technol., vol. 7, no. 1, pp. 2–11, Jan. 1999.
[3] S. Brennan and A. Alleyne, “Using a scale testbed: Controller design and
evaluation,” IEEE Control Syst. Mag., vol. 21, no. 3, pp. 15–16, Jun. 2001. A. B. Rad received the B.Sc. (Eng.) degree from
[4] H. Mouri and H. Furusho, “Automatic path tracking using linear quadratic Abadan Institute of Technology, Abadan, Iran,
control theory,” in Proc. IEEE Conf. Intell. Transp. Syst., Nov. 1997, in 1977, the M.Sc. degree in control engineer-
pp. 948–953. ing from University of Bradford, Bradford, U.K.,
[5] J. Ackermann, J. Guldner, W. Sienel, R. Steinhauser, and V. I. Utkin, in 1986, and the Ph.D. degree in control engineering
“Linear and nonlinear controller design for robust automatic steering,” from University of Sussex, Brighton, U.K., in 1989.
IEEE Trans. Control Syst. Technol., vol. 3, no. 1, pp. 132–143, Mar. 1995. He is currently a Professor with the Department of
[6] H.-S. Tan, J. Guldner, S. Patwardhan, C. Chen, and B. Bougler, “Devel- Electrical Engineering, The Hong Kong Polytechnic
opment of an automated steering vehicle based on roadway magnets—A University, Kowloon, Hong Kong. His current re-
case study of mechatronic system design,” IEEE/ASME Trans. Mecha- search interests include mechatronic systems, intel-
tronics, vol. 4, no. 3, pp. 258–272, Sep. 1999. ligent control and soft-computing methods, system
[7] T. Hessburg and M. Tomizuka, “Fuzzy logic control for lateral vehicle identification, and adaptive control.
guidance,” IEEE Control Syst. Mag., vol. 14, no. 4, pp. 55–63, Aug. 1994.
[8] K. R. S. Kodagoda, W. S. Wijesoma, and E. K. Teoh, “Fuzzy speed and
steering control of an AGV,” IEEE Trans. Control Syst. Technol., vol. 10,
no. 1, pp. 112–120, Jan. 2002.
[9] B. Hu, G. K. I. Mann, and R. G. Gosine, “New methodology for analytical
and optimal design of fuzzy PID controllers,” IEEE Trans. Fuzzy Syst.,
vol. 7, no. 5, pp. 521–539, Oct. 1999.
[10] M. Depoorter, “Development, experimentation, and control of a small
scale vehicle dynamics and control laboratory,” M.S. thesis, Univ. Illinois
at Urbana-Champaign, Urbana-Champaign, IL, 1997.
[11] L. Ljung, System Identification, Theory for the User. Upper Saddle
River, NJ: Prentice-Hall, 1999. Wai-Lok Chan received the B.Sc.(Eng) and M.Phil.
[12] C.-C. Fuh and P.-C. Tung, “Robust stability analysis of fuzzy control degrees from University of Hong Kong, Hong Kong.
systems,” Fuzzy Sets Syst., vol. 88, no. 3, pp. 289–298, Jun. 1997. In 2000, he received the Ph.D. degree from City
[13] H. Miyagi and K. Yamashita, “Robust stability of Luré systems with University, London, London, U.K.
multiple nonlinearities,” IEEE Trans. Autom. Control, vol. 37, no. 6, He is currently with the Department of Electri-
pp. 883–886, Jun. 1992. cal Engineering, Hong Kong Polytechnic University,
[14] T. Wada and M. Ikeda, “Extended popov criteria for multivariable Luré Kowloon, Hong Kong. His research interests include
systems,” in Proc. 32th Conf. Decis. Control, Dec. 1993, pp. 20–21. microprocessor applications in power systems and
[15] H. K. Khalil, Nonlinear Systems, 3rd ed. Upper Saddle River, NJ: applications of artificial intelligence.
Prentice-Hall, 2002.

S-ar putea să vă placă și