SICE Journal of Control, Measurement, and System

Integration

ISSN: 1882-4889 (Print) 1884-9970 (Online) Journal homepage: https://www.tandfonline.com/loi/tmsi20

Extended High-Gain Observers as Disturbance

Estimators

Hassan K. Khalil

To cite this article: Hassan K. Khalil (2017) Extended High-Gain Observers as Disturbance
Estimators, SICE Journal of Control, Measurement, and System Integration, 10:3, 125-134, DOI:
10.9746/jcmsi.10.125

To link to this article: https://doi.org/10.9746/jcmsi.10.125

Published online: 18 Jan 2021.

Submit your article to this journal

Article views: 474

View related articles

View Crossmark data

Citing articles: 21 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tmsi20
SICE Journal of Control, Measurement, and System Integration, Vol. 10, No. 3, pp. 125–134, May 2017

Extended High-Gain Observers as

Disturbance Estimators

Hassan K. KHALIL ∗

Abstract : This paper reviews recent results on the use of extended high-gain observers as disturbance estimators. It
shows how the use of these observers enables recovering the performance of feedback linearization controllers. It shows
also how the observer can be combined with dynamic inversion. Three experimental applications are described.

Key Words : nonlinear control, high-gain observers, disturbance estimators.

1. Introduction
Disturbance estimators are widely used in feedback control;
see [1] for a comprehensive coverage of the topic; see also [2]
and [3] for the use of disturbance estimators in Active Distur-
bance Rejection Control (ADRC). One of the common meth-
ods to estimate disturbance is to use extended observers. For a
system of the form
ẋ = f (x, u, w), y = h(x)
Motor with uncertain parameters and unknown load, and an in-
verted pendulum on a cart with uncertain parameters. Finally,
some concluding remarks are given in Section 7.
2. Preliminaries
A single-input–single-output nonlinear system of the form
ẋ = f (x) + g(x)u, y = h(x) (1)
has relative degree n in an open set R if, for all x ∈ R,

f h(x) = 0, for 1 ≤ i ≤ n − 1; Lg L f h(x)  0

Lg Li−1
with constant disturbance w, the state equation can augmented n−1

with the equation ẇ = 0. Provided the extended system is ob-

servable, an observer is designed to estimate the extended state
(x, w). While such observer is designed for constant disturbance
w, it will work for slowly varying disturbance under appropri-
ate conditions. The extended high-gain observer presented in
this paper is also based on extending the state vector by treat-
ing the disturbance as an additional state, but it differs from
other extended observers in two aspects. First, it does not re-
quire the disturbance to be slowly varying because the observer
dynamics are designed to be fast relative to the dynamics of
the extended system. Second, the observer estimates a matched
disturbance term, which results from model uncertainty and ex-
ternal disturbances.
The papers starts by recalling some background informa-
tion in Section 2. Section 3 presents the results of [4] on
the use of extended high-gain observers to recover the perfor-
mance of feedback-linearization controllers in the presence of
uncertainties. Section 4 presents the results of [5], which ex-
tends the results [4] to multi-input–multi-output systems. Sec-
tion 5 presents the results of [6], which combines extended
high-observers with the technique of dynamic inversion [7] to
handle systems with uncertain control coefficient as well as sys-
tems that depend nonlinearly on the control signal. The results
of Sections 3 and 5 have been tested experimentally with suc-
cessful results. Section 6 describes applications to a DC mo-
tor with nonlinear friction, a Permanent Magnet Synchronous
∗
Department of Electrical and Computer Engineering, Michigan
State University, 428 S. Shaw Lane, East Lansing, MI 48824,
USA
E-mail: khalil@egr.msu.edu
(Received February 9, 2017)
(Revised February 19, 2017)
where L f h(x) = ∂h∂x f is the Lie derivative of h with respect to the
vector field f . Under appropriate smoothness conditions on f ,
g, and h, there is a change of variables that transforms a relative
degree n system into the normal form [8, Chapter 4]
η̇ = f0 (η, ξ)
ξ̇i = ξi+1 , for 1 ≤ i ≤ n − 1
ξ̇n = a(η, ξ) + b(η, ξ)u
y = ξ1
at least locally. Under some stronger conditions, the change
of variables will hold globally [8, Chapter 9]. The equation
η̇ = f0 (η, ξ) is called the internal dynamics because it repre-
sents the dynamics of the system when state feedback lineariza-
tion is used to reduce the input-output map, from u to y, to
a chain of integrators. The equation η̇ = f0 (η, 0) is called the
zero dynamics because it represents the dynamics of the system
when the output y is identically zero. The system is minimum
phase when the zero dynamics have an asymptotically stable
equilibrium point. The results of this paper are presented for
minimum-phase systems or systems where the internal dynam-
ics, with input ξ and state η, are bounded-input–bounded-state
stable.
An m-input–m-output nonlinear system of the form

m
ẋ = f (x) + gi (x)ui , yi = hi (x), for 1 ≤ i ≤ m
i=1
has vector relative degree {n1 , n2 , . . . , nm } in an open set R if,
for all x ∈ R,
f hi (x) = 0,
Lg j Lk−1 for 1 ≤ k ≤ ni − 1

JCMSI 0003/17/1003–0125 
c 2017 SICE
 126 SICE JCMSI, Vol. 10, No. 3, May 2017

for all 1 ≤ i ≤ m and 1 ≤ j ≤ m, and the m × m matrix state estimates could peak to a negative power ε before they
⎡ n −1 n −1 ⎤ decay quickly to O(ε) values. In feedback control, the effect
⎢⎢⎢ Lg1 L f 1 h1 · · · · · · Lgm L f 1 h1 ⎥⎥⎥ of the observer peaking on the plant is overcome by saturating
⎢⎢⎢ ⎥
⎢⎢⎢ Lg1 Lnf 2 −1 h2 · · · · · · Lgm Lnf 1 −1 h2 ⎥⎥⎥⎥⎥ the state estimates or the control signal outside a compact set of
⎢⎢⎢⎢ .. .. .. ..
⎥⎥⎥
⎥⎥⎥
⎢⎢⎢ . . . . interest; see [9].
⎢⎣⎢ ⎥⎥⎦⎥
Lg1 Lnf m −1 hm · · · · · · Lgm Lnf m −1 hm 3. Single-Input–Single-Output Systems
is nonsingular. Under appropriate smoothness conditions on Consider a single-input–single-output nonlinear system in
f , g, and h, there is a change of variables that transforms the the globally defined normal form:
system into the multivariable normal form [8, Chapters 5 and 9] η̇ = f0 (η, ξ, w) (2)
η̇ = f0 (η, ξ) ξ̇i = ξi+1 , for 1 ≤ i ≤ n − 1 (3)
ξ̇ij = ξij+1 , for 1 ≤ j ≤ ni − 1, 1 ≤ i ≤ m ξ̇n = a(η, ξ, w) + b(η, ξ, w)u (4)
ξ̇ni i = fi (η, ξ) + Gi (η, ξ)u, for 1 ≤ i ≤ m e = ξ1 (5)
yi = ξ1i , for 1 ≤ i ≤ m where η ∈ R p and ξ = col(ξ1 , . . . , ξn ) ∈ Rn are the state vari-
ables, u is the control input, e is the measured regulation error,
where ξi = col(ξ1i , ξ2i , . . . , ξni i ), ξ = col(ξ1 , ξ2 , . . . , ξm ), and and w ∈ R is an exogenous input. The functions a, b, and f0
n = n1 + · · · + nm . and the signal w could be unknown.
A scalar continuous function α(r), defined for r ∈ [0, a), be-
longs to class K if it is strictly increasing and α(0) = 0. It Assumption 1 w(t) and ẇ(t) are bounded and w(t) belongs to
belongs to class K∞ if it is defined for all r ≥ 0 and α(r) → ∞ a known compact set W ⊂ R .
as r → ∞. A scalar continuous function β(r, s), defined for Assumption 2 The functions a and b are continuously differ-
r ∈ [0, a) and s ∈ [0, ∞), belongs to class KL if, for each entiable with locally Lipschitz derivatives, f0 is locally Lips-
fixed s, β(r, s) belongs to class K with respect to r and, for each chitz, and b(η, ξ, w) ≥ b0 > 0, for all (η, ξ, w) ∈ R p × Rn × W.
fixed r, β(r, s) is decreasing with respect to s and β(r, s) → 0 as
s → ∞. The system ẋ = f (x, u) is input-to-state stable (ISS) Assumption 3 There is a continuously differentiable function
if there exist a class KL function β and a class K function γ V0 (η), class K∞ functions 1 and 2 , and a nonnegative contin-
such that for any t0 ≥ 0, any initial state x(t0 ), and any bounded uous non-decreasing function χ such that
input u(t), the solution x(t) exists for all t ≥ t0 and satisfies 1 ( η ) ≤ V0 (η) ≤ 2 ( η )


x(t) ≤ max β( x(t0 ) , t − t0 ), γ sup u(τ) ∂V0
f0 (η, ξ, w) ≤ 0, ∀ η ≥ χ( ξ + w )
t0 ≤τ,≤t ∂η
for all t ≥ t0 . If this inequality holds only on compact sets of x for all (η, ξ, w) ∈ R p × Rn × W.
and u, the system is said to be regionally input-to-state-stable;
Assumption 3 ensures that the system η̇ = f0 (η, ξ, w), with in-
see, for example, [9] and [10].
put (ξ, w), is bounded-input–bounded-state stable. It is satisfied
A high-gain observer for the system
whenever η̇ = f0 (η, ξ, w) is ISS because ISS is equivalent to
ẋi = xi+1 , for 1 ≤ i ≤ n − 1 the existence of an ISS Lyapunov function V0 (η) that satisfies
ẋn = φ(x, u) the foregoing inequality with [∂V0 /∂η] f0 (η, ξ, w) ≤ −γ0 ( η )
for some class K functions γ0 [11]. Assumption 3, however, is
y = x1
less restrictive than ISS because it does not require the origin of
is given by η̇ = f0 (η, 0, 0) to be asymptotically stable.
The goal is to design an output feedback controller to asymp-
x̂˙i = x̂i+1 + (αi /εi )(y − x̂1 ), for 1 ≤ i ≤ n − 1 totically regulate the output e(t) to zero while meeting cer-
x̂˙n = φ0 ( x̂, u) + (αn /ε )(y − x̂1 )
n tain requirements on the transient response. Had (η, ξ, w) been
available for feedback and the functions a and b been known,
where φ0 is a nominal model of φ, ε is a sufficiently small pos- we could have used the feedback linearization control
itive constant, and α1 to αn are chosen such that the roots of −a(η, ξ, w) + v
u=
sn + α1 sn−1 + · · · + αn−1 s + αn = 0 b(η, ξ, w)
to reduce the input-output map to
have negative real parts. Under appropriate smoothness con-
ditions on φ and φ0 ; see for example [9], the estimation errors ξ̇ = Aξ + Bv, e = Cξ (6)
x̃i = xi − x̂i , for i = 1, . . . n, satisfy the inequality where the triple (A, B, C) represents a chain of n integrators;
 that is,
b
| x̃i | ≤ max i−1 x̃(0) e−at/ε , εn+1−i cM ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
ε ⎢⎢⎢0 1 0 ··· 0⎥⎥⎥ ⎢⎢⎢0⎥⎥⎥ ⎢⎢⎢1⎥⎥⎥
⎢⎢⎢⎢0 0 1 ··· 0⎥⎥⎥⎥
⎥ ⎢⎢⎢ ⎥⎥⎥
⎢ 0 ⎥
⎢⎢⎢ ⎥⎥⎥
⎢⎢0⎥⎥
⎢⎢⎢ ⎢⎢⎢ ⎥⎥⎥
.. ⎥⎥ , B = ⎢⎢ .. ⎥⎥ , C T = ⎢⎢⎢⎢⎢ .. ⎥⎥⎥⎥⎥
⎥
for some positive constants a, b, c, and M. The first term in this ⎥
⎢ ⎥ ⎢ ⎥
inequality is due to the initial estimation error and the second A = ⎢⎢⎢⎢ ... ... ..
.
..
. . ⎥⎥⎥ ⎢⎢⎢ . ⎥⎥⎥ ⎢⎢⎢ . ⎥⎥⎥
⎢⎢⎢ ⎥⎥⎥ ⎢⎢⎢ ⎥⎥⎥ ⎢⎢⎢ ⎥⎥⎥
term gives an ultimate bound due to the error φ − φ0 . The first ⎢⎢⎢0 0 ··· 0 1⎥⎥ ⎥ ⎢ ⎥
⎢⎢⎣0⎥⎥⎦ ⎢⎢⎢0⎥⎥⎥
⎢⎣ ⎦ ⎣ ⎦
term shows the peaking phenomenon of the observer where the 0 0 ··· ··· 0 1 0
 SICE JCMSI, Vol. 10, No. 3, May 2017 127
 
−a(η, ξ, w) − Kξ 
The auxiliary control v could then be designed as a feedback M> max   (13)
function of ξ to stabilize the origin ξ = 0 and meet the transient (η,ξ)∈Ω, w∈W b(η, ξ, w)
response specifications. Let v = −Kξ where K is chosen using Saturating the foregoing expression of u at ±M using the satu-
a linear control design method to make (A − BK) Hurwitz and ration function sat(·) yields the output feedback controller
shape the transient response. Hence, the origin of

ψ(ξ̂, σ̂)
ξ̇ = (A − BK)ξ (7) u = M sat (14)
M
is exponentially stable and V1 (ξ) = ξT P1 ξ is a Lyapunov func- where sat(y) = min{1, |y|} sign(y) and ψ is defined by (12).
tion, where P1 = PT1 > 0 is the solution of the Lyapunov equa- Let
tion P1 (A − BK) + (A − BK)T P1 = −Q1 , for some Q1 = QT1 > 0.  
Let  b(η, ξ, w) − b̂(ξ) 
ku = max  
(η,ξ)∈Ω,w∈W  b̂(ξ)
Ω = {V0 (η) ≤ c0 } × {V1 (ξ) ≤ c}
αn+1
H(s) =
where c > 0 and c0 ≥ 2 (χ(κ1 (c) + κ2 )) in which κ1 (c) =
maxξ∈{V1 (ξ)≤c} ξ and κ2 = maxw∈W w . The set Ω is compact
and positively invariant with respect to the system
η̇ = f0 (η, ξ, w), ξ̇ = (A − BK)ξ
The constants c and c0 can be chosen arbitrarily large so that
any compact set of Rn+p can be put in the interior of Ω.
Let â(ξ) and b̂(ξ) ≥ b1 > 0 be twice continuously differ-
entiable, globally bounded functions that model a(η, ξ, w) and
b(η, ξ, w), respectively. It is allowed to take â = 0 and b̂ > 0 as
a constant. The ξ̇n -equation can be written as
ξ̇n = σ + â(ξ) + b̂(ξ)u
where
σ = a(η, ξ, w) − â(ξ) + [b(η, ξ, w) − b̂(ξ)]u
Augmenting σ as an additional state to the chain of integrators
(3)–(5), a high-gain observer for the extended system is taken
as
ξ̂˙i = ξ̂i+1 + (αi /εi )(e − ξ̂1 ), for 1 ≤ i ≤ n − 1 (8)
ξ̂˙n = σ̂ + â(ξ̂) + b̂(ξ̂)u + (αn /εn )(e − ξ̂1 ) (9)
σ̂˙ = (αn+1 /εn+1 )(e − ξ̂1 ) (10)
where α1 to αn+1 are chosen such that the roots of
sn+1 + α1 sn + · · · + αn+1 = 0
have negative real parts and ε > 0 is a small parameter. Global
boundedness of â and b̂ is required to overcome the peaking
phenomenon of high-gain observers. It does not exclude lin-
ear or unbounded functions because global boundedness can be
achieved by saturating the control outside a compact set.
From the properties of high-gain observers, it is anticipated
that the terms (αi /εi )(e − ξ̂1 ) for 1 ≤ i ≤ n will be O(ε) after a
short transient period. Then, equations (8)–(9) can be viewed
as a perturbation of the system
 ing ε small enough. The third bullet shows that the controller
ξ̂˙ = Aξ̂ + B σ̂ + â(ξ̂) + b̂(ξ̂)u (11)
which can be made to coincide with the target system (7) by
taking
−σ̂ − â(ξ̂) − K ξ̂ def
u= = ψ(ξ̂, σ̂) (12)
b̂(ξ̂)
To protect the system from peaking in the observer’s transient
response, the control is saturated outside the compact set Ω. Let
sn+1 + α1 sn + · · · + αn+1
and H ∞ = maxω |H( jω)|. Because H(0) = 1, H ∞ ≥ 1. It
can be verified that H ∞ = 1 if all the poles of H(s) are real.
Theorem 1 Consider the closed-loop system formed of the
plant (2)–(5), the observer (8)–(10) and the controller (14).
Suppose Assumptions 1 to 3 are satisfied and
1
ku < (15)
H ∞
Let Ω0 be a compact set in the interior of Ω and S a compact
subset of Rn+1 . Then, for all initial conditions (η(0), ξ(0)) ∈ Ω0
and (ξ̂(0), σ̂(0)) ∈ S ,
• there exists ε∗1 > 0 such that for every 0 < ε ≤ ε∗1 , the
trajectories of the closed-loop system are bounded for all
t ≥ 0;
• given any μ > 0 there exist ε∗2 > 0, dependent on μ, such
that for every 0 < ε ≤ ε∗2 ,
ξ(t) − ξr (t) ≤ μ, ∀t≥0 (16)
where ξr is the solution of the target system (7) with
ξr (0) = ξ(0).
• given any μ > 0 there exist ε∗3 > 0 and T 1 > 0, both
dependent on μ, such that for every 0 < ε ≤ ε∗3 ,
ξ(t) ≤ μ, ∀ t ≥ T1 (17)

The proof of Theorem 1 is given in [4].
Remark 1 The second bullet shows performance recovery be-
cause the trajectories of ξ under output feedback can be made
arbitrarily close to the trajectories of the target system by choos-
achieves practical regulation because the ultimate bound on ξ
can be made arbitrarily small by choosing ε small enough. 
Remark 2 Condition (15) can be always satisfied by appropri-
ate choice of b̂ and the observer eigenvalues if an upper bound
on b(x, z, w) is known. Let bm ≥ max{(η,ξ)∈Ω,w∈W} b(η, ξ, w).
Taking b̂ > bm ensures that ku < 1. Assigning the observer
eigenvalues to be real results in H ∞ = 1; hence ku H ∞ < 1.

 128 SICE JCMSI, Vol. 10, No. 3, May 2017

Remark 3 There is interplay between ε and the constants c and integrators in the ith chain. The system can be equivalently rep-
c0 that define Ω. The larger are c and c0 , the larger is the satura- resented by the equations
tion level M, which, in general, requires a smaller ε to achieve
enough separation of times scales between the observer and the η̇ = f0 (η, ξ, w) (22)
plant. This interplay can be interpreted the other way around ξ̇ = Aξ + B[ f (η, ξ, w) + G(η, ξ, w)u] (23)
by saying that a smaller ε would lead to a larger region of at- e = Cx (24)
traction. 
where
Theorem 1 ensures only practical regulation. The next the-
A = block diag[A1 , . . . , Am ]
orem shows that if w is constant, then, under additional con-
ditions on the zero dynamics, the controller provides integral ⎡ ⎤
⎢⎢⎢ 0 1 ··· ··· 0 ⎥⎥⎥
action that ensures regulation of e(t) (in fact the whole vector ⎢⎢⎢ ⎥⎥⎥
⎢⎢⎢ 0 0 1 ··· 0 ⎥⎥⎥
ξ(t)) to zero even when the model uncertainty does not vanish ⎢⎢ .. .. ⎥⎥⎥
Ai = ⎢⎢⎢⎢ . . ⎥⎥⎥
at ξ = 0. The following assumption states the additional re- ⎢⎢⎢ ⎥⎥⎥
⎢⎢⎢ 0 ··· ··· 0 1 ⎥⎥⎥
quirements on η̇ = f0 (η, ξ, w) when w is constant. ⎢⎣ ⎥⎦
0 ··· ··· ··· 0 ni ×ni
Assumption 4
⎡ ⎤
⎢⎢⎢ 0 ⎥⎥⎥
⎢⎢⎢ .. ⎥⎥⎥
⎢⎢ ⎥⎥⎥
• f0 (η, 0, w) is continuously differentiable in η; B = block diag[B1 , . . . , Bm ], Bi = ⎢⎢⎢⎢ . ⎥⎥⎥
⎢⎢⎢ 0 ⎥⎥⎥
⎢⎣ ⎥⎦
• for each w ∈ W, the system η̇ = f0 (η, 0, w) has a unique 1 ni ×1
equilibrium point η ss = η ss (w) ∈ {V0 (η) ≤ c0 }.
C = block diag[C1 , . . . , Cm ]
• With ζ = η − η ss , there is a continuously differentiable

Lyapunov function Va (ζ), possibly dependent on w, and Ci = 1 0 · · · 0
1×ni
class K functions γ1 to γ4 , independent of w, such that
,
γ1 ( ζ ) ≤ Va (ζ) ≤ γ2 ( ζ ) ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
⎢⎢⎢ u1 ⎥⎥⎥ ⎢⎢⎢ e1 ⎥⎥⎥ ⎢⎢⎢ f1 ⎥⎥⎥ ⎢⎢⎢ G1 ⎥⎥⎥
⎢⎢⎢⎢ . ⎥⎥⎥⎥ ⎢⎢⎢ . ⎥⎥⎥ ⎢⎢⎢ . ⎥⎥⎥ ⎢⎢ . ⎥⎥
∂Va u = ⎢⎢⎢ . ⎥⎥⎥ , e = ⎢⎢⎢ . ⎥⎥⎥ , f = ⎢⎢⎢ . ⎥⎥⎥ , G = ⎢⎢⎢⎢⎢ .. ⎥⎥⎥⎥⎥
. ⎢ . ⎥ ⎢ . ⎥
f0 (ζ + η ss , ξ, w) ≤ −γ3 ( ζ ), ∀ ζ ≥ γ4 ( ξ ) ⎢⎣ ⎥⎦ ⎢⎣ ⎥⎦ ⎢⎣ ⎥⎦ ⎢⎣ ⎥⎦
∂ζ um em fm Gm
The functions f0 , f , and G and the signal w could be unknown.
for all (η, ξ, w) ∈ Ω × W
Suppose w satisfies Assumption 1.
• the origin of ζ̇ = f0 (ζ + η ss , ξ, w) is exponentially stable.
Assumption 5 The functions f and G are continuously dif-
Theorem 2 Consider the closed-loop system formed of the ferentiable with locally Lipschitz derivatives, f0 is locally Lip-
plant (2)–(5), the observer (8)–(10) and the controller (14), un- schitz, and the matrix G(η, ξ, w) is nonsingular with bounded
der the assumptions of Theorem 1. Suppose w is constant and inverse, for all (η, ξ, w) ∈ R p × Rn × W.
Assumption 4 is satisfied. Then, there exists ε̄ > 0 such that Similar to the previous section, a target system is taken as
for ε ∈ (0, ε̄), ξ(t) converges to zero as t → ∞. Furthermore,
if f0 (0, 0, w) = 0, then the origin (η, ξ, ϕ) = (0, 0, 0) is an expo- ξ̇ = (A − BK)ξ (25)
nentially stable equilibrium point and Ω0 × S is a subset of the
where K is designed to make (A − BK) Hurwitz and shape the
region of attraction. 
transient response of (25). A Lyapunov function for (25) can
The proof of Theorem 2 is given in [4]. be taken as V1 (ξ) = ξT P1 ξ, where P1 = PT1 > 0 is the solution
of the Lyapunov equation P1 (A − BK) + (A − BK)T P1 = −Q1 ,
4. Multi-Input–Multi-Output Systems for some Q1 = QT1 > 0. Let Ω = {V0 (η) ≤ c0 } × {V1 (ξ) ≤ c},
The globally-defined normal form where c and c0 are positive constants and c0 is chosen as in the
previous section. The set Ω is compact and positively invariant
η̇ = f0 (η, ξ, w) (18) with respect to the system
ξ̇ij = ξij+1 , for 1 ≤ j ≤ ni − 1, 1 ≤ i ≤ m (19)
η̇ = f0 (η, ξ, w), ξ̇ = (A − BK)ξ
ξ̇ni i = fi (η, ξ, w) + Gi (η, ξ, w)u, for 1 ≤ i ≤ m (20)
Let the m-dimensional function fˆ(ξ) and the m × m matrix
ei = ξ1i , for 1 ≤ i ≤ m (21)
Ĝ(ξ) be twice continuously differentiable, globally bounded
represents a nonlinear system with vector relative de- models of f (η, ξ, w) and G(η, ξ, w), respectively, where Ĝ(ξ) is
gree {n1 , . . . nm }, where ξi = col(ξ1i , ξ2i , . . . , ξni i ), ξ = nonsingular with bounded inverse. It is allowed to take fˆ = 0
col(ξ1 , ξ2 , . . . , ξm ), n = n1 + · · · + nm , η ∈ R p and ξ ∈ Rn are and Ĝ as a constant matrix. The ξ̇-equation can be written as
the state variables, u ∈ Rm is the control input, e1 to em are the ξ̇ = Aξ + B[σ + fˆ(ξ) + Ĝ(ξ)u], e = Cξ
measured regulation errors, and w ∈ R is an exogenous input.
Equations (19)–(21) represents m chains of integrators with ni where
 SICE JCMSI, Vol. 10, No. 3, May 2017 129

σ = f (η, ξ, w) − fˆ(ξ) + [G(η, ξ, w) − Ĝ(ξ)]u • given any μ > 0 there exist ε∗2 > 0, dependent on μ, such
that for every 0 < ε ≤ ε∗2 ,
Equivalently, for each 1 ≤ i ≤ m,

ξ̇ij = ξij+1 , for 1 ≤ j ≤ ni − 1 ξ(t) − ξr (t) ≤ μ, ∀t≥0 (34)

ξ̇ni i = σi + fˆi (ξ) + Ĝi (ξ)u, where ξr is the solution of the target system (25) with
ei = ξ1i ξr (0) = ξ(0).

where Ĝi is the ith row of Ĝ. Augmenting σi as an additional • given any μ > 0 there exist ε∗3 > 0 and T 1 > 0, both
state to the ith chain of integrators, a high-gain observer of the dependent on μ, such that for every 0 < ε ≤ ε∗3 ,
extended system is taken as
ξ(t) ≤ μ, ∀ t ≥ T1 (35)
ξ̂˙ij = ξ̂ij+1 + (αij /ε j )(ei − ξ̂1i ), 1 ≤ j ≤ ni − 1 (26)
ξ̂˙ni i = σ̂i + fˆi (ξ̂) + Ĝi (ξ̂)u + (αini /εni )(ei − ξ̂1i ) (27) 
σ̂ ˙ i = (αi /εni +1 )(ei − ξ̂i ) (28) The main part of the proof of Theorem 3 is given in [5]; the rest
ni +1 1
follows the proof of Theorem 1.
for 1 ≤ i ≤ m, where αi1 to αini +1 are chosen such that the
polynomial Remark 4 Because the observer can be designed to have
s ni +1
+ αi1 sni + · · · + αni +1 H ∞ = 1, it is sufficient to require ku < 1. 

is Hurwitz, and ε > 0 is a small parameter. Let The next theorem extends Theorem 2 to the multi-input–
multi-output case.
ψ(ξ, σ) = Ĝ−1 (ξ)[− fˆ(ξ) − σ − Kξ] (29)
Theorem 4 Consider the closed-loop system formed of the
hi (η, ξ, w) be the ith component of G−1 (η, ξ, w)[− f (η, ξ, w) − plant (18)–(21), the observer (26)–(28) and the controller (31),
Kξ], and under the assumptions of Theorem 3. Suppose w is constant
Mi > max |hi (η, ξ, w)| (30) and Assumption 4 is satisfied. Then, there exists ε̄ > 0 such
(η,ξ)∈Ω, w∈W that for ε ∈ (0, ε̄), ξ(t) converges to zero as t → ∞. Further-
The output feedback controller is taken as more, if f0 (0, 0, w) = 0, then the origin (η, ξ, ϕ) = (0, 0, 0) is an

 exponentially stable equilibrium point of the closed-loop sys-
ψi (ξ̂, σ̂) tem and Ω0 × S is a subset of the region of attraction. 
ui = Mi sat , for 1 ≤ i ≤ m (31)
Mi
Let 5. Dynamic Inversion

 
[G(η, ξ, w) − Ĝ(ξ)]DĜ−1 (ξ) (32)
Consider a multi-input–multi-output system represented by
ku = max max
|di |≤1 (η,ξ)∈Ω,w∈W the equations

where D = diag(d1 , . . . , dm ), η̇ = f0 (η, ξ, w) (36)

H(s) = diag[H1 (s), . . . Hm (s)] ξ̇ij = ξij+1 , for 1 ≤ j ≤ ni − 1, 1 ≤ i ≤ m (37)

ξ̇ni i = fi (η, ξ, w, u), for 1 ≤ i ≤ m (38)
αini +1
Hi (s) = ei = ξ1i , for 1 ≤ i ≤ m (39)
sni +1 + αi1 sni + · · · + αini +1
where ξi = col(ξ1i , ξ2i , . . . , ξni i ), ξ = col(ξ1 , ξ2 , . . . , ξm ), n = n1 +
and H ∞ = max1≤i≤m maxω |Hi ( jω)|. Because Hi (0) = 1, · · · + nm , η ∈ R p and ξ ∈ Rn are the state variables, u ∈ Rm is
H ∞ ≥ 1. It can be verified that H ∞ = 1 if for each the control input, e1 to em are the measured regulation errors,
1 ≤ i ≤ m, all the poles of Hi (s) are real. and w ∈ R is an exogenous input. The system is defined for
The following theorem reads almost the same as Theorem 1. η ∈ Dη ⊂ R p and ξ ∈ Dξ ⊂ Rn ⊂ Rn for some domains Dη
and Dξ . The system can be equivalently represented by the
Theorem 3 Consider the closed-loop system formed of the equations
plant (18)–(21), the observer (26)–(28) and the controller (31).
Suppose Assumptions 1, 3 and 5 are satisfied and η̇ = f0 (η, ξ, w) (40)
1 ξ̇ = Aξ + B f (η, ξ, w, u) (41)
ku < (33)
H ∞ e = Cx (42)
Let Ω0 be a compact set in the interior of Ω and S a compact where the block diagonal matrices {A, B, C} are defined in the
subset of Rn+m . Then, for all initial conditions (η(0), ξ(0)) ∈ Ω0 previous section. Equations (36)–(39) are similar to equations
and (ξ̂(0), σ̂(0)) ∈ S , (18)–(21) except that the right-hand side of (38) depends non-
• there exists ε∗1 > 0 such that for every 0 < ε ≤ ε∗1 , the linearly on u.
trajectories of the closed-loop system are bounded for all Suppose w satisfies Assumption 1 and f and f0 satisfy the
following assumption.
t ≥ 0;
 130 SICE JCMSI, Vol. 10, No. 3, May 2017

Assumption 6 The function f is continuously differentiable An estimate σ̂ of σ is provided by the extended high-gain
with locally Lipschitz derivatives and f0 is locally Lipschitz, for observer
all (η, ξ, w) ∈ Dη × Dξ × W.
ξ̂˙ij = ξ̂ij+1 + (αij /ε j )(ei − ξ̂1i ), 1 ≤ j ≤ ni − 1 (48)
Consider the target system ξ̂˙ni i = σ̂i + fˆi (ξ̂, u) + (αini /εni )(ei − ξ̂1i ) (49)
ξ̇ = (A − BK)ξ (43) σ̂ ˙ i = (αi /εni +1 )(ei − ξ̂i ) (50)
ni +1 1

where (A−BK) is Hurwitz. The goal is to design an output feed- for 1 ≤ i ≤ m, where αi1 to αini +1 are chosen such that the
back controller to asymptotically regulate the regulation error e polynomial
to zero while bringing ξ(t) of the closed-loop system arbitrarily sni +1 + αi1 sni + · · · + αni +1
close to the response of the target system when both systems
start from the same initial states. is Hurwitz, and ε > 0 is a small parameter. Let

Assumption 7 For (η, ξ, w) ∈ Dη × Dξ × W, the equation Mi > max |Ki ξ|

ξT P1 ξ≤c

f (η, ξ, w, u) = −Kξ (44) where Ki is the ith row of K,

has a unique solution u = φ(η, ξ, w) and φ is continuously dif-
ferentiable.
It follows that φ̇ is bounded on compact sets of (η, ξ).
Assumption 8 There is a known, locally Lipschitz, m × m non-
singular matrix P(ξ) such that
zT P(ξ)[ f (η, ξ, w, z + φ) − f (η, ξ, w, φ)] ≥ β z 2 (45)
for all (η, ξ, w, z) ∈ Dη × Dξ × W × Rm where β is a positive
constant.
Remark 5 When f (η, ξ, w, u) = h(η, ξ, w) + G(η, ξ, w)u, (45) is
satisfied if
P(ξ)G(η, ξ, w) + (P(ξ)G(η, ξ, w))T ≥ 2βI (46)
 is defined by the extended-high-gain observer (48)–(50) and the
Let V1 (ξ) = ξT P1 ξ, where P1 = PT1 > 0 is the solution of
the Lyapunov equation P1 (A − BK) + (A − BK)T P1 = −Q1 ,
for some Q1 = QT1 > 0. Let c be a positive constant such that
{V1 (ξ) ≤ c} ⊂ Dξ . Suppose Assumption 3 is satisfied. Then,
the compact set Ω = {V0 (η) ≤ c0 } × {V1 (ξ) ≤ c} is positively
invariant with respect to the system
η̇ = f0 (η, ξ, w), ξ̇ = (A − BK)ξ (47)
Let fˆ(ξ, u) be a twice continuously differentiable, globally
bounded function that models f (η, ξ, w, u). The choice fˆ = 0 is
allowed. Set σ = f (η, ξ, w, u) − fˆ(ξ, u). If ξ and σ were avail-
able for feedback, the dynamic inversion algorithm could have
been taken as
μu̇ = −P(ξ)[ fˆ(ξ, u) + σ + Kξ]
 
Ni > max  fi (η, ξ, w, z + φ) − fˆi (ξ, z + φ)
(η,ξ)∈Ω,w∈W,||z ≤r
and define the saturated estimates (K ξ̂) s and σ̂ s by
K1 ξ̂ Km ξ̂
(K ξ̂) s = col M1 sat , . . . , Mm sat
M1 Mm
σ̂1 σ̂m
σ̂ s = col N1 sat , . . . , Nm sat
N1 Nm
The dynamic inversion algorithm is defined by
μu̇ = −P(ξ̂ s )[ fˆ(ξ̂, u) + σ̂ s + (K ξ̂) s ] (51)
where μ is a small positive constant. The estimates on the right-
hand-side of (51) are saturated to protect against peaking of
the high-gain observer. The arguments of fˆ are not saturated
because fˆ is globally bounded. The output feedback controller
dynamic inversion algorithm (51).
Remark 6 The output feedback controller has two small pa-
rameters μ and ε, which determine the time scales of the dy-
namic inversion and the observer, respectively. The observer
dynamics should be faster than the dynamic inversion so that
the estimates ξ̂ and σ̂ approach ξ and σ rapidly. The dynamic
inversion should be faster than the dynamics of the system (47).
Thus, 0 < ε << μ << 1. In tuning these parameters, one would
consider first the closed-loop system under state feedback and
reduce μ until its trajectories are brought close enough to the
trajectories of (47). Then, the closed-loop system under out-
put feedback is considered and ε is reduced until its trajectories
are brought close enough to the trajectories of (47). The tuning
procedure would use simulation while allowing the uncertain
quantities to vary over their respective ranges. 

and z = u − φ(η, ξ, w) would satisfy the equation Theorem 5 Consider the closed-loop system formed of the
plant (36)–(39), the observer (48)–(50) and the dynamic in-
μż = −P(ξ)[ f (η, ξ, w, z + φ) − f (η, ξ, w, φ)] − μφ̇
version algorithm (51). Suppose Assumptions 1, 3, 6, 7 and
Then, from Assumption 8, 8 are satisfied. Let Ω0 be a compact set in the interior of Ω,
U a compact subset of Rm and S a compact subset of Rn+m .
μzT ż ≤ −β z 2 + μ z φ̇ Then, for all initial conditions (η(0), ξ(0)) ∈ Ω0 , u(0) ∈ U, and
(ξ̂(0), σ̂(0)) ∈ S ,
For all (η, ξ, w) ∈ Ω × W and u ≤ r0 , φ̇ is bounded. There-
fore, the set { z ≤ r}, for any r > 0, is positively invariant for • there exist μ∗1 > 0 and λ∗1 > 0 such that for each μ ∈
sufficiently small μ. Taking r ≥ r0 + max(η,ξ,w)∈Ω×W φ(η, ξ, w) (0, μ∗1 ] and ε/μ ∈ (0, λ∗1 ], the trajectories of the closed-loop
ensures that z(t) ≤ r for all u(t) ≤ r0 . system are bounded for all t ≥ 0;
 SICE JCMSI, Vol. 10, No. 3, May 2017 131

• given any ς > 0 there exist μ∗2 > 0 and λ∗2 > 0 such that for where μ0 > 0, σ1 > 0, and σ2 > 0 are positive constants and
each μ ∈ (0, μ∗2 ] and ε/μ ∈ (0, λ∗2 ], g(v) is the Stribeck curve, modeled by

Fc+ + (F s+ − Fc+ ) e−(v/vs ) , v > 0
2

ξ(t) − ξr (t) ≤ ζ, ∀t≥0 (52) g(v) =

Fc− + (F s− − Fc− ) e−(v/vs ) , v < 0
2

where ξr is the solution of the target system (43) with with Fc± and F s± being the Coulomb and static friction values,
ξr (0) = ξ(0).  and v s being the Stribeck velocity.
The goal of the experiment is to have θ asymptotically track
The proof of Theorem 5 is given in [6]. a smooth reference signal r(t), which has bounded derivatives
up to the third order. The system is transformed into the normal
Remark 7 Since limt→∞ ξr (t) = 0, (52) implies that for every
form by the change of variables
ς > 0 there is T > 0 such that for sufficiently small μ and ε/μ,
ξ(t) ≤ ς for all t ≥ T .  x1 = θ − r(t), x2 = θ̇ − ṙ(t), w1 = r̈(t), w2 = ṙ(t)

As in Theorem 2, when w is constant the dynamic inversion which results in

controller provides integral action and stabilization under ap-
ẋ1 = x2 , ẋ2 = au + b(x, z, w), ż = f0 (x, z, w), y = x1
propriate conditions. These results are stated in Theorem 6.
where a = kc /J, b(x, z, w) = −w1 − h(x, z, w)/J,
Theorem 6 Consider the closed-loop system formed of the
|x2 + w2 |
plant (36)–(39), the observer (48)–(50) and the dynamic in- h(x, z, w) = z + (σ1 + σ2 )(x2 + w2 ) − σ1 z
version algorithm (51), under the assumptions of Theorem 5. g(x2 + w2 )
Suppose w is constant and Assumption 4 is satisfied. Then,

1 |x2 + w2 |
there exist μ̄ > 0 and λ̄ > 0 such that for each μ ∈ (0, μ̄] f0 (x, z, w) = (x2 + w2 ) − z
ε0 g(x2 + w2 )
and ε/μ ∈ (0, λ̄], ξ(t) converges to zero as t → ∞. Further-
more, if f (0, 0, w, φ(0, 0, w)) = 0, fˆ(0, φ(0, 0, w)) = 0, and The equation ż = f0 (x, z, w) is not ISS because the origin z =
f0 (0, 0, w) = 0, then the origin (η, ξ, z, ϕ) = (0, 0, 0, 0), where 0 is not asymptotically stable when x2 + w2 = 0. However,
z = u − φ(η, ξ, w), is an exponentially stable equilibrium point Assumption 3 is satisfied with V0 (z) = ε0 z2 because V̇0 ≤ 0
and Ω0 × U × S is a subset of the region of attraction.  for |z| > g(x2 + w2 ). With â as the nominal value of kc /J, the
extended observer is taken as

6. Experimental Testbeds x̂˙1 = x̂2 + 3(y − x̂1 )/ε

The theory described in the foregoing sections has been ex-
perimentally tested in three experiments, which are described in
this section. The DC motor with nonlinear friction is presented

in [4], the Permanent Magnet Synchronous Motor in [12], and
the inverted pendulum on a cart in [13]. We describe the con-
troller derivation in each case, but the reader is referred to the
afore-mentioned papers to see the experimental results. In all
three cases the experiments confirm the theory. The extended
high-gain observer in all cases is a third-order observer with all
the observer eigenvalues at −1/ε. In the first two experiments,
ε = 0.01 and the sampling frequency is 10 kHz. In the third ex-
periment ε = 0.005 and the sampling frequency is 1.667 kHz.
6.1 DC Motor with Friction
The experimental set-up is a DC-motor with an attached rigid
arm, equipped with a tacho generator servo amplifier and an
incremental encoder whose resolution is 1024 imp/turn. Mat-
lab/Simulink in combination with a dSPACE board are used to
obtain a discrete-time version of the controller and implement
it in real time with the sampled period of 0.1 ms. The system is
modeled by
J θ̈ = kc u − F
where J is the moment of inertia, kc is the motor constant, u
is the control signal, and F is an unknown disturbance, mostly
due to friction torque, which is modeled by the dynamic LuGre
model
|θ̇| |θ̇|
μ0 ż = θ̇ − z, F = z + (σ1 + σ2 )θ̇ − σ1
g(θ̇) g(θ̇)
x̂˙2 = â u − r̈ + σ̂ + 3(y − x̂1 )/ε2
σ̂
˙ = (y − x̂1 )/ε3
With K = 25 7 , the control is given by
u = M sat((−σ̂ + r̈ − K x̂)/(âM))
where M = 0.5 is determined by simulation.
6.2 Permanent Magnet Synchronous Motor with Uncer-
tain Parameters and Unknown Load
The mathematical model of a Permanent Magnet Syn-
chronous Motor (PMSM) in the rotor frame of reference is
given by
did
L = −Rid + n p Lωiq + ud
dt
diq
L = −Riq − n p Lωid − km ω + uq
dt
dω
J = km iq − Dω − T L
dt
dθ
= ω
dt
where θ is the rotor position, ω the rotor speed, id and ud are the
direct axis current and voltage, iq and uq are the quadrature axis
current and voltage, and T L is a load torque. The parameters L,
R, n p , km , J, and D are the stator inductance, stator resistance,
number of pole pairs, rotor magnetic flux linkage, moment of
inertia, and damping coefficient, respectively. The goal is to
design a feedback controller, using only measurement of θ, id ,
z and iq , to regulate the speed ω to a reference signal ωre f in
 132 SICE JCMSI, Vol. 10, No. 3, May 2017

the presence of both unknown bounded load T L and parameter regulation, we take it to be zero; the ẋd -equation shows that
uncertainty. This is to be done while shaping the speed transient xd (t) converges to zero. So we are left with the design of a feed-
response and meeting the constraints u2d +u2d ≤ Vmax
2
and i2d +i2q ≤ back controller for iqre f to regulate the speed ω to ωre f . Since
2
Imax . the load T L is unknown and the parameters α, γ, and μ are un-
The control algorithm consists of three parts, fast inner cur- certain, we rewrite the ω̇-equation as
rent loops, speed and disturbance estimation using the mea-
dω
sured position via an extended high-gain observer, and speed = α̂iqre f − γ̂ω + μ̂xq + σ
dt
regulation via feedback linearization. The fast current loops are
made possible by the smallness of the electrical time constant where α̂, γ̂, and μ̂ are nominal values of α, γ, and μ, respec-
L/R and the use of PI controllers. Define the current tracking tively, and
errors ed and eq by 1
σ = (α − α̂)iqre f − (γ − γ̂)ω + (μ − μ̂)xq − TL
ed = idre f − id , eq = iqre f − iq J
Had ω and σ been available for feedback we could have used
where idre f , and iqre f are the direct and quadrature current ref-
feedback linearization to regulate the speed tracking error eω =
erence signals, respectively. The control inputs ud and uq are
ωre f − ω to zero while shaping its transient response. In partic-
taken as
 t ular, taking
ud (t) = k p ed (t) + ki ed (λ) dλ  
1 dωre f
0 iqre f = + γ̂ωre f + (kω − γ̂)eω − μ̂xq − σ
 α̂ dt
t
uq (t) = k p eq (t) + ki eq (λ) dλ yields the target system
0
de∗ω
where k p and ki are the proportional and integral gains, respec- = −kω e∗ω
tively. By substitution of ud and uq in the motor equations, it dt
can be seen that ed and eq satisfy the equations whose transient response is shaped by choosing the positive
gain kω . The choice of kω is limited by the current constraint.
ded didre f R
τ = τ + id − ed Now, the extended high-gain observer
dt dt R + k p re f
1 dθ̂ 3
− τn p ω(iqre f − eq ) − xd = ω̂ + (θ − θ̂)
R + kp dt ε
dω̂ 3
deq diqre f R = α̂iqre f − γ̂ω̂ + μ̂xq + σ̂ + 2 (θ − θ̂)
τ = τ + iq − eq dt ε
dt dt R + k p re f dσ̂ 1
1 km = 3 (θ − θ̂)
+ τn p ω(idre f − ed ) − xq + ω dt ε
R + kp R + kp
  is used to estimate ω and σ by ω̂ and σ̂, respectively. The
where τ = L/(R + k p ), xd = ki ed , and xq = ki eq . By condition (15) is satisfied provided |(α − α̂)/α̂| < 1. With ω and
proper choice of the gains k p and ki , we can meet the voltage σ replaced by their estimates, the output feedback controller is
constraint while making the time constant τ small enough so given by
that ed and eq become much faster than the other state variables.  
1 dωre f
By singular perturbation theory [14], we can reduce the order iqre f = + γ̂ωre f + (kω − γ̂)êω − μ̂xq − σ̂
of the model by replacing ed and eq by their quasi-steady-state α̂ dt
values, obtained by setting τ = 0 in the foregoing equation. The where êω = ωre f − ω̂. Analysis of the closed-loop system
resulting reduced-order model is given by shows that |eω (t) − e∗ω (t)| can be made arbitrarily small for all
dxd ki t ≥ 0 by choosing ε and τ/ε sufficiently small. The need for a
= (Ridre f − xd ) small ratio τ/ε is expected because of the model order reduc-
dt R + kp
tion that is performed before the observer design. The nominal
dxq ki
= (Riqre f + km ω − xq ) value of τ is 2.145 × 10−4 seconds and ε is taken as 0.01, so that
dt R + kp
τ/ε = 2.145 × 10−2 . The controller was tested experimentally
dω 1 with 10 kHz sampling frequency and 2500 PPR incremental en-
= αiqre f − γω + μxq − T L
dt J coder mounted on the motor’s shaft. The electric currents are
dθ
= ω measured with Hall Effect sensors. The control algorithm is
dt implemented using a National Instruments’ Real-Time system.
where idre f and iqre f are viewed as the control inputs, and the
parameters α, γ, and μ are defined by 6.3 Inverted Pendulum on a Cart in the Presence of Un-
certainties
D km
α = k p μ, γ = km μ + , μ= The mathematical model of the inverted pendulum on a cart
J J(R + k p )
is given by
The variables xd and xq are available for feedback because they
are integrations of ed and eq , which are defined in terms of the ẋ1 = x2 , ẋ2 = f x (α1 , α2 , F)
measured currents id and iq . Since idre f does not affect the speed α̇1 = α2 , α̇2 = fα (α1 , α2 , F)
 SICE JCMSI, Vol. 10, No. 3, May 2017 133

where x1 is the position of the cart, x2 its velocity, α1 is the an- satisfy the equations f x − u = 0 and fα + β1 (α1 − αr ) + β2 α2 = 0,
gular displacement of the pendulum measured clockwise from respectively, the dynamic inversion algorithm is taken as
the vertical position, α2 is the angular velocity of the pendulum,   
Ḟ − fx + u
and F is the force applied to the cart. The functions f x and fα ε2
u̇ fα + β1 (α1 − αr ) + β2 α2
are given by
F where 0 < ε2  1. Let f¯x (α1 , α2 , F) and f¯α (α1 , α2 , F) be nom-
fx = + Gx inal models of f x and fα . Define the uncertain terms σ x and σα
m p + mc − m p cos2 α1
by σ x = f x − f¯x and σα = fα − f¯α . The velocities x2 , α2 and the
−F cos α1
fα = + Gα uncertain terms σ x , σα are estimated by the extended high-gain
(m p + mc − m p cos2 α1 )
observers
where
x̂˙1 = x̂2 + (3/ε3 )(x1 − x̂1 )
m p α22 sin α1 − m p g cos α1 sin α1 x̂˙2 = f¯x (α̂1 , α̂2 , F) + σ̂ x + (3/ε23 )(x1 − x̂1 )
Gx =
m p + mc − m p cos2 α1
σ̂
˙ x = (1/ε3 )(x1 − x̂1 )
3
g sin α1 − G x cos α1
Gα = α̂˙ 1 = α̂2 + (3/ε3 )(α1 − α̂1 )

α̂˙ 2 = f¯α (α̂1 , α̂2 , F) + σ̂α + (3/ε23 )(α1 − α̂1 )
m p and mc are the masses of the pendulum and the cart, re-
spectively, g is the acceleration due to gravity, and  is the σ̂
˙ α = (1/ε3 )(α1 − α̂1 )
3
length of the pendulum. The angle α1 is restricted to the do- The observer parameter ε3 is chosen such that ε3 /ε2  1 to
main −π/2 < α1 < π/2. The goal is to stabilize the pendulum make the observer dynamics faster than the dynamic inversion.
at the origin using measurements of x1 and α1 , despite uncer- The output feedback controller is given by
tainties in the system parameters.    
Taking Ḟ − fˆx + u
ε2 = ˆ
u̇ fα + β1 (α1 − α̂r ) + β2 M1 sat (α̂2 /M1 )
F = (m p + mc − m p cos2 α1 )(u − G x )
where
with
fˆx = f¯x (α1 , M1 sat (α̂2 /M1 ), F) + M2 sat (σ̂ x /M2 )
[β1 (α1 − αr ) + β2 α2 ]
u = g tan α1 + fˆα = f¯α (α1 , M1 sat (α̂2 /M1 ), F) + M3 sat (σ̂α /M3 )
cos α1

−1 −γ1 x1 − γ2 x̂2
where β1 and β2 are positive constants and αr is a reference α̂r = tan
g
trajectory for α1 , results in
The estimates α̂2 , σ̂ x , and σ̂α are saturated outside the compact
ẋ1 = x2 set of operation with the saturation levels M1 , M2 , and M3 , re-


 spectively. The estimate x̂2 is not saturated because it enters
ẋ2 = g tan α1 + [β1 (α1 − αr ) + β2 α2 ]
cos α1 through the globally bounded arctan function.
α̇1 = α2 In the experimental setup, a 6V-DC motor with a planetary
α̇2 = −β1 (α1 − αr ) − β2 α2 gearhead (reduction ratio 3.71 : 1) drives the cart on a track.
The angle of the pendulum and the position of the cart are
The control strategy is to choose the controller parameters to measured by optical encoders with resolution 1024 lines per
make the pendulum dynamics much faster than the cart dynam- revolution. The experimental hardware was interfaced with
ics so that α1 and α2 converge quickly to αr and 0. Then, the dSPACE board and the controller was implemented in the MAT-
cart dynamics can be represented by the reduced model LAB/Simulink environment with sampling period 0.6 ms. To
reduce the effect of measurement noise, the encoder signals
ẋ1 = x2 , ẋ2 = g tan αr
were passed through a low-pass filter with 1 kHz cutoff fre-
With quency. The parameters ε1 to ε3 were taken as ε1 = 0.2,

 ε2 = 0.01, ε3 = 0.005.1
−γ1 x1 − γ2 x2
αr = tan−1
g 7. Conclusions
the cart dynamics are given by This paper reviewed some recent results on the use of ex-
tended high-gain observers as disturbance estimators. Three
ẋ1 = x2 , ẋ2 = −γ1 x1 − γ2 x2 experimental testbeds are described. High-gain observers are
With positive gains γ1 and γ2 , x1 and x2 converge to zero; con- known to be sensitive to measurement noise [15]. The ex-
sequently, αr converges to zero, showing that the whole state perimental results presented here show that successful results
vector converges to the origin. The two time-scale behavior is can be obtained without using very large gains (equivalently,
achieved by taking γ1 = ε21 k1 and γ2 = ε1 k2 , where k1 , k2 , and very small ε). In all three experiments, the observer’s order
ε1 are positive constants with ε1  1.
In the presence of parameter uncertainties, f x and fα are un-
known. These functions are estimated using extended high-gain
observers. In preparation for that, dynamic inversion is used to
compute F and u in terms of f x and fα . Noting that F and u
was three. For higher-order observers, the issue of measure-
ment noise could be challenging. More experimental testing is
needed to determine how high the observers’s order could be
and still achieves successful results.
1
See [13] for the other controller parameters.
 134 SICE JCMSI, Vol. 10, No. 3, May 2017

Acknowledgments
This work was supported by the National Science Foundation Hassan K. KHALIL
under grant number ECCS–1128476. He received the B.S. and M.S. degrees in electrical
References engineering from Cairo University, Egypt, in 1973 and
1975, respectively, and the Ph.D. degree from the Univer-
[1] S. Li, J. Yang, W-H. Chen, and X. Chen: Disturbance
sity of Illinois, Urbana-Champaign, in 1978, all in electri-
Observer-Based Control: Methods and Applications, CRC
cal engineering. Since 1978, he has been with Michigan
Press, 2014.
State University (MSU), where he is currently University
[2] Z. Gao: Active disturbance rejection control: A paradigm shift
Distinguished Professor of Electrical and Computer En-
in feedback control system design, Proc. American Control
gineering. He has consulted for General Motors and Delco Products, and
Conf., pp. 2399–2405, 2006.
published over 100 papers on singular perturbation methods and nonlinear
[3] J. Han: From PID to active disturbance rejection control, IEEE
control. He is the author of Nonlinear Control (Pearson 2015), Nonlin-
Trans. Ind. Electron., Vol. 56, No. 3, pp. 900–906, 2009.
ear Systems (Macmillan 1992; Prentice Hall 1996 & 2002) and coauthor
[4] L.B. Freidovich and H.K. Khalil: Performance recovery of
of Singular Perturbation Methods in Control: Analysis and Design (Aca-
feedback-linearization-based designs, IEEE Trans. Automat.
demic Press 1986; SIAM 1999). He was named IEEE Fellow in 1989 and
Contr., Vol. 53, No. 10, pp. 2324–2334, 2008.
IFAC Fellow in 2007. He received the 1989 IEEE-CSS George S. Axelby
[5] L. Wang, A. Isidori, and H. Su: Output feedback stabilization
Outstanding Paper Award, the 2000 AACC Ragazzini Education Award,
of nonlinear mimo systems having uncertain high-frequency
the 2002 IFAC Control Engineering Textbook Prize, the 2004 AACC O.
gain matrix, Syst. Contr. Lett., Vol. 83, pp. 1–8, 2015.
Hugo Schuck Best Paper Award, the 2009 AGEP Faculty Mentor of the
[6] J. Lee, R. Mukherjee, and H.K. Khalil: Output feedback per-
Year Award, and the 2015 IEEE-CSS Bode Lecture Prize. At MSU he re-
formance recovery in the presence of uncertainties, Syst. Contr.
ceived the 1983 Teacher Scholar Award, the 1994 Withrow Distinguished
Lett., Vol. 90, pp. 31–37, 2016.
Scholar Award, and the 1995 Distinguished Faculty Award. He was named
[7] N. Hovakimyan, E. Lavretsky, and C. Cao: Dynamiv inver-
University Distinguished Professor in 2003. He served as Associate Editor
sion for multivariable non-affine-in-control system via time-
of the IEEE Transactions on Automatic Control, Automatica, and Neural
scale separation, Int. J. Contr., Vol. 81, No. 12, pp. 1960–1967,
Networks, and as Editor of Automatica for nonlinear systems and control.
2008.
He was Registration Chair of the 1984 CDC, Finance Chair of the 1987
[8] A. Isidori: Nonlinear Control Systems, 3rd edition, Springer-
ACC, Program Chair of the 1988 ACC, and General Chair of the 1994
Verlag, 1995.
ACC.
[9] H.K. Khalil: Nonlinear Control, Pearson, 2015.
[10] E.D. Sontag: Mathematical Control Theory, Springer Verlag,
1990.
[11] E.D. Sontag and Y. Wang: On characterizations of the input-
to-state stability property, Syst. Contr. Lett., Vol. 24, No. 5,
pp. 351–359, 1995.
[12] A.A. Alfehaid, E.G. Strangas, and H.K. Khalil: Speed control
of permanent magnet synchronous motor using extended high-
gain observer, Proc. American Control Conf., pp. 2205–2210,
2016.
[13] J. Lee, R. Mukherjee, and H.K. Khalil: Output feedback sta-
bilization of inverted pendulum on a cart in the presence of
uncertainties, Automatica, Vol. 54, pp. 146–157, 2015.
[14] P.V. Kokotović, H.K. Khalil, and J. O’Reilly: Singular Pertur-
bations Methods in Control: Analysis and Design, Academic
Press, 1986. Republished by SIAM, 1999.
[15] J.H. Ahrens and H.K. Khalil: High-gain observers in the pres-
ence of measurement noise: A switched-gain approach, Auto-
matica, Vol. 45, No. 4, pp. 936–943, 2009.