2011 American Control Conference
on O'Farrell Street, San Francisco, CA, USA
June 29 - July 01, 2011
Comparison of the DOB Based Control, A Special Kind of PID Control
and ADRC
Wenchao Xue and Yi Huang
Abstract— In this paper, the methods for estimating uncer-
tainties in the disturbance observer (DOB), a special kind
of PID (SPID) and the active disturbance rejection control
(ADRC) are discussed. The stability of ADRC is proven for
a class of SISO nonlinear systems with unknown dynamics
and disturbance. The comparison study and the analysis for
ADRC show that under some conditions, DOB and SPID can
be generalized for nonlinear systems with mixed internal and
external uncertainties.
Index Terms— disturbance observer (DOB), A special kind
of PID, active disturbance rejection control (ADRC).
I. INTRODUCTION
Usually, one of the main objects of control is to deal with
the uncertainties including internal (parameter or unmodeled
dynamics) uncertainties and external (disturbance) uncer-
tainties. The idea of the invariant principle [1] gives some
suggestions for the problem of controlling uncertain systems:
the uncertainties causing changes in the controlled variable
can be used to generate an activating signal which will tend
to cancel the effect of the same uncertainties, no matter
they are internal or external. Obviously, the activating signal
to attenuate uncertainties can be easily constructed if the
uncertainties are measurable. However, most uncertainties
are not measurable. Hence how to estimate uncertainties by
the known information, for example, the control input and
the output of system, become a significant problem. In the
past years, lots of approaches have been proposed to estimate
uncertainties from the input-output data, such as the distur-
bance accommodation control (DAC) [2], the unknown input
disturbance observer (UIDO) [3], the disturbance observer
(DOB)[4]-[9], some special PID (SPID) [10]-[11] and the
active disturbance rejection control (ADRC) [12]-[16]. DAC,
UIDO and DOB are initially proposed to deal with external
disturbance for linear time-invariant systems. The survey in
[4] addressed that by the transfer function from disturbance
to its estimation, DAC and UIDO can be viewed as a special
form of DOB. SPID was proposed to deal with internal
dynamics uncertainty for some kinds of linear and nonlinear
systems. The ADRC was proposed by Han to deal with
the nonlinear systems with mixed uncertain dynamics and
disturbances [12]-[13]. The common idea of these methods is
This work was supported by NSFC60821091, NSFC60736022, KJCX3-
SYW-S01.
Wenchao Xue is with the Key Laboratory of Systems
and Control, Academy of Mathematics and Systems Science,
Chinese Academy of Sciences, Beijing 100190, P. R. China,
xuewensuper1985@yahoo.com.cn
Yi Huang is with the Key Laboratory of Systems and Control, Academy
of Mathematics and Systems Science, Chinese Academy of Sciences,
Beijing 100190, P. R. China, yhuang@amss.ac.cn
978-1-4577-0081-1/11/$26.00 ©2011 AACC
to divide the process of controller design into two parts: one
is to compensate for the uncertainties, which is reconstructed
by the input-output data via certain kind of observer, the
other is to realize the desired performance (tracking or
regulating) for the compensated system.
ADRC’s framework does not set strict mathematical con-
straints on the uncertainties to be estimated. There have been
lots of application researches [17]-[22] and some stability
analysis [23]-[25] since ADRC was introduced. In this paper,
the methods for estimating uncertainties in DOB, SPID and
ADRC are compared. The comparison illustrates that for a
simplified benchmark plant, DOB, SPID and ADRC employ
similar structure for estimating uncertainties. Furthermore,
the paper proves that ADRC can stabilize a general class of
nonlinear system with unknown dynamics and external dis-
turbance. Hence, the frame of ADRC suggests the possibility
to generalize DOB and SPID for estimating both internal
and external uncertainties for nonlinear systems under some
conditions. Although this statement has been indicated in
some literature [6]-[9], the rigorous theoretical analysis has
not been given.
The rest of paper is organized as follows. In Section II,
the basic ideas and some theoretical results for DOB, SPID
and ADRC are briefly introduced. Then these controllers
are compared in Section III. Some simulations are given
in Section IV. The proof of Theorem 1 which presents the
performance and stability of ADRC is given in Section V,
and Section VI are conclusions.
II. THE BASIC IDEAS OF DOB, SPID AND ADRC
Consider the following SISO nonlinear system
(
ẋ = Ax + B( f (x,t) + g(x,t)u
,t ≥ t0 (1)
y = x1
     
x1 0 1 0
x2   0  ...
where x =  ... , A = 
   , B =   , f (x,t),
... 1 0
xn 0 1
g(x,t) can be linear or nonlinear, time-varying or time-
invariant functions which may contain unknown dynamics
and external disturbance, u is the control input and y is the
measured output.
Next the basic ideas and some stability results of DOB,
SPID and ADRC will be introduced.
4373
A. Disturbance Observer (DOB) where k1 + k2 s + ... + kn sn−1 + sn is a stable polynomial, and
Usually, the Disturbance Observer (DOB) is designed for dˆS is given by
the following LTI system with an unknown disturbance d(t) 
n−1
ˆS = ξ + ∑ qi xi+1

d
( 

ẋ = Ax + B (a1 x1 + ... + an xn + b(d(t) + u))


i=0
,t ≥ t0 (2) (9)
y = x1  n−1 n−2
ξ̇ = −q ξ − qn−1 ∑ i i+1
q x − ∑ i i+2 n−1
q x − q u


 n−1
which can be seen as a special case of (1) with

i=0 i=0

f (x,t) = a1 x1 + ... + an xn + bd(t), g(x,t) = b. where qi , i = 0, ..., n − 2 are arbitrary constant and qn−1 =
σ (g1 (x))µ with σ (g1 (x)) being the sign of g1 (x) and µ > 0
The open-loop transfer function for (2) is: being a suitable constant.
y b The stability analysis of the system (5) with the control
= , P(s).
u sn − an sn−1 − ... − a1 (8)-(9) can be shown by the following lemma.
Let Pn (s) denote the nominal system and uD denote the Lemma 1 [11] . Under the assumptions:
DOB based control. The following DOB can be designed to A.1 x1 , ..., xn can be available;
estimate d(t) by the information of Pn (s), the control input A.2 f1 (x), g1 (x) are pth differentiable, where p > 0 being a
uD and the output y. suitable integer;
A.3 g1 (x) satisfies g1 (x) 6= 0, |g1 (x)| ≥ b > 0 where b is a
dˆD = Q(s)(Pn−1 (s)y − uD ) (3) constant and σ (g1 (x)) is known in the domain of interest,
there exists a constant µ ∗ > 0 such that if µ > µ ∗ , then the
where Q(s) is a low-pass filter. To guarantee the realization
closed-loop system (5), (8)-(9) is asymptotically stable.
of (3), Q(s) should satisfy ∂ Q(s) ≥ ∂ Pn (s), where ∂ Q(s) (or
Lemma 1 shows that the internal dynamics uncertainty
∂ Pn (s)) stands for the order of Q(s) (or P(s)). Then uD can
dS (·) satisfying A.2-A.3 can be estimated by dˆS (·) of the
be designed as
observer (9).
uD = −dˆD + un (4)
where un = Kn (s)y can stabilize the nominal system Pn (s). C. Active Disturbance Rejection Control (ADRC)
In most research on DOB, Pn (s) is assumed to be equal to The key of ADRC is to design an extended state observer
P(s) and DOB is used mainly for estimating the external (ESO) to estimate not only states but also the ”total dis-
disturbance. However, in practice, the parameters in P(s) turbance” which contains internal uncertain dynamics and
usually contain some uncertainty. In some literature [6]-[9], external disturbance [12]-[16].
it’s declared that the estimation of the plant uncertainties can Let ĝ(t) be the estimation of g(x,t), then the following
be included into dˆD , and [7] analyzed the case of Pn 6= P for ESO, which employs a linear structure, is designed for (1),
some special kinds of system. However, in general case, how 
˙
to choose the nominal system Pn to assure the closed-loop x̂1 = −β1 (x̂1 − x1 ) + x̂2


stability is still an open problem. x̂˙2 = −β2 (x̂1 − x1 ) + x̂3



B. A Special Kind of PID (SPID) ... (10)
x̂˙n = −βn (x̂1 − x1 ) + x̂n+1 + ĝ(t)u


[10]-[11] proposed a special PID control for a class of 

x̂˙

minimal-phase system. To simplify the analysis, we consider n+1 = −βn+1 (x̂1 − x1 )
the following SISO system without zero dynamics:
( where the parameters βi are designed to satisfy
ẋ = Ax + B ( f1 (x) + g1 (x)u)
,t ≥ t0 (5) β̄i
y = x1 βi = , i ∈ n + 1, ε ∈ (0, ∞).
εi
which can also be a special form of (1) with and
f (x,t) = f1 (x), g(x,t) = g1 (x) n+1 n+1
L(s) , sn+1 + ∑ β̄i sn+1−i = Π (s + λi ), Re(λi ) < 0 (11)
Rewrite the nth equation of (5) as i=1 i=1

ẋn = dS (x, u) + u (6) is the characteristic polynomial of ESO (10) when ε = 1.

The bandwidth of ESO (10) can be adjusted by ε .
where When ESO (10) is well tuned, its output x̂ = [x̂1 , x̂2 , ...x̂n ]T
dS (·) = f1 (x) + (g1 (x) − 1)u (7) and x̂n+1 can be used as the estimation of x and the total
is an unknown term in (6). If x can be available, the influence disturbance
of dS (·) can be rejected by the following PID controller [10]- xn+1 , f (x,t) + (g(x,t) − ĝ(t))u. (12)
[11]:
n
uS = − ∑ ki xi − dˆS (8) For ESO (10), when ε is small, the peaking phenomenon
i=1 will happen if the observer’s initial errors êi 6= 0(i ∈ n). In

4374
this paper, the following scheme is designed in the initial III. A COMPARISON FOR DOB, SPID AND ADRC
phase of the control to avoid peaking: From the above introduction, DOB faces the problem that
 0, t0 6 t < tu how to choose Pn (s), Q(s) to guarantee the stability of closed-


u = −K T x̂ − x̂n+1 (13) loop system, and SPID can only be implemented under the
 , t ≥ tu assumption that x1 , ..., xn are all available. ADRC provides a
ĝ(t)
most systematic frame to estimate both external disturbance
where K = [k1 k2 .. kn ]T and tu is a constant to be and internal uncertainty.
designed. Next using the system (2) with the following assumptions:
Under the following assumptions: A.6 x1 , ..., xn can all be available;
A.4 ∂ f∂(x,t) ∂ f (x,t) ∂ g(x,t) ∂ g(x,t)
x , ∂t , ∂x , ∂t are locally Lipschitz with A.7 σ (b) is known and a1 , .., an , d(t), b are unknown,
respect to (x,t), and ∀x ∈ {x|kxk ≤ ρ }, there is as a benchmark, we will give a further comparison for DOB,
∂ f (x,t) ∂ f (x,t) SPID and ADRC. Since the discussion for σ (b) = 1 and
| f (x,t)| ≤ F1 (ρ ), | | ≤ F2 (ρ ), | | ≤ F3 (ρ ),
∂x ∂t σ (b) = −1 is similar, let σ (b) = 1 in the following analysis.
|g(x,t)| ≤ F4 (ρ ), |
∂ g(x,t)
| ≤ F5 (ρ ), |
∂ g(x,t)
| ≤ F6 (ρ ),
Let k1 + k2 s + ... + kn sn−1 + sn be a stable polynomial. Set
∂x ∂t Pn (s) = s1n . In this setting, the disturbance to be handed by
where Fi , i ∈ 6 are known functions dependent on ρ ; DOB will be
A.5
|g(x,t)| ≥ α1 > 0, (14) dD (·) = a1 x1 + ... + an xn + bd(t) + (b − 1)u. (20)

where α1 is constant, Form (3)-(4), the DOB based control can be given by
the performance of the closed-loop system (1), (10)-(13) can uD = −(k1 + k2 s + ... + kn sn−1 )y − dˆD (21)
be presented by the following theorem. ˆ −1
Theorem 1. Consider the system (1) with the control (10)- dD = Q(s)(Pn (s)y − uD ) = Q(s)(sxn − uD ) (22)
(13) under the assumption A.4-A.5. Design ĝ(t) to satisfy where dˆD is utilized as an estimation for the disturbance
˙ dD (·). To guarantee the realization of (22), Q(s) is a low-
|ĝ(·)| ≤ α2 , (15)
pass filter satisfying ∂ Q(s) ≥ 1. However, the stability of the
β̄n+1 DOB based control (21)-(22) by setting Pn (s) = s1n has not
k k∞ · |Λ(·)| ≤ αΛ < 1 (16)
been studied for the general case.
L(s)
Next consider the SPID design. According to (8)-(9), the
where Λ(·) = g(·)− ĝ(·)
ĝ(·) and α2 , αΛ are positives. Then there SPID based control for system (2) is equal to:
exists ε > 0 such that for ε ∈ (0, ε ∗ ]
∗
n
uS = − ∑ ki xi − dˆS (23)
1). kx(t) − x∗ (t)k ≤ O(|ε lnε |), t ∈ [t0 , ∞) i=1
qn−1 µ
2). lim kx(t)k ≤ O(ε ), dˆS = (sxn − uS ) = (sxn − uS ) (24)
t→∞ s + qn−1 s+µ
where x∗ (t) is the trajectory of the reference system where the uncertain term to be approximated by dˆS is
(
ẋ∗ (t) = f (x∗ ,t), t ∈ [t0 ,tu )
(17) dS (·) = a1 x1 + ... + an xn + bd(t) + (b − 1)u (25)
∗ ∗
ẋ (t) = Ac x , t ∈ [tu , ∞)
which is the same as dD (·) in (20). However, Lemma 1 dose
with x(t0 ) = x∗ (t0 ) and Ac = A + BK T being Hurwitz. The not cover this case due to dS (·) is time-varying.
proof of Theorem 1 will be given in Section V. Now, let us analyze ADRC. Since x are available, the ESO
The result 1) of Theorem 1 shows that the performance (18) with ĝ = 1 is used. Considering the peaking will not
of closed-loop system can be controlled to be close to the happen, the ADRC (18) and (19) for the system (2) can be
reference system (17) by tuning ε . And the result 2) of simplified as:
Theorem 1 means the ultimate bound of the states can be n
small enough by tuning ε . uA = − ∑ ki xi − dˆA (26)
Remark 1. If the state x is available, then a reduced order i=1
ESO can be designed as follows to estimate xn+1 β2
dˆA = (sxn − uA ) (27)
s2 + β 1 s + β2
(
x̂˙n = −βn (x̂n − xn ) + x̂n+1 + ĝ(t)u
(18)
x̂˙n+1 = −βn+1 (x̂n − xn ) where the uncertain term to be approximated by dˆA is
Furthermore, the peaking phenomenon will not happen in dA (·) = a1 x1 + ... + an xn + bd(t) + (b − 1)u (28)
ESO (18). Hence, the control can be simply designed as
which equals to dD (·) and dS (·) in (20) and (25).
−K T x − x̂n+1 (21)-(27) clearly show that for this simple benchmark
u= . (19)
ĝ(t) plant, the difference in DOB, SPID and ADRC lies in the

4375
 The curve of x1 (Case 1)
difference in the observers which are designed for the same 1.5
uncertainty dD = dS = dA : 1

dˆD = Q(s)(sxn − uD )


1
0.5

x

 dˆ = µ (sx − u )



0
S n S
s+µ (29)
−0.5
β

 0 2 4 6 8 10 12 14 16
2
 dˆA =

 (sxn − uA )
s2 + β1 s + β2 1
dD
Then some meaningful conclusions can be obtained from 0.5 dˆD

dˆD and dD
the above comparison: 0
1). Although the DOB based control (21) lacks theoretical
−0.5
base, Theorem 1 provides one for DOB when the nominal
model Pn is simply chosen to be a chain of integrator −1
0 2 4 6 8 10 12 14 16
and Q(s) = s2 +ββ2s+β . Furthermore, Theorem 1 suggests the time
1 2
possibility of generalizing DOB to the system (1) where Fig. 1. The simulation result for DOB based control (36)-(37)
f (x,t), g(x,t) are nonlinear and time-varying as follows:
if y is available and ẏ, ..., y(n−1) are known, the following
The curve of x1 (Case 2)
generalized DOB based control can be developed: 1.5
ADRC
dˆD (s) = Q(s)(Pn−1 (s)y − Ũ(s)), 1 SPID

1 βn+1 (30)

1
0.5

x
Pn (s) = n , Q(s) = n+1 ,
s s + β1 sn + ... + βn+1 0

where Ũ(s) is the Laplace transform of ĝ(t)u(t), and −0.5

0 2 4 6 8 10 12 14 d 16
A
1
u(t) = (−dˆD (t) − k1 y − k2 ẏ − ... − kn y(n−1) ). (31) 2
x̂3
dA , x̂3 , dS and dˆS

ĝ(t) dS
1
2) Theorem 1 also presents the enlightenment of general- dˆS
0
izing SPID to the system (1).
−1
2.1) If x is available, SPID can be used for the system (1).
−2
However, if only y is available and the the system order is
0 2 4 6 8 10 12 14 16
greater than 2, then SPID no longer works. time
2.2) [23] proved that ESO (10) can be independently used
as a filter to estimate uncertainty if the uncertainty or its Fig. 2. The simulation result for SPID (38) (40), and ADRC (39) (41)
derivative is bounded. However, the observer (9) only works
in the closed-loop system and can not be used as a filter
when σ (g(·)) = −1. However, when x is available, a first- discussion in Section III, the following generalized DOB and
order ESO can be designed as follows: SPID can be utilized to deal with the system (33).
Case 1). Only y = x1 is available.
(
ż = −β z − β 2 xn − β ĝ(t)u
(32) Set Pn (s) = s1n , ĝ(t) = −1, from (30), the following gener-
x̂n+1 = β xn + z
alize DOB based control is designed,
where β > 0 and σ (ĝ) = σ (g). (32) can be viewed as another β3
design of SPID which can be used as an independent filter. dˆD (s) = (s2 y + u) (35)
s3 + β1 s2 + β2 s + β3
IV. S IMULATION
where β1 = 3/ε , β2 = 3/ε 2 , β3 = 1/ε 3 , ε = 0.1 and dˆD is
Consider the following nonlinear time-varying system: an estimation of the total disturbance dD = f (x1 , x2 ,t) +
(g(x1 , x2 ,t) + 1)u. Rewrite (35) in the form of state space,


 ẋ1 = x2

ẋ2 = f (x1 , x2 ,t) + b(x1 , x2 ,t)u

˙

(33) x̂1 = −β1 (x̂1 − x1 ) + x̂2

y = x1 ˙x̂2 = −β2 (x̂1 − x1 ) + dˆD − u .

 (36)
x1 (0) = 1, x2 (0) = 0

 ˙ˆ

dD = −β3 (x̂1 − x1 )
where the uncertainties are set to be
( Obviously, (36) is equal to the 3rd-order ESO, and x̂1 , x̂2
f (x1 , x2 ,t) = sin(0.25π t)x1 + x22 + 0.2cos(0.2π t), are the estimation of x1 , x2 . Choose the initial condition as
(34)
g(x1 , x2 ,t) = −1 + 0.2sin(0.2π x2 ). x̂1 (0) = 0, x̂2 (0) = 0, dˆD (t0 ) = 0 and take K = [2 1], then we
can get the following control based on DOB (36)
The control object is x1 → 0. Although neither DOB nor
SPID in its old frame can handle this problem, from the u = dˆD + x1 + 2x̂2 . (37)
4376
 ∂f ∂f
· ¸ · ¸
0 K
The fig. 1 gives the simulation results for system (33) with η1 = (Ax + B f ) + , B2 = T , Ke = ,
∂x ∂t B 1
the control (36)-(37). ∂f ∂ f 1 ∂g
µ
∂g ˙
¶³ ´
Case 2). x1 , x2 are available. η2 = Ac x + − Ac x + − ĝ − Λ(·)ĝ K T X + f
˙
∂x ∂t g ∂x ∂t
Set q1 = −µ , q0 = 0, µ > 0 in (8)-(9) and ĝ(t) = −1 in T
− Λ(·)K Ac x,
(32), the following SPID and 1st-order ESO are designed 1 ∂g
µ
∂g ˙
¶
for the system (33), respectively: η3 = (Ac x − BK T x − B f ) + − ĝ − Λ(·)ĝ˙ KeT T1 (ε )
g ∂x ∂t
∂f
+ BK T1 (ε ) − Λ(·)K BKe T1 (ε ) − Λ(·)K T T2 (ε ),
T T T
(
dˆS = ξ − µ x2 , ∂x e
SPID: (38)
ξ̇ = µξ − µ 2 x2 + µ u, 1 ∂g
η4 = BT1 (ε )Ke KeT T1 (ε ),
( g ∂x
ż = −β z − β 2 x2 + β u, 
β̄1 ε n−1 −ε n−1 0 ...

1st-order ESO: (39)
x̂3 = β x2 + z, β̄ ε n−2
2 0 −ε n−2 ...
T2 (ε ) = 
 ...
.
... ... ...
β̄n 0 ... 1
where dˆS and x̂3 are the estimations of the total disturbance
dS = f (·) + (g(·) − 1)u and dA = f (·) + (g(·) + 1), respec- Suppose the initial condition be
tively. The controls based on (38) and (39) are taken as kx(t0 )k = ρ0 , kÊe (t0 )k = ρ ∗ .

SPID: u = −x1 − 2x2 − dˆS , (40) Design tu such that

ADRC: u = x1 + 2x2 + x̂3 . (41) x(t) ∈ {x|kxk ≤ ρ1 }, t ∈ [t0 ,tu ), (47)
where ρ1 > ρ0 is the scope accepted in practice.
Take µ = 10, β = 10 and dˆS (0) = 0, x̂3 (0) = 0. The fig. 2 is Using A.4, we get for ∀x ∈ {x|kxk ≤ ρ1 }
the simulation results for system (33) with SPID and ADRC.
Fig. 1 and Fig. 2 illustrate both generalized DOB and SPID |η1 (·)| ≤ γ1 (ρ1 )
can stabilize the nonlinear uncertain system (33), which where γ1 (ρ1 ) , F2 (ρ1 )(kAkρ + F1 (ρ1 )) + F3 (ρ1 ). And simi-
validates our theoretical results. larly
|η2 (·) + η3 (·)ξe + ξeT η4 (·)ξe | ≤
(48)
V. T HE P ROOF OF T HEOREM 1 γ2 (ρ1 , ε ) + γ3 (ρ1 , ε )kξe k + γ4 (ρ1 , ε )kξe k2
where γi , i = 2, 3, 4 are positives dependent on (ρ1 , ε ) and are
Before the proof of Theorem 1, we introduce a transfor- analytical with respect to ε .
mation As for the properties of Ac , Ãe , Ae , we have the following
Lemmas.
Eˆe = T1 (ε )ξe (42) Lemma 2.[26] If Ac , Ãe are Hurwitz, then there exist
      positive matrices P1 , P̃2 and positives c11 , c12 , c̃21 , c̃22 such
ê1 x1 − x̂1 ξ1 that
 ...  ...  ... 
where Êe =   =   , ξe =   and ATc P1 + P1 Ac = −I, c11 I ≤ P1 ≤ c12 I,
 
ên+1 xn+1 − x̂n+1 ξn+1 ÃTe P̃2 + P̃2 Ãe = −I, c̃21 I ≤ P̃2 ≤ c̃22 I.
T1 (ε ) = diag{ε n , ..., ε , 1}. Using the control (10)-(13) and
Lemma 3.[27]-[28] If Ãe is Hurwitz and the condition (16)
the transformation (42), we get the closed-loop system of
is satisfied, then there exist positive matrix P2 , and positives
(x, ξe ) for t < tu and t ≥ tu , respectively:
c0 , c21 , c22 such that
ẋ = Ax + B f (x,t), t < tu (43) ATe P2 + P2 Ae ≤ −c0 I, c21 I ≤ P2 ≤ c22 I.
1
ξ̇e = Ãe ξe + B2 η1 (x,t), t < tu (44) Proof of Theorem 1.
ε Define V1 = xT P1 x, Ṽ2 = ξeT P̃2 ξe ,V2 = ξeT P2 ξe and E = x −
ẋ = Ac x + BKeT ξe , t ≥ tu (45) x∗ . According to (17) and (43), if x(t0 ) = x∗ (t0 ), then
1
ξ̇e = Ae ξe + B2 η2 (x,t) (46) E(t) = 0,t ∈ [t0 ,tu ).
ε¡
+B2 η3 (x,t, ε )ξe + ξeT η4 (x,t, ε )ξe ,t ≥ tu
¢
When t ≥ tu , since
Ė = Ac E + BKe ξe , E(tu ) = 0, (49)
where
    there is
−β̄1 1 −β̄1 1 ³ ´
 ...   ...  V̇1 (E) ≤ −kEk kEk − 2kP1 BKeT kkξe k . (50)
Ãe =  ,A =  ,
 −β̄n 1 e  −β̄n 1
−β̄n+1 −β̄n+1 (1 + Λ) Next the property of ξe will be analyzed by two steps.

4377
 Step 1: The bounds of ξe (t) in [t0 ,tu ) and x̂n+1 satisfies
We will illustrate ∃t p0 (ε ) ≤ tu such that when t ≥ t p0 , ξe (t)
can converge to a small region. kx̂n+1 (tu )k = k lim x̂n+1 (t)k
t→tu
By Lemma 2, the derivative of the Lyapunov function ζ0 (61)
= k lim− f (x,t) − lim− ξn+1 (t)k ≤ F1 (ρ1 ) + √ ε .
Ṽ2 (ξe ) along the trajectories of (44) will be t→tu t→tu c̃21
1
Ṽ˙2 (ξe ) ≤ − kξe k2 + γ5 (ρ1 ) kξe k, t ∈ [t0 ,tu ) (51) Form (58)-(61), there exists γ6 (ρ1 ) such that ∀ε ∈ (0, ε1 ]
ε
where γ5 = 2 kP2 B2 k γ1 (ρ1 ). kξe (tu )k = kξ (tu )k + kξn+1 (tu )k ≤ γ6 (ρ1 )
Then
d Ṽ˙2 (ξe ) 1 kξe k2 γ kξ e k Thus the initial condition (x(tu ), ξe (tu )) satisfies:
q
Ṽ2 (ξe ) = p ≤− p + p5
dt 2 Ṽ2 (ξe ) ε 2 Ṽ2 (ξe ) 2 Ṽ2 (ξe )
p kx(tu )k ≤ ρ1 , kξe (tu )k ≤ γ6 . (62)
1 Ṽ2 (ξe ) γ
≤− + √5 . (52) √
ε 2c̃22 2 c̃21 √
Define ρ2 , c12 max{ρ1 , 2kP1 BKeT k √cc2221 γ6 }. Next we will de-
According to Gronwall-Bellman inequality, we can obtain sign ε such that
³ ´
c̃22 γ 1
√
q q
− (t−t0 )
Ṽ2 (ξe (t)) ≤ √ 5 ε + Ṽ2 (ξe (t0 ))e 2ε c̃22
p p
(53) Ω2 = {(x, ξe )| V1 (x) ≤ ρ2 , V2 (ξe ) ≤ c22 γ6 }
c̃21

Since Êe (t0 ) = T1 (ε )ξe (t0 ), the second term of (53) satisfies is a positive invariant set for (45)-(46) in t ∈ [tu , ∞). The
proof can be achieved by the following 2 steps.
√ ρ ∗ − 2t−t
³ ´ ³ ´
− 2ε c̃1
q 0
(t−t0 ) n −1
Ṽ2 (ξe (t0 ))e 22 ≤ c̃22 kε T1 (ε )k n e ε c̃22 (54) i) Let
ε p p √
V1 (x) = ρ2 , V2 (ξe ) ≤ c22 γ6 , (63)
where ε n T1−1 (ε ) is analytical with respect to ε .
When t p0 (ε ) , max{t0 − 2c̃22 (n + 1) ε ln ε ,t0 }, there is then along the trajectories of (45)
³ ´
t−t
− 2ε c̃ 0
≤ ε n+1 ,t ≥ t p0 (ε ).
³ ´
e 22 (55) V̇1 (x) ≤ −kxk kxk − 2kP1 BKeT kkξe k
µ √ √ ¶ (64)
Since lim t p0 (ε ) = t0 , there exists ε1 such that ∀ε ∈ (0, ε1 ], 1 c12 c22
ε →0 ≤ − √ kxk ρ2 − 2kP1 BKeT k √ γ6 ≤ 0.
c12 c21
t p0 (ε ) ≤ tu .
ii) Let
Remark 2. If ε > 1, then t p0 (ε ) = t0 . Otherwise, t p0 (ε ) = p p √
t0 − 2c̃22 (n + 1) ε ln ε . V1 (x) ≤ ρ2 , V2 (ξe ) = c22 γ6 , (65)
From (53), (54) and (55), we can obtain
q then
Ṽ2 (ξe (t)) ≤ εζ0 (ρ1 , ρ ∗ ), t ∈ [t p0 ,tu ] (56)
c0 kξe k
V̇2 (ξe ) ≤ −kξe k + 2kξe kkP2 B2 k|η2 (·)|
³ √ ° n −1 ° ´ ε
where ζ0 = max c̃22√γ5c̃(ρ1 ) + c̃22 °°ε T1 (ε )° ρ ∗ .
°
+ 2kξe kkP2 B2 k|η3 (·)ξe + ξeT η4 (·)ξe |
ε ∈(0,ε1 ] 21
(66)
Combining (56) and (47), it can be concluded that ∀ε ∈ c0 kξe k
≤ −kξe k + 2kξe kkP2 B2 kγ2 (ρ2 , ε )
(0, ε1 ] ε
( + 2kξe kkP2 B2 k(γ3 (ρ2 , ε )kξe k + γ4 (ρ2 , ε )kξe k2 ).
kx(t)k ≤ ρ1 , t ∈ [t0 ,tu )
. (57)
kξe (t)k ≤ √ζc̃0 ε , t ∈ [t p0 ,tu ) By
p √
V2 (ξe ) = c22 γ6 , we can find ε2 such that ∀ε ∈ (0, ε2 ]
21

Step 2: The bounds of ξe (t) in [tu , ∞) V̇2 (ξe ) ≤ 0. (67)

According to A.4,
ζ0 Using (64) and (67), we can get Ω2 is a positive invariant
kx(tu )k ≤ ρ1 , kξ (tu )k ≤ √ ε . (58)
c̃21 set for (45)-(46).
£ ¤T Define
where ξ = ξ1 ... ξn . According to (12) and (13),
2c22
kξn+1 (tu )k = kxn+1 (tu ) − x̂n+1 (tu )k t p1 (ε ) , max{tu − (n + 1) ε ln ε ,tu }. (68)
c0
= k f (tu ) + Λ(tu )(−K x̂(tu ) − x̂n+1 (tu )) − x̂n+1 (tu )k (59)
= k f (tu ) − Λ(tu )K x̂(tu ) − (Λ(tu ) + 1)x̂n+1 (tu )k Similar to the deduction from (51)-(56), there exist ζ1 (ρ2 )
and ε3 ≤ ε2 such that for ∀ε ∈ (0, ε3 ]
where x̂(tu ) can be written as
p
x̂(tu ) = x(tu ) − diag{ε n , ..., ε }ξ (tu ) (60) V2 (ξe (t)) ≤ ζ1 ε , t ∈ [t p1 , ∞). (69)

4378
 Since ξe ∈ Ω2 and ξe can be constrained by (69), according [6] E. Schrijver and J.V. Dijk, ”Disturbance observers for rigid mechanical
to (50) and Gronwall-Bellman inequality, we have systems: Equivalence, stability, and design”, Journal of Dynamic
Systems, Measurement, and Control , Vol. 124, No. 4, 2002. pp. 539-
t 548.
√
Z c11 (t−τ )
V1 (E(t)) ≤ c12 kP1 BKeT k e− 2 kξe (τ )kd τ
p
[7] R. Bickel and M. Tomizuka, ”Passivity-Based Versus Disturbance
Observer Based Robot Control: Equivalence and Stability”, ASME J.
tu Dyn. Syst., Meas., Control, Vol. 121, 1999, pp. 41C47.
t p1 [8] T. Umeno and Y. Hori, ”Robust speed control of dc servomotors using
√
Z c11
(t−τ )
≤ c12 kP1 BKeT k e− 2 kξe (τ )kd τ modern two degrees-of-freedom controller design”, IEEE Trans. Ind.
Electron., Vol. 38, No. 5, 1991, pp. 363-368.
tu (70) [9] S. Kwon and W.K. Chung, ”A Discrete-Time Design and Analysis of
Zt Perturbation Observer for Motion Control Applications”, IEEE Trans.
√ c
− 11 (t−τ )
+ c12 kP1 BKeT k e 2 kξe (τ )kd τ Control Syst. Technol. , Vol. 11, No. 3, 2003, pp. 399-407.
[10] A. Tornambe and P. Valigi, ”A PID Controller for the Robust Stabi-
t p1
√ lization of SISO Linear Systems”, Applied Mathematics Letters, Vol.
2ζ1
µ ¶
√ c22 c22 5, No. 4, 1992, pp. 15-18.
≤ c12 kP1 BKe k 2 (n + 1) √ γ6 |ε ln ε | + √ ε . [11] A. Tornambe and P. Valigi, ”A Decentralized Controller for the Robust
c0 c21 c11 c21
Stabilization of a Class of MIMO Dyanmics Systems”, Journal of
Since O(|ε lnε |) ≥ Dynamics Systems, Measurement, and Control, Vol. 116, 1994, pp.
pO(ε ), from (70) there exists a positive 293-304.
γ7 (ρ2 ) such that V1 (E(t)) ≤ |ε ln(ε )|γ7 for t ∈ [tu , ∞). There- [12] J. Han, ”Auto-Disturbance Rejection Control and its Applications”,
fore, Control and Decision , Vol. 13, No. 1, 1998, pp. 19-23. (In Chinese)
[13] J. Han, ”From PID to Active Disturbance Rejection Control”, IEEE
kx(t) − x∗ (t)k ≤ O(|ε lnε |), t ∈ [tu , ∞). (71) Trans. Ind. Electron. , Vol. 56, No. 3, 2009, pp. 900-906.
[14] J. Han, ”A class of extended state observers for uncertain systems”,
Since Control and Decision , Vol. 10, No. 1, 1995, pp. 85-88. (In Chinese)
[15] Y. Huang and J. Han, ”Analysis and design for nonlinear continuous
t p1
extended state observer”, Chinese Bulletin , Vol. 45, No. 21, 2000, pp.
√
Z c11
(t−τ )
lim c12 kP1 BKeT k e− 2 kξe (τ )kd τ = 0, 1938-1944.
t→∞ [16] Z. Gao, ”Scaling and Bandwidth-Parameterization Based Controller
tu
Tuning”, in Proc. of the 2003 American Control Conference , Vol. 6,
Zt 2003, pp. 4989-4996.
√ c11
(t−τ ) 2ζ1
c12 kP1 BKeT k e− 2 kξe (τ )kd τ ≤ √ ε, [17] Y. Huang, K. Xu and J. Han and J. Lam, ”Flight control design using
c11 c21 extended state observer and non-smooth feedback”, in Proc. of the
t p1
2001 IEEE Conference on Decision and Control, Vol. 1, 2001, pp.
and lim kx∗ (t)k = 0, from (70), there is 223-228.
t→∞ [18] Y. Hou, Z. Gao, F. Jiang and B.T. Boulter, ”Active Disturbance
Rejection Control for Web Tension Regulation”, in Proc. of the 2001
lim kx(t)k = lim kE(t)k ≤ O(ε ). (72) IEEE Conference on Decision and Control, Vol. 5, 2001, pp. 4974-
t→∞ t→∞
4979.
VI. C ONCLUSION [19] Y. Huang, Z.W. Luo, M.Svinin, T. Odashima and S. Hosoe, ”Extended
state observer based technique for control of robot systems”, in Proc.
In this paper, the strategies of DOB, SPID and ADRC for of the 4th World Congress on Intelligent Control and Automation, Vol.
estimating the uncertainties are compared. Originally, DOB 4, 2002, pp. 2807-2811.
[20] G. Feng, Y. Liu, and L. Huang, ”A New Robust Algorithm to Improve
is usually used to deal with external disturbances, SPID is the Dynamic Performance on the Speed Control of Induction Motor
proposed for nonlinear time-invariant systems with internal Drive”, IEEE Tran. Power Electronics, Vol. 19, No. 6, 2004, pp. 1614-
uncertainties, and ADRC is proposed for more general kinds 1627.
[21] D. Wu, K. Chen and X. Wang, ”Tracking control and active distur-
of nonlinear systems with mixed internal uncertainties and bance rejection with application to noncircular machining”, Interna-
external disturbances. The comparison study and Theorem 1 tional Journal of Machine Tools and Manufacture, Vol. 47, Issue 15,
show that ADRC provides a base for generalizing both DOB 2007, pp. 2207-2217.
[22] B. Yan, Z. Tian, S. Shi and Z. Weng, ”Fault diagnosis for a class of
and SPID for nonlinear system with mixed uncertainties. nonlinear systems via ESO”, ISA Transactions, Vol. 47, No, 4, 2008,
In other words, ADRC is a breakthrough in the research pp. 386-394.
of controlling uncertain systems. The examples validates [23] X. Yang and Y. Huang, ”Capability of extended state observer for
estimating uncertainties”, in Proc. of the 2009 American Control
the efficiency of the generalized DOB and SPID which are Conference , 2009, pp. 3700-3705.
designed according to the frame of ADRC. [24] W.Xue and Y.Huang, ”Stability analysis of ADRC for nonlinear
systems with unknown dynamics and disturbance”, Theory and Ap-
R EFERENCES plications of Complex Systems, 2010.
[25] Q. Zheng, L. Q. Gao, and Z. Gao, ”On Stability Analysis of Active
[1] G. V. Schipanov, ”Theory and methods of designing automatic regu- Disturbance Rejection Control for Nonlinear Time-Varying Plants
lators”, Automatika in Telemekhanika, vol. 4, no. 1, 1939, pp. 49-66. with Unknown Dynamics”, in Proc. of the 46th IEEE Conference on
[2] C.D. Johnson, ”Accommodation of External Disturbances in Linear Decision and control, 2007, pp. 3501-3506.
Regulator and Servomechanism Problems”, IEEE Trans. Automat. [26] Khalil, H.K. Nonlinear Systems. 3rd ed, Prentice Hall: Upper Saddle
Contr., Vol. 16, No. 6, 1971, pp. 635-644. River, New Jersey, 2002.
[3] G.H. Hostetter and J. Meditch, ”On the generalization of observers [27] L.B. Freidovich and H.K. Khalil, ”Performancerecovery of feedback-
to systems with unmeasurable, unknown inputs”, Automatica , Vol. 9, linearization based designs”, IEEE Trans. Automat. Contr. , Vol. 53,
1973, pp. 721-724. No. 10, 2008, pp. 2324-2334.
[4] J. Profeta, W. Vogt and M. Mickle, ”Disturbance Estimation and [28] L. Praly, Z.P. Jiang, ” Further results on robust semiglobal stabilization
Compensation in Linear Systems”, IEEE Trans. Aerosp. Electron. Syst. with dynamic input uncertainties”, in Proc. of the 37th IEEE Conf.
, Vol. 26, No. 2, 1990, pp. 225-231. Decision Control , Vol. 1, 1998, pp. 891-896.
[5] K. Ohishi, M. Nakao, and K. Miyachi, ”Microprocessor-controlled
DC motor for load-insensitive position servo system”, IEEE Trans.
Ind. Electron. , Vol. IE-34, Issue. 1, 1987, pp. 44-49.

4379