You are on page 1of 11

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 27, NO.

9, SEPTEMBER 2019 1703

Fuzzy Adaptive State-Feedback Control Scheme of


Uncertain Nonlinear Multivariable Systems
Abdesselem Boulkroune , Loubna Merazka, and Hongyi Li , Senior Member, IEEE

Abstract—In this article, we propose a new fuzzy adaptive state- approximation feature of fuzzy systems [10]. In these control
feedback control strategy for unknown nonlinear multivariable strategies, the stability study associated to the closed-loop
systems whose the input-gains matrix is not necessarily symmetric system is established using a Lyapunov’s approach. To handle
and is characterized by nonzero leading principle minors. A lin-
early parameterized fuzzy system is used to appropriately model the fuzzy approximation (reconstruction) errors and eventual
the uncertainties. When designing our control scheme and study- disturbances, the main controller (which is the fuzzy control
ing the stability analysis, a decomposition property of the input- term) is generally improved by adding a robust control term
gain matrix is employed. A proportional-integral adaptation law to the latter, namely a sliding mode control term and/or H∞
is suggested to enhance the adaptive parameter convergence. An control term [1]–[7], [9]. In [8], this issue has been differently
appropriate Lyapunov function is exploited to study the stability
of the corresponding closed-loop control system as well as to derive solved by adding to the integral update laws a proportional
the adaptation laws. Numerical simulations and a detailed com- term derived from the Lyapunov stability theory. In spite of this
parison study are given to evaluate the efficiency of the proposed important extension and improvement, the control scheme in
control methodology. [8] presents unfortunately some drawbacks.
Index Terms—Adaptation proportional-integral (PI) law, fuzzy 1) In the adaptive laws, the proportional term has been de-
adaptive control, matrix SDU decomposition. signed as a nonderivable function, while it should be
derivable.
2) The presence of the well-known algebraic-loop problem
I. INTRODUCTION in these updates laws.
HE MAJORITY of systems in control engineering is nat- Although the control schemes in [2]–[6] can give, in general,
T urally multivariable and displays an uncertain nonlinear
behavior resulting unfortunately to complex control issues. We
good performances, they are unfortunately limited to MIMO
systems, which are characterized by an input-gains matrix being
are convinced that the developed control theory for multivari- positive definite (PD) or negative definite (ND). It is worth
able (linear or nonlinear) systems will have an immediate appli- noting that this assumption is paramount in the control design
cation in many practical systems, space technology, chemical as well as in the stability study, thereby it limits the scope of
processes, robotics, and so on. Due to the existence of a natural applications of these fundamental results. How to extend these
coupling between the plant inputs and outputs, the control prob- results to MIMO systems with an input-gains matrix which is
lem of multivariable nonlinear plants is also far from an easy not PD or ND is an important issue. Motivated by the early
task [1]. works in [11]–[13], this problem has been solved in [8], by
Over the past three decades, many adaptive fuzzy con- exploiting a so-called matrix SDU-factorization lemma. The
trol strategies have been developed for uncertain nonlinear latter allows to factorize the input-gains matrix into a product of
multivariable systems [1]–[9], by exploiting the universal three useful matrices, namely, a PD symmetric matrix (denoted
by S), a diagonal matrix (denoted by D) having the values +1 or
−1 on its main diagonal. Next, the matrix S can be incorporated
Manuscript received June 24, 2017; revised March 29, 2018 and July 28, 2018;
accepted November 12, 2018. Date of publication November 29, 2018; date of in the Lyapunov’s function candidate, as to the matrix U, it
current version August 30, 2019. This work was supported in part by the National allows us to sequentially design the control signals by avoiding
Natural Science Foundation of China under Grant 61622302, in part by the algebraic-loop problem.
Department of Education of Guangdong Province under Grant 2016KTSCX030
and Grant 2017KZDXM027, in part by the Science and Technology Planning Motivated by the aforementioned discussions, the purpose of
Project of Guangdong Province under Grant 2017B010116006, and in part this paper is to design a new fuzzy adaptive state-feedback con-
by the Innovative Research Team Program of Guangdong Province Science troller for multivariable nonlinear systems with uncertainties
Foundation under Grant 2018B030312006. (Corresponding author: Abdesselem
Boulkroune.) and eventual dynamical external disturbances. The so-called
A. Boulkroune and L. Merazka are with the LAJ, University of Jijel, Jijel matrix SDU factorization is used when designing the control
18000, Algeria (e-mail:,boulkroune@2002@yahoo.fr; lou_meraz@yahoo.fr). scheme. A new proportional-integral (PI) update law is pro-
H. Li is with the Guangdong Province Key Laboratory of Intelligent Deci-
sion and Cooperative Control, Guangzhou 510006, China, and also with the posed to estimate the unknown fuzzy parameters. It is theoret-
College of Engineering, Bohai University, Jinzhou 121013, China (e-mail:, ically demonstrated that our controller can ensure the tracking
lihongyi2009@gmail.com). errors’ convergence asymptotically to the origin. Compared to
Color versions of one or more of the figures in this article are available online
at http://ieeexplore.ieee.org. the available results [8], our contributions can be summarized
Digital Object Identifier 10.1109/TFUZZ.2018.2883369 as follows.

1063-6706 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
1704 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 27, NO. 9, SEPTEMBER 2019

1) By designing a low-pass filter, new PI update laws are Let us define the tracking error of the system’s output,
introduced, in which the proportional part is designed as a e1 (t) = [e11 , e12 , . . . , e1m ]T ∈ Rm , as
derivable function and the famous algebraic-loop problem
e1 = yd − y (4)
is avoided.
2) A detailed comparison study is conducted to investigate In addition, we introduce the following error variables ei ∈
the influence of each term in both the main control law Rm , i = 2, . . . , n, [13]:
and the adaptation law.
e2 = ė1 + e1

II. PRELIMINARIES AND SYSTEM DESCRIPTION e3 = ė2 + e2 + e1


Consider the uncertain multivariable nonlinear systems ..
.
y (n ) = F (x) + G (x) u + P (x, t) (1) en = ėn −1 + en −1 + en −2 . (5)
(n −1) T T
where x = [y T ẏ T . . . (y ) ] ∈ Rr denotes the state vec- From (5), the errors, ei , i = 2, . . . , n, can be simply ex-
tor with r = m ∗ n, u(t) ∈ Rm is the input, y(t) ∈ Rm is the pressed as [13] follows:
output. The functions F (x) ∈ Rm and G(x) = [gij ] ∈ Rm ×m

i−1
are smooth and uncertain, and P (x, t) ∈ Rm is an unknown ei = aij e1
(j )
∀ i = 2, 3, . . . , n. (6)
disturbance. j =0
Our goal is to design a state-feedback control scheme that en-
Where the constants aij can be determined via the well-known
sures the boundedness of all variables in the closed-loop system
Fibonacci number series [12].
and guarantees the output’s tracking to the desired trajectory
Define the error E as follows:
yd (t) ∈ Rm , in spite of the presence of uncertainties in the
model and eventual disturbances. E = en + en −1 . (7)
The next assumptions will be helpful to design our control
From (6) and (7), the time-derivative of E can be expressed
system.
as follows:
Assumption 1: The vector of the desired trajectories,
n −1

(n −1) T
xd = [ydT ẏdT . . . (yd ) ]T ∈ Rm ∗n , is continuous and Ė = ėn + ėn −1 = an j e1
(j +1)
+ ėn −1
bounded. j =0
Assumption 2:
n −2

a) The signs of the leading principal minors of input-gains (j +1) (n )
= an j e1 + e1 + ėn −1
matrix G(x), are known.
(n −1) j =0
b) ∂gij (x) /∂yj = 0 ∀ i = 1, . . ., m and j = 1,
n −2

. . ., m. (j +1) (n )
Assumption 3: The disturbance vector, P (x, t) ∈ Rm , is = an j e1 + yd − y (n ) + ėn −1 . (8)
j =0
bounded and uncertain.
Remark 1: Assumptions 1 and 3 are not restrictive. They are Multiplying (8) by S1 and using (3), we get the following:
standard assumptions for adaptive control theory [1], [5], ]10]. ⎛ ⎞
n −2

Remark 2: Assumption 2 can be considered as a condition S1 Ė = S1 ⎝
(j +1)
an j e1 + yd + ėn −1 ⎠ − S1 y (n )
(n )
of controllability and is not restrictive as it is fulfilled by numer- j =0
ous practical systems: e.g., robotic systems, chaotic systems, ⎛ ⎞
mechanical systems, electrical machines, and so on. n −2
Lemma 1: (SDU-factorization): Any matrix G(x) ∈ Rm ×m = S1 ⎝ + yd + ėn −1 ⎠
(j +1) (n )
an j e1
having nonzero leading-principle minors can be decomposed as j =0
a product of three matrices [7], [8]
− H (x) − DT (x) u
G (x) = S (x) DT (x) (2) ⎛ ⎞
n −2
= S1 ⎝ + yd + ėn −1 ⎠
(j +1) (n )
with S(x) ∈ Rm ×m is a PD symmetric matrix, D ∈ Rm ×m is a an j e1
diagonal matrix, and T (x) ∈ Rm ×m is a unity upper triangular j =0
matrix. The diagonal elements of the matrix D (which are +1
− H(x) − Du − D (T (x) − Im ) u. (9)
or −1) represent the ratios of the signs of the leading-principal
minors of G(x). We can also write (9) as follows:
By using Lemma 1, we can rewrite the system (1) as follows: ⎛ ⎞
n −2

D−1 S1 D D−1 Ė = D−1 S1 ⎝ + yd + ėn −1 ⎠
(j +1) (n )
S1 (x) y (n ) = H (x) + DT (x) u (3) an j e1
j =0
where S1 (x) = S −1 (x) is PD and symmetric,
H(x) = S −1 (x)(F (x) + P (x, t)) ∈ Rm . − D−1 H (x) − (T (x) − Im ) u − u
BOULKROUNE et al.: FUZZY ADAPTIVE STATE-FEEDBACK CONTROL SCHEME OF UNCERTAIN NONLINEAR MULTIVARIABLE SYSTEMS 1705

or equivalently as
⎛ ⎞
n −2

S̄1 Ē˙ = D−1 S1 ⎝ + ėn −1 ⎠
(j +1) (n )
an j e1 + yd
j =0

− D−1 H (x) − (T (x) − Im ) u − u (10)


where S̄1 = D−1 S1 D and Ē = [Ē1 , . . . , Ēm ]T = D−1 E.
Finally, by posing
α (z) = [α1 (z1 ) , . . . , αm (zm )]T
⎛ ⎞
n −2

= D−1 S1 ⎝yd + + ėn −1 ⎠
(n ) (j +1) Fig. 1. Basic configuration of a FLS.
an j e1
j =0

1
− D−1 H (x) + S̄˙ 1 Ē − (T (x) − Im ) u + Den −1 . where Ai1 , Ai2 , . . . , and Ain are fuzzy sets and f i is the fuzzy
2
(11) singleton for the ith rule. The use of the singleton fuzzifier, prod-
uct inference, and center-average defuzzifier allows to express
Equation (10) becomes simply the fuzzy system output as follows:
1 M i n 
S̄1 Ē˙ + S̄˙ 1 Ē = α (z) − u − Den −1 (12) f μ
j =1 A ji (x j )
2 i=1
 = θT ψ (x)
fˆ (x) = (14)
M n
where z = [z1T , z2T , . . . , zm
T T
] . Due to the special structure of i=1 μ
j =1 A ji (x j )
αi (zi ), one can select the argument vector zi as follows [7], [8]
 T where μA ij (xj ) represent the degree of membership of xj to Aij ,
z1 = xT , u2 , . . . , um
M is the number of fuzzy rules, θT = [f 1 , f 2 , . . . , f M ] stands
 T for the adaptable parameter vector (the consequent parameters),
z1 = xT , u2 , . . . , um
and ψ T = [ψ 1 , ψ 2 , . . . , ψ M ] where
 T
z2 = xT , u3 , . . . , um 
n
j =1 μ A i (xj )

j
.. ψ i (x) = (15)
. M n
μ i (x )
i=1 j =1 A j j
 T
zm −1 = xT , um is the fuzzy basis function (FBF).
zm = x . T
(13) It is worth noting that the FLS (14), which is broken down
into the product of two main terms θ and ψ(x), allows us to
Note that this helpful triangular structure of zi permits us to show well the concept of fuzzy adaptive control and helps us
sequentially design control inputs ui , ∀ i = 1, 2, . . . , m, with- in the control design and stability analysis. Via an update law
out any algebraic loop problem. usually determined based on the Lyapunov stability theorem,
Now, let us denote the following compact sets all the elements of the vector θ are prone to be changed in
 T
order to ameliorate the fuzzy approximation, [10], [17]–[19].
Ωz i = xT , ui+1 , . . . , um |x ∈ Ωx ⊂ Rr
However, the FBF vector ψ(x) should be adequately specified
∀ i = 1, . . . , m − 1 by the designer.

Ωz m = {x |x ∈ Ωx ⊂ Rm } . IV. ADAPTIVE FUZZY CONTROL DESIGN


In what follows, adaptive fuzzy-logic systems will be The uncertain function αi (zi ) may be expressed in terms of
employed to model the uncertain function αi (zi ). the fuzzy system (14) as follows:

III. FUZZY APPROXIMATION α̂i (zi , θi ) = θiT ψi (zi ) , i = 1, . . . , m (16)

The main structure of a fuzzy logic system (FLS), as presented where ψi (zi ) is the FBF vector. The latter should be properly
in Fig. 1, consists in a fuzzifier, certain fuzzy IF–THEN rules, a determined by the designer. And θi is the vector of the updated
fuzzy inference engine, and a defuzzifier [10], [14]–[16]. online parameters.
A mapping from an input vector xT = [x1 , x2 , . . . , xn ] ∈ Rn Now, denote the parametric estimation error as follows:
to an output fˆ ∈ R is performed by the fuzzy inference engine θ̃i = θi − θi∗ with i = 1, . . . , m (17)
using a set of fuzzy IF–THEN [1]. The ith fuzzy rule can be
expressed as and the fuzzy reconstruction error as [10], [20]–[25]
R(i) : IF x1 is Ai1 and . . . and xn is Ain THEN fˆ is f i εi (zi ) = αi (zi ) − α̂i (zi , θi∗ ) (18)
1706 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 27, NO. 9, SEPTEMBER 2019

with α̂i (zi , θi∗ ) = θi∗ ψi (zi ). tracking errors. In fact, the latter introduces, in the time-
According to the universal approximation theorem [10], we derivative of the Lyapunov function, a significant negative term
have [see (26)].
|εi (zi )| ≤ ε̄i , ∀ zi ∈ Ωz i , Remark 5:
1) The adaptation law mechanism (22) is also equipped by
where ε̄i is an uncertain constant. Also, let us denote an useful robust term, namely: an e-modification term,
α (z) = θT ψ (z) = [α1 (z1 , θ1 ) , . . . , αm (zm , θm )]T σi γ1i |Ēi |θi , which guarantees the uniform boundedness
of the adaptive parameters.
ε (z) = [ε1 (z1 ) , . . . , εm (zm )]T , ε̄ = [ε̄1 , . . . , ε̄m ]T . 2) Contrary to [8], because the proportional term (−γ2i δf i )
From the above considerations, we can get is obtained via the low-pass filter (23), the algebraic-loop
problem and the time-differentiability requirement of the
α̂ (zi , θ) − α (zi ) = α̂ (zi , θi ) − α̂ (zi , θi∗) + α̂ (zi , θi∗ ) − α (zi) proportional term are indeed solved in the modified adap-
= α̂ (zi , θi ) − α̂ (zi , θi∗ ) − εi (zi ) tive law (22). Recall that the PI adaptive law in [8] uses
only the term δi = −(−σi |Ēi |θi + Ēi ψi (zi )) as a propor-
= θ̃T ψ (zi ) − εi (zi ) (19)
tional term.
with θ̃T ψ(z) = [θ̃1T ψ1 (z1 ), . . . , θ̃m
T
ψm (zm )]T , and By considering (21), the dynamics (20) become

θ̃i = θi − θi where i = 1, . . . , m. 1  m
Using (17), (18) and (19), (12) can be expressed as follows: Ē T S̄1 Ē˙ + Ē T S̄˙ 1 Ē = − Ēi θ̃iT ψi (zi )
2 i=1
1 m m
 t 
Ē T S̄1 Ē˙ + Ē T S̄˙ 1 Ē = − Ēi θ̃iT ψi (zi ) + Ēi εi (zi )
2 − Ē T Ki Ēdt − Ē T Kp Ē
i=1 i=1 o

m

m
   m
+ Ēi θiT ψi (zi) − Ē T u − Ē T Den −1 . −  
kr i Ēi + Ēi εi (zi )
i=1 i=1 i=1
(20)
− Ē T Den −1 . (24)
The control law u can be determined based on (20) as follows:
 t Theorem: Let’s consider the system (1) subject to Assump-
 
u = θT ψ (z) + Kr sign Ē + Kp Ē + Ki Ēdτ. (21) tions 1–3, and with the control law (21) and its update law given
o by (22) and (23), one can ensures the following properties.
With Ki = diag{ki1 }, Kp = diag{kpi }, and Kr = diag{kr i }. 1) The boundedness of all signals involved in the closed-loop
Ki , Kp , and Kr ∈ Rm ×m are PD design matrices. control system.
Remark 3: Note that the control input (21) is essentially 2) The underlying tracking errors, e1 , asymptotically vanish
comprised of four terms, namely: at the origin.
1) a fuzzy term, θiT ψi (zi ), employed to model the uncertain Proof: Consider the following candidate quadratic
functions αi (zi ); Lyapunov function:

1  2
2) a robust control term, Kr sign(Ē), designed to deal with n −1
1 T
m
the uncertain term σ2i θ∗ 2 + ε̄i ;
1  
V = ei ei + Ē T S1 Ē + θ̃i + γ2i δf i 
3) a linear control term, Kp Ē, introduced for stability pur- 2 i=1 2 2γ1i i=1
poses, and finally m  t T  t 
t 2 1
4) an integral control term, Ki o Ēdτ , added to compensate + δf i  + Ēdτ Ki Ēdτ . (25)
2 o o
for the steady-state error. i=1
The PI update laws associated to (21) are given by the By considering (22)–(24), the derivative of V is
following:
n −1
 
m
 
θ̇i + γ2i δ̇f i = −σi γ1i Ēi  θi + γ1i Ēi ψi (zi ) (22) V̇ ≤ − eTi ei − Ē T Kp Ē − γ2i δf i 2
i=1 i=1
δ̇f i = −γ2i δf i + γ2i δi (23)

m
   m
  
m
 
where δi = −(−σi |Ēi |θi + Ēi ψi (zi )), with γ1i , γ2i , σi > 0 − kr i Ēi  − σi Ēi  θ̃iT θi + Ēi  ε̄i . (26)
are design parameters. i=1 i=1 i=1

Remark 4: The adaptation law mechanism (22) can be also Using the inequality
written as follows:    T  σi    
 2  σi   ∗ 2
m m m
 t  
    − σi Ēi θ̃i θi ≤ − Ēi θ̃i  + Ēi θi 
θ = θI + θ P = −σi γ1i Ēi  θi + γ1i Ēi ψi (zi ) dτ − γ2i δf i i=1 i=1
2 i=1
2
0 (27)
t
where θI = 0 (−σi γ1i |Ēi |θi + γ1i Ēi ψi (zi ))dτ is an adapta- and if we select kr∗i ≥ σ2i θ∗i 2 + ε̄i , (26) becomes
tion integral term, and θP = −γ2i δf i is an adaptation propor- n −1

tional term. Note that the added proportional term (−γ2i δf i ) V̇ ≤ − eTi ei − Ē T Kp Ē. (28)
allows a rapid convergence to adaptive fuzzy parameters and i=1
BOULKROUNE et al.: FUZZY ADAPTIVE STATE-FEEDBACK CONTROL SCHEME OF UNCERTAIN NONLINEAR MULTIVARIABLE SYSTEMS 1707

TABLE I
COMPARISON BETWEEN THE CONTROL SCHEME OF THIS PAPER AND THAT OF [8]

The system (29) can be transformed to (1) as follows:

ÿ = F1 (x) + G1 (x) u + P1 (x, t) (30)

with
 −1   
1 + q22 0 q2 q̇2 0 q̇1
F1 (x) = −
0 1 −q̇1 q2 1 q̇2
 −1  
1 + q22 0 (1 + q2 ) cos(q1 )

0 1 sin(q1 )
 −1
1 + q22 0
G1 (x) =
0 1

Fig. 2. 2-DOF polar manipulator robot. where y = [x1 , x2 ]T = [q1 , q2 ]T , x = [x1 , x2 , x3 , x4 ]T =


[q1 , q2 , q̇1 , q̇2 ]T , u = [u1 , u2 ]T is the system input, and
Then, from (28), we can easily conclude about the bound- P1 (x, t) = [p11 (x, t) , p12 (x, t)]T is an added disturbance.
edness of all signals in the closed-loop control system, includ- It is assumed that the model (30) includes the dynamics of
ing the control input u. By the Barbalat’s lemma [26], we both drives and the desired trajectories are given by the follow-
can establish that the errors e1 , Ēj ( for j = 1, . . . , m, and ing: qd1 = 0.5 sin(2t) + 0.5 and qd2 = 0.2 sin(4t) + 1, [27].
ei (for i = 2, . . . , n), can vanish asymptotically at the origin. For the proposed controller (21), two adaptive fuzzy systems,
Remark 6: To highlight the theoretical contributions of this namely θ1 T ψ1 (z1 ) and θ2 T ψ2 (z2 ), have been constructed,
paper, a comparison between this work and that of Boulkroune where z1 = [xT , u2 ]T and z2 = x. One defines for each input
et al. in [8] is given in Table I. variable, as in [28], three membership functions (one triangular
and two trapezoidal) being uniformly distributed on the fol-
lowing selected intervals: [−2,2] for q̇2 and q̇1 ; [0.5,2] for q2 ;
V. SIMULATION RESULTS AND COMPARATIVE STUDY
[−0.5,1.5] for q1 ; and [−10,10] for u2 . Thus, the number of the
In this section, one will give some simulation examples to fuzzy rules (for both fuzzy system) is, respectively, 35 and 34 . It
show the efficiency of the above theoretical results and will is worth noting that the fuzzy rule bases are full bases. Therefore,
conduct a detailed comparison to investigate the influence of the sizes of the FBF vectors (ψ1 (z1 ) and ψ2 ) are 243 × 1 and
each term in both main control law and adaptation law. 81 × 1, respectively. The consequent parameters are determined
through the PI adaptive laws (22) and (23), and they are initial-
A. Simulation Examples ized as follows: θ1i (0) = 0, for i = 1, . . . , 243, θ2j (0) = 0,
for j = 1, . . . ,81.
Example 1: Consider a 2-degree of freedom (DOF) po-
The control parameters are chosen as follows:
lar manipulator (see Fig. 2). Its mathematical model is given
γ11 = γ12 = 100, γ21 = γ22 = 35, σ1 = σ2 = 0.1, λ1 = λ2 =
by [27]:
      2, kr 1 = kr 2 = 0.2, ki1 = ki2 = 0.2, kp1 = kp2 = 0.2. The ini-
1 + q22 0 q̈1 q2 q̇2 0 q̇1 tial conditions are chosen as follows: x(0) = [0, 0.5, 0, 0]T .
+
0 1 q̈2 −q̇1 q2 1 q̇2 The external disturbances p11 (x, t) and p12 (x, t) are square
    signals with an amplitude of ∓1 and a frequency of 1/2π Hz.
(1 + q2 ) cos (q1 ) u1 The simulation results obtained are presented in Fig. 3. It
+ = . (29)
sin (q1 ) u2 can be seen from this figure that the system effectively tracks
1708 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 27, NO. 9, SEPTEMBER 2019

Fig. 3. Simulation results for a 2-DOF polar manipulator robot (example 1): (a) Tracking of q1 . (b) Tracking of q2 . (c) Tracking of q̇1 . (d) Tracking of q̇2 . (e)
Tracking errors e1 1 and ė1 1 . (f) Tracking errors e1 2 and ė1 2 . (g) Norm of fuzzy parameters: θ1  and θ2 . (h) Control signals: u1 and u2 .

where M11 = a1 + 2a3 cos(q2 ) + 2a4 sin(q2 ), M22 = a2 ,


M21 = M12 = a2 + a3 cos(q2 ) + a4 sin(q2 ), and h = a3 sin
(q2 ) − a4 cos(q2 ), with a1 = I1 + m1 lc1 2
+ Ie + me lce 2
+
2 2
me l1 , a2 = Ie + me lce , a3 = me l1 lce cos(δe ), and
a4 = me l1 lce sin(δe ).
The robot’s parameter values used are as follows:
m1 = 1, me = 2, l1 = 1, lc1 = 0.5, lce = 0.6, I1 = 0.12,
Ie = 0.25, and δe = 30.
Let y = [q1 , q2 ]T , u = [u1 , u2 ]T , x = [q1 , q2 , q̇1 , q̇2 ]T .
System (31) can be rewritten in the form of system (1) as
follows:
ÿ = F2 (x) + G2 (x) u + P2 (x, t) (32)
Fig. 4. Two links manipulator robot. where
    
f1 (x) −hq̇2 −h (q̇1 + q̇2 ) q̇1
the desired reference signals and also the control signals are F2 (x) = = −M −1
f2 (x) hq̇1 0 q̇2
bounded and admissible.    
Example 2: Consider the dynamical equations of a two-link g11 g12 −1 M11 M12
G2 (x) = =M =
rigid manipulator robot moving on a horizontal plane shown in g21 g22 M21 M22
Fig. 4, this MIMO system is given by [8]
and P2 (x, t) = [p21 (x, t) p22 (x, t)]T .
   −1
q̈1 M11 M12 The desired trajectories are yd1 = qd1 = sin(t) and
= yd2 = qd2 = sin(t).
q̈2 M21 M22
      For the proposed controller (21), two adaptive fuzzy sys-
u1 −hq̇2 −h (q̇1 + q̇2) q̇1 tems, namely θ1 T ψ1 (z1 ) and θ2 T ψ2 (z2 ), have been con-
× − (31)
u2 hq̇1 0 q̇2 structed, where z1 = [xT , u2 ]T and z2 = x. One defines for each
BOULKROUNE et al.: FUZZY ADAPTIVE STATE-FEEDBACK CONTROL SCHEME OF UNCERTAIN NONLINEAR MULTIVARIABLE SYSTEMS 1709

Fig. 5. Simulation results for a two links manipulator robot (example 2): (a) Tracking of q1 . (b) Tracking of q2 . (c) Tracking of q̇1 . (d) Tracking of q̇2 .
(e) Tracking errors e1 1 and ė1 1 . (f) Tracking errors e1 2 and ė1 2 . (g) Norm of fuzzy parameters: θ1 and θ2 . (h) Control signals: u1 and u2 .

input variable, as in [28], three membership functions (one tri-


angular and two trapezoidal) being uniformly distributed on the
following selected intervals: [−2,2] for q1 and q̇1 ; [−2,2] for
q2 , q̇2 ; and [−10,10] for u2 . Thus, the number of the fuzzy
rules (for both fuzzy system) is, respectively, 35 and 34 . It is
worth noting that the fuzzy rule bases are full bases. Therefore,
the sizes of the FBFs vector (ψ1 (z1 ) and ψ2 (z2 )) are 243 × 1
and 81 × 1, respectively. The consequent parameters are deter-
mined through the PI adaptive laws (22) and (23), and they are
initialized by θ1i (0) = 0, for i = 1, . . . , 243, θ2j (0) = 0, for
j = 1, . . . ,81.
The controller parameters are selected as follows:
γ11 = γ12 = 100, γ21 = γ22 = 100, σ1 = σ2 = 0.005,
λ1 = λ2 = 2, kp1 = kp2 = 0.2, ki1 = ki2 = 2, kr 1 = kr 2 =
0.1. The initial conditions are given as x(0) = [1, 1, 0, 0]T .
The external disturbances p21 (x, t) and p22 (x, t) are square
signals with an amplitude of ∓1 and a frequency of 1/2π Hz. Fig. 6. 2-DOF helicopter (CE150).
Simulation results of this example are displayed in Fig. 5.
Despite the presence of disturbances and uncertainties, it is clear
from this figure, that this MIMO system successfully follows with
the desired trajectories. The control signals associated are also
bounded and admissible.    
cos(φIl2 )2 0 0
Example 3: Let us consider in this example a 2-DOF heli- M (φ, ϑ) = , G(φ, ϑ) =
0 Il2 mgIc cosϑ
copter (CE150) shown in Fig. 6. The dynamics of this MIMO  
nonlinear system are given by [29], [30] the following: −cosϑ sinϑϑ̇Il2 −cosϑ sinϑφ̇Il2
C(φ, ϑ, φ̇, ϑ̇) =
cosϑ sinϑφ̇Il2 0
   φ̇  
φ̈ u1
M (φ, ϑ) + C φ, ϑ, φ̇, ϑ̇ + G (φ, ϑ) = u (33) and u= .
ϑ̈ ϑ̇ u2
1710 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 27, NO. 9, SEPTEMBER 2019

Fig. 7. Simulation results for a 2-DOF helicopter (example 3): (a) Tracking of φ. (b) Tracking of ϑ. (c) Tracking of φ̇. (d) Tracking of ϑ̇. (e) Tracking errors e1 1
and ė1 1 . (f) Tracking errors e1 2 and ė1 2 . (g) Norm of fuzzy parameters: θ1  and θ2 . (h) Control signals: u1 and u2 .

TABLE II TABLE III


EFFECT OF EACH TERM IN THE CONTROL LAW (35) EFFECT OF THE PI TERMS IN THE ADAPTATION LAW (36)

Where where
 
ml (l13 + l23 ) −cosϑ sinϑφ̇Il2
−cosϑ sinϑϑ̇Il2
Il2 = + m1 l12 + m2 l22 and F3 (x) = −M −1
3 (l1 + l2 ) cosϑ sinϑφ̇Il2 0
   
lc = (ml (l1 − l2 ) + m1 l1 − m2 l2 )/m. φ̇ g11 g12
× − M −1 G G3 (x) = = M −1
ϑ̇ g21 g22
The helicopter given by (33) can be expressed in the following
form: and P3 (x, t) = [p31 (x, t) , p32 (x, t)]T .
The desired trajectories are yd1 = φd = sin(t) and
ÿ = F3 (x) + G3 (x) u + P3 (x, t) (34) yd2 = ϑd = sin(t).
BOULKROUNE et al.: FUZZY ADAPTIVE STATE-FEEDBACK CONTROL SCHEME OF UNCERTAIN NONLINEAR MULTIVARIABLE SYSTEMS 1711

Fig. 8. Comparative study 2: Norm of fuzzy parameters: θ1  and θ2 , obtained by using the (a) Proposed PI adaptation law. (b) E-modification adaptive law.
(c) σ-modification adaptive law. (d) Standard adaptation law.

t
For the proposed controller (21), two adaptive fuzzy systems, where uPI = Kp Ē + Ki o Ēdτ is a PI control term,
namely θ1 T ψ1 (z1 ) and θ2 T ψ2 (z2 ), have been constructed, ur = Kr sign(Ē) is a robust control term, and uf = θT ψ(z)
where z1 = [xT , u2 ]T and z2 = x. One defines for each input is a fuzzy control term.
variable, as in [28], three membership functions (one triangular Second, one investigates the influence of both main terms in
and two trapezoidal) being uniformly distributed on the fol- the adaptation laws (22) (i.e., the adaptation integral term and
lowing selected intervals: [−2, 2] for φ and φ̇; [−2, 2] for the adaptation proportional term). From Remark 4, one has
ϑ, ϑ̇; and [−10, 10] for u2 . Thus, the number of the fuzzy
rules (for both fuzzy system) is, respectively, 35 and 34 . The θ = θI + θP (36)
consequent parameters are determined through the PI adaptive t
laws (22) and (23), and they are initialized by θ1i (0) = 0, for where θI = 0 (−σi γ1i |Ēi |θi + γ1i Ēi ψi (zi ))dτ is an adapta-
i = 1, . . . , 243, θ2j (0) = 0, for j = 1, . . . ,81. tion integral term, and θP = −γ2i δf i is an adaptation propor-
The helicopter’s parameters are: m1 = 0.42(kg), tional term.
m2 = 0.16(kg), ml = 0.35(kg), m = ml + m1 + m2 , Moreover, to quantify the tracking performances in all tests,
l1 = 0.198(m), l2 = 0.174(m), g = 9.8. the following performance index (the integral of absolute value
The controller parameters used are selected as follows: of the tracking error) is used
γ11 = γ12 = 100, γ21 = γ22 = 100, σ1 = σ2 = 0.005, λ1 =  T
λ2 = 2, kp1 = kp2 = 0.2, ki1 = ki2 = 2, kr 1 = kr 2 = 0.1. IIAE = |ydi − yi | dt. (37)
The initial conditions are chosen as: [0.5, −0.5, 0, 0]T . The o
external disturbances, p31 (x, t) and p32 (x, t) are square signals
The obtained results are given in Tables II and III. From
with an amplitude of ∓1 and a frequency of 1/2π Hz.
Table II, in the third test (i.e., when the robust term is added to
In this third example, the simulation results are presented in
the control law), we can clearly observe a notable improvement
Fig. 7. These results show good tracking as well as the efficiency
in the tracking performances.
of our proposed control strategy.
From Table III, the superiority of the PI adaptation law over
other adaptation laws (i.e., in the test 6) is clearly shown in these
B. Comparative Study 1 results.
In this section, one will conduct a series of tests to evaluate
the performances and the contribution of each term in the main C. Comparative Study 2
control law and in the adaptive law by applying the developed
control law in different situations on our robotic system (29). In this section, one will conduct a test of our PI update law
First, one evaluates the effect of each term in the main control (22) and (23), with three well-known update laws, namely σ −
laws (21). The latter can be expressed as follows: modification [31], e − modification [32], and standard adaptive
law [33]. For that, for all update laws, the control law given by
u = uPI + ur + uf (35) (21) is applied to the system (29).
1712 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 27, NO. 9, SEPTEMBER 2019

Fig. 9. Comparative study 2: The underlying tracking errors e1 and e2 , obtained by using the (a) Proposed PI adaptation law. (b) E-modification adaptive law.
(c) σ-modification adaptive law. (d) Standard adaptation law.

Simulation results for all update laws are displayed in REFERENCES


Figs. 8 and 9. From Fig. 8, the PI adaptation law proposed [1] A. Boulkroune, M. M’Saad, and H. Chekireb, “Design of a fuzzy adaptive
in this paper allows for a rapid convergence of adaptive fuzzy controller for MIMO nonlinear time-delay systems with unknown actuator
parameters and the final values of the steady state are well ac- nonlinearities and unknown control direction,” Inf. Sci., vol. 180, no. 24,
pp. 5041–5059, 2010.
ceptable. However, the behavior of fuzzy parameters tuned with [2] S. C. Tong, J. Tang, and T. Wang, “Fuzzy adaptive control of multivariable
σ − modification is well similar to that of the of fuzzy parame- nonlinear systems,” Fuzzy Sets Syst., vol. 111, no. 2, pp. 153–167, 2000.
ters adjusted only with e − modification. For these two update [3] S. C. Tong and H. X. Li, “Fuzzy adaptive sliding model control for MIMO
nonlinear systems,” IEEE Trans. Fuzzy Syst., vol. 11, no. 3, pp. 354–360,
laws, the fuzzy parameters are bounded and their final values Jun. 2003.
are quite acceptable. However, for the standard adaptation law, [4] S. C. Tong, B. Chen, and Y. Wang, “Fuzzy adaptive output feedback
we can remark that the adaptive parameters tend to diverge, due control for MIMO nonlinear systems,” Fuzzy Sets Syst., vol. 156, no. 2,
pp. 285–299, 2005.
to the pure integral action. [5] S. Labiod, M. S. Boucherit, and T. M. Guerra, “Adaptive fuzzy control
Also, Fig. 9, which shows the underlying tracking errors ob- of a class of MIMO nonlinear systems,” Fuzzy sets Syst., vol. 151, no. 1,
tained by applying the four update laws, confirms well the su- pp. 59–77, 2005.
[6] N. Essounbouli, A. Hamzaoui, and J. Zaytoon, “An improved robust adap-
periority of our PI update laws. The standard adaptation law tive fuzzy controller for MIMO systems,” Control Intell. Syst., vol. 34,
also gives a satisfactory tracking performance, but at the cost no. 1, pp. 12–21, 2006.
of a great value for fuzzy parameters (i.e., an important control [7] A. Boulkroune, M. M’Saad, M. Tadjine, and M. Farza, “Adaptive fuzzy
control of MIMO nonlinear systems with unknown hysteresis and control
effort). gain matrix sign,” in Proc. 16th Mediterranean Conf. Control Automat.,
2008, pp. 380–385.
[8] A. Boulkroune, M. Tadjine, M. M’Saad, and M. Farza, “Fuzzy adaptive
controller for MIMO nonlinear systems with known and unknown control
VI. CONCLUSION direction,” Fuzzy Sets Syst., vol. 161, no. 6, pp. 797–820, 2010.
[9] A. Boulkroune, M. M’Saad, and M. Farza, “Fuzzy approximation-based
A fuzzy state-feedback controller for a class of uncertain mul- indirect adaptive controller for multi-input multi-output non-affine sys-
tems with unknown control direction,” IET Control Theory Appl., vol. 6,
tivariable nonlinear systems has been proposed. Adaptive fuzzy no. 17, pp. 2619–2629, 2012.
systems have been used to online model uncertain functions. [10] L. X. Wang, A Course in Fuzzy Systems. Englewood Cliffs, NJ, USA:
To enhance the convergence of the updated fuzzy parameters Prentice-Hall, 1999.
[11] A. S. Morse, “A gain matrix decomposition and some of its applications,”
as well as the tracking performances, a PI update law has been Syst. Control Lett., vol. 21, pp. 1–10, 1993.
introduced. Simulation results of three practical examples have [12] B. Xian, D. Dawson, M. deQueiroz, and J. Chen, “A continuous asymptotic
been given to emphasize the performance of the proposed con- tracking control strategy for uncertain nonlinear systems,” IEEE Trans.
Autom. Control, vol. 49, no. 7, pp. 1206–1206, Jul. 2004.
troller. A comparative study has been conducted also to evalu- [13] J. Chen, A. Behal, and D. M. Dawson, “Adaptive output feedback control
ate the effect of each terms in the main control and PI update for a class of MIMO nonlinear systems,” in Proc Amer. Control Conf.,
laws. 2006, pp. 5300–5305.
BOULKROUNE et al.: FUZZY ADAPTIVE STATE-FEEDBACK CONTROL SCHEME OF UNCERTAIN NONLINEAR MULTIVARIABLE SYSTEMS 1713

[14] E. Lughofer and S. Kindermann, “Sparse FIS: Data-driven learning of Abdesselem Boulkroune received the Engineering
fuzzy systems with sparsity constraints,” IEEE Trans. Fuzzy Syst., vol. 18, degree from Setif University, Setif, Algeria, in 1995,
no. 2, pp. 396–411, Apr. 2010. the master’s degree from the Military Polytechnic
[15] M. Luo, F. Sun, and H. Liu, “Hierarchical structured sparse representation School (EMP) of Algiers, Bordj-El-Bahri, Algeria,
for T–S fuzzy systems identification,” IEEE Trans. Fuzzy Syst., vol. 21, in 2002, and the Ph.D. degree in automatic from the
no. 6, pp. 1032–1043, Dec. 2013. National Polytechnic School (ENP) of Algiers, El
[16] M. Luo, F. Sun, and H. Liu, “Joint block structure sparse representation Harrach, Algeria, in 2009.
for multi-input–multi-output (MIMO) T–S fuzzy system identification,” In 2003, he joined the automatic control de-
IEEE Trans. Fuzzy Syst., vol. 22, no. 6, pp. 1387–1400, Dec. 2014. partment, Jijel University, Jijel, Algeria, where he
[17] W. Liu, C. C. Lim, P. Shi, and S. Xu, “Backstepping fuzzy adaptive control is currently a Full Professor. His research interests
for a class of quantized nonlinear systems,” IEEE Trans. Fuzzy Syst., include in nonlinear control, chaos synchronization,
vol. 25, no. 5, pp. 1090–1101, Oct. 2017. fractional-order systems, induction motor drives, and fuzzy control and adaptive
[18] Y. Li, S. Tong, and T. Li, “Hybrid fuzzy adaptive output feedback control control.
design for uncertain MIMO nonlinear systems with time-varying delays
and input saturation,” IEEE Trans. Fuzzy Syst., vol. 24, no. 4, pp. 841–853,
Aug. 2016.
[19] G. Lai, Z. Liu, Y. Zhang, C. P. Chen, S. Xie, and Y. Liu, “Fuzzy adaptive
inverse compensation method to tracking control of uncertain nonlinear
systems with generalized actuator dead zone,” IEEE Trans. Fuzzy Syst., Loubna Merazka received the bachelor’s degree in
vol. 25, no. 1, pp. 191–204, Feb. 2017. automatic control, in 2011, the master’s degree in
[20] L. Merazka, F. Zouari, and F. A. Boulkroune, “Fuzzy state-feedback con- automation and industrial informatics, in 2013, from
trol of uncertain nonlinear MIMO systems,” in Proc. 6th Int. Conf. Syst. the University of Jijel, Jijel, Algeria. She is currently
Control, 2017, pp. 103–108. working toward the Ph.D. degree in automatic control
[21] Y. J. Liu, J. Li, S. Tong, and C. L. Philip Chen, “Neural network with the University of Jijel. Her research interests in-
control-based adaptive learning design for nonlinear systems with full- clude nonlinear control, fuzzy control, adaptive con-
state constraints,” IEEE Trans. Neural Netw. Learn. Syst., vol. 27, no. 7, trol and observer-based control.
pp. 1562–1571, Jul. 2016.
[22] B. Xu, F. Sun, and Y. Pan, “ Disturbance observer based composite learning
fuzzy control of nonlinear systems with unknown dead zone,” IEEE Trans.
Syst., Man, Cybern.: Syst., vol. 47, no. 8, pp. 1854–1862, Aug. 2017.
[23] Y. C. Hsueh and S.-F. Su, “Learning error feedback design of di-
rect adaptive fuzzy control systems,” IEEE Trans. Fuzzy Syst., vol. 20,
no. 3, pp. 536–545, Jun. 2012.
[24] J. Fu, H. He, and X. Zhou, “Adaptive learning and control for MIMO
system based on adaptive dynamic programming,” IEEE Trans. Neural Hongyi Li (SM’17) received the Ph.D. degree in in-
Netw., vol. 22, no. 7, pp. 1133–1148, Jul. 2011. telligent control from the University of Portsmouth,
[25] Y. J. Liu, L. Tang, S. Tong, and C. L. Philip Chen, “Reinforcement learning Portsmouth, U.K., in 2012.
design-based adaptive tracking control with less learning parameters for He was a Research Associate with the Depart-
nonlinear discrete-time MIMO systems,” IEEE Trans. Neural Netw. Learn. ment of Mechanical Engineering, University of Hong
Syst., vol. 26, no. 1, pp. 165–176, Jan. 2015. Kong, Hong Kong and Hong Kong Polytechnic Uni-
[26] H. K. Khalil, Nonlinear Systems. 3rd., New Jersey, NY, USA: Prentice versity, Hong Kong. He was a Visiting Principal
Hall, 2001. Fellow with the Faculty of Engineering and Infor-
[27] A. Vitecek and A. M. Viteckova, “Control of mechanical systems with mation Sciences, University of Wollongong, Wol-
uncertainties,” in Proc. 17th Int. Carpathian Control Conf., 2016. longong, Australia. His research interests include
[28] A. Boulkroune, M. Tadjine, M. Msaad, and M. Farza, “How to design a adaptive control, fuzzy control, sliding mode control,
fuzzy adaptive control based on observers for uncertain affine nonlinear and their applications.
systems,” Fuzzy Sets Syst., vol. 159, pp. 926–948, 2008. Dr. Li was a recipient of the National Science Fund for Excellent Young
[29] Humusoft, CE 150 Helicopter Model User’s Manual, 1989. Scholars of China, in 2016, the New Century Excellent Talents in University
[30] L. Merazka and A. Boulkroune, “Adaptive fuzzy state-feedback control of Ministry of Education of China, in 2013, the Distinguished Professor of
for a class of multivariable nonlinear systems,” in Proc. 8th Int. Conf. Liaoning Province, China, in 2015, and the 2016 Andrew P. Sage Best Transac-
Model., Identification Control, 2016, pp. 925–931. tions Paper Award from IEEE System, Man, Cybernetics Society and the Best
[31] P. Ioannou and P. Kokotovic, “Instability analysis and improvement of Paper Award in Theory from ICCSS 2017. He has been in the Editorial Board
robustness of adaptive control,” Automatica, vol. 20, no. 5, pp. 583–594, of several international journals, including IEEE TRANSACTIONS ON NEURAL
Sep. 1984. NETWORKS AND LEARNING SYSTEMS, IEEE TRANSACTIONS ON SYSTEMS, MAN
[32] K. Narendra and A. Annaswamy, “A new adaptive law for robust adapta- AND CYBERNETICS: SYSTEMS, IEEE/CAA JOURNAL OF AUTOMATICA SINICA,
tion without persistent excitation,” IEEE Trans. Autom. Control, vol. 32, Asian Journal of Control, International Journal of Fuzzy Systems, International
no. 2, pp. 134–145, Feb. 1987. Journal of Control, Automation and Systems, and Circuits, Systems, and Signal
[33] H. P. Whitaker and A. Kezer, “New developments in the design of adap- Processing, etc. He has been a Guest Editor for the IEEE TRANSACTIONS ON
tive control systems,” Developments in the Design of Adaptive Control CYBERNETICS and IET Control Theory and Applications. He is a member of the
Systems, no. 61–39, 1961. IFAC Technical Committee on Computational Intelligence in Control.

You might also like