IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 65, NO.

2, FEBRUARY 2020 831

Design Procedure for Linear Unknown Input Functional Observers

Imane Sakhraoui , Baptiste Trajin , and Frédéric Rotella

Abstract—This paper considers an iterative procedure to design mechanisms to increase the order of a functional observer or a UIFO,
a minimum possible order unknown input functional observer for such that an asymptotically stable observer can be obtained even under
linear time-invariant systems. Necessary and sufficient conditions
for the existence of the observer are given. The proposed procedure nondetectability of the system [21], [22]. However, to cope with the
is simple and is based neither on the solution of a generalized design of minimum possible order functional observer a methodology
Sylvester equation nor canonical forms. This procedure can be to obtain a UIFO through an increase of the number of functional to
easily implemented due to its incremental form. The main feature be observed is presented in [22]. However, between the fixed-pole ob-
of our procedure is then the simplicity of the design. Moreover, in
the case of a single functional to be observed, a minimum order
server problem where the poles are specified at the outset and the stable
stable linear unknown input functional observer is obtained. observer problem where the poles are permitted to lie anywhere in the
left half-plane, finding the lowest possible order linear unknown input
Index Terms—Linear systems, time-invariant systems, unknown observer (LUIFO) is one of the main challenges in this field of research
input functional observers.
that is still not fully addressed [22].
Rewriting the necessary and sufficient conditions for the existence
of a stable UIFO for a linear system expressed in [23], a direct and
I. INTRODUCTION iterative procedure to get a minimal order for a single stable linear
functional observer has been proposed in [24] and for multifunctional
For a dynamic system, the observation problem consists in the esti-
observers in [25]. The herein presented paper extends these results
mation of some internal variables, that are not measured, from measure-
towards minimality of LUIFO. The proposed procedure is simple, iter-
ments given by sensors. From the seminal works of Luenberger [1], [2],
ative and is not based on the solution of so-called Sylvester equation or
solutions of the observation problem for a linear time-invariant system
on the use of canonical state space forms. The term iterative indicates
have been conducted in two main ways. On the one hand, methods have
that an increasing sequence for the order of the observer is tested to
been developed for the design of linear functional observers (LFO) of
obtain a possible minimal order. The main feature of the proposed de-
the state vector of the system. This way of designing LFO, which has
sign procedure is the highlighting of some degrees of freedom to place
been considered in the first beginning of the observers theory, leads to
some poles of the obtained observer.
reduced order observers. On the other hand, especially for diagnosis
This paper is organized as follows. In Section II, the well known
[3] or for robustness purposes, the observation problem of the whole
necessary and sufficient conditions for the existence of a LUIFO are
state with unknown inputs or disturbances has led to design methods
outlined. It is underlined that these conditions are focussed on the ex-
for unknown input linear observers (UILO). Existence conditions and
istence of a linear relationship between the state of the system and the
algorithms of these observers have been reported in [4]–[9] and in the
state of the observer. To overcome this difficulty, which has been the
books [10], [11]. The cited papers are only milestones and the inter-
keypoint in the design of linear observers over many years, Darouach
ested reader can find more details in the books and in the references
has proposed seminal criteria in [26] for LFO and in [23] for LUIFO.
therein.
In Section III, the main results are claimed. A necessary and sufficient
Functional observers are a generalized version of ordinary full or-
condition is detailed for the existence of an asymptotic observer. This
der or reduced order Luenberger observers that aim at reconstructing
necessary and sufficient condition has been stated for linear functional
single or multiple functions of the states of the system [4], [11]–[13].
observer where the inputs are completely measured in [7] and [27].
These observers offer the advantage to be of lower order compared to
The sufficient condition is proved here in the unknown input case. As
the ordinary reduced order observers of the whole state. Nevertheless,
a specific feature, these conditions can be tested on the observation
unknown or unmeasured inputs or disturbances acting on the system
problem. Section IV is devoted to the constructive proof of the main
have been recently considered in the design of functional observers
result and the design procedure of the LUIFO. From the use of succes-
[14]–[18] and latest developments can be found in [19] and [20]. Min-
sive derivatives of state functionals, an usual realization method leads
imality of the order of these observers is yet an open problem. Indeed,
to obtain the observer. It can be remarked that the presented results
it depends on the wanted objective with respect to the desired eigenval-
extend the results in [28] in the case where unknown inputs are present.
ues of the observer. It is well known that the minimum order required
Finally, an example illustrates the procedure in Section V and points
for a functional observer cannot be lower than the number of func-
out the advantage of the proposed design regarding the previous ones.
tions that have to be estimated [11]. In fact, there might exist some
Moreover, this example offers the opportunity to illustrate the notion
of degrees of freedom to, eventually, place some of the eigenvalues of
Manuscript received December 24, 2018; accepted May 18, 2019. the observer.
Date of publication June 6, 2019; date of current version January 28, In all the following, the following notations are used. M (m × n)
2020. This work was supported by the framework of ANR-CAPTIF is a constant matrix with m rows and n columns. In is the iden-
project (ANR-14-CE05-0044). Recommended by Associate Editor L. tity matrix of size n. 0(m × n) is a null matrix with m rows and n
Menini. (Corresponding author: Baptiste Trajin.)
columns and, shortly, 0n is the square null matrix of size n. M [1 ] is
The authors are with the Laboratoire Génie de Production (LGP),
Université de Toulouse, INP-ENIT, 65000 Tarbes, France (e-mail:, any of the generalized inverses of the matrix M , it fulfills the matrix
[email protected]; [email protected]; [email protected]). equation M XM = M , and, M † is the pseudoinverse, namely it is
Digital Object Identifier 10.1109/TAC.2019.2921435 the unique generalized inverse of M that fulfills the Moore–Penrose

0018-9286 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Beijing University of Chemical Technology. Downloaded on December 09,2024 at 04:19:51 UTC from IEEE Xplore. Restrictions apply.
 832 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 65, NO. 2, FEBRUARY 2020

set of equations (XM X = X, M XM = M , (XM ) = XM , and such that

(M X) = M X) [29]. A matrix is a Hurwitz matrix if all its eigen-
values have strictly negative real part. Finally, f (i ) denotes the ith F T + HC − T A = 0 (4)
derivative of the function f , N is the set of natural integers including 0, L − PT − V C = 0 (5)
N ∗ is the set of natural integers excluding 0, and [k1 ; k2 ] is the closed
set of natural integers from k1 to k2 . G − TB = 0 (6)
T D = 0. (7)
II. LINEAR UNKNOWN INPUT FUNCTIONAL OBSERVERS
Proof: The proof of the sufficiency is well known and is summed up
Let us consider a system described by the linear state space equations below for completeness. Let us denote the errors (t) = z(t) − T x(t)
with unknown inputs and e(t) = v(t) − v̂(t). We get from (1) and (3)
⎧
⎨ ẋ(t) = Ax(t) + Bu(t) + Dd(t) ˙ = F (t) + (G − T B)u(t) + (F T + HC − T A)x(t) − T Dd(t).
(t)
(1) (8)
⎩
y(t) = Cx(t)
Thus, when (4), (6), and (7) are fulfilled and F is a Hurwitz matrix we
where ∀t ∈ R+ , x(t) is the n-dimensional state vector, u(t) is a p- have limt →+ ∞ (t) = 0. The estimation error e(t) = v(t) − v̂(t), can
dimensional control vector supposed to be measured, y(t) is a m- be written using (2) and (3)
dimensional measured output vector, and d(t) is a r-dimensional un-
known input vector. A(n × n), B(n × p), C(m × n), and D(n × r) e(t) = (L − P T − V C)x(t) − P (t).
are constant matrices. Without loss of generality C and D are, respec-
tively, of full row and of full column ranks. Thus, when L − P T − V C = 0, we obtain limt →+ ∞ e(t) = 0.
The objective of this paper is to propose a simple design of a linear We insist afterwards on the proof of the necessity. Note that the
observer to estimate the vector v(t) related to the state x(t) by observer must be observable. Conversely, when (3) is a LUIFO of (2) for
the system (1), we have from linearity, ∀i ∈ N ∗ , limt →+ ∞ e(i ) (t) = 0.
v(t) = Lx(t) (2) With M = L − V C, when t → +∞, the following relationships are
deduced:
where L(l × n) is a constant full row rank matrix.
To design a functional observer, the triplet (A, C, L) has to be func- lim e(t) = lim (M x(t) − P z(t)) = 0
t →+ ∞ t →+ ∞
tionally observable [28], i.e.,
  lim ė(t) = lim ((M A − P HC) x(t)
t →+ ∞ t →+ ∞
OA , C , n
rank = rank (OA , C , n ) . + (M B − P G) u(t)
L
+ M Dd(t) − P F z(t)) = 0.
Moreover, in order to avoid a trivial algebraic part in the observer
(where observed functionals can be estimated through linear combi- As the last relation must be fulfilled for every u(t) and d(t), we
nation of measured outputs), it is supposed without loss of generality get M B − P G = 0 and M D = 0. Consequently, limt →+ ∞ ė(t) =
that (M A − P HC) x(t) − P F z(t) = 0. Continuing this procedure with
  the successive derivative of e(t) we get, for i ∈ [[0 ; ν − 1]]
C
rank = m + l.
L lim e(i ) (t) = lim Mi x(t) − P F i z(t) = 0 (9)
t →+ ∞ t →+ ∞

The observation of v(t) can be carried out by a LUIFO, which is a where the matrices Mi are recursively defined as M0 = M and
Luenberger observer described by the following state space equation
[1] and [2]: Mi + 1 = Mi A − P F i HC.
⎧
⎨ ż(t) = F z(t) + Gu(t) + Hy(t) Moreover, we have also Mi B = P F i G and Mi D = 0. Gathering the
(3) relationships (9) we get, for i ∈ [[0 ; ν − 1]]
⎩
v̂(t) = P z(t) + V y(t) ⎛⎡ ⎤ ⎡ ⎤ ⎞
M0 P
where ∀t ∈ R+ , z(t), and v̂(t) are, respectively, the ν-dimensional ⎜⎢ ⎥ ⎢ ⎥ ⎟
⎜⎢ M1 ⎥ ⎢ PF ⎥ ⎟
state vector and the l-dimensional output vector of the observer. The ⎜⎢ ⎥ ⎢ ⎥ ⎟
lim ⎜⎢ . ⎥ x(t) − ⎢ .. ⎥ z(t)⎟ = 0.
constant matrices F (ν × ν), G(ν × ν), H(ν × m), P (l × ν), V (l × t →+ ∞ ⎜ ⎢ . ⎥ ⎢ ⎥ ⎟
⎝⎣ . ⎦ ⎣ . ⎦ ⎠
m), and the order ν have to be determined such that limt →+ ∞ (v(t) −
v̂(t)) = 0. Moreover, it must be kept in mind that a possible minimal Mν −1 P F ν −1
order observer is looked for. It is well known that the necessary and
As the observer (3) is supposed to be observable,
sufficient conditions for the existence of an asymptotic observer (3) are
given by the following result [10], [11]. Furthermore, a LUIFO cannot ⎡ ⎤
P
be designed if there are unstable transmission zeros from the unknown ⎢ ⎥
⎢ .. ⎥
input to the output [17]. ⎢ . ⎥
⎣ ⎦
Theorem 1: The observable system (3) is a LUIFO of (2) for the ν −1
system (1) if and only if F is Hurwitz and there exists a matrix T (ν × n) PF

Authorized licensed use limited to: Beijing University of Chemical Technology. Downloaded on December 09,2024 at 04:19:51 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 65, NO. 2, FEBRUARY 2020 833

is a full column rank matrix. Thus, there exists a unique matrix T , Remark 4: Equation (4) is the so-called “constrained Sylvester
written equation” of the observer problem. Let us notice that it is a
⎡ ⎤[1 ] ⎡ ⎤ nonlinear equation to solve due to the fact that F , T , and H are un-
P M0 known. Nevertheless, until [23], [26], a great number of design methods
⎢ ⎥ ⎢ ⎥
⎢ PF ⎥ ⎢ M1 ⎥ have been devoted to the determination of the matrix T from canonical
⎢ ⎥ ⎢ ⎥
T =⎢ .. ⎥ ⎢ . ⎥ form or by imposing F . Let us mention that our procedure does not need
⎢ ⎥ ⎢ . ⎥
⎣ . ⎦ ⎣ . ⎦ to determine T as a prerequisite of the method. For the completeness
of proof, T will be given afterwards.
P F ν −1 Mq −1

such that limt →+ ∞ (T x(t) − z(t)) = 0. On the one hand, this property
III. MAIN RESULT
must be fulfilled for every u(t), x(t), and d(t). Thus, from (8), we get
the relationships (4), (6), and (7) and the fact that F is a Hurwitz This section is devoted to our claim. First, let us define recursively
matrix. On the other hand, limt →+ ∞ e(t) = 0 for every x(t) leads the matrices Kq and Σq , q ∈ N, which are as follows:

to (5).  1) K0 = In and for q ≥ 1, Kq = AKq −1 D ;
Remark 2: From Darouach work [23], [26], several works (see for 2) Σ0 = C and for q ≥ 1
instance [11]) have studied or used a particular case of Luenberger ob- ⎡ ⎤
servers (3) where P = Iν , which in the following are named Darouach Σq −1
⎢ 0(q(m + l) × r) ⎥
observers (10), are observable. In the case of unknown inputs Σq = ⎣ LKq −1 ⎦.
⎧ CKq
⎨ ż(t) = F z(t) + Gu(t) + Hy(t)
(10)
⎩ More explicitly, we get
v̂(t) = z(t) + V y(t)
⎡ ⎤
is an asymptotic observer of (2) for the⎛system
⎡ (1) if and
⎤⎞only if C 0 ··· 0 0 0
⎛⎡ ⎤⎞ C 0 ⎢ ⎥
0 ⎢ L 0 ··· 0 0 0 ⎥
C ⎜⎢ ⎥⎟ ⎢ ⎥
⎜⎢ ⎥⎟ ⎜⎢ L 0 ⎥⎟ ⎢ ⎥
1) rank ⎝⎣ L 0 ⎦⎠ = rank ⎜⎢ ⎜ ⎢ ⎥⎟ . ⎢ CA CD ··· 0 0 0 ⎥
⎥⎟ ⎢ ⎥
⎝⎣ CA CD ⎦⎠ ⎢ LA ⎥
CA CD ⎢ LD ··· 0 0 0 ⎥
LA LD ⎢ ⎥
⎢ .. .. .. .. .. ⎥
2) ∀s, such as (s) ≥ 0 ⎢ ⎥
Σq = ⎢ . . . . . ⎥
⎢ ⎥
⎛⎡ ⎤⎞ ⎢ .. .. ⎥
C 0 ⎢ ⎥
⎢ . . 0 0 0 ⎥
⎜⎢ ⎥⎟ ⎢ ⎥
rank ⎝⎣ CA CD ⎦⎠ ⎢ CAq −1 q −2 ⎥
⎢ CA D ··· CAD CD 0 ⎥
sL − LA −LD ⎢ ⎥
⎛⎡ ⎤⎞ ⎢ LAq −1 LA q −2
D ··· LAD LD 0 ⎥
C 0 ⎣ ⎦
⎜⎢ ⎥⎟ CAq CAq −1 D ··· CA2 D CAD CD
= rank ⎝⎣ L 0 ⎦⎠ .
CA CD where the “0” blocks are of adapted dimensions.
In the following, we use the notation:
On the one hand, when fulfilled, the first condition ensures the ex-
istence of the Darouach structure (10), i.e., (4)–(7) have a solution ⎡ ⎤
Cq , 0
considering P = Iν . On the second hand, the second condition leads
⎢ L ⎥
to the asymptotic tracking of the functional v(t). The second condi- ⎢ q, 0 ⎥
⎢ ⎥
tion, which is a Hautus type condition [30], can be only used for linear ⎢ Cq , 1 ⎥
⎢ ⎥
time-invariant systems. In the case of D = 0, expressed in [24], the ⎢ ⎥
⎢ Lq , 1 ⎥
second condition has the following form: ⎢ ⎥
Σq = ⎢
⎢ .. ⎥
⎥ (11)
LA = Γ0 C + Λ0 L + Γ1 CA. ⎢ . ⎥
⎢ ⎥
⎢ ⎥
⎢ Cq , q −1 ⎥
The system (10) is an observer of (2) if and only if Λ0 is a (l × l) ⎢ ⎥
⎢ ⎥
Hurwitz matrix. This standpoint has been extended to linear time- ⎣ Lq , q −1 ⎦
varying systems in [31] and [32]. In the following, this criterion is Cq , q
extended for the unknown input case.
Remark 3: Darouach’s method [23] has been used in the majority where the matrices Cq , i , for i ∈ [[0; q]], and Lq , i , for i ∈ [[0; q − 1]],
of the papers dealing with the design of unknown input functional are, respectively, of dimensions (m × (n + rq)) and (l × (n + rq)).
observers. Indeed, a lot of design procedures are based on the extension Theorem 5: There exists an observable unknown input functional
of the functional to observe. Namely, these methods consist in looking observer (3) for the system (1) to asymptotically observe the functional
for a R(ρ × n) matrix with ν = l + ρ and P = Iν such that (3) is a (2), if there exist matrices Γi for i ∈ [[0; q]], and Λi for i ∈ [[0; q − 1]],
Darouach observer of v(t) = [ RL ]. Nevertheless, as it is underlined in such that
[11], to find R is an “intriguing and challenging problem” and some
attempts have been proposed in [19], [20], [22], and[33]–[35]. As it 
q q −1

will be explained in the following section, our design principle differs LKq = Γi Cq , i + Λi Lq , i (12)
a lot from this standpoint. i= 0 i= 0

Authorized licensed use limited to: Beijing University of Chemical Technology. Downloaded on December 09,2024 at 04:19:51 UTC from IEEE Xplore. Restrictions apply.
 834 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 65, NO. 2, FEBRUARY 2020

and the matrix (12) that can then be explicitly written as

⎡ ⎤
0l ··· ··· 0l Λ0 q −1
⎢ ⎥ 
q

⎢ .. .. ⎥ LAq = Γi CAi + Λi LAi
⎢ Il . . Λ1 ⎥
⎢ ⎥ i= 0 i= 0
⎢ ⎥
⎢ .. .. .. .. ⎥
⎢ 0l . . . . ⎥ (13) 
q q −1

⎢ ⎥ LAq −1 D = Γi CAi −1 D + Λi LAi −1 D
⎢ ⎥
⎢ . .. .. ⎥
⎢ .. . . 0l Λq −2 ⎥
i= 1 i= 1
⎣ ⎦
..
0l ··· 0l Il Λq −1 .
q −1
is a Hurwitz one. 
q

The proof of this result is constructive and leads to the design of the LAq −k D = Γi CAi −k D + Λi LAi −k D
LUIFO. Thus, it will be detailed in the following section. i= k i= k

Remark 6: On the one hand (12), that can be written as ..

.
 
Σq LAD = Γq CAD + Γq −1 CD + Λq −1 LD
rank Σq = rank (14)
LKq
LD = Γq CD. (15)
leads to the design of the state space equations of the observer (3).
Indeed, as indicated in the Remark 10 of Section IV the knowledge of Second, the design of the observer can be considered only when
matrices F , G H, P , and V may lead to determine a matrix T that the matrix (13) is an Hurwitz matrix, that can be tested from the pre-
solves (4)–(7). On the other hand, the Hurwitz condition ensures that vious decomposition for LKq on Σq . So, when the conditions of the
limt →+ ∞ (v(t) − v̂(t)) = 0. Thus, the previous theorem appears as an Theorem 5 are satisfied, we can go further to the determination of the
extension of the reformulated Darouach criterion for LUIFO. matrices in (3).
Remark 7: We insist here that the observable condition of (3) en-
sures the necessity in the Theorem 5. The proposed design procedure
leads to an observable observer. B. Second Step
In the following section, the sufficient condition for the existence
The design of the observer uses the successive derivations of v(t).
of a minimum order UIFO is used in order to develop a method for
After q derivations of v(t) = Lx(t), we obtain
designing a LUIFO. When the obtained observer structure does not
lead to an asymptotically stable observer, the procedure will consist in
increasing the integer q. This standpoint is detailed in the example in v̇(t) = LAx(t) + LBu(t) + LDd(t)
the Section V.
v (2 ) (t) = LA2 x(t) + LABu(t) + LB u̇(t) + LADd(t)
˙
+ LD d(t)
IV. DESIGN OF AN UNKNOWN INPUT FUNCTIONAL OBSERVER
This section deals with the design of an observable LUIFO for the ..
.
observation problem defined by (1) and (2) when condition (12) is
fulfilled. Moreover, the procedure is given to get a possible stable q −2

observer of least order. v (q −1 ) (t) = LAq −1 x(t) + LAi Bu(q −2 −i ) (t)
i= 0

q −2
A. First Step 
+ LAi Dd(q −2 −i ) (t)
This step consists in the determination of the integer q candidate for i= 0
the structure (3). First, q is given by the condition (14), which must be q −1
fulfilled. Let us suppose that the relation (14) is fulfilled for the smallest 
v (q ) (t) = LAq x(t) + LAi Bu(q −i −1 ) (t)
q. Then, the matrices Γi for i ∈ [[0; q]] and Λi for i ∈ [[0; q − 1]] are i= 0
obtained from the solution of the consistent linear system LKq = XΣq
q −1

X= LKq Σ[1 ]
+ Z Ir q − Σq Σ[1 ] + LAi Dd(q −i −1 ) (t). (16)
q q
i= 0

where rq = m(q + 1) + lq is the number of rows of Σq , Z is any

[1 ]
matrix with the same size than X, and, Σq is any of the generalized From the last equation of (16), LAk x(t) is expressed, for q ∈ N ∗ as
inverses of Σq . Due to numerical properties and implementations in
[1 ]
numerical software, we can chose as a particular case Σq = Σ†q . In k
−1

the case where Σq is of full row rank, the solution is unique and, conse- LAk x(t) = v (k ) (t) − LAi Bu(k −i −1 ) (t)
i= 0
quently, the matrices Γi and Λi also. Partitioning X with respect to the 
k −1 (17)
row partition of Σq in (11) (X = [Γ0 Λ0 Γ1 Λ1 . . . Γq −1 Λq −1 Γq ]), − LAi Dd(k −i −1 ) (t).
we are led to the matrices Γi for i ∈ [[0; q]] and Λi for i ∈ [[0; q − 1]] in i= 0

Authorized licensed use limited to: Beijing University of Chemical Technology. Downloaded on December 09,2024 at 04:19:51 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 65, NO. 2, FEBRUARY 2020 835

Moreover, the last equation of (16) can also be written, using the in case of detectable systems, by increasing step by step the integer q,
first equation of (15), as we are led, at least, to the well-known reduced order state observer.
q −1 Remark 9: When Σq is of full row rank, the eigenvalues of the

q

v (q ) (t) = Γi CAi x(t) + Λi LAi x(t) matrix F are fixed. In the opposite, when rank (Σq ) = ρq < rq , rq −
i= 0 i= 0
ρq degrees of freedom can be defined to place some of the eigenvalues
q −1 q −1 (18)
  of F . It leads to the possibility of increasing the integer q step by step
(q −i −1 ) (q −i −1 )
+ i
LA Bu (t) + i
LA Dd (t). to choose the decay rate of the designed observer.
i= 0 i= 0 Remark 10: From the expressions (20) for the matrices Φi and the
In order to eliminate x(t) and d(t) from (18) we use the derivative relationship G = T B (6), it can be deduced, as it has been indicated
of v(t) and y(t). Indeed, from y(t) = Cx(t) we get, for k ∈ N ∗ for LFO in [28], that
⎡ ⎤

k −1 T1
CAk x(t) = y (k ) (t) − CAi Bu(k −1 −i ) (t) ⎢ ⎥
⎢ T2 ⎥
i= 0 ⎢ ⎥

k −1 ⎢ . ⎥
− CAi Dd(k −1 −i ) (t). T =⎢ .
⎢ . ⎥
⎥
⎢ ⎥
i= 0 ⎢ ⎥
⎣ Tq −1 ⎦
Moreover, taking into account the expression of the matrices LAi D Tq
from (15) and the terms LAi x(t) from (17), for i ∈ [[0; q − 1]], it leads
to where for j ∈ [[0; q − 1]]
q −1 q −1

q
  q −1
v (q ) (t) = Γi y (i ) (t) + Λi v (i ) (t) + Φi u(i ) (t) (19)  
q
Tj = LAq −j − Λi LAi −j − Γi CAi −j
i= 0 i= 0 i= 0
i= j i= j
where for i ∈ [[0; q − 2]]
 and Tq = L − Γq C. From (15), it is easy to verify that T D = 0.

q
q −1 −i j −i −1
Φi = LA − Γj CA
j = i+ 1 V. ILLUSTRATIVE EXAMPLE
q −1
 (20)
 According to [19], let us consider the system (1) and the single
− Λj LAj −i −1 B
j = i+ 1
functional (3) defined by
⎡ ⎤
and Φq −1 = [L − Γq C] B. −2.51 0.33 0.68 1.12 −0.25
A realization of the input–output differential equation (19) [24], [28], ⎢ ⎥
⎢ 0.14 −0.23 −0.31 0.91 0.36 ⎥
[36] leads to the observer ⎢ ⎥
⎢ ⎥
⎧ ⎡ ⎤ A = ⎢ 0.51 −1.18 0.41 0.63 −0.77 ⎥
⎪ Φ0 ⎢ ⎥
⎪
⎪ ⎢ 0.22 0.33 0.46 0.65 −0.77 ⎥
⎪
⎪ ⎢ ⎥ ⎣ ⎦
⎪
⎪ ⎢ Φ1 ⎥
⎪ ⎢ ⎥
⎪
⎪
⎪ ż(t) = F z(t) + ⎢ . ⎥ u(t) ⎡ 0.23 ⎤ 0.33 ⎡ 3.97 0.06 ⎤ 0.69
⎪
⎪ ⎢ . ⎥ 0.43 1.00 0.00
⎪
⎪ ⎣ . ⎦ ⎢ ⎥ ⎢ ⎥
⎪
⎪ ⎢ 0.00 ⎥ ⎢ −3.00 −1.00 ⎥
⎪
⎪ ⎢ ⎥ ⎢ ⎥
(22)
⎨ Φq −1
⎡ ⎤ ⎢ ⎥ ⎢ ⎥
Γ0 + Λ0 Γq (21) B = ⎢ 0.92 ⎥ , D = ⎢ 0.00 0.50 ⎥
⎪
⎪ ⎢ ⎥ ⎢ ⎥
⎪
⎪ ⎢ ⎥ ⎢ 1.20 ⎥ ⎢ 0.45 0.00 ⎥
⎪
⎪ ⎢ Γ1 + Λ1 Γq ⎥ ⎣ ⎦ ⎣ ⎦
⎪
⎪ ⎢ ⎥
⎪
⎪ + ⎢ .. ⎥ y(t) −1.27 0.00 0.00
⎪
⎪ ⎢ ⎥  
⎪
⎪ ⎣ . ⎦ 1.00 0.00 0.00 0.00 0.60
⎪
⎪ C=
⎪
⎪ Γ + Λ Γ
⎪
 0.00 0.00 0.00 1.00 0.00
−1 −1
⎩ q q q
v̂(t) = [ 0l · · · 0l Il ]z(t) + Γq y(t) L = 2.00 0.00 0.00 9.00 0.30 .
with
⎡ ⎤ In [19], it is remarked that the system matrix has two detectable invari-
0l ··· ··· 0l Λ0 ant zeros at −2.16 ± 2.02i. Consequently, they will appear as poles of
⎢ ⎥
⎢ .. .. ⎥ the observer. The following cases illustrate our proposed design of the
⎢ Il . . Λ1 ⎥
⎢ ⎥ LUIFO. In a first attempt, a minimal order observer will be obtained.
⎢ ⎥
⎢ .. .. .. .. ⎥ The second attempt illustrates the introduction of degrees of freedom
F = ⎢0 . . . . ⎥
⎢ l ⎥ to change some of the poles of the observer. From the invariant ze-
⎢ ⎥
⎢ . .. .. ⎥ ros, it is assumed that the minimal order of an observer is two. Let us
⎢ .. . . 0l Λq −2 ⎥
⎣ ⎦ mention here that in [19], a third-order observer whose eigenvalues are
0l ··· 0l Il Λq −1 {−5, −2.16 ± 2.02i} has been designed.

and the observer design is complete.

Remark 8: When the Hurwitz condition is satisfied, it is demon- A. Design of a Minimal-Order Observer
strated that (21) is an asymptotic observer of the linear functional Σ1
As rank(Σ1 ) = 5 and rank([ ]) = 6, as expected, a first-
Lx(t). Otherwise, it becomes necessary to increase the integer q and to LA LD
do again the design procedure with a higher order [13], [37]. Note that order minimum observer cannot be designed.

Authorized licensed use limited to: Beijing University of Chemical Technology. Downloaded on December 09,2024 at 04:19:51 UTC from IEEE Xplore. Restrictions apply.
 836 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 65, NO. 2, FEBRUARY 2020

We have
⎡
1.00 0.00 0.00 0.00 0.60
⎢
⎢ 0.00 0.00 0.00 1.00 0.00
⎢
⎢ 2.00 0.00 0.00 9.00 0.30
⎢
⎢
⎢ −2.37 0.53 3.06 1.16 0.16
⎢
Σ2 = ⎢
⎢ 0.22 0.33 0.46 0.65 −0.77
⎢
⎢
⎢ −2.97 3.73 6.69 8.11 −7.22
⎢
⎢
⎣ 7.88 −4.08 0.66 0.51 −2.35
−0.31 −0.59 −2.52 1.21 −1.32
⎤
0.00 0.00 0.00 0.00
⎥
0.00 0.00 0.00 0.00 ⎥
⎥
0.00 0.00 0.00 0.00 ⎥
Fig. 1. Simulation results for the second and third-order observers.
⎥
⎥
1.00 0.00 0.00 0.00 ⎥
⎥
⎥
0.45 0.00 0.00 0.00 ⎥
⎥ B. Design of a Third-Order Observer
⎥
6.05 0.00 0.00 0.00 ⎥ In this section, the design of a LUIFO of a third-order observer with
⎥
⎥ q = 3 is detailed. The appearance of a degree of freedom in the design
−3.44 1.00 1.00 0.00 ⎦
procedure is here illustrated.
−0.48 −0.10 0.45 0.00 The size of Σ3 is (11, 11) and we obtain rank(Σ3 ) = rank([ LΣK3 3 ]) =
and 10. Consequently, there exist ten rows in Σ3 that are independent. Let
 us consider Σ∗3 , where the component L3 , 2 of Σ3 has been eliminated
LK2 = 11.51 −9.44 −25.38 9.12 −14.29
 ⎡ ⎤
C 0 0 0
−10.51 −0.38 6.05 0.00 .
⎢ L 0 0 0 ⎥
⎢ ⎥
We get rank(Σ2 ) = 8 and rank([ LΣA22 ]) = 8. Thus, q = 2 is the small- ⎢ ⎥
⎢ CA CD 0 0 ⎥
est integer that has to be considered. Consequently, as l = 1, a mini- ∗ ⎢ ⎥
Σ3 = ⎢ ⎥.
⎢ LA LD 0 0 ⎥
mal second-order observer can be considered as a candidate observer. ⎢ ⎥
From the unique solution X = LK2 Σ†2 , we obtain Λ0 = −8.77 and ⎢ ⎥
⎣ CA2 CAD CD 0 ⎦
Λ1 = −4.32. That yields to
  CA3 CA2 D CAD CD
0 −8.77
F = . Thus, Σ∗3 has 10 rows and rank(Σ∗3 ) = 10. On the one hand, (12) can
1 −4.32 be written as
 
  Σ∗3
As expected the eigenvalues of F are (−2.16 ± 2.02i). LK3 = Γ0 Λ0 Γ1 Λ1 Γ2 Γ3 Λ2 (23)
Moreover, to design the second-order observer we get, from X = L3 , 2
LK2 Σ†2 that is nonunique. As the row L3 , 2 has been eliminated from Σ3 , Λ2
  appears as a degree of freedom. On the other hand, we get
Γ0 = 13.24 75.14
 
  LK3 = Γ3 0 Λ3 0 Γ3 1 Λ3 1 Γ3 2 Γ3 3 Σ∗3 (24)
Γ1 = 3.96 44.38   ∗
L3 , 2 = Π0 Δ0 Π1 Δ1 Π2 Π3 Σ3 (25)
 
Γ2 = 0.78 11.70 . where the matrices Γi j , Λi j , Πi , and Δi are unique. We have, from
Using (21), the following matrices of the observer are deduced: (23) and (25)
 
    LK3 = Γ0 Λ0 Γ1 Λ1 Γ2 Γ3 Σ∗3
−8.25 6.36 −27.55   (26)
G= ,H= + Λ2 Π0 Δ0 Π1 Δ1 Π2 Π3 Σ∗3 .
−2.50 0.57 −6.21
    Consequently, identifying (24) and (26), it yields
P = 0 1 , V = 0.78 11.70 . ⎧
⎪
⎪ Γ0 = Γ3 0 − Λ2 Π0
The observer design is complete. A simulation result with different ⎪
⎪
⎪
⎪
initial conditions is displayed in Fig. 1. Initial conditions of the system ⎪
⎪ Λ0 = Λ3 0 − Λ2 Δ0
⎪
⎪
(resp. the observer) are set to x(t = 0) = [0 0 0 0 0] (resp. z(t = 0) = ⎪
⎪
⎪
⎪
[500 200] ). Identical inputs than those in [19] are applied for t ≥ 0: ⎪
⎨ Γ1 = Γ3 1 − Λ2 Π1
u(t) = 0.2 + e−0 . 4 t cos (2t), d1 (t) = 0.1 + 0.2e−0 . 1 sin (t ) tanh(2t), (27)
⎪
⎪
and d2 (t) = 2. It can be seen that the observer output converges to ⎪
⎪ Λ1 = Λ3 1 − Λ2 Δ1
⎪
⎪
v(t) = Lx(t) with the decay rate given by the eigenvalues of F . It ⎪
⎪
⎪
⎪ Γ2 = Γ3 2 − Λ2 Π2
can be noticed that in [19], it is found that the order of unknown input ⎪
⎪
⎪
⎪
functional observer is q = 3. With our procedure we have obtained a ⎪
⎩
minimal second-order LUIFO. Γ3 = Γ3 3 − Λ2 Π3

Authorized licensed use limited to: Beijing University of Chemical Technology. Downloaded on December 09,2024 at 04:19:51 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 65, NO. 2, FEBRUARY 2020 837

Due to invariant zeros, Λ2 can place only one eigenvalue for F . In

order to compare with [19] where the roots of pF (λ) are (−2.16 ±
2.02i, −5), we take Λ2 = −9.32, which leads to Λ1 = −30.34 and
Λ0 = −43.78. For these values, we get the third-order Luenberger
observer defined by
⎡ ⎤ ⎡ ⎤
0 0 −43.78 −41.77
F = ⎣ 1 0 −30.34 ⎦ , G = ⎣ −26.04 ⎦
0 1 −9.32 −2.50
⎡ ⎤
31.81 −137.79
H = ⎣ 9.21 −58.64 ⎦
0.57 −6.22
   
P = 0 0 1 , V = 0.78 11.70 .
Simulation results are displayed in Fig. 1 where the second and third
Fig. 2. Simulation results of the estimation error for the second and order observer outputs are compared to the functional to be observed.
third-order observers. Inputs and initial conditions of the system are identical for the two
simulations. Initial conditions of the third-order observer are set to
z(t = 0) = [500 − 300 200] .
As shown in Fig. 2, the tracking property of the observers is ensured
by the proposed design methods.
As it can be seen in Fig. 3, the value of the parameter Λ2 influ-
ences the decay rate of the error. A comparison is performed between
third-order observer with Λ2 = −9.32 and Λ2 = −24.32 leading to
eigenvalues of the third-order observer of (−2.16 ± 2.02i, −20). As
predicted, the response time of the estimation error is reduced by choos-
ing a parameter Λ2 leading to an eigenvalue with a lower negative real
part.

VI. CONCLUSION
This paper presents a new and direct design procedure to obtain a
possible minimal unknown input functional observer for LTI systems.
Necessary and sufficient conditions for the existence of a stable LUIFO
have been detailed. The proposed procedure is based on linear algebraic
Fig. 3. Simulation results of the estimation error for the third-order operations in a state space setting and only needs to solve linear equa-
observers.
tions. The order of the obtained observer can be increased with some
degrees of freedom to choose the dynamic response of the observed
functional. Finally, the proposed procedure could be extended to linear
where Λ2 is arbitrary. In this example, we get time-varying systems.
  We demonstrated, in this paper, that for the observation of a multi-
Γ3 0 = −57.28 −324.98 , Λ3 0 = 37.95 functional Lx(t), q is the smallest integer that fulfils the Theorem 5.
  From the proposed design procedure, we can assert that the minimal
Γ3 1 = −3.88 −116.79 , Λ3 1 = 9.92
    order ν of the observer is such that ν ≤ ql. Nevertheless, to obtain a
Γ3 2 = 0.57 −6.21 , Γ3 3 = 0.78 11.70 minimal stable observer and determining ν as small as possible is a
  challenging problem largely beyond the scope of the proposed paper.
Π0 = 13.24 75.14 , Δ0 = −8.77 This specific point will be the topics of our future works.
 
Π1 = 3.96 44.38 , Δ1 = −4.32
    REFERENCES
Π2 = 0.78 11.70 , Π3 = 0.00 0.00 .
[1] D. Luenberger, “Observers for multivariable systems,” IEEE Trans. Au-
Then, the matrix F expressed as tom. Control, vol. 11, no. 2, pp. 190–197, Apr. 1966.
[2] D. Luenberger, “An introduction to observers,” IEEE Trans. Autom. Con-
⎡ ⎤ trol, vol. 16, no. 6, pp. 596–602, Dec. 1971.
0 0 Λ3 0 − Λ2 Δ0
[3] C. Edwards, “A comparison of sliding mode and unknown input observers
F = ⎣ 1 0 Λ3 1 − Λ2 Δ1 ⎦ . for fault reconstruction,” in Proc. 43rd IEEE Conf. Decis. Control, 2004,
0 1 Λ2 vol. 5, pp. 5279–5284.
[4] J. B. Moore and G. F. Ledwich, “Minimal order observers for estimating
The eigenvalues of the matrix F are the roots of the characteristic linear functions of a state vector,” IEEE Trans. Autom. Control, vol. 20,
polynomial pF (λ) = λ3 − Λ2 λ2 − Λ1 λ − Λ0 , which depends on the no. 5, pp. 623–632, Oct. 1975.
parameter Λ2 . Here we obtain, with Λ0 = 37.95 + 8.77Λ2 and Λ1 = [5] F. W. Fairman and R. D. Gupta, “Design of multi-functional reduced order
observers,” Int. J. Syst. Sci., vol. 11, no. 9, pp. 1083–1094, 1980.
9.92 + 4.32Λ2 [6] P. A. Fuhrmann and U. Hemke, “On the parametrization of conditioned
invariant subspaces and observer theory,” Linear Algebr. Appl., vol. 332,
pF (λ) = (λ2 + 4.32λ + 8.77)(λ − 4.32 − Λ2 ). pp. 265–353, 2001.

Authorized licensed use limited to: Beijing University of Chemical Technology. Downloaded on December 09,2024 at 04:19:51 UTC from IEEE Xplore. Restrictions apply.
 838 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 65, NO. 2, FEBRUARY 2020

[7] J. Trumph, “Observers for linear time-varying systems,” Linear Algebra [23] M. Darouach, “Functional observers for systems with unknown inputs,”
Appl., vol. 425, no. 2/3, pp. 303–312, 2007. in Proc. 16th Int. Symp. Math. Theory Netw. Syst., Leuven, Belgium, Jul.
[8] L. S. Jennings, T. L. Fernando, and H. M. Trinh, “Existence conditions 5–9, 2004.
for functional observability from an eigenspace perspective,” IEEE Trans. [24] F. Rotella and I. Zambettakis, “Minimal single linear functional observers
Autom. Control, vol. 56, no. 12, pp. 2957–2961, Dec. 2011. for linear systems,” Automatica, vol. 47, pp. 164–169, 2011.
[9] M. Lungu and R. Lungu, “Design of full-order observers for systems with [25] F. Rotella and I. Zambettakis, “A direct design procedure for linear state
unknown inputs by using the eigenstructure assignment,” Asian J. Control, functional observers,” Automatica, vol. 70, pp. 211–216, 2016.
vol. 16 , no. 5, pp. 1470–1481, 2014. [26] M. Darouach, “Existence and design of functional observers for linear
[10] J. O’Reilly, Observers for Linear Systems. New York, NY, USA: Aca- systems,” IEEE Trans. Autom. Control, vol. 45, no. 5, pp. 940–943, May
demic, 1983. 2000.
[11] H. Trinh and T. Fernando, Functional Observers for Dynamical Systems. [27] T. Fortmann and D. Williamson, “Design of low-order observers for lin-
Berlin, Germany: Springer Science and Business Media, 2012. ear feedback control laws,” IEEE Trans. Autom. Control, vol. 17, no. 3,
[12] H. R. Sirisena, “Minimal-order observers for linear functions of a state pp. 301–308, Jun. 1972.
vector,” Int. J. Control, vol. 29, no. 2, pp. 235–254, 1979. [28] F. Rotella and I. Zambettakis, “A note on functional observability,” IEEE
[13] F. Rotella and I. Zambettakis, “Some new standpoints in the design of Trans. Autom. Control, vol. 61, no. 10, pp. 3197–3202, Oct. 2016.
asymptotic functional linear observers,” ESAIM: Proc. Surv., vol. 49, [29] A. Ben-Israel and T. N. Greville, Generalized Inverses: Theory and Ap-
pp. 102–114, 2015. plications. Berlin, Germany: Springer Science and Business Media, 2003.
[14] P. M. Frank, “Fault diagnosis in dynamic systems using analytical and [30] M. L. J. Hautus, “Strong detectability and observers,” Linear Algebra
knowledge-based redundancy: A survey and some new results,” Automat- Appl., vol. 50, pp. 353–368, 1983.
ica, vol. 26, no. 3, pp. 459–474, 1990. [31] F. Rotella and I. Zambettakis, “On functional observers for linear time-
[15] M. Hou and P. C. Mller, “Design of observers for linear systems with varying systems,” IEEE Trans. Autom. Control, vol. 58, no. 5, pp. 1354–
unknown inputs,” IEEE Trans. Autom. Control, vol. 37, no. 6, pp. 871– 1360, May 2013.
875, Jun. 1992. [32] L. T. Li and G. R. Duan, “Observer design for a class of linear time-varying
[16] M. Hou and P. C. Mller, “Design of decentralized linear state function systems,” in Proc. 36th IEEE Chin. Control Conf., 2017, pp. 116–121.
observers,” Automatica, vol. 30, no. 11, pp. 1801–1805, 1994. [33] H. Trinh, T. D. Tran, and T. Fernando, “Disturbance decoupled observers
[17] M. Hou, A. C. Pugh, and P. C. Mller, “Disturbance decoupled functional for systems with unknown inputs,” IEEE Trans. Autom. Control, vol. 53,
observers,” IEEE Trans. Autom. Control, vol. 44, no. 2, pp. 382–386, Feb. no. 10, pp. 2397–2402, Nov. 2008.
1999. [34] H. Trinh and Q. Ha, “Design of linear functional observers for linear
[18] Y. Xiong and M. Saif, “Unknown disturbance inputs estimation based on systems with unknown inputs,” Int. J. Syst. Sci., vol. 31, no. 6, pp. 741–
a state functional observer design,” Automatica, vol. 39, no. 8, pp. 1389– 749, 2000.
1398, 2003. [35] R. Mohajerpoor, H. Abdi, and S. Nahavandi, “A new algorithm to design
[19] R. Mohajerpoor, H. Abdi, and S. Nahavandi, “Minimal unknown-input minimal multi-functional observers for linear systems,” Asian J. Control,
functional observers for multi-input multi-output LTI systems,” J. Process vol. 18, no. 3, pp. 842–857, 2015.
Control, vol. 35, pp. 143–153, 2015. [36] T. Kailath, Linear Systems, Englewood Cliffs, NJ, USA: Prentice-Hall,
[20] R. Mohajerpoor, H. Abdi, and S. Nahavandi, “On unknown-input func- 1980.
tional observability of linear systems,” in Proc. Amer. Control Conf., 2015, [37] H. Trinh, T. D. Tran, and S. Nahavandi, “Design of scalar functional
pp. 3534–3539. observers of order less than (ν- 1),” Int. J. Control, vol. 79, pp. 1654–
[21] T. Fernando, H. M. Trinh, and l. S. Jennings, “Functional observability and 1659, 2006.
the design of minimum order linear functional observers,” IEEE Trans.
Autom. Control, vol. 55, no. 5, pp. 1268–1273, May 2010.
[22] T. Fernando, S. MacDougall, V. Sreeram, and H. M. Trinh, “Existence
conditions for unknown input functional observers,” Int. J. Control, vol. 86,
no. 1, pp. 22–28, 2013.

Authorized licensed use limited to: Beijing University of Chemical Technology. Downloaded on December 09,2024 at 04:19:51 UTC from IEEE Xplore. Restrictions apply.