1326 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 38, NO.
5, OCTOBER 2008
Adaptive Fuzzy Neural Network Control Design
via a T–S Fuzzy Model for a Robot Manipulator
Including Actuator Dynamics
Rong-Jong Wai, Senior Member, IEEE, and Zhi-Wei Yang
Abstract—This paper focuses on the development of adaptive
fuzzy neural network control (AFNNC), including indirect and
direct frameworks for an n-link robot manipulator, to achieve
high-precision position tracking. In general, it is difficult to adopt
a model-based design to achieve this control objective due to
the uncertainties in practical applications, such as friction forces,
external disturbances, and parameter variations. In order to cope
with this problem, an indirect AFNNC (IAFNNC) scheme and
a direct AFNNC (DAFNNC) strategy are investigated without
the requirement of prior system information. In these model-free
control topologies, a continuous-time Takagi–Sugeno (T–S) dy-
namic fuzzy model with online learning ability is constructed to
represent the system dynamics of an n-link robot manipulator. In
the IAFNNC, an FNN estimator is designed to tune the nonlinear
dynamic function vector in fuzzy local models, and then, the
estimative vector is used to indirectly develop a stable IAFNNC
law. In the DAFNNC, an FNN controller is directly designed to
imitate a predetermined model-based stabilizing control law, and
then, the stable control performance can be achieved by only
using joint position information. All the IAFNNC and DAFNNC
laws and the corresponding adaptive tuning algorithms for FNN
weights are established in the sense of Lyapunov stability analyses
to ensure the stable control performance. Numerical simulations
and experimental results of a two-link robot manipulator actu-
ated by dc servomotors are given to verify the effectiveness and
robustness of the proposed methodologies. In addition, the superi-
ority of the proposed control schemes is indicated in comparison
with proportional–differential control, fuzzy-model-based control,
T–S-type FNN control, and robust neural fuzzy network control
systems.
Index Terms—Adaptive control, fuzzy neural network (FNN),
Lyapunov stability analyses, robot manipulator, Takagi–Sugeno
(T–S) fuzzy model.
I. I NTRODUCTION
O VER the last decade, there has been tremendous progress
in the development of controllers for robot manipulators,
which have inherent physical constraints such as saturation
Manuscript received July 25, 2007; revised December 13, 2007 and
March 5, 2008. This work was supported in part by the National Science
Council of Taiwan, R.O.C., under Grant NSC 95-2221-E-155-085. This paper
was recommended by Associate Editor A. Tayebi.
R.-J. Wai is with the Department of Electrical Engineering and Fuel
Cell Center, Yuan Ze University, Chung-Li 32003, Taiwan, R.O.C. (e-mail:
rjwai@saturn.yzu.edu.tw).
Z.-W. Yang was with the Department of Electrical Engineering, Yuan Ze
University, Chung-Li 32003, Taiwan, R.O.C. He is now with the Air Force,
Ministry of National Defense, Taipei 10048, Taiwan, R.O.C.
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSMCB.2008.925749
1083-4419/$25.00 © 2008 IEEE
nonlinearities of actuators and friction phenomena at the robotic
joints. Saturation may lead to electromechanical actuator dam-
age, and friction will cause steady-state tracking error and os-
cillations. These constraints deteriorate the system performance
and stability, and it is difficult to establish a precise mathe-
matical model for the design of a model-based control system.
Santibanez et al. [1] proposed a novel global asymptotic stable
set-point fuzzy controller with bounded torques for robot ma-
nipulators. Although the friction effect and the phenomenon of
torque saturation were considered in this robot dynamic model,
some constrained conditions and prior system information were
required in the control process. Kim [2] developed the output
feedback tracking control of robot manipulators with model
uncertainty by using adaptive fuzzy logic. However, partial
system parameters were used in the designed control law, and
the total errors, including the observation error, tracking error,
and fuzzy estimation error, were only ensured to be uniformly
ultimately bounded. Gao and Selmic [3] investigated a neural
network control of a class of nonlinear systems with actuator
saturation. Unfortunately, in order to focus on the effect of
actuator saturation nonlinearity, gravity and joint friction were
neglected in its system model. Jafarov et al. [4] introduced a
new variable structure proportional–integral–differential (PID)-
controller design for robot manipulators. Even though the
global asymptotic stability of the controlled robot system was
analyzed, the bounds of system parameter matrices need to be
known in the control design.
For the most part of the robot manipulator control in the
published literature [1]–[4], actuator dynamics are typically
excluded from the robot dynamic behavior to simplify the
control design. However, actuator dynamics perform an im-
portant part of the complete robotic dynamics, particularly
in the factors of high-velocity moment, highly varying loads,
friction, and actuator saturation [5]. Thus, there exist some
interactions between the robot and actuator dynamics that can-
not be neglected in the dynamic modeling. Several modeling
methods based on fuzzy reasoning have been investigated in
recent years [6]–[13]. The Takagi–Sugeno (T–S) fuzzy model,
which is often used in the literature, can approximate a wide
class of nonlinear dynamic systems [13]. Feng [6] presented a
survey on recent developments of controller analysis and design
based on the T–S fuzzy models or fuzzy dynamic models. Lam
and Leung [7] developed a fuzzy controller, including stability
and performance fuzzy rules, through two designed steps for
nonlinear systems, i.e., the system stability was determined
WAI AND YANG: ADAPTIVE FUZZY NN CONTROL DESIGN VIA T–S FUZZY MODEL
by stability fuzzy rules, and the control performance was ad-
justed by performance fuzzy rules. Kumar et al. [8] proposed
two novel control schemes with variable state-feedback gains
to stabilize a T–S fuzzy system. One can note that the pa-
rameter matrices of subsystems in local fuzzy models [7], [8]
are required to calculate by linearization method so that the
requirement of prior system information is indisputable, and the
T–S fuzzy model dynamics are not governed by all the linear
subsystems at each operating point. In virtue of this reason,
this static fuzzy model is impractical under the occurrence of
unpredictable system uncertainties. Tseng et al. [9] investigated
fuzzy tracking control design for nonlinear dynamic systems
via the T–S fuzzy model. Only one simple fuzzy controller was
used in this approach without a feedback linearization tech-
nique and complicated adaptive scheme. Unfortunately, a huge
amount of system parameters and control gains should be cal-
culated in advance for controller action. Lin et al. [10] studied
H-infinite tracking control for nonlinear systems via T–S fuzzy
model approach. It depends on feasible sets of solutions to a
group of strict linear matrix inequalities (LMIs). Lian et al.
[11] introduced a new control scheme to improve the system
performance for a T–S fuzzy system using control grade func-
tions tuned by neural networks. In these control schemes [10],
[11], one needs to solve the LMIs to satisfy sufficient conditions
of the closed-loop stability, and prior system information is
usually required. Cheng and Chien [12] developed an adaptive
sliding-mode control technique based on T–S fuzzy system
model for a class of perturbed nonlinear multiinput–multioutput
(MIMO) systems in order to solve tracking problems. Besides
the requirement of prior system information to obtain a local
fuzzy model, all system state variables to be measurable may
limit the applicability of the control scheme in [12]. The moti-
vation of this paper is to design a continuous-time T–S dynamic
fuzzy model with online learning ability for dealing with the
aforementioned problems.
Recently, the concept of incorporating fuzzy logic into
a neural network has grown into a popular research topic
[14]. The integrated fuzzy neural network (FNN) system pos-
sesses the merits of both fuzzy systems [13] (e.g., humanlike
IF–THEN rule thinking and ease of incorporating expert knowl-
edge) and neural networks [15] (e.g., learning and optimization
abilities and connectionist structures). In this way, one can
bring the low-level learning and computational power of neural
networks into fuzzy systems and also high-level humanlike
IF–THEN rule thinking and reasoning of fuzzy systems into
neural networks. Moreover, adaptive control schemes of non-
linear systems that incorporate the techniques of FNN have
also grown rapidly [16]–[19]. Gao and Joo [16] proposed a
robust adaptive fuzzy neural controller (AFNC) for the iden-
tification and control of a class of uncertain MIMO nonlinear
systems. With the use of T–S fuzzy rules, the AFNC system
had good features such as fast learning speed and fast con-
vergence of tracking errors. However, prior system information
and constraint conditions were required in the control process.
Leu et al. [18] developed an observer-based direct adaptive
fuzzy neural control for nonaffine nonlinear systems. Although
this control system can achieve fine tracking performance,
some strict assumptions and auxiliary controller/observer de-
1327
sign were necessary. Wai and Chen [17], [19] devised T–S FNN
control (T-FNNC) and robust neural fuzzy network control
(RNFNC) for robot manipulators including actuator dynamics.
In these two approaches, they can offer favorable tracking
performance for complex robot systems. Unfortunately, the
control performance would be easily influenced by measure
noise because the joint velocity and the acceleration were
necessary in the control process. As a result, the applicability
of the control strategies in [17] and [19] will be substan-
tially limited in practice. The aim of this paper is to design
an adaptive FNN control (AFNNC) with indirect and direct
frameworks for the position control of an n-link robot manip-
ulator including actuator dynamics to cope with the problem
of static fuzzy modeling and to manipulate the control design
without the requirement of system information and constraint
conditions.
This paper is organized into five sections. Following the
introduction, the dynamic model of an n-link robot manipulator
including actuator dynamics introduced from [17] and [19] is
described briefly, and a T–S dynamic fuzzy modeling is investi-
gated to represent nonlinear dynamics of this robot manipulator
in Section II. In Section III, an indirect AFNNC (IAFNNC)
scheme with an FNN estimator is designed to evaluate the
nonlinear dynamic function vectors in local fuzzy models, and
a direct AFNNC (DAFNNC) strategy with an FNN controller is
used to imitate a predetermined model-based stabilizing control
law. All the design procedures of the proposed control schemes
are described in detail. The corresponding control laws and
adaptive tuning algorithms for FNN weights are derived in the
sense of Lyapunov stability analyses [20], [21] to ensure the
system-tracking stability. Numerical simulations and experi-
mental results of a two-link robot manipulator under the pos-
sible occurrence of uncertainties are provided to demonstrate
the robust control performance of the proposed schemes in
Section IV. Finally, conclusions are drawn in Section V.
II. S YSTEM D ESCRIPTION AND F UZZY M ODELING
A. Robotic Dynamic Model
In general, a robot arm consists of a number of links
connected and motor driving rotational joints. The governed
equation [17], [19] of an n-link robot manipulator including
actuator dynamics can be expressed as
...
q = α(q, q̇, q̈, ν t ) + βν t (1)
where q = [q1 q2 · · · qn ]T ∈ Rn , q̇ ∈ Rn , and q̈ ∈ Rn
are the vectors of joint position, velocity, and acceleration,
respectively; ν t = [νt1 νt2 · · · νtn ]T ∈ Rn is the con-
trol input vector representing the motor terminal voltages;
α(q, q̇, q̈, ν t ) ∈ Rn is a nonlinear dynamic function vector
related to the components of moment inertia, centripetal and
Coriolis forces, gravitational and frictional forces, and external
disturbance and unmodeled dynamics; and β ∈ Rn×n is a
positive-definite constant matrix. For convenience, a two-link
robot manipulator, as shown in Fig. 1, is utilized to verify the
positive-definite property of β, and the corresponding elements
in α(q, q̇, q̈, ν t ) and β are given in the Appendix.
1328 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 38, NO. 5, OCTOBER 2008

derivative of the filtered error vector with respect to time can be

represented as
 
ṡ = Cė = C Ax + Ω(x, ν t ) + Bν t − ẋd
= CAx + α(x, ν t ) + βν t − Cẋd . (6)

B. T–S Dynamic Fuzzy Modeling

Fig. 1. Architecture of two-link robot manipulator.
According to new state definitions of x1 = q, x2 = q̇, and
x3 = q̈, one can rewrite the robotic dynamic equation (1) as
ẋ1 = x2 ẋ2 = x3 ẋ3 = α(x1 , x2 , x3 , ν t ) + βν t .
By collecting state variables into the state vector x =
[x1 x2 x3 ]T ∈ R3n , (2) can be further represented in the
following state-space form:
ẋ = Ax + Ω(x, ν t ) + Bν t
Due to the complexity of the electric machinery structure
and the possible occurrence of parametric and environmental
uncertainties, the accurate mathematical model of a realistic
robot manipulator is generally difficult to obtain in many
practical applications. In this paper, the T–S fuzzy modeling
method [13] is utilized to represent the dynamic model of the
n-link robot manipulator including actuator dynamics in (2). It
is worth mentioning that the newly designed T–S fuzzy model
employed in this paper, which is a piecewise interpolation of
several subsystems through membership functions, belongs to a
continuous-time and dynamic type. In the continuous-time T–S
(2) dynamic fuzzy model, it includes fuzzy membership functions
in the premise part and local models in the consequent part. The
oth rule of the fuzzy model considered here is represented in the
following form [6]–[13]:
Ro : IF z1 is F1o and z2 is F2o . . . and zn is Fno ,
(3) THEN ẋ3o = αo (x, ν t )+β o ν t , o = 1, . . . , m (7)

where where zi |i=1,...,n denotes the input variables for the fuzzy
⎡ ⎤ model, m is the number of fuzzy model rules, and
0 I 0 Fio |i=1,...,n;o=1,...,m denotes the fuzzy sets, namely, linguistic
A = ⎣0 0 I ⎦ ∈ R3n×3n values associated with the input variable zi . In this paper, the
0 0 0 triangular membership function is adopted in the fuzzy model.
⎡ ⎤ Note that we let β 1 = β 2 = · · · = β m = β in local models
0 because β in (2) is a positive-definite constant matrix and its
Ω(x, ν t ) = ⎣ 0 ⎦ ∈ R3n
value can be predetermined by dynamic analyses. The domain
α(x, ν t ) F o = F1o × F2o × · · · × Fno defines a fuzzy region within the
⎡ ⎤
0 premise space associated with each fuzzy model rule. The fuzzy
B = ⎣ 0 ⎦ ∈ R3n×n Cartesian product is quantified by means of t-norm operator,
β which is usually implemented either as the min-operator or
the product operator [13]. In this paper, the algebraic product
in which I is an n × n identity matrix. The control objective operator is employed, and the corresponding rule firing strength
is to find a suitable control law so that the state vector x can is given by
track a specific command vector xd = [q d q̇ d q̈ d ]T . Define 
n
a tracking error vector and its filtered error vector as μF o (z) = μFio (zi ) (8)
i=1
e = x−xd = [q−q d q̇− q̇ d q̈−q̈ d ]T ≡ [q̃ q̃˙ ¨T
q̃] (4)
where μFio (zi ) is the triangular membership value with re-
¨ + Ka q̃˙ + Kb q̃ ≡ Ce
s = q̃ (5)
m to the respective input variable zi , μF (z) >
spect 0, and
o
T
o=1 μF o (z) = 1, in which z = [z1 z 2 . . . zn ] . By us-
where ing the weight average fuzzy inference approach, the global
fuzzy state-space linear model can be indicated as
C = [ Kb Ka I ] ∈ Rn×3n

m
ẋ = Ax + ϕo Ωo (x, ν t ) + Bν t (9)
in which Ka ∈ Rn×n and Kb ∈ Rn×n are the diagonal o=1
positive-definite matrices to be designed by the user. Once the
filtered error vector s = 0 occurs, the tracking error vector will where Ωo (x, ν t ) = [0 0 αo (x, ν t )]T , and ϕo |o=1,...,m ∈
approach to zero as time tends to infinite. Based on (3)–(5), the [0, 1] denotes the normalized rule firing value, which is
WAI AND YANG: ADAPTIVE FUZZY NN CONTROL DESIGN VIA T–S FUZZY MODEL 1329

Fig. 2. Block diagram of IAFNNC scheme.

Fig. 3. Block diagram of DAFNNC strategy.

defined by a four-layer FNN structure [19], as shown in Fig. 4, which

is composed of input, membership, rule, and output layers,
μF o (z) is adopted to implement the FNN estimator in IAFNNC and
ϕo (z) =
m , o = 1, . . . , m. (10)
μF o (z) the FNN controller in DAFNNC. The signal propagation and
o=1 the basic function in each layer of the FNN are introduced as
follows.
Because the nonlinear dynamic function vector is easily influ-
enced by system uncertainties, a static fuzzy modeling in (9) 1) Input layer transmits the input linguistic variables
may not meet the realistic situation in practice. In order to deal zb |b=1,...,r to the next layer.
with this problem, the AFNNC, including indirect and direct 2) Membership layer represents the input values with the
frameworks, is proposed in the following section. following Gaussian membership functions:
  2   2 
III. AFNNC
μjb (zb ) j
= exp − zb − mb tjb (11)
In the IAFNNC scheme, as shown in Fig. 2, an FNN estima-
tor is designed to tune the nonlinear dynamic function vector
in fuzzy local models, and then, the estimative vector is used where exp[·] is the exponential function, and mjb and
to indirectly develop a stable IAFNNC law. In the DAFNNC tjb (b = 1, . . . , r; j = 1, . . . , z) are the mean and the stan-
strategy, as shown in Fig. 3, an FNN controller is directly dard deviation of the Gaussian function in the jth term
designed to imitate a predetermined model-based stabilizing of the bth input variable zb to the node of membership
control law, and then, the stable control performance can be layer, respectively. It can be referred as the fuzzification
achieved by only using joint position information. In this paper, procedure.
1330 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 38, NO. 5, OCTOBER 2008

Fig. 4. Structure of four-layer FNN.

3) The output of each node in the rule layer is determined tion procedure. Moreover, it can be rewritten in the following
by fuzzy AND  operation. Each node in this rule layer is vector form:
denoted by , which multiplies the input signals and
outputs the result of the product. The product operation is y = [y1 y2 ··· yn ]T = Wl (14)
utilized to determine the firing strength. It can be referred
as the fuzzy inference mechanism. The output of this where
layer is given as ⎡w
11 w12 ··· w1p ⎤

r ⎢ w21 w22 ··· w2p ⎥
lk = k j
wjb μb (zb ) (12) W=⎢
⎣ .. .. .. .. ⎥
⎦
. . . .
b=1
wn1 wn2 ··· wnp
where lk |k=1,...,p represents the kth output of the rule Δ
k = [w1 w2 ··· wn ]T ∈ Rn×p
layer; wjb , which represents the weights between the
membership layer and the rule layer, is assumed to be l = [l1 l2 ··· lp ]T ∈ Rp×1 .
unity; and p is the number of rules.
4) Final layer is the output layer, and nodes in this layer
represent the output linguistic variables. Each node in A. IAFNNC Design
the output layer yi (i = 1, . . . , n) is labeled as , which
computes the overall output as the summation of all input In the IAFNNC scheme, the FNN estimator is constructed to
signals, and can be represented as regulate αo (x, ν t ) in local models to further improve the fuzzy
model accuracy. Moreover, the IAFNNC law and adaptive

p tuning algorithms for FNN weights are introduced from the

yi = wik lk . (13) stability analyses of the closed-loop system by using Lyapunov
k=1 method. The inputs of the FNN estimator are the elements in
the filtered error vector, and the outputs of the FNN estimator
The output node, together with the links connected to it, acts are the nonlinear dynamic function vectors in local models. For
as a defuzzifier. It can be referred as the normal defuzzifica- ease of notation, define the regulative values of the nonlinear
WAI AND YANG: ADAPTIVE FUZZY NN CONTROL DESIGN VIA T–S FUZZY MODEL 1331

 m  
dynamic function vectors in local models by collecting all the

outputs of the FNN in the following form: =s T o∗

ϕo Wik l + e + CAx + βν t − Cẋd
o=1
···
1
 ˙ o
Y = [α̂1 (s) α̂2 (s) α̂m (s)] (15) m
oT
+ tr Ŵik W̃ik
o
where α̂o (s)|o=1,...,m = Ŵik l, in which Ŵik o
|k=1,...,p ∈ ρ o=1
R n×p
denotes the adjustable weight matrix to be introduced  m 

m
later. = sT − o
ϕo W̃ik l+ o
ϕo Ŵik l + e
Based on the powerful approximation ability [14], there o=1 o=1
exists an optimal FNN estimator to approximate the nonlinear 
1
 ˙ o
m
dynamic function vector α(x, ν t ) in (2) such that
+ CAx + βν t − Cẋd + oT
tr Ŵik W̃ik
ρ o=1

m
o∗
α(x, ν t ) = ϕo Wik l + e (16)
m
o=1 =− ϕo sT W̃ik
o
l + sT
o∗ o=1
where the optimal weight matrix Wik and the estimative error m 
vector e ∈ R n×1
are assumed to be given by

  × ϕo α̂o (s) + e + CAx + βν t − Cẋd



m  o=1
o∗  o 
Wik = arg min sup α(x, ν t )− ϕo α̂o ( s| Ŵik )
1
 ˙ o
m
o ∈M
Ŵik w x∈Mx
  oT
o=1 + tr Ŵik W̃ik . (22)
ρ o=1
(17)
e  ≤ εe (18) Based on the fact that ϕo sT W̃ik o
l = tr(ϕo sT W̃ik
o
l) =
T
tr(ϕo sl W̃ikoT
) and on the inequality of ab ≤ a b, (22)
in which  ·  is the Euclidean norm, Mx and Mw are the pre-
o can be rewritten as
defined compact sets of x and Ŵik , and the positive constant

m 
m  
ε̄e can be reduced arbitrarily by increasing the number of rules. 1 ˙ o
Theorem 1: Consider an n-link robot manipulator including V̇ = tr −ϕo slT W̃ik oT
+ tr oT
Ŵik W̃ik
o=1 o=1
ρ
actuator dynamics represented by (1). If the IAFNNC law is  
designed as (19) and the weight update law is designed as + sT α̂(s) + CAx + βν t − Cẋd + sT e
(20), then the stability of the proposed IAFNNC system can

m   
be assured 1 ˙ o
≤ tr −ϕo slT + Ŵik W̃ik oT
  ρ
ν t = β −1 −α̂(s)−CAx+Cẋd −k1 s−k2 sgn(s) (19)
o=1
 
+sT α̂(s)+CAx+βν t −Cẋd +sT  e . (23)
˙ o
Ŵik = ρϕo slT (20)
m By substituting (19) and (20) into (23), then
where α̂(s) = o=1 ϕo α̂o (s); sgn(·) is a sign function vector;   
ρ is a positive learning rate; and k1 and k2 are the control
m
1 ˙ o
V̇ ≤ tr −ϕo slT + Ŵik W̃ik
oT
gains, in which the value of k2 should be selected to satisfy the ρ
o=1
condition of k2 > ε̄e and the selection of k1 > k2 is helpful to
alleviate the possible chattering phenomena caused by the sign − k1 sT s − sT (k2 − εe ) ≤ −k1 sT s. (24)
function in (19).
Proof: Define a Lyapunov function candidate as
o
Since V̇ (s(t), W̃ik ) ≤ 0, V̇ (s(t), W̃ik
o
) is a negative semidef-
inite function, i.e., V (s(t), W̃ik ) ≤ V (s(0), W̃ik
o o
). It implies
 1 1

m
o that s and W̃ik are bounded functions. Let function h(t) ≡
o
V s(t), W̃ik = sT s + o
tr(W̃ik oT
W̃ik ) (21)
2 2ρ o=1 k1 sT s ≤ −V̇ , and integrate function h(t) with respect to time

where W̃ik o o
= Ŵik − Wik o∗
, and tr(·) is a trace operator. By t  
differentiating (21) with respect to time and using (6), (16), and h(τ ) dτ ≤ V s(0), W̃ik
o
− V s(t), W̃ik
o
. (25)
(18), one can obtain 0

1
 ˙ o
m o
Because V (s(0), W̃ik ) is a bounded function, and
V̇ = sT ṡ + oT
tr Ŵik W̃ik o
ρ o=1 V (s(t), W̃ik ) is a nonincreasing and bounded function,
  the following result can be concluded:
= sT CAx + α(x, ν t ) + βν t − Cẋd
t
1
 ˙ o
m
+ oT
tr Ŵik W̃ik lim h(τ ) dτ < ∞. (26)
t→∞
ρ o=1 0
1332 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 38, NO. 5, OCTOBER 2008

In addition, ḣ(t) is bounded; thus, by Barbalat’s lemma [20], where K ∈ Rn×n and Ke ∈ Rn×n are the diagonal positive-
[21], it can be shown that limt→∞ h(t) = 0. It can imply that s definite matrices. Based on (29) and (30), one can obtain
will converge to zero as time tends to infinite. Based on (5), the  
tracking error vector e will approach to zero by the appropriate

m
...
−1
choice of parameter matrices Ka and Kb . Based on (19) and q̃ f = (sI + K) L(s) ϕo αo (x, ν t ) + βν t − qd +KT
ce
(24), it is obvious that the selection of k1 > k2 is helpful to o=1
(33)
alleviate the possible chattering phenomena caused by the sign
function in (19). In other words, the term of −k1 sT s via a large where L(s) = Ke [(sI + Ki )n−1 ]−1 is a stable filter. From (31),
value of k1 in (24) is enough to ensure the system stability, (33) can be rewritten as
even though a smaller value of k2 is selected to unsatisfy the m 
condition of k2 > εe . In this paper, the saturation function is
...
unnecessary and is not applied for the simulation and experi- q̃˙ f + Kq̃ f = L(s) ϕo αo (x, ν t ) + βν t − q d +Kc e
T

o=1
ment to replace the sign function. As a result, the stable control
behavior of the proposed IAFNNC system can be ensured ≡ ω(x, q̃ f , ν t ) + ν t (34)
without the requirement of detailed system information. m ...
where ω(x, q̃ f , ν t ) = {L(s)[ o=1 ϕo αo (x, ν t )+βν t− q d +
KTc e] − ν t }. Note that, although the right-hand side of (34)
B. DAFNNC Design contains unknown parameters and signals, the vector q̃ f can
be obtained via filtering q̃, which is available [i.e., using (31)].
Based on (4) and (9), the output error vector (q̃) and the In other words, the dynamical equation (34) is only used for
derivative of the tracking error vector (e) with respect to time analytic purposes.
can be represented as Based on (34), a stabilizing control law (ν sc ) for the arma-
ture input-voltage vector ν t can be designed as
q̃ = Ce (27)
  ν sc = −ω(x, q̃ f , ν t ) − kf sgn(q̃ f ) (35)
ė = A − BKT
c e
m  where kf ∈ R is a constant value. By considering a Lyapunov

+B
...
ϕo αo (x, ν t ) + βν t − q d +KT function candidate Vsc = q̃ T f q̃ f /2 and taking the derivative of
ce (28)
o=1 Vsc with respect to time, there exists a stabilizing control law
in (35) such that V̇sc ≤ −q̃ T
f Kq̃ f < 0. Based on the Lyapunov
where stability theorem [20], [21] and (32), it can imply that the output
error vector q̃ will converge to zero as t → ∞. However, if the
B = [0 0 I]T ∈ R3n×n system dynamics and the joint velocity/acceleration informa-
T
tion are perturbed or unknown, it is difficult to implement this
C = [I 0 0]T ∈ R3n×n stabilizing control law in practical applications, and the stability
of the controlled system may be destroyed. Moreover, the chat-
in which KT c ∈R
n×n tering control effort could be induced due to the conservative
is a feedback gain matrix. Based on (27)
and (28), the output error vector can be rewritten as selection of a larger control gain kf . To ensure the stability of
the controlled system despite the existence of uncertainties, an
 

m
...
FNN controller is investigated to mimic the stabilizing control
q̃ = H(s) ϕo αo (x, ν t ) + βν t − q d +KT
ce (29) law in (35) to ensure the robust control performance. Based on
o=1 the powerful approximation ability [14], there exists an optimal
FNN controller to approximate the stabilizing control law (ν sc )
where s is the Laplace operator, and H(s) = C[sI − (A − in (35) such that
−1
BKT c )] B. If the feedback gain matrix is selected as Kc =
T
∗
[λnn (Ki )n , . . . , λ2n (Ki )2 , λ1n Ki ] with λin = n!/[i!(n − i)!], it ν sc = Wik l+ (36)
can easily show that
where  is a minimum reconstructed-error vector and is as-
sumed to be bounded by  < ε, in which ε is a given positive
H(s) = [(sI + Ki )n ]−1 (30) ∗
constant, and Wik is the optimal weight matrix of Wik in
the FNN. Thus, the FNN control law is assumed to take the
in which Ki ∈ Rn×n is a diagonal positive-definite matrix. following form:
Define a new variable q̃ f ∈ Rn [22] as
ν t = Ŵik l (37)
q̃ f = (sI + K)−1 Ke (sI + Ki )q̃ (31)
where Ŵik is the estimate of the optimal weight matrix, as
provided by tuning algorithms to be introduced later. By sub-
or
tracting (36) from (37), an approximation error ν̃ is defined as
q̃˙ f + Kq̃ f = Ke (q̃˙ + Ki q̃) (32) ν̃ = ν t − ν sc = W̃ik l −  (38)
WAI AND YANG: ADAPTIVE FUZZY NN CONTROL DESIGN VIA T–S FUZZY MODEL 1333

∗
where W̃ik = Ŵik − Wik . In this paper, the inputs of the
FNN controller in the DAFNNC strategy are q̃ and q̃ f , and
the output of the FNN controller is ν t . Based on (29), the
output error vector (q̃) can be represented as a function vector
of x and ν t . Hence, this FNN is a reasonable design because
approximating the optimal stabilizing control law ν sc in (35)
depends on x, q̃ f , and ν t .
Theorem 2: Consider an n-link robot manipulator including
actuator dynamics represented by (1). If the DAFNNC law is
designed as (37) and the weight update law is designed as
(39), then the stability of the proposed DAFNNC system can
be assured with only the joint position information.
By substituting (39) into (42) and using the selection of kf > ε,
(42) can be represented as
V̇d ≤ −q̃ T
f Kq̃ f − q̃ f sgn(q̃ f )(kf − ε) ≤ −q̃ f Kq̃ f .
T T
(43)
Since V̇d (q̃ f (t), W̃ik ) ≤ 0, V̇d (q̃ f (t), W̃ik ) is negative semi-
definite function, i.e., Vd (q̃ f (t), W̃ik ) ≤ Vd (q̃ f (0), W̃ik ). It
implies that q̃ f and W̃ik are bounded functions. Let function
f Kq̃ f ≤ −V̇d , and integrate function f (t) with re-
f (t) ≡ q̃ T
spect to time
t 


˙ f (τ ) dτ ≤ Vd q̃ f (0), W̃ik − Vd q̃ f (t), W̃ik . (44)
Ŵik = −ηq̃ f lT (39)
0
where η is a positive learning rate.
Proof: Define a Lyapunov function candidate as Because Vd (q̃ f (0), W̃ik ) is a bounded function and
1 
Vd (q̃ f (t), W̃ik ) is a nonincreasing and bounded function,
1 T T
Vd (q̃ f (t), W̃ik ) = q̃ f q̃ f + tr W̃ik W̃ik . (40) the following result can be concluded:
2 2η
t
By differentiating (40) with respect to time and using (34), (35),
and (38), one can obtain lim f (τ ) dτ < ∞. (45)
t→∞
0

˙ 1  ˙
V̇d = q̃ T
f q̃ f
T
+ tr Ŵik W̃ik
η In addition, f˙(t) is bounded; thus, by Barbalat’s lemma [20],
  1  ˙
[21], it can be shown that lim f (t) = 0. It can imply that
= q̃ T
f −Kq̃ f + ω(x, q̃ f , ν t ) + ν t + tr Ŵik W̃ikT
t→∞
 η  q̃ f will converge to zero as time tends to infinite. Based on
= q̃ f −Kq̃ f + ω(x, q̃ f , ν t ) + ν sc + W̃ik l − 
T
(32), the output error vector q̃ will approach to zero by the
1  ˙ T
appropriate choice of parameter matrices K, Ke , and Ki . As
+ tr Ŵik W̃ik a result, the stable control behavior of the proposed DAFNNC
η  system can be ensured without the requirement of detailed
f −Kq̃ f − kf sgn(q̃ f ) + W̃ik l − 
= q̃ T system dynamics and joint velocity/acceleration information.
1  ˙ T

The effectiveness of the proposed control schemes is veri-
+ tr Ŵik W̃ik
η fied by the following numerical simulations and experimental
 
= −q̃ T f Kq̃ f + q̃ f −kf sgn(q̃ f ) − 
T results.
1  ˙
+ q̃ Tf W̃ik l + tr Ŵ ik W̃ T
ik . (41)
η IV. N UMERICAL S IMULATIONS AND
E XPERIMENTAL R ESULTS
Based on the fact that q̃ T T
f W̃ik l = tr (q̃ f W̃ik l) =
A two-link robot manipulator, as shown in Fig. 1, is utilized
tr (q̃ f l W̃ik ) and on the inequality of ab ≤ a b,
T T
in this paper to verify the validity of the proposed control
(41) can be rewritten as
schemes. The detailed system parameters of this robot manip-
   ulator actuated by dc servomotors [17], [19] are given in (46),
1 ˙
V̇d ≤ tr T T
q̃ f l + Ŵik W̃ik shown at the bottom of the page, where the suffixes “1” and “2”
η
 T denote the ith actuated motor or link of the robot manipulator.
T   .
f Kq̃ f − kf q̃ f sgn(q̃ f ) + q̃ f
− q̃ T (42) The control objective is to control the joint angles of a two-link

Jm1 = 3.7 × 10−5 kg · m2 Bm1 = 1.3 × 10−5 N · m/s KT1 = 0.21 N · m/A
Jm2 = 1.47 × 10−4 kg · m2 Bm2 = 2 × 10−5 N · m/s KT2 = 0.23 N · m/A
Ra1 = 2.8 Ω La1 = 3 mH KE1 = 2.42 × 10−4 s/radV
Ra2 = 4.8 Ω La2 = 2.4 mH KE2 = 2.18 × 10−4 s/radV (46)
lt1 = 205 mm lc1 = 154.8 mm m1 = 3.55 kg
lt2 = 210 mm lc2 = 105 mm m2 = 0.75 kg
g = 9.8 m/s2 gr1 = 60 gr2 = 30
1334 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 38, NO. 5, OCTOBER 2008
robot manipulator to follow periodic sinusoidal trajectories.
In this AFNNC system, the inputs of the fuzzy model are q1
and q2 , and three fuzzy model rules (m = 3) with the type of
(7) are adopted because this fuzzy model has online dynamic
learning ability. Moreover, the inputs of the FNN estimator
in the IAFNNC scheme are s1 and s2 , and the output of the
FNN estimator is α̂o (s)|o=1,2,3 . By dividing the inputs of the
FNN estimator into five fuzzy sets equally and considering full-
connection estimator rules, the FNN estimator has two, ten,
twenty-five, and six neurons at the input, membership, rule, and
output layer, respectively. It can be regarded that the associated
fuzzy sets with Gaussian function for each input signal are
divided into negative big, negative small, zero (Z), positive
small, and positive big. In addition, the inputs of the FNN
controller in the DAFNNC strategy are q̃1 , q̃2 , q̃f 1 , and q̃f 2 , and
the output of the FNN controller is ν t . By dividing the inputs of
the FNN controller into three fuzzy sets equally and considering
full-connection controller rules, the FNN controller has four,
twelve, eighty-one, and two neurons at the input, membership,
rule, and output layer, respectively. It can be regarded that the
associated fuzzy sets with Gaussian function for each input
signal are divided into negative, Z, and positive. Usually, some
heuristics can be used to roughly initialize the parameters of
the FNN for practical applications; for example, the mean and
standard deviation of Gaussian functions can be determined
according to the maximum variations of s, q̃, and q̃ f . The effect
due to the inaccurate selection of the initialized parameters can
be retrieved by the online training methodology.
In order to exhibit the superior control performance of
the proposed IAFNNC and DAFNNC schemes, proportional–
differential control (PDC), fuzzy-model-based control (FMBC)
in [12], T-FNNC in [17], and RNFNC in [19] are examined in
the meanwhile. The PDC law can be expressed as
˙
ν t = Kp q̃(t) + Kd q̃(t)
where Kp and Kd denote the proportional and the differential
gain matrices, respectively. Moreover, the gains in these control
systems are given as
    
100 0 1000 0 1
Ka = Kb = β=
0 50 0 1500 0 1
   
600 0 20 500 0 71.1%, and 29.7% position-tracking improvements than the
K= Ke =
0 800 0 19 500
   
0.03 0 1500 0 tracking responses of the DAFNNC strategy without joint
Ki = Kp =
0 0.01 0 800
 
100 0
Kd =
0 30
ρ = 0.5 k1 = 1 k2 = 0.2 η = 7.5.
All the gains are chosen to achieve the superior transient control
performance in the simulation and experimentation considering
the limitation of control effort, the requirement of stability,
and the possible operating conditions. For recording respective
control performance, the mean-square-error (mse) measure of
the position-tracking response is defined as
1
T
mse = [qi (μ) − qdi (μ)]2 (49)
T μ=1
where T is the total sampling instant, and qi and qdi are the
elements in the vectors q and q d . Based on (49), the normalized
mse (nmse) value of the position-tracking response using a per-
unit value with 1 rad is used to examine the control perfor-
mance. In this paper, the simulation and the experimentation are
carried out by using Matlab and Turbo C software, respectively.
A. Numerical Simulations
The most important parameters that affect the control perfor-
mance of the robotic system are the external disturbance tl (t),
the friction term f (q̇), and the parameter variation of the mass
of link 2 m2 . In numerical simulations, three circumstances,
including nominal situation (m2 = 0.75 kg and n = 0) at the
beginning, parameter variation situation occurring at 4 s, and
disturbance situation occurring at 7 s, are considered. The
parameter variation situation is that 1 kg weight is added to the
mass of link 2, i.e., m2 = 1.75 kg. The disturbance situation is
that external forces are injected into the robotic system and that
their shapes are expressed as follows:
tl (t) = [5 sin(5t) 0.5 sin(5t)]T . (50)
In addition, friction forces are also considered in this simula-
tions and are given as
f (q̇) = [20q̇1 + 0.8sgn(q̇1 ) 4q̇2 + 0.16sgn(q̇2 )]T . (51)
(47)
Hence, two simulation cases (i.e., cases S1 and S2), including
the aforementioned three situations, are adopted to demon-
strate the robust property of the proposed control schemes.
Case S1 denotes the three situations without joint friction, and
the friction forces are considered in case S2.
The simulated comparisons of PDC, FMBC in [12], T-FNNC
 in [17], RNFNC in [19], and IAFNNC and DAFNNC systems
0
are summarized in Table I. Based on the nmse measures,
the proposed DAFNNC strategy has over 83.3%, 56%, 58%,
PDC, FMBC, T-FNNC, and RNFNC and IAFNNC schemes,
respectively. From the records in Table I, the joint position-
velocity/acceleration measurements can be controlled to more
closely follow specific reference trajectories under a wide range
of operating conditions and the occurrence of uncertainties than
the other ones.
(48)
B. Experimental Results
Some experimental results are provided here to further verify
the effectiveness of the proposed methodologies. An image of
a practical control system for a two-link robot manipulator
WAI AND YANG: ADAPTIVE FUZZY NN CONTROL DESIGN VIA T–S FUZZY MODEL
TABLE I
SIMULATED COMPARISONS OF PDC, FMBC, T-FNNC, RNFNC, IAFNNC, AND DAFNNC SYSTEMS
Fig. 5. Image of practical control system.
driven by two dc servomotors is shown in Fig. 5. A servo
control card is installed in the control computer, which includes
multichannels of digital/analog and encoder interface circuits.
Digital filter and frequency multiplied by four circuits are built
into the encoder interface circuit to increase the precision of
position feedback. All the proposed methodologies are realized
in a Pentium CPU, and the control interval of the position
control loop is set at 1 ms. Two perturbed factors are given
to verify the system robustness. One is the parameter variation
factor, i.e., the nominal mass with a tip load of 1 kg weight, and
the other is the external disturbance factor, i.e., the increasing
of plastic disks between joint connections. To the end, four
experimental conditions, including two perturbed factors, are
adopted to demonstrate the robust property of the proposed
control scheme. Conditions E1 and E2 denote the absence of
joint friction with no tip load and 1-kg loading, respectively;
conditions E3 and E4 represent the existence of joint friction
with no tip load and 1-kg loading, respectively.
The experimental position responses, control efforts, and
tracking errors of the PDC system in (47) at four condi-
tions are shown in Fig. 6. The ones at condition E1 are
shown in Fig. 6(a)–(f), the ones at condition E2 are shown
in Fig. 6(g)–(l), the ones at condition E3 are shown in
1335
Fig. 6(m)–(r), and the ones at condition E4 are shown in
Fig. 6(s)–(x). By analyzing the experimental results of the
PD control system, it can be discovered that if there is a
tip load or joint friction inserted into the robotic system, the
tracking performance is degenerated. Although the selection
of large control gains can improve this phenomenon, it will
result in large control efforts and may excite the unstable
system dynamics, particularly in tuning the differential gain.
In industry, the conventional PD or PID controllers have been
widely used due to their simple control structure, ease of
design, and inexpensive cost. However, successful applications
of these model-free simple controllers usually require manual
parameters retuning before being transferred to the process
under control. Fig. 7 shows the experimental results of the
FMBC in [12] to verify the necessity of the FNN estimator for
dynamic model regulation. As can be seen from vibrate tracking
errors in Fig. 7(c), (f), (i), (l), (o), (r), (u), and (x) at four
conditions, the tracking errors do not diminish with time grad-
ually. By observing Figs. 6 and 7, it is obvious that the FMBC
scheme shows a better performance than the PDC system due to
the powerful control ability of the fuzzy-model-based strategy.
Nevertheless, the control performance of the FMBC system is
sensitive to system uncertainties and measure noise because the
T–S fuzzy model dynamics in [12] may not be governed by
all the linear subsystems at each operating point and all system
state variables should be measurable in [12].
For comparison, the experimental position responses, control
efforts, and tracking errors of the IAFNNC and DAFNNC sys-
tems at four conditions are shown in Figs. 8 and 9, respectively.
By observing the experimental results in Figs. 8 and 9, both
the tracking errors can converge to an acceptable region, and
the control performance is excellent both in the conditions of
joint friction and parameter variation. Because all parameters
in the IAFNNC and DAFNNC are roughly initialized, suitable
control efforts are regulated through online training process
whether system uncertainties exist or not. Although the pro-
posed IAFNNC scheme can perform superior control property
than the PDC and FMBC in [12], the requirement of precise
velocity and acceleration sensors will increase the hardware
cost, and the measure noise may result in degenerate control
performance. The experimental comparisons of PDC, FMBC
in [12], T-FNNC in [17], RNFNC in [19], and IAFNNC and
1336 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 38, NO. 5, OCTOBER 2008

Fig. 6. Experimental position responses, control efforts, and tracking errors of PDC system: (a)–(f) at joints 1 and 2 for condition E1; (g)–(l) at joints 1 and 2 for
condition E2.
WAI AND YANG: ADAPTIVE FUZZY NN CONTROL DESIGN VIA T–S FUZZY MODEL 1337

Fig. 6. (Continued.) Experimental position responses, control efforts, and tracking errors of PDC system: (m)–(r) at joints 1 and 2 for condition E3; and
(s)–(x) at joints 1 and 2 for condition E4.
1338 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 38, NO. 5, OCTOBER 2008

Fig. 7. Experimental position responses, control efforts, and tracking errors of FMBC system: (a)–(f) at joints 1 and 2 for condition E1; (g)–(l) at joints 1 and 2
for condition E2.
WAI AND YANG: ADAPTIVE FUZZY NN CONTROL DESIGN VIA T–S FUZZY MODEL 1339

Fig. 7. (Continued.) Experimental position responses, control efforts, and tracking errors of FMBC system: (m)–(r) at joints 1 and 2 for condition E3; and
(s)–(x) at joints 1 and 2 for condition E4.
1340 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 38, NO. 5, OCTOBER 2008

Fig. 8. Experimental position responses, control efforts, and tracking errors of IAFNNC system: (a)–(f) at joints 1 and 2 for condition E1; (g)–(l) at joints 1 and 2
for condition E2.
WAI AND YANG: ADAPTIVE FUZZY NN CONTROL DESIGN VIA T–S FUZZY MODEL 1341

Fig. 8. (Continued.) Experimental position responses, control efforts, and tracking errors of IAFNNC system: (m)–(r) at joints 1 and 2 for condition E3; and
(s)–(x) at joints 1 and 2 for condition E4.
1342 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 38, NO. 5, OCTOBER 2008

Fig. 9. Experimental position responses, control efforts, and tracking errors of DAFNNC system: (a)–(f) at joints 1 and 2 for condition E1; (g)–(l) at joints 1 and 2
for condition E2.
WAI AND YANG: ADAPTIVE FUZZY NN CONTROL DESIGN VIA T–S FUZZY MODEL 1343

Fig. 9. (Continued.) Experimental position responses, control efforts, and tracking errors of DAFNNC system: (m)–(r) at joints 1 and 2 for condition E3; and
(s)–(x) at joints 1 and 2 for condition E4.
1344 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 38, NO. 5, OCTOBER 2008
TABLE II
EXPERIMENTAL COMPARISONS OF PDC, FMBC, T-FNNC, RNFNC, IAFNNC, AND DAFNNC SYSTEMS
DAFNNC systems are summarized in Table II, respectively.
Based on the nmse measures, the proposed DAFNNC strategy
has over 86.9%, 32.5%, 36.4%, 64.7%, and 16.6% position-
tracking improvements than the PDC, FMBC in [12], T-FNNC
in [17], RNFNC in [19], and IAFNNC systems, respectively.
From both the simulation and the experimental results, the
control performance of the proposed DAFNNC at any test
conditions is better than that of the IAFNNC. The major reason
is the implementation of joint velocities and accelerations in the
IAFNNC by the differential approximation method (i.e., with-
out tachometer and accelerator in practice). It will result in irre-
sistible velocity and acceleration measure errors in comparison
with actual signals, particularly in practical applications. On the
other hand, it will further cause the bandwidth reduction and the
slow response although one can use a filter mechanism to cope
with the possible occurrence of noise. Relatively, the proposed
DAFNNC strategy with only joint position measurement indeed
needs fewer sensors and yields favorable control performance
than the other ones.
V. C ONCLUSION
This paper has successfully implemented AFNNC, including
indirect and direct frameworks, to control the joint position of
an n-link robot manipulator including actuator dynamics for
achieving high-precision position control. Numerical simula-
tions and experimental results of a two-link robot manipulator
via various existing control frameworks were also provided in
this paper to compare and display the superior performance of
the proposed control schemes. Compared with previous works,
the major contributions of this paper are recited as follows:
1) the successful construction of the entire system behavior
of the n-link robot manipulator including actuator dynamics to
redeem the incomplete robotic dynamics for controller design
in the published literature [1]–[4]; 2) the successful implemen-
tation of an FNN estimator in the IAFNNC scheme to form a
continuous-time T–S dynamic fuzzy model so that there are no
strict constraints in the control design and that all the system
dynamics could be unknown in comparison with [7] and [8];
3) the successful development of a major FNN controller via
only joint position information in the DAFNNC strategy to
enhance the system applicability and to cope with the problem
cause by the assumption of all system state variables to be mea-
surable in previous researches [12], [17], [19]; 4) the successful
guarantee of the system stability for the proposed DAFNNC
strategy without auxiliary controllers, strict constraints, and
detailed system knowledge to relax the complex calculation
to solve sufficient conditions for the closed-loop stability in
[10] and [11]; and 5) the successful comparisons of individual
control performances to provide designers with preliminary
guideline for manipulating the n-link robot manipulator effi-
ciently and the convenient extension of this DAFNNC strategy
with fewer precise sensors (i.e., only position sensor) and
simple control framework (i.e., only one FNN) to other control
problems.
A PPENDIX
From Fig. 1 [17], [19], the nonlinear dynamic function vector
in (1) can be represented as
 
α(q, q̇, q̈, ν t ) = −(M∗ )−1 [D(q, q̇, q̈)+ d]+ (M∗ )−1 − β ν t
(A1)
with
M∗ = L
n [M n + Jn ]  (A2)
D(q, q̇, q̈) = Ln Ṁθ (q, q̇) + C(q, q̇) + Bn
+ Rn [M(q) + Jn ] q̈

+ Ln Ċ(q, q̇, q̈)q̇ + Rn C(q, q̇)q̇

+ Rn Bn q̇ + KEn q̇ + Ln ġ(q, q̇)+ Rn g(q)
(A3)
WAI AND YANG: ADAPTIVE FUZZY NN CONTROL DESIGN VIA T–S FUZZY MODEL 1345

2 2 2 2

m1 lc1 + m2 lt1 + m2 lc2 + 2m2 lt1 lc2 cos q2 m2 lc2 + m2 lt1 lc2 cos q2
M(q) = 2 2
m2 lc2 + m2 lt1 lc2 cos q2 m2 lc2
 2 2 2 2
  
m1 lc1 + m2 lt1 + m2 lc2 m2 lc2 2m2 lt1 lc2 cos q2 m2 lt1 lc2 cos q2
Mn = 2 2 Mθ (q) =
m2 lc2 m2 lc2 m2 lt1 lc2 cos q2 0
   La1 
−m2 lt1 lc2 sin q2 q̇2 −m2 lt1 lc2 sin q2 (q̇1 + q̇2 ) gr1 KT1 0
C(q, q̇) = Ln = La2
m2 lt1 lc2 sin q2 q̇1 0 0 gr2 KT2
   2

m1 glc1 cos q1 + m2 glt1 cos q1 + m2 glc2 cos(q1 + q2 ) gr1 Jm1 0
g(q) = Jn = 2
m2 glc2 cos(q1 + q2 ) 0 gr2 Jm2
 Ra1     
gr1 KT1 0 2
gr1 Bm1 0 KE1 gr1 0
Rn = Ra2 Bn = 2 KEn =
0 gr2 KT2
0 gr2 Bm2 0 KE2 gr2

...
d = Ln Mθ (q) q +Ln ṅ + Rn n (A4)
β11 β12
β= (A5)
β21 β22
ν t = [νt1 νt2 ]T (A6)
where νt1 and νt2 are the armature input voltages for joint
dc motors, and the expressions M(q), Mn , Mθ (q), C(q, q̇),
Ln , g(q), Jn , Rn , Bn , and KEn are shown at the top of
the page, in which q, q̇, and q̈ are the joint position, velocity,
and acceleration vectors, respectively; M(q) = Mn + Mθ (q)
is the inertia matrix; C(q, q̇) is the matrix of centripetal
and Coriolis forces; g(q) is the gravity vector; n represents
the vector of external disturbance tl , friction term f (q̇), and
unmodeled dynamics; q1 and q2 are the angles of joints 1
and 2; m1 and m2 are the masses of links 1 and 2; lt1
and lt2 are the total lengths of links 1 and 2; lc1 and lc2
are the center-of-gravity lengths of links 1 and 2; g is the
gravity acceleration; gr1 and gr2 are the gear ratios for joints
1 and 2; (KT1 , KT2 ), (Jm1 , Jm2 ), (Bm1 , Bm2 ), (La1 , La2 ),
(Ra1 , Ra2 ), and (KE1 , KE2 ) denote the torque constants,
moment inertias, torsional damping coefficients, armature
inductances, armature resistances, and back electromotive-
force coefficients of joint motors, respectively. By observing
(A1), one can see that the constant matrix β is particularly
introduced into (1) to manipulate the model-free AFNNC
design. Thus, the positive-definite property of the designed
matrix β can be easily ensured according to the selection
of β11 > 0 and det[β] > 0, where det[·] is the determinant
operator.
R EFERENCES
[1] V. Santibanez, R. Kelly, and M. A. Llama, “A novel global asymptotic
stable set-point fuzzy controller with bounded torques for robot manipu-
lators,” IEEE Trans. Fuzzy Syst., vol. 13, no. 3, pp. 362–372, Jun. 2005.
[2] E. Kim, “Output feedback tracking control of robot manipulators with
model uncertainty via adaptive fuzzy logic,” IEEE Trans. Fuzzy Syst.,
vol. 12, no. 3, pp. 368–378, Jun. 2004.
[3] W. Gao and R. R. Selmic, “Neural network control of a class of nonlinear
systems with actuator saturation,” IEEE Trans. Neural Netw., vol. 17,
no. 1, pp. 147–156, Jan. 2006.
[4] E. M. Jafarov, M. N. A. Parlakçý, and Y. Istefanopulos, “A new vari-
able structure PID-controller design for robot manipulators,” IEEE Trans.
Control Syst. Technol., vol. 13, no. 1, pp. 122–130, Jan. 2005.
[5] B. S. Chen, H. J. Uang, and C. S. Tseng, “Robust tracking enhancement
of robot systems including motor dynamics: A fuzzy-based dynamic
game approach,” IEEE Trans. Fuzzy Syst., vol. 6, no. 4, pp. 538–552,
Nov. 1998.
[6] G. Feng, “A survey on analysis and design of model-based fuzzy con-
trol systems,” IEEE Trans. Fuzzy Syst., vol. 14, no. 5, pp. 676–697,
Oct. 2006.
[7] H. K. Lam and F. H. F. Leung, “Fuzzy controller with stability and
performance rules for nonlinear systems,” Fuzzy Sets Syst., vol. 158, no. 2,
pp. 147–163, Jan. 2007.
[8] P. P. Kumar, I. Kar, and L. Behera, “Variable-gain controllers for nonlinear
systems using the T–S fuzzy model,” IEEE Trans. Syst., Man, Cybern. B,
Cybern., vol. 36, no. 6, pp. 1442–1449, Dec. 2006.
[9] C. H. Tseng, B. S. Chen, and H. J. Uang, “Fuzzy tracking control design
for nonlinear dynamic systems via T–S fuzzy model,” IEEE Trans. Fuzzy
Syst., vol. 9, no. 3, pp. 381–392, Jun. 2001.
[10] C. Lin, Q. G. Wang, and T. H. Lee, “H-infinite tracking control for
nonlinear systems via T–S fuzzy model approach,” IEEE Trans. Syst.,
Man, Cybern. B, Cybern., vol. 36, no. 2, pp. 450–457, Apr. 2006.
[11] K. Y. Lian, C. H. Su, and C. S. Huang, “Performance enhancement for T–S
fuzzy control using neural networks,” IEEE Trans. Fuzzy Syst., vol. 14,
no. 5, pp. 619–627, Oct. 2006.
[12] C. C. Cheng and S. H. Chien, “Adaptive sliding mode controller design
based on T–S fuzzy system models,” Automatica, vol. 42, no. 6, pp. 1005–
1010, Jun. 2006.
[13] L. X. Wang, A Course in Fuzzy Systems and Control. Englewood Cliffs,
NJ: Prentice-Hall, 1997.
[14] C. T. Lin and C. S. George Lee, Neural Fuzzy Systems. Englewood
Cliffs, NJ: Prentice-Hall, 1996.
[15] O. Omidvar and D. L. Elliott, Neural Systems for Control. New York:
Academic, 1997.
[16] Y. Gao and M. Joo, “Online adaptive fuzzy neural identification and
control of a class of MIMO nonlinear systems,” IEEE Trans. Fuzzy Syst.,
vol. 11, no. 4, pp. 462–477, Aug. 2003.
[17] R. J. Wai and P. C. Chen, “Intelligent tracking control for robot manip-
ulator including actuator dynamics via TSK-type fuzzy neural network,”
IEEE Trans. Fuzzy Syst., vol. 12, no. 4, pp. 552–559, Aug. 2004.
[18] Y. G. Leu, W. Y. Wang, and T. T. Lee, “Observer-based direct adap-
tive fuzzy-neural control for nonaffine nonlinear systems,” IEEE Trans.
Neural Netw., vol. 16, no. 4, pp. 853–861, Jul. 2005.
[19] R. J. Wai and P. C. Chen, “Robust neural-fuzzy-network control for robot
manipulator including actuator dynamics,” IEEE Trans. Ind. Electron.,
vol. 53, no. 4, pp. 1328–1349, Aug. 2006.
[20] K. J. Astrom and B. Wittenmark, Adaptive Control. New York: Addison-
Wesley, 1995.
[21] H. K. Khalil, Nonlinear Systems. Englewood Cliffs, NJ: Prentice-Hall,
1996.
[22] A. Boulkroune, M. Tadjine, M. M’Saad, and M. Farza, “How to design a
fuzzy adaptive controller based on observers for uncertain affine nonlinear
systems,” Fuzzy Sets Syst., vol. 159, no. 8, pp. 926–948, Apr. 2008.
1346 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 38, NO. 5, OCTOBER 2008

Rong-Jong Wai (A’00–M’02–SM’05) was born in Zhi-Wei Yang was born in Taichung, Taiwan,
Tainan, Taiwan, R.O.C., in 1974. He received the R.O.C., in 1982. He received the B.S. degree in elec-
B.S. degree in electrical engineering and the Ph.D. trical engineering from Tamkang University, Tamsui,
degree in electronic engineering from Chung Yuan Taiwan, in 2005, and the M.S. degree in electrical
Christian University, Chung Li, Taiwan, in 1996 and engineering from Yuan Ze University, Chung Li,
1999, respectively. Taiwan, in 2008.
Since 1999, he has been with the Department of He currently serves with the Air Force, Ministry
Electrical Engineering, Yuan Ze University, Chung of National Defense, Taipei, Taiwan. His research
Li, where he is currently a Professor and is the Direc- interests include motor servo drives, fuzzy theory,
tor of the Electric Control and System Engineering neural network, and robot manipulators.
Laboratory. He is also the Dean of General Affairs,
Yuan Ze University, Chung Li. He is a chapter author of Intelligent Adaptive
Control: Industrial Applications in the Applied Computational Intelligence Set
(CRC Press, 1998) and the coauthor of Drive and Intelligent Control of Ultra-
sonic Motor (Tsang-Hai, 1999), Electric Control (Tsang-Hai, 2002), and Fuel
Cell: New Generation Energy (Tsang-Hai, 2004). He has authored numerous
published journal papers in the area of control system applications. His research
interests include power electronics, motor servo drives, mechatronics, energy
technology, and control theory applications.
Dr. Wai’s biography was listed in Who’s Who in Science and Engineering
(Marquis Who’s Who) in 2004–2009, Who’s Who (Marquis Who’s Who)
in 2004–2009, Leading Scientists of the World (International Biographical
Centre) in 2005, Who’s Who in Asia (Marquis Who’s Who), Who’s Who of
Emerging Leaders (Marquis Who’s Who) in 2006–2009, and Asia/Pacific Who’s
Who (Rifacimento International) in Vols. VII and VIII. He was the recipient
of the Excellent Research Award in 2000, and the Wu Ta-You Medal and
Young Researcher Award in 2003 from the National Science Council, R.O.C.
In addition, he was the recipient of the Outstanding Research Award in 2003
and 2007 from the Yuan Ze University, R.O.C.; the Excellent Young Electrical
Engineering Award in 2004 from the Chinese Electrical Engineering Society,
R.O.C.; the Outstanding Professor Award in 2004 and 2008 from the Far
Eastern Y.Z. Hsu-Science and Technology Memorial Foundation, R.O.C.; the
International Professional of the Year Award in 2005 from the International
Biographical Centre, U.K.; the Young Automatic Control Engineering Award
in 2005 from the Chinese Automatic Control Society, R.O.C.; the Yuan-Ze
Lecture Award in 2007 from the Far Eastern Y. Z. Hsu-Science and Technology
Memorial Foundation, R.O.C.; and the Invent Silver Metal Award in 2007 from
the International Invention Show and Technomart, Taipei, R.O.C.