A New Extended State Observer For Uncertain Nonlinear Systems
A New Extended State Observer For Uncertain Nonlinear Systems
Automatica
journal homepage: www.elsevier.com/locate/automatica
Brief paper
article info a b s t r a c t
Article history: Extended state observer (ESO) is a class of high-gain observers and a powerful tool for output feedback
Received 4 December 2019 control of uncertain nonlinear systems. In this paper, we propose a new ESO design technique which is
Received in revised form 23 December 2020 based on cascading a series of first-order ESOs. Benefited from the cascade design, saturations can be
Accepted 18 May 2021
inserted into the observer internal variables to limit the maximal implemented gain. The convergence
Available online 5 July 2021
of the new ESO with a class of nonlinear gain functions is established, and it is shown that adopting
Keywords: appropriate nonlinear gain mechanisms leads to improvement in measurement noise tolerance.
Extended state observer (ESO) What is more, the new ESO based output feedback control has stronger uncertainty estimation
Uncertain nonlinear systems and compensation capability than the standard ESO based output feedback control. Specifically, the
Output feedback control proposed approach only relies on the knowledge of the control direction of the uncertain nonlinear
system. Finally, the obtained results are applied to suppress the wing rock motion of a slender delta
wing.
© 2021 Elsevier Ltd. All rights reserved.
1. Introduction
However, there are some limitations for the application of
Output feedback control for uncertain nonlinear systems is ESO. The first is the peaking phenomenon. The peaking occurs
one of the central problems in control theory and applications. in the transient period of the observer and may cause destabi-
The high-gain observer is an efficient and powerful tool for state lization of the overall system (Khalil & Praly, 2014). Also, the
estimation in uncertain nonlinear systems (Khalil & Praly, 2014). gain may increase to be very large for high dimensional systems
The advantages of the high-gain methodology are that it provides and thus its digital implementation becomes demanding. The
an intrinsic robustness with respect to system uncertainties, and second is the high sensitivity to high-frequency measurement
it can recover the performance achieved by state feedback if noise. It is now well-recognized that increasing the gain of the
the gain of the observer is sufficiently high (Khalil, 2017b). The observer will incredibly deteriorate the estimation performance
extended state observer (ESO, or extended high-gain observer) in the presence of high-frequency measurement noise. There is a
can be regarded as an extension of the general high-gain observer. tradeoff between the state estimation speed and the robustness
The basic idea of an ESO is to view the system total uncertainty to measurement noise (Ahrens & Khalil, 2009). The third is the
as an extended state of the system, and then estimate the state requirement of a matching condition between the actual control
and the defined extended state simultaneously. Based on the gain of the system and the nominal control gain utilized in the
output of an ESO, the system total uncertainty can be compen- observer.1 When the ESO is implemented in a feedback loop, the
sated directly in the control action in real time. The theoretical bound information of the actual control gain is required to select
verifications of the ESO-based control were given in Freidovich the nominal control gain to fulfill the matching condition. This
and Khalil (2008) and Guo and Zhao (2013), and its successful ap- matching condition was theoretically derived in Freidovich and
plications in a broad range of control engineering can be found in Khalil (2008) and Guo and Zhao (2013), and has also been verified
the literature, known as the active disturbance rejection control by many practical experiments (see, e.g., Wu & Chen, 2009).
(ADRC) (Gao, 2003; Han, 2009). Many solutions have been developed to alleviate the limi-
tations of the ESO, most of which have focused on reducing
the peaking phenomenon and improving the robustness to mea-
✩ The material in this paper was not presented at any conference. This paper surement noise. The simplest way to prevent the effects of the
was recommended for publication in revised form by Associate Editor Daniele peaking phenomenon is to saturate the control to be injected
Astolf under the direction of Editor Daniel Liberzon.
∗ Corresponding author.
E-mail addresses: [email protected] (M. Ran), [email protected] 1 In this paper, for a system ẋ = f (x) + bu with state x and input u, the
(J. Li), [email protected] (L. Xie). parameter b is referred to as the control gain and sign(b) the control direction.
https://doi.org/10.1016/j.automatica.2021.109772
0005-1098/© 2021 Elsevier Ltd. All rights reserved.
M. Ran, J. Li and L. Xie Automatica 131 (2021) 109772
into the system outside a compact set of interest (Freidovich & 2. New extended state observer
Khalil, 2008; Ran, Wang, & Dong, 2017). However, the peaking
in the internal variables of the observer cannot be prevented 2.1. Observer design
by this method. The basic idea of mitigating the sensitivity to
measurement noise is to use a smaller gain in the steady period of Consider the following uncertain nonlinear system
the observer (Ahrens & Khalil, 2009; Prasov & Khalil, 2013). The
⎨ ż =f0 (x, z , ω),
⎧
existing approaches can be generally classified into four groups
(Cocetti, Tarbouriech, & Zaccarian, 2018): (1) multi-observer ap- ẋ =Ax + B[f (x, z , ω) + bu], (1)
proach (Chong, Nešić, Postoyan, & Kuhlmann, 2015); (2) switched ⎩
y =Cx,
gain approach (Ahrens & Khalil, 2009); (3) adaptive gain approach
(Sanfelice & Praly, 2011); (4) nonlinear or time-varying output where x = [x1 , . . . , xn ]T ∈ Rn and z ∈ Rp are the states, ω ∈ R
injection approach (Andrieu, Praly, & Astolfi, 2008). Usually these is the external disturbance, u ∈ R is the control input, y ∈ R is
approaches behave better than the standard high-gain observer, the measured output, b is the bounded uncertain nonzero control
but the performance heavily depends on the new parameters gain, f0 : Rn × Rp × R → Rp and f : Rn × Rp × R → R are unknown
introduced by the modified schemes. What is more, the resulting continuously differentiable functions, and matrices A, B, and C are
observers are generally more complicated to implement than the given by
standard one.
0 1 ··· 0 0
⎡ ⎤ ⎡ ⎤
In Astolfi and Marconi (2015), a new high-gain observer,
⎢ .. .. .. .. ⎥ ⎢.⎥
which can be regarded as a cascade connection of n − 1 second-
A=⎢ . . . . ⎥ ∈ Rn×n , B = ⎢ .. ⎥ ∈ Rn×1 ,
order high-gain observers with feedback injection from one chan- ⎣
0 0 ··· 1
⎦
0
⎣ ⎦
nel to the previous one, was proposed for nth order systems. 0 0 ··· 0 1
The key feature of the observer proposed in Astolfi and Marconi
(2015) is that its maximum implemented gain is 1/ε 2 , while C = [1 0 · · · 0] ∈ R1×n .
for a standard linear nth order high-gain observer it is 1/ε n ,
where 1/ε is the high-gain parameter. Motivated by Astolfi and The following assumptions are made throughout the paper.
Marconi (2015), a cascade high-gain observer was developed in
Khalil (2017a) by cascading a second-order observer with first- Assumption A1. The external disturbance ω and its time deriva-
order ones, and inserting saturations between different channels. tive ω̇ are bounded.
Note that the observers in Astolfi and Marconi (2015) and Khalil
(2017a) cannot estimate the system uncertainty, and the ob- Assumption A2. The zero-dynamics ż = f0 (x, z , ω) with input
servers’ gain functions are limited to be linear. However, now it is (x, ω) is bounded-input-bounded-state stable.
well-realized that using nonlinear gain mechanisms has the po-
For system (1), let xn+1 ≜ f (x, z , ω) + (b − b0 )u be the extended
tential of improving the performance, especially in measurement
state, where b0 is a bounded nonzero nominal value of the control
noise tolerance (Cocetti et al., 2018; Li, Xia, Qi, & Gao, 2017; Zhao
gain b. Then a standard ESO can be designed as (Freidovich &
& Guo, 2018).
Khalil, 2008; Guo & Zhao, 2013)
In this paper, a new ESO is proposed for uncertain nonlinear ⎧ ( )
systems. Motivated by Astolfi and Marconi (2015), Han (2009), ⎪
⎪ ˙ i = x̂i+1 + ε n−i gi x1 − x̂1 , 1 ≤ i ≤ n − 1,
x̂
and Khalil (2017a), the basic philosophy of the proposed ESO εn
⎪
⎪
⎪
is to cascade a series of first-order ESOs. A framework for the
⎪
⎪ ( )
x1 − x̂1
⎨
design, analysis, and application in output feedback control of the x̂˙ n = x̂n+1 + gn + b 0 u, (2)
new ESO is established. With respect to the existing ESOs, the ⎪
⎪ εn
⎪ ( )
x1 − x̂1
⎪
advantages of the new ESO are threefold: ⎪˙
⎩ x̂n+1 = ε−1 gn+1 ,
⎪
⎪
εn
(1) Inherited from the cascade design in Khalil (2017a), by
inserting saturations, the peaking of the internal variables where [x̂1 , . . . , x̂n+1 ]T ∈ Rn+1 is the observer state, ε < 1 is a
of the new ESO can be limited to the order of 1/ε , no matter small positive constant, gi : R → R, 1 ≤ i ≤ n + 1, are some
how large the system order n is. odd continuous functions. The simplest form of (2) is to let the
(2) The convergence of the new ESO with a class of nonlinear functions gi (·), 1 ≤ i ≤ n + 1, be linear, that is (Gao, 2003),
gain functions is established, and hence one can adopt
appropriate nonlinear gain mechanisms with convergence gi (ν ) = li ν, (3)
guarantees to improve the performance in measurement
where L = [l1 , l2 , . . . , ln+1 ]T is the linear ESO (LESO) gain selected
noise tolerance.
such that the following polynomial is Hurwitz:
(3) The new ESO based output feedback control has stronger
capability to estimate and compensate the system uncer- λn+1 + l1 λn + · · · + ln λ + ln+1 . (4)
tainty than the standard ESO based control. Specifically, we
show that the new ESO based control relaxes the matching Note that there are several drawbacks of the ESO (2) in the
condition and only needs to know the control direction. standard form: (1) it faces a numerical challenge if the system
order, n, is high (the maximal implemented gain is of the order of
The rest of the paper is organized as follows. The structure and 1/ε n+1 ); (2) it lacks a flexible scheme to apply different nonlinear
convergence analysis of the new ESO are presented in Section 2. gain functions gi (·), 1 ≤ i ≤ n + 1, with convergence and
A numerical example is also given in Section 2 to illustrate the performance guarantees; (3) when implemented in a feedback
effectiveness and advantages of the proposed observer. The new loop, the actual control gain b and nominal control ⏐ gain⏐ b0 need
ESO based output feedback control is presented in Section 3. In ⏐ b−b0 ⏐
to satisfy a matching condition in the form of ⏐ b0 ⏐
≤ ∆0 ,
Section 4, the proposed approach is applied to suppress the wing
rock motion with largely uncertain control gain. Finally, Section 5 where 0 < ∆0 < 1 (Freidovich & Khalil, 2008). Motivated
concludes the paper. by these observations, in this paper, we present a new ESO for
2
M. Ran, J. Li and L. Xie Automatica 131 (2021) 109772
system (1), Assumption A4. The states x and z, and the control input u and
⎧
g1 (e1 ) g1 (e1 ) its derivative u̇ are bounded.
⎪ ζ̇1 = ε , e1 = y − ζ1 , x̂2 = ε ,
⎪
⎪
Theorem 1. Consider the uncertain nonlinear system (1) and the
⎪
⎪
gi (ei ) gi (ei )
⎨
ζ̇i = , ei = x̂i − ζi , x̂i+1 = , 2 ≤ i ≤ n − 1, (5) proposed ESO (5). Suppose Assumptions A1–A4 are satisfied, and the
⎪
⎪ ε ε initial condition of the observer belongs to a compact subset of Rn .
⎩ ζ̇n = gn (en ) + b0 u, en = x̂n − ζn , x̂n+1 = gn (en ) ,
⎪
Then for any σ > 0 and T > 0, there exists ε ∗ > 0 such that
⎪
⎪
ε ε ∀ε ∈ (0, ε∗ ),
where ζ = [ζ1 , ζ2 , . . . , ζn ]T ∈ Rn is the observer state, ε < 1
|xi (t) − x̂i (t)| ≤ σ , 2 ≤ i ≤ n + 1, ∀t ≥ T . (10)
is a small positive constant, gi : R → R, 1 ≤ i ≤ n, are odd
continuous functions.
Proof. Since y and u are bounded by Assumption A4 and each
stage of the observer (5) is a stable system, the existence of x̂2
Remark 1 (Design Philosophy). Compared with the standard ESO
to x̂n+1 is guaranteed. Consider the scaled estimation error η =
(2), the main feature of the proposed ESO (5) is that there exists
a connection between channel i and channel i + 1, i.e., the output
[η2 , . . . , ηn+1 ]T , with
of channel i is the input of channel i + 1. Since each channel of ηi = ε i−1 (xi − x̂i ), 2 ≤ i ≤ n + 1. (11)
(5) is a first-order ESO, the proposed observer can be regarded as
a cascade connection of n first-order ESOs. For subsequent use, denote η̃j = [η2 , . . . , ηj ] , Hj = [0 · · · 0 1] ∈
T
R1×(j−1) , 3 ≤ j ≤ n.
2.2. Convergence analysis
The proof will proceed by an iterative procedure. The conver-
gence of the estimation error will be proved step by step, starting
Before establishing the convergence of the proposed ESO (5),
we need some preliminary results on linear time-varying sys- from η2 and ending with ηn+1 .
tems. Step (1): The dynamics of η2 can be formulated as
1
Definition 1 (Rugh, 1996). Consider the linear time-varying sys- η̇2 = ε(ẋ2 − x̂˙ 2 ) = ε x3 − η2 g1′ (e1 ). (12)
ε
tem
Note that this equation falls into the form of system (9), 0 <
ξ̇ (t) = M(t)ξ (t), t ∈ R+ , (6) m1 ≤ g1′ (e1 ) ≤ m1 by Assumption A3, and x3 is bounded by
Assumption A4. Therefore, by Lemma 2, one can conclude that
where M(t) ∈ C (R+ , Rn×n ). The system (6) is said to be uni-
for any σ > 0 and T1 ∈ (0, T ), there exists ε1 > 0 such that for
formly exponentially stable (UES) if and only if there exist positive
any ε ∈ (0, ε1 ) and t ∈ [T1 , ∞), |η2 (t)| ≤ σ ε . This proves the case
constants γ1 and γ2 such that
i = 2 of (10).
∥Θ (t , t0 )∥ ≤ γ1 e−γ2 (t −t0 ) , t ≥ t0 , t0 ∈ R+ , (7) Step (2): The dynamics of η3 reads
where Θ (t , t0 ) is the state transition matrix of (6). η̇3 =ε 2 (ẋ3 − x̂˙ 3 )
( ( ) )
1 1 1
Lemma 1. Consider the following blocked linear time-varying =ε 2 x4 − 3 η2 g1′ (e1 )g2′ (e2 ) + x3 − 2 η3 g2 (e2 )
′
system ε ε ε
1 1
= − g2′ (e2 )η3 − g1′ (e1 )g2′ (e2 )η2 + ε g2′ (e2 )x3 + ε 2 x4 .
[ ]
M11 (t) 0 (13)
ξ̇ (t) = ξ (t), t ∈ R+ , (8) ε ε
M21 (t) M22 (t)
Then by (12) and (13), the dynamics of η̃3 can be written into the
where Mij (t) ∈ C (R+ , Rni ×nj ), i, j = 1, 2, i ≥ j, and n1 + n2 = n. following compact form:
If the subsystems ξ̇1 (t) = M11 (t)ξ1 (t) and ξ̇2 (t) = M22 (t)ξ2 (t) are
1
UES, and ∥M21 (t)∥ ≤ γ21 , ∀t ∈ R+ , where γ21 ∈ R+ . Then the η̃˙ 3 = Φ31 η̃3 + ε Φ32 x3 + ε 2 Φ33 x4 , (14)
system (8) is UES. ε
where
Proof. See Appendix.
[ ] [ ]
−g1′ (e1 ) 0 1
Φ31 = ′ ′ , Φ32 = ′ ,
−g1 (e1 )g2 (e2 ) −g2′ (e2 ) g2 (e2 )
Lemma 2. Consider the perturbed time-varying system ]T
Φ33 = 0 1 .
[
1 (15)
ξ̇ (t) = M(t)ξ (t) + δ (t), t ≥ t0 , t0 ∈ R+ , (9)
ε Eq. (14) can be used to show a bound on η̃3 of the form σ ε . By the
where M(t) ∈ C (R≥t0 , Rn×n ), δ (t) is the perturbation term, and the definition of η3 , this bound does not yield (10). We need a sharper
initial condition ξ (t0 ) belongs to a compact subset of Rn . Assume the bound on η̃3 . Note that Φ31 is invertible for any t ∈ [0, ∞). To
linear time-varying system ξ̇0 (t) = M(t)ξ0 (t) is UES, and ∥δ (t)∥ ≤ eliminate the ε Φ32 x3 term in (14), we apply the transformation
δ̄εm for some ε -independent positive constant δ̄ and integer m ≥ 0. ψ3 = η̃3 + ε 2 Φ31−1
Φ32 x3 , which yields
Then for any σ > 0 and T > 0, there exists ε0 > 0 such that for
d Φ31
−1
Φ32
( )
any ε ∈ (0, ε0 ) and t ∈ [t0 + T , ∞), the solution of (9) satisfies 1
ψ̇3 = Φ31 ψ3 + ε 2 x3 + ε 2 (Φ31
−1
Φ32 + Φ33 )x4 . (16)
∥ξ (t)∥ ≤ σ εm . ε dt
( )
−1
d Φ31 Φ32 (2)
(x2 −x̂2 )g1 (e1 )
Proof. See Appendix. It can be computed that = [ 0]T . By the
dt g ′2 (e )
1 1
results in Step (1), the sum of the second and third terms on the
Assumption A3. The functions gi (·), 1 ≤ i ≤ n, are nth order con-
right hand side of (16) is of the order of O(ε 2 ) for any t ≥ T1 . This
tinuously differentiable, and there exist ε -independent constants
together with Lemma 1 yields that the dynamics of ψ3 fall into
0 < m1 ≤ m1 and mj > 0 such that m1 ≤ gi′ (ei ) ≤ m1 and
dj gi (ei ) the form of system (9) with m = 2. Thus one can conclude that
|gi(j) (ei )| ≤ mj , 2 ≤ j ≤ n, where gi′ (ei ) ≜ dgi (ei )
dei
(j)
and gi (ei ) ≜ j . for any σ > 0 and T2 ∈ (T1 , T ), there exists ε2 ∈ (0, ε1 ] such that
dei
3
M. Ran, J. Li and L. Xie Automatica 131 (2021) 109772
for any ε ∈ (0, ε2 ) and t ∈ [T2 , ∞), ∥ψ3 (t)∥ ≤ ε 2 σ . On the other where
hand, η3 can be obtained by d
(
d
(
Γ (·) = Φj−+11,1 Φj−+11,1 · · ·
η3 = H3 η̃3 = H3 ψ3 − ε 2 H3 Φ31
−1
Φ32 x3 . (17) dt dt
( ) ))
d d ( −1
The second term ] of (17) vanishes because H3 Φ31 Φ32 = [0 1] −1
Φj+1,1
−1
Φj+1,1 Φj+1,2 x3 · · ·
)
− g ′ (e1 ) dt dt
[
0
[ ]
1
1 1 = 0. Thus η3 = H3 ψ3 , and one
( (
1 ′ d d
1 − g ′ (e ) g2 (e2 ) + Φj−+11,1 Φj−+11,1 · · ·
2 2
dt dt
can conclude that for any σ > 0 and T2 ∈ (T1 , T ), there exists ( ) ))
d d ( −2
ε2 ∈ (0, ε1 ] such that for any ε ∈ (0, ε2 ) and t ∈ [T2 , ∞), Φj−+11,1 Φj+1,1 Φj+1,2 + Φj−+11,1 Φj+1,3 x4 · · ·
)
|η3 (t)| ≤ ε 2 σ . This proves the case i = 3 of (10). dt
( (
dt
Step (j) (3 ≤ j ≤ n − 1): In this step, we consider the d −(j−1)
convergence of ηj+1 with the conditions that ηj converges over + ··· + Φj+1,1 Φj+1,2 + · · ·
dt
[Tj−1 , ∞) for any ε ∈ (0, εj−1 ), and the dynamics of η̃j is given ))
by + Φj−+21,1 Φj+1,j−1 + Φj−+11,1 Φj+1,j xj+1 .
j−1
1 ∑
η̃˙ j = Φj1 η̃j + ε i−1 Φji xi+1 + ε j−1 Φjj ẋj , (18) The second term of (22) is of the order of O(ε j ) in the time
ε interval [Tj−1 , ∞) due to the convergence of η2 to ηj , and the
i=2
where Φj1 to Φjj are matrices whose expressions with j = 3 are lower triangular structure of Φj+1,1 . Thus by Lemmas 1 and 2, one
given by (15). has that for any σ > 0 and Tj ∈ (Tj−1 , T ), there exists εj ∈ (0, εj−1 ]
Consider the dynamics of ηj+1 , such that for any ε ∈ (0, εj ) and t ∈ [Tj , ∞), ∥ψj+1 (t)∥ ≤ ε j σ .
( ) What is more,
η̇j+1 =ε j ẋj+1 − x̂˙ j+1
( ) ηj+1 =Hj+1 η̃j+1
1 ( 1−j
[ ]
=ε j ẋj+1 − −ε η̇j + ε−j ηj+1 gj′ (ej ) d(Φj−+11,1 Φj+1,2 )
)
ε − ε 3
Hj+1 Φj−+11,1 x3 + · · ·
dt
1
= − gj′ (ej )ηj+1 + gj′ (ej )η̇j + εj ẋj+1 . (19) j
( k
)
ε ∑ ∑ −(k+1−i)
− ε k
Hj+1 Φj+1,1 Φj+1,i xk+1 . (23)
By (18) and the relationships η̇j = Hj η̃˙ j and η̃j+1 = [η̃j ηj+1 ]T , one k=2 i=2
has
By a straightforward iterative procedure, one can verify that
Φj1 η̃j
[ ][ ]
˙η̃j+1 = 1 0 −(k+1−i)
Hj+1 Φj+1,1 Φj+1,i = 0, 2 ≤ k ≤ j, 2 ≤ i ≤ k. In addition,
ε Hj Φj1 gj (ej ) −gj (ej ) ηj+1
′ ′
since the first j − 1 entries of Hj+1 are 0, the second term of
j
Φ the right hand side of (23) equals 0. Therefore ηj+1 = Hj+1 ψj+1 ,
[ ] [ ]
∑ 0
+ ε i−1 H Φ gji′ (e ) xi+1 + ε j (j−1)×1 ẋj+1 and as a consequence, for any σ > 0 and Tj ∈ (Tj−1 , T ), there
j ji j j 1
i=2 exists εj ∈ (0, εj−1 ] such that for any ε ∈ (0, εj ) and t ∈ [Tj , ∞),
1
j
∑ |ηj+1 (t)| ≤ ε j σ . This proves the cases 4 ≤ i ≤ n of (10).
≜ Φj+1,1 η̃j+1 + ε i−1 Φj+1,i xi+1 Step (n): Computing the time derivative of ηn+1 yields
ε
i=2
1
+ ε j Φj+1,j+1 ẋj+1 , (20) η̇n+1 = − gn′ (en )ηn+1 + gn′ (en )η̇n + ε n ẋn+1 . (24)
ε
Φj1 Φji
[ ] [ ]
0 The differentiation of the extended state xn+1 can be specified as
where Φj+1,1 = , Φj+1,i = ,
Hj Φj1 gj (ej ) −gj (ej )
′ ′
Hj Φji gj (ej )
′
df
(x, z , ω) + (b − b0 )u̇
[ ]
0(j−1)×1 ẋn+1 =
2 ≤ i ≤ j, Φj+1,j+1 = . Similar to Step (2), to eliminate dt
1 n−1
∑ ∂f ∂f
the second term in (20), consider the transformation = xi+1 (x, z , ω) + (f (x, z , ω) + bu) (x, z , ω)
∂ xi ∂ xn
ψj+1 =η̃j+1 + ε 2 Φj−+11,1 Φj+1,2 x3 i=1
∂f ∂f
d(Φj−+11,1 Φj+1,2 ) + f0 (x, z , ω) (x, z , ω) + ω̇ (x, z , ω) + (b − b0 )u̇. (25)
+ε 3
Φj−+11,1 x3 ∂z ∂ω
dt By Assumptions A1, A2, and A4, it can be observed that ẋn+1 is
+ ε 3 Φj−+21,1 Φj+1,2 + Φj−+11,1 Φj+1,3 x4 + · · ·
( )
bounded by some ε -independent positive constant. Thus, based
(
+ ε j Φj−+(j1−,11) Φj+1,2 + Φj−+(j1−,12) Φj+1,3 + · · · on the boundedness of ẋn+1 , and following a similar line of the
previous steps, one can obtain that for any σ > 0 and T > 0, there
exists ε ∗ ∈ (0, εn−1 ] such that for any ε ∈ (0, ε ∗ ) and t ∈ [T , ∞),
)
+ Φj−+21,1 Φj+1,j−1 + Φj−+11,1 Φj+1,j xj+1 . (21)
|ηn+1 (t)| ≤ ε n σ . This proves the case i = n + 1 of (10).
Applying (21) to (20) yields Finally, summarizing Step (1) to Step (n) leads to (10). This
1 completes the proof of Theorem 1. ■
ψ̇j+1 = Φj+1,1 ψj+1 To overcome the effects of peaking, for a standard ESO, one
ε
can saturate its output before applying the estimates to control.
+ ε j Γ (x3 , . . . , xj+1 , Φj+1,1 , . . . , Φj+1,j ) However, such saturation does not prevent the peaking in the
( j+1 )
∑ −(j+1−k) internal dynamics of the observer. For the proposed ESO (5),
+ε j
Φj+1,1 Φj+1,k ẋj+1 , (22) benefited from its cascade structure, the peaking in the output
k=2 and in the internal dynamics can be prevented simultaneously
4
M. Ran, J. Li and L. Xie Automatica 131 (2021) 109772
5
M. Ran, J. Li and L. Xie Automatica 131 (2021) 109772
which, contradicts (30). Thus there exists ε ∗ > 0 such that for any respectively (see Table 2 in Ordón̄ez & Passino, 2000).
ε ∈ (0, ε∗ ) and t ∈ [0, ∞), x(t) ∈ Ω1 . This completes the proof of The angle of attack is assumed to be generated by an exoge-
Lemma 3. ■ nous dynamical system
Based on the boundedness of the system state x, we are in a
α̇1 −25 α1
[ ] [ ][ ] [ ] [ ]
0 0 0
position to state the convergence of the closed-loop system. = + + r, (39)
α̇2 −25 −10 α2 500 62.5
Theorem 3. Consider the closed-loop system formed by plant (1), where α1 = α and r is a command input that takes values
ESO (5), and control (29). Suppose bb0 > 0, Assumptions A1–A3 are between −1 and 1 every 0.5 s. The system (39) represents the
satisfied, the initial condition of the observer belongs to a compact effects of the aircraft itself which receives the pilot’s commands
subset of Rn , and the initial condition of the plant (x(0), z(0)) belongs through the input r (Ordón̄ez & Passino, 2000). It was shown
to Ω0 × Z . Let the saturation level M be selected as in Lemma 3. in Nayfeh et al. (1989) that the wing rock dynamics has a sta-
Then for any σ > 0 and T > 0, there exists ε † > 0 such that for ble focus at the origin for α ≤ 19.5◦ . For higher angles, the
any ε ∈ (0, ε † ), (10) holds, and origin becomes an unstable equilibrium and a limit cycle be-
havior occurs. Note that according to (39), the angle of attack α
lim ∥x(t)∥ ≤ σ . (37)
t →∞ varies between 15◦ and 25◦ . Thus the qualitative behavior of (38)
changes periodically as α becomes larger or smaller than 19.5◦ .
Proof. From Lemma 3, there exists ε ∗ > 0 such that for any In the present work, we assume that the complicated uncer-
ε ∈ (0, ε∗ ), x(t) ∈ Ω1 holds for all t ∈ [0, ∞). This together tain nonlinear aerodynamics of the slender delta wing is totally
with the fact that the control u is bounded and each stage of the unknown, and the control input gain g is also uncertain.
( That is,
ρj (α ) −ωj2 x1 + µj1 x2 +
∑7
observer (5) is exponentially stable guarantees the existence of x̂2 the extended state is defined as x3 = j=1
to x̂n+1 for all t ∈ [0, ∞). Then following the first n − 1 steps in j j j
)
Theorem 1, one can conclude the cases 2 ≤ i ≤ n of (10), and by b1 x32 + µ2 x21 x2 +b2 x1 x22 + (g − g0 )δa , where g0 is a nominal value
(33) the case i = n + 1 of (10). Based on the convergence of the of g. The new ESO is designed in the linear form, with l1 = l2 = 1
observer, (37) can be concluded readily from (35). This completes and ε = 0.01. The new ESO based control (29) is designed with
the proof of Theorem 3. ■ K = [−4 − 4] and M = 20. Initial conditions are set as x1 (0) = 5,
x2 (0) = 0, x̂2 (0) = 0, and x̂3 (0) = 0. To verify the feature that
Remark 3 (Prior Knowledge of the Control Gain). Theorem 3 shows the new ESO based control does not rely on an accurate nominal
that the proposed ESO based output feedback control does not control gain, we conduct simulations with different values of g0 .
rely on a carefully selected nominal control gain b0 . Specifically, Figs. 3 and 4 show the simulation results with g0 = 1.0 and
7
M. Ran, J. Li and L. Xie Automatica 131 (2021) 109772
Fig. 3. Simulation results of the proposed ESO based controller with g0 = 1: Fig. 5. Simulation results of the proposed ESO based controller (blue full line)
(a) the roll angle; (b) the roll rate and its estimate; (c) the total uncertainty and and the standard ESO based controller (red dashed line) with g0 = 20: (a) the
its estimate; (d) the aileron deflection. roll angle; (b) the roll rate; (c) the aileron deflection.
Acknowledgments
Appendix
Fig. 4. Simulation results of the proposed ESO based controller with g0 = 0.01:
(a) the roll angle; (b) the roll rate and its estimate; (c) the total uncertainty and
its estimate; (d) the aileron deflection.
Proof of Lemma 1. Let Θii (t , t0 ) be the state transition matrices
g0 = 0.01, respectively. One can see that, in both cases, the new for the subsystems ξ̇i (t) = Mii (t)ξi (t), i = 1, 2. Then the state
ESO based control suppresses the wing rock motion effectively. transition matrix for system (8) can be given by
For comparison, we also design a standard LESO based control. Θ11 (t , t0 )
[ ]
0
The standard LESO is designed with L = [3 3 1]T and ε = 0.01. Θ (t , t0 ) = , (41)
Θ21 (t , t0 ) Θ22 (t , t0 )
The gain L = [3 3 1]T places the poles of the polynomial (4) at
−1. The control is designed as where
t
( )
K x̂ − x̂3
∫
δa = Msat , (40) Θ21 (t , t0 ) = Θ22 (t , τ )M21 (τ )Θ11 (τ , t0 )dτ . (42)
Mg0 t0
where x̂ = [x̂1 , x̂2 ]T and x̂3 are the output of the standard LESO, Since the two subsystems are UES, one has
control gain K = [−4 − 4], and saturation bound M = 20. {
Initial conditions of the standard LESO are set as x̂1 (0) = 5, ∥Θ11 (t , t0 )∥ ≤ γ1 e−γ2 (t −t0 ) , t0 ∈ R+ , t ≥ t0 ,
(43)
x̂2 (0) = 0, and x̂3 (0) = 0. Our simulation shows that for the ∥Θ22 (t , t0 )∥ ≤ γ3 e−γ4 (t −t0 ) , t0 ∈ R+ , t ≥ t0 ,
considered wing rock motion control problem, if there is an accu-
rate nominal value of the control gain g, the standard ESO based where γi , 1 ≤ i ≤ 4, are positive constants. By (41)–(43),
control (40) achieves comparable performance with the new ESO ∥M21 (t)∥ ≤ γ21 , and some straightforward manipulations, one
based control. However, the performance of the standard ESO can verify that the state transition matrix Θ (t , t0 ) satisfies (7).
based control degrades rapidly as g0 deviates from the actual Here we omit the details due to the space limitation. ■
value of g. Fig. 5 shows the comparison of the two approaches
with nominal input gain g0 = 20. It can be clearly observed that Proof of Lemma 2. Let Θ (t , t0 ) be the state transition matrix for
the new ESO based control achieves better performance. What the system ξ̇0 (t) = 1ε M(t)ξ0 (t). Then one has
is more, the closed-loop system under the standard ESO based
γ2
control becomes unstable when g0 is larger than 28.74 or smaller ∥Θ (t , t0 )∥ ≤ γ1 e− ε (t −t0 ) , t ≥ t0 , t0 ∈ R+ , (44)
than 0.183.
where γ1 and γ2 are positive constants. The solution of (9) is given
5. Conclusion by
∫ t
A new ESO design technique is presented for uncertain nonlin-
ξ (t) = Θ (t , t0 )ξ (t0 ) + Θ (t , τ )δ (τ )dτ . (45)
ear systems. The proposed observer provides a novel scheme to t0
8
M. Ran, J. Li and L. Xie Automatica 131 (2021) 109772
By (44) and (45), one has Rugh, W. J. (1996). Linear System Theory (2nd ed.). Upper Saddle River, New
∫ t
Jersey: Prentice Hall.
γ2 γ2
Sanfelice, R. G., & Praly, L. (2011). On the performance of high-gain ob-
∥ξ (t)∥ ≤γ1 e− ε (t −t0 ) ∥ξ (t0 )∥ + γ1 e− ε (t −τ ) δ̄εm dτ servers with gain adaptation under measurement noise. Automatica, 47(10),
t0
2165–2176.
2 γ δ̄γ1 ( γ2 ) Wu, D., & Chen, K. (2009). Design and analysis of precision active distubance re-
=γ1 e− ε (t −t0 ) ∥ξ (t0 )∥ + 1 − e− ε (t −t0 ) ε m+1 jection control for noncircular turning process. IEEE Transactions on Industrial
γ2 Electronics, 56(7), 2746–2753.
γ2 δ̄γ 1 m+1 Zhao, Z. L., & Guo, B. Z. (2018). A novel extended state observer for output
≤γ1 e− ε (t −t0 ) ∥ξ (t0 )∥ + ε . (46) tracking MIMO systems with mismatched uncertainty. IEEE Transactions on
γ2 Automatic Control, 63(1), 211–218.
If ∥ξ (t0 )∥ < ε m+1 , the right hand side of (46) is of the order
of O(ε m+1 ) for all t ≥ t0 . If ∥ξ (t0 )∥ ≥ ε m+1 , the right hand
side of (46) is of the order of O(ε m+1 ) for all t ≥ t0 + tε with Maopeng Ran received the B.E. degree in automation,
(m+1)ε
ln ε + γε ln ∥ξ (t0 )∥. Note that tε → 0 as ε → 0.
the Ph.D. degree in navigation, guidance, and control
tε = − γ from Beihang University, Beijing, China, in 2012 and
2 2
Therefore, for any σ > 0 and T > 0, there exists ε0 > 0 such that 2018, respectively. He is currently a Research Fellow
in the School of Electrical and Electronic Engineer-
for any ε ∈ (0, ε0 ) and t ∈ [t0 + T , ∞), ∥ξ (t)∥ ≤ σ ε m holds. ■
ing, Nanyang Technological University, Singapore. His
research interests include nonlinear control, machine
References learning, observers, multi-agent systems, and control
for flight vehicles and mobile robots.
Ahrens, J. H., & Khalil, H. K. (2009). High-gain observers in the presence of Dr. Ran received the 2019 Beihang University Ex-
measurement noise: a switched-gain approach. Automatica, 45, 936–943. cellent Doctoral Dissertation Award, and the Guan
Andrieu, V., Praly, L., & Astolfi, A. (2008). Homogeneous approximation, re- Zhao-Zhi Award at the 38th Chinese Control Conference.
cursive observer design, and output feedback. SIAM Journal on Control and
Optimization, 47(4), 1814–1850.
Astolfi, D., & Marconi, L. (2015). A high-gain nonlinear observer with limited
gain power. IEEE Transactions on Automatic Control, 60(11), 3059–3064.
Chong, M. S., Nešić, D., Postoyan, R., & Kuhlmann, L. (2015). Parameter Juncheng Li received the B.S. degree from the Depart-
and state estimation of nonlinear systems using a multi-observer under ment of Mechanical Engineering, Tsinghua University,
the supervisory framework. IEEE Transactions on Automatic Control, 60(9), Beijing, China, in 2017. He is currently pursuing
2336–2349. the Ph.D. degree in School of Electrical and Elec-
Cocetti, M., Tarbouriech, S., & Zaccarian, L. (2018). On dead-zone observer for tronic Engineering, Nanyang Technological University
linear plants. In American Control Conference, USA: (pp. 5138-5143). (NTU), Singapore. His current research interests include
Freidovich, L. B., & Khalil, H. K. (2008). Performance recovery of feedback- nonlinear control and its application in mobile robots.
linearization-based designs. IEEE Transactions on Automatic Control, 53(10),
2324–2334.
Gao, Z. (2003). Scaling and bandwidth-parameterization based controller tuning.
In American Control Conference, USA: (pp. 4989-4996).
Guo, B. Z., & Zhao, Z. L. (2013). On convergence of the nonlinear active
disturbance rejection control for MIMO systems. SIAM Journal on Control and
Optimization, 51(2), 1727–1757.
Han, J. (2009). From PID to active disturbance rejection control. IEEE Transactions Lihua Xie received the B.E. and M.E. degrees in elec-
on Industrial Electronics, 56(3), 900–906. trical engineering from Nanjing University of Science
Khalil, H. K. (2017a). Cascade high-gain observers in output feedback control. and Technology in 1983 and 1986, respectively, and
Automatica, 80, 110–118. the Ph.D. degree in electrical engineering from the
Khalil, H. K. (2017b). High-Gain Observers in Nonlinear Feedback Control. Society University of Newcastle, Australia, in 1992. Since 1992,
for Industrial and Applied Mathematics (SIAM). he has been with the School of Electrical and Elec-
Khalil, H. K., & Praly, L. (2014). High-gain observers in nonlinear feedback control. tronic Engineering, Nanyang Technological University,
International Journal of Robust and Nonlinear Control, 24(6), 993–1015. Singapore, where he is currently a Professor and Direc-
Li, J., Xia, Y., Qi, X., & Gao, Z. (2017). On the necessity, scheme, and basis of tor, Delta-NTU Corporate Laboratory for Cyber–Physical
the linear-nonlinear switching in active disturbance rejection control. IEEE Systems. He served as the Head of Division of Control
Transactions on Industrial Electronics, 64(2), 1425–1435. and Instrumentation from July 2011 to June 2014. He
Nayfeh, A. H., Elzebda, J. M., & Mook, D. T. (1989). Analytical study of the held teaching appointments in the Department of Automatic Control, Nanjing
subsonic wing-rock phenomenon for slender delta wings. Journal of Aircraft, University of Science and Technology from 1986 to 1989.
26(9), 805–809. Dr Xie’s research interests include robust control and estimation, networked
Ordón̄ez, R., & Passino, K. M. (2000). Wing rock regulation with a time-varying control systems, multi-agent networks, localization and unmanned systems. He
angle of attack. In Proceedings of the 15th IEEE International Symposium on is an Editor-in-Chief for Unmanned Systems and an Associate Editor for IEEE
Intelligent Control Greece: (pp. 145-150). Transactions on Network Control Systems. He has served as an editor of IET Book
Prasov, A. A., & Khalil, H. K. (2013). A nonlinear high-gain observer for systems Series in Control and an Associate Editor of a number of journals including IEEE
with measurement noise in a feedback control framework. IEEE Transactions Transactions on Automatic Control, Automatica, IEEE Transactions on Control
on Automatic Control, 58(3), 569–580. Systems Technology, and IEEE Transactions on Circuits and Systems-II. He is an
Ran, M., Wang, Q., & Dong, C. (2017). Active disturbance rejection control elected member of Board of Governors, IEEE Control System Society (Jan 2016–
for uncertain nonaffine-in-control nonliner systems. IEEE Transactions on Dec 2018), and a Fellow of Academy of Engineering Singapore. Dr Xie is a Fellow
Automatic Control, 62(11), 5830–5836. of IEEE, and a Fellow of IFAC.