International Journal of Systems Science

Vol. 00, No. 00, January 2005, 1–24

Super twisting algorithm based step-by-step sliding mode

observers for nonlinear systems with unknown inputs

T. FLOQUET† , §,? and J. P. BARBOT‡ , §

†
LAGIS, UMR CNRS 8146, Ecole Centrale de Lille,
Cité scientifique, BP 48, 59651 Villeneuve d’Ascq Cedex, France
‡
ECS, ENSEA, 6 avenue du Ponceau, 95014 Cergy, France
§Projet ALIEN, INRIA-Futurs
(Received 00 Month 200x; In final form 00 Month 200x)

This paper highlights the interest of step-by-step higher order sliding mode observers for MIMO
nonlinear systems with unknown inputs. A structural matching condition, stating on the possibility
to design such observers and to reconstruct the unknown inputs, is derived. A finite time sliding
mode observer, based on the hierarchical use of the super twisting algorithm, is developed. Then, it
is shown that this observer is of interest in the field of hybrid systems and systems with observability
singularities. Lastly, it is shown through an example how to relax the usual matching condition by
the means of this type of finite time sliding mode observer.

Keywords: Nonlinear systems, unknown inputs, sliding mode observers, super twisting algorithm,
matching condition.

1 Introduction

The problem of designing an observer for a multivariable system partially

driven by unknown inputs has been widely studied. Such observers can be of
important use for systems subject to disturbances or with inaccessible inputs,
and in many applications such as fault detection and identification or parame-
ter identification. Many different approaches have been considered to design
unknown input observers: conventional Luenberger design procedure under
some decoupling and detectability conditions (Guan and Saif 1991, Hou and
Müller 1992, Darouach et al. 1994) for linear systems, or high gain observers
(Gauthier et al. 1992, Busawon and Saif 1999, Hou et al. 2000) and algebraic
methods (Reger et al. 2005) in the nonlinear case.

? Corresponding author
c
International Journal of Systems Science ISSN 0020-7721 print/ ISSN 1464-5319 online °2005 Taylor & Francis Ltd
http://www.tandf.co.uk/journals
DOI: 10.1080/00207720xxxxxxxxxxxxx
2 Step-by-step sliding mode observers for nonlinear systems with unknown inputs

Other works rely on the use of discontinuous observers and sliding mode
methodology (see Utkin 1992, Edwards and Spurgeon 1998 or Perruquetti and
Barbot 2002 for further information on sliding mode control and observation).
The system trajectories are constrained to reach and stay, after a finite time,
on a given sliding manifold for which the output error is zero. The sliding
motion provides an estimation of the system state, asymptotically (see e.g.
Walcott and Żak 1987, Slotine et al. 1987, Utkin 1992, Edwards and Spurgeon
1994, Tan and Edwards 2001, Poznyak 2003, Shtessel et al. 2003) or in finite
time (for instance, one can refer to the works of Drakunov 1992, Drakunov
and Utkin 1995, Barbot et al. 1996, Haskara et al. 1998, Utkin et al. 1999 for
design methods based on first order sliding mode algorithms and the works of
Levant 1998, Bartolini et al. 2000, Saadaoui et al. 2006 for second order ones).
Sliding mode observers have been shown to be a robust observation scheme
in robotics (Canudas de Wit and Slotine 1991, Hernandez and Barbot 1996,
Manamanni et al. 2001), electrical engineering (Djemaı̈ et al. 1999, Utkin et al.
1999) or chemical reactors (Martinez-Guerra et al. 2004). Specific applications
of asymptotic and finite time observers for systems with unknown inputs can
be found in a large number of fields such as fault detection and identification
(Edwards et al. 2000, Goh et al. 2002, Tan and Edwards 2003, Floquet et al.
2004), parameter identification (Floret-Pontet and Lamnabhi-Lagarrigue 2001,
Shtessel and Poznyak 2004), cryptography (Barbot et al. 2003), or mechanical
systems with Coulomb friction (Davila et al. 2005, Garrido et al. 2004, Orlov
et al. 2003).
The first part of this paper deals with the possibility to design finite time
sliding mode observers for autonomous MIMO nonlinear systems subject to
unknown inputs. Finite time sliding mode observers, are usually designed un-
der the assumptions that the system can be put in a set of triangular ob-
servable forms, where the unknown inputs act only on the last dynamics of
each triangular form. This assumption is known as the observability matching
condition (it can be seen as the analogue of the well-known matching condi-
tion given in Drazenovic (1969) for a sliding mode controller to be insensitive
to matched perturbations). This condition was explicitly given in the case of
SISO observable linear systems in Perruquetti and Barbot (2002), Chapter 4.
Here, the condition is derived for MIMO nonlinear systems and is expressed
in terms of the observability indices of the system and its relative degrees with
respect to the unknown inputs. Then, a finite time higher order sliding mode
observers, based on the iterative use of the super twisting algorithm (a second
order sliding mode algorithm), is introduced in order to obtain the state and
the unknown inputs in finite time.
The finite time convergence property of sliding mode observers is often de-
sirable in the framework of observation and particularly for the purpose of
observer-based controller design for nonlinear systems. Then, for a large class
 T. Floquet and J.-P. Barbot 3

of nonlinear systems, the observer can be designed separately from the con-
troller and the separation principle does not need to be proved. As shown in
many papers, it can also be of paramount importance in applications that
require fast estimations of some unknown inputs like fault detection and iden-
tification or on-line parameter identification. In the second part of this paper,
other relevant applications are described. It is shown that finite time higher
order sliding mode observers can cope with the problem of observability singu-
larities that may occur in nonlinear systems. A straightforward application can
be found in the domain of secure communication and cryptography, where the
confidential information (the unknown inputs) can be hidden in observability
bifurcations. Variable structure observers can also be of high interest to solve
the observation problem of some class of hybrid systems (switched systems).
Lastly, it is shown through an example that the matching conditions com-
monly assumed when designing sliding mode observers can be relaxed using
the given sliding mode observer.
Notations: LX ψ denotes the classical Lie derivative of the function ψ along
the vector field X and dψ is the differential of the function ψ. The codistrib-
ution ∆⊥ stands for the annihilator of the distribution ∆.

2 Observability form and super twisting algorithm based step-by-step

sliding mode observer

Consider the following autonomous nonlinear system subject to unknown in-

puts:
½ P
ẋ = f (x) + m
i=1 gi (x)wi
(1)
y = h(x) = [h1 (x), .., hp (x)]T

where x ∈ U , a neighborhood of Rn , is the state vector, y ∈ Rp is the output

vector and where w = [w1 , ..., wm ] ∈ Rm represents the unknown inputs.
The vector fields f and g1 , ..., gm , and the functions h1 , ..., hp are assumed
to be sufficiently smooth on U . It is assumed that p ≥ m, and without loss
of generality that, for all x ∈ U , the codistribution span {dh1 , . . . , dhp } is
nonsingular. The system:
½
ẋ = f (x)
y = h(x)

is supposed to be locally weakly observable (see Hermann and Krener 1977

for a definition in the single output case). Hereafter, the definition given in
Krener and Respondek (1985) for the case of several outputs is recalled.
4 Step-by-step sliding mode observers for nonlinear systems with unknown inputs

Definition 2.1 The system is locally weakly observable at x0 if there exists

V(x0 ), a neighborhood of x0 , and p integers (k1 , k2 , . . . , kp ) that form, after
a possible reordering of the outputs, the smallest p-tuple with respect to the
lexicographic ordering, such that:
(i) k1 ≥ k2 ≥ ... ≥ kp ≥ 0
(ii)

p
X
ki = n
i=1

(ii)
 
dh1 (x)
 
dLf h1 (x)
 .. 
 
 . 
 k1 −1 
 dLf h1 (x) 
 
rank  .. =n
 . 
 
 dh p 
 
 .. 
 . 
kp −1
dLf hp (x)

for all x ∈ V(x0 ).

The integers (k1 , k2 , . . . , kp ) are called the observability indices of (1).

2.1 Observability form

Finite time sliding mode observers are usually designed under the assumption
that the system can be put into a specific block triangular observable form
where the unknown inputs appear in the last dynamics of each block only. Here,
we give conditions for the MIMO nonlinear system (1) to be transformed into
such a form. Let us first define the following integers {ρ1 , ..., ρp } such that, for
i = 1 : p:

Lgj Lkf hi (x) = 0, for k < ρi − 1, and for all j = 1 : m,

Lgj Lρfi −1 hi (x) 6= 0, for at least one j = 1 : m,

for all x ∈ V(x0 ).

 T. Floquet and J.-P. Barbot 5

Proposition 2.2 Assume the system (1) is locally weakly observable. There
exists a change of coordinates such that (1) is locally transformed into

ξ˙i = Ai ξ i + Fi (ξ) + Gi (ξ)w (2)

£ ¤
yi = 1 0 . . . 0 ξ i (3)

where
h¡ ¢ iT
T
ξ= ξ 1 . . . (ξ p )T , ξ i ∈ Rki

   
0 1 0 0 0 0
0 0 1 00  .. 
. .. 
Ai =  .. .. .. ∈ Rki ×ki , Fi (ξ) =  . 
 ∈ Rki ,
 .. . . . . 
 0 
0 0 0 1 0
ki
Lf hi (x)
0 0 0 0 0
 
0 0 0
 .. .. .. 
 . . . 
Gi (ξ) =   ∈ Rki ×m
 0 0 0 
Lg1 Lkf i −1 hi (x) . . . Lgm Lfki −1 hi (x)

if and only if ki ≤ ρi for all i = 1 : p and for all x ∈ V(x0 ).

Proof Since the system (1) is locally weakly observable, the following change
of coordinates
n o
ξ = φ(x) = col hi (x), ..., Lkf i −1 hi (x)
i=1:p

is a local diffeomorphism. Since ki ≤ ρi for all i = 1 : p, one has:

 
m
X
dhi (x) ∂hi (x) 
= f (x) + gj (x)wj  = Lf hi (x)
dt ∂x
j=1
 
m
X
d (Lf hi (x)) ∂Lf hi (x) 
= f (x) + gj (x)wj  = L2f hi (x)
dt ∂x
j=1

..
.
³ ´
d Lkf i −1 hi (x) m
X
= Lkf i hi (x) + Lgj Lkf i −1 hi (x)wj
dt
j=1
6 Step-by-step sliding mode observers for nonlinear systems with unknown inputs

Thus, the system is described by (2)-(3) in the new coordinates, with Gi (ξ) 6= 0
if and only if ki = ρi . ¤
As a consequence, the system (1), even if it is weakly observable,
Pp can not be
put into the block triangular observable form (2)-(3) if i=1 ρi < n.
Remark 1 The condition ki ≤ ρi can also be stated in terms of codistribution.
Indeed,

ki ≤ ρi ⇐⇒ gj (x) ∈ Ω⊥ 0
i (x) , for all x ∈ V(x ), for all j = 1 : m,

where
n o
Ωi (x) = span dhi (x), dLf hi (x), . . . , dLkf i −2 hi (x) .

Remark 2 For sake of simplicity, the case of nonlinear systems with known
inputs is not discussed here. It is known that observability of nonlinear systems
depends on known control inputs u. However, the result given in Proposition
2.2 can be extended for uniformly observable systems, i.e. systems which are
observable independently of the inputs (see Gauthier and Bornard (1981) for
a work on uniformly observable systems and Barbot et al. (1996) for a sliding
mode observer design for such systems).
Finite time sliding mode observers are usually designed for system in the
form (2)-(3). One can refer to the works Haskara et al. (1998), Utkin et al.
(1999) for linear and linear time varying systems or to Drakunov (1992),
Drakunov and Utkin (1995), Barbot et al. (1996), Xiong and Saif (2001)1
for nonlinear systems. All of them are based on the analyze of the so-called
equivalent output injection (which is the counterpart of the equivalent control
in the design of sliding mode control). The estimation of the state is based
on a procedure using the successive filtered values of equivalent output injec-
tions obtained from recursive first order sliding mode observers. However, the
approximation of the equivalent output injections by low pass filters at each
step may introduce some delays that could lead to inaccurate estimates or to
instability for high order systems2 . To overcome this problem, a second order
sliding mode algorithm can be used in the innovative term of the observer. In
the next part, a super twisting based sliding mode observer is designed for the
finite time estimation of a n-order system in the form (2)-(3).

1 Notethat in Xiong and Saif (2001), a first order sliding mode observer is designed for a system with
asymptotically stable zero dynamics. This results in a partial finite time stabilization of the state.
2 In Haskara and Özgüner (1999), the implementation of such filters was studied and a relation
between the estimation accuracy, the filter time constants and the sampling time was given.
 T. Floquet and J.-P. Barbot 7

2.2 A finite time step-by-step sliding mode observer

A robust and finite time exact differentiator based on the super twisting algo-
rithm (a second order sliding mode algorithm (Levant 1993)) was introduced
in Levant (1998) and successfully applied in Garrido et al. (2004) and Pisano
and Usai (2004). This approach was extended to the design of arbitrary or-
der robust exact differentiators with finite time convergence in Levant (2003,
2005) using homogeneity properties and in Bejarano et al. (2006) using some
hierarchical procedure. In Barbot et al. (1996), a so-called step-by-step first
order sliding mode observer for the finite time estimation of the state vari-
ables was developed. Hereafter, a similar observer based on the super twisting
algorithm is given.

2.2.1 SISO case. Let us consider a SISO nonlinear system in triangular

observable form:

½
ẋ1 = x2 , . . . , ẋn−1 = xn
(4)
ẋn = α(x) + β(x)ω

where x = [x1 , · · · , xn ]T ∈ Rn is the state vector, y = x1 ∈ R is the output vec-

tor and ω ∈ R is the unknown input. α(x) and β(x) are bounded smooth scalar
functions. Let us assume that the state of the system is uniformly bounded,
i.e. for all t > 0, |xi (t)| < di , and that for all t > 0:

0
|ω(t)| < K1 , |ω̇(t)| < K1 ,
0
|α(t)| < K2 , |α̇(t)| < K2 ,
¯ ¯
¯ ¯ 0
|β(t)| < K3 , ¯β̇(t)¯ < K3 ,
8 Step-by-step sliding mode observers for nonlinear systems with unknown inputs

0
where Ki and Ki are some known positive scalars. Let us design the following
observer:


 x̂˙ = x̃2 + λ1 |e1 |1/2 sign(e1 )
 ˙1


 x̃2 = α1 sign(e 1)

 ˙ h i

 x̂ = E x̃ + λ |e |1/2
sign(e )

 2 1 3 2 2 2




 x̃˙ 3 = E1 αh2 sign(e2 ) i

 x̂˙ = E x̃ + λ |e |1/2 sign(e )

 3 2 4 3 3 3


..
. (5)



 ˙
x̃n−1 = En−3 αhn−2 sign(en−2 )

 i

 x̂˙ n−1 = En−2 x̃n + λn−1 |en−1 |1/2 sign(en−1 )




 x̃˙ n = En−2 αn−1 sign(en−1 )


 h i

 ˙ 1/2
x̂
 n
 = E n−1 θ̃ + λ |e
n n | sign(e n )


˙
θ̃ = En−1 αn sign(en )
h iT h iT
where ei = x̃i − x̂i for i = 1, .., n, with x̃1 = x1 and x̃, θ̃ = x̃1 , x̃2 , ...x̃n , θ̃
is the output of the observer. For i = 1, ..., n − 1, the scalar functions Ei are
defined as

Ei = 1 if |ej | = |x̃j − x̂j | ≤ ε, for all j ≤ i else Ei = 0.

where ε is a small positive constant. The observer gains λi and αi are positive
scalars.
The convergence of the state observation error is obtained in (n − 1) steps
and in finite time.
First step: Assume e1 (0) 6= 0, the error dynamics is given by:

ė1 = x2 − x̃2 − λ1 |e1 |1/2 sign(e1 ) (6)

x̃˙ 2 = α1 sign(e1 ) (7)
h i
ė2 = α1 sign(e1 ) − E1 x̃3 + λ2 |e2 |1/2 sign(e2 )
h i
ėi = Ei−2 αi−1 sign(ei−1 ) − Ei−1 x̃i+1 + λi |ei |1/2 sign(ei ) , i = 3, ..., n − 1
h i
ėn = En−2 αn−1 sign(en−1 ) − En−1 θ̃ + λn |en |1/2 sign(en )

Equations (6)-(7) coincide with the so-called super twisting algorithm and the
remaining dynamics is bounded in finite time. The second time derivative of
 T. Floquet and J.-P. Barbot 9

e1 is given by:

1 1
ë1 = x3 − α1 sign (e1 ) − λ1 ė1 |e1 |− 2 .
2

Sufficient conditions for the finite time convergence on the second order sliding
set {e1 = ė1 = 0} are:

α1 > d3 ,
α1 + d3
λ21 > 4d3 ,
α1 − d3

and one can find an upper bound of the convergence time (see Levant 1998,
Davila et al. 2005, Saadaoui et al. 2006). The equivalent dynamics provides
after a finite time T1 a continuous estimation of x2 (without any introduction
of low pass filters): x̃2 = x2 . The functions Ei are introduced so that the ei ’s,
i = 2, ..., n, do not escape two far before the good information on x̂1 and x̃2 is
obtained. Indeed,

ė2 = α1 sign(e1 ) (8)

ėi = 0, i = 3, ..., n

except on a finite number of finite time intervals where |e1 | may be less than
ε before the second order sliding set {e1 = ė1 = 0} is reached.
Second step: after a finite time, one has E1 = 1 and the observer dynamics
becomes:

ė1 = 0
ė2 = x3 − x̃3 − λ2 |e2 |1/2 sign(e2 )
x̃˙ 3 = α2 sign(e2 )
h i
ė3 = α2 sign(e2 ) − E2 x̃4 + λ3 |e3 |1/2 sign(e3 )
h i
ėi = Ei−2 αi−1 sign(ei−1 ) − Ei−1 x̃i+1 + λi |ei |1/2 sign(ei ) , i = 4, ..., n − 1
h i
ėn = En−2 αn−1 sign(en−1 ) − En−1 θ̃ + λn |en |1/2 sign(en )

Again, the structure of a super twisting algorithm appears. Thus, the trajecto-
ries converge in finite time onto {e2 = ė2 = 0}. As a consequence, x̃3 provides
an estimation of x3 .
10 Step-by-step sliding mode observers for nonlinear systems with unknown inputs

Following the same scheme till the (n − 1)-th step, an estimation of the
whole state x is obtained in finite time. The finite time estimation of the
unknown inputs can be obtained with the n-th step of the observer.
n-th step: the observer dynamics is given by:

 ė1 = . . . = ėn−1 = 0

ėn = α(x) + β(x)ω − θ̃ − λn |en |1/2 sign(en )

 θ̃˙ = α sign(e )
n n

and a sliding motion appears after a finite time on the sliding manifold en =
ėn = 0. One has:

1 1
ën = α̇(x) + β̇(x)ω + β(x)ω̇ − αn sign (en ) − λn ėn |en |− 2 .
2

Thus, choosing
0 0 0
αn > K2 + K3 K1 + K3 K1 ,
³ 0 ´ α + K0 + K0 K + K K0
0 0 n 3 1 3 1
λ2n > 4 K2 + K3 K1 + K3 K1 ¡ 20 0 0
¢,
αn − K2 + K3 K1 + K3 K1

a second order sliding motion appears on {en = ėn = 0}. Then, a continuous
approximation ω̃ of the unknown variables, on every subset of Rn where β(x) 6=
0, is:
h i
ω̃ = β −1 (x̃) θ̃ − α(x̃) .

The successive equivalent output injections on the sliding manifolds are ob-
tained in a continuous way (they are all outputs of some integrators and so
they are less affected by the chattering phenomenon) and without any intro-
duction of low pass filters.

2.2.2 MIMO case. Consider a system with p outputs and m unknown in-
puts in the form (2)-(3). It is constituted of p blocks in a triangular observable
form similar to (4). The design of an observer (5) for each block results in the
finite time estimation of the state variables and the following relations:
m
X
Lkf i hi (x) + Lgj Lkf i −1 hi (x)wj = θ̃i , i = 1 : p (9)
j=1
 T. Floquet and J.-P. Barbot 11

where θ̃i is the equivalent output injection of the last observer variable of each
block. The relations (9) can be rewritten as:

   
Lg1 Lkf 1 −1 h1 (x) ... Lgm Lkf 1 −1 h1 (x) θ̃1 − Lkf 1 h1 (x)
 .. ..   .. 
 . . w =  . 
k −1 k −1 k
Lg1 Lf p hp (x) ... Lgm Lf p hp (x) θ̃p − Lf p hp (x)

Let us assume that the distribution span {g1 (x), . . . , gm (x)} is nonsingular for
all x ∈ V(x0 ). Then, the matrix

 
Lg1 Lkf 1 −1 h1 (x) ... Lgm Lkf 1 −1 h1 (x)
 .. .. 
Λ(x) =  . . 
k −1 k −1
Lg1 Lf p hp (x) ... Lgm Lf p hp (x)

has full column rank since:

 
dLkf 1 −1 h1 (x)
 .. ¡ ¢
Λ(x) =  .  g1 (x) ... gm (x)
k −1
dLf p hp (x)

Thus, an estimation of the unknown input is given by:

 
θ̃1 − Lkf 1 h1 (x̃)
 .. 
w̃ = Λ+ (x̃)  . ,
k
θ̃p − Lf p hp (x̃)

where Λ+ is a well defined pseudo-inverse of Λ.

Remark 3 In this case, the p × m matrix
 
Lg1 Lρf1 −1 h1 (x) ... Lgm Lρf1 −1 h1 (x)
 .. .. 
Γ(x) =  . . 
ρ −1 ρ −1
Lg1 Lfp hp (x) ... Lgm Lfp hp (x)

has also full column rank. Then, the set of integers {ρ1 , ..., ρp } is exactly the
vector relative degree as it is defined in Isidori (1995), p. 220.
12 Step-by-step sliding mode observers for nonlinear systems with unknown inputs

3 Some applications of higher order sliding mode observers

3.1 Variable structure observer and observability singularity

Loss of observability may locally appear in nonlinear systems because of the
nonlinear structure itself (Barbot et al. 2003) or due to singular inputs (i.e.
inputs that make the system unobservable, see Gauthier and Bornard (1981)).
This section highlights the efficiency of finite time sliding mode observers in
the case of systems with local observability singularity. Roughly speaking, the
variable structure of the sliding mode observer is very well suited to the change
of structural properties like the transient between full state observability and
partial state observability. This point is illustrated in the case of the following
nonlinear system with an unknown input m:

   
ẋ1 x2
 ẋ2  =  −x1 + x2 x3  := f (x) + g(x)m (10)
ẋ3 2 2 3
x3 − (1 + x1 + x2 )x3 + m
y = x1 := h(x) (11)

This system is locally weakly observable because the rank condition given in
Hermann and Krener (1977) is satisfied, i.e.:

   
dh 1 0 0
 dLf h   0 1 0 
 2   −1 
 
rank  dLf3 h  = rank 
 ∗
x3 x2
2 2 2 
=3
 dLf h   ∗ −x1 + x2 (1 + 2x3 − 3x3 (1 + x1 + x2 )) 
  .. .. ..
..
. . . .

Actually, for x2 = 0, the observability of the system is obtained at least with

the third time derivative of y. Nevertheless, (10) does not satisfy the observ-
ability matching condition given in Proposition 2.2 if x2 = 0. This means
that usual classical sliding mode observers1 can only be designed if the system
trajectories always
© evolve outside
ª the following manifold of observability sin-
gularity S = x ∈ R3 : x2 = 0 . In the vicinity of S, the observer must change
of structure. Thus, the following modified finite time sliding mode observer is

1 but also nonlinear observers based on other methodologies.

 T. Floquet and J.-P. Barbot 13

introduced in this paper:



 x̂˙ 1 = x̃2 + λ1 |ẽ1 |1/2 sign(ẽ1 )

 x̃˙ 2

 = α1 sign(ẽ
h 1) i

 ˙ 1/2
x̂2 = E1 −x1 + υ2 + λ2 |ẽ2 | sign(ẽ2 )
(12)

 υ̇2 = E1 α£ 2 sign(ẽ2 )



 x̂˙ 3

 = E2 (Es x̃3 + (1 − Es )x̂3 ) − (1 + x21 + x̂22 )(Es x̃33 + (1 − Es )x̂33 )
+KEs sign(x̃3 − x̂3 )]

where ẽi = x̃i − x̂i , with x̃1 = x1 (i = 1, 2, 3). λ1 , α1 , λ2 , α2 , and K are positive
constant and the functions Ei (i = 1, 2) and Es are given by:

Ei = 1 if |ẽj | = |x̃j − x̂j | ≤ ε for all j ≤ i, else Ei = 0

and

Es = 0 if |x̃2 | ≤ ε, else Es = 1.

x̃3 is defined as:

υ2
x̃3 = E1 Es + (1 − Es )x̂3 (13)
x̃2

Provided that the system trajectories evolve outside the manifold of singularity
S, the output of this observer provides a finite time estimation of x and m:
Step 1: From the section 2.2, there exists a time t1 > 0 such that, ∀t > t1 ,
the first exact differentiator gives x̂1 = x1 and x̃2 = x2 .
Step 2: For t > t1 , E1 = 1. Thus ∃t2 > t1 such that ∀t > t2 , the second exact
differentiator gives x̂2 = x̃2 = x2 and υ2 = x2 x3 . If |x2 | > ε, then Es = 1 and
the relation (13) gives x̃3 = x3 . Otherwise, Es = 0 and x̃3 = x̂3 .
Step 3: For t > t2 , E2 = 1. If Es = 1, the observation error dynamics of
e3 = x3 − x̂3 is equal to:

ė3 = m − Ksign(e3 ) (14)

and, for sufficiently large K, a first order sliding motion appears on e3 = 0.

The equivalent dynamics provides an estimation of the unknown input m.
Let us now assume that the system is on S. In that case, Es = 0 and

ė3 = m + e3 − (1 + x21 + x22 )(x33 − x̂33 ) (15)

£ ¤
= m + e3 1 − (1 + x21 + x22 )(x23 + x̂3 x3 + x̂23 )
 14 Step-by-step sliding mode observers for nonlinear systems with unknown inputs

Thus, the observer can be viewed as an “estimator”. The system (15) can
be shown to be stable for sufficiently large x̂3 and, consequently, e3 remains
always bounded.
Since the manifold of observability bifurcation S is not an invariant set
except for the trivial solution m = x1 = x2 = 0 (or very particular m), the
observer will converge in finite time to e3 = 0. Note that for m = x2 = x1 = 0,
both the state x3 and the observed state x̂3 asymptotically converge to 1.
This type of observability singularity have to be overcome in applications
such as bioprocess (Rapaport and Maloum 2000) or induction motor without
mechanical sensor (Canudas de Wit et al. 2000). On another hand, it can be
very useful for particular application such as secure data transmission (Barbot
et al. 2003). Indeed, the confidential message can be hidden in some observabil-
ity singularity and one has to design a variable structure observer to recover
it.

3.2 Switched system

A particular model of interest in hybrid systems are continuous-time systems
with discrete switching events from a certain class. Such systems are called
switched systems and there is an abundant literature on the observation prob-
lem (see e.g. Balluchi et al. 2002, Bemporad et al. 2000, De la Sen and Luo
2000, Vidal et al. 2003 and the references therein). Let us consider the following
nonlinear switched system without jump constituted, for a sake of simplicity,
of two nonlinear subsystems:
½
ẋ = f1 (x) and y = h1 (x) if σ(x) ≤ 0
(16)
ẋ = f2 (x) and y = h2 (x) if σ(x) > 0

where x ∈ Rn , y ∈ R is the output, fi (x) are smooth vector fields, and hi (x),
σ(x) are smooth functions. The continuous state space Rn is partitioned into
two operating regions by means of the switching surface σ(x) = 0. Let us
denote the switching times t1 , t2 ,... A difficulty arises from the fact that the
observer has to converge in a time less than the evolution duration ti+1 − ti
(before each switching of the system). Here, it is shown that the use of the
step-by-step sliding mode observer can be of high interest for the observation
problem of switched systems.
It is assumed that:
A.1 the system (16) is bounded state in finite time1 and without jump.

1 Theassumption of bounded state must concern the whole system. Indeed, all the subsystems can
be perfectly stable, while the global system can be unstable due to the switching phenomenon.
 T. Floquet and J.-P. Barbot 15

A.2 each subsystem is observable i.e. for i = 1 : 2,

n o
rank dhi , dLfi hi , ...., dLn−1
fi hi = n.

A.3 ti+1 − ti > δ for all i. δ is called the dwell time Liberzon (2003).
Assumption A.3 means that systems with Zeno phenomenon are not con-
sidered. Under those assumptions, one can conclude that the global system
(16) is observable if the dynamics of evolution of the system is known. Let us
consider the state transformation defined by:
j
zi+1 = φji+1 (x) = Lifj hj (x) for 0 ≤ i ≤ n − 1, j = 1 : 2.

In these local coordinate charts, each subsystem of (16) can be written in a

triangular observable form:
½
żi1 = zi+1
1 for i = 1 : n − 1
.1 (17)
z n = g1 (z1 , z2 , ..., zn1 )
1 1

if σ1 := (φ1 )−1 (z11 , z21 , ..., zn1 ) ≤ 0, and

( .2 2
z i = zi+1 for i = 1 : n − 1
.2 (18)
z n = g2 (z1 , z2 , ..., zn2 )
2 2

if σ2 := (φ2 )−1 (z12 , z22 , ..., zn2 ) > 0.

Remark 1 The assumption of a system without jump is necessary because the
diffeomorphism φ1 linked to the system i = 1 is not generally the same than
the diffeomorphism φ2 linked to the system i = 2. Then at each switch, the
state in triangular observable form may jump even if the state in the origi-
nal coordinates does not jump. Jumps can be taken into account by a new
structure of sliding mode observer that will be given in a forthcoming pa-
per. Nevertheless, in the case of switched Lagrangian systems, the considered
assumption is satisfied (Saadaoui et al. 2006). Note also that when the dif-
feomorphisms φi are the same (again, it is the case for switched Lagrangian
systems), the step-by-step sliding mode observer is a common observer to all
subsystems and does not require the knowledge of the switching function σ.
The finite time sliding mode observer (5) is of interest for the two main
following reasons:
(i) Under assumptions A.1 and A.2, it can provide an estimation z̃ij of the
state in a finite time Tf . Thus, if the observer gains can be chosen such
16 Step-by-step sliding mode observers for nonlinear systems with unknown inputs

that Tf < δ, one has the knowledge of the state between each switching of
the system.
A lot of observers for switched systems in the literature assume the
knowledge of the operating regions (i.e. the value of the switching function
σ). Here, under the assumption that gi (x) and ġi (x) are bounded, one
can get the continuous equivalent output injection (in the n-th step of the
convergence of the observer) θ̃ = gj (z1j , z2j , ..., znj ) that can be compared
to each value of gj (z̃1j , z̃2j , ..., z̃nj ), j = 1, 2 in order to know the operating
region where the system evolves (see Boutat et al. 2004 or Chaib et al. 2005
for more details on the geometrical conditions establishing this region).
(ii) In Djemaı̈ et al. (2005), a step-by-step first order sliding mode observer was
mainly employed for the same reasons: the finite time convergence and the
ability to take naturally into account the variable structure of the hy-
brid system. Nevertheless, some difficulties occurred due to the chattering
phenomena. It induced some irrelevant decision of switching between the
subsystems when the trajectory was in the neighborhood of the switching
manifold. This problem was bypassed by using a low pass filter during the
computation of the equivalent vector. However, this solution introduced a
delay in the observed state (asymptotic convergence) and again that could
lead to false information about the discrete state of the system. Thus, the
second order sliding mode observer can be a powerful and efficient tool
to solve this problem of observation because of its ability to provide a
continuous and accurate finite time estimation.
Further work may deal with the design of observers for switched or im-
pact systems with jumps and non autonomous systems which turns out to be
particularly useful in many applications.

3.3 Extension of the matching condition for nonlinear systems

In Floquet and Barbot (2006a), the authors presented an observation algo-
rithm in order to put a linear system with unknown inputs in a set of block
triangular observable forms, even if the matching condition was not fulfilled,
and a finite time sliding mode observer was designed in order to estimate the
state and the unknown inputs in Floquet and Barbot (2006b). The main idea
of this algorithm is to take advantage of some equivalent output injections in
order to generate fictitious outputs such that the system can be transformed
into a set of block observable triangular forms. The first block of the triangular
form is fed by the original inputs of the system while the subsequent blocks
are fed by the fictitious outputs. Here, it is shown via an example that this
idea can also be applied for some class of nonlinear systems that can not be
 T. Floquet and J.-P. Barbot 17

put in the form (2)-(3). Let us consider the following nonlinear system:

   
ẋ1 x2
 ẋ2   sin(x1 ) − x4 + m1 
 =  (19)
 ẋ3   x4 + m1 
ẋ4 cos(x3 ) + m2
µ ¶ µ ¶ µ ¶
y1 h1 x1
= = (20)
y2 h2 x3

where all the states are bounded ∀t ≥ 0 and where m1 , m2 are some bounded
unknown inputs. Figure 1 is a block diagram of (19)-(20).

Sin(x1)

m1
X2 X1 y1
1/p 1/p

x3 y2
1/p

X4
1/p
m2
Cos(x3)

Figure 1. System representation of (19)-(20)

For this system, one has ρ1 = 2 and ρ2 = 1 with respect to the outputs y1 and
y2 . Thus, according to Proposition 2.2, (19)-(20) can not be transformed into
a form similar to (2)-(3), and as a consequence usual sliding mode observers
fail to estimate the state in finite time. Furthermore, the matrix
à ! µ ¶
Lg1 Lρf1 −1 h1 (x) Lg2 Lρf1 −1 h1 (x) 10
Γ(x) = =
Lg1 Lρf2 −1 h2 (x) Lg2 Lρf2 −1 h2 (x) 10

is singular ∀x ∈ R4 and the method developed in Xiong and Saif (2001), that
could eventually lead to the asymptotic estimation of x, can neither be applied.
Let us show that one can even though estimate both state and unknown in-
puts using the finite time sliding mode observer described in Section 2.2. Define
a fictitious output as y3 = y2 − ẏ1 = x3 − x2 . One has ρ1 = 2, ρ2 = 1 and
ρ3 = 2 with respect to the fictitious output ȳ = [y1 , y2 , y3 ]T , while the observ-
ability indices are {1, 1, 2}. The change of coordinates φ(x) = [ξ1 , ξ2 , ξ3 , ξ4 ]T =
[y1 , y2 , y3 , x4 ]T = [x1 , x3 , x3 − x2 , x4 ]T , defined for all x ∈ R4 , transforms the
18 Step-by-step sliding mode observers for nonlinear systems with unknown inputs

system into a set of block triangular observable forms with respect to ȳ:
˙   
ξ1 ξ2 − ξ3
 ξ˙2   ξ4 + m1 
 =  (21)
 ξ˙3   −sin(ξ1 ) + 2ξ4 
ξ˙4 cos(ξ2 ) + m2

Figure 2 is a representation scheme of (21).

x1 y1

1/2 1/p

sin(x1 )

x4 x3 y3
m2
1/p 1/p
2

cos(x 2 )

m1 x2 y2
2 1/p

Figure 2. System representation of (21)

Following the lines of Section 2.2, a state observer can be designed as:



 x̂˙ 1 = x̃2 + λ1 |ẽ1 |1/2 sign(ẽ1 )

 x̃˙ 2 = α1 sign(ẽ
h 1) i
˙ (22)

 ξˆ3 = E1 −sin(y1 ) + 2ξ˜4 + 2λ3 |ẽ3 |1/2 sign(ẽ3 )


 ˙˜
ξ4 = E1 α3 sign(ẽ3 )

where ẽ1 = y1 − x̂1 , ẽ3 = ξ3 − ξˆ3 and with:

E1 = 1 if |ẽ1 | = |y1 − x̂1 | ≤ ε, else E1 = 0.

It can be seen that (22) is made of two exact differentiators.

Step 1: ∃t1 > 0 such that ∀t > t1 the first exact output differentiator provides
x̂1 = x1 , x̃2 = x2 and E1 = 1. The previously introduced fictitious output
y3 = ξ3 = y2 − x̃2 is henceforth available.
Step 2: Similarly to the first step, ∃t2 > t1 such that ∀t > t2 , the second
 T. Floquet and J.-P. Barbot 19

exact differentiator gives ξˆ3 = ξ3 and ξ˜4 = ξ4 . Thus, the full original state
[x1 , x2 , x3 , x4 ]T = [y1 , y2 − ξ3 , y2 , ξ4 ]T has been obtained in finite time.
Now, let us assume that ṁ1 , ṁ2 are bounded. The following observer pro-
vides a finite time estimation of m1 and m2 :

 ˙

 ξˆ2 = Eobs [x̃4 + m̃1 + λm1 |ẽ2 |1/2 sign(ẽ2 )]

 m̃˙1 = Eobs αhm1 sign(ẽ2 ) i
˙ (23)


 ξˆ4 = Eobs cos(y2 ) + m̃2 + λm2 |ẽ4 |1/2 sign(ẽ4 )

 ˙
m̃2 = Eobs αm2 sign(ẽ4 )

where

Eobs = 0 if t < t2 ,
else Eobs = 1.

For t > t2 , the first part of the observer has converged, all the state is available
and Eobs = 1. Then, the output of the two exact differentiators in (23) gives
a finite time estimation of m̃1 = m1 and m̃2 = m2 .
The observer design can be represented by the scheme given in Figure 3.

Figure 3. Observer scheme

Remark 2 The finite time sliding mode observers designed in Levant (2005)
and in Bejarano et al. (2006) can be also used in this kind of problem.

4 Conclusion

In this paper, we considered the problem of finite time estimation of MIMO

nonlinear systems with unknown inputs. Structural conditions stating on the
possibility to design sliding mode observers for finite time estimation of both
20 Step-by-step sliding mode observers for nonlinear systems with unknown inputs

state and unknown inputs have been derived. Then, a finite time sliding mode
observer based on the iterative use of the super twisting algorithm has been
proposed. It was shown that this kind of observer could be a powerful tool
to solve quite difficult problems of observation, like the design of observers
for switched systems. A modified version of the sliding mode observer was
designed for systems with observability singularities. It should be stressed
that few papers are devoted to the problem of sliding mode observer design for
these two promising fields of investigation. Lastly, by the way of an illustrative
example, a finite time observer was designed for a nonlinear system that do
not satisfies the matching condition. Further works aim at generalizing this
idea.

REFERENCES

A. Balluchi, L. Benvenuti, M. D. Di Benedetto and A.L. Sangiovanni-

Vincentelly, “Design of Observer for Hybrid Systems”, in: Lecture notes in
computer Science 2289, C. J. Tomlin and M. R. Greenstreet, Eds, Springer
Verlag, pp. 77–89, 2002.
J.-P. Barbot, T. Boukhobza and M. Djemaı̈, “Sliding mode observer for trian-
gular input form”, IEEE Conference on Decision and Control, Kobe, Japan,
1996.
J.-P. Barbot, I. Belmouhoub and L. Boutat-Baddas, “Observability Normal
Forms”, in: New trends in Nonlinear dynamics and control, LNCIS 295, W.
Kang et al., Eds, Springer Verlag, pp. 1–24, 2003.
G. Bartolini, A. Pisano and E. Usai, “First and second derivative estimation
by sliding mode technique”, Journal of Signal Processing, 4, pp. 167–176,
2000.
A. Bemporad, G. Ferrari-Trecate and M. Morari, “Observability and control-
lability of piecewise affine and hybrid systems”, IEEE Transactions on Au-
tomatic Control, 45, pp. 1864–1876, 2000.
A. Bejarano, F. Poznyak and L. Fridman “Observer for linear time invari-
ant systems with unknown inputs based on the hierarchical super-twisting
concept”, International Workshop on Variable Structure Systems, Alghero,
Sardinia, 2006.
D. Boutat, A. Benali and J.-P. Barbot, “About the observability of piecewise
dynamical systems”, NOLCOS, Stuttgart, Germany, 2004.
K. Busawon and M. Saif, “A state observer for nonlinear systems”, IEEE
Transactions on Automatic Control, 44, pp. 2098–2103, 1999.
C. Canudas de Wit and J. J. E. Slotine, “Sliding observers in robot manipu-
lators”, Automatica, 27, pp. 859–864, 1991.
C. Canudas De Wit, A.Youssef, J.-P. Barbot, Ph. Martin and F. Malrait,
 T. Floquet and J.-P. Barbot 21

“Observability conditions of induction motors at low frequencies”, IEEE

Conference on Decision and Control, Sydney, Australia, 2000.
S. Chaib, D. Boutat, A. Benali and J.-P. Barbot, “Observability of the discrete
state for dynamical piecewise hybrid systems”, Nonlinear Analysis: Theory,
Methods & Applications, 63, pp 423–438, 2005.
M. Darouach, M. Zasadzinski and S. J. Xu, “Full-order observers for linear
systems with unknown inputs”, IEEE Transactions on Automatic Control,
39, pp. 606–609, 1994.
J. Davila, L. Fridman and A. Levant, “Second-order sliding-mode observer
for mechanical systems”, IEEE Transactions on Automatic Control, 50, pp.
1785–1790, 2005.
M. De la Sen and N. Luo, “Design of linear observers for a class of linear
hybrid systems”, Int. Journal of Systems Science, 31, pp. 1077–1090, 2000.
M. Djemaı̈, J.-P. Barbot, A. Glumineau and R. Boisliveau, “Nonlinear flux
sliding mode observer”, Proceedings of the IEEE CSCC99, IMACS, Athens,
Greece, 1999.
M. Djemaı̈, N. Manamani and J. P. Barbot, “Sliding mode observer for trian-
gular input hybrid system”, IFAC World congress, Praha, 2005.
B. Drazenovic, “The invariance conditions in variable structure systems”, Au-
tomatica, 5, pp. 287–295, 1969.
S. V. Drakunov, “Sliding-mode Observer Based on Equivalent Control
Method”, IEEE Conference on Decision and Control, Tucson, Arizona,
1992.
S. V. Drakunov and V. I. Utkin, “Sliding mode observers. Tutorial”, IEEE
Conference on Decision and Control, New-Orleans, LA , 1995.
C. Edwards and S. K. Spurgeon, “On the development of discontinuous ob-
servers”, International Journal of control, 59, pp. 1211–1229, 1994.
C. Edwards and S. K. Spurgeon, Sliding mode control: theory and applications,
Taylor and Francis, Eds, 1998.
C. Edwards and S. K. Spurgeon and R. J. Patton, “Sliding mode observers for
fault detection and isolation”, Automatica, 36, pp. 541–553, 2000.
T. Floquet, J.-P. Barbot, W. Perruquetti and M. Djemaı̈, “On the robust fault
detection via a sliding mode disturbance observer”, International Journal
of control, 77, pp. 622–629, 2004.
T. Floquet and J.-P. Barbot, “An observability form for linear systems with
unknown inputs”, International Journal of control, 79, pp. 132–139, 2006a.
T. Floquet and J.-P. Barbot, “A canonical form for the design of unknown
input sliding mode observers”, in: Advances Variable Structure and Sliding
Mode Control, Lecture Notes in Control and Information Sciences, Vol. 334,
C. Edwards, E. Fossas Colet, L. Fridman, Eds., Springer Edition, 2006b.
F. Floret-Pontet and F. Lamnabhi-Lagarrigue, “Parametric identification
methodology using sliding modes observer”, International Journal of Con-
22 Step-by-step sliding mode observers for nonlinear systems with unknown inputs

trol, 74, pp. 1743–1753, 2001.

R. Garrido, Y. B. Shtessel, L. M. Fridman, “Application of VS differentiators
to DC servomechanisms”, International Workshop on Variable Structure
Systems, Barcelona, Spain, 2004.
J-P. Gauthier and G. Bornard, “Observability for any u(t) of a class of bilinear
systems”, IEEE Transactions on Automatic Control, 26, No 6, pp. 922–926,
1981.
J. P. Gauthier, H. Hammouri and S. Othman, “A simple observer for nonlinear
systems with applications to bioreactors”, IEEE Transactions on Automatic
Control, 37, pp. 875–880, 1992.
K. B. Goh, S. K. Spurgeon, N. B. Jones, “Fault diagnostics using sliding mode
techniques”, Control Engineering Practice, 10, pp. 207–217, 2002.
Y. Guan and M. Saif, “A novel approach to the design of unknown input ob-
servers”, IEEE Transactions on Automatic Control, 36, pp. 728–740, 1991.
I. Haskara, Ü. Özgüner and V. I. Utkin, “On sliding mode observers via equiva-
lent control approach”, International Journal of control, 71, pp. 1051–1067,
1998.
I. Haskara and Ü. Özgüner, “Equivalent value filters in disturbance estima-
tion and state observation”, in: Lecture Notes in Control and Information
Science, Vol. 247, pp. 167–179, 1999.
R. Hermann and A. J. Krener, “Nonlinear controllability and observability”,
IEEE Transactions on Automatic Control, 22, pp. 728–740, 1977.
J. Hernandez and J-P. Barbot, “Sliding observer-based feedback control for
flexible joints manipulator”, Automatica, 32, pp. 1243–1254, 1996.
M. Hou and P. C. Müller, “Design of observers for linear systems with unknown
inputs”, IEEE Transactions on Automatic Control, 37, pp. 871–875, 1992.
M. Hou, K. Busawon, and M. Saif, “Observer design for nonlinear systems
via injective mapping”, IEEE Transactions on Automatic Control, 45, pp.
1350–1355, 2000.
A. Isidori, Nonlinear Control Systems, Communication and Control Engineer-
ing Series, Third edition, Springer-Verlag, 1995.
A. J. Krener and W. Respondek, “Nonlinear observers with linearizable error
dynamics”, SIAM J. Control Optim., 23, pp. 197-216, 1985.
A. Levant, “Sliding order and sliding accuracy in sliding mode control”, In-
ternational Journal of Control, 58, pp. 1247–1263, 1993.
A. Levant, “Robust exact differentiation via sliding mode technique”, Auto-
matica, 34, pp. 379–384, 1998.
A. Levant, “Higher-order sliding modes, differentiation and output-feedback
control”, International Journal of Control, 76, pp. 924–941, 2003.
A. Levant, “Homogeneity approach to high-order sliding mode design”, Auto-
matica, 41, pp. 823–830, 2005.
D. Liberzon, Switching in systems and control, Birkhäuser, 2003.
 T. Floquet and J.-P. Barbot 23

N. Manamanni, M. Djemaı̈, T. Boukhobza and N. M’Sirdi, “Nonlinear sliding

observer based control for a pneumatic robot leg”, Int. J. of Robotics and
Automation, 16, pp.100–112, 2001.
R. Martinez-Guerra, R. Aguilar-Lopez, A. Poznyak, “Robust sliding-mode ob-
servers for on-line reaction heat monitoring in continuous chemical reactors”,
ASME Journal of Dynamics Systems, Measurement & Control, 126, pp.
473–478, 2004.
Y. Orlov, L. Aguilar, and J. C. Cadiou, “Switched chattering control vs. back-
lash/friction phenomena in electrical servo-motors,”International Journal of
Control, 76, pp. 959–967, 2003.
W. Perruquetti and J.-P. Barbot, Sliding Mode Control in Engineering, Ed.
Marcel Dekker, 2002.
A. Pisano and E. Usai, “Output-feedback control of an underwater vehicle
prototype by higher-order sliding modes,” Automatica, 40, pp. 1525–1531,
2004.
A. S. Poznyak, “Stochastic output noise effects in sliding mode estimations”,
International Journal of Control, 76, pp. 986–999, 2003.
A. Rapaport and A. Maloum,“Embedding for exponential observers of nonlin-
ear systems”, IEEE Conference on Decision and Control, Sydney, Australia,
2000.
J. Reger, H. Sira-Ramirez and M. Fliess, “On non-asymptotic observation of
non-linear systems”, IEEE Conference on Decision and Control, Sevilla,
2005.
H. Saadaoui, N. Manamanni, M. Djemaı̈, J.P. Barbot and T. Floquet, “Exact
differentiation and sliding mode observers for switched Lagrangian systems”,
Nonlinear Analysis: Theory, Methods & Applications, 65, pp. 1050–1069,
2006.
Y. B. Shtessel, I. A. Shkolnikov, and M. D. J. Brown, “An asymptotic second-
order smooth sliding mode control,” Asian Journal Control, 5, pp. 498–504,
2003.
Y. B. Shtessel and A. S. Poznyak, “Parameter identification of linear time
varying systems via traditional and high order sliding modes”, International
Workshop on Variable Structure Systems, Barcelona, Spain, 2004.
J. J. Slotine, J. K. Hedrick, and E. A. Misawa, “On sliding observers for
nonlinear systems”, ASME J. Dyn. System Measurement Control, 109, pp.
245–252, 1987.
C. P. Tan and C. Edwards, “An LMI approach for designing sliding mode
observers”, International Journal of Control, 74, pp. 1559–1568, 2001.
C. P. Tan and C. Edwards, “Sliding mode observers for robust detection and
reconstruction of actuator and sensor faults”, Int. J. of robust and nonlinear
control, 13, pp. 443–463, 2003.
V. I. Utkin, Sliding Modes in Control and Optimization, Berlin, Germany,
24 Step-by-step sliding mode observers for nonlinear systems with unknown inputs

Springer-Verlag, 1992.
V. I. Utkin, J. Guldner and J. Shi, Sliding mode control in electromechanical
systems, Taylor and Francis, London, 1999.
R. Vidal, A. Chiuso, S. Soatto and S. S. Sastry, “Observability of linear hybrid
systems, in: Hybrid Systems: Computation and Control, Lecture Notes in
Computer Science, Springer Verlag, Vol. 2623, pp. 526–539, 2003.
B. L. Walcott and S. H. Żak, “State observation of nonlinear uncertain dynam-
ical systems”, Transactions on Automatic Control, 32, pp. 88–104, 1987.
Y. Xiong and M. Saif, “Sliding Mode Observer for Nonlinear Uncertain Sys-
tems”, IEEE Transactions on Automatic Control, 46, pp. 2012–2017, 2001.