Professional Documents
Culture Documents
DOI 10.1007/s11045-010-0118-z
Abstract The notions of input-containing and detectability subspaces are developed within
the context of observer synthesis for two-dimensional (2-D) Fornasini-Marchesini models.
Specifically, the paper considers observers which asymptotically estimate the local state, in
the sense that the error tends to zero as the reconstructed local state evolves away from possibly mismatched boundary values, modulo a detectability subspace. Ultimately, the synthesis
of such observers in the absence of explicit input information is addressed.
Keywords Fornasini-Marchesini models Detectability subspaces Observers
Unknown-input observation
1 Introduction
Controlled invariant subspaces were originally introduced by Basile and Marro (1969) to
solve decoupling and tracking problems for one-dimensional (1-D) systems. These subspaces
were subsequently studied by Wonham and Morse (1970). Conditioned invariant subspaces
for 1-D systems were also introduced by Basile and Marro (1969), as the duals of controlled
invariant subspaces. The role of such subspaces in solving state estimation problems was
first investigated in Basile and Marro (1969a). Later, conditioned invariance was studied by
Willems in terms of the existence of a class of observers Willems (1981). Specifically, for
any conditioned invariant subspace S that can be externally stabilised by output-injection,
This work was partially supported by the Australian Research Council (Discovery Grant DP0986577).
L. Ntogramatzidis (B)
Department of Mathematics and Statistics, Curtin University of Technology, Perth, WA 6845, Australia
e-mail: L.Ntogramatzidis@curtin.edu.au
M. Cantoni
Department of Electrical and Electronic Engineering, The University of Melbourne, Parkville,
VIC 3010, Australia
e-mail: cantoni@unimelb.edu.au
123
80
there exists an observer that asymptotically recovers the state modulo S ; see also the textbooks (Basile and Marro, 1992, Chapter 4) and (Trentelman et al., 2001, Chapter 5). Unlike
the classic Luenberger observer, this class of observers does not directly exploit the input in
reconstructing the state. The relevance of this for problems in fault-detection and isolation
is well established, (Massoumnia 1986; Verghese et al. 1989).
Over the last twenty years, several extensions of important geometric concepts, such as
controlled invariance, have been proposed for 2-D systems, including the so-called
Fornasini-Marchesini and Roesser models, (Conte and Perdon 1988; Karamancioglu and
Lewis 1990, 1992).
While definitions of controlled invariance are not difficult to establish for FornasiniMarchesini models, definitions of conditioned invariance are less obvious because duality
cannot be exploited as in the 1-D case. The definitions of conditioned invariance in Karamancioglu and Lewis (1990), Karamancioglu and Lewis (1992) have the disadvantage of being
defined for models with very special structures; that is, with duality properties absent in the
case of the standard Fornasini-Marchesini model. In Conte and Perdon (1988), a definition of
conditioned invariance is proposed for the standard (first order) Fornasini-Marchesini model
(Fornasini and Marchesini 1978), governed by
xi+1, j+1 = A1 xi+1, j + A2 xi, j+1 + B1 u i+1, j + B2 u i, j+1 ,
yi, j = C xi, j + D u i, j .
(1)
The definition of conditioned invariance in Conte and Perdon (1988) is related therein to
the reconstruction of local state trajectories given a record of the output, in the case of exact
knowledge of the boundary conditions for (1). The problem of local state estimation with
decaying error in the case of unknown boundary conditions has not been considered. Motivated by this, we develop a systematic procedure for the external stabilisation of conditioned
invariant subspaces via output injections. This will lead to a notion of 2-D detectability
subspaces, i.e. conditioned invariant subspace that can be externally stabilised by output
injection. An algorithm for computing a stabilising output injection matrix is also provided
in terms of matrix inequalities. The approach is similar to Ntogramatzidis et al. (2008), where
internal stabilisability of output-nulling controlled invariant subspaces is studied. Ultimately,
the notion of a detectabilty subspace is linked to the existence of a local state observer
i+1, j+1 = K 1 i+1, j + K 2 i, j+1 + L 1 yi+1, j + L 2 yi, j+1 .
(2)
It is required that the size of the estimation error ei, j = Q xi, j i, j asymptotically
approaches zero as i + j , for some full row-rank matrix Q and arbitrary boundary conditions for xi, j and i, j .The local state observer (2), like its 1-D counterparts defined
in Willems (1981), does not exploit knowledge of the input. As such, it is structurally different from the Luenberger-type 2-D observers discussed in Bisiacco (1986), since its local
state does not explicitly depend on the input u i, j . For more details on the synthesis of 2-D
observers, see also Kaczorek (1992), Kaczorek (2001) and the references therein. As an
important application of the concepts described above, we present a solution to the so-called
unknown-input observation problem for the standard 2-D system model (2), based on a
sufficient constructive condition given in terms of a standard subspace inclusion. The relevance in fault detection and non-interaction are well known, (Bisiacco and Valcher 2006,a).
In Bisiacco and Valcher (2004), a polynomial approach is employed to develop necessary and
sufficient conditions for the solution of this problem under the requirement that the observer
error exhibit dead-beat dynamics. The conditions involve Bzout equations, which can be
difficult to solve. In this paper, the unknown-input observation problem is considered under
def
123
81
the weaker requirement that the estimation error only asymptotically converge to zero away
from the location of the unknown boundary conditions.
Notation. Throughout this paper, we denote by N the set of positive integers including zero.
The image and the kernel of matrix M Rn m are denoted by im M and ker M, respectively.
The n m zero matrix is denoted by 0nm . We also denote by M and by M the transpose
and the Moore-Penrose pseudoinverse of M, respectively. Given a subspace Y of Rm , the
symbol M 1 Y stands for the inverse image of Y with respect to the linear transformation M.
For the sake of brevity, we define M D := diag(M, M), and, accordingly, given a subspace
J of Rn , the symbol J D will identify the subspace J J of R2 n . Finally, given the vector Rn , the symbol /J denotes the canonical projection of on the quotient space Rn /J .
(3)
where, for all integers i, j, the vector xi, j Rn is the local state of the system. Here,
A1 , A2 Rnn . Defining for each k Z the separation set
Sk = {(i, j) Z Z | i + j = k},
def
it is easily seen from (3) that Xk can be uniquely expressed in terms of Xk1 , (Fornasini and
Marchesini 1978). In particular, if we fix the values of xi, j on S0 , i.e. fix X0 as a boundary
condition, (3) uniquely determines Xk for k > 0 (i.e., xi, j for i + j > 0).1 Indeed, these are the
boundary conditions usually associated with the FM model (3). In the sequel, given a subspace
W Rn , by a W -valued boundary condition we intend the set {xi, j W | (i, j) S0 }.
Similarly, for each k > 0, the global state Xk is said to be W -valued when xi, j W for all
(i, j) Sk . We also define the set of indexes for which the local state xi, j of (3) is uniquely
determined by fixing X0 as boundary condition:
def
S+ =
Sk = {(i, j) Z Z | i + j 0}.
kN
1 As shown in Fornasini and Marchesini (1980), other separation sets can be defined so that boundary condi-
tions specified over them uniquely determine a local-state trajectory solution of (3) over a region of Z Z.
def
An interesting and useful example is the separation set Sk = {(i, j) {0}[1, ) [1, ){0}}, which
def
with corresponding boundary conditions uniquely determines xi, j for (i, j) S+ = N N \ {(0, 0)}. Most
of the considerations in this paper can be adapted to such separations sets.
123
82
Ai = S Ai S =
.
0(nr )r A i22
Matrix S is any basis matrix of Rn adapted to J , i.e., such that its first columns
span J .
Sc
, where
Equivalently, S can be constructed as the square non-singular matrix S =
Q
ker Q = J and the rows of Sc are linearly independent from those of Q. With respect to this
new set of coordinates, model (3) can be written as
11
11
xi+1, j+1
xi+1, j
xi, j+1
A 1 A 12
A 2
A 12
1
2
=
+
.
(5)
xi+1,
xi+1,
xi,
j+1
0
A 22
0
A 22
j+1
j
1
2
Any boundary condition {xi, j = bi, j J | (i, j) S0 } is such that xi,
j = 0 for (i, j) S0 .
Therefore, by (5), xi,
j = 0 for all (i, j) S+ , which implies that xi, j J for all i, j S+ .
123
83
xi,
j
= S xi, j , the component xi,
j is the projecxi,
j
tion of the local state xi, j onto the invariant subspace J , while xi,
j is the canonical projection
on to the quotient space Rn /J . Thus, we refer to xi,
j of xi, j as the internal (or inner) component of the local state (with respect to J ), and to xi,
j of xi, j as the external (or outer)
component of the local state (with respect to J ).
In the basis corresponding to S, whereby
for any boundary condition satisfying X0 < , the corresponding sequence {Xi }i=0
converges to zero (Fornasini and Marchesini 1978). This property is clearly invariant under
coordinate transformation and with a slight abuse of nomenclature, the system matrix pair
(A1 , A2 ) is called asymptotically stable in this case. It is well-known that the pair (A1 , A2 )
is asymptotically stable if, and only if,
def
det(In A1 z 2 A2 z 1 ) = 0 (z 1 , z 2 ) P
(6)
is imposed, for all k 0 the global state Xk
associated with the external dynamics (5)
satisfies Xk
= 0. Hence, the internal dynamics on J satisfy
11
11
xi+1,
j+1 = A1 x i+1, j + A2 x i, j+1 .
(9)
123
84
11 alone is also asymptotically stable, then the global state X
associated with
If A 11
,
A
1
2
k
in this case, the elements of the global state Xk associated with (3) approach the invariant
subspace J , as k . Finally, in view of the discussion above, note that the model (3) is
asymptotically stable if, and only if, any (A1 , A2 )-invariant subspace is both internally and
externally stable. The following corollary will be important in the sequel.
Corollary 2.1 Given an r -dimensional subspace J of Rn , let Q R(nr )n be a full rowrank matrix such that ker Q = J . Then J is an externally stable (A1 , A2 )-invariant subspace if, and only if, an asymptotically stable pair (L 1 , L 2 ) exists such that Q Ai = L i Q
for i {1, 2}.
Proof
With respect to a basis of Rn adapted to J , a matrix whose columns span J is
Ir r
, and therefore Q = [ 0(nr )r I(nr ) ] is a full row-rank matrix such that
0(nr )r
ker Q = J . Writing the identities Q Ai = L i Q for i {1, 2} with respect to this basis
yields
Ai12
Ai11
0(nr )r Inr
= L i 0(nr )r Inr ,
22
0(nr )r Ai
leading to L i = Ai22 for all i {1, 2}.
(10)
def
As for the 1-D case, it is easily seen that the set of conditioned invariant subspaces is
closed under subspace intersection but not under subspace addition. Its smallest element
(with respect to the partial relation of subspace inclusion ) is {0}, its largest element is Rn .
In the following lemma, the most important properties of 2-D conditioned invariance are
given.
123
85
(11)
(12)
QD
Proof (1) (2). Inclusion (10) can be written in matrix notation as ker
CD
ker Q [ A1 A2 ]. Hence, matrices R(ns)2(ns) and R(ns)2 p exist such that
Q [ A1 A2 ] = Q D + C D .
(2) (3). Equation (12) follows from (11) with any G such that = Q G. Such a
matrix G always exists as Q is of full row-rank.
(3) (1). This follows from the definition.
Remark 3.1 Notice that property (2) in Lemma 3.1 can be written equivalently as
Q Ai = i Q + i C for i {1, 2}.
As a consequence, inclusion (10) in Definition 3.1 can also be written as Ai (S ker C) S
for i {1, 2}, which coincides with the definition of 2-D conditioned invariance given in
Conte and Perdon (1988). Moreover, (3) in Lemma 3.1 is equivalent to
(Ai + G i C) S S for i {1, 2}.
This means that S is conditioned invariant for (1) if and only if there exists an output-injection matrix G = [ G 1 G 2 ] Rn2 p such that S is an (A1 + G 1 C, A2 + G 2 C)-invariant
subspace.
3.1 Construction of stabilising output-injection matrices
Our aim now is to establish a procedure that enables an output-injection matrix G to be
determined such that S is an externally stable (A1 + G 1 C, A2 + G 2 C)-invariant subspace.
As for the 1-D case, we say that S is externally stabilisable if we can find an output-injection
matrix G such that S is an externally stable (A1 + G 1 C, A2 + G 2 C)-invariant subspace.
To find all the output-injection matrices associated with the conditioned invariant subspace
S , let and be such that (11) holds, which can be written as the linear equation
QD
Q AH =
.
(13)
CD
This equation can be solved for and . The solutions of (13) are given by
QD
= Q AH
+ K H,
(14)
CD
QD
, while K is an arbitrary
where H has linearly independent rows and ker H = im
CD
matrix of suitable size. As it will become clear in the sequel, K represents a first degree of
123
86
freedom in the construction of the output-injection matrix, that can be exploited to externally
QD
is
stabilise the 2-D conditioned invariant subspace S . Notice that in the case when
CD
QD
full-rank, the only solution of (13) is [ ] = Q A H
, and this degree of freedom
CD
disappears.
By (12), = [ 1 2 ] exists such that
Q (A H + G C D ) = Q D ,
(15)
or, equivalently, such that Q (Ai + G i C) = i Q, for i {1, 2}. We now investigate the rela satisfying (13) and (15), respectively. Given a pair
tion between the pairs (, ) and (G, )
such that (15) holds, then (13) is satisfied with = and = Q G. Conversely,
(G, )
given a pair of matrices (, ) such that (13) holds, then (15) is satisfied with = and
with any G such that = Q G. As such, no generality is lost by assuming = , and by
representing the set of all output-injection matrices associated with the conditioned invariant
subspace S as the set of matrices G Rn2 p satisfying = Q G, where R(ns)2 p
is any matrix for which R(ns)2(ns) exists so that (13) holds. For any pair (, ) such
that (13) holds, the solutions of the linear equation = Q G are parameterised as
G = G + U,
(16)
S (Ai + G i C) S
i11 (K , U )
i12 (K , U )
i22 (K , U )
.
(17)
123
87
(22)
Proof First, we show that i (K , U ) does not depend on U . To this end, let U1 , U2 be two
arbitrary matrices of suitable size. From (17) we find that
i11 (K , U1 ) i11 (K , U2 ) i12 (K , U1 ) i12 (K , U2 )
0
i22 (K , U1 ) i22 (K , U2 )
= S (Ai + G C + U1 C) S 1 S (Ai + G C + U2 C) S 1
Sc
=
(U1 U2 ) C S 1 ,
Q
so that
which is equal to zero since Q is the zero matrix by definition of . Hence, i22 (K , U1 ) =
i22 (K , U2 ), which implies that the term i22 (K , U ) in (17) does not depend on U . Now we
show that i11 (K , U ) does not depend on K . First, from (17), it is found that
(18)
Sc Ai Q (Q Q )1 C + U C = i11 (K , U ) Sc + i12 (K , U ) Q,
where U is an arbitrary matrix of suitable size. Let 1 and 1 be the solution of (14) with K =
K 1 , and let 2 and 2 be the solution of (14) with K = K 2 . Since Q A H = j Q D + j C D ,
for j {1, 2}, we get (1 2 ) C D = (1 2 ) Q D . By partitioning (1 2 ) as
[ 1 2 ], we get (1 2 ) C = i Q. Writing (18) with respect to 1 and 2 and by
computing the difference yields
Sc Q (Q Q )1 (1 2 ) C = i11 (K 1 , U ) i11 (K 2 , U ) Sc
+ i12 (K 1 , U ) i12 (K 2 , U ) Q,
so that
Sc Q (Q Q )1 i Q= i11 (K 1 , U )i11 (K 2 , U ) Sc + i12 (K 1 , U )i12 (K 2 , U ) Q.
i11 (K 1 , U ) i11 (K 2 , U ) Sc = 0,
the second yielding i11 (K 1 , U ) = i11 (K 2 , U ) since Sc has linearly independent rows.
Therefore, the term i11 (K , U ) in (17) does not depend on K .
Now we want to find a method to design the output-injection matrix G = [ G 1 G 2 ] such
that S is an externally stable (A1 + G 1 C, A2 + G 2 C)-invariant subspace; i.e., such that
there exists an asymptotically stable pair (1 , 2 ) for which Q (A H + G C D ) = Q D .
For a given a conditioned invariant S , write (14) as
[1 2 ] = [V1 V2 V ] + K [H1 H2 H ],
(19)
QD
and [ H1 H2 H ] = H are partitioned comformably
where [ V1 V2 V ] = Q A H
CD
with [ 1 2 ], i.e., i = Vi +K Hi for i = 1, 2 and = V +K H . If S D +ker C D = R2n ,
123
88
there is only one solution to (13), so that there are no degrees of freedom in the choice of the
pair (1 , 2 ). In this case, if (1 , 2 ) is asymptotically stable, then with the corresponding
def
= V , the matrix G = Q (Q Q )1 = [ G ,1 G ,2 ] is such that S is an externally stable (A1 + G ,1 C, A2 + G ,2 C)-invariant subspace. On the other hand, if the pair
(1 , 2 ) is not asymptotically stable, the subspace S is not externally stabilisable.
Now, when S D + ker C D R2n , the problem we need to solve is to find a matrix K
such that the resulting pair (1 , 2 ) = (V1 + K H1 , V2 + K H2 ) is asymptotically stable; the corresponding = V + K H , for which (, ) is a solution of (11), is such that
def
G = Q (Q Q )1 , yielding Q (A H + G C D ) = Q D , so that S is an externally
stable (A1 + G ,1 C, A2 + G ,2 C)-invariant subspace.
Towards characterising a subset of such matrices K , we can virtually exploit any stability criterion for 2-D Fornasini-Marchesini models. As mentioned, necessary and sufficient
conditions have recently appeared in the literature that characterise stability in finite terms,
(Ebihara et al. 2006; Zhou 2006). For the sake of simplicity, however, we consider the sufficient condition recalled in Lemma 2.2, whose structure appears to be much less involved.
Let us rewrite this condition for asymptotic stability in Lemma 2.2 for the pair (1 , 2 ) as
1
0
1 2 > 0,
0
2
for some
= P1 > 0 and = P1 + P2 > 0. Standard manipulation and i = Vi + K Hi ,
for i = 1, 2, yield the equivalent condition
0
( V1 + H1 )
(20)
0
( V2 + H2 ) > 0
V1 + H1 V2 + H2
def
def
for some
> 0, > 0, and of suitable dimensions, where = K . We have just
proved the following result
Theorem 3.1 Let S be a conditioned invariant subspace for (1). Then, S is an externally
stabilisable conditioned invariant subspace if there exist
=
> 0, = > 0 and
of suitable dimensions such that (20) holds. Moreover, given a triple (,
, ) in the convex
set defined by (20), a matrix K for which the pair (1 , 2 ) is asymptotically stable is given
by K = 1 .
A H B H (S D R2m ) ker C D D D S .
As for the 1-D case, it is easy to see that the intersection of two input-containing subspaces
is input-containing. It follows that the set of input-containing subspaces for (1) is closed under
subspace intersection. The same is not true for subspace addition. This is due to the fact that
123
89
S i = A H BH
(S Di1 R3 m ) ker C D D D ,
Q A H B H = Q D 02 (ns)2 m + C D D D ;
(21)
A H + G C D B H + G D D S D R2m S
(22)
Proof The result follows in the same way as the result in Lemma 3.1.
Following the procedure outlined for 2-D conditioned invariant subspaces, to find the set
of ouput-injection matrices associated with the input-containing subspace S , we first solve
(21) with respect to and , obtaining
QD
0
= Q A H BH
+ K H,
C D DD
QD 0
where H is full row-rank, ker H = im
, and K is an arbitrary matrix of suitable
C D DD
size. Using (22), we compute the solutions of = Q G as G = G + U . As for conditioned invariant subspaces, K represents the degree of freedom that can be used to assign the
external dynamics of the input-containing subspace S , e.g. by means of an LMI condition
similar to that given in Theorem 3.1. As such, we say that S is a detectability subspace if
an output-injection matrix G exists (or, equivalently, if K exists) such that (22) holds and S
is an externally stable (A1 + G 1 C, A2 + G 2 C)-invariant subspace. It can be straightforwardly established that an exact equivalent of Theorem 3.1 holds for detectability subspaces
by simply writing (19) with
QD
QD
0
0
A
B
=
Q
=
im
and
ker
.
V1 V2 V
H1 H2 H
H
H
C D DD
C D DD
(23)
123
90
Detectability input-containing subspaces can be linked to the existence of certain observers (Ntogramatzidis and Cantoni 2007). Consider a system
governed by a FornasiniMarchesini model (1). Given a subspace S of Rn , the 2-D system
O ruled by (2) is said to
be an S -quotient observer if for any boundary condition of
and
O , the local state of
O
asymptotically reconstructs the local state xi, j of
modulo the components of this vector on
S . In other words, on the basis of the observations y, the vector i, j asymptotically converges
to xi, j /S , as the indexes i and j evolve away from the boundary, regardless of the boundary
conditions of
and
O .
Obviously, given an arbitrary subspace S of Rn , an S -quotient observer does not necessarily exists. But if this subspace is a detectability subspace, the existence of such an observer
is guaranteed. Let Q be a full row-rank matrix such that ker Q = S . Define the new variable
ei, j = Q xi, j i, j , along with the vectors x(i,
j) = [ xi,j xi+1,
j) =
j x i, j+1 ] , u(i,
[ u i, j u i+1, j u i, j+1 ] and (i,
j) = [ i, j i+1, j i, j+1 ] , (i, j) S+ . Let and
be such that (21) holds. Let system (2) be defined by K H = and L H = . It is found that
ei+1, j+1 = Q xi+1, j+1 i+1, j+1
= Q A H x(i,
j) + Q B H u(i,
j) (i,
j) Q G C D x(i,
j) Q G D D u(i,
j)
x(i,
j)
= Q A H + G C D BH + G DD
(i,
j) = 1 ei+1, j + 2 ei, j+1 ,
u(i,
j)
where (22) has been used. Moreover, since S is a detectability subspace, the pair (1 , 2 )
is asymptotically stable. Therefore, the estimation error converges to zero as the index (i, j)
evolves away from S0 , so that i, j asymptotically converges to Q xi, j . Since ker Q = S ,
this means that
O recovers the external components of xi, j with respect to S . Notice that if
S is an input-containing subspace but not a detectability subspace, the estimation error does
not converge to zero, unless the error is zero over the entire boundary region.
Note that the characterisation of external stabilisability for conditioned invariant and inputcontaining subspaces is essential in employing these ideas in the construction of local state
observers. Indeed, the fact that the subspace S is input-containing alone can only guarantee
that
O gives rise to an estimation error that only depends on the boundary conditions. Therefore,
O can only guarantee that when i, j = Q xi, j for (i, j) S0 , then the estimation error
is identically zero, which means that i, j = Q xi, j for all S+ .
Example 4.1 Consider (1) with
0.35 0 0.01 0
0
0 0.05 0
0
0.5 0.01
0.1 0
0
0
0
A1 =
0.01 0 0.4 0.4 , A2 = 0.35 0
0.35 0.05
0
0 0.02 0.15
0
0 0.45 0.15
0 0
0 2
1 6
7 0
B1 =
0 5, B2 = 0 0 ,
0 2
0 0
0.4 0 0 0
0 2
C =
, D=
.
0 0 0 2
0 7
123
91
Using Algorithm 4.1, which in this case converges in one step, it is found that S =
im [ 0 1 0 0 ] , so that the kernel of
1 0 0 0
Q = 0 0 1 0
0 0 0 1
QD 0
def
is exactly S . By defining W =
, a matrix H whose kernel is the image of W can
C D DD
be easily computed:
0 0
0
25 57 0
0
1 0 10
7
.
H=
5
0
0
57
0 0
0
1 0 10
7
2
Using (23), it is easy to verifyfor example exploiting the MATLAB routines setlmis.m,
lmiterm.m, getlmis.m and feasp.mthat the LMI condition (20) is satisfied with
0.1101 0.9151
0.0314
0.1717 , which yields K = 1 = 0.9209
= 0.9581
0.4129
0.4502
0.1925
As such,
1 2
1 2
0.1346
0.1252
0.2691
0.0100
0.4000
0.0200
0.0959
0.0593
0.0290
0.3058
0.4026
0.2934
0.0500
0.3500
0.4500
0.1678
= 3.0962
0.6868
0.0479
0.1703
0.0895
0.7645
0.1315
0.7334
0.0673
0.0376 .
0.2095
0.2829
= 1.2485
0.2747
and
0.2017
0.0956 .
0.2715
ei, j = Q xi, j i, j =
xi, j i, j .
(4)
(3)
xi, j i, j
If for simplicity of representation we consider as boundary the set of indexes ({0} [1, 20])
([1, 20] {0}), and we assign random boundary conditions (using randn.m) on this region,
using the randomly generated input functions of Fig. 1,we obtain that the estimation error is
the one given in Fig. 2. Notice that indeed the estimation errors converge to zero as the index
(i, j) moves away from the boundary.
123
92
400
300
200
200
100
0
100
200
200
400
20
300
20
15
10
5
10
0 0
15
20
15
10
5
0
15
10
20
Fig. 1 Inputs u 1 and u 2 in the bounded frame [0, 20] [0, 20]
20
10
0
10
20
30
20
15
10
5
00
10
15
20
30
20
10
0
10
20
30
20
20
10
0
10
20
15
10
5
00
10
15
20
30
20
15
10
5
00
10
15
20
Fig. 2 Estimation errors over the bounded frame [0, 20] [0, 20]
5 Unknown-input observers
In this section, we use the geometric notions developed so far for the solution of the unknowninput observation problem, which plays an important role in signal reconstruction problems,
fault-detection and identification, non-interaction control. Consider the Fornasini-Marchesini
model
xi+1, j+1 = A1 xi+1, j + A2 xi, j+1 + B1 u i+1, j + B2 u i, j+1
yi, j = C xi, j + D u i, j
(24)
z i, j = R xi, j + S u i, j
where, for all integers i, j, vector u i, j Rm represents an input which is not accessible for
measurement. The variable yi, j R p1 represents an output that can be measured and the
variable z i, j R p2 is an output that we want to estimate on the basis of the measurement
y. All matrices appearing in (24) are of appropriate dimensions. Consider the block diagram
depicted in Fig. 3. Let the observer
O be described by the equations
i+1, j+1 = K 1 i+1, j + K 2 i, j+1 + L 1 yi+1, j + L 2 yi, j+1 ,
i, j = M i, j + N yi, j
(25)
denote the overall system from the input u to the output e := z . With the choice
and let
of the structure of the observer
O , the overall system is governed by
123
93
xi+1, j+1
i+1, j+1
A1 0
xi+1, j
A2 0
xi, j+1
+
L 1 C K1
L 2 C K2
i+1, j
i, j+1
B2
B1
u i+1, j +
u i, j+1 ,
+
L1 D
L2 D
xi, j
= R N CM
+ (S N D) u i, j .
i, j
ei, j
R S =
Q 0 + C D .
(26)
The solutions
and of (26) are parameterised as
= R
Q
C
0
D
+ K H,
Q C
and K is an arbitrary matrix of
0 D
suitable
size.Furthermore, the matrices
and satisfying (26) are unique if and only if the
Q0
map
is surjective. When [ C D ] is full row-rank, this condition is equivalent to
C D
where the rows of H span the null-space of
123
94
R D SD =
D Q D 0 + D C D D D .
(27)
Q C
is zero if and only if such is the kernel of
Now, since obviously the kernel of
0 D
Q D CD
, it turns out that in the case where C S + im D = R p1 , equation
0 S D
Q A H B H = Q D 0 + C D D D
admits a unique solution, so that the four matrices , ,
and can be uniquely determined.
Since S is a detectability subspace, there exists an output-injection matrix G such that (22)
holds with an asymptotically stable pair (1 , 2 ). We show that the dynamical system
O
ruled by (25) with K k = k , L k = Q G k , (k {1, 2}), M =
, and N = solves the
unknown-input observation problem. First, note that in view of (26)
xi, j
ei, j = R N CM
+ (S N D) u i, j
i, j
= (R C) xi, j
i, j + (S D) u i, j
xi, j
= R S C D
i, j
u i, j
xi, j
i, j =
(Q xi, j i, j ).
=
Q0
u i, j
Define i, j := Q xi, j i, j using the same notation of Sect. 4. Given the signal s : S+ Rh
def
for some h, let also s (i, j) = [ si,j si+1,
j si, j+1 ] . It follows that
i+1, j+1 = Q xi+1, j+1 i+1, j+1
x(i,
j)
= Q A H BH
(i,
j)
u(i,
j)
+Q G 1 (C xi+1, j + D u i+1, j ) + Q G 2 (C xi, j+1 + D u i, j+1 )
x(i,
j)
= Q A H BH + G C D DD
(i,
j)
u(i,
j)
x(i,
j)
= Q 0
(i,
j) = 1 i+1, j + 2 i, j+1 .
u(i,
j)
Hence, the signal i, j is independent of u i, j , and since ei, j =
i, j , such is also the estimation error ei, j . It follows that if i, j = z i, j for all (i, j) S0 , then i, j = z i, j for all
(i, j) S+ . Moreover, in view of the asymptotic stability of the pair (1 , 2 ), it also follows
that for all boundary conditions i, j and xi, j , (i, j) S0 , the estimation error ei, j converges
to zero as (i, j) moves away from S0 .
6 Concluding remarks
The paper develops notions of conditioned invariant and detectability subspaces for 2-D
Fornasini-Marchesini models. By contrast with earlier work, the development here leads to
2 Recall that C 1 im D = {x Rn | C x im D}.
123
95
an LMI based procedure for the synthesis of observers which asymptotically estimate the
local state of a standard Fornasini-Marchesini model, in the sense that the error tends to zero
as the reconstructed local state evolves away from unknown boundary conditions. The geometric notions and results presented here complement those in Ntogramatzidis et al. (2008),
where notions of controlled-invariance and stabilisability are developed within the context
of 2-D disturbance decoupling problems. It is expect that the results of this paper will lead
to 2-D extensions of techniques for the detection and identification of faults, as developed in
Massoumnia (1986) and Verghese et al. (1989).
References
Basile, G., & Marro, G. (1969). Controlled and conditioned invariant subspaces in linear system theory. Journal of Optimization Theory and Applications, 3(5), 306315.
Basile, G., & Marro, G. (1969). On the observability of linear, time-invariant systems with unknown
inputs. Journal of Optimization Theory and Applications, 3(6), 410415.
Basile, G., & Marro, G. (1992). Controlled and conditioned invariants in linear system theory. Englewood
Cliffs, New Jersey: Prentice Hall.
Bisiacco, M. (1986). On the structure of 2D observers. IEEE Transactions on Automatic Control, AC31(7), 676680.
Bisiacco, M., & Valcher, M. E. (2004). Unknown input observers for 2D state-space models. International
Journal of Control 77(9), 861876(16)
Bisiacco, M., & Valcher, M. E. (2006). Observer-based fault detection and isolation for 2D state-space
models. Multidimensional Systems and Signal Processing, 17, 219242.
Bisiacco, M., & Valcher, M. E. (2006). The general fault detection and isolation problem for 2D state-space
models. Systems & Control Letters, 55, 894899.
Bliman, P. A. (2002). Lyapunov equation for the stability of 2-D systems. Multidimensional Systems and
Signal Processing, 13, 201222.
Conte, G., & Perdon, A. (1988). On the geometry of 2D systems. Proc. IEEE Int. Symp. on Circuit and
Systems, Helsinki, Finland.
Ebihara, Y., Ito, Y., & Hagiwara, T. (2006). Exact stability analysis of 2-D systems using LMIs. IEEE
Transactions on Automatic Control, AC-51(9), 15091513.
Fornasini, E., & Marchesini, G. (1976). State-space realization theory of two-dimensional filters. IEEE
Transactions on Automatic Control, AC-21(4), 484492.
Fornasini, E., & Marchesini, G. (1976). Two dimensional filters: New aspects of the realization theory.
Proceedings of the 3rd Conference on Pattern Recognition, Coronado, CA.
Fornasini, E., & Marchesini, G. (1978). Doubly-indexed dynamical systems: State-space models and
structural properties. Mathematical System Theory, 12, 5972.
Fornasini, E., & Marchesini, G. (1980). Stability analysis of 2-D systems. IEEE Transactions on Circuits
and Systems, CAS-27(12), 12101217.
Hu, G. D., & Liu, M. (2005). Simple criteria for stability of two-dimensional linear systems. IEEE
Transactions on Signal Processing, 53(12), 47204723.
Kaczorek, T. (1988). The singular general model of 2-D systems and its solution. IEEE Transactions on
Automatic Control, AC-33, 10601061.
Kaczorek, T. (1992). Linear control systems. NY: Wiley.
Kaczorek, T. (2001). Perfect observers for singular 2-D Fornasini-Marchesini models. IEEE Transactions
on Automatic Control, AC-40, 16711675.
Kar, H., & Sigh, V. (2003). Stability of 2-D systems described by the Fornasini-Marchesini first model. IEEE
Transactions on Signal Processing, 51(6), 16751676.
Karamancioglu, A., & Lewis, F. L. (1990). A geometric approach to 2-D implicit systems. In Proceedings
of the 29th Conference on Decision and Control, Honolulu, Hawaii, December.
Karamancioglu, A., & Lewis, F. L. (1992). Geometric theory for the singular Roesser model. IEEE
Transactions on Automatic Control, AC-37(6), 801806.
Kung, S.-Y., Levy, B. C., Morf, M., & Kailath, T. (1977). New results in 2-D systems theory, Part II:
2-D state-space models-realization and the notions of controllability, observability and minimality,
In Proceedings of the IEEE, 65(6), 945961, June.
Kurek, J. (1985). The general state-space model for a two-dimensional linear digital system. IEEE
Transactions on Automatic Control AC, 30(6), 600602.
123
96
Author Biographies
Lorenzo Ntogramatzidis received the Laurea degree, cum laude,
in Computer Engineering in 2001 from the University of Bologna, Italy. He received the Ph.D. degree in Control and Operations Research in 2005. From 2005 to 2008, he was a post-doctoral Research Fellow at the Department of Electrical and Electronic
Engineering, The University of Melbourne, Australia. Since 2009,
he has been with the Department of Mathematics and Statistics at
Curtin University of Technology, Perth, Australia, where he is currently an ARC APD Research Fellow. His research interests are in
the area of systems and control theory.
Michael Cantoni received the B.E. (Hons. I) in electrical engineering and the B.Sc. degree in applied mathematics from the University
of Western Australia in 1995. He subsequently received the Ph.D.
degree in engineering from the University of Cambridge, U.K., in
1998. From 1998 to 2000, Dr. Cantoni was a Research Associate
with the Department of Engineering at The University of Cambridge
and a Junior Research Fellow with St. Johns College, Cambridge.
Since 2000, he has been with the Department of Electrical and Electronic Engineering, The University of Melbourne, Australia, where
is currently Associate Professor. His research interests lie in the
broad area of systems and control theory.
123