Proceedings of the 20th World Congress

Proceedings
The of
of the
International
Proceedings 20th
20th World
Federation
the Worldof Congress
Automatic Control
Congress
Proceedings
The of the
International 20th World
Federation of Congress Control
Toulouse, France,Federation
The International of Automatic
July 9-14, 2017
Automatic Control
Available online at www.sciencedirect.com
The International Federation
Toulouse, of Automatic Control
Toulouse, France,
France, July
July 9-14,
9-14, 2017
2017
Toulouse, France, July 9-14, 2017
ScienceDirect
IFAC PapersOnLine 50-1 (2017) 6265–6270
Interval
Interval Estimation
Estimation for
for Linear
Linear Switched
Switched
Interval Estimation
Interval Estimation for
for Linear Switched
Linear Switched
System
System 
System
System
∗,∗∗ ∗,∗∗ ∗,∗∗
Djahid Rabehi ∗,∗∗ Denis Efimov ∗,∗∗ Jean-Pierre Richard ∗,∗∗
Djahid
Djahid Rabehi
Rabehi ∗,∗∗ Denis Efimov ∗,∗∗ Jean-Pierre Richard ∗,∗∗
Djahid
∗
Rabehi ∗,∗∗ Denis Denis EfimovEfimov ∗,∗∗ Jean-Pierre
Jean-Pierre Richard Richard ∗,∗∗
∗ Inria, Non-A team, Parc Scientifique de la Haute Borne, 40 av.
∗ Inria, Non-A team, Parc Scientifique de la Haute Borne, 40 av.
∗ Inria, Non-A team, Parc Scientifique de la Haute Borne,
∗∗ Inria, Non-A Halley,
team,
Halley,
59650 Villeneuve
ParcVilleneuve
59650 Scientifique d’Ascq,
de la Haute
d’Ascq,
France.
France. Borne, 40 40 av. av.
∗∗ CRIStAL (UMR-CNRS
Halley, 59650
Halley, 59650 Villeneuve
9189), Ecole
Villeneuve d’Ascq,
Centrale
d’Ascq, France. France.
de Lille, BP 48, Cité
∗∗ CRIStAL (UMR-CNRS 9189), Ecole Centrale de Lille, BP 48, Cité
∗∗ CRIStAL Scientifique,
(UMR-CNRS 9189), Ecole
CRIStAL Scientifique,
(UMR-CNRS 59651
9189),
59651 Ecole Centrale
Centrale de
Villeneuve-d’Ascq,
Villeneuve-d’Ascq,
Lille,
Lille, BP
deFrance.
France. BP 48, 48, Cité
Cité
Scientifique,
Scientifique, 5965159651 Villeneuve-d’Ascq,
Villeneuve-d’Ascq, France. France.
Abstract: In this paper, the problem of state estimation is investigated for linear switched
Abstract:
Abstract: In
In this paper, the problem
systems. of It state estimation is investigated
the intervalfor linear switched
system,
Abstract:
system, a In this
a subclassthis paper,
subclass
of hybrid
paper,
of
the
the problem
hybrid problem
systems.
of
of
It
state
will be
state
will
estimation
shown that
estimation
be shown
is
is investigated
that investigated
the interval
for linear
linear switched
forobserver
observer
is very
switched
is very
system,
often
system, a
exists subclass
under
a subclass of
moderatehybrid systems.
conditions
of hybridconditions
systems. at It
at will
least
It least be
in
will in shown
discrete
bediscrete
shown time that
time the
instants
that instants interval
the interval from observer
continuous-time
observer is
is very
very
often
often exists
exists
measurements. under
under moderate
moderate
Themoderate conditions
novelty consists at least
in proposing in discrete time instants from
from continuous-time
continuous-time
often exists under
measurements. The novelty conditions
consists in at least innew
proposing discrete
new
conditions
conditions time instantsof cooperativity
of cooperativity
for switched
from continuous-time
for switched
measurements.
systems
measurements.in discrete The
Thetimenovelty
novelty consists
instants,
consists in
in proposing
which guaranteenew
proposing errors
new conditions
nonnegativity
conditions of
of cooperativity
of interval for
cooperativity switched
observation.
for switched
systems
systems
The in
in
efficiencydiscrete
discrete
of the time
time instants,
instants,
interval which
which
observers isguarantee
guarantee
shown errors
errors
through nonnegativity
nonnegativity
simulation of
of
examples. interval
interval observation.
observation.
systems
The in
efficiencydiscrete
of time
the intervalinstants, which
observers isguarantee
shown through errors nonnegativity
simulation of
examples. interval observation.
The
The efficiency
efficiency of
of the
the interval
interval observers
observers is
is shown through
shown Control)
throughHosting simulation
simulation examples.
examples.
© 2017, IFAC (International Federation of Automatic by Elsevier Ltd. All rights reserved.
Keywords: Nonnegative dynamics, Interval observer, Switched system.
Keywords: Nonnegative
Keywords: Nonnegative dynamics, Interval Interval observer, Switched Switched system.
Keywords: Nonnegative dynamics, dynamics, Interval observer, observer, Switched system. system.
1. INTRODUCTION new condition of cooperativity is proposed. The stability
1. INTRODUCTION
1. INTRODUCTION new
new thecondition
interval of
condition of cooperativity
cooperativity is
is proposed. The stability
1. INTRODUCTION of
new
of the condition
interval
observer
of cooperativity
observer
is guaranteed
is is proposed.
guaranteed proposed. by using
by using
The stability
TheLyapunov
stability
Lyapunov
Switched systems are very flexible modeling tools, which stability of
of the
the interval
analysis
interval observer
based
observer is
on
is guaranteed
Common
guaranteed by using
Lyapunov
by using Lyapunov
Function
Lyapunov
Switched systems are very
very flexible modeling tools, stability
which stability
(CLF) that analysis
analysis
decreases basedindependently
based on Common
on Common Lyapunov Lyapunov
switchingFunction Function
Switched
appear
Switched
appear
in systems
in
several fields
systems
several
are such flexible
are very
fields such flexible
as
modeling
as networked
modeling
networked
tools,
control
tools,
control
which
systems,
which stability
systems, (CLF) that analysis
decreases on Common of
basedindependently Lyapunov
of switching
(Liber-
Function
(Liber-
appear
electricalin several fields
devices/circuits, such as networked
congestion control
modeling systems,
(Liberzon, (CLF)
zon
(CLF) et that
al.,
that decreases
2004).
decreases independently
independently of
of switching
switching (Liber-
(Liber-
appear
electricalin devices/circuits,
several fields suchcongestionas networked control
modeling systems, zon
(Liberzon, zon et
et al.,
al., 2004).
2004).
electrical
2003;
electrical devices/circuits,
Briat, 2015).
devices/circuits,This congestion
class
congestionof modeling
systems
modeling is (Liberzon,
viewed
(Liberzon, as zon et al., 2004).
2003;
2003;
an Briat, 2015).
Briat,
abstraction 2015). Thissystems,
This
of hybrid class of
class of systems
systems by is viewed
is viewed as
as
2003;
an Briat, 2015).
abstraction of hybrid class of obtained
Thissystems, systems by
obtained
neglecting
is viewed
neglecting as 2. PRELIMINARIES
2. PRELIMINARIES
an abstraction of hybrid systems, obtained by neglecting 2.
the
an
the
discrete
abstraction
discrete
dynamics
of hybrid
dynamics
(Lin
(Lin
and
systems,
and
Antsaklis, 2009;
obtained2009;
Antsaklis,
Liberzon
by neglecting
Liberzon 2. PRELIMINARIES
PRELIMINARIES
the
et
the discrete
al., dynamics
2004).dynamics
discrete (Lin
State estimation and Antsaklis,
of switched
(Lin and Antsaklis, 2009; Liberzon
systems
2009; Liberzonhas Euclidean norm for a vector x ∈ Rnn will be denoted as
et al., 2004). State estimation of switched systems has
the Euclidean norm for a vector ∈ Rnn essentially
et
et al.,
al., 2004).
received
received 2004). State
State estimation
considerable
considerable
attention over
estimation
attention
of
of switched
over switched systems
past decades.
systems
past decades.
decades.
In has
has Euclidean
|x|e , and for
the Euclidean
In and
norm
norm for
for aa vector
a measurable vector andx x ∈ R
∈
xlocally
will
will
will be
R essentially
be denoted
be denoted bounded
denoted
as
as
as
received
case of considerable
switched system attention
with over
state past
jumps, necessary In the |x|
|x|
inpute ,
, and
andu : for
R
for a
a →measurable
R
measurable(R = and
and
{τ ∈locally
R
locally : τessentially
≥ 0}) the bounded
bounded
symbol
received
case of considerable
switched system attention
with over
state past
jumps, decades.
necessary In the
and |x|
input
e
e , andu : for
R + a →measurable
R (R +
= and
{τ ∈locally
R : τessentially
≥ 0}) the bounded
symbol
case of
of switched system for with state
state jumps, necessary
been and R + → R
∈R
sufficient conditions the observability jumps, have es- input {τ
case
sufficient
sufficient
tablished
switched
conditions
conditions
based
system
on for with
for the observability
the
graph-theoretic observability
approach
necessary
have
have been and
been
(Boukhobza es-
||u||[t0 ,tu
es- input
||u|| u1 ]:: denotes
R+
[t0 ,t1 ] denotes+ → R
its(R
its (RL+
L ∞=
+ =norm:
∞ norm:
+ {τ ∈ R :: ττ ≥ ≥ 0})0}) the the symbol
symbol
sufficient conditions for the observability have ||u||
been es- ||u||[t00 ,t11 ] ||u||
[t ,t ] denotes
denotes its
its L
L ∞ norm:
norm:
tablished
tablished based
based on
on graph-theoretic approach (Boukhobza = ess∞ sup{|u(t)| , t ∈ [t , t ]},
on graph-theoretic
and geometricapproach (Boukhobza
and Hamelin, [t0 ,t1 ] e 0 1
tablished based2011), graph-theoretic approach
approach (Tanwani
(Boukhobza ||u||
||u|| [t0 ,t1 ] =
] =
ess
ess sup{|u(t)|
sup{|u(t)| e ,, t
e, t
∈
∈ [t
[t 0 ,, t
0, t
1 ]},
andal.,Hamelin,
and
et
and
Hamelin,
2013).
Hamelin,
2011),
2011),
In (Barbot
2011),
and etgeometric
and
and
geometric
al.,
geometric 2007; approach
approach
Rı́os
approach et (Tanwani
(Tanwani
al., 2015)
(Tanwani if t 1 = +∞ ||u|| [t0 ,t1we
then
[t ,t ] = ess
will sup{|u(t)|
simply write e t ∈ [t
||u||. 0By t11 ]},
]},
L∞ we will
et al.,
al., 2013).
2013). In (Barbot
(Barbot etbeen
al., 2007;
2007; Rı́os et al., 2015)
2015) if tt1 = +∞ then 0 1
we will simply write ||u||.
||u||. By L||u||
∞ we <will
et
sliding mode
et al., 2013).
sliding mode
In
observers
In (Barbot
observers
have
have
et al.,
etbeen designed
al., 2007;
designed
Rı́os
to
Rı́os
to
et
et al.,
estimate
al., 2015)
estimate
the
the
if
denote
1 =
if t1 = +∞
denote
+∞
the
the
then
set
then
set
of
of
we
all
we
all
will
inputs
will
inputs
simply
simplyu
u
with
with
write
write the
the ||u||. By L
By
property
property
∞ we
L||u||
∞ we <
will
∞.
will
∞.
sliding
continuousmode andobservers
discrete have
states beenfor designed
the to
switched estimate
system the
in denote
By
denote 1, k the
we
the set
denote
set of
of all
all inputs
the
inputs sequenceu
u with
with of the
the property
integers
property 1, ||u||
...,
||u||k. <
< ∞.
Any
∞.
sliding
continuousmodeand observers
discrete have
states beenfor designed
the to estimate
switched system the
in By 1, k we denote the sequence of integers 1, ..., k. Any
continuous
observability
continuous and
and discrete
canonical states
form. for the switched
states for the switched system in p
discreteform. system in By × 1, 1, k we
mk matrix denote whose the sequence
elements of
are integers
all ones or 1, ..., k.
zeros Any
are
observability
observability canonical
canonical form. By
p × m we
matrix denote whose the sequence
elements of
are integers
all ones 1,
or ..., k.
zeros Any
are
p× ×m m denoted
matrix whose whose elements are all all ones
onesIp or or zeros the are
observability
The estimation canonical
problemform. becomes much more involved if simply p
simply matrix
denoted
by Ep,m
by E p,m. or
or 0, respectively.
elements 0, are
respectively.
denotes
IIp denoteszeros are
the
The
The estimation
estimation problem
problem becomes
whichbecomes much
much more involved if simply
identity denoted
matrix by
in R
E p×p
p,m or 0, respectively. p denotes the
we
The consider
estimation
we consider
consider
systems
systemsproblemwhich becomes much more
are subjected
are subjected
subjected
to model
more involved
and/or
involved
toestimation
model and/or
if
if simply
identity
identity
Throughout
denoted
matrix
matrix
by REp×p
thisin
in R p,m. or 0, respectively. Ip denotes the
p×p
p×p . the inequalities must be under-
paper
we
signals
we systems
uncertainties.
consider systems which which are
Therefore, the
are subjected stateto model
toestimation and/or
model and/or ap- identity
Throughout matrix this in R
paper . the inequalities must be under-
signals
signals
proaches uncertainties.
uncertainties.
based on theTherefore,
Therefore,
set-membership the state
the stateandestimation
interval ap- Throughout
ap-
ob- stood
Throughout component-wise, this
this paper
paper the
for
the inequalities
matrices
inequalities must
as well
mustas be
befor under-
vec-
under-
signals uncertainties.
proaches based on theTherefore,
set-membership the stateand estimation
interval ap- tors,
ob- stood
stood component-wise,
component-wise,
i.e. A = (a ) ∈ for
R
for p×mmatrices
matrices
and B as
as = well
well
(b as
as
) ∈ for
forR vec-
p×m
vec-
proaches based on the set-membership and interval ob- stood component-wise, i,j for matrices as well i,j as for vec-
servers
proaches
servers
get
get
more
based
more onattention
attention
for
for
uncertain
the set-membership
uncertain
systems
and
systems
(Efimov
interval
(Efimov ob- tors,tors,
such i.e.
i.e.
that A
A A =
= ≥ (a
(a i,j
B )
) ∈
∈
if R
R
and
p×m
p×m
onlyand
and if, B
B a =
= (b
(b≥ i,j b)
) ∈
∈ R
R
for
p×m
p×m
all
servers
and
servers get
Raı̈ssi, more
2016).
get more attention
In the
attention for uncertain
literature,
for uncertain the systems
interval (Efimov
observers
(Efimov such
systemsobservers tors, i.e.
that A A = ≥ (a i,j
B
i,j ) ∈
if R
and
p×m
onlyand if, B a i,j
= (b≥
i,j
i,j )i,j
bbi,j∈ R
for
p×m
all
and
and Raı̈ssi,
Raı̈ssi, 2016).
2016). In
In the
the literature, the interval such
i ∈ that
{1, . . . A
, p},
Ap},≥j ∈≥j ∈B if
{1, and
. . . , only
m}. M if, = a i,j
max{A,
i,j ≥
≥ bi,j B}
i,j for
is all
the
are
and applied
Raı̈ssi, to estimate
2016). In theliterature,
the literature, the
state for several
the interval
classes
interval observers
of non- such
observers i ∈ that
{1, . . . , B {1, if .and. . , onlyMif,=ai,j
m}. max{A, B} for
is all
the
are
are applied
applied to
to estimate
estimate the
the state
state for
for several
several classes
classes of
of non-
non- i
i ∈
matrix
∈ {1,
{1, .
. .
where
. .
. ,
, p},
p}, j
each
j ∈
∈ {1,
entry
{1, .
. .
. .
is
. ,
, m}.
m
m}.i,j M
=
M max{a=
= max{A,
i,j
max{A, , b B}
i,j
B}}. is
Let
is the
us
the
linear
are systems
applied (Raissi
to estimate et al.,
theal., 2012;
state Efimov
for Efimov et
several classesal., 2013c),
of non- matrix matrix where
+ each entry is− m += max{a , b }.
i,j }. Let us
linear
linear systems
systems (Raissi
(Raissi et
et al., 2012;
2012; Efimov et
et al.,
al., 2013c),
2013c), define
matrix A where
where = each
max{A,
each entry
0},
entry A is
is−
m
m= i,j
A
i,j =
= − max{a
max{aA; i,j , the
i,j , b
thus, b i,j }. Let
element-
Let us
us
time-delay
linear systems systems
(Raissi (Efimov
et al., etet al.,
2012; 2013b)
Efimov and
et sampled-
al.,sampled-
2013c), wise define A +
+ = max{A, 0}, A − = A
i,j +
+ − A; thus,
i,j thei,j +element- −
time-delay systems (Efimov al., 2013b) and define A + = max{A,
absolute value 0},
will Abe − = A + − A;
denoted as thus,
|A| =the A element-
+ A −.
time-delay
data systems systems
(Mazenc (Efimov
et al., et al.,
2014; 2013b)
Efimov and
et sampled-
al., 2013a). define
wise A
absolute = max{A, 0}, A = A − A; thus,
|A| the +element-
time-delay
data systems
data systems systems
(Mazenc
(Mazenc
(Efimov
et al.,
et al., et al., Efimov
2014;
for2014;
2013b) et
Efimov
and
et al.,sampled-
al., a class A
2013a).
2013a). wise
wise absolute
matrix
absolute P ∈value Rn×n
value
value
will
will
willis
be
be
be saiddenoted
to be as
denoted
denoted as |A|
|A| =
negative
as
= A
= A
+ + A− .
+ + A−if.
definite
A
Several interval observers impulsive systems, R n + A if .
n×n
data
Severalsystems
interval(Mazenc
observerset al.,
for 2014; Efimov
impulsive et al.,
systems, 2013a).
a class A
A
υ T matrix
matrix
P υ < 0 P
P for∈
∈ R
all n×n
non-zerois
is said
said realto
to be
be
vectors negative
negative
υ ∈ definite
R
definite
and it if
Several
of linear interval
hybrid observers
systems, for
have impulsive
been systems,
recently a
applied class
in A
υ T matrix
P υ < 0 P for∈ R
all
n×n
non-zerois said realto be
vectors negative
υ ∈ definite
R n
and if
it
Several
of linear
of linearetinterval
hybrid observers
systems, have for impulsive
have been recently systems,
recently a class
applied in will T
υT P
υ P beυ
υ < 00 for
denoted for all by P non-zero
≺ 0. Similarly,real
real vectors P  0(P υ ∈
υ ∈ R
R0)
n
n and
means it
of linearethybrid
(Degue
(Degue
al., 2016).
hybrid
al.,
systems,
systems, have been
2016). been recently applied applied in in willwill be
be
semi-negative
<denoted
denoted
all
by
by P
P
(positive)
non-zero
≺
≺ 0.
0. Similarly,
Similarly,
definite
vectors
matrix. P
P 
 0(P
0(P 
 0)
0)
and
means
means
it
(Degue
(Degue et al., 2016).
2016). the switching among subsystems can semi-negative
et al.,systems will be denoted by P ≺
(positive) definite0. Similarly,
matrix. P  0(P  0) means
For switched semi-negative
semi-negative (positive) (positive) definite definite matrix.
matrix.
For switched
For
affectswitched
the systems the
systems
cooperativity theproperty
switching
switching of among
among
interval subsystems
subsystems
estimation. can
canIn
For
affectswitched
the systems theproperty
cooperativity switching of among
interval subsystems
estimation. canIn 2.1 Cooperativity
affect the 2.1 Cooperativity
this
affect the cooperativity
work,
thisswitched
work,
we deal with the
cooperativity
wesystem
deal with
with
property
synthesis
property
the synthesis
of
of interval
interval estimation.
of interval estimators
estimation.
of interval
interval estimators
In
In 2.1 2.1 Cooperativity
Cooperativity
this
for work, we deal underthe synthesis
bounded of
unknown estimators
input, and a 1. (Minc, 1988) A matrix M = (mi,j ) ∈ Rn×n
this
for
work, wesystem
for switched
switched deal with
system under
under
thebounded
synthesisunknown
bounded
of interval
unknown input,
input, and aa Definition
estimators
and Definition 1. (Minc, 1988) A matrix M = (m i,j ) ∈ R n×n
forThis
 switched
work system
was under
partially bounded
supported by theunknown
Governmentinput,of and a is
Russian Definition
said
Definition to 1.
be
1. (Minc,
Metzler
(Minc, 1988)
if
1988) all A
A matrix
its off-diagonal
matrix M
M =
= (m
(m i,j )
elements
) ∈
∈ R
R
n×n
are
n×n
 This is
is said
said to
to be
be Metzler
Metzler
mi,j ≥if if all
0,all its
its off-diagonal
off-diagonal
i = j. Andelements elements
it is saidare
i,j
are

 This work
Federation was
was partially
work(Grant 074-U01)
partially supported
and theby
supported the
the Government
byMinistry of
of Russian
of Education
Government and
Russian nonnegative
is said to be i.e.
Metzler if all ∀(i,
itsj),off-diagonal elements to
are
This work(Grant
Federation
Science
was partially
of Russian
Federation 074-U01)
(Grant Federation
supported
074-U01) and the
(Project
and
byMinistry
the Government
Ministry of
the 14.Z50.31.0031).
of Russian
of Education
Education and and
nonnegative
nonnegative
be Nonnegative i.e.
i.e. m
m
if i,j ≥
every ≥ 0,
0, ∀(i,
∀(i,
entries j),
j),
are =
ii nonnegative:

= j.
j. And
And it
it is
is
M said
said
≥ to
to
0.
i,j
Federation
Science (Grant 074-U01) and the Ministry of Education and nonnegative
be Nonnegative i,j ≥ 0,
i.e. ifmevery ∀(i, j),
entries are = j. And it isMsaid
i nonnegative: ≥ to
0.
Science of
of Russian
Russian Federation
Federation (Project
(Project 14.Z50.31.0031).
14.Z50.31.0031). be Nonnegative if every entries are nonnegative: M ≥ 0.
Science of Russian Federation (Project 14.Z50.31.0031). be Nonnegative if every entries are nonnegative: M ≥ 0.
Copyright
2405-8963 ©© 2017,
2017 IFAC 6454Hosting by Elsevier Ltd. All rights reserved.
IFAC (International Federation of Automatic Control)
Copyright
Copyright © 2017
©under IFAC
2017 responsibility
IFAC 6454
6454Control.
Peer review
Copyright © 2017 IFAC of International Federation of Automatic
6454
10.1016/j.ifacol.2017.08.855
Proceedings of the 20th IFAC World Congress
6266
Toulouse, France, July 9-14, 2017 Djahid Rabehi et al. / IFAC PapersOnLine 50-1 (2017) 6265–6270
Definition 2. (Farina and Rinaldi, 2000) A continuous-
time linear system ẋ(t) = Ax(t) (discrete-time linear
system x(t + 1) = Ax(t)), with the state x ∈ Rn and
A ∈ Rn×n , is said to be cooperative if A is a Metzler
(Nonnegative) matrix.
The solutions of cooperative systems, initiated from
x(0) ≥ 0, stay nonnegative: x(t) ≥ 0 for all t ≥ 0.
2.2 Intervals
Lemma 1. (Efimov et al., 2012) Let A ∈ Rp×m be a
constant matrix and the vector x ∈ Rm be variable with
some bounds x, x ∈ Rm such that x ≤ x ≤ x, then
A+ x − A− x ≤ Ax ≤ A+ x − A− x.
2.3 Stability
A continuous-time switched linear time-invariant (LTI)
system can be represented as
ẋ(t) = Aσ(t) x(t), σ(t) ∈ I = {1, . . . , N }
where x ∈ Rn is the state, the finite set I is an index
set and it stands for the collection of discrete modes
Ai ∈ Rn×n for all i ∈ I. The logical rule that arranges the
switching between the subsystems generates a switching
signal, a piecewise constant function σ(t) : R+ → I, the
index i = σ(t) specifies, at each instant of time, the system
that currently being followed.
By saying that the switching signal is piecewise constant
we mean that it has a finite number of discontinuities on
any finite interval of time. We assume that σ(t) is contin-
uous from the right everywhere: σ(t) = limα→t+ σ(α).
We will say that the switching signals satisfy the so-called
dwell-time property if tk+1 − tk ≥ τD for some dwell time
constant τD > 0, where tk is the instant of k th switch.
Lemma 2. (Liberzon et al., 2004) Let P ∈ Rn×n be a
symmetric positive definite matrix that satisfies the Linear
Matrix Inequalities (LMIs)
ATi P + P Ai ≺ 0, i ∈ I = {1, . . . , N }
T
Then V (x) = x P x is a Common Quadratic Lyapunov
Function (CQLF) for the systems (1).
This lemma establishes conditions of the internal stability
(without taking into account the effect of external inputs).
For switched systems with inputs and dwell-time switching
signals, the overall system is input-to-state stable (ISS) if
the individual subsystems are ISS (Vu et al., 2007). For
linear switched systems with additive inputs, which we
will consider below:
ẋ(t) = Aσ(t) x(t) + Bσ(t) u(t) + d(t), (3)
where u(t) ∈ Rm , u ∈ L∞ is a known input and Bi ∈
Rn×m for all i ∈ I, d(t) ∈ Rn , d ∈ L∞ is external
disturbance, the conditions of Lemma 2 imply ISS.
3. MAIN RESULTS
Consider the switched linear system (3) with the output
expression given by
y(t) = Cσ(t) x(t) + ν(t) (4)
where y(t) ∈ Rp is the output of the system, and ν(t) is
the output measurement noise.
The models of subsystems can be represented as
ẋ(t) = Ai x(t) + Bi u(t) + d(t) i ∈ I
= (Ai − Li Ci )x(t) + d(t) + Bi u(t) + Li y(t) − Li ν(t)
(5)
and Li ∈ Rn×p is an observer gain that will be defined
later.
The goal here is to compute two bounds for the state of
the considered system. To do so, several approaches are
considered to satisfy the nonnegativity of the estimation
error dynamics. A new cooperativity condition based on
Trotter approximation is proposed.
3.1 Simple Interval Observer
In this section we will consider a structure of interval
observer for (5) based on the classical Luenberger observer
with proper assumptions that the estimation error dynam-
ics are cooperative.
(1)
Assumption 1. There exist matrices gains Li and a sym-
metric positive-definite matrix P such that ∀i ∈ I:
(1) The LMIs
(Ai − Li Ci )T P + P (Ai − Li Ci ) ≺ 0 (6)
(2) The matrices (Ai − Li Ci ) are Metzler.
If the Assumption 1.1 is satisfied, then V (x(t)) = xT P x is
a CQLF for the system (5).
Assumption 2. Let two functions d, d : R+ → Rn , d, d ∈
Ln∞ and the scalar constant V ≥ 0 be given such that
d(t) ≤ d(t) ≤ d(t), |ν(t)| ≤ VEp
are verified ∀t ∈ R+ .
Under assumptions 1 and 2 the interval observer can be
designed as
ẋ(t) = (Ai − Li Ci )x(t) + d(t) + Li y(t) + Bi u(t) − |Li |VEp
(2) ẋ(t) = (Ai − Li Ci )x(t) + d(t) + Li y(t) + Bi u(t) + |Li |VEp
(7)
where |Li | = L+ −
i + Li .
Theorem 1. Let Assumptions 1, 2 hold and x ∈ Ln∞ , then
in (5) with the interval observer (7) the relations
x(t) ≤ x(t) ≤ x(t)
are satisfied provided that x(0) ≤ x(0) ≤ x(0). In addition,
x, x ∈ Ln∞ .
All proofs are omitted due to space limitations.
3.2 Relaxation of Cooperativity
The conditions of the cooperativity of the observer given in
Assumption 1.2 are rather restrictive, and further we will
consider several ways to relax them. The main approach
deals with transformation of coordinates in order to ob-
tain the estimation error in a cooperative form. A time-
invariant interval observer can be designed by applying a
change of coordinates proposed in (Raissi et al., 2012).
In this context the interval observer will be constructed
6455
Proceedings of the 20th IFAC World Congress

Toulouse, France, July 9-14, 2017 Djahid Rabehi et al. / IFAC PapersOnLine 50-1 (2017) 6265–6270 6267

in the new state variables z = Si−1 x, where Si is a non- Theorem 2. Based on Assumptions 2 and 3, if the system
singular matrix, selected such that the matrices Di = (5) with some given x(0), x(0) ∈ Rn fulfills x(0) ≤ x(0) ≤
Si−1 (Ai − Li Ci )Si are Metzler. x(0), then the established interval observer (10) satisfies
Applying the transformation on the system (5) we obtain the inclusion
ż(t) = Di z(t) + Si−1 Bi u(t) + Si−1 Li y(t) + δi (t) (8) x(t) ≤ x(t) ≤ x(t) ∀t ≥ 0
In addition, x, x ∈ Ln∞ .
where δi (t) = Si−1 (d(t)
− Li ν(t)).
However, it is worth noting that when we adopt the change 3.3 An alternative of interval observer without coordinates
of coordinates x = Si z, Si ∈ Rn×n , jumps in the new change
state may arise. This is due to the fact that, generally, the
transformation matrices Si assigned to subsystems are not An alternative solution to avoid the requirement that the
the same. Therefore, after switching, the new state is matrices (Ai − Li Ci ) for i ∈ I are Metzler has been
z(tk ) = ϕi z(t−
k ), ∀i ∈ I,
proposed in (Cacace et al., 2015) (for linear systems), it
−1 is based on the fact that the system may be represented
where matrix ϕi = Si+1 Si is not identity matrix. in its internally positive representation. However, this ap-
Consequently, instead of synthesis an interval observer for proach has more restrictive condition on stability. For the
continuous state, the interval observer will be synthesised, switched system (5) the state matrices can be presented
in this new variables, for a switched system with state as follows (Ai − Li Ci ) = (Ai − Li Ci )d + (Ai − Li Ci )+ o −
jumps, i.e. for a hybrid system. Thus, based on the expres- (Ai − Li Ci )− with i ∈ I where (·) and (·) are the
d o
sion of the output y(tk+1 ) = Ci+1 Si z(t− k+1 ) + ν(tk+1 ), at
o
matrices that contain only the main diagonal elements and
the switching instant, the reset equation can be rewritten
as follows the off-diagonal elements, respectively. Thus the matrices
  (Ai − Li Ci )d + (Ai − Li Ci )+o ] for i ∈ I are Metzler. In the
z(tk+1 ) = ϕi − Mi+1 Ci+1 Si z(t− k+1 )
(9) sequel, we shall denote for short Ãi = (Ai − Li Ci ).
+ Mi+1 y(tk+1 ) − Mi+1 ν(tk+1 ) Assumption 4. There exists a symmetric positive-definite
where Mi ∈ Rn×p is another observer gain defined later. matrix PΥ that satisfies the LMIs
From (8), the interval observer is then given by ΥTi PΥ + PΥ Υi ≺ 0 ∀i ∈ I (12)
x(t) =Si+ z(t) − Si− z(t), where  
Ãid + Ã+ Ã−
x(t) =Si+ z(t) − Si− z(t), Υi = io io .
Ã−
io Ãid + Ã+ io
ż(t) =Di z(t) + Si−1 (Bi u(t) + Li y(t))
Theorem 3. Let Assumptions 2, 4 be satisfied, x ∈ Ln∞
+ δ (t), ∀t ∈ [tk , tk+1 ), ∀k ∈ N, in (5) and for some x(0), x(0) ∈ Rn the inclusion x(0) ≤
 i 
z(tk+1 ) = ϕi − Mi+1 Ci+1 Si z i (t− k+1 ) x(0) ≤ x(0) is valid, then the interval observer
(10) −
+ Mi+1 y(tk+1 ) − |Mi+1 |VEp ẋ(t) =(Ãid + Ã+
io )x(t) + Ãio x(t) + d(t)
ż(t) =Di z(t) + Si−1 (Bi u(t) + Li y(t)) + Li y(t) + Bi u(t) − |Li |VEp
−
(13)
+ δ i (t), ∀t ∈ [tk , tk+1 ), ∀k ∈ N ẋ(t) =(Ãid + Ã+
io )x(t) + Ãio x(t) + d(t)
  + Li y(t) + Bi u(t) + |Li |VEp
z(tk+1 ) = ϕi − Mi+1 Ci+1 Si z i (t− k+1 )
+ Mi+1 y(tk+1 ) + |Mi+1 |VEp satisfies the order relations x(t) ≤ x(t) ≤ x(t) and
x(t), x(t) ∈ Ln∞ .
where z is the state variable of (10), which is the same
variable after and before jump, and 3.4 Trotter Formula Based Interval Observer
   
−1 + −1 −
δ i (t) = (Si ) d(t) − |Li |VEp − (Si ) d(t) + |Li |VEp ,
    In this section we consider a special case when the gain
δ i (t) = (Si−1 )+ d(t) + |Li |VEp − (Si−1 )− d(t) − |Li |VEp matrices of the observer do not exist in a way to satisfy
the cooperativity of the estimation error dynamics, neither
with the initial conditions by direct application nor by the means of coordinate
z(0) = (Si−1 )+ x(0) − (Si−1 )− x(0), transformation. In this case the following Trotter formula
z(0) = (Si−1 )+ x(0) − (Si−1 )− x(0). result can be used.
Assumption 3. There exist matrices Mi , Li ∈ R n×p
, i ∈ I Theorem 4. (Trotter, 1959) For two given matrices B, C ∈
and a positive-definite matrix Pϕ ∈ R n×n
such that Di are R n×n
, the exponentials of B and C are related to that of
 B + C as follows
Metzler and ϕi − Mi+1 Ci+1 Si are nonnegative matrices,  B C m
which satisfy the LMIs lim e m e m = e(B+C) . (14)
 T D T θ Di θ
  m→∞
ϕi −Mi+1 Ci+1 Si e i Pϕ e ϕi −Mi+1 Ci+1 Si −Pϕ ≺ 0
A
  (11) The Trotter result can be used to evaluate e by splitting
for all θ ∈ θmin , θmax and 0 ≤ θmin ≤ ti+1 − ti ≤ θmax ≤ A into B + C and then using the approximation
 B C m
+∞ for all i ≥ 0. eA  e m e m
Note that in order to formulate stability conditions in for a sufficiently big m ≥ 0.
the assumption above we need to introduce a kind of To apply this formula, we consider the system (5) without
minimal θmin and maximal θmax dwell-time restrictions of the input u(t) and perturbations d(t), ν(t)
the switching signal under consideration. ẋ(t) = (Ai − Li Ci )x(t) + Li y(t) (15)

6456
Proceedings of the 20th IFAC World Congress
6268
Toulouse, France, July 9-14, 2017 Djahid Rabehi et al. / IFAC PapersOnLine 50-1 (2017) 6265–6270

In addition in this subsection we will assume that I = Assumption 5. There exist two matrices L1 , L2 ∈ Rn×p
{1, 2} and that the switching is periodical and tk+1 −tk = τ such that the matrix A1 − L1 C1 + A2 − L2 C2 is Metzler.
for some period τ > 0. The state trajectories of the system
between two switching times can be calculated analytically It has been shown in (Minc, 1988; Mitkowski, 2008) that
from the equation (15) of the activated subsystem the matrix eAt ≥ 0 is nonnegative if and only if A ∈ Rn×n
is Metzler. Thereby, Assumption 5 allows us to conclude
x(tk+1 ) = eÃσ(tk ) (tk+1 −tk ) x(tk ) T
 tk+1 that the matrix e(Ã1 +Ã2 ) 2 is nonnegative for all T ≥ 0.
+ eÃσ(tk ) (tk+1 −s) Lσ(tk ) y(s)ds Assumption 6. There exist two matrices L1 , L2 ∈ Rn×p ,
tk a symmetric positive-definite matrix PΓ ∈ R2n×2n and a
where Ãσ(tk ) = Aσ(tk ) −Lσ(tk ) Cσ(tk ) and σ(tk ) corresponds scalar m ∈ N such that the nonnegative matrix
to the index of the activated system. After switching 2m  
T
times between subsystems represented by (15), the state e(Ã1 +Ã2 ) 2 + ∆− ∆+
variables will be obtained through the analytical solution Γ= T
∆+ e(Ã1 +Ã2 ) 2 + ∆−
2m−1

x(T ) = eÃσ(tk ) τ x(t0 ) satisfies the LMI
k=0 ΓT PΓ Γ − PΓ ≺ 0.
2m  2m−1
    Theorem 5. Let Assumptions 5 and 6 be satisfied and
+ eÃσ(tj ) τ (16)
x ∈ Ln∞ , then in the system (20) with the interval observer
k=1 j=k (22) the relations
 tk 
× eÃσ(tk−1 ) (tk −s) Lσ(tk−1 ) y(s)ds x(lT ) ≤ x(lT ) ≤ x(lT ), ∀l ∈ Z+
tk−1 are satisfied provided that x(0) ≤ x(0) ≤ x(0), and the
where T = 2mτ . To simplify, the equation (16) can be estimates x(lT ), x(lT ) are bounded.
represented as follows
2m−1 To generalize the synthesis of the interval observer to the

x(T ) = eÃσ(tk ) τ x(t0 ) + Λ(y(t)) (17) perturbed case the conditions imposed in Assumption 2.1
will be used. The system (15) with input disturbance can
k=0
be rewritten in the following form:
where
2m  2m−1
   ẋ(t) = (Ai − Li Ci )x(t) + d(t) + Li y(t) i ∈ I (23)

Λ(y(T )) = eÃσ(tj ) τ The state variables, after period of time T , can be calcu-
k=1 j=k lated with the same as previously analytical method
 tk  2m−1
 2m  2m−1
   Ã
× Ãσ(tk−1 ) (tk −s)
e Lσ(tk−1 ) y(s)ds 
tk−1
x(T ) = e Ãσ(tk ) τ
x(t0 ) + e σ(tj ) τ
k=0 k=1 j=k (24)
Using the fact that Ãσ(tk ) ∈ {Ã1 , Ã2 }. tk 
× eÃσ(tk−1 ) (tk −s) d(t)ds + Λ(y(T ))
tk−1
2m−1
  m
eÃσ(tk ) τ = eÃ1 τ eÃ2 τ Proposition 1. Under Assumption 2.1 the inequality
k=0 (18) D k ≤ Dk ≤ D k
 T T
m
= e Ã1 2m Ã2 2m
e is satisfied, where
 tk 
 Ãσ(t ) (tk −s) +
At this stage by using Theorem 4 we know Dk = e k−1 d(s)
 T T
m T tk−1
lim eÃ1 2m eÃ2 2m = e(Ã1 +Ã2 ) 2 (19)  − 
− eÃσ(tk−1 ) (tk −s) d(s) ds,
m→∞
By coupling (17) and (19), we find 
 T
 tk
x((l + 1)T ) = e(Ã1 +Ã2 ) 2 − ∆ x(lT ) + Λ(y((l + 1)T )) Dk = eÃσ(tk−1 ) (tk −s) d(s)ds, (25)
tk−1
(20)  tk  +
where   m  Dk = eÃσ(tk−1 ) (tk −s) d(s)
(Ã1 +Ã2 ) T2 T
Ã1 2m T
Ã2 2m tk−1
∆= e − e e (21)  − 
− eÃσ(tk−1 ) (tk −s) d(s) ds.
And due to Trotter result we know that the matrix ∆
converges to zero by increasing m. Then the interval
observer for (20) can be designed in the following form: And the new-estimated bounds can be expressed as follows
T

x((l + 1)T ) =e (Ã1 +Ã2 ) T2

x(lT ) − ∆+ x(lT ) + ∆− x(lT ) x((l + 1)T ) =e(Ã1 +Ã2 ) 2 x(lT ) + ∆− x(lT ) − ∆+ x(lT )
+ Λ(y((l + 1)T )) + Π2m + Λ(y((l + 1)T ))
T
T
x((l + 1)T ) =e(Ã1 +Ã2 ) 2 x(lT ) − ∆+ x(lT ) + ∆− x(lT ) x((l + 1)T ) =e(Ã1 +Ã2 ) 2 x(lT ) + ∆− x(lT ) − ∆+ x(lT )
+ Λ(y((l + 1)T )) + Π2m + Λ(y((l + 1)T ))
(22) (26)
provided that the next Assumption is valid: where Π2m , Π2m are calculated by

6457
Proceedings of the 20th IFAC World Congress

Toulouse, France, July 9-14, 2017 Djahid Rabehi et al. / IFAC PapersOnLine 50-1 (2017) 6265–6270 6269

 +
Ψk Πk−1 − Ψ− k Πk−1 + D k for k ≥ 2
Πk =
Dk for k = 1
 + −
Ψk Πk−1 − Ψk Πk−1 + Dk for k ≥ 2
Πk =
Dk for k = 1
 Ãσ(t ) τ 
where Ψk = e k−1 and the quantities Dk , Dk are
defined by expressions (25).
Theorem 6. Let Assumptions 5, 6 and 2.1 be satisfied and
x ∈ Ln∞ , then in the system (24) with the interval observer
(26) the relations
x(lT ) ≤ x(lT ) ≤ x(lT ), ∀l ∈ Z+ Fig. 2. Simulation results of the second approach: state
are satisfied provided that x(0) ≤ x(0) ≤ x(0), in addition variables x.
the estimates x(lT ) and x(lT ) are bounded. 4.2 Example for the third approach

4. ILLUSTRATIVE EXAMPLES Consider system (5) with

       
For the sake of simplicity, −10 −4 5 −8 4 10
 we  consider switched systems A1 =
4 −3
, B1 =
0
, A2 =
−4 −7
, B2 =
0
,
with 2 modes, i.e. I = 1, 2 .
C1 = C2 = [0 1]
4.1 Example for the change of coordinates based interval and the model and output perturbations
observer  T
d(t) = 0.5 sin(5t), 0.5 sin(5t) , v(t) = 0.5 sin(2t) +
 T
Consider system (5) with 0.5 sin(5t) + 0.2 with V = 1.2, d = − 0.5, −0.5 , d =
       T
0 1 1 0 1 0.5, 0.5 .
A1 = , B1 = , A2 = ,  T
−0.35 −1.2 0 −0.8 −1.8
    For the observer gain matrices L1 = 1 4 , L2 =
2 0  T
B2 = , C1T = C2T = 2 6 :
0 1    
−10 0 − 0 5
and u(t) = 10 sin(2t), d(t) = 0.1 sin(5t)E2,1 , and v(t) = Ã1d + Ã+
1o = , Ã1o = ;
0.05 sin(6t) + 0.05 sin(t), which satisfy 4 −7 0 0
   
−0.1 · E2,1 = d ≤ d(t) ≤ d = 0.1 · E2,1 , |v(t)| ≤ V = 0.1. −8 2 − 0 0
Ã2d + Ã+
2o = , Ã2o =
For the observable pairs (Ai , Ci ) we choose the observer 0 −13 4 0
 T  T are Metzler. The stability of switched interval observer
gain matrices L1 = − 4.43 1.7 , L2 = − 2.75 1.7 .
   
−1 0.0 −1.5 0 15 State x
D1 = , D2 = , 10
Lower bound

0.0 −1.9 0 −2 5
Upper bound
x1

Furthermore, the matrices for coordinates change are 0

    −5

1.9669 −1.8877 2.7854 −2.6471 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
S1 = , S2 = 20

−0.3623 0.6607 −1.1142 1.4118 15

 T  10
x2

And the gain matrices M1 = 0.5 1.8 , M2 = − 2.5 − 5

T 0

2.5 are chosen in a way to guarantee the positivity of −5

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

reset matrices in the new coordinates. The stability of

the observer is fulfilled for all θ ∈ [0.2086, +∞). Fig. 3. Simulation results of the third approach.
The results of simulation are shown in Figure 1 for the
interval observer in the new coordinates z(t) by dashed (13) is verified by satisfying the LMIs (12). The simulation
lines and in Figure 2 the bounds of interval observer for result is shown in Figure 3.
the system are presented in red and blue color for upper
4.3 Example of the Trotter approximation based interval
and lower bounds respectively.
observer
40 New state z
Lower bound
30
Upper bound Consider system (23) with
20
     
z1

10
−5 3 −2 −1 T T 1
0
A1 = , A2 = , C1 = C2 =
−10 −1 −3 1 −4 0
40
5 10 15 20 25
 T
20
and d(t) = 2 sin(5t) + 10 sin(0.6t), 5 cos(t) + 3 sin(5t) ,
which satisfy d(t) ≤ d(t) ≤ d(t) where
2
z

−20    
−40 −2 + 10 sin(0.6t) 2 + 10 sin(0.6t)
0 5 10 15 20 25 30 d(t) = , d(t) = .
−3 + 5 cos(t) 3 + 5 cos(t)
Fig. 1. Simulation results of the second approach: new For the cooperativity of estimated error bounds, the pa-
state variables z. rameter m = 2 in the expression (21) guarantees the

6458
Proceedings of the 20th IFAC World Congress
6270
Toulouse, France, July 9-14, 2017 Djahid Rabehi et al. / IFAC PapersOnLine 50-1 (2017) 6265–6270
10
5 Monterey.
x1
0
−5
0 2 4 6 8 10 12 14 16 18
5
0 Efimov, D., Perruquetti, W., Raı̈ssi, T., and Zolghadri, A.
2
x
State x
−5
Lower estimate
Upper estimate
−10
0 2 4 6 8 10 12 14 16 18
Fig. 4. Simulation results of the Trotter approximation
based approach; the red (blue) points are the upper
(lower) estimates of the state at discrete time instants.
nonnegativity of ∆ and the Assumption 5 is fulfilled due
to the matrix  
−9.0 2.0
A1 − L1 C1 + A2 − L2 C2 =
1.5 −7.0
is Metzler. And the time of activation of each mode is
τ = 0.25. The stability of discrete interval observer (26) is
verified by satisfying the LMI in the Assumption 6. The
result of simulation is presented in Figure 4.
5. CONCLUSION
In this paper the interval observation for class of switched
systems has been considered. The uncertainties are pre-
sented as unknown inputs (also called model perturba-
tion), and output measurement noise. It is supposed that
the values of these uncertainties belong to certain known
intervals at any moment of time. Some sufficient condi-
tions of positivity of estimated error dynamics have been
adopted. Several approaches of designing interval observer
have been proposed based on cooperative system dynam-
ics. The three first approaches give continuous estimation,
where the second one is based on a coordinates transforma-
tion approach (Raissi et al., 2012) and the third one uses
the fact that any system can be presented in its internal
nonnegative form (Cacace et al., 2015). The last approach
is a sampled-time estimation from continuous measure-
ments based on the Trotter approximation (Trotter, 1959).
Demonstration of developed approaches have been shown
through academic examples.
REFERENCES
Barbot, J., Saadaoui, H., Djemai, M., and Manamanni, N.
(2007). Nonlinear observer for autonomous switching
systems with jumps. Nonlinear Analysis: Hybrid Sys-
tems, 1(4), 537–547.
Boukhobza, T. and Hamelin, F. (2011). Observability
of switching structured linear systems with unknown
input. a graph-theoretic approach. Automatica, 47(2),
395–402.
Briat, C. (2015). Convex conditions for robust stabi-
lization of uncertain switched systems with guaranteed
minimum and mode-dependent dwell-time. Systems &
Control Letters, 78, 63–72.
Cacace, F., Germani, A., Manes, C., et al. (2015). A new
approach to design interval observers for linear systems.
IEEE Trans. Automat. Contr., 60(6), 1665–1670.
Degue, K., Efimov, D., and Richard, J.P. (2016). Interval
observers for hybrid linear systems. In 10th IFAC
Symposium on Nonlinear Control Systems (NOLCOS).
Efimov, D., Fridman, L., Raissi, T., Zolghadri, A., and
Seydou, R. (2012). Interval estimation for lpv systems
20
applying high order sliding mode techniques. Automat-
ica, 48(9), 2365–2371.
(2013a). Interval observers for time-varying discrete-
time systems. IEEE Transactions on Automatic Con-
20
trol.
Efimov, D., Perruquetti, W., and Richard, J.P. (2013b).
Interval estimation for uncertain systems with time-
varying delays. International Journal of Control, 86(10),
1777–1787.
Efimov, D. and Raı̈ssi, T. (2016). Design of interval
observers for uncertain dynamical systems. Automation
and Remote Control, 77(2), 191–225.
Efimov, D., Raı̈ssi, T., and Zolghadri, A. (2013c). Control
of nonlinear and lpv systems: interval observer-based
framework. IEEE Transactions on Automatic Control,
58(3), 773–778.
Farina, L. and Rinaldi, S. (2000). Positive Linear Systems:
Theory and Applications, volume 50. John Wiley &
Sons.
Liberzon, D. (2003). Switching in Systems and Control.
Springer Science & Business Media.
Liberzon, D., Tempo, R., et al. (2004). Common lyapunov
functions and gradient algorithms. IEEE Transactions
on Automatic Control, 49(6), 990–994.
Lin, H. and Antsaklis, P.J. (2009). Stability and stabi-
lizability of switched linear systems: a survey of recent
results. IEEE Transactions on Automatic control, 54(2),
308–322.
Mazenc, F., Dinh, T.N., and Niculescu, S.I. (2014). In-
terval observers for discrete-time systems. International
Journal of Robust and Nonlinear Control, 24(17), 2867–
2890.
Minc, H. (1988). Nonnegative matrices. Wiley.
Mitkowski, W. (2008). Dynamical properties of metzler
systems. Technical Sciences, 56(4).
Raissi, T., Efimov, D., and Zolghadri, A. (2012). Interval
state estimation for a class of nonlinear systems. Auto-
matic Control, IEEE Transactions on, 57(1), 260–265.
Rı́os, H., Mincarelli, D., Efimov, D., Perruquetti, W.,
and Davila, J. (2015). Continuous and discrete state
estimation for switched lpv systems using parameter
identification. Automatica, 62, 139–147.
Tanwani, A., Shim, H., and Liberzon, D. (2013). Ob-
servability for switched linear systems: characterization
and observer design. IEEE Transactions on Automatic
Control, 58(4), 891–904.
Trotter, H.F. (1959). On the product of semi-groups of
operators. Proceedings of the American Mathematical
Society, 10(4), 545–551.
Vu, L., Chatterjee, D., and Liberzon, D. (2007). Input-
to-state stability of switched systems and switching
adaptive control. Automatica, 43(4), 639–646.
6459