Professional Documents
Culture Documents
R
J
w
V
x
u 4
Eq. (4) clearly shows that, during braking, the slip ratio is
dependent on the input torque u and the vehicle velocity V
x
. In
state space, the system state variables are: x
1
S
x
, x
2
V
x
, x
3
l,
where S
x
is the stopping distance. The state space equations are
_ x
1
x
2
_ x
2
m
r
F
N
m
_ x
3
m
r
F
N
x
2
1x
3
m
R
2
J
w
_ _
R
J
w
x
2
u 5
This state space model in Eq. (5) has been used in the
simulation tests to evaluate the performance of the ABS, using
different PID control strategies.
During braking, it is assumed that the wheel radius is constant.
Also, the vehicle speed V
x
and the wheel angular velocity o are
available signals through transducers mounted on suitable places.
So that the slip ratio l, is an available parameter for the ABS
closed-loop system.
2.2. Braking basics and problem denition
The ability of the ABS to maintain vehicle stability and
steerability, and still produce shorter stopping distances than
those from a locked wheel stop, comes from the relation of the
adhesion coefcient m
r
versus wheel slip ratio l. The friction
coefcient can vary in a very wide range, depending on factors
like:
(a) road surface conditions (dry or wet),
(b) tire side-slip angle,
(c) tire brand (summer tire, winter tire),
(d) vehicle speed, and
(e) the slip ratio between the tire and the road.
In this paper, the tire friction model introduced by Burckhardt
(1993) and adopted in Hari et al. (2008) has been used. It
provides the tire-road coefcient of friction m
r
as a function of the
wheel slip l and the vehicle velocity V
x
.
m
r
l,V
x
bC
1
1e
C
2
l
C
3
lce
C
4
lVx
6
The parameters in Eq. (6) denote the following: C
1
is the
maximum value of friction curve; C
2
the friction curve shape; C
3
the friction curve difference between the maximum value and the
value at l1; and C
4
is the wetness characteristic value and in the
range 0.020.04 s/m. Table 1 shows the friction model parameters
for different road conditions.
Dependence of the road friction coefcient on surface condi-
tions and slip ratio is shown in Fig. 2, (von Altrock Nov., 1997).
The lateral force is essential to the steering of vehicle. It is obvious
when slipping is equal to one; this force is equal to zero, which
explains why the steering ability is lost during wheel lockup. The
effective coefcient of friction between the tire and the road has
an optimum value which differs according to the road type and
the worst performance occurs at l1 (locked wheel). Most
manufacturers use a set point for the slipping ratio l
d
equal to 0.2,
which is a good compromise for all road conditions, (Chih-Min Lin
and Hsu March, 2003). So that the control problem can be
described as a set-point control system that may implement PID
controller; Fig. 3. Next section describes several PID strategies for
this closed-loop system.
Table 1
Friction model parameters (Burckhardt (Burckhardt, 1993)).
Surface conditions C
1
C
2
C
3
Dry asphalt 1.2801 23.99 0.52
Wet asphalt 0.857 33.822 0.347
Dry concrete 1.1973 25.168 0.5373
Snow 0.1946 94.129 0.0646
Ice 0.05 306.39 0
Vehicle Body
N
F
N r
F
u J
w
w
R
ma
x
x
V
Fig. 1. The quarter vehicle dynamic model.
A.B. Sharkawy / Engineering Applications of Articial Intelligence 23 (2010) 10411052 1043
3. PID tuning with fuzzy logic
In the early 1940s, after extensive manual experimentation by
the way of trial and error, Ziegler and Nichols invented the
well-known ZN formula for off-line tuning (Ziegler and Nichols,
1942). In time domain for example, the method yielded PID
coefcients directly from the three important parameters of a
stable plant to be controlled, namely the plant gain, time-constant
and transport delay. These parameters can be easily obtained
graphically from a step response of the open loop plant. The
concept of fuzzy systems instead may be used to emulate human
expertise and on-line tune the control gains using ZN formula.
The coming sub-sections demonstrate some of the previous works
in this area. They are used in later sections for comparison
purposes.
3.1. Standard and nonlinear PID controllers
There are many types of PID controllers, e.g., PID plus gravity
compensator, PID plus friction compensator, PID plus disturbance
observer, etc. (Astrom, 1996). Here, we shall consider the basic
forms for PID controller which is placed in a unity feedback
control system. A typical PID control law in its standard form is
ut K
p
etT
d
det
dt
1
T
i
_
t
0
etdt
_ _
7
where u(t) is the control variable, e(t)l
d
(t)l(t) is the system
error (difference between the demand input l
d
and the system
output, l), K
p
is the proportional gain, T
d
the derivative time
constant and T
i
is the integral time constant. Eq. (7) can be
rewritten as
ut K
p
et K
i
_
etdt K
d
det
dt
8
K
d
K
p
T
d
the derivative gain and K
i
K
p
/T
i
the integral gain.
In the design and tuning of a PID controller, the P, I and D
actions need to be coordinated. This is not a trivial process, since
coefcients of these three actions interact mutually and cannot be
simply tuned individually in a de-coupled manner. An experiment
is carried out on the process using ZN tuning method, can be
stated as follows (Yu, 1999; Ziegler and Nichols, 1942). First, the
process is controlled using the proportional gain K
p
. The value of
K
p
is slowly increased until continuous oscillations are happened.
At the time of oscillation, the values of the gain K
u
and the
oscillation period t
u
are noted. The method assumes that K
p
is 60%
of the gain at the time of oscillation. The integral time constant T
i
is 50% of the oscillation period t
u
and T
d
the derivative time
constant is 12.5% of the oscillation period. The ZN method is
devised for off-line tuning of continuous systems and can also be
used on discrete cases for a fast sampling time.
Nonlinear PID controllers have been widely considered in
literature, (Yu, 1999; Mann et al., 1999; Visioli January, 2001;
Tzafestas and Papanikolopoulos, 1990). A large class of nonlinear
PID controllers can be generalized as follows:
ut K
p
f e, _ eK
d
det
dt
K
i
_
t
0
etdt 9
where f e, _ e is a nonlinear function that depends on the closed-
loop error and the process delay (Mann et al., 1999). The f e, _ e
may be represented by a fuzzy logic system, in which the inputs
are e(t) and _ et (Tzafestas and Papanikolopoulos, 1990). It comes
out that the proportional gain depends on the current error and
other parameters contained on f e, _ e. Effectiveness of this non-
linear PID has been discussed in Mann et al. (1999) for linear
systems.
3.2. Fuzzy self-tuning of a single parameter (SSP)
Because of its simplicity, this scheme has been widely
considered in literature (S.-Z. et al., 1993; Visioli January, 2001;
Bhattacharya et al., 2003). The method devised by He et al. S.-Z.
et al. (1993), consists of parameterizing the ZieglerNichols
formula by means of a single parameter a, then using an online
fuzzy inference system to self-tune the parameter. In this method,
the three PID parameters can be expressed as
K
p
1:2atK
u
T
i
0:75
1
1at
t
u
T
d
0:25T
i
10
where K
u
and t
u
are the ultimate gain and ultimate period,
respectively. The value of a(t) is determined recursively with the
following equation:
at 1
atght1at for at40:5
atghtat for atr0:5
_
11
where g is a positive constant that has to be chosen in the range
[0.2,0.6] and h(t) is the output of the fuzzy inference system. The
fuzzy system has seven membership functions for each of the two
input (e and _ e) and seven for the output, i.e. the rule-base consists
of 49 rules. The initial value of a(t) is set equal to 0.5, which
corresponds to the ZieglerNichols formula. With respect to the
method however, the tuning of the scaling coefcient of the fuzzy
modules and of the parameter g is left to the user and no rule of
thumb is given for this task.
3.3. Fuzzy set-point weighting (FSW)
This approach proposed by Visioli (1999) consists of fuzzifying
the set-point weight, leaving xed the other three parameters
(again, determined with the ZieglerNichols method). In this way,
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
Wet Road
Icy Road
Design Region
Lateral Force
Dry Road
r
Fig. 2. Coefcient of road friction versus wheel slip ratio.
d
u
Wheel
Brakes
(ABS)
PID
Fig. 3. The closed-loop control of ABS.
A.B. Sharkawy / Engineering Applications of Articial Intelligence 23 (2010) 10411052 1044
the control law can be written as
ut K
p
btl
d
tlt
_
K
d
det
dt
K
i
_
t
0
etdt
bt wf t 12
where w is a positive constant less than or equal to 1, and f(t) is
the output of the fuzzy inference system, which consists of ve
triangular membership functions for each of the two inputs e(t)
and _ et and nine triangular membership functions for the output.
The fuzzy rules are based on the Macvicar-Whelan matrix
(Macvicar-Whelan, 1976), as shown in Table 2. The linguistic
variables are negative small NS, negative big NB, negative very big
NVB, zero Z, positive small PS, positive big PB, and positive very
big, PVB.
The method however, is not a straightforward one and large
number of arithmetic operations is needed for on-line tuning.
3.4. Incremental fuzzy expert PID control (IFE)
Tzafestas and Papanikolopoulos (Tzafestas and Papanikolo-
poulos, 1990), introduced a procedure for scaling the values
of the three control parameters (initially determined by the
ZieglerNichols formula) during the transient period based on
the system error and its rate. In other words, the current values of
the proportional, integral and derivative gains are increased or
decreased by means of a fuzzy inference system according to the
following relations:
K
p
K
p
FSc
1
et,c
2
_ et k
1
K
i
K
i
FSc
1
et,c
2
_ et k
2
K
d
K
d
FSc
1
et,c
2
_ et k
3
13
where the basic tuning method is that of the ZieglerNichols;
FSc
1
et,c
2
_ et is the output of fuzzy inference system, based on
Macvicar-Whelan fuzzy rule matrix (Macvicar-Whelan, 1976)
(see Table 2), which reects the typical action of a human control.
For example, the integral action has to be increased at the
beginning of the transient response to decrease the rise time, and
then has to be decreased when the system error is negative to
reduce the overshoot. The range of the input membership
functions is normalized between 1 and 1. Finally, c
1
, c
2
are
scaling factors and k
i
, i 1,2,3 are constant parameters that
determine the range of variation of each term. The whole fuzzy
system involve 14 quantization levels for both e and _ e. Similar
approach has been followed in Bhattacharya et al. (2003), where
the authors have to rely on trial and error procedure in order to
identify parameters of the fuzzy systems.
However, tuning the three parameters k
i
and the two scaling
factors (c
1
, c
2
) that multiply the inputs e and _ e, is left to the
user, and it is not clear how these parameters inuence the
performance of the overall controller.
4. The proposed fuzzy self-tuning PID control scheme
The proposed control scheme does not rely on ZN method nor
on a previous rule-base, especially designed for some plants.
Instead, it uses rst order TS fuzzy systems as the tuning-tool for
each of the PID control modules. A modied genetic algorithm is
used to optimally select parameters of the fuzzy systems.
The main control objectives of an ABS are:
minimization of the stopping distance S
x
, and
maintaining the fastest possible response with no/low over-
shoot and close to zero steady state error.
The rst objective is a key issue related to ABS, while the
second is a common objective for any closed-loop control system.
In this work, optimality of the second objective is based on
conventional cost functions; i.e. integrated absolute error (IAE)
and the integrated time multiplied by the absolute error (ITAE).
For example, in an IAE, the hatched area in Fig. 4 represents the
cost function. A minimum area is achieved at the fastest
physically possible response, no/small overshoot and close to
zero steady state error. It means that decreasing overshoot will
not result in an increase in the rise time or vise versa, as it is the
case with ZN method. The expected tuning result is, simply, the
best physically possible response.
In short, the objective of an ABS design is to minimize a
performance index that contains the stopping distance and an IAE
or ITAE.
4.1. The control system architecture
Motivated by the work of Bhattacharya et al. (2003); Tzafestas
and Papanikolopoulos (1990), three decoupled fuzzy systems
constitute the proposed self-tuning system; each for one para-
meter of the PID controller, i.e. K
p
, K
i
and K
i
; Eq. (8). The error e
and change in error _ e are used as behavior-recognizers of the
closed-loop performance. They are available signals in the closed-
loop system of the ABS and do not require extra hardware. The
self-tuner can be expressed as
K
p
FS
1
et, _ et
K
i
FS
2
et, _ et
K
d
FS
3
et, _ et 14
where a fuzzy P controller, a fuzzy I controller and a fuzzy D
controller are connected in parallel to give the resultant controller
signal. With this structure, independent control actions can be
generated, which should necessarily eliminate the problems
associated with most practical two-terms or three-terms FLCs.
The basic approach is summarized in Fig. 5, where each fuzzy
system is trying to recognize when the corresponding parameter
E
r
r
o
r
,
e
(
t
)
0
) ( dt t e A
0
A
Time, sec
Fig. 4. A typical closed-loop error time-history.
Table 2
Rule table of the FSW (Macvicar-Whelan, 1976).
f(t) Change in error, _ e
NB NS Z PS PL
Error, e NB NVB NB NM NS Z
NS NB NM NS Z PS
Z NM NS Z PS PM
PS NS Z PS PM PB
PB Z PS PM PB PVB
A.B. Sharkawy / Engineering Applications of Articial Intelligence 23 (2010) 10411052 1045
is not properly tuned and then seeks to adjust it to obtain
improved performance. In such a way, each fuzzy system can be
looked at as gain scheduler (module). TS type fuzzy systems are
used to synthesize each of the self-tuning systems.
The TS fuzzy system (also known as functional fuzzy system
(Passino and Yurkovich, 1998)) was proposed in an effort to
develop a systematic approach to generating fuzzy rules from a
given inputoutput data set; (Takagi and Sugeno, 1985; Jang et al.,
1997). A typical rule has the following form
IF x
1
IS B
1
AND x
2
IS B
2
THEN z f x
1
,x
2
where B
1
and B
2
are fuzzy sets in the antecedent, while zf(x
1
,x
2
)
is a crisp function in the consequent. With this form, the fuzzy
system can be characterized as two input one output fuzzy
systems.
Usually f(x
1
,x
2
) is a polynomial in the input variables x
1
and x
2
,
but it can be any function as long as it can appropriately describe
the output of the model within the fuzzy region specied by
the antecedent of the rule. Although there are no restrictions on
the form of the input membership functions, Gaussians are used
in the premise through out this work. A Gaussian membership
function is specied by two parameters (c,s)
m
B
l
j
x
j
gaussianx
j
;c,s exp
1
2
x
j
c
s
_ _
2
_ _
15
where m is the membership grade, c represents the membership
functions center, s determines its spread; B is the membership
which represents a linguistic variable, i 1,2, y, n is the rule
number, j 1,2 is subscript of the input variables.
In the proposed self-tuner, the inputs e and _ e are normalized
using three Gaussian membership functions; negative N, zero Z,
and positive P. So that nine rules constitute the rule-base for each
module. For simplicity, the consequent part has been chosen to be
a rst order function of e and _ e. So that the rules have the
following form:
Rule
j
: IF e IS B
1
AND e IS B
2
THEN K
i
a
i,j,1
ea
i,j,2
_ e a
i,j,3
16
where both B
1
and B
2
are positive (P), zero (Z) or negative (N),
K
i
f e, _ e is the gain to be tuned, i.e. K
p
, K
i
or K
d
. a
i,1,j
, a
i,2,j
are the
constants, and j 1,2, y, 9 is the rule number.
Fig. 6 shows the fuzzy reasoning procedure for a rst order TS
model. The fuzzy part is only in its antecedent. Each rule has a
crisp output and the overall output is obtained via weighted
average. This fuzzy procedure avoids the time-consuming process
of defuzzication required in a Mamdani fuzzy model. The
weighted sum has been also used in Jang et al. (1997) to further
reduce computations.
With the above structure, the data-base of each module
consists of 27 free parameters (coefcients of the rst order
polynomial) and 12 free parameters for the inputs membership
functions. So that, the total number of free parameters is 117. If an
identical input membership is chosen for the three modules, the
total number of parameters is reduced to 93; i.e. parameters of
the input membership functions c
N1
, s
N1
, c
Z1
, s
Z1
c
P1
, s
P1
for e, c
N2
,
s
N2
, c
Z2
, s
Z2
, c
P2
, s
P2
for _ e, and a
ijk
where i 1,2,3 are the
coefcients of the rst order polynomial, j 1,2, , 9 the rule
number for each tuner and k1,2,3 the number of modules.
Determination of optimal values for these parameters is the
subject of the next sub-section.
4.2. Genetic algorithm-based parameter learning
GAs are optimization stochastic technique mimicking the
natural selection, which consists of three operations, namely,
reproduction, crossover, and mutation (Jang et al., 1997). The
most general considerations about GA can be stated as follows:
(i) The searching procedure of the GA starts from multiple initial
states simultaneously and proceeds in all of the parameter
subspaces simultaneously.
(ii) GA requires almost no prior knowledge of the concerned
system, which enables it to deal with completely unknown
systems that other optimization methods may fail.
(iii) GA cannot evaluate the performance of a system properly at
one step. For this reason, generally, it cannot be used as an
on-line optimization strategy and is more suitable for fuzzy
modeling.
In practice, training data can be obtained by experimentation
or by the establishment of an ideal model. In this work, the ABS
model in Section 2 is used to emulate the behavior of the ABS, in
order to collect training data. Fig. 7 shows the training process for
each fuzzy module involved in the self-tuning system.
d
e
.
dt
d
dt
dt
d
Update Kp, K
i
and K
d
d
K
p
K
i
K
e
e Fuzzy System
1
Fuzzy System
2
Fuzzy System
3
ABS
+
+
+
+
Fig. 5. The self-tuning fuzzy PID controller.
Product inference
P
N
Weighted
Average
......
......
Fig. 6. TakagiSugeno fuzzy system for a module.
A.B. Sharkawy / Engineering Applications of Articial Intelligence 23 (2010) 10411052 1046
The following two closed-loop performance indices have been
examined. The rst uses the stopping distance and the integrated
absolute error (IAE). The second uses the stopping distance and
the integrated time multiplied by the absolute error (ITAE). These
two performance indices are dened as follows:
J
IAE
Z
1
S
x
Z
2
IAE Z
1
S
x
Z
2
_
T
0
et
dt
Z
1
S
x
Z
2
M
k 0
9l
d
lk9Dt 17
J
ITAE
Z
1
S
x
Z
2
ITAE Z
1
S
x
Z
2
_
T
0
t9et9dt
Z
1
S
x
Z
2
M
k 0
k9l
d
lk9Dt
2
18
where el
d
l is the closed-loop error, l
d
is the desired slip ratio,
l(k) is the current slip ratio and Dt is the time step. M is the
number of training samples. In addition, the coefcients Z
1
and Z
2
are the weighting factors to emphasize the relative importance of
the associated terms.
Because GA endeavors to maximize the tness function, the
tness function of each gene is calculated as follows:
F
1
1J
19
where J is the performance index and 1 is introduced at the
denominator to prevent the tness function from becoming
innitely large. The overall training procedure is shown in Fig. 8.
To simplify the presentation, let us denote F in Eq. (19) as F
IAE
when J J
IAE
and F
ITAE
when J J
ITAE
. With this notation, the
controller is called PID-IAE when the performance index is
calculated by Eq. (17) and PID-ITAE when the performance index
is given by Eq. (18).
Coding of the parameters to be adjusted can be stated as
follows:
c
N1
,c
N2
, , s
N1
,s
N2
, , a
111
,a
121
, , a
339
20
where a certain number of binary bits stands for each parameter.
The combined string composes a gene (possible solution) in a
population. Evaluation of each possible solution is performed via
IAE/ITAE and genes of best solutions are allowed to reproduce.
Although, genetic algorithms were developed a few decades
ago, concrete theoretical analysis of the algorithm have not been
provided until recent years (Golberg, 1989; Rudolph, 1994).
Reference (Rudolph, 1994) concludes that the canonical GA
cannot always nd the optimal solution within the denite time.
Furthermore, the paper pointed out that if the chromosome with
the best performance in each generation is reserved for the next
generation, the algorithm will globally converge. Inspired by these
conclusions, the following two measures have been considered in
this work:
(i) In reproduction, we stochastically introduce a randomly
generated gene at a probability of R
h
to replace one of the
two parents selected for reproduction.
(ii) Select the best performed genes in the current population at a
rate of R
e
and place them directly in the next generation.
If the reproduction is carried out in the traditional way, the
best gene will globally be lost, and thus convergence cannot
be guaranteed. However, if we only adopt the second measure,
the procedure of the evolution is no longer a Markov process, and
thus it does not satisfy the assumption of the convergence theory.
Despite our successful application of these measures, mathema-
tical analysis of them still lacks. The parameters R
h
and R
e
are
adjusted such that at the beginning of the learning, R
h
is relatively
large and R
e
is relatively small, and later, vice versa.
Fuzzy Expert
ABS PID
Genetic Fuzzy
Optimizer
Modify membership
functions
Modify coefficients of
the consequent parts
Update the gains
d