Professional Documents
Culture Documents
, Yi-Jen Mon
Department of Electrical Engineering, Yuan-Ze University, Chung-Li, Tao-Yuan, Taipei, 32026 Taiwan,
Republic of China
Received 10 December 2001; received in revised form 2 September 2002; accepted 15 October 2002
Abstract
This paper presents a hybrid adaptive fuzzy control (HAFC) design methodology, where the nonlinear
system is controlled by a state feedback controller and an adaptive fuzzy controller. In the adaptive fuzzy
controller, an adaptive law is developed to tune a robust gain of the sliding-mode controller (SMC) so as to
cope with the uncertainties and model errors. The proposed design method is applied to investigate the position
and tracking control of a two-link robot. A comparison between the proportional-derivative control, SMC,
adaptive control and the proposed HAFC for a two-link robot is made. The simulation results demonstrate
that, by using the HAFC, the system performances is improved and the system also exhibits stability and
robustness.
c 2002 Elsevier Science B.V. All rights reserved.
Keywords: Fuzzy control; Sliding-mode control; Adaptive law; Robotic systems
1. Introduction
Fuzzy logic control (FLC) is a human knowledge-based design methodology which is driven ac-
cordingly by fuzzy membership functions and fuzzy rules. The applications of FLC mainly focus
on certain systems that are structurally dicult to model due to their inherent natural nonlinearities
and other modeling complexities [10]. However, the conventional FLC still has two main draw-
backs: (1) The fuzzy rules are obtained from human trial-and-error work, and (2) it still lacks the
systematic or mathematical methodologies to ensure the system stability. In order to avoid the time-
consuming trial-and-error process for the design of the fuzzy rules, several self-tuning processes have
been proposed. The tuning processes can be carried out by the following methods: least-square-error
and n6j
where j is a positive constant and denotes the induced norm.
The necessity of coping with the uncertainty in the model calls for a hybrid control input u of
the form
u = u
fb
+ u
rb
, (3)
where the feedback controller u
fb
is determined on the basis of the nominal model considering the
disturbance-free case, while the robust controller u
rb
compensates for the system uncertainty. The
concept diagram of the control system is shown in Fig. 1, and the design procedure is described in
the following.
The linear feedback control law is designed as
u
fb
= Gx, (4)
where GR
n
is the state feedback gain matrix which can be found from standard methods such
as pole placement. And the closed-loop dynamic for the disturbance-free system is given by
x = A
0
x + B
0
u = (A
0
+ B
0
G)x, (5)
where the closed-loop dynamics matrix
A
c
=A
0
+ B
0
G (6)
has eigenvalues specied in the left-half s-plane for the systems stability and desired responses.
Dene denote the sliding surface as
s = Cx = [s
1
s
2
s
]
T
= 0 t
t
t
s
(7)
where t
s
is the time at which the sliding motion starts and CR
n
is set as a constant matrix to
correspond the desired sliding-mode responses.
4 C.-M. Lin, Y.-J. Mon / Fuzzy Sets and Systems ( )
ARTICLE IN PRESS
Nonlinear Plant
C
x
rb
u
fb
u
+
u
) sgn(
s
rb
k
s
G
Adaptive law (19)
Fuzzy inference (13)
Fig. 1. HAFC design concept diagram.
In (3), the robust controller is designed as
u
rb
= k
rb
sgn(s), (8)
where k
rb
0 is utilized to cope with the system uncertainties and sgn() is a sign function. When
the sign function is used as a robustness term, the control law in (8) becomes an SMC. However, the
control law in (8) is usually discontinuous across the sliding surface. The application of the robust
controller needs to know the bound of uncertainty. However, this bound is dicult to measure in
practical applications; thus, in general, the bound of the uncertainty is chosen large enough to ensure
robust stability. However, a large k
rb
will result in substantial chattering of the control eort. On the
other hand, if the bound is chosen too small, the robust stability cannot be guaranteed. To relax the
requirement for the bound of uncertainty, an adaptive design method for the robust gain is proposed.
Assume there is an optimal gain k
rb
for the existence of sliding-mode condition, i.e. nk
rb
.
However, the optimal gain k
rb
cannot be obtained exactly because of the unknown of uncertainty.
Therefore an adaptive fuzzy control scheme is developed to estimate this optimal gain. The basic
idea for constructing the fuzzy rules is that when the system states are far away from the sliding
surfaces then a large robust gain is needed and vice versa. Meanwhile, in the proposed fuzzy rules
the antecedent parts are predetermined with appropriate membership functions and the consequent
parts are dened to be the singletons and should be self-generated by the adaptive laws developed
later. For this, the robust controller in (8) is modied to be an adaptive fuzzy controller
u
rb
=
k
rb
sgn(s), (9)
where
k
rb
0 can be estimated by the following fuzzy inference mechanism:
Rule
i
: If s is F
i
1hen
k
rb
is
0
i
, (10)
ARTICLE IN PRESS
C.-M. Lin, Y.-J. Mon / Fuzzy Sets and Systems ( ) 5
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
s w
s
w
c
w
d
w
c
1
+
w
d
w
c
1
)
(
s
F
i
Fig. 2. A generalized Gaussian-type membership function.
where
0
i
i =1, 2, . . . , r, are the adjustable singleton control actions and F
i
is the label of the fuzzy
set characterized by a generalized Gaussian-type fuzzy membership function [20]:
j
F
i
(s) = e
ln(0.5)(w
s
sw
c
)
2
w
2
d
, (11)
where s is the fuzzy input, and w
s
, w
c
and w
d
are the parameters of the generalized Gaussian-type
function. The equivalent membership function is demonstrated as in Fig. 2, in which the parameter
w
c
represents the center value and the parameter w
d
denotes the reciprocal value of deviation from
the center to which the value on the standardized support set has 0.5. The input s is scaled by
the parameter w
s
.
The center-of-gravity method is used for the defuzzication
k
rb
=
r
i=1
i
0
i
r
i=1
i
, (12)
where
i
is the ring weight of the ith rule. From (12),
k
rb
can be rewritten as
k
rb
=
X
T
w, (13)
where
X=[
0
1
0
2
0
r
]
T
is a parameter vector and w=[w
1
w
2
w
r
]
T
is a regressive vector with
w
i
dened as
w
i
=
i
r
i=1
i
. (14)
6 C.-M. Lin, Y.-J. Mon / Fuzzy Sets and Systems ( )
ARTICLE IN PRESS
Dene the parameter error vector as
X =
X X
, (15)
where the constant vector X
rb
= X
T
w, (16)
Choose the Lyapunov function
J =
1
2
s
T
s +
1
:
X
T
, (17)
where : is a positive constant. Dierentiating J with respect to time yields
J = s
T
s +
1
:
X
T
X
= s
T
C x +
1
:
X
T
X
= s
T
C[A
0
x + B
0
u + d] +
1
:
X
T
X
= s
T
C[A
0
x + B
0
(Gx
k
rb
sgn(s)) + d] +
1
:
X
T
X
= s
T
C[A
0
+ B
0
G]x s
T
CB
0
k
rb
sgn(s) + s
T
Cd +
1
:
X
T
X
= s
T
CA
c
x
k
rb
s
T
CB
0
sgn(s) + s
T
CB
0
n +
1
:
X
T
X
6A
c
s
T
s
k
rb
CB
0
s
T
sgn(s) +sCB
0
n +
1
:
X
T
X
= A
c
s
T
s sCB
0
(
k
rb
n) +
1
:
X
T
X
= A
c
s
T
s sCB
0
(
k
rb
k
rb
n + k
rb
) +
1
:
X
T
X
= A
c
s
T
s sCB
0
(k
rb
n) sCB
0
(
k
rb
k
rb
) +
1
:
X
T
X
= A
c
s
T
s sCB
0
(k
rb
n) sCB
0
(
X
T
w) +
1
:
X
T
X
= A
c
s
T
s sCB
0
(k
rb
n)
X
T
1
:
X sCB
0
w
. (18)
The adaptive law is chosen as
X = :sCB
0
w, (19)
ARTICLE IN PRESS
C.-M. Lin, Y.-J. Mon / Fuzzy Sets and Systems ( ) 7
Fig. 3. An articulated two-link manipulator.
then inequality (18) becomes
J = A
c
s
T
s sCB
0
(k
rb
n)
6A
c
s
T
s
62A
c
J. (20)
Since all the eigenvalues of A
c
are chosen to be on the left-half s-plane, J will converge exponen-
tially to zero, i.e. J(t) 0 at t . This guarantees the closed-loop stability.
In summary, the learning algorithm of the consequents of the fuzzy rules in (10) is given in (13)
where the regressive vector w is determined from (14) and the adaptive law of the parameter vector
X is derived in (19).
3. Controller design for robotic systems
The robot considered is a two-link, articulated manipulator as shown in Fig. 3, whose position
can be described by a joint angle vector q=[q
1
q
2
]
T
, and whose actuator input is the torque vector
t=[t
1
t
2
]
T
applied at the manipulator joints. The dynamics of such a manipulator are strongly
nonlinear, and can be written in a general form
H(q) q + C(q, q) q + g(q) = t, (21)
where H(q) is the 2 2 manipulator inertia matrix, C(q, q) q is the vector of centripetal and Coriolis
torques, and g(q) is the gravitational torque vector. The control problem for such a system is to
8 C.-M. Lin, Y.-J. Mon / Fuzzy Sets and Systems ( )
ARTICLE IN PRESS
design the control law such that the required actuator inputs to the robot are sucient to perform
the desired task, e.g. follow a desired trajectory. Assume that the manipulator of Fig. 3 is in the
horizontal plane (g(q) 0), then the dynamics can be written explicitly as [16]
H
11
H
12
H
21
H
22
q
1
q
2
h q
2
h( q
1
+ q
2
)
h q
1
0
q
1
q
2
t
1
t
2
, (22)
where
H
11
= a
1
+ 2a
3
cos q
2
+ 2a
4
sin q
2
,
H
12
= H
21
= a
2
+ a
3
cos q
2
+ a
4
sin q
2
,
H
22
= a
2
,
h = a
3
sin q
2
a
4
cos q
2
(23)
with
a
1
= I
1
+ m
1
!
2
c1
+ I
e
+ m
e
!
2
ce
+ m
e
!
2
1
,
a
2
= I
e
+ m
e
!
2
ce
,
a
3
= m
e
!
1
!
ce
cos o
e
,
a
4
= m
e
!
1
!
ce
sin o
e
(24)
and the parameters of the robot are shown in Fig. 3 and their values are given as m
1
=1, !
1
=1,
m
e
=2, o
e
=30
, I
1
=0.12, !
c1
=0.5, I
e
=0.25, !
ce
=0.6.
Eq. (22) can be rewritten as
q
1
q
2
H
11
H
12
H
21
H
22
h q
2
h( q
1
+ q
2
)
h q
1
0
q
1
q
2
H
11
H
12
H
21
H
22
t
1
t
2
, (25)
where
H
11
H
12
H
21
H
22
1
is assumed to exist. Eq. (25) can be expressed as a state equation
q = A q + Bu, (26)
where
A =
H
11
H
12
H
21
H
22
h q
2
h( q
1
+ q
2
)
h q
1
0
A
11
A
12
A
21
A
22
, (27)
ARTICLE IN PRESS
C.-M. Lin, Y.-J. Mon / Fuzzy Sets and Systems ( ) 9
B =
H
11
H
12
H
21
H
22
1
=
B
11
B
12
B
21
B
22
, (28)
u =
t
1
t
2
, (29)
where A
ij
and B
ij
imply the element values in the index of the ith row and the )th column for the
matrices A and B, respectively. The joint angle error is dened as
q = q q
d
(30)
and q
d
is the desired joint angles vector. Because the state of (26) is q, it is necessary to make
suitable transformation for (26) such that the controlled states q can be induced. Dierentiating (30)
with respect to time and from (26), it is revealed that
q = A(
q + q
d
) + Bu q
d
, (31)
i.e.
q = A
q + Bu + (A q
d
q
d
). (32)
Dene the states as x
1
= q
1
, x
2
= q
2
, x
3
=
q
1
, x
4
=
q
2
and the disturbance d =A q
d
q
d
=[d
1
d
2
]
T
. Then,
Eq. (32) can be transformed to a fourth-order dynamic system:
x
1
x
2
x
3
x
4
0 0 1 0
0 0 0 1
0 0 A
11
A
12
0 0 A
21
A
22
x
1
x
2
x
3
x
4
0 0 0 0
0 0 0 0
0 0 B
11
B
12
0 0 B
21
B
22
0
0
t
1
t
2
0
0
d
1
d
2
, (33)
i.e.
x = A
x + B
u + d (34a)
=
0 0 1 0
0 0 0 1
0 0 A
11
A
12
0 0 A
21
A
22
x
1
x
2
x
3
x
4
0 0 0 0
0 0 0 0
0 0 B
11
B
12
0 0 B
21
B
22
0
0
t
1
t
2
0
0
d
1
d
2
, (34b)
where
A
0 0 1 0
0 0 0 1
0 0 8.41 10
6
5.19 10
6
0 0 2.43 10
5
8.41 10
6
, (35)
10 C.-M. Lin, Y.-J. Mon / Fuzzy Sets and Systems ( )
ARTICLE IN PRESS
B
0 0 0 0
0 0 0 0
0 0 0.50 0.81
0 0 0.81 2.34
, (36)
x = [x
1
x
2
x
3
x
4
]
T
, (37)
u = [0 0 t
1
t
2
] (38)
and the nominal states are given as q
1
=q
2
= q
1
= q
2
=0.
Based on the system dynamic of (33), the proposed HAFC design procedures are described as
follows:
Step 1: In order to design the u
fb
of (4), choose the eigenvalues of A
+ B
0 0 0 0
0 0 0 0
27241 6283.5 1044.2 345.42
9420.7 3881.2 361.1 213.41
. (39)
Step 2: In order to design the u
rb
of (9), rst choose a constant matrix C for (7) such that the
sliding surface s=Cx=0 can be dened. In this paper, C is chosen as
C =
10 0 0 0
0 10 0 0
0 0 10 0
0 0 0 10
. (40)
In this paper, the Gaussian-type function is utilized as the membership functions of fuzzy system. Five
Gaussian-type membership functions are constructed from (11). The parameters are chosen based on
human knowledge and through some trials. The centers are set at w
c
= 0.5, 0.25, 0, 0.25 and
0.5, respectively. And for each membership functions, w
s
and w
d
are set as w
s
=1 and w
d
=0.15.
The consequent parts characterized by adjustable singletons
0
i
are all initialized at zeros and : is
set as 0.1.
ARTICLE IN PRESS
C.-M. Lin, Y.-J. Mon / Fuzzy Sets and Systems ( ) 11
A
n
g
l
e
e
r
r
o
r
q
1
(
d
e
g
)
HAFC
PD control
0.0 0.5 1.0 1.5 2.0 2.5 3.0
-100
-80
-60
-40
-20
0
20
A
n
g
l
e
e
r
r
o
r
q
2
(
d
e
g
)
Time (sec)
Time (sec)
0.0 0.5 1.0 1.5 2.0 2.5 3.0
-30
-20
-40
-50
-60
-70
-10
0
10
20
HAFC
PD control
(a) (c)
(d) (b)
0.0 0.5 1.0 1.5 2.0 2.5 3.0
-2000
0
2000
4000
6000
8000
10000
C
o
n
t
r
o
l
t
o
r
q
u
e
1
(
N
-
m
)
Time (sec)
HAFC
PD control
0.0 0.5 1.0 1.5 2.0 2.5 3.0
-1000
0
1000
2000
3000
4000
C
o
n
t
r
o
l
t
o
r
q
u
e
2
(
N
-
m
)
Time (sec)
HAFC
PD control
Fig. 4. (a) Angle error q
1
for position control; (b) angle error q
2
for position control; (c) control torque t
1
for position
control; and (d) control torque t
2
for position control.
Step 3: Finally, combining the results of Step 1 and Step 2, then the control law can be designed
as
u = Gx
k
rb
sgn(s). (41)
4. Simulation results
The performance of position and tracking control from the simulations are presented and com-
pared with the simulation results of Slotine and Li [16]. In [16], the PD control has been utilized
for position control; and the adaptive control and sliding control have been utilized for tracking
control. In the simulations, mass uncertainties for m
1
and m
e
, and model error of A in (26) are
considered.
(a) position control: The robot, initially at rest at (q
1
=0
, q
2
=0
, q
d2
=90
1
(
N
-
m
)
Time (sec)
HAFC
PD control
-4000
-3000
-2000
-1000
0
1000
2000
3000
4000
5000
6000
7000
8000
C
o
n
t
r
o
l
t
o
r
q
u
e
2
(
N
-
m
)
Time (sec)
HAFC
PD control
0.0 0.5 1.0 1.5 2.0 2.5 3.0
(c)
(d)
Fig. 5. (a) Angle error q
1
for position control with 50% mass uncertainties and 25% model error; (b) joint angle error
q
2
for position control with 50% mass uncertainties and 25% model error; (c) control torque t
1
for position control with
50% mass uncertainties and 25% model error; and (d) control torque t
2
for position control with 50% mass uncertainties
and 25% model error.
control [16]. For the case without mass uncertainties, the simulation results are shown in Fig. 4.
It can be seen that the HAFC can achieve better control performance. For the case of 50% mass
uncertainties for m
1
and m
e
, and 25% model error of A in (26), the simulation results are shown in
Fig. 5. The results show that the proposed HAFC controller still can cope with the mass uncertainties
and model error to achieve better performance.
(b) robust tracking control: The robot, initially at rest at (q
1
=0
, q
2
=0
), is commanded to follow
a desired trajectory q
d1
(t)=30
(1 cos(2t)) and q
d2
(t)=45
1
(
N
-
m
)
0.0 0.5 1.0 1.5 2.0 2.5 3.0
-60
-40
-20
0
20
40
60
80
100
C
o
n
t
r
o
l
t
o
r
q
u
e
2
(
N
-
m
)
Time (sec)
HAFC
Sliding control
Adaptive control
HAFC
Sliding control
Adaptive control
0.0 0.5 1.0 1.5 2.0 2.5 3.0
Time (sec)
-150
-100
-50
0
50
100
150
200
250
(c)
(d)
Fig. 6. (a) Angle tracking error q
1
for tracking control; (b) angle tracking error q
2
for tracking control; (c) control torque
t
1
for tracking control; and (d) control torque t
2
for tracking control.
performance, as shown in Fig. 7. This simulation also shows that the proposed HAFC can achieve
the best control performance with favorable tracking performance.
5. Conclusions
A hybrid adaptive fuzzy controller (HAFC) has been demonstrated to be an ecient control
technology for a robotic system. The adaptive fuzzy controller includes a state feedback control
and an adaptive fuzzy robust controller to treat the multi-input multi-output robotic system. The
advantage of this approach is that the adaptive fuzzy control law can tune the robust gain of the
SMC so as to cope with the uncertainties and modeling error of the nonlinear robotic system. Thus
the controller can improve the system performance for the nonlinear robotic system subject to system
uncertainties. In the simulation example, the controller developed here is applied to a two-link robot.
It is shown that, by using the proposed HAFC, the system performance is considerably improved
and the system also exhibits stability and robustness.
14 C.-M. Lin, Y.-J. Mon / Fuzzy Sets and Systems ( )
ARTICLE IN PRESS
0.0 0.5 1.0 1.5 2.0 2.5 3.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0
-4
-2
0
2
4
6
8
10
A
n
g
l
e
e
r
r
o
r
q
1
(
d
e
g
)
Time (sec)
5.0
4.5
4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
-0.5
A
n
g
l
e
e
r
r
o
r
q
2
(
d
e
g
)
Time (sec)
HAFC
Sliding control
Adaptive control
HAFC
Sliding control
Adaptive control
(a) (c)
(d) (b)
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0
-700
-600
-500
-400
-300
-200
-100
0
100
200
C
o
n
t
r
o
l
t
o
r
q
u
e
1
(
N
-
m
)
Time (sec)
HAFC
Sliding control
Adaptive control
0.0 0.5 1.0 1.5 2.0 2.5 3.0
-240
-220
-200
-180
-160
-140
-120
-100
-80
-60
-40
-20
0
20
40
60
C
o
n
t
r
o
l
t
o
r
q
u
e
2
(
N
-
m
)
Time (sec)
HAFC
Sliding control
Adaptive control
Fig. 7. (a) Angle tracking error q
1
for tracking control with 50% mass uncertainties and 25% model error; (b) angle
tracking error q
2
for tracking control with 50% mass uncertainty and 25% model error; (c) control torque t
1
for tracking
control with 50% mass uncertainties and 25% model error; and (d) control torque t
2
for tracking control with 50% mass
uncertainties and 25% model error.
Acknowledgements
This work was supported by the National Science Council of the Republic of China under grant
NSC-89-2218-E-155-011.
References
[1] Y.C. Chang, B.S. Chen, A nonlinear adaptive H