530 International Journal of Control,

Yan-Ming Automation,
Fu, Guang-Ren and and
Duan, Systems, vol. 2,Song
Shen-Min no. 4, pp. 530-535, December 2004

Design of Unknown Input Observer for Linear Time-delay Systems

Yan-Ming Fu, Guang-Ren Duan, and Shen-Min Song

Abstract: This paper deals with the unknown input observer (UIO) design problem for a class
of linear time-delay systems. A case in which the observer error can completely be decoupled
from an unknown input is treated. Necessary and sufficient conditions for the existences of
such observers are present. Based on Lyapunov stability theory, thedesign of the observer with
internal delay is formulated in terms of linear matrix inequalities (LMI). The design of the
observer without internal delay is turned into a stabilization problem in linear systems. Two
design algorithms of UIO are proposed. The effect of the proposed approach is illustrated by
two numerical examples.

Keywords: Linear matrix inequalities, Lyapunov stability, stabilization, time-delay systems,

unknown input observer.

1. INTRODUCTION design of the observer with internal delay is solved in

terms of linear matrix inequality; and the design of the
Time delay is commonly encountered in various observer without internal delay is turned into a
engineering systems, such as chemical processes, and stabilization problem in a linear system.
long transmission lines in pneumatic, hydraulic, This paper is divided into six sections. In Section 2,
economic and rolling mill systems. It usually results the problem to be solved in this paper is formulated.
in unsatisfactory performance and is frequently a Section 3 provides the observer conditions. The
source of instability, so control of time-delay systems design methods are proposed in Section 4. Numerical
is practically important, especially state feedback examples are presented in Section 5. Conclusions
control. An unknown input observer has important follow in Section 6.
applications in realization of state feedback control. A
UIO is an estimator that is decoupled from the 2. PROBLEM FORMULATION
unknown input (disturbances, faults, etc.) that may be
acting on the system. Consider the following continuous-time linear time-
Observer design theory for time-delay systems has delay system described by
been most widely considered in the last decade [1-7].
Various methods have been used in the observer  x (t ) = Ax(t ) + Ad x(t − d (t )) + Bu (t ) + Ew(t ),
design, for example, coordinate change approach [1], 
LMI method [2], reducing transformation technique  y (t ) = Cx(t ), (1)
 x(t ) = φ (t ), t ∈ − d (t ), 0 ,
[3], factorization approach [4], and polynomial  [ ]
approach [5]. However, little research has been
focused on design of UIO for linear systems with
time-varying delay. where x(t ) ∈ R n is the state vector, u (t ) ∈ R r is the
In this paper, design of UIO for linear systems with input vector, w(t ) ∈ R q is the unknown input vector,
time-varying delay is proposed. Necessary and
sufficient conditions for the existence of the observer y (t ) ∈ R m is the measured output vector, and d (t )
are derived. Based on Lyapunov stability theory, the is a known time delay satisfying
__________
Manuscript received November 27, 2003; revised October 18, 0 ≤ d (t ) ≤ d 0 , 0 ≤ d (t ) ≤ d1 < 1, (2)
2004; accepted November 2, 2004. Recommended by Editorial
Board member Min-Jea Tahk under the direction of Editor-in-
Chief Myung Jin Chung. This work was supported by the where d0 and d1 are known constant scalars. φ (t )
National Outstanding Youth Foundation of China (No. is a continuous vector-valued initial function.
69925308). A, Ad , B, C and E are real matrices with appropriate
Yan-Ming Fu, Guang-Ren Duan, and Shen-Min Song are
with the Center for Control Theory and Guidance Technology, dimensions. Without loss of generality, it can be
Harbin Institute of Technology, P.O. Box 416, Harbin, 150001, assumed that E has full column rank and C has
China. (e-mails: {fuyanming, g.r.duan}@hit.edu.cn, songsm full row rank.
[email protected]).
 Design of Unknown Input Observer for Linear Time-delay Systems 531

For the system (1), an observer that reconstructs the (Sufficiency) If the last four conditions in the theorem
state x(t ) without the knowledge of the unknown are met, then the error (6) reduces to
input w(t ) can be constructed using measured output
e(t ) = Fe(t ) + Ge(t − d ) . (7)
y (t ) and input u (t ) as
Further, in view of the first condition in the theorem,
xˆ (t ) = Fxˆ (t ) + Gxˆ (t − d (t )) + Hu (t ) + L1 y (t ) we know that e(t ) → 0 for any initial conditions and
(3)
+ L2 y (t ) + L3 y (t − d (t )), u (t ) . So xˆ (t ) is an estimate of x(t ) .
(Necessity) Let (2) be a UIO for system (1). If
with the initial state xˆ(t ) = ϕ (t ), ∀t ∈ [ − d (t ),0] , condition 1) is not satisfied, then for x(t ) = 0 ,
u (t ) = 0 and w(t ) = 0, ∀t , we have e(t ) ≠ 0 as
where xˆ (t ) ∈ R n is the state vector. F , G, H ,
t → ∞ . This contradicts with (4).
L1 , L2 and L3 are the coefficient matrices with
If conditions 2) and 3) are not satisfied, then we can
appropriate dimensions. find u (t ) such that e(t ) ≠ 0 as t → ∞ . This
With the above preparation, the problem of
contradicts with (4).
unknown input observer design to be solved in the
If conditions 4) and 5) are not satisfied, there exist
paper can be stated as follows.
u (t ) and w(t ) such that e(t ) ≠ 0 as t → ∞ . This
Problem UIO: Given the system (1) and the
observer (3), design the observer parameters contradicts with (4).
F , G , H , L1 , L2 and L3 , such that Remark 1: From the error equation (6) and
Theorem 1, we know that the condition 5) in Theorem
1 guarantees that the estimate error can completely be
lim[ xˆ (t ) − x(t )] = 0 (4) decoupled from the unknown input.
t →∞
The observer (3) is dependent of xˆ (t − d (t )) . If G = 0 ,
for any initial functions. the observer becomes the following observer without
internal delay:
3. UIO CONDITIONS OF TIME-DELAY
SYSTEMS xˆ (t ) = Fxˆ (t ) + Hu (t ) + L1 y (t ) + L2 y (t ) + L3 y (t − d (t )),
(8)
First, denote by e(t ) the error between x(t ) and
its estimate xˆ (t ) , that is, which is independent of xˆ (t − d (t )) . The initial state
is again xˆ(t ) = ϕ (t ), ∀t ∈  − d (t ), 0] . Based on Theorem
e(t ) = x(t ) − xˆ (t ) , (5)
1, we have the following corollary.
Corollary 1: The observer in the form of (8) is a
then, we have the following theorem. UIO for the system (1) if and only if the following
Theorem 1: The observer in the form of (3) is a relations are satisfied:
UIO for system (1) if and only if the following a) F is Hurwitz stable.
relations are satisfied: b) F = ( I − L1C ) A − L2 C .
1) e(t ) = Fe(t ) + Ge(t − d ) is asymptotically stable.
c) ( I − L1C ) Ad − L3C = 0 .
2) F = ( I − L1C ) A − L2 C .
d) H = ( I − L1C ) B .
3) G = ( I − L1C ) Ad − L3C .
e) ( I − L1C ) E = 0 .
4) H = ( I − L1C ) B .
5) ( I − L1C ) E = 0 . 4. UIO DESIGN OF TIME-DELAY SYSTEMS
Proof: From (1) and (3) we obtain the following
error equation In this section, we study the design of a UIO for a
class of time-delay systems based on the observer
e(t ) = Fe(t ) + Ge(t − d (t )) + [ ( I − L1C ) A − L2C − F ] x(t ) conditions proposed in Section 3. The following
lemma is needed.
+ [ ( I − L1C ) B − H ] u (t ) + ( I − L1C ) Ew(t ) (6) Lemma 1: [8] Condition 5) in Theorem 1 is
+ [ ( I − L1C ) Ad − L3C − G ] x(t − d (t )), solvable if and only if the following relation holds:

rank(CE ) = q , q ≤ m . (9)
with the initial condition e(t ) = x(t ) − xˆ (t ),
∀t ∈ [− d (t ),0] . Furthermore, when q < m , the general solution of the
532 Yan-Ming Fu, Guang-Ren Duan, and Shen-Min Song

equation in condition 5) can be expressed as L2 = P −1Y , (16)

−1
L3 = P Z . (17)
L1 = E (CE )+ + K ( I m − CE (CE )+ ) , (10)
Proof: Consider the following Lyapunov function:
where K ⊂ R n×m is an arbitrary matrix, (CE )+
t
denotes the generalized inverse matrix of CE . V ( e, t ) = eT (t ) Pe(t ) + ∫ eT ( s )Qe( s )ds, (18)
In the following we investigate the design of a UIO. t − d (t )
Under the condition in Lemma 1, two cases are
considered. where P = PT > 0 , Q = QT > 0 . By differentiating
Case 1: q < m
V (e, t ) along the solution of 1) of Theorem 1, we can
Substituting (10) into conditions 2) and 3) in
Theorem 1, gives obtain

F = Φ1 − K Φ 2 − L2C , (11) V (e, t ) = eT (t )[ F T P + PF + Q]e(t )

G = Ψ1 − K Ψ 2 − L3C , (12) − (1 − d (t ))eT (t − d (t ))Qe(t − d (t )) (19)

+ eT (t − d (t ))GT Pe(t ) + eT (t ) PGe(t − d (t )).

where

Φ1 = A − E (CE )+ CA , Φ 2 = [ I m − CE (CE )+ ]CA , Letting Qd = (1 − d1 )Q and using 0 ≤ d (t ) ≤ d1 < 1 ,

we get
Ψ1 = Ad − E (CE )+ CAd , Ψ 2 = [ I m − CE (CE )+ ]CAd .
 F T P + PF + Q PG 
Using Theorem 1, relations (11) and (12), we can V (e, t ) ≤ eˆT (t )   eˆ(t ) , (20)
obtain the observer error equation as  GT P −Qd 

e(t ) = (Φ1 − K Φ 2 − L2 C )e(t )

(13) where eˆT (t ) = eT (t ) eT (t − d (t ))  , if V (e, t ) < 0 ,
+ (Ψ1 − K Ψ 2 − L3C )e(t − d (t )).  
when eˆ(t ) ≠ 0 then e(t ) → 0 as t → ∞ and 1) of
From the above equation, we know that the problem Theorem 1 is globally asymptotically stable
of the observer design is reduced to the determination independent of delay. From (20), V (e, t ) < 0 if
of the observer parameters L2 , L3 and the free
parameter K such that condition 1) of Theorem 1 is  F T P + PF + Q PG 
satisfied. Based on Theorem 1 and Lyapunov stability   <0. (21)
theory, we can give the independent of delay  GT P −Qd 
conditions for the stability of the observer in the form
of (3). Substituting the conditions 2)-4) in Theorem 1 into
Theorem 2: The observer in the form of (3) is a (21) and replacing X , Y and Z by (15)-(17), it can
UIO for the system (1) if there exist be easily obtained that (21) is equivalent to the LMI
matrices P = PT > 0 , Q = QT > 0 , X , Y and Z (14).
satisfying the following LMI: If the delay d is a constant, we obtain the
following corollary.
Corollary 2: The observer in the form of (3) is a
 M1 M2 
 T  < 0, (14) UIO for the system (1) if there exist matrices
 M 2 −(1 − d1 )Q  P = PT > 0 , Q = QT > 0 , X , Y and Z satisfying
the following LMI:
where
M1 = Φ1T P + PΦ1 − ΦT2 X T − X Φ 2 − C T Y T − YC + Q ,  M1 M2 
 T  <0. (22)
M 2 = PΨ1 − X Ψ 2 − ZC .  M 2 −Q 

In this case, the free parameter K and the observer Under this condition, the free parameter K and the
parameters L2 , L3 are given by observer parameters L2 , L3 are given by

K = P −1 X , (15) K = P −1 X , (23)
 Design of Unknown Input Observer for Linear Time-delay Systems 533

L2 = P −1Y , (24) L1 = E (CE )−1 . (28)

−1
L3 = P Z . (25)
Substituting (28) into conditions 2) and 3) in Theorem
Now, we consider the design of the observer 1, yields
without internal delay. Corollary 1 provides us not
F = Ξ1 − L2C , (29)
only with the existence conditions for UIO in the form
of (8) but also an approach to the selection of the G = Ξ 2 − L3C , (30)
observer parameters.
First we provide the following theorem about the where
existence condition for solution to condition c).
Theorem 3: Condition c) in Corollary 1 is solvable Ξ1 = A − E (CE )−1 CA , Ξ 2 = Ad − E (CE )−1 CAd .
if and only if the following relation holds By using Theorem 1 and relations (29) and (30), the
observer error equation can be derived as
Ψ 2 
Ψ 2 
rank   = rank  C  . (26) e(t ) = (Ξ1 − L2 C )e(t ) + (Ξ 2 − L3C )e(t − d (t )) . (31)
C   Ψ1 
From (31), we know that the problem of the observer
Proof: Substituting equation (10) into the condition is reduced to the determination of the observer
c), we obtain parameters L2 , L3 such that the condition 1) in
Theorem 1 is satisfied. Based on Theorem 2, we can
Ψ 
[− K L3 ]  2  = Ψ1 . (27) give the independent of delay conditions for the
C  stability of the observer in the form of (3) in this case.
Corollary 3: The observer in the form of (3) is a
Using the result of linear algebra, matrix equation UIO for the system (1) if there exist
(27)is solvable if and only if (26) holds. matrices P = PT > 0 , Q = QT > 0 , Y and Z
Now let us consider the condition a) in Corollary 1. satisfying the following LMI:
Based on the observer knowledge in linear systems
theory, the condition a) in Corollary 1 can be satisfied
 N1 N2 
if and only if the matrix pair ( ( I − L1C ) A , C ) is  T  < 0, (32)
 N 2 −(1 − d1 )Q 
detectable.
Based on the above analysis, an algorithm for UIO
where
design can be given as follows.
Algorithm UIO: ( q < m ) N1 = Ξ1T P + PΞ1 − C T Y T − YC + Q , N 2 = PΞ 2 − ZC .
Step 1: Calculate matrix L1 by (10).
In this case, the observer parameters L2 and L3 are
Step 2: Check the rank condition (26). If (26) holds,
calculate L3 and K by solving (27). Otherwise go given by
to Step 4. L2 = P −1Y , (33)
Step 3: If the matrix pair ( ( I − L1C ) A , C ) is
L3 = P −1Z . (34)
detectable, obtain L2 by using any stabilization
approach in linear systems theory such that the matrix Similar to Theorem 3, we present the following
F given in condition b) in Corollary 1 is stable, and corollary concerning the existence condition for
then go to Step 5. Otherwise proceed to Step 4. solution to condition c) in Corollary 1.
Step 4: Solve LMI (14). If LMI (14) is insolvable, the Corollary 4: The condition c) in Corollary 1 is
problem has no solution, or the algorithm fails. solvable if and only if the following relation holds:
Otherwise, we can obtain matrices P , X , Y and Z ,
and can further calculate L1 , L2 and L3 according C 
rank C = rank   . (35)
to (10) and (15)-(17). Ξ1 
Step 5: Calculate F , G, H by their expressions in the
conditions 2), 3) and 4) in Theorem 1. Based on Corollaries 3 and 4, an algorithm for UIO
Case 2: q = m design can be given as follows.
If q = m , based on condition 5) in Theorem 1, we Algorithm UIO: ( q = m )
can obtain Step 1: Calculate matrix L1 by (28).
534 Yan-Ming Fu, Guang-Ren Duan, and Shen-Min Song

Step 2: Check the rank condition (35), if (35) holds, Step 3: It is easy to check that [ ( I − L1C ) A , C ] is
calculate L3 by solving equation c) in Corollary 1. detectable. So we can design the observer in the form
Otherwise go to Step 4. of (8). Based on some stabilization approach in linear
Step 3: Based on L1 , if the matrix pair systems theory, we obtain
( ( I − L1C ) A ,
C ) is detectable, obtain L2 by any  1.2 −0.2 
stabilization approach in linear systems theory such L2 =  .
1.48 0.32 
that matrix F given in condition b) in Corollary 1 is
stable, and then go to Step 5. Otherwise proceed to Step 4: Using the equations in conditions 2), 3) and 4)
Step 4. in Theorem 1, we can derive
Step 4: Solve LMI (32). If LMI (32) is insolvable, the
problem has no solution, or the algorithm fails.  0.6 −0.2   −1 0   1.2 
L1 =   , F =  , H = .
Otherwise, we can obtain matrices P , X , Y and Z , 0.74 0.32   0 −2   −0.32 
and can further calculate L2 and L3 according to
(33) and (34). Example 2: ( q = m )
Step 5: Calculate F , G, H by their expressions in Consider the following system:
the conditions 2), 3) and 4) in Theorem 1.
 −1 0  1 0 
x (t ) =   x(t ) + 0 1  x(t − d (t ))
5. NUMERICAL EXAMPLES  0 − 2   
1  1 
To illustrate the design methods for UIO, in this +   u (t ) +   w(t ),
section we present two examples. 0 2
Example 1: ( q < m ) y (t ) = [1 0] x(t ) ,
Consider the following system:
where d (t ) satisfies 0 ≤ d (t ) ≤ 0.2 .
 −1 0  1 0  It is clear that q = m . So we can design the observer
x (t ) =   x(t ) +   x(t − d (t ))
 0 −2  0 1  for this example system using algorithm UIO ( q = m ).
1  1  Step 1: From (28), we can obtain
+   u (t ) +   w(t ),
0 2 L1 = [1 2] .
T

0 1 
y (t ) =   x(t ) , Step 2: Check the rank condition (35). We can obtain
1 0 
C 
where d (t ) satisfies 0 ≤ d (t ) ≤ 0.2 . rank C ≠ rank   .
Ξ1 
It is clear that q < m . So we can design the
observer for this example system using algorithm UIO So we can’t design the observer in the form of (8).
( q < m ). Step 3: Solve LMI (32) by the LMI toolbox in
MATLAB.
Step 1: Calculate L1 by (10).
 55.3031 −0.0000   69.1288 −0.0000 
P=  , Q= ,
0.4 0.2   0.2 −0.4   −0.0000 26.9949   −0.0000 59.0970 
L1 =   +K .
 0.8 0.4   −0.4 0.8  62.2160   −0.0000 
Y =  , Z = .
Step 2: Note that 59.9889   −53.9898

Ψ 2  According to (33)-(34), we have

Ψ 2 
rank   = rank  C  = 2 , 1.1250   0.0000 
C   Ψ1 
L2 =   , L3 =  .
 2.0000   −2.0000 
and the rank condition (26) holds. Following from Step 4: Using the equations in conditions 2), 3) and 4)
(27), we have in Theorem 1, we can derive

 0.2 −0.4   −0.2 0.4   −1.1250 0  0 0  0 

K =  , L3 =  . F =  , G=  , H =  .
 −0.06 0.12   0.14 −0.08  −0.0000 −2.0000  0 1.0  2
 Design of Unknown Input Observer for Linear Time-delay Systems 535

6. CONCLUSIONS Conference on Decision and Control, pp. 1433-

1437, 2001.
In this paper, the design problem of UIO for a class [11] M. Saif and Y. P. Guan, “A new approach to
of linear time-delay systems is studied. Necessary and robust fault detection and identification,” IEEE
sufficient conditions to its existence are presented. Trans. on Aerospace and Electronic Systems, vol.
Based on these conditions and Lyapunov stability 29, no. 3, pp. 685-695, 1993.
theory, two design algorithms are proposed. [12] H. L. Yang and M. Saif, “Observer design and
Demonstrative examples show the effect of the fault diagnosis for state-retarded dynamical
proposed algorithms. systems,” Automatica, vol. 34, no. 2, pp. 217-
Future research will focus on the design problem of 227, 1998.
UIO when the estimate error can’t completely be [13] H. L. Yang and M. Saif, “Robust observation
decoupled from unknown inputs. and fault diagnosis in a class of time-delay
control systems,” Proc. of the American Control
REFERENCES Conference, pp. 478-482, June 1997.
[1] M. Hou, P. Zitek, and R. J. Patton, “An observer [14] S. X. Ding, M. Y. Zhong, B. Y. Tang, and P.
design for linear time-delay systems,” IEEE Zhang, “An LMI approach to the design of fault
Trans. on Automatic Control, vol. 47, no. 1, pp. detection filter for time-delay LTI systems with
121-125, January 2002. unknown input,” Proc. of the American Control
[2] M. Darouch, “Linear functional observers for Conference, pp. 2137-2142, June 2001.
systems with delays in state variables,” IEEE
Trans. on Automatic Control, vol. 46, no. 3, pp.
491-496, March 2001. Yan-Ming Fu received the B.S.
[3] A. E. Pearson and Y. A. Fiagbedzi, “An observer degree in Automatic Control from the
for time lag systems,” IEEE Trans. on Automatic. Northeast Institute of Electric Power,
Control, vol. 34, no. 4, pp. 775-777, July 1989. China, in 2001, and the M.S. degree in
Control Theory and Control
[4] Y. X. Yao and Y. M. Zhang, “Parameterization of Engineering from Harbin Institute of
observers for time delay systems and its Technology, China, in 2003. He is
application in observer design,” IEE Proc. currently working toward the Ph.D.
Control Theory Appl., vol. 143, no. 3, pp. 225- degree at the Center for Control Theory and Guidance
232, May 1996. Technology, Harbin Institute of Technology, China. His
[5] O. Sename, “Unknown input robust observer for research interests include robust control and filtering,
time delay system,” Proc. of the 36th IEEE jumping systems, and time-delay systems.
Conference on Decision and Control, pp. 1629-
1630, 1997. Guang-Ren Duan received the B.S.
[6] J. Leyva-Ramos and A. E. Pearson, “An degree in Applied Mathematics and
asymptotic model observer for linear the M.S. degree and Ph.D. degrees
autonomous time lag systems,” IEEE Trans. on from Harbin Institute of Technology,
Automatic control, vol. 40, no. 7, pp. 1291-1294, China. He is currently a Professor at
July 1995. the Harbin Institute of Technology,
[7] Z. D. Wang, James Lam, and K. J. Burnham, where he is also Director of the Center
“Stability Analysis and observer design for for Control Theory and Guidance
Technology. His research interests
neutral delay systems,” IEEE Trans. on
include robust control, eigenstructure assignment, missile
Automatic Control., vol. 47, no. 3, pp. 478-482, autopilot design, magnetic bearing control, and descriptor
March 2002. systems.
[8] S. H. Lee, J. Hong, and J. H. See, “Observers for
bilinear systems with unknown inputs and
application to superheater temperature control,” Shen-Min Song received the B.S. and
Control Eng. Practice, vol. 5, no. 4, pp. 493-506, M.S. degrees in Applied Mathematics
1997. from Harbin Institute of Technology in
1990 and 1993, respectively, and the
[9] X. Guan, Z. Lin, and G. Duan, “Robust
Ph.D degree in Control Theory and
guaranteed cost control for discrete-time Application from Harbin Institute of
uncertain systems with delay,” IEE Proc.- Technology in 1996. From 2000 to
Control Theory Appl., vol. 146, no. 6, pp. 598- 2002, he was a Visiting Researcher at
602, 1999. Tokyo University. He is currently employed as an an
[10] W. Michiels and D. Roose, “Time-delay Associate Professor at the Harbin Institute of Technology.
compensation in unstable plants using delayed His research interests include robust control and filtering,
state feedback,” Proc. of the 40th IEEE and soft computing theory and application.