Systems & Control Letters 59 (2010) 277–283

Contents lists available at ScienceDirect

Systems & Control Letters

journal homepage: www.elsevier.com/locate/sysconle

Delay-adaptive feedback for linear feedforward systems

Nikolaos Bekiaris-Liberis ∗ , Miroslav Krstic
Department of Mechanical and Aerospace Engineering, University of California at San Diego, La Jolla, CA 92093-0411, USA

article info abstract

Article history: Predictor techniques are an indispensable part of the control design toolbox for plants with input and
Received 8 October 2009 state delays of significant size. Yet, they suffer from sensitivity to the design values. Explicit feedback laws
Received in revised form were recently introduced by Jankovic for a class of feedforward linear systems with simultaneous state
13 March 2010
and input delays. For the case where the delays are of unknown length, using the certainty equivalence
Accepted 13 March 2010
Available online 13 April 2010
principle, we design a Lyapunov-based adaptive controller, which achieves global stability and regulation,
for arbitrary initial estimates for the delays. We consider a two-block sub-class of linear feedforward
Keywords:
systems. A generalization to the n-block case involves a recursive application of the same techniques.
Delays systems © 2010 Elsevier B.V. All rights reserved.
Adaptive control
Feedforward systems

1. Introduction in [20] the result is extended to also incorporate unknown plant

parameters. The aforementioned designs are based on predictor
The design of stabilizing controllers for systems with delays feedback together with tools that come from the adaptive control
continues to be an active area of research. Controllers for both of parabolic PDEs [21].
linear and nonlinear systems exist in the literature, many of which In this paper we develop a delay-adaptive version of the design
are based on predictor-like techniques [1–9,24–38]. However, introduced by Jankovic in [4] for linear feedforward systems with
systems with simultaneous input and state delay remain a simultaneous state and input delays. In [4], for the system
challenge, even for linear systems [10,2–4,11–13].
Ẋ1 (t ) = F1 X1 (t ) + H1 X2 (t − D1 ) + B1 U (t ) (1)
An even more challenging problem is the adaptive control
of systems with simultaneous input and state delays. From the Ẋ2 (t ) = F2 X2 (t ) + B2 U (t ), (2)
practical point of view, controllers for delay systems should be
a predictor-based controller is designed as
robust to parametric uncertainties, including plant parameters � 1
and delays. The importance of designing robust controllers when
the delays are the unknown parameters was highlighted in the U (t ) = K1 D1 e−F1 D1 θ H1 X2 (t + D1 (θ − 1)) dθ
0
control problems considered in [14] and [15]. On the other hand,
since there already exists a rich literature for the control of time + K1 X1 (t ) + K2 X2 (t ). (3)
delay systems, adaptive control schemes that are based on existing The above controller is based on a transformation that reduces
control techniques are of interest. Since many control schemes are the system to an equivalent system without state delay. This
based on predictor-like techniques, which are known to be very transformation is
sensitive to delay uncertainties [16], designing adaptive versions � 1
of these control schemes is crucial for making them usable in Z1 (t ) = X1 (t ) + D1 e−F1 D1 θ H1 X2 (t + D1 (θ − 1)) dθ (4)
scenarios with uncertain delays. 0
Adaptive control schemes can be found in [17,18]. Yet, the Z2 (t ) = X2 (t ), (5)
adaptive control problem when the delays are the unknown
parameters had not been solved until recently with the works and it transforms the system (1)–(2) to
in [19,20]. In [19] the problem of designing an adaptive control Ż1 (t ) = F1 Z1 (t ) + e−F1 D1 H1 Z2 (t ) + B1 U (t ) (6)
scheme for a linear system with unknown input delay is solved, and
Ż2 (t ) = F2 Z2 (t ) + B2 U (t ). (7)
The importance of the previous transformation, besides transform-
∗ Corresponding author. ing the original system to an equivalent one without state delay, is
E-mail addresses: [email protected], [email protected] that the system can be linearly parameterized in the state delay,
(N. Bekiaris-Liberis). which is the key for designing an adaptive control law.
0167-6911/$ – see front matter © 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.sysconle.2010.03.001
278 N. Bekiaris-Liberis, M. Krstic / Systems & Control Letters 59 (2010) 277–283
� � ����
e−D1 F1 H1
Then, assuming that the pair , BB12 is com-
F1
0 F2
3. Controller design
pletely controllable, a state feedback controller U (t ) = K1 Z1 (t ) +
K2 Z2 (t ) is designed such that the transformed system is asymptot- We first rewrite (12)–(13) using a PDE representation of the
ically stable. It can be shown that controllability of the original sys- delayed states and control as
tem is equivalent to controllability of the transformed system [22].
This design can be also applied in the case where there is a delay Ẋ1 (t ) = F1 X1 (t ) + H1 ξ (0, t ) + B1 u(0, t ) (15)
in the input, say D2 . In this case after employing the state transfor- D1 ξt (x, t ) = ξx (x, t ) (16)
mation a predictor feedback is needed for the transformed system.
ξ (1, t ) = X2 (t ) (17)
In this case the controller that compensates for D2 is given by
� Ẋ2 (t ) = F2 X2 (t ) + B2 u(0, t ) (18)
1
U (t ) = K e AD2
Z (t ) + KD2 e AD2 (1−θ )
BU (t − D2 (θ − 1)) dθ , (8) D2 ut (x, t ) = ux (x, t ) (19)
0
u(1, t ) = U (t ), (20)
where
� � where x ∈ [0, 1]. We assume that the infinite-dimensional states
F e−D1 F1 H1 ξ (x, t ), u(x, t ), x ∈ [0, 1] are available for measurement. This as-
A= 1 (9)
0 F2 sumption is not in contradiction with the assumption that the con-
� � vection speeds 1/D1 and 1/D2 are unknown. As restrictive as the
B1
B= (10) requirement for measurement of ξ (x, t ), u(x, t ), x ∈ [0, 1] may
B2
appear, we do not believe that the delay-adaptive problem without
� �
K = K1 K2 . (11) such measurements is solvable globally because it cannot be for-
mulated as linearly parameterized in the unknown delays D1 and
The controller (8) is the basis of our adaptive design.
D2 .
The transport PDE states can be expressed in terms of the past
2. Problem formulation values of X2 and U as

In this paper we consider both the state and input delays to be ξ (x, t ) = X2 (t + D1 (x − 1)) (21)
unknown, that is, we consider the system u(x, t ) = U (t + D2 (x − 1)). (22)
Ẋ1 (t ) = F1 X1 (t ) + H1 X2 (t − D1 ) + B1 U (t − D2 ) (12) Using the certainty equivalence principle the controller (8) is taken
Ẋ2 (t ) = F2 X2 (t ) + B2 U (t − D2 ), (13) as
 � 1

with D1 and D2 unknown. Since D1 and D2 are unknown, in addition
A(D̂1 )D̂2 (t ) X1 (t ) + D̂1 (t ) e−F1 D̂1 (t )y H1 ξ (y, t )dy
to the predictor based controller, we must design two estimators, U (t ) = K (D̂1 )e 0
one for each of the delays. We employ projector operators and X2 (t )
assume a bound on the length of the delays to be known. � 1

Assumption 1. There exist known constants D1 , D1 and D2 such + K (D̂1 )D̂2 (t ) eA(D̂1 )D̂2 (t )(1−y) Bu(y, t )dy. (23)
0
that D1 ∈ [D1 , D1 ] and D2 ∈ [0, D2 ].
The update laws for the estimations of the unknown delays D1 and
Our controller is based on the transformed system (i.e., on the D2 are given by
system
�� without� state
� ��delay). As indicated in Section 1, the pair
F1 e−D1 F1 H1
, B1
must be completely controllable. Under
˙
D̂1 (t ) = γ1 Proj[D ,D1 ] {τD1 } (24)
0 F2 B2 1
this assumption we can find a stabilizing state feedback. In the case ˙
of unknown state delay D1 , we must assume that there exists a D̂2 (t ) = γ2 Proj[0,D2 ] {τD2 }, (25)
stabilizing state feedback for all values of the state delay in a given
where the projector operators are defined as
interval. We thus make the following assumption.
�� � � �� 
Assumption 2. The pair F1 e−D1 F1 H1
, BB12 is completely � � 0 if D̂i = Di and τDi < 0
0 F Proj[D ,Di ] τDi = (26)
0 if D̂i = Di and τDi > 0 ,
2
i
controllable ∀D1 ∈ [D1 , D1 ]. Furthermore, we assume that there τDi else
exists a triple of vector/matrix-valued
� � functions (K�(D1 ), P (D� 1 ),
Q (D1 )) such that K (D1 ) ∈ C 1 [D1 , D1 ] , P (D1 ) ∈ C 1 [D1 , D1 ] , Q and where
� �
(D1 ) ∈ C 0 [D1 , D1 ] , the matrices P (D1 ) and Q (D1 ) are positive �1 2 T
definite and symmetric, and the following Lyapunov equation is 0
(1 + x)w(x, t )K (D̂1 )eA(D̂1 )D̂2 (t )x dx − a2
Z (t )P (D̂1 )
τD1 =
satisfied ∀D1 ∈ [D1 , D1 ]: Γ (t )
T
× R2 ( t ) (27)
(A(D1 ) + BK (D1 )) P (D1 ) + P (D1 ) (A(D1 ) + BK (D1 )) �1
= −Q (D1 ), ∀D1 ∈ [D1 , D1 ]. (14)
(1 + x)w(x, t )K (D̂1 )eA(D̂1 )D̂2 (t )x dx
τD2 = − 0
� Γ (t ) �
Our final assumption is needed in the choice of the normaliza- × Bu(0, t ) + A(D̂1 )Z (t ) (28)
tion coefficients in the adaptation laws for the delay estimates.
� 1
Assumption 3. The quantities λ = infD1 ∈[D ,D1 ] min{λmin (Q (D1 )), Γ (t ) = 1 + Z T (t )P (D̂1 )Z (t ) + a2 (1 + x)w 2 (x, t )dx
1 0
� 1
λmin (P (D1 ))} and λ = supD1 ∈[D1 ,D1 ] λmax (P (D1 )) exist and are
known. +k (1 + x)ξ T (x, t )ξ (x, t )dx, (29)
0
 N. Bekiaris-Liberis, M. Krstic / Systems & Control Letters 59 (2010) 277–283 279

D̂ D̃1
with Using relations (16) and (17), the fact that D1 = 1 − and
1 D1
λD1 integrating by parts the last integral in (42) we obtain
k≤ (30)
8
Ż1 (t ) = F1 X1 (t ) + H1 ξ (0, t ) + B1 u(0, t )
D2 sup |P (D̂1 )B|2 � 1

D̂1 ∈[D1 ,D1 ] + D̂˙ 1 (t ) e−D̂1 (t )F1 y H1 ξ (y, t )dy

a2 ≥ . (31) 0
� 1
λ
In the above relation we use the following signals which are − D̂1 (t )D̂˙ 1 (t ) F1 ye−D̂1 (t )F1 y H1 ξ (y, t )dy
derived in the stability analysis of the closed-loop system � � 0
� D̃1 � −D̂1 (t )F1 �
1 + 1− e H1 X2 (t ) − H1 ξ (0, t )
Z1 (t ) = X1 (t ) + D̂1 (t ) e−D̂1 (t )F1 y H1 ξ (y, t )dy (32) D1
0 � ��
1
Z2 (t ) = X2 (t ) (33) D̃1
  + 1− D̂1 (t )F1 e−D̂1 (t )F1 y H1 ξ (y, t )dy (44)
� 1 D1 0
−D̂1 (t )F1
e H1 Z2 (t ) − H1 ξ (0, t ) + D̂1 (t )F1 e−D̂1 (t )F1 y H1 ξ (y, t )dy
R2 (t ) = (34)
0
0
Ż2 (t ) = F2 X2 (t ) + B2 u(0, t ). (45)
and the transformed infinite dimensional state of the actuator Using (32)–(33) and after some algebra we arrive at
� � �
w(x, t ) = u(x, t ) − K (D̂1 ) eA(D̂1 )D̂2 (t )x Z (t ) + D̂2 (t ) Ż1 (t )
Ż (t ) =
Ż2 (t )
� x
�
eA(D̂1 )D̂2 (t )(x−y) Bu(y, t )dy . D̃1
× (35) = A(D̂1 )Z (t ) + Bu(0, t ) + D̂˙ 1 (t )R1 (t ) − R2 (t ), (46)
0 D1
where R2 (t ) is defined in (34) and
4. Stability analysis � �
A(D̂1 ) = F1 e−D̂1 (t )F1 H1 (47)
This section is devoted to the proof of the main result. We start 0 F2
by giving the main theorem and in the rest of the section we prove � �
it using a series of technical lemmas. B1
B= (48)
B2
Theorem 1. Let Assumptions 1–3 hold. Then system (12)–(13) with � 
the controller (23) and the update laws (24)–(25) is stable in the sense
1 � �
−D̂1 (t )F1 y
I − D̂1 (t )F1 y e H1 ξ (y, t )dy 
that there exist constants R and ρ such that R1 ( t ) =  0
. (49)
� � 0
Ω (t ) ≤ R eρ Ω (0) − 1 , (36)
Using relation (35) for x = 0 we get that
where � �
2 2 2 Ż (t ) = A(D̂1 ) + BK (D̂1 ) Z (t ) + Bw(0, t )
Ω (t ) = |X (t )| + �ξ (t )� + �u(t )� + D̃21 (t ) + D̃22 (t ), (37)
and D̃1
� � + D̂˙ 1 (t )R1 (t ) − R2 (t ). (50)
1
1 t D1
2 T T
�ξ (t )� = ξ (y, t ) ξ (y, t )dy = X2 (θ ) X2 (θ )dθ (38)
Moreover the transformation of the actuator state w satisfies
0 D1 t −D 1
� 1 � t D̃1
1 D2 wt (x, t ) = wx (x, t ) + D2 p1 (x, t ) − D̃2 p2 (x, t )
�u(t )�2 = u2 (y, t )dy = U 2 (θ )dθ . (39) D1
0 D2 t −D2

Furthermore − D2 D̂˙ 1 (t )q1 (x, t ) − D2 D̂˙ 2 (t )q2 (x, t ), (51)

lim X (t ) = 0 (40) where

t →∞

lim U (t ) = 0. (41) p1 (x, t ) = K (D̂1 )eA(D̂1 )D̂2 (t )x R2 (t ) (52)

t →∞
� x� �
∂ K (D̂1 ) ∂ A(D̂1 )
We start proving the above theorem by first transforming the q1 (x, t ) = + K (D̂1 ) (x − y)
system (15)–(20) using the transformations (32)–(33) and (35). By 0 ∂ D̂1 ∂ D̂1
differentiating with respect to time (32) and (33) and by using (15) × D̂�2 (t )eD̂2 (t )A(D̂1 )(x−y) Bu(y, t )dy �
and (18) we get
∂ K (D̂1 ) ∂ A(D̂1 )
+ + K (D̂1 ) D̂2 (t )x
Ż1 (t ) = F1 X1 (t ) + H1 ξ (0, t ) + B1 u(0, t ) ∂ D̂1 ∂ D̂1
� 1
+ D̂˙ 1 (t ) e−D̂1 (t )F1 y H1 ξ (y, t )dy × eD̂2 (t )A(D̂1 )x Z (t ) + K (D̂1 )eD̂2 (t )A(D̂1 )x R1 (t ) (53)
0
� 1
p2 (x, t ) = K (D̂1 )eD̂2 (t )A(D̂1 )x Bu(0, t ) + K (D̂1 )A(D̂1 )
− D̂1 (t )D̂˙ 1 (t ) F1 ye−D̂1 (t )F1 y H1 ξ (y, t )dy
� 1 0 × eD̂2 (t )A(D̂1 )x Z (t ) (54)
� x� �
+ D̂1 (t ) e−D̂1 (t )F1 y H1 ξt (y, t )dy (42)
0 q2 (x, t ) = K (D̂1 ) I + D̂2 (t )(x − y)A(D̂1 ) eA(D̂1 )D̂2 (t )(x−y)
0
Ż2 (t ) = F2 X2 (t ) + B2 u(0, t ). (43) × Bu(y, t )dy. (55)
280 N. Bekiaris-Liberis, M. Krstic / Systems & Control Letters 59 (2010) 277–283

Thus now, system (15)–(20) is mapped to the target system that is Using relations (52)–(55) together with Young’s and Cauchy–
comprised of (50) and (51). Moreover, the inverse transformation Schwartz’s inequalities, relations (65)–(66) and (57) we get the
of the state X (t ) is easily obtained from Eqs. (32)–(33) and the bounds of the lemma with
inverse transformation of (35) is given by
� � � Mp1 = MK2 e2MA D2 MR2 (73)
A(D̂1 )+BK (D̂1 ) D̂2 (t )x � �
u(x, t ) = w(x, t ) + K (D̂1 ) e Z (t )
Mp2 = 3 max MK2 e2MA D2 |B|2 , MK2 MA2 e2MA D2 (74)
� x � � � ��
A(D̂1 )+BK (D̂1 ) D̂2 (t )(x−y) �2
+ D̂2 (t ) e Bw(y, t )dy . (56) Mq1 = 3 max MK� D2 + D2 MK MA� |B|2 e2MA D2 ,
0
� �2 2M D �
We first prove that the signals in (50) and (51) that multiply the MK2 e2MA D2 MR1 , MK� + MK MA D2 �
e A 2
(75)
˙
‘‘disturbances’’ D̂i and D̃i , i = 1, 2., are bounded with respect to � �2
the system’s transformed states Z (t ), ξ (x, t ) and w(x, t ). Before Mq2 = MK2 1 + D2 MA e2MA D2 . � (76)
doing that, we point out that boundness of the transformed states
is equivalent to boundness of the original states.
Lemma 3. There exist constants k, a2 , γ1 and γ2 such that for the
Lemma 1. There exist constants Mu , Mw , MX and MZ such that Lyapunov function
� �
�u(t )�2 ≤ Mu �w(t )�2 + |Z (t )|2 (57) D̃21 (t ) D̃22 (t )
� � V (t ) = D2 log (1 + Ξ (t )) + a2 D2 + a2 , (77)
|X (t )|2 ≤ MX |Z (t )|2 + �ξ (t )�2 (58) D1 γ1 γ2
� � where
�w(t )�2 ≤ Mw �u(t )�2 + |Z (t )|2 (59) �
� � � � 1
T
|Z (t )|2 ≤ MZ |X (t )|2 + �ξ (t )�2 . (60) Ξ (t ) = Z (t ) P D̂1 Z (t ) + k (1 + x)ξ T (x, t )ξ (x, t )dx
0
� 1
Proof. First observe that the signals K (D̂1 ), P (D̂1 ) and A(D̂1 ) are
+ a2 (1 + x)w 2 (x, t )dx, (78)
continuously differentiable with respect to D̂1 . Moreover, since D̂1 0
and D̂2 are uniformly bounded, the signals K (D̂1 ), P (D̂1 ) and A(D̂1 ) the following holds
and their derivatives are also uniformly bounded. Denote by MK ,
MP and MA the bounds of K (D̂1 ), P (D̂1 ) and A(D̂1 ) respectively, and V (t ) ≤ V (0). (79)
with MK� , MP� and MA� the bounds of their derivatives. From relations
(32)–(33) and (35), (56) and using Young’s and Cauchy–Schwartz’s Proof. Taking the time derivative of the above function we obtain
inequalities it easy to show that the above bounds hold with
� � D̃1 � ˙ � D̃2 � ˙ �
2 V̇ (t ) = −2a2 D2 D̂1 (t ) − γ1 τD1 − 2a2 D̂2 (t ) − γ2 τD2
Mu = 3 max 1 + MK2 D2 e2D2 (MA +|B|MK ) 2
|B| , MK2 e2D2 (MA +|B|MK ) (61) D1 γ1 γ2
� � �
2 D2
MZ = 2 max 1, D1 e2|F1 |D1 |H1 | 2
(62) + −Z T (t )Q (D̂1 )Z (t ) + 2Z T (t )P (D̂1 )Bw(0, t )
1 + Ξ (t )
� 2
�
MX = 2 max 1, D1 e2|F1 |D1 |H1 |2 (63) 2k k
+ Z2T (t )Z2 (t ) − |ξ (0, t )|2
� 2
� D1 D1
Mw = 3 max 1 + MK2 D2 e2MA D2 |B|2 , MK2 e2MA D2 . � (64) � 1
a2 k
− w2 (0, t ) − ξ T (x, t )ξ (x, t )dx
D2 D1 0
We are now ready to state the following lemma � �
1
a2 ˙ ∂ P (D̂1 )
Lemma 2. There exist constants MR1 , MR2 , Mp1 , Mp2 , Mq1 and Mq2 − w (x, t )dx + D̂1 (t ) Z T (t )
2
Z (t )
D2 0 ∂ D̂1
such that the following bounds hold � 1 �
T
|R1 (t )|2 ≤ MR1 �ξ (t )�2 (65) + 2Z (t )P (D̂1 )R1 (t ) − 2a2 (1 + x)w(x, t )q1 (x, t )dx
0
� � �
|R2 (t )|2 ≤ MR2 |Z (t )|2 + |ξ (0, t )|2 + �ξ (t )�2 (66) � 1
� � − 2a2 D̂˙ 2 (t ) (1 + x)w(x, t )q2 (x, t )dx . (80)
p21 (x, t ) ≤ Mp1 |Z (t )|2 + |ξ (0, t )|2 + �ξ (t )�2 (67) 0
� �
p22 (x, t ) ≤ Mp2 |Z (t )|2 + u2 (0, t ) (68) Using the properties of the projector operators and relations
� � (24)–(25) we get
q21 (x, t ) ≤ Mq1 |Z (t )|2 + �u(t )�2 + �ξ (t )�2 (69)
�
q22 (x, t ) ≤ Mq2 �u(t )� , 2
(70) D2 2k
V̇ (t ) ≤ −Z T (t )Q (D̂1 )Z (t ) + Z2T (t )Z2 (t )
1 + Ξ (t ) D1
for all x ∈ [0, 1].
Proof. From relations (49) and (34) and by using Young’s and k a2
− |ξ (0, t )|2 − w 2 (0, t ) + Z T (t )P (D̂1 )Bw(0, t )
Cauchy–Schwartz’s inequalities we get the bounds for R1 (t ) and D1 D2
R2 (t ) with
� 1 � 1
k T a2
� �2 − ξ (x, t )ξ (x, t )dx − w2 (x, t )dx
D1 D2
MR1 = 1 + D1 |F1 | e2D2 |F1 | |H1 |2 (71) 0
�
0

� � ∂ P (D̂1 )
MR2 = 3|H1 |2 max e2D2 |F1 | |F1 |2 , 1, D1 |F1 |2 e2D2 |F1 | .
2
(72) + D̂˙ 1 (t ) Z T (t ) Z (t ) + 2Z T (t )P (D̂1 )R1 (t )
∂ D̂1
 N. Bekiaris-Liberis, M. Krstic / Systems & Control Letters 59 (2010) 277–283 281
� 1
�
max {B1 , B2 }
− 2a2 (1 + x)w(x, t )q1 (x, t )dx m2 = � �, (91)
0 min λ, k, a2
� �
1 we get
− 2a2 D̂˙ 2 (t ) (1 + x)w(x, t )q2 (x, t )dx . (81) �
0 D2
V̇ (t ) ≤ − (m1 − m2 (γ1 + γ2 )) |Z (t )|2 + w 2 (0, t )
1 + Ξ (t )
Then using Young’s inequality and (30)–(31) we get �
� + �w(t )�2 + |ξ (0, t )|2 + �ξ (t )�2 . (92)
D2 λ k k
V̇ (t ) ≤ − |Z (t )|2 − |ξ (0, t )|2 − �ξ (t )�2
1 + Ξ (t ) 2 D1 D1 Thus when γ1 + γ2 ≤
m1
, V̇ (t ) is negative definite and thus
m2
�
a2 2 a2 ∂ P (D̂1 )
− w (0, t ) − �w(t )�2 + D̂˙ 1 (t ) Z T (t ) Z (t ) V (t ) ≤ V (0). � (93)
2D2 D2 ∂ D̂1
� 1 �
To prove the stability bound of Theorem 1 we use the following
+ 2Z T (t )P (D̂1 )R1 (t ) − 2a2 (1 + x)w(x, t )q1 (x, t )dx lemma.
0
� �
1 Lemma 4. There exist constants M and M such that
− 2a2 D̂˙ 2 (t ) (1 + x)w(x, t )q2 (x, t )dx . (82)
0 M Ξ (t ) ≤ Π (t ) ≤ M Ξ (t ), (94)
Using bounds (65)–(70) together with relations (26) and where
(24)–(25), and by employing Young’s inequality one more time we
get Π (t ) = |X (t )|2 + �ξ (t )�2 + �u(t )�2 . (95)

|D̂˙ 1 (t )| ≤ γ1 |τD1�| �
Proof. Immediate, using (57)–(60) with
|Z (t )|2 + |ξ (0, t )|2 + �ξ (t )�2 + �w(t )�2 M = max {Mu + MX , MX + 1} (96)
≤ γ1 M1 (83)
1 + Ξ (t ) M = max {Mw + MZ , MZ + 1} . � (97)
� �
˙ |Z (t )|2 + w2 (0, t ) + �w(t )�2
|D̂2 (t )| ≤ γ2 |τD2 | ≤ γ2 M2 , (84) We are now ready to derive the stability estimate of Theorem 1.
1 + Ξ (t )
Using (77) it follows that
where � �
� � V (t )
1 1 Ξ (t ) ≤ e D2 − 1 (98)
M1 = max Mp1 + MP + MP MR2 , 1 (85)
a2 a2 � �
� � V (t ) V (t )
M2 = max 1, 2Mp2 2Mp2 MK2 , + Mp2 . (86) D̃21 + D̃22 ≤ C2 ≤ C2 e D2 − 1 , (99)
D2
Plugging in the above bound to (82) (and applying once more
where
Young’s and Cauchy–Schwartz’s inequalities) we get
� �
� γ2 D2 + γ1 D1
D2 λ k k C2 = . (100)
V̇ (t ) ≤ − |Z (t )|2 − |ξ (0, t )|2 − �ξ (t )�2 a2
1 + Ξ (t ) 2 D1 D1
Consequently
a2 2 a2 2
− w (0, t ) − �w(t )� � �
2D2 D2 � � V (t )
� � Ω ( t ) ≤ M + C2 e D2
−1 . (101)
|Z (t )| + |ξ (0, t )|2 + �ξ (t )�2 + �w(t )�2
2
+ B1 γ1
1 + Ξ (t ) Moreover, from (77) we take
� � � �� �
× |Z (t )|2 + �ξ (t )�2 + �w(t )�2 V (0) ≤ max λ, k, a2 |Z (0)|2 + �ξ (0)�2 + �w(0)�2
� �
� � a2 a2 D2 � 2 �
|Z (t )|2 + w2 (0, t ) + �w(t )�2 + max , D̃1 (0) + D̃22 (0) , (102)
+ B2 γ2 γ2 γ1 D1
1 + Ξ (t )
� �� and using Lemma 4 we have
2 2
× |Z (t )| + �w(t )� , (87) V (0) ≤ C3 Ω (0), (103)
where
where � � � � ��
� � 1 D2 max λ, k, a2
B1 = M1 max MP + MP + a2 Mq1 + 2a2 Mq1 Mu + 2a2 Mq1 , C3 = max max , , . (104)
� 2γ2 2γ1 D1 M
MP MR1 + 2a2 Mq1 , 2a2 + 2a2 Mq1 Mu (88)
� � Thus, by setting
B2 = 2M2 a2 1 + Mq2 Mu . (89)
Now by defining the constants R = M + C2 (105)
� � ρ = C3 , (106)
λ k a2
m1 = min , , (90)
2 D1 2D2 we get the stability result in Theorem 1.
282 N. Bekiaris-Liberis, M. Krstic / Systems & Control Letters 59 (2010) 277–283

Fig. 2. The estimations of the unknown delays for the cases D̂2 (0) = 0 and D̂2
(0) = 1.6.
Fig. 1. System’s response for the cases D̂2 (0) = 0 and D̂2 (0) = 1.6.

We now turn our attention to proving the convergence of X (t ) at 0.25. The lower bounds for the unknown delays are D1 = 0.1
and U (t ) to zero. We use here an alternative to Barbalat’s Lemma for D1 and 0 for D2 . Analogously the upper bounds are chosen as
from [23]. We first point out that from (93) it follows that |Z (t )|, twice the real values of the delays i.e., D1 = 0.8 and D2 = 1.6.
�w(t )�, �ξ (t )�, D̂1 and D̂2 are uniformly bounded. Moreover, The initial conditions are chosen as X1 (0) = 0.5, X2 (0) = 0.5 and
using (57) and (58) we get the uniform boundness of |X (t )| and X2 (θ ) = 0.5, ∀θ ∈ [−D1 , 0], and finally D̂1 (0) = D1 = 0.1. The
�u(t )�. Using (23) it follows that U (t ) is uniformly bounded. From � chosen as a2 = 200, k �= 0.005, γ1 = 25,
controller parameters are
dX 2 (t )
(24)–(25), (83)–(84) and (50) we conclude that dt is uniformly γ2 = 25 and K (D̂1 ) = −10.0625e0.25D̂1 −8.5 .
bounded. Finally, since from (92) it turns out that |Z (t )| and �ξ (t )�
Figs. 1–3 show two distinct simulations, starting from two
are square integrable, using (58) and the alternative to Barbalat’s
extreme initial values for the input delay estimate, one at zero, and
Lemma from [23], we conclude that limt →∞ X (t ) = 0. We now
the other at twice the true delay value. In Figs. 1 and 3 we observe
turn our attention to proving the convergence of U (t ). Using (92)
that, as Theorem 1 predicts, convergence to zero is achieved for the
we also get that �w(t )� is square integrable in time. Thus, with the
states and the input, despite starting with initial estimate for the
help of (57) and by the square integrability of |Z (t )| we conclude
input delay at the two extreme values and for the state delay at the
using (23) that U (t ) is square integrable. It only remains to show
lower bound. In Fig. 2 one can see the evolution of the estimates for
dU 2 (t )
that dt
is uniformly bounded. Hence, it is sufficient to show that the two distinct simulation cases. The estimates for the two delays,
˙
U̇ (t ) is uniformly bounded. From (23) one can observe that since D̂1 after a transient response, converge to stabilizing for the system
˙ values.
and D̂2 are uniformly bounded, with the help of (16) and (19) we
dU 2 (t )
conclude the uniform boundness of dt
.
6. Conclusions
5. Simulations
In this paper we presented an adaptive control design for a
We give here a simulation example to illustrate the effective- system in feedforward form with simultaneous unknown input
ness of the proposed adaptive scheme. We choose a second order and state delays. The design of the controller is based on predictor
feedforward system with parameters F1 = F2 = 0.25, H1 = 1, feedback. The update laws for the estimation of the unknown
D1 = 0.4 and D2 = 0.8. This is an unstable system with two poles delays are based on the construction of a Lyapunov function with
 N. Bekiaris-Liberis, M. Krstic / Systems & Control Letters 59 (2010) 277–283 283

[14] S. Diop, I. Kolmanovsky, P.E. Moraal, M. van Nieuwstadt, Preserving stabil-

ity/performance when facing an unknown time-delay, Control Engineering
Practice 9 (2001) 1319–1325.
[15] M. Krstic, A. Banaszuk, Multivariable adaptive control of instabilities arising in
jet engines, Control Engineering Practice 14 (2006) 833–842.
[16] J-P. Richard, Time-delay systems: an overview of some recent advances and
open problems, Automatica 39 (2003) 1667–1694.
[17] S. Evesque, A.M. Annaswamy, S. Niculescu, A.P. Dowling, Adaptive control
of a class of time-delay systems, ASME Transactions on Dynamics, Systems,
Measurement, and Control 125 (2003) 186–193.
[18] S.-I. Niculescu, A.M. Annaswamy, An adaptive Smith- controller for time-delay
systems with relative degree n ≥ 2, Systems and Control Letters 49 (2003)
347–358.
[19] D. Bresch-Pietri, Miroslav Krstic, Delay-adaptive full-state predictor feedback
for systems with unknown long actuator delay, in: 2009 American Control
Conference, St. Louis, Missouri, 2009.
[20] D. Bresch-Pietri, Miroslav Krstic, Adaptive trajectory tracking despite un-
known input delay and plant parameters, Automatica 45 (2009) 2074–2081.
[21] M. Krstic, A. Smyshlyaev, Adaptive boundary control for unstable parabolic
PDEs—Part I: Lyapunov design, IEEE Transactions on Automatic Control 53
(2008) 1575–1591.
[22] K.P. Bhat, H.N. Koivo, Model characterization of controllability and observ-
ability in time-delay systems, IEEE Transactions on Automatic Control 21 (2)
(1976) 292–293.
Fig. 3. The control effort for the cases D̂2 (0) = 0 and D̂2 (0) = 1.6. [23] W.-J. Liu, M. Krstic, Adaptive control of Burgers’ equation with unknown
viscosity, International Journal of Adaptive Control and Signal Processing 15
(2001) 745–766.
normalization. Convergence to zero is then proved using the linear [24] K. Gu, S.-I. Niculescu, Survey on recent results in the stability and control of
boundness of the relative signals. time-delay systems, ASME Transactions on Dynamics, Systems, Measurement,
and Control 125 (2003) 158–165.
[25] M. Jankovic, Control Lyapunov–Razumikhin functions and robust stabilization
References of time delay systems, IEEE Transactions on Automatic Control 46 (2001)
1048–1060.
[1] Z. Artstein, Linear systems with delayed controls: a reduction, IEEE [26] M. Jankovic, Control of nonlinear systems with time delay, in: 2003 IEEE
Transactions on Automatic Control 27 (1982) 869–879. Conference on Decision and Control, Maui, Hawai, 2003.
[2] Y.A. Fiagbedzi, A.E. Pearson, Feedback stabilization of linear autonomous time [27] M. Jankovic, Control of cascade systems with time delay—The integral cross-
term approach, in: 2006 IEEE Conference on Decision and Control, San Diego,
lag systems, IEEE Transactions on Automatic Control 31 (1986) 847–855.
CA, 2006.
[3] M. Jankovic, Forwarding, backstepping, and finite spectrum assignment for
[28] I. Karafyllis, Finite-time global stabilization by means of time-varying
time delay systems, in: 2006 American Control Conference, Minneapolis,
distributed delay feedback, Siam Journal on Control and Optimization 45 (1)
Minnesota, 2006.
(2006) 320–342.
[4] M. Jankovic, Recursive predictor design for linear systems with time delay, in:
[29] I. Karafyllis, Z.P. Jiang, Control Lyapunov functionals and robust stabilization
2008 American Control Conference, Seattle, Washington, 2008. of nonlinear time-delay systems, in: 2008 IEEE Conference on Decision and
[5] M. Krstic, On compensating long actuator delays in nonlinear control, IEEE Control, Cancun, Mexico, 2008.
Transactions on Automatic Control 53 (2008) 1684–1688. [30] M. Krstic, A. Smyshlyaev, Backstepping boundary control for first-order
[6] M. Krstic, Input delay compensation for forward complete and feedforward hyperbolic PDEs and application to systems with actuator and sensor delays,
nonlinear systems, IEEE Transactions on Automatic Control 55 (2010) Systems and Control Letters 57 (2008) 750–758.
287–303. [31] F. Mazenc, S. Mondie, S.I. Niculescu, Global asymptotic stabilization for chains
[7] W.H. Kwon, A.E. Pearson, Feedback stabilization of linear systems with delayed of integrators with a delay in the input, IEEE Transactions on Automatic
control, IEEE Transactions on Automatic Control 25 (1980) 266–269. Control 48 (1) (2003) 57–63.
[8] S. Mondie, W. Michiels, Finite spectrum assignment of unstable time-delay [32] F. Mazenc, S. Mondie, R. Francisco, Global asymptotic stabilization of
systems with a safe implementation, IEEE Transactions on Automatic Control feedforward systems with delay at the input, IEEE Transactions on Automatic
48 (2003) 2207–2212. Control 49 (2004) 844–850.
[9] A.W. Olbrot, Stabilizability, detectability, and spectrum assignment for [33] F. Mazenc, P.-A. Bliman, Backstepping design for timedelay nonlinear systems,
linear autonomous systems with general time delays, IEEE Transactions on IEEE Transactions on Automatic Control 51 (2004) 149–154.
Automatic Control 23 (1978) 887–890. [34] L. Mirkin, On the approximation of distributed-delay control laws, Systems
[10] N. Bekiaris-Liberis, M. Krstic, On stabilizing strict-feedback linear systems and Control Letters 51 (2004) 331–342.
with delayed integrators, in: 2009 Mediterranean Conference on Control and [35] O.J.M. Smith, A controller to overcome dead time, ISA Transactions 6 (1959)
Automation, Thessaloniki, Greece, 2009. 28–33.
[11] J.J. Loiseau, Algebraic tools for the control and stabilization of time-delay [36] Q.-C. Zhong, L. Mirkin, Control of integral processes with dead time—Part 2:
systems, Annual Reviews in Control 24 (2000) 135–149. quantitative analysis, IEE Proceedings-Control Theory and Applications 149
[12] A.Z. Manitius, A.W. Olbrot, Finite spectrum assignment for systems with (2002) 291–296.
delays, IEEE Transactions on Automatic Control 24 (1979) 541–553. [37] Q.-C. Zhong, On distributed delay in linear control laws—Part I: discrete-
[13] K. Watanabe, E. Nobuyama, T. Kitamori, M. Ito, A New Algorithm for delay implementation, IEEE Transactions on Automatic Control 49 (2006)
Finite Spectrum Assignment of Single-Input Systems with Time Delay, IEEE 2074–2080.
Transactions on Automatic Control 37 (1992) 1377–1383. [38] Q.-C. Zhong, Robust Control of Time-delay Systems, Springer, 2006.