DOI: 10.2478/amcs-2013-0053
ROBUST OBSERVER DESIGN FOR SUGENO SYSTEMS WITH INCREMENTAL QUADRATIC NONLINEARITY IN THE CONSEQUENT
H ODA MOODI, M OHAMMAD FARROKHI
Department of Electrical Engineering
Iran University of Science and Technology, Narmak, Farjam St., Tehran 16846, Iran e-mail: {moodi,farrokhi}@iust.ac.ir
This paper is concerned with observer design for nonlinear systems that are modeled by T–S fuzzy systems containing parametric and nonparametric uncertainties. Unlike most Sugeno models, the proposed method contains nonlinear functions in the consequent part of the fuzzy IF-THEN rules. This will allow modeling a wider class of systems with smaller modeling errors. The consequent part of each rule contains a linear part plus a nonlinear term, which has an incremental quadratic constraint. This constraint relaxes the conservativeness introduced by other regular constraints for nonlinearities such as the Lipschitz conditions. To further reduce the conservativeness, a nonlinear injection term is added to the observer dynamics.
Simulation examples show the effectiveness of the proposed method compared with the existing techniques reported in well-established journals.
Keywords: nonlinear Sugeno model, incremental quadratic constraint, robust observer.
1. Introduction
Observer and observer-based controller design for uncertain Sugeno systems has been widely addressed by many researchers over the last decades (Tseng et al., 2009;
Yoneyama, 2009; Xu et al., 2012; Ichalal et al., 2012).
Sugeno systems are popular for their local linear form, which allows one to use powerful existing tools (e.g., Linear Matrix Inequality (LMI)) for analysis and design of these systems. When uncertainties exist in the model, the observer-based controller design for Takagi–Sugeno (T–S) systems becomes harder as the problem results in Bilinear Matrix Inequalities (BMIs) instead of LMIs.
Some researchers have tried to overcome this problem;
examples of such works are reported by Asemani and Majd (2013), Chadli and Guerra (2012), as well as Dong et al. (2010; 2011).
As the complexity of the system increases, the number of rules in the fuzzy model and hence the number and dimensions of LMIs (used for stability analysis) increase and become harder to solve. Many works in the literature are devoted to decreasing the conservativeness of these LMIs in order to apply them to a wider class of systems (Guerra et al., 2012; Abdelmalek et al., 2007; Bernal and Huˇsek, 2005). Another possible solution is to use nonlinear local subsystems for the T–S
model. Although it seems that this method increases the complexity of the fuzzy model, it decreases the number of rules and at the same time increases the model accuracy.
The key idea of using nonlinear terms in the subsystems is to employ some kind of nonlinearity, which is less complicated than the nonlinearities of the main system.
A very simple form of these nonlinear T–S models is used by Rajesh and Kaimal (2007). The authors used linear form for the consequence part plus a sinusoidal term. A more advanced work is performed by Dong et al.
(2010; 2011), who employed sector-bounded functions in the subsystems. Tanaka et al. (2009a; 2009b) proposed a T–S model with polynomial subsystems. For stability analysis, they used the Sum Of Squares (SOS) approach.
This was the first use of the SOS instead of the LMI in fuzzy systems analysis. Sala and Arino (2009) as well as Sala (2009) represented a similar form of the Sugeno model and used the Taylor series expansion of the system for construction of the polynomial subsystems.
The authors state that the nonlinear consequent part in the T–S model not only reduces the number of rules, but also reduces the conservativeness in the controller design.
In this paper, a similar form of Dong’s model is
employed. In other words, every subsystem in the
Sugeno model contains a linear plus a nonlinear term
in the consequent part of the fuzzy IF-THEN rules.
712
However, unlike in the previous works, this nonlinear term is not assumed to be Lipschitz, which is a mild condition but results in conservative designs. Instead, in this paper, the incremental Quadratic Constraint (δQC) is adopted (Ac¸ikmese and Corless, 2011). This constraint is less conservative compared with the Lipschitz condition.
Hence, it can encompass a larger class of nonlinearities.
In addition, for the first time, a nonlinear injection term is added to the fuzzy observer that provides more degrees of freedom to the design procedure. For further reduction in conservativeness, Fuzzy Lyapunov Functions (FLFs) are employed.
The FLF is one of the three classes of Lyapunov functions that are used to analyze T–S systems. The other two classes are the traditional quadratic Lyapunov functions and piecewise Lyapunov functions, which are usually more conservative than the FLF. A complete review of recent Lyapunov functions for discrete fuzzy systems is presented by Guerra et al. (2009). The FLF, as a non-quadratic Lyapunov function, has been of increasing interest in recent years. In this case, quadratic Lyapunov functions share the same membership functions with the T–S fuzzy model. For continuous-time systems, it is more difficult to obtain LMI conditions using the FLF, compared with discrete-time systems, because the stability conditions depend on the time derivative of the membership functions, which are usually handled with very conservative bounds.
Other approaches have been also investigated. Rhee and Won (2006) proposed a method which does not depend on the derivative of the membership functions.
Mozelli et al. (2009) derived LMI conditions for state feedback controller design by adding some slack matrices.
Results of Guerra and Bernal (2012), as well as Guerra et al. (2012), overcome the aforementioned deficiency by providing local asymptotic conditions at the price of computationally demanding LMIs. In this paper, the FLF is used based on the work by Faria et al. (2012).
The reminder of the paper is organized as follows. In Section 2, a nonlinear Sugeno model and an incremental quadratic constraint are introduced. Section 3 provides the problem of observer design for nonlinear T–S systems along with analytical results. Numerical examples are given in Section 4 to show effectiveness of the proposed method. Section 5 concludes the paper.
2. Problem statement
Consider the class of nonlinear systems described by
˙x(t) = f a (x(t)) + f b (x(t))ϕ
x(t), u(t), t + g(x(t))u(t),
y(t) = f ya (x(t)) + f yb (x(t))ϕ
x(t), u(t), t ,
(1)
where x(t) ∈ R n
xis the state, u(t) ∈ R n
uis the control input, y(t) ∈ R n
yis the measurable output, f n (x(t)) :
n ∈ {a, b, ya, yb} and g(x(t)) ∈ R (n
x×n
u) are nonlinear functions, and ϕ
x(t), u(t), t
∈ R (n
x×n
ϕ) is a vector of nonlinear functions.
2.1. Incremental quadratic constraint. Suppose that the following relation exists:
ϕ
x(t), u(t), t
= φ
s(t), q(t) , q(t) = C q x(t) + D q ϕ
x(t), u(t), t
, (2)
where q ∈ R n
qand C q and D q are constant matrices with proper dimensions and s(t) =
t, u(t), y(t) . For simplicity, in the rest of this paper, s(t) and q(t) are shown with s and q, respectively. Note that the term D q is included to treat systems where the nonlinear term depends on the derivative of a state variable.
Characterization of the nonlinear element φ(s, q) is based on a set of symmetric matrices M, which is referred to as incremental multiplier matrices (Ac¸ikmese and Corless, 2011). Specifically, for all M ∈ M the following incremental quadratic constraint holds:
q 2 − q 1 φ(s, q 2 ) − φ(s, q 1 )
T M
q 2 − q 1 φ(s, q 2 ) − φ(s, q 1 )
≥ 0.
(3) Defining v := C q x, we have
φ
s, v + D q ϕ
x(t), u(t), t
= ψ(s, v), ϕ
x(t), u(t), t
= ψ
s, C q x(t) .
(4)
The implicit description of ϕ arises in many situations, for instance, when the nonlinear term depends on ˙x. As an example, consider the following plant (Ac¸ikmese and Corless, 2011):
˙ x 1 = x 2 ,
˙
x 2 = 0.5 sin(x 1 + ˙x 2 ).
Letting ϕ = sin(x 1 + ˙x 2 ) yields ˙ x 1 = x 2 , ˙x 2 = 0.5ϕ.
Now, let q = x 1 + ˙x 2 to obtain ϕ = φ(q) = sin(q) with q = C q x + D q ϕ, where C q = [1, 0] and D q = 0.5.
For each v = C q x, there exists a unique solution for ϕ = sin(v + 0.5ϕ), which can be denoted by ϕ = ψ(v). In general, obtaining a δQC characterization for a nonlinearity is easier by using the function φ, rather than ψ when D q = 0 (i.e., when ϕ is implicitly defined). In some cases, the only way to obtain ψ is via numerical methods, where φ may readily be shown to satisfy δQC.
Note that ψ(s, v) satisfies the incremental quadratic constraint. That is,
δv δψ
T N
δv δψ
≥ 0, (5)
where
N =
I D q 0 I
T M
I D q 0 I
. (6)
713 2.2. Nonlinear Sugeno model. The system (1) can be
represented by a T–S fuzzy system with local nonlinear models and uncertainties as follows:
Plant Rule i :
if z 1 (t) is μ i1 (z), . . . , and z p (t) is μ ip (z) then:
˙x(t) = (A i + ΔA i )x(t) + (G xi + ΔG xi )ϕ
x(t), u(t), t + (B i + ΔB i )u(t) + D 1i ν(t), y(t) = (C i + ΔC i )x(t)
+ (G yi + ΔG yi )ϕ
x(t), u(t), t + D 2i ν(t),
(7)
where A i ∈ R (n
x×n
x) , B i ∈ R (n
x×n
u) , C i ∈ R (n
y×n
x) , G xi ∈ R (n
x×n
ϕ) , G yi ∈ R (n
y×n
ϕ) , D 1i ∈ R (n
x×n
ν) , and D 2i ∈ R (n
y×n
ν) (i = 1, . . . , r) are constant matrices, in which r is the number of rules, n x is the number of states, n u is the number of inputs, n y is the number of outputs, n ϕ is the number of nonlinear functions in the vector ϕ, and n ν is the dimension of ν. Moreover, z 1 (t), . . . , z p (t) are the premise variables, the μ ij ’s denote the fuzzy sets, and ν(t) is a band-limited white noise.
The uncertainties are defined as ΔA i := M 1i F 1 N 1 , ΔB i := M 1i F 2 N 2 , ΔG xi := M 1i F 3 N 3 , ΔC i := M 2i F 4 N 1 , ΔG yi := M 2i F 5 N 3 ,
(8)
where F i T F i < 1 (i = 1, . . . , 5), in which F i ∈ R (n
F×n
F) . In this case, the whole fuzzy system can be represented as
˙ x(t) =
r i=1
ω i (z)
(A i + ΔA i )x(t) + (B i + ΔB i )u(t) + (G xi + ΔG xi )ϕ(x(t), u(t), t) + D 1i ν(t)
, y(t) =
r i=1
ω i (z)
(C i + ΔC i )(x(t))
+ (G yi + ΔG yi )ϕ(x(t), u(t), t) + D 2i ν(t) ,
(9) where
ω i (z) = h i (z) r
k=1 h k (z) , h i (z) = Π p j=1 μ ij (z).
(10)
3. Observer design
The observer used in this paper is as follows:
Observer Rule i :
if z 1 (t) is μ i1 (z), . . . , and z p (t) is μ ip (z) then:
˙ˆx(t) = A i x(t) + G ˆ xi ϕ ˆ ˆ
x(t), u(t), t
+ B i u(t) + L i
ˆ
y(t) − y(t) , ˆ
y(t) = C i x(t) + G ˆ yi ϕ ˆ ˆ
x(t), u(t), t , ˆ
ϕ ˆ
x(t), u(t), t
= ψ
s, C q x(t) + L ˆ n [ˆ y(t) − y(t)] . (11) Unlike other Sugeno observers, here the nonlinear injection term L n [ˆ y(t) − y(t)] is used for better estimation of ϕ(x(t), u(t), t), which in turn provides better estimation for all states of the system. For the system (7) and the observer (11), the following error dynamic can be stated:
˙e(t)
=
r i=1
r j=1
ω ij (z)
(A i + L i C j )e(t) + (G xi + L i G yj )δϕ(t) −
(ΔA i + L i ΔC j )x(t) + ΔB i u(t) + (ΔG xi + L i ΔG yj )ϕ
x(t), u(t), t + (D 1i + L i D 2j )ν(t)
,
(12) where
e(t) = ˆ x(t) − x(t), δϕ(t) = ˆ ϕ
ˆ
x(t), u(t), t
− ϕ
x(t), u(t), t , ω ij (z) = ω i (z)ω j (z).
(13)
When there is no uncertainty in the model, from (11) it follows that ˆ y(t) − y(t) = r
i=1 ω i (z)
C i e(t) + G yi δϕ(t)
. Hence, by defining two variables v 1 and v 2 as v 1 := C q x(t),
v 2 := C q x(t) + L ˆ n r
i=1
ω i (z)
C i e(t) + G yi δϕ(t) , (14) and based on (5), it follows that
e(t) δϕ(t)
T Φ T M Φ
e(t) δϕ(t)
≥ 0, (15)
where Φ :=
r i=1
ω i (z)
C q + L n C i D q + L n G xi
0 I
. (16)
In order to analyze the system using LMIs, it is assumed that matrix M has the following form:
M =
X 0 0 −Y
, (17)
where X ∈ R (n
q×n
q) , Y ∈ R (n
ϕ×n
ϕ) , X=X T >0, and Y =Y T >0. Then, by defining C z := r
i=1 ω i (z)C i and
714
using the same definition for G yz , Φ T M Φ
=
(C q + L n C z ) T (D q + L n G yz ) T
X
C q + L n C z D q + L n G yz
−
0 I
Y
0 I .
(18) 3.1. Observer analysis. In this section, the conditions for the asymptotic convergence of the observer states in (11) to the system states in (7) will be given. The following lemmas are used in this paper.
Lemma 1. (Tuan et al., 2001) If
M ii < 0, 1 < i < r, 1
r − 1 M ii + 1
2 (M ij + M ji ) < 0, 1 < i = j < r, (19)
then r
i=1
r j=1
α i α j M ij < 0, (20) where 0 ≤ α i ≤ 1 and r
i=1 α i = 1.
Lemma 2. (Boyd et al., 1994) For any positive defi- nite matrix Π with appropriate dimensions, the following property holds:
X T Y + Y T X ≤ X T ΠX + Y T Π −1 Y. (21)
In the following theorem, sufficient conditions for the stability of the error dynamic (12) will be given.
Theorem 1. Assume | ˙ω i (z) | < κ i for known positive real numbers κ i , where ˙ ω i (z) is the derivative of ω i (z) with respect to time. The error dynamic (12) is asymptoti- cally stable and with an H ∞ performance bound γ > 0 if there exist matrices P 1ρ = P 1ρ T > 0, P 2ρ = P 2ρ T > 0 (1 ≤ ρ ≤ r), X, Y, X 1 = X 1 T , X 2 = X 2 T , R n , S i , S Li (1 ≤ i ≤ 6), and a scalar η > 0 such that P 1ρ , P 2ρ , X 1 , X 2 , S 1 , S 2 , S 3 , S 4 ∈ R n
x×n
x, X ∈ R n
q×n
q, Y ∈ R n
ϕ×n
ϕ, S 5 ∈ R n
ϕ×n
x, S 6 ∈ R n
u×n
x, S Li ∈ R n
x×n
y, R n ∈ R n
q×n
y, and
P 2ρ + X 2 − P 2ξ ≥ 0, ∀ρ ∈ 1, . . . , r − ξ, P 1ρ + X 1 − P 1ξ ≥ 0, ∀ρ ∈ 1, . . . , r − ξ,
Ξ ii < 0, 1 < i < r, 1
r − 1 Ξ ii + 1
2 (Ξ ij + Ξ ji ) < 0, 1 < i = j < r, (22) where ξ is an arbitrary value in 1, . . . , r and
Ξ ij =
R ij Ψ T
Ψ −X
, Ψ =
XC q + R n C i 0 1 XD q + R n G yi 0 2 , (23)
in which 0 1 and 0 2 are zero vectors with dimensions n x × (n x + 2n F ) and n x × (2n x + n ϕ + n u + 4n F ), respectively, and
R ij =
⎛
⎜ ⎝
R ij 11 ∗ ∗ .. . . .. ∗ R ij 14,1 · · · R 14,14 ij
⎞
⎟ ⎠ , (24)
where
A ∗
B C
:=
A B T
B C
and
R ij 1,1 = E 1 − S 3 A i − S Li C j − (S 3 A i + S Li C j ) T + I, R ij 2,1 = ( √
3S 3 M 1i + √
2S Li M 2j ) T , R ij 3,1 = P 1i − ηS 3 A i − ηS Li C j + S T 3 , R ij 5,1 = −G T xi S 3 T − (S Li G yj ) T ,
R ij 14,1 = −D T 1i S 3 T − (S Li D 2j ) T , R ij 2,2 = −I, R ij 3,3 = η(S 3 + S 3 T ),
R ij 4,3 = (η √
3S 3 M 1i + η √
2S Li M 2j ) T , R ij 4,4 = −I, R ij 5,3 = −ηG T xi S 3 T − η(S Li G yj ) T ,
R ij 14,3 = −ηD T 1i S 3 T − η(S Li D 2j ) T , R ij 5,5 = −Y, R ij 6,6 = E 2 − S 1 A i − A T i S 1 T + 8N 1 T N 1 ,
R ij 7,6 = √
3M 1i T S 1 T , R ij 8,6 = P 2i + S 1 T − S 2 A i , R ij 10,6 = −S 5 A i − G T xi S 1 T , R ij 7,7 = −I, R ij 12,6 = −B T i S 1 T − S 6 A i , R ij 14,6 = −D T 1i S 1 T ,
R ij 8,8 = S 2 + S 2 T , R ij 9,8 = √
3M 1i T S 2 T , R ij 9,9 = −I, R ij 10,8 = S 5 − G T xi S 2 T ,
R ij 12,8 = −B T i S 2 T + S 6 , R 14,8 ij = −D T 1i S 2 T , R ij 10,10 = −S 5 G xi − G T xi S 5 T + 8N 3 T N 3 , R ij 11,10 = √
3M 1i T S 5 T , R ij 11,11 = −I,
R ij 12,10 = −B T i S 5 T − S 6 G xi , R ij 14,10 = −D T 1i S 5 T , R ij 12,12 = −S 6 B i − B i T S 6 T + 6N 2 T N 2 ,
R ij 13,12 = √
3M 1i T S 6 T , R ij 13,13 = −I, R ij 14,12 = −D T 1i S 6 T , R ij 14,14 = −γI,
(25) and
E 1 = ±κ ξ X 1 +
r ρ=1 ρ=ξ
κ ρ (P 1ρ + X 1 − P 1ξ ),
E 2 = ±κ ξ X 2 +
r ρ=1 ρ=ξ
κ ρ (P 2ρ + X 2 − P 2ξ ). (26)
715 Then, the observer gains are
L i = S 3 −1 S Li , L n = X −1 R n . (27) The ± sign means that the LMIs must be checked for both positive and negative signs. Note that parameters η and κ i should be given in advance and LMIs can be solved to find the best value for γ.
Proof. Let us define the following fuzzy Lyapunov function:
V (t) :=
r i=1
ω i (z)
e T (t)P 1i e(t) + x T (t)P 2i x(t)
. (28)
Its time derivative is V (t) = x ˙ T (t)
r
i=1
˙ ω i (z)P 2i
x(t)
+ 2x T (t) r
i=1
ω i (z)P 2i
˙x(t)
+ e T (t) r
i=1
˙ ω i (z)P 1i
e(t)
+ 2e T (t) r
i=1
ω i (z)P 1i
˙e(t).
(29)
In order to make the LMI representation possible, two zero terms are added to (29), which results in
V (t) = x ˙ T (t) r
i=1
˙ ω i (z)P 2i
x(t)
+ 2x T (t) r
i=1
ω i (z)P 2i
˙ x(t)
+ e T (t) r
i=1
˙ ω i (z)P 1i
e(t)
+ 2e T (t) r
i=1
ω i (z)P 1i
˙e(t) + 2
x T (t)S 1 + ˙ x T (t)S 2
+ ϕ(x(t), u(t), t) T S 5 + u T (t)S 6 S x1 + 2
e T (t)S 3 + ˙e T (t)S 4 S e1 ,
(30)
where
S x1 = ˙x(t) − r
i=1
ω i (z)
(A i + ΔA i )x(t) + (G xi + ΔG xi )ϕ(x(t), u(t), t) + (B i + ΔB i )u(t) + D 1i ν(t) ,
S e1 = ˙e(t) − r
i=1
r j=1
ω ij (z)
(A i + L i C j )e(t) + (G xi + L i G yj )δϕ(t)
−
(ΔA i + L i ΔC j )x(t) + (ΔG xi + L i ΔG yj )ϕ(x(t), u(t), t)
+ ΔB i u(t) + (D 1i + L i D 2j )ν(t) .
(31) Using Lemma 2, (8) and (30), ˙ V (t) becomes
V (t) ˙ ≤ x T (t) r
i=1
˙ ω i (z)P 2i
x(t)
+ 2x T (t) r
i=1
ω i (z)P 2i
˙x(t)
+ e T (t) r
i=1
˙
ω i (z)P 1i | e(t)
+ 2e T (t) r
i=1
ω i (z)P 1i
˙e(t) + 2
x T (t)S 1 + ˙x T (t)S 2
+ ϕ T (x(t), u(t), t)S 5 + u T (t)S 6 S x2 + 2
e T (t)S 3 + ˙e T (t)S 4 S e2 +
r i=1
r j=1
ω ij (z)
3x T (t)S 1 M 1i M 1i T S 1 T x(t)
+ 3 ˙x T (t)S 2 M 1i M 1i T S 2 T x(t) ˙ + 3u T (t)S 6 M 1i M 1i T S 6 T u(t)
+ 3ϕ(x(t), u(t), t) T S 5 M 1i M 1i T S 5 T ϕ(x(t), u(t), t) + e T (t)S 3 (3M 1i M 1i T + 2L i M 2j M 2j T L T i )S 3 T e(t) + ˙e T (t)S 4 (3M 1i M 1i T + 2L i M 2j M 2j T L T i )S 4 T ˙e(t) + 8x T (t)N 1 T N 1 x(t) + 6u T (t)N 2 T N 2 u(t)) + 8ϕ(x(t), u(t), t) T N 3 T N 3 ϕ(x(t), u(t), t)
, (32) where
S x2 = ˙ x(t) − r
i=1
ω i (z)
A i x(t) + G xi ϕ(x(t), u(t), t)
+ B i u(t) + D 1i ν(t) ,
S e2 = ˙e(t) − r
i=1
r j=1
ω ij (z)
(A i + L i C j )e(t) + (G xi + L i G yj )δϕ(t)
− (D 1i + L i D 2j )ν(t) .
(33)
716
Then ˙ V (t) ≤ r
i=1
r
j=1 ω ij K T R ¯ ij K, where K =
e(t) ˙e(t) δϕ(t) x(t) ˙ x(t) ϕ(x(t), u(t), t) u(t) ν(t) T
R ¯ ij =
⎛
⎜ ⎝
R ¯ ij 11 ∗ ∗ .. . . . . ∗ R ¯ ij 81 · · · R ¯ ij 88 ,
⎞
⎟ ⎠
(34) in which
R ¯ ij 11 = r
ρ=1 ω ˙ ρ (z)P 1ρ − S 3 (A i + L i C j )
− (A i + L i C j ) T S 3 T
+ S 3 (3M 1i M 1i T + 2L i M 2j M 2j T L T i )S 3 T , R ¯ ij 21 = P 1i − S 4 (A i + L i C j ) + S 3 T ,
R ¯ ij 31 = −(G xi + L i G yj ) T S 3 T , R ¯ ij 81 = −(D 1 i + L i D 2 j) T S 3 T ,
R ¯ ij 22 = S 4 + S 4 T + S 4 (3M 1i M 1i T + 2L i M 2j M 2j T L T i ) T S 4 , R ¯ ij 32 = −(G xi + L i G yj ) T S 4 T ,
R ¯ ij 82 = −(D 1 i + L i D 2 j) T S 4 T , R ¯ ij 44 = r
ρ=1 ω ˙ ρ (z)P 2ρ − S 1 A i − A T i S T 1 + 8N 1 T N 1 + 3S 1 M 1i M 1i T S 1 T ,
R ¯ ij 54 = P 2i + S 1 T − S 2 A i , R ¯ ij 64 = −S 5 A i − G T xi S T 1 , R ¯ ij 74 = −B i T S 1 T − S 6 A i , R ¯ ij 84 = −D 1i T S 1 T ,
R ¯ ij 55 = S 2 + S 2 T + 3S 2 M 1i M 1i T S 2 T , R ¯ ij 65 = S 5 − G T xi S 2 T ,
R ¯ ij 75 = −B i T S 2 T + S 6 , R ¯ ij 85 = −D T 1i S T 2 ,
R ¯ ij 66 = −S 5 G xi − G T xi S 5 T + 8N 3 T N 3 + 3S 5 M 1i M 1i T S 5 T , R ¯ ij 76 = −B i T S 5 T − S 6 G xi , R ¯ ij 86 = −D T 1i S 5 T ,
R ¯ ij 77 = −S 6 B i − B i T S T 6 + 6N 2 T N 2 + 3S 6 M 1i M 1i T S 6 , R ¯ ij 87 = −D T 1i S 6 T .
(35) The other terms in ¯ R ij are equal to zero. Based on (10), we get
˙ ω ξ = −
r ρ=1,ρ=ξ
˙
ω ρ (z), (36)
and hence
r ρ=1
˙ ω ρ (z)P 1ρ
= ˙ ω ξ X 1 +
r ρ=1,ρ=ξ
˙
ω ρ (z)[P 1ρ + X 1 − P 1ξ ]. (37)
Moreover, based on (22),
˙
ω ρ (z)[P 1ρ + X 1 − P 1ξ ] ≤ κ ρ [P 1ρ + X 1 − P 1ξ ]. (38)
By defining S 4 := ηS 3 and S 3 L i := S Li and using the Schur complement lemma (Boyd et al., 1994) on the diagonal elements of ¯ R ij , and based on (26), (37) and (38), we have
V (t) + e ˙ T (t)e(t) − γν T (t)ν(t)
≤
r i=1
r j=1
ω ij K ¯ T R ˜ ij K, ¯ (39)
where ˜ R ij is equal to R ij defined in (24) except that R ˜ ij 5,5 = 0. Moreover, ¯ K is the same as K with augmented terms equal to one in order to have proper dimensions.
Adding the term e(t)
δϕ(t)
T
Φ T M Φ e(t)
δϕ(t)
to both the sides of (39), introducing R n = XL n and using Schur complements on the right-hand side of the resulting inequality and based on the Schur complement of (18), we have
r i=1
r j=1
ω ij ( ¯ K T R ˜ ij K) + ¯
e(t) δϕ(t)
T Φ T M Φ
e(t) δϕ(t)
=
r i=1
r j=1
ω ij K ¯ T Ξ ij K, ¯ (40)
in which Ξ ij is defined in (23). Then, from Lemma 1 and (22), we get
V (t) + e ˙ T (t)e(t) − γν T (t)ν(t) +
e(t) δϕ(t)
T Φ T M Φ
e(t) δϕ(t)
< 0, (41) which, based on (15), guarantees that the error dynamic is asymptotically stable with an H ∞ performance bound γ.
That is,
e T (t)e(t) < γν T (t)ν(t). (42)
Remark 1. One of the main drawbacks of using the FLF for continuous-time systems is the existence of a derivative of the membership functions, which can be written as follows (Manai and Benrejeb, 2011):
˙
ω i (z) = ∂ω i
∂z(t) · ∂z(t)
∂x(t) · dx(t)
dt . (43)
This derivative, if it exists at all, is usually hard to
calculate. However, it can be assumed that it is limited by
a bound (κ i ), which is used here. For the existence of a
bound on ∂ω i /∂z(t), it is assumed that the membership
functions are continuous. In addition, ∂z(t)/∂x(t) is
known a priori. However, in general, it is hard to find a
limit on ˙ x(t). Nevertheless, when ˙ x(t) increases from its
bounds, it means that x(t) has also been increased from
its predefined bounds. Hence, the first derivative must
717 approach to zero, which results in ˙ ω i (z) → 0. For more
detail, the reader may refer to the work of Guerra and Bernal (2012).
Remark 2. It is also possible to assume the multiplier matrix M with the following form:
M =
0 X
X T −Y
. (44)
In this case, the equality in (18) changes to Φ T M Φ =
Υ T 11 X −1 Υ 11 Υ T 11 Υ 11 Υ 22 + Υ T 22 − Y
, (45) where
Υ 11 = XC q + R n C z ,
Υ 22 = XD q + R n G yz . (46) Hence, it is possible to use Theorem 1 with the following changes:
R ij 5,1 = R ij 5,1 + (XC q + R n C i ), R ij 1,5 = R ij 1,5 + (XC q + R n C i ) T ,
R ij 5,5 = R ij 5,5 + (XD q + R n G yi ) T + (XD q + R n G yi ), Ψ =
XC q + R n C i 0 0 0 .
(47) In order to further enlarge the class of nonlinearities, the matrix M can have the following forms as well:
M = T T
0 X
X T −Y
T ,
M = T T
X 0 0 −Y
T , (48)
where
T =
T 11 T 12 T 21 T 22
. (49)
In this case, T 22 + T 21 D q should be nonsingular. Then, it is possible to formulate the problem using LMIs.
However, in order to use Theorem 1, L n cannot be selected arbitrarily. This means that there are fewer degrees of freedom for the LMI solver. To solve this problem, an extra term should be added to the observer (Ac¸ikmese and Corless, 2011).
Remark 3. One essential point in Theorem 1 is the injection term L n
ˆ
y(t) − y(t) in ˆ ϕ
ˆ
x(t), u(t), t in (11). It should be mentioned that, if the nonlinearity ϕ
x(t), u(t), t
is Lipschitz (instead of incremental quadratic), then L n cannot be introduced. This case is shown in the following theorem.
Theorem 2. Assume that | ˙ω i (z) |<κ i for known positive real numbers κ i . The error dynamic (12) is asymptoti- cally stable and with an H ∞ performance bound γ > 0 if
ϕ
x(t), u(t), t
satisfies the Lipschitz condition e T (t)Γ T θΛΓe(t)
− δϕ T
x(t), u(t), t Λδϕ
x(t), u(t), t
≥ 0, (50) where θ is the Lipschitz constant and Γ is a constant matrix with proper dimensions and there exist matrices P 1i = P 1i T > 0, P 2i = P 2i T > 0 (1 ≤ i ≤ r), X 1 , X 2 , S i , SL i (1 ≤ i ≤ 6), Λ = diag [λ 1 , . . . , λ s ] and a scalar η > 0 such that
P 2ρ + X 2 − P 2ξ ≥ 0, ∀ρ ∈ [1, . . . , r] − ξ, P 1ρ + X 1 − P 1ξ ≥ 0, ∀ρ ∈ [1, . . . , r] − ξ,
R ii < 0, 1 < i < r, 1
r − 1 R ii + 1
2 ( R ij + R ji ) < 0, 1 < i = j < r, (51) where R ij is defined in (24) with the following changes:
R ij 1,1 = R ij 1,1 + Γ T θΛΓ,
R ij 5,5 = −Λ. (52)
Then, the observer gains are
L i = S 3 −1 S Li . (53) Proof. The proof is similar to that of Theorem 1. The only difference is adding (50), instead of (15), to both the sides of (39). Moreover, the nonlinear injection term is no longer defined here and hence (23) changes to Ξ ij = R ij .
Remark 4. It should be noted that a common case for nonlinear T–S observers, which is considered in the literature, is a T–S model with Lipschitz nonlinearities along with a traditional Luenberger observer (Theorem 2).
On the other hand, in this paper, the nonlinearity is of the δQC type, which encompasses a wider class of systems (Theorem 1). Moreover, by introducing a nonlinear injection term to the observer, better state estimates can be achieved. In the following section, it will be shown through simulations that Theorem 2 cannot be used for some systems, while Theorem 1 is a more general case and can be applied to a wider class of systems with almost no extra computation time.
4. Simulation examples
Example 1. In this example, the performance of
two theorems in this paper is compared. Consider the
system shown in Fig. 1, which represents a Translational
Oscillator with an eccentric Rotational Actuator (TORA)
system (Lee, 2004; Karagiannis et al., 2005). The
nonlinear coupling between the rotational motion of the
actuator and the translational motion of the oscillator
provides a mechanism for control. Let x 1 and x 2 denote
718
the translational position and velocity of the cart, and x 3 and x 4 denote the angular position and velocity of the rotational mass, respectively. Then, the system dynamics can be described as (Tanaka and Wang, 2001)
˙x =
⎛
⎜ ⎜
⎜ ⎜
⎜ ⎝
x 2
−x 1 + x 2 4 sin x 3 1 − 2 cos 2 x 3
x 4
cos x 3 (x 1 − x 2 4 sin x 3 ) 1 − 2 cos 2 x 3
⎞
⎟ ⎟
⎟ ⎟
⎟ ⎠
+
⎛
⎜ ⎜
⎜ ⎜
⎜ ⎝
− cos x 0 3 1 − 2 cos 2 x 3
0 1 1 − 2 cos 2 x 3
⎞
⎟ ⎟
⎟ ⎟
⎟ ⎠ u + d 1 ,
y =
x 1 x 2 T + d 2 ,
(54)
where
d 1 =
⎛
⎜ ⎜
⎜ ⎜
⎜ ⎝ 0 1 1 − 2 cos 2 x 3
− cos x 0 3
1 − 2 cos 2 x 3
⎞
⎟ ⎟
⎟ ⎟
⎟ ⎠ ν(t),
d 2 =
0.1 0.2
ν(t), = 0.05,
(55)
in which ν(t) represents band-limited white noise with power of 0.001. This system can be modeled as follows:
A 1 =
⎛
⎜ ⎜
⎝
0 1 0 0
1− −1
20 0 0 0 0 0 1
1−
20 0 0
⎞
⎟ ⎟
⎠ ,
A 2 =
⎛
⎜ ⎜
⎝
0 1 0 0
1− −1
20 0 0 0 0 0 1
1− −
20 0 0
⎞
⎟ ⎟
⎠ ,
A 3 =
⎛
⎜ ⎜
⎝
0 1 0 0
−1 0 0 −a
0 0 0 1
0 0 0 0
⎞
⎟ ⎟
⎠ ,
A 4 =
⎛
⎜ ⎜
⎝
0 1 0 0
−1 0 0 a
0 0 0 1 0 0 0 0
⎞
⎟ ⎟
⎠ ,
B 1 T =
0 −
1 − 2 0 1 1 − 2
, B 2 T = B 4 T =
0 0 0 1 , B 3 T =
0
1 − 2 0 1 1 − 2
, G x1 = G x3 =
0 0 0 1 T , G x2 = −G x4 =
0 0 0 T , C =
1 0 0 0 0 0 1 0
, ϕ(x) = x 2 4 + ax 4 , a = 4.
(56)
Also x 4 ∈ [−a a]. To introduce uncertainty in the model, is equal to 0.01. Note that the system is stabilized first and then an observer is designed for the stable system.
The state feedback gains for the stabilizer are K 1 =
−0.2000 −5.9282 −0.3682 0.6357 , K 2 =
−0.2443 −6.3255 −0.4499 0.7859 , K 3 =
−0.3245 −7.3220 −0.5987 1.0194 , K 4 =
−0.2278 −5.7132 −0.4192 0.7102 . (57)
Fig. 1. TORA system.
The following gains are obtained for the observer based on Theorem 1:
L 1 =
⎛
⎜ ⎜
⎝
−5.42 2.71
−11.96 0.94 3.09 −1.53 2.02 −0.95
⎞
⎟ ⎟
⎠ ,
L 2 =
⎛
⎜ ⎜
⎝
−5.57 2.81
−13.53 2.06 3.44 −1.79 2.45 −1.31
⎞
⎟ ⎟
⎠ ,
L 3 =
⎛
⎜ ⎜
⎝
−5.46 2.77
−12.27 1.66 3.14 −1.70 2.03 −1.23
⎞
⎟ ⎟
⎠ ,
L 4 =
⎛
⎜ ⎜
⎝
−5.47 2.76
−12.40 1.56 3.20 −1.69 2.14 −1.18
⎞
⎟ ⎟
⎠ ,
L n =
0.5825 −1.0597 ,
(58)
719
0 1 2 3 4 5 6 7 8 9 10
−1.5
−1
−0.5 0 0.5 1 1.5
Time(s)
States
x2
x1
x3 x4
Fig. 2. States (solid lines) and their estimaties (dotted lines) of the TORA system based on Theorem 1.
0 1 2 3 4 5 6 7 8 9 10
−2
−1 0 1
IQC
e1 e
2 e
3 e
4
0 1 2 3 4 5 6 7 8 9 10
−2
−1 0 1
Lipschitz
Time(s)
e1 e2 e3 e4