Preprints of the 18th IFAC World Congress

Milano (Italy) August 28 - September 2, 2011

Extended State Observer for Nonlinear

Systems with Uncertainty
Bao-Zhu Guo ∗ Zhi-Liang Zhao ∗∗
∗
Academy of Mathematics and Systems Science, Academia Sinica,
Beijing 100190, P.R. China, and School of Computational and Applied
Mathematics, University of the Witwatersrand, Wits 2050,
Johannesburg, South Africa (e-mail: [email protected])
∗∗
Department of Mathematics,University of Science and Technology of
China, Hefei 230061, P.R. China (e-mail: [email protected])

Abstract: The extended state observer first proposed by Jingqing Han in [J. Q. Han, The
extended state observer for a class of uncertain systems, Control and Decision, 10(1)(1995),
85-88] is the main interim loop of active disturbance rejection control that is taking off as a
technology after numerous successful applications in engineering. Unfortunately, there is no a
rigorous proof of convergence to date. In this paper, we attempt to tackle this long unsolved
extraordinary problem.
The main idea is to transform the error equation of objective system with its extended state
observer into a disturbed system of a asymptotical stable system, with which the effect of total
disturbance error is eliminated by high-gain.

Keywords: Extended state observer, nonlinear systems, uncertainty.

1. INTRODUCTION

 ẋ1 (t) = x2 (t), x1 (0) = x10 ,
 ẋ2 (t) = x3 (t), x2 (0) = x20 ,



.

The observer design is always one of fundamental issues ..
(2)
in nonlinear systems control. The Luenberger observer- 
 ẋn (t) = f (t, x1 (t), x2 (t), . . . , xn (t))
based approach is the main method. They are many 

 +w(t) + u(t), xn (0) = xn0 ,
other approaches like slide mode based approach. Some

y(t) = x1 (t),

of them are of advantage in robustness, but most of
them come up against the problems like adaptability where u is the input (control), y the output (measure-
and chattering. We refer to Corless, Tu [1998], Darouach ment), f an usually unknown system function, and w
et al. [1994], Gourshankar et al. [1997], Koshkouei, Zinober the uncertain external disturbance. f + w is called “to-
[1998], Viswanadham, Ramakrishna [1980], Slotine et al. tal disturbance” in the elaborated expository paper Han
[1987], Walcott et al. [1987], Walcott, Zak [1987] and [2009]. (x10 , x20 , . . . , xn0 ) is the initial state, and αi , i =
references therein, and the recent book Besancon [2007]. 1, 2, . . . , n+1 are regulable constants. The main idea of the
A particular attention is paid to Huang, Han [1999] where extended state observer is that for the appropriately cho-
the comparison of the approach discussed in present paper sen functions gi , the state of the observer x̂i , i = 1, 2, . . . , n
with other existing approaches were presented in details. and x̂n+1 can be, through regulating αi , considered as
In his seminal work Han [1995], Han proposed the follow- the approximations of the corresponding state xi for i =
ing extended state observer (ESO): 1, 2, . . . , n, and the total disturbance f + w, respectively.
The last remarkable fact is the source in which the external
x̂˙ 1 (t) = x̂2 (t) − α1 g1 (x̂1 (t) − y(t)),

 state observer is rooted. This idea is essentially the same
with the high-gain observer with augmented state variable


x̂˙ 2 (t) = x̂3 (t) − α2 g2 (x̂1 (t) − y(t)),


used in Freidovich, Khalil [2008] (an expository survey can





.. be found in Khalil [2008]). The numerical studies (e.g., Han
. (1) [1995]) and many other studies the years followed show

that for some nonlinear functions gi and parameter αi ,

 x̂˙ n (t) = x̂n+1 (t) − αn gn (x̂1 (t) − y(t)) + u(t),




 the observer (1) performs a very satisfactory adaptability,
x̂˙ n+1 (t) = −αn+1 gn+1 (x̂1 (t) − y(t)), robustness, and anti-chattering. For other perspectives of


this remarkable series researches, we refer the reader to
for an n-dimensional SISO nonlinear system Han [2009].
Unfortunately, although huge applications have been car-
½
x(n) (t) = f (x(t), ẋ(t), . . . , x(n−1) (t)) + w(t),
y(t) = x(t), ried out in engineering applications since then (see e.g.,
Hou et al. [2001], Miklosovic, Gao [2005], Zheng et al.
which can be written as [2008], Zheng, Dong et al. [2007], Zheng, Gao [2010],

Copyright by the 1855

International Federation of Automatic Control (IFAC)
Preprints of the 18th IFAC World Congress
Milano (Italy) August 28 - September 2, 2011

Zheng, Cheng et al. [2007], Zhou, Gao [2007]) , the choice Assumption (H1). The possibly unknown functions f, w
of functions gi is essentially experiential. In order to apply are differentiable with respect to their variables, and
conveniently in practice, Gao [2003] proposed the following ¯ ¯ ¯ ¯ n
linear extended state observer (LESO in short in what ¯ ∂f ¯ ¯ ∂f ¯ X
|u| + |f | + |ẇ| + ¯ ¯ + ¯
¯ ¯ ¯ ¯ ≤ c0 + cj |xj |k (5)
follows), a special case of (1): ∂t ∂xi ¯ j=1
α1
x̂˙ 1 (t) = x̂2 (t) +

(y(t) − x̂1 (t)), for some positive constants cj and positive integer k.
ε




 α2 We proceed as follows. In next section, Section 2, we
 x̂˙ 2 (t) = x̂3 (t) + 2 (y(t) − x̂1 (t)),





 ε show that the NLESO (4) is convergent under additional

.. assumptions. Various special NLESO are presented in Sec-
. (3) tion 3. Meanwhile, some Examples and numerical simula-

tions are presented from time to time for illustrations.

 αn
x̂˙ n (t) = x̂n+1 (t) + n (y(t) − x̂1 (t)) + u(t),



ε





 x̂˙
 αn+1
n+1 (t) = n+1 (y(t) − x̂1 (t)),
ε
2. THE CONVERGENCE OF EXTENDED STATE
where αi , i = 1, 2, . . . , n + 1 are pertinent constants, and ε OBSERVER
is regulable gain constant.
The LESO (3) is essentially similar with the “extended In this section, we consider the convergence of the NLESO
high-gain observer” studied in Freidovich, Khalil [2008]. (4). To this end, we make some additional assumptions.
High-gain observer has been extensively used to design
the stabilizing output feedback control by using singu- Assumption (H2). w and the solution of (2) satisfy |w|+
lar perturbation method, see, for instance Atassi, Khalil |xi (t)| ≤ B for some constant B > 0 and all i = 1, 2 . . . , n,
[2000], Esfandiari, Khalil [1992], Freidovich, Khalil [2008], and t ≥ 0.
Oh, Khalil [1997] and references therein. The convergence Assumption (H3). There exist constants λi (i = 1, 2, 3, 4),
of the high-gain observer for determined systems without α, β and positive definite, continuous differentiable func-
augmented state variable was given in Theorem 6 of Be- tions V, W : Rn+1 → R such that
sancon [2007] on page 18. The high-gain observer with
augmented state variable was used in Freidovich, Khalil • λ1 kyk2 ≤ V (y) ≤ λ2 kyk2 , λ3 kyk2 ≤ W (y) ≤ λ4 kyk2 ,
[2008] for tracking problems. A nice survey of high-gain n
observer with or without augmented state variable can
X ∂V ∂V
• (yi+1 − gi (y1 )) − gn+1 (y1 ) ≤ −W (y),
be found in Khalil [2008]. However, the convergence of i=1
∂yi ∂yn+1
high-gain observer with uncertainties, particularly with ¯ ¯
augmented state variable is, to the best of our knowledge, ¯ ∂V ¯
• ¯ ∂yn+1 ¯ ≤ βkyk,
¯ ¯
not available in literature. For linear extended state ob-
server (3), Zheng et al. [2007] presented some convergence
results where either f + w is completely known or f + w where y = (y1 , y2 , . . . , yn+1 ), k · k denotes the Euclid norm
is completely unknown. The convergence results for some of Rn+1 .
special LESO were also available in Yang, Huang [2009]. Theorem 2.1. Suppose that Assumption (H1)-(H3) are
satisfied. Then
In this paper, we are concerned with the convergence of
nonlinear extended state observe (NLESO in short in what (i). For every positive constant a,
follows) of the following:
lim |xi (t) − x̂i (t)| = 0 uniformly in t ∈ [a, ∞).
 µ ¶ ε→0
˙ n−1 y(t) − x̂1 (t)
 x̂1 (t) = x̂2 (t) + ε g1 ,


εn

lim |xi (t) − x̂i (t)| ≤ O εn+2−i ,
¡ ¢
(ii).



 µ ¶ t→∞
˙ 2 (t) = x̂3 (t) + εn−2 g2 y(t) − x̂1 (t) ,


x̂

where xi , x̂i denote the solution of (2) and (4) respectively,

εn



i = 1, 2, . . . , n + 1, xn+1 = f + w is the extended state


..

. (4) variable for system (2).


Proof. By extra state variable xn+1 = f +w (Han [2009]),
 µ ¶
 y(t) − x̂1 (t)
 x̂˙ n (t) = x̂n+1 (t) + gn

+ u(t),

system (2) can be written as



 εn

ẋ1 (t) = x2 (t), x1 (0) = x10 ,
 µ ¶ 
 1 y − x̂1 (t)
 x̂˙ n+1 (t) = gn+1

,
 
ẋ2 (t) = x3 (t), x2 (0) = x20 ,
 
ε εn


 ..


which is a special case of (1) and an nonlinear general- . (1)
ization of LESO (3) for regulable gain parameter ε and 

 ẋn (t) = xn+1 (t) + u(t), xn (0) = xn0 ,
pertinent functions gi , i = 1, 2, . . . , n + 1. Notice that the  ẋn+1 (t) = L̇(t), xn+1 (0) = L(0),



solution of (4) depends on parameter ε, but here we miss y(t) = x1 (t),
ε by abuse of notation without confusion.
where L(t) = f (t, x1 (t), x2 (t), . . . , xn (t)) + w(t). We first
The following assumptions are presumed: notice that

1856
Preprints of the 18th IFAC World Congress
Milano (Italy) August 28 - September 2, 2011

identity matrix I. Define the Lyapunov function V, W :

¯
d ¯
∆(t) = f (s, x1 (s), . . . , xn (s))¯¯ + ẇ(εt) Rn+1 → R by
ds s=εt
∂ V (η) = hP η, ηi, W (η) = hη, ηi. (10)
= f (εt, x1 (εt), . . . , xn (εt)) 2 2
∂t Then λmin (P )kηk ≤ V (η) ≤ λmax (P )kηk ,
n (2) n
X ∂ ∂V ∂V
+ xi+1 (εt) f (εt, x1 (εt), . . . , xn (εt))
X
∂x (ηi+1 − αi η1 ) − αn+1 η1 = −kηk2 = −W (η),
i=1 i
i=1
∂ηi ∂ηn+1
∂ and
+u(εt) f (εt, x1 (εt), . . . , xn (εt)) + ẇ(εt).
∂xn ¯ ¯ ¯¯
¯ ∂V ¯ ¯¯ ∂V ¯¯
¯¯
⊤
¯ ∂ηn+1 ¯ ¯¯ ∂η ¯¯ = k2η P k ≤ 2kP kkηk = 2λmax (P )kηk.
From Assumptions (H1)-(H2), |∆(t)| ≤ M is uniformly ¯ ¯ ≤ ¯¯ ¯¯
bounded for some M > 0 and all t ≥ 0.
where λmax (P ) and λmin (P ) are the maximal and minimal
Set eigenvalues of P , respectively. So V, W defined by (10)
ei (εt) satisfies Assumption (H3). Therefore, as a consequence of
ei (t) = xi (t) − x̂i (t), ηi (t) = , i = 1, 2, . . . , n + 1.
εn+1−i Theorem 2.1, we have improved the stability of LESO (3)
(3) presented in Zheng et al. [2007] where the a in the following
Then a direct computation shows that η = (η1 , η2 , . . . , Corollary 2.1 is required to be large.
ηn+1 )⊤ satisfies
Corollary 2.1. If E is Hurwitz and assumptions (H1)-(H2)
e1 (0)


 η̇1 = η2 − g1 (η1 ), η1 (0) = , are satisfied, then
εr



(i). For every positive constant a,

e2 (0)



 η̇2 = η3 − g2 (η1 ), η2 (0) = r−1 ,
lim |xi (t) − x̂i (t)| = 0 uniformly in t ∈ [a, ∞).

ε


ε→0

.. (4)
.
lim |xi (t) − x̂i (t)| ≤ O εn+2−i ,
 ¡ ¢


 (ii).
t→∞
en (0)




 η̇n = ηn+1 − gn (η1 ), ηn (0) = , where xi , x̂i denote the solutions of (2), (3) respectively,
ε



 i = 1, 2 · · · , n + 1, xn+1 = f + w is the extended state
η̇n+1 = −gn+1 (η1 ) + ε∆(t), ηn+1 (0) = er+1 (0).

variable for system (2).
By Assumption (H3), finding the derivative of V (η(t)) with Example 2.1. For the system following
respect to t along the solution η(t) of system (4) gives 
ẋ (t) = x2 (t),
√  1

d λ3 λ1 p
V (η(t)) ≤ − V (η) + εM β V (η). (5) ẋ2 (t) = −x1 (t) − x2 (t) + w(t) + u(t), (11)
dt λ2 λ1 
y(t) = x1 (t),

It follows that √
dp λ3 p λ1 εM β we design an nonlinear extended state observer following:
V (η(t)) ≤ − V (η(t)) + . (6)
dt 2λ2 2λ1 µ ¶
3 y(t) − x̂1 (t)

By Assumption (H3) again, we have  ˙
x̂1 (t) = x̂2 (t) + (y(t) − x̂1 (t)) − εϕ ,
ε2


 ε
εM β t − 2λ
p
λ1 V (η(0)) − 2λ
Z 
λ3 λ3 
t (t−s) 
kηk ≤ e 2 + e 1 ds. (7) ˙ 3
λ1 2λ1 0  x̂2 (t) = x̂3 (t) + ε2 (y(t) − x̂1 (t)) + u(t),
This together with (3) yields


 x̂˙ (t) = 1 (y(t) − x̂ (t)).


"p 
λ1 V (η(0)) − 2λ 3 1
|ei (t)| ≤ ε n+1−i
e
λ3 t
2ε
ε3
λ1 (12)
(8) 1
Z t # where ϕ : R → R is defined as ϕ(r) = − for r ∈
+
εM β ε − 2λ
e
λ3
1
(t/ε−s)
ds . ³ π´ 1 ³ π π ´4
2λ1 0 −∞, − , ϕ(r) = sin r for r ∈ − , , ϕ(r) =
2³ 4 2 2
Both (i) and (ii) of Theorem 2.1 then follow from (8). The 1 π ´
for r ∈ , −∞ . In this case, gi in (4) can be taken
proof is complete. 2 4 2
The result of Theorem 2.1 enables us to deduce the g1 (y1 ) = −3y1 −ϕ(y1 ), g2 (y1 ) = −3y1 , g3 (y1 ) = −y1 . (13)
convergence of LESO (3) immediately. Actually, if the The Lyapunov function V : R3 → R for this case is defined
matrix following is Hurwitz: by
−α1 1 0 · · · 0
  Z x1
 −α2 0 1 · · · 0  V (y) = hP y, yi + ϕ(s)ds, (14)
 . .. .. . . ..  0
E=  .. . . . . , (9)

 −α 0 0 · · · 1  where P is the positive definite solution of!Lyapunov
−3 1 0
n
Ã
−αn+1 0 0 · · · 0 equation P E + E ⊤ P = −I with E = −3 0 1 . A direct
let P be the positive definite matrix solution of the Lya- −1 0 0
punov equation P E + E ⊤ P = −I for (n + 1)-dimensional computation shows that

1857
Preprints of the 18th IFAC World Congress
Milano (Italy) August 28 - September 2, 2011


2
µ ¶
X ∂V ∂V  ˙ 1 (t) = x̂2 (t) + εn−1 g1 v(t) − x̂1 (t) ,
x̂
(yi+1 + gi (y1 )) + g3 (y1 )

εn


i=1
∂yi ∂y3 

(15)

 µ ¶
µ 2 

˙ n−2 v(t) − x̂1 (t)
7y 2 3y 2
¶
y1 x̂2 (t) = x̂3 (t) + ε g2 ,


≤− + 2 + 3 , −W (y1 , y2 , y3 ). εn



8 8 4 

..

. (18)
So all conditions of Assumption (H3) are satisfied. There- 
fore, (12) serves as a well defined NLESO for (11) accord-

 µ ¶
˙ n (t) = x̂n+1 (t) + gn v(t) − x̂1 (t) ,


ing to Theorem 2.1. Now take the data as x̂


εn




x1 (0) = x2 (0) = 1, x̂1 (0) = x̂2 (0) = x̂3 (0) = 0,

 µ ¶
(16)


˙ 1 v(t) − x̂1 (t)
u(t) = sin t, w(t) = cos t, ε = 0.01. 
x̂
 n+1
 (t) = gn+1 ,
ε εn
The simulations for NLESO (12) are plotted as Fig. 1.
The convergence result in this special case is stated as
It is seen from Fig. 1 that the NLESO (12) is well tracking Theorem 3.1 below.
the state of the system (11). Theorem 3.1. [Output tracking differentiator] Sup-
In what follows, we relax the conditions of Assumption pose that Assumption (H3) holds and v (n) (t) is bounded.
(H3) by Assumption (H4) following. Then

Assumption (H4). There exist constant R > 0, α > 0 (i). For every positive constant a,
and positive definite, continuous differentiable function
V, W : Rn+1 → R such that lim |v (i−1) (t) − x̂i (t)| = 0, uniformly on [a, ∞).
ε→0

• {y|V (y) ≤ d} is bounded for any d > 0,

lim |v (i−1) (t) − x̂i (t)| ≤ O εn+2−i ,
¡ ¢
(ii).
n t→∞
X ∂V ∂V
• (yi+1 − gi (y1 )) − gn+1 (y1 ) ≤ −W (y), where x̂i is the solution of (18), i = 1, 2 · · · , n + 1.
i=1
∂y i ∂y n+1

The next example is different to system (2) and those

¯ ¯
¯ ∂V ¯
• ¯ ∂yn+1 ¯ ≤ αW (y) for kyk > R.
¯ ¯ discussed in Zhou et al. [2009], in which the system
function f is known. The unknown part is only the external
disturbance w. In this case, our NLESO can be modified
We can obtain the weak convergence result stated in as
Theorem 2.2 below. µ ¶
x1 (t) − x̂1 (t)

Theorem 2.2. Under Assumptions (H1),(H2) and (H4), ˙ n−1
 x̂1 (t) = x̂2 (t) + ε g1 ,

εn

the nonlinear extended state observer (4) is convergent 


in the sense that for any σ ∈ (0, 1), there exists εσ ∈ (0, 1)
 µ ¶
˙ 2 (t) = x̂3 (t) + εn−2 g2 x1 (t) − x̂1 (t) ,


x̂

such that for any ε ∈ (0, εσ ), 
εn





|xi (t) − x̂i (t)| < σ, ∀ t ∈ (Tε , ∞), 
 ..


. (19)
where Tε > 0 depends on ε, xi , x̂i denote the solutions of µ ¶
(2), (4) respectively, i = 1, 2 · · · , n + 1, xn+1 = f + w is x (t) − x̂ (t)

1 1
 x̂˙ n (t) = x̂n+1 (t) + gn



the extended state variable for system (2). 

 εn


Duo to page limiitation, we omit the proof. 2



 +f (t, x̂1 (t), x̂2 (t), . . . , x̂n (t)) + u(t),
 µ ¶
x1 (t) − x̂1 (t)


 x̂˙ n+1 (t) = gn+1

 .
εn
3. SOME SPECIAL EXTENDED STATE OBSERVER
Using the same notation as in (3) and setting xn+1 = w
in this case, we get
In this section, we state some applications of extended
e (0)

state observer. The first special case is what so called out-  η̇1 = η2 − g1 (η1 ), η1 (0) = 1 ,

put tracking differentiator, a special kind of tracking differ- εn




entiator (Han [1989, 2009]). Suppose that v is the tracked 

 e2 (0)
signal. Let xi (t) = v (i−1) (t). Then xi , i = 1, 2, . . . , n satisfy

 η̇2 = η3 − g2 (η1 ), η2 (0) = n−1 ,
ε




ẋ1 (t) = x2 (t), ..

 ẋ2 (t) = x3 (t),
 

 .
 
.. en (0)
 

(17)

.  η̇n = ηn+1 − gn (η1 ) + δ1 (t), ηn (0) = ε ,



(n−1)

 ẋn (t) = v (t),

 

η̇n+1 = −gn+1 (η1 ) + εδ2 (t), ηn+1 (0) = en+1 (0),
 
y(t) = x1 (t) = v(t).
(20)
The corresponding NLESO (4) becomes where

1858
Preprints of the 18th IFAC World Congress
Milano (Italy) August 28 - September 2, 2011

δ1 (t) = f (t, x1 (εt), . . . , xn (εt)) − f (t, x̂1 (εt), . . . , x̂n (εt)), Matlab to solve the Lyapunov equation P E + E ⊤ P =
δ2 (t) = ẇ(εt). −I, to find that the eigenvalues of P satisfy λmax (P ) ≈
2.3230 < π. Let V, W : Rn+1 → R be defined as
(21) 2
V (η) = hP η, ηi, W (η) = ¯−hη, ¯¯ Then λmin (P )kηk ≤
ηi.
Theorem 3.2. [Modified extended state observer] In ¯
¯
¯ ∂V ¯¯
¯ ¯
addition
¯ to ¯ the conditions in Assumption (H3), we assume kV (η)k ≤ λmax (P )kηk2 , ¯¯¯¯ ≤ 2λmax (P )kηk, and
¯ ∂V ¯ ∂η ¯¯
that ¯ ∂y ¯ ≤ αkyk, αL < λ3 , where L is the Lipschitz ∂V ∂V ∂V
n
(η2 − 6η1 ) + (η3 − 11η1 ) − 6η1 = −W (η). Now
constant of f : ∂η1 ∂η2 ∂η1
|f (t, x1 , . . . , xn ) − f (t, y1 , . . . , yn )| ≤ Lkx − yk, (22) f (x1 , x2 ) = sin(x1 (t))+sin(x
4π
2 (t))
, and we find that L = 2π 1
.
∀ t ≥ 0, x = (x1 , x2 , . . . , xn )⊤ , y = (y1 , y2 , . . . , yn )⊤ ∈ Rn . 1
Hence Lλmax (P ) < 2 . Therefore, for any bounded control
u and bounded disturbance w, ẇ (let’s say the Pp finite super-
If ẇ is bounded in (2), then
position of sinusoidal disturbance w(t) = i=1 ai sin bi t),
(i). For every positive constant a, by Theorem 3.2, for any a > 0 limε→0 |xi (t) − x̂i (t)| =
lim |xi (t) − x̂i (t)| = 0 uniformly in t ∈ [a, ∞) 0 uniformly for t ∈ [a, ∞) and lim |xi (t) − x̂i (t)| ≤
ε→0 t→∞
O εn+2−i . Once again, we use the same data presented
¡ ¢
lim |xi (t) − x̂i (t)| ≤ O εn+2−i ,
¡ ¢
(ii). in (16) to simulate Example 3.1. The results are plotted
t→∞
in Fig. 2. It is seen from Fig.2 that the modified extended
where xi , x̂i denote the solutions of (2), (19) respectively, state observer (28) tracks very satisfactorily the state of
i = 1, 2, . . . , n + 1, xn+1 = w. system (27).
Proof. Finding the derivative of V (η(t)) with respect to t The last case is more special that the system function f
along the solution η(t) of system (20) yields is known and the external w = 0. In this case, system
√
d λ3 − αL λ1 p (2) becomes a special deterministic nonlinear system. To
V (η(t)) ≤ − V (η) + εM β V (η). (23) design the observer, we make the following assumption.
dt λ2 λ1
It follows that √ Assumption (H5). f is locally Lipschitz continuous, and
dp λ3 − αL p λ1 εM β ∂
n
∂ ∂f (t, x)
V (η(t)) ≤ − V (η(t)) + . (24) h(t, x) = f (t, x) +
X
xi+1 (t) f (t, x) + u(t).
dt 2λ2 2λ1 ∂t ∂x ∂xn
i=1 i
This together with Assumption (H3) gives
(29)
εM β t − λ32λ
p
λ1 V (η(0)) − λ32λ
Z
is globally Lipschitz continuous in x = (x1 , x2 , . . . , xn )
−αL
t
−αL
(t−s)
kηk ≤ e 2 + e 1 ds.
λ1 2λ1 0 uniformly for t ∈ (0, ∞), where xn+1 (t) = f (t, x(t)).
(25)
By (3), it follows that The state observer for (2) now is designed as following:
 µ ¶
"p
˙ n−1 y(t) − x̂1 (t)
λ1 V (η(0)) − (λ32λ−αL)t x̂1 (t) = x̂2 (t) + ε g1 ,


|ei (t)| ≤ εn+1−i εn

e 2ε


λ1 

 µ ¶
(26) ˙ 2 (t) = x̂3 (t) + εn−2 g2 y(t) − x̂1 (t) ,


x̂
Z t # 

εM β ε − λ32λ εn
−αL

(t/ε−s)

+ e ds .

1 

2λ1 0 
..
. (30)
The (i) and (ii) of Theorem 3.2 follow from (26). The proof 

 µ ¶
is complete. 2 

˙ y(t) − x̂1 (t)
 x̂n (t) = x̂n+1 (t) + gn + u(t),


Example 3.1. For the system following εn



 ẋ1 (t) = x2 (t),
 
 µ ¶
 x̂˙ n+1 (t) = h(t, x̂) + 1 gn+1 y − x̂1 (t) .



(27)
 ẋ2 (t) = sin(x1 (t)) + sin(x2 (t)) + w(t) + u(t), ε εn

4π Theorem 3.3. Under assumptions (H3) and (H5), it has
where w is the external disturbance, we design the corre-
sponding modified linear extended state observer as: (i). There exists an ε0 > 0 such that for any ε ∈ (0, ε0 )

6 lim |xi (t) − x̂i (t)| = 0,

 x̂˙ 1 (t) = x̂2 (t) + (y(t) − x̂1 (t)), t→∞

 ε
(ii). For any a > 0,


 11
 x̂˙ 2 (t) = x̂3 (t) + 2 (y(t) − x̂1 (t))



ε lim |xi (t) − x̂i (t)| = 0 uniformly in t ∈ [a, ∞)
(28) ε→0
 sin(x̂1 (t)) + sin(x̂2 (t)) where xi , x̂i are the solutions of (2), (30) respectively,
 + + u(t),
i = 1, 2, . . . , n + 1, xn+1 (t) = f (t, x(t)).

4π




 x̂˙ 3 (t) = 6 (y(t) − x̂1 (t)).


Duo to page limitation, we omit the proof. 2

ε3
−6 1 0
à !
4. CONCLUDING REMARK
For this example, the associated matrix E = −11 0 1
−6 0 0 In this paper, the convergence of various high gain non-
has three eigenvalues {−1, −2, −3}, so it is Hurwitz. Use linear extended state observers for a kind of nonlinear

1859
Preprints of the 18th IFAC World Congress
Milano (Italy) August 28 - September 2, 2011

systems with uncertainty both from internal and external S. Oh and H.K. Khalil. Nonlinear output-feedback track-
is proved. The high gain is used to eliminate the influence ing using high-gain observer and variable structure con-
of uncertainties. Examples and numerical analysis are fol- trol. Automatica, 33:1845-1856, 1997.
lowed to illustrate the effectiveness of the corresponding J. Slotine, J. Hedrick, and E. Misawa. On sliding observers
observers. Due to high gain and the difference between the for nonlinear systems. ASME J. Dynam. Sys. Measur.
initial states of the tracked system and the observer, the & Contr., 109:245-252, 1987.
peaking phenomena observed in Freidovich, Khalil [2008] B.L. Walcott, M.J. Corless, and S.H. Zak. Comparative
in the beginning of the time interval is also observed from study of non-linear state-observation technique. Int. J.
the numerical analysis. Contr., 45:2109-2132, 1987.
B. L. Walcott and S.H. Zak. State observation of nonlinear
uncertain dynamical systems. IEEE Trans. Auto.
REFERENCES Contr., 32:166-170, 1987.
X. X. Yang and Y. Huang. Capability of extended state
A. N. Atass and H.K. Khalil. Separation results for the observer for estimating uncertainties. American Control
stabilization of nonlinear systems using different high- Conferencce, 3700-3705, 2009.
gain observer designs. Sys. & Contr. Lett., 39:183–191, Q. Zheng, L.Gao, and Z. Gao. On stability analysis
2000. of active disturbance rejection control for nonlinear
G. Besancon. Nonlinear Observers and Applications. time-varying plants with unknow dynamics. IEEE
Springer Verlag, New York, 2007. Conference on Decision and Control, 3501-3506, 2007.
M.J. Corless and J. Tu. State/input estimation for a class Q. Zheng, L. Gong, D.H. Lee, and Z. Gao. Active
of uncertain systems. Automatica, 34:757–764, 1998. disturbance rejection control for MEMS gyroscopes.
M. Darouach, M. Zasadzinski, and J.S. Xu. Full-order American Control Conference, 4425-4430, 2008.
observers for linear systems with uncertain inputs. IEEE Q. Zheng, L.L. Dong, and Z. Gao. Control and rota-
Trans. Auto. Contr., 39:606–609, 1994. tion rate estimation of vibrational MEMS gyroscopes.
F. Esfandiari and H.K. Khalil. Output feedback stabiliza- IEEE Multi-conference on Systems and Control, 118-
tion of fully linearizable systems. Int. J. Contr., 56: 123, 2007.
1007–1037, 1992. Q. Zheng and Z.Gao. On applications of active disturbance
L. B. Freidovich and H. K. Khalil. Performance recovery rejection control. Chinese Control Conference, 6095-
of feedback-linearization-based Designs. IEEE. Trans. 6100, 2010.
Auto. Contr., 53:2324–2334, 2008. Q. Zheng, Z. Chen, and Z. Gao. A dynamic decoupling
Z. Gao. Scaling and bandwith-parameterization based control approach and tts applications to chemical pro-
controller tuning. American Control Conference, 4989– cesses. American Control Conference, 5176-5181, 2007.
4996, 2003. W. Zhou and Z. Gao. An active disturbance rejection
V. Gourshankar, P. Kudva, and K. Ramar. Reduced or- approach to tension and velocity regulations in Web
der observer for multivariable systems with inaccessible processing lines. IEEE Multi-conference on Systems
disturbance inputs. Int. J. Contr., 25:311-319, 1977. and Control, 842-848, 2007.
J. Q. Han. Control theory: Model approach or control W. Zhou, S. Shao, and Z. Gao. A stability study of
approach. J. Sys. Sci. & Math. Scis., 9(4):328–335, the active disturbance rejection control problem by a
1989(in Chinese). singular perturbation approach. Appl. Math. Sci., 3:
J. Q. Han. A class of extended state observers for uncertain 491-508, 2009.
systems. Control & Decision, 10(1):85-88, 1995(in
Chinese). 2

1.5
4

3
4

J. Q. Han. From PID to active disturbance rejection 1

0.5
2

1
2

control. IEEE Trans. Ind. Electron., 56:900-906, 2009. 0

0

−1
0

−1

Y. Hou, Z. Gao, F. Jiang, and B.T. Boulter. Active −0.5

−1
−2

−3
−2

−3

disturbance rejection control for web tension regulation. −1.5

0 2 4 6 8 10
−4
0 2 4 6 8 10
−4
0 2 4 6 8 10

IEEE Conference on Decision and Control , 4974-4979, (a) (b) (c)

2001. x1 (green), x̂1 ( red), x2 (green), x̂2 (red), x3 (green), x̂3 ( red),
Y. Huang and J. Q. Han. A new synthesis method for x1 − x̂1 (blue) x2 − x̂2 (blue) x3 − x̂3 (blue)
uncertain systems-the self-stable region approach. In-
ternational Journal of Systems Science, 30:33-38, 1999. Fig. 1. NLESO (12) for system (11)
H. K. Khalil. High-gain observers in nonlinear feedback
control. Int. Conf. Control, Automation and Systems, 30 4 4

Oct.14-17, Korea, 2008.

25
3

2
20
2

A. Koshkouei and A. Zinober. Sliding mode controller

1
15
1 0
10

observer design for SISO linear systems. Int. J. Sys. 5

0
−1

−2

Sci., 29:1363-1373, 1998.

−1
0 −3

−5 −2 −4
0 2 4 6 8 10 0 1 2 3 4 5 6 7 8 9 10 0 2 4 6 8 10

P. Kudva, N. Viswanadham, and A. Ramakrishna. Ob-

servers for linear systems with unknown inputs. IEEE (a) (b) (c)
x1 (green), x̂1 (red), x2 (green), x̂2 (red), x3 (green), x̂3 (red),
Trans. Auto. Contr., 25:113-115, 1980.
x1 − x̂1 (blue) x2 − x̂2 (blue) x3 − x̂3 (blue)
R. Miklosovic and Z. Gao. A dynamic decoupling
method for controlling high performance turbofan en-
gines. Proc. of the 16th IFAC World Congress, July Fig. 2. The modified extended state observer (28) for
4-8, 2005, Czech Republic. system (27)

1860