Automatica 116 (2020) 108902

Contents lists available at ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

Brief paper

Delay-adaptive control for linear systems with distributed input

delays✩
∗
Yang Zhu a , Miroslav Krstic b , Hongye Su c ,
a
School of Electrical Engineering, Tel-Aviv University, Israel
b
Department of Mechanical and Aerospace Engineering, University of California, San Diego, USA
c
State Key Laboratory of Industrial Control Technology, Institute of Cyber-Systems and Control, Zhejiang University, China

article info a b s t r a c t

Article history: A majority of existing literature on adaptive method of time-delay systems concentrate on uncertainty
Received 13 April 2019 in plant parameters or discrete input delays. This paper proposes a systematic adaptive control
Received in revised form 30 September 2019 approach to solve stabilization problems of linear systems with unknown distributed input delays.
Accepted 31 January 2020
Under the rescaled unity-interval notation, the uncertain delay leads the input vector to consist of
Available online xxxx
unknown functions and unknown parameters as well. To resolve the coexistent uncertainties in delay
Keywords: and input vector, a reduction-based change of variable and a backstepping–forwarding transformation
Delay of the finite-dimensional plant state and the infinite-dimensional actuator state are introduced. Making
Adaptive use of these conversions, the certainty-equivalence-based control law and the Lyapunov-based update
Predictor law are developed for adaptive stabilization.
Distributed delays
© 2020 Elsevier Ltd. All rights reserved.

1. Introduction are known, is a more difficult problem than all of delay-free

1.1. Literature review
This year is the 60th anniversary of Otto J. Smith’s 1959 publi-
cation (Smith, 1959) of a control design idea for the compensation
of input delays, which is commonly referred to as the Smith
predictor. Over the past 60 years, the predictor feedback has been
demonstrated to be quite effective in compensating large delays,
and major breakthroughs have been successively reported, such
as the framework of the ‘‘reduction’’ approach in Artstein (1982).
The standard predictor-based compensation for input delay has
the premise that the delay value of every actuator channel is
known. However, the fact is that such a prior knowledge of delay
length may be hard to acquire in practice. Early work on adaptive
control of time-delay systems mainly focused on the uncertainty
of plant parameters rather than actuator delays (Bekiaris-Liberis,
Jankovic, & Krstic, 2013; Niculescu & Annaswamy, 2003; Ortega
& Lozano, 1988; Yildiz, Annaswamy, Kolmanovsky, & Yanakiev,
2010; Zhou, Wen, & Wang, 2009). One could argue that the situ-
ation with a highly uncertain delay, even if the plant parameters
✩ This work was supported by National Natural Science Foundation of China
(61621002). The material in this paper was not presented at any conference.
This paper was recommended for publication in revised form by Associate Editor
Tong Zhou under the direction of Editor Richard Middleton.
∗ Corresponding author.
E-mail addresses: yangzhu@mail.tau.ac.il (Y. Zhu), krstic@ucsd.edu
(M. Krstic), hysu@iipc.zju.edu.cn (H. Su).
adaptive control of the 1970–1990s.
Due to the infinite-dimensional nature of time-delay systems,
conventional adaptive control methods for finite-dimensional
ordinary differential equation (ODE) systems (Goodwin & Sin,
2014) cannot be directly applied to address uncertain delays.
Instead, borrowing ideas from adaptive control approaches for
infinite-dimensional partial differential equation (PDE) systems
(Anfinsen & Aamo, 2019; Smyshlyaev & Krstic, 2010), a few of
progresses on adaptive control to handle unknown discrete input
delays were recently reported in Bresch-Pietri, Chauvin, and Petit
(2012), Bresch-Pietri and Krstic (2009), Zhu, Krstic, and Su (2017,
2018a, 2018b, 2019) and Zhu, Su, and Krstic (2015).
Unfortunately, the results for unknown discrete input de-
lays (Bresch-Pietri et al., 2012; Bresch-Pietri & Krstic, 2009; Zhu
et al., 2017, 2018a, 2018b, 2019, 2015) are not applicable to
the case of unknown distributed input delays, as the finite-
dimensional state of the plant and the infinite-dimensional
actuator state are not in the strict-feedback form. References
(Bekiaris-Liberis & Krstic, 2011) provided a novel backstepping–
forwarding transformation for stability analysis of linear systems
with distributed input delays. However, it assumes the delay to
be known.
1.2. Contribution and organization
In this paper, we proposes an adaptive control scheme to
deal with stabilization problems of linear systems with unknown

https://doi.org/10.1016/j.automatica.2020.108902
0005-1098/© 2020 Elsevier Ltd. All rights reserved.
2 Y. Zhu, M. Krstic and H. Su / Automatica 116 (2020) 108902

distributed input delays. To lay a foundation for adaptive sta- Taking the time-derivative of (6) along (3)–(5) and using the
bilization, a predictor-based feedback scheme in the rescaled integration by parts in x, we get
unity-interval notation under the known delay is provided in ∫ 1
Section 2, in which the control scheme is parameterizable in Ż (t) = AX (t) + D B(Dx)u(x, t)dx
delay and the boundary of spatial variable of PDE is fixed. 0
Section 3 handles the unknown delay. When the delay is uncer-
∫ 1 ∫ x

tain, a key challenge is that the input vector may also contain +D e−AD(x−y) B(Dy)dyux (x, t)dx
0 0
unknown functions and unknown parameters. Thus two different ∫ 1
cases of the input vector are taken into account: (1) the input = AX (t) + D B(Dx)u(x, t)dx
vector is a constant vector independent of delay (Section 3.1), 0
(2) the input vector is a continuous vector-valued function of ∫ 1
delay (Section 3.2). To solve the coexistent uncertainties in de- +D e−AD(1−y) B(Dy)dyu(1, t)
lay and input vector, the reduction-based change of variable 0
∫ 1
and the backstepping–forwarding transformation, of certainty-
equivalence type, are introduced to convert the ODE–PDE cascade −D B(Dx)u(x, t)dx
0
consisting of the finite-dimensional plant state and the infinite- ∫ 1∫ x
dimensional actuator state into a ‘target’ system. Making use of + D2 Ae−AD(x−y) B(Dy)dyu(x, t)dx
these conversions, the certainty-equivalence-based control law 0 0
and the Lyapunov-based update law are developed for adaptive
∫ 1
stabilization. = AZ (t) + D e−AD(1−x) B(Dx)dxU(t) (7)
0
Notation: (
• For a finite-dimensional ODE vector X (t), its Euclid norm is Assumption 1. For the system (3) and (7), the pair A, D
)
denoted by |X (t)|. ∫1
e−AD(1−x) B(Dx)dx is stabilizable. There exist a vector K to
• For an infinite-dimensional PDE scalar function u(x, t), u : 0
∫1 −AD(1−x)
[0, 1] × R+ → R, make A + D 0
e B(Dx)dxK Hurwitz. Namely, there exist
(∫ 1 )1/2 matrices P = P T > 0 and Q = Q T > 0 such that
∥u(x, t)∥ = u(x, t)2 dx ■ ( ∫ 1
)T
0
A+D e−AD(1−x) B(Dx)dxK P
0
2. Predictor feedback under known delay ( ∫ 1
)
−AD(1−x)
+P A+D e B(Dx)dxK = −Q ■ (8)
Consider linear systems with distributed input delays as fol- 0
lows:
The controller is designed as
∫ D
Ẋ (t) = AX (t) + B(D − σ )U(t − σ )dσ (1) U(t) = u(1, t) = KZ (t) (9)
0

where X (t) ∈ Rn is the plant state, U(θ ) ∈ R for θ ∈ [t − D, t ] Thus the system (7) becomes
is the actuator state, U(t) ∈ R is the control input, D > 0 is the Ż (t) = Acl Z (t) (10)
constant delay, and A, B are the system matrix and input vector
of appropriate dimensions, respectively. The input vector B(·) is where
a continuous real-valued vector function defined on [0, D]. For ∫ 1
notational simplicity, the system is assumed to be single input. Acl = A + D e−AD(1−x) B(Dx)dxK (11)
The results of this paper can be straightforwardly extended to the 0

multi-input case, when the delays are the same in each individual For stability analysis, the invertible backstepping–forwarding
input channel. transformation is brought in as
In this section, we consider the simplest stabilization problem
of (1) by assuming the delay D is known.
w(x, t) = u(x, t) − KeAcl D(x−1) Z (t) (12)
Through a multi-variable function in rescaled unity interval Take the partial derivatives of (12) with respect to x and t,
notation respectively,
u(x, t) = U(t + D(x − 1)) = U(t − σ ), x ∈ [0, 1], σ ∈ [0, D], (2) wx (x, t) = ux (x, t) − KeAcl D(x−1) Acl DZ (t) (13)
the system (1) is converted into the ODE–PDE cascade wt (x, t) = ut (x, t) − Ke Acl D(x−1)
Acl Z (t) (14)
∫ 1
Multiplying (14) with D and minus (13), we have
Ẋ (t) = AX (t) + D B(Dx)u(x, t)dx (3)
0
Dwt (x, t) = wx (x, t) (15)
Dut (x, t) = ux (x, t), x ∈ [0, 1] (4)
Combining (10) with (15), substituting x = 1 into (12) and
u(1, t) = U(t) (5) utilizing (9), the target system for analysis is obtained as follows:
where ut (x, t) and ux (x, t) denote the partial derivatives of u(x, t) Ż (t) = Acl Z (t) (16)
with respect to t and x, respectively. The finite-dimensional ODE
state X (t) and infinite-dimensional PDE state u(x, t) for x ∈ [0, 1] Dwt (x, t) = wx (x, t), x ∈ [0, 1] (17)
are assumed to be measurable. It is evident (2) is a solution of the w(1, t) = 0 (18)
transport PDE (4)–(5). The control objective is to stabilize (3)–(5).
The reduction-based change of variable is introduced as
Remark 1. As illustrated in Fig. 1, a few of conversions are
∫ 1 ∫ x
employed in above control scheme. Firstly, through the multi-
Z (t) = X (t) + D 2
e −AD(x−y)
B(Dy)dyu(x, t)dx (6) variable function (2), the original system with distributed input
0 0
 Y. Zhu, M. Krstic and H. Su / Automatica 116 (2020) 108902 3

Fig. 1. Conversion between the original system (1) and target system (16)–(18).

∫ 1
delay (1) is represented by the ODE–PDE cascade (3)–(5). Sec-
= −Z T (t)QZ (t) + 2(1 + x)w (x, t)wx (x, t)dx
ondly, by the reduction-based change of variable (6), the stabiliza- 0
tion problem of ODE plant (3) is reduced to the stabilization prob-
= −Z T (t)QZ (t) + 4w (1, t)2 − 2w(0, t)2
lem of delay-free system (7). Finally, under the backstepping– ∫ 1 ∫ 1
forwarding transformation (12) and the control law (9), the ODE
− 2w (x, t) dx −
2
2(1 + x)wx (x, t)w (x, t)dx
(7) and PDE (4)–(5) are transformed into the target system (16)– 0 0
(18), which is convenient for stability analysis. ■ ∫ 1
= −Z T (t)QZ (t) − w(0, t)2 − w (x, t)2 dx
Remark 2. An alternative representation of the control scheme ∫ 1
0

(2)–(18) is listed as follows:

≤ −λmin (Q )|Z (t)|2 − w(x, t)2 dx
u(x, t) = U(t + x − D) = U(t − σ ), x ∈ [0, D], σ ∈ [0, D], (19) 0
λmin (Q )
∫ 1
D 1
λmax (P)|Z (t)|2 − (1 + x)w (x, t)2 dx
∫
≤− D
Ẋ (t) = AX (t) + B(x)u(x, t)dx (20) λmax (P) 2D 0
0
λmin (Q ) 1
{ }
ut (x, t) = ux (x, t), x ∈ [0, D] (21) ≤ − min , V (t) (30)
λmax (P) 2D
u(D, t) = U(t) = KZ (t) (22)
It is evident that (30) implies∫ the exponential stability in the
∫ D ∫ x 1
Z (t) = X (t) + e −A(x−y)
B(y)dyu(x, t)dx (23) sense of the norm |X (t)|2 + 0 w (x, t)2 dx by the comparison
0 0 principle. Then making use of the inverse transformations of
w(x, t) = u(x, t) − KeAcl (x−D) Z (t) (24) (6) and (12), the exponential stability in the sense of the norm
∫1
∫ D |X (t)|2 + 0
u(x, t)2 dx is derived. ■
Ż (t) = Acl Z (t), Acl = A + e−A(D−x) B(x)dxK (25)
0 3. Predictor feedback under unknown delay
wt (x, t) = wx (x, t), x ∈ [0, D] (26)
In this section, we consider that the delay D is unknown.
w(D, t) = 0 (27) Since D is uncertain, the vector B(Dx) for x ∈ [0, 1] in (3) may
It is evident that (19)–(27) is equivalent to (2)–(18). The main contain unknown variables so that it cannot be employed directly
difference is that (2)–(18) is parameterized in D, whereas (19)– for control design. Two different cases are taken into account as
(27) is non-parameterized in D. When the delay D is unknown follows:
and a time-varying signal D̂(t) is employed to estimate D, it • B(Dx) is a constant vector independent of Dx,
is inconvenient for the adaptive control to be applied to (19)– • B(Dx) is a continuous vector-valued function of Dx.
(27). For example, a moving boundary u(D̂(t), t) appears in (24)
which renders the boundary condition (27) non-homogeneous 3.1. B(Dx) is a constant vector independent of Dx
such that w (D̂(t), t) = u(D̂(t), t) − u(D, t) ̸ = 0. The estimate
appears in the limit of integration of (23) such that Z (t) = X (t) + This section addresses the relatively simple case where the
vector B(Dx) for x ∈ [0, 1] in (3) (i.e., B(D − σ ) for σ ∈ [0, D]
∫ D̂(t) ∫ x
0
e−A(x−y) B(y)dyu(x, t)dx, which makes it difficult to get the
0
error D̃(t) = D − D̂(t) for the estimator design. That is the reason in (1)) is a known constant vector independent of Dx such that
why the feedback scheme in rescaled unity-interval notation (2)– B(Dx) = B(D − σ ) = B (31)
(18) is introduced to lay a foundation for delay-adaptive control,
rather than applying adaptive control directly to (19)–(27). ■ Then the system (1) is reduced to
∫ D
Theorem 1. The closed-loop system consisting of the plant (3)–(5) Ẋ (t) = AX (t) + B U(t − σ )dσ (32)
and the controller (9) is exponentially stable in the sense of the norm 0

∫ 1 and the ODE–PDE cascade (3)–(5) is accordingly reduced to

|X (t)|2 + u2 (x, t)dx ■ (28) ∫ 1
0 Ẋ (t) = AX (t) + DB u(x, t)dx (33)
0
Proof. Consider the Lyapunov candidate Dut (x, t) = ux (x, t) (34)
1
u(1, t) = U(t)
∫
(35)
V (t) = Z T (t)PZ (t) + D (1 + x)w (x, t)2 dx (29)
0 The control objective is to stabilize (33)–(35) when the delay D is
Bearing (8) and (18) in mind, taking the time derivative of (29) unknown.
along the target system (16)–(18) and using the integration by
parts in x, we get Remark 3. On the basis of the framework (2)–(18), the key
T idea of certainty-equivalence-based adaptive control is to use an
ATcl P
( )
V̇ (t) = Z (t) + PAcl Z (t) estimate to replace the unknown delay in (6), (9) and (12). And
∫ 1
the Lyapunov-based update law is employed to cancel the esti-
+D 2(1 + x)w (x, t)wt (x, t)dx mation error term in the time-derivative of Lyapunov function.
0
4 Y. Zhu, M. Krstic and H. Su / Automatica 116 (2020) 108902

A non-trivial problem is to deal with the unknown D2 in (6). hD1 (x, t) = K (β̂ (t))eAcl (β̂ (t))D̂(t)(x−1) fD1 (t) (51)
One intuitive method is to estimate D2 by D̂(t)2 where D̂(t) is ∫ 1
an estimate of D. The other possible method is to treat D2 as a fD (t) = D̂1 (t) Bu(x, t)dx
whole body such that θ = D2 and estimate θ by θ̂ (t). However, 0
∫ 1
both of methods will produce the mismatch term for update law
− D̂1 (t) e−AD̂(t)(1−x) Bdxu(1, t)
design. The available method is to regard D2 as a product of two 0
parameters such that D2 = D · D1 where D1 = D and estimate D ∫ 1 ∫ x
and D1 with different update laws, respectively. ■ − D̂(t)D̂1 (t) Ae−AD̂(t)(x−y) Bdyu(x, t)dx (52)
0 0
For using projector operators in the update law later, we make ( )
the following assumption. hD (x, t) = K (β̂ (t))eAcl (β̂ (t))D̂(t)(x−1) fD (t) + Acl (β̂ (t))Z (t) (53)

and the projector operator is defined as

Assumption 2. There exist known constants D and D such that
⎨0, τ <0
⎧
D̂1 (t) = D and
0 < D ≤ D = D1 ≤ D ■ (36) Proj[D,D] {τ } = 0, D̂1 (t) = D and τ >0 (54)
τ,
⎩
In order to stabilize (33)–(35), the following assumption is else
required.
⎨0, and τ < 0
⎧
D̂(t) = D
Assumption 3. For the system (33), the pair (A, β) is stabilizable Proj[D,D] {τ } = 0, D̂(t) = D and τ >0 (55)
τ,
⎩
where else
∫ 1 ∫ 1
β=D e −AD(1−x)
Bdx = D1 e−AD(1−x) Bdx (37) Theorem 2. Consider the closed-loop system consisting of the plant
0 0 (33)–(35) and the adaptive controller (41)–(55). All the states (X (t),
There exists a vector K (β ) to make A + β K (β ) Hurwitz. Namely, u(x, t), D̂(t), D̂1 (t)) of the closed-loop system are globally bounded
there exist matrices P(β ) = P(β )T > 0 and Q (β ) = Q (β )T > 0 and the regulation of X (t) and U(t) such that limt →∞ X (t) =
such that limt →∞ U(t) = 0 is achieved. ■
(A + β K (β ))T P(β ) + P(β ) (A + β K (β )) = −Q (β ) ■ (38)
Proof. Taking the time-derivative of (43) along (33)–(35), and
Denote D̂(t) and D̂1 (t) as the estimates of D and D1 with using the integration by parts in x, we obtain
estimation errors satisfying ∫ 1
˙ ˙
( )
Ż (t) = AX (t) + D Bu(x, t)dx + φ D̂(t), D̂1 (t)
D̃(t) = D − D̂(t) (39) 0
∫ 1 ∫ x
D̃1 (t) = D1 − D̂1 (t) = D − D̂1 (t) (40) D̂(t)D̂1 (t)
+ e−AD̂(t)(x−y) Bdyux (x, t)dx
The delay-adaptive control scheme is listed as follows: D 0 0
∫ 1
The control law is ˙ ˙
( )
= AX (t) + D Bu(x, t)dx + φ D̂(t), D̂1 (t)
U(t) = u(1, t) = K (β̂ (t))Z (t) (41) 0
∫ 1
D̂(t)D̂1 (t)
where K (β̂ (t)) is chosen to let + e−AD̂(t)(1−y) Bdyu(1, t)
D 0
∫ 1 ∫ 1
Acl (β̂ (t)) = A + D̂1 (t) e−AD̂(t)(1−x) Bdx K (β̂ (t)) (42) −
D̂(t)D̂1 (t)
Bu(x, t)dx
0 D
   0
β̂ (t)
∫ 1 ∫ x
D̂(t)D̂1 (t)
+ AD̂(t)e−AD̂(t)(x−y) Bdyu(x, t)dx
be Hurwitz and D 0 0
∫ 1 ∫ x
∫ 1
˙ ˙
( )
Z (t) = X (t) + D̂(t)D̂1 (t) e−AD̂(t)(x−y) Bdyu(x, t)dx (43) = AX (t) + D Bu(x, t)dx + φ D̂(t), D̂1 (t)
0 0 0
( ) ∫ 1
The update laws are D̃(t)
+ 1− D̂1 (t) e−AD̂(t)(1−x) Bdxu(1, t)
˙ D
D̂1 (t) = γD1 Proj[D,D] {τD1 (t)}, γD1 > 0 (44) ( )
0
∫ 1
1 D̃(t)
1/gZ T (t)P(β̂ (t))fD1 (t) − 0 (1 + x)w (x, t)hD1 (x, t)dx
∫
− 1− D̂1 (t) Bu(x, t)dx
τD1 (t) = (45) D
1 + Ξ (t) ( )
0

˙
D̂(t) = γD Proj[D,D] {τD (t)}, γD > 0 (46) D̃(t)
+ 1− D̂1 (t)D̂(t)
∫1 D
1/gZ (t)P(β̂ (t))fD (t) − 0 (1 + x)w (x, t)hD (x, t)dx
T
τD (t) = (47) ∫ 1 ∫ x
1 + Ξ (t) × Ae−AD̂(t)(x−y) Bdyu(x, t)dx (56)
where P(β̂ (t)) satisfies (38) and g > 0 is a designing coefficient,
0 0
where
w(x, t) = u(x, t) − K (β̂ (t))eAcl (β̂ (t))D̂(t)(x−1) Z (t) (48) ˙ ˙
( )
∫ 1 φ D̂(t), D̂1 (t)
Ξ (t) = Z (t)P(β̂ (t))Z (t) + g
T
(1 + x)w (x, t)2 dx (49)
∫ 1 ∫ x [(
˙ ˙ ˙
) ]
0 = D̂(t)D̂1 (t) + D̂(t)D̂1 (t) I − D̂(t)D̂1 (t)AD̂(t)(x − y)
∫ 1 0 0
fD1 (t) = Bu(x, t)dx (50) × e−AD̂(t)(x−y) Bdyu(x, t)dx (57)
0
 Y. Zhu, M. Krstic and H. Su / Automatica 116 (2020) 108902 5

1
∫
Under the control law (41), the formula (56) becomes ˙ ˙
( )
− 2Dg (1 + x)w (x, t)ψ D̂(t), D̂1 (t) dx
0
D̃(t)
Ż (t) = Acl (β̂ (t))Z (t) + D̃1 (t)fD1 (t) +
∫ 1
fD (t) (
D − 2g (1 + x)w (x, t) DD̃1 (t)hD1 (x, t)
˙ ˙
( )
+ φ D̂(t), D̂1 (t) (58) 0
]
)
+D̃(t)hD (x, t) dx
where Acl (β̂ (t)) is given in (42), fD1 (t) and fD (t) have been defined
in (50) and (52).
2gD ˙ 2g ˙
Taking partial derivatives of (48) with respect to x and t, − D̃1 (t)D̂1 (t) − D̃(t)D̂(t)
respectively, we get γD1 γD
[
Acl (β̂ (t))D̂(t)(x−1) 1
wx (x, t) = ux (x, t) − K (β̂ (t))e = −DZ T (t)Q (β̂ (t))Z (t)
1 + Ξ (t)
× Acl (β̂ (t))D̂(t)Z (t) (59)
(
˙ ˙
) − g w(0, t)2 − g ∥w(x, t)∥2
wt (x, t) = ut (x, t) − ψ D̂(t), D̂1 (t)
˙ ˙
( )
+ DZ T (t)ϕ D̂(t), D̂1 (t) Z (t)
− K (β̂ (t))eAcl (β̂ (t))D̂(t)(x−1)
˙ ˙
( )
(
D̃(t)
) + 2DZ T (t)P(β̂ (t))φ D̂(t), D̂1 (t)
× Acl (β̂ (t))Z (t) + D̃1 (t)fD1 (t) + fD (t) (60) ]
D ∫ 1
˙ ˙
( )
− 2Dg (1 + x)w (x, t)ψ D̂(t), D̂1 (t) dx
where 0
˙ ˙
( )
ψ D̂(t), D̂1 (t) 2gD ˙
( )
− D̃1 (t) D̂1 (t) − γD1 τD1 (t)
( ) γD1
∂ K (β̂ (t)) ˙ ∂ K (β̂ (t)) ˙
D̂1 (t) eAcl (β̂ (t))D̂(t)(x−1) Z (t) 2g ˙
( )
= D̂(t) +
∂ D̂(t) ∂ D̂1 (t) − D̃(t) D̂(t) − γD τD (t) (65)
γD
+ K (β̂ (t))eAcl (β̂ (t))D̂(t)(x−1) where τD1 (t) )and τD (t) have been defined in (45) and (47), and
˙ ˙ ∂ P(β̂ (t)) ˙ ∂ P(β̂ (t)) ˙
(
ϕ D̂(t), D̂1 (t) =
[( )
∂ Acl (β̂ (t)) ˙ ∂ Acl (β̂ (t)) ˙ D̂(t) + D̂ (t).
∂ D̂(t) ∂ D̂1 (t) 1
× D̂(t) + D̂1 (t) D̂(t)
∂ D̂(t) ∂ D̂1 (t) Please note that D̂1 (t) and D̂(t) are bounded as the projector
] operators ensure them to stay in the interval (36). Making use of
˙ Young’s and Cauchy–Schwarz inequalities, it is evident that the
+ Acl (β̂ (t))D̂(t) (x − 1)Z (t)
inverse transformation of (48) implies
∥u(x, t)∥2 ≤ Mu |Z (t)|2 + ∥w(x, t)∥2
( )
˙ ˙
( )
+ K (β̂ (t))eAcl (β̂ (t))D̂(t)(x−1) φ D̂(t), D̂1 (t) (61) (66)
where Mu > 0 is a constant.
Multiplying (60) by D and minus (59), and employing the control ε2
Utilizing (66) and inequalities 0 < < 1 and 0 < < 1,
|ε|
law (41), we get 1+ε 2 1+ε 2
it is easy to show that
˙ ˙
( )
Dwt (x, t) = wx (x, t) − Dψ D̂(t), D̂1 (t) ⏐
⏐˙
⏐ |Z (t)|2 + ∥w(x, t)∥2
⏐D̂1 (t)⏐ ≤ γD1 MD1 ≤ γD1 M̄D1 (67)
⏐
− DD̃1 (t)hD1 (x, t) − D̃(t)hD (x, t) (62) 1 + Ξ (t)
⏐
⏐˙ ⏐
⏐ |Z (t)|2 + ∥w(x, t)∥2
w(1, t) = 0, x ∈ [0, 1] (63) ⏐D̂(t)⏐ ≤ γD MD ≤ γD M̄D (68)
1 + Ξ (t)
where hD1 (x, t) and hD (x, t) have been defined in (51) and (53).
Build the Lyapunov candidate such that where MD1 , MD , M̄D1 and M̄D are positive constants.
Thus we have
gD g
V (t) = D log 1 + Ξ (t) + D̃1 (t)2 + D̃(t)2
( ) [
(64) 1
γD1 γD V̇ (t) ≤ −Dλmin (Q )|Z (t)|2 − g w(0, t)2 − g ∥w(x, t)∥2
1 + Ξ (t)
where Ξ (t) has been defined in (49). ]
+ (γD1 + γD )M |Z (t)| + ∥w(x, t)∥
2 2
( )
Taking the time derivative of (64) along the target closed-loop (69)
system (58) and (62)–(63), we have
[ where M > 0 is a constant. By carefully selecting design coeffi-
1 cients λmin (Q ), g, γD1 and γD , it is easy to get
DZ (t) P(β̂ (t))Acl (β̂ (t))
T
(
V̇ (t) =
1 + Ξ (t) N
[ ]
V̇ (t) ≤ −|Z (t)|2 − w(0, t)2 − ∥w(x, t)∥2 (70)
+ ATcl (β̂ (t))P(β̂ (t)) Z (t) 1 + Ξ (t)
)

˙ ˙ where N > 0 is a constant. Thus Z (t) and w (x, t) are bounded

( )
+ DZ T (t)ϕ D̂(t), D̂1 (t) Z (t)
and converge to zero. By inverse conversions of (43) and (48), the
original states X (t) and u(x, t) are bounded and converge to zero.
( ( )
˙ ˙
+ 2Z T (t)P(β̂ (t)) Dφ D̂(t), D̂1 (t)
Then the proof of Theorem 2 is completed. ■
)
+ DD̃1 (t)fD1 (t) + D̃(t)fD (t) 3.2. B(Dx) is a continuous vector-valued function of Dx
∫ 1
This section is concerned with the more challenging case
+ 2g (1 + x)w (x, t)wx (x, t)dx where the n-dimensional input vector B(Dx) for x ∈ [0, 1] in (3)
0
6 Y. Zhu, M. Krstic and H. Su / Automatica 116 (2020) 108902

is a continuous function of Dx such that The delay-adaptive control scheme is designed as follows:
⎡
ρ1 (Dx)
⎤ The control law is
⎢ρ2 (Dx)⎥ U(t) = u(1, t) = K (β̂ (t))Z (t) (83)
⎣ .. ⎦
B(Dx) = ⎢ (71)
⎥
. where K (β̂ (t)) is chosen to let
ρn (Dx) ∫ 1
where ρi (Dx) for i = 1, . . . , n are unknown components of the Acl (β̂ (t)) = A + e−AD̂(t)(1−x) B̂ (x, t)dxK (β̂ (t))
vector-valued function B(Dx). 0
1 n
On the basis of (71), we further denote
∫ ∑
=A+ e−AD̂(t)(1−x) b̂i (x, t)Bi dx K (β̂ (t)) (84)
n n
∑ ∑ 0 i=1
B (x) = DB(Dx) = Dρi (Dx)Bi = bi (x)Bi (72)   
i=1 i=1 β̂ (t)

where bi (x) = Dρi (Dx) for i = 1, . . . , n are unknown scalar be Hurwitz and
continuous functions of x, and Bi ∈ Rn for i = 1, . . . , n are the ∫ 1 ∫ x
unit vectors accordingly. A three-dimensional example of (72) is Z (t) = X (t) + D̂(t) e−AD̂(t)(x−y) B̂ (y, t)dyu(x, t)dx (85)
given below. 0 0

The update laws are

Example 1.
˙
⎡ 1
⎤ D̂(t) = γD Proj[D,D] {τD (t)}, γD > 0 (86)
1 0
[ ] [ ]
Dx+1 D ∫1
DB(Dx) = D ⎣sin Dx⎦ = 0 + D sin Dx 1 1/gZ (t)P(β̂ (t))fD (t) − 0 (1 + x)w (x, t)hD (x, t)dx
T

eDx
Dx + 1 0 0 τD (t) = (87)
1 + Ξ (t)
0 ˙
[ ]
Dx b̂i (x, t) = γb Proj{τbi (x, t)}, γb > 0 (88)
+ De 0
1 1/gZ (t)P(β̂ (t))fbi (x, t)
T
τbi (x, t) =
= b1 (x)B1 + b2 (x)B2 + b3 (x)B3 ■ (73) 1 + Ξ (t)
∫1
0
(1 + y)w (y, t)hbi (y, t)dyfbi (x, t)
As a result, with (72), the system (3)–(5) is rewritten as − (89)
∫ 1
1 + Ξ (t)
Ẋ (t) = AX (t) + B (x)u(x, t)dx (74) where P(β̂ (t)) satisfies (79) and g > 0 is a designing coefficient,
0

Dut (x, t) = ux (x, t), x ∈ [0, 1] (75) w(x, t) = u(x, t) − K (β̂ (t))eAcl (β̂ (t))D̂(t)(x−1) Z (t) (90)
u(1, t) = U(t)
∫ 1
(76)
Ξ (t) = Z (t)P(β̂ (t))Z (t) + g
T
(1 + x)w (x, t)2 dx (91)
For using projector operators later, the following assumption ∫ 1
0
is assumed.
fD (t) = B̂ (x, t)u(x, t)dx
0
Assumption 4. There exist known constants D, D, b̄i , and known ∫ 1
continuous functions b∗i (x) such that − e−AD̂(t)(1−x) B̂ (x, t)dxu(1, t)
∫ 1 0
)2 ∫ 1∫ x
0 < D ≤ D ≤ D, 0< bi (x) − b∗i (x)
(
dx ≤ b̄i (77)
0
− D̂(t) Ae−AD̂(t)(x−y) B̂ (y, t)dyu(x, t)dx (92)
0 0
for i = 1, . . . , n.
( )
To stabilize (74)–(76), the following assumption is required. hD (x, t) = K (β̂ (t))eAcl (β̂ (t))D̂(t)(x−1) fD (t) + Acl (β̂ (t))Z (t) (93)

fbi (x, t) = Bi u(x, t) (94)

Assumption 5. For the system (74), the pair (A, β) is stabilizable
Acl (β̂ (t))D̂(t)(x−1)
where hbi (x, t) = K (β̂ (t))e (95)
1
for i = 1, . . . , n, and the projector operator are defined as
∫
β= e−AD(1−x) B (x)dx (78)
⎨0, D̂(t) = D and τ < 0
0
⎧
There exists a vector K (β ) to make A + β K (β ) Hurwitz. Namely, Proj[D,D] {τ } = 0, D̂(t) = D and τ > 0 (96)
there exist matrices P(β ) = P(β )T > 0 and Q (β ) = Q (β )T > 0
τ , else
⎩
such that ⎧ ( ) ∫ 1 (b̂i (x)−b∗ (x))τ (x)dx
(A + β K (β ))T P(β ) + P(β ) (A + β K (β )) = −Q (β ) ■ (79)
⎪
⎪
⎪ τ (x) − b̂i (x) − bi (x) ∫ 1 (
∗ 0 i
)2 ,
b∗
⎪ b̂ (x)− (x) dx
i
⎪ 0 i
Denote D̂(t) and b̂i (x, t) as the estimates of D and bi (x) (for
⎪
⎪
⎨ ∫1( )2
i = 1, . . . , n) with estimation errors satisfying Proj{τ (x)} = if 0
b̂i (x) − b∗i (x) dx = b̄i (97)
⎪ ∫1( )
τ (x)dx > 0,
⎪ ∗
D̃(t) = D − D̂(t) (80) and b̂ (x) − b (x)
⎪
i
⎪ i
⎪
⎪ 0

τ (x), else
⎪
b̃i (x, t) = bi (x) − b̂i (x, t)
⎩
(81)
and
n
Theorem 3. Consider the closed-loop system consisting of the plant
∑ (74)–(76) and the adaptive controller (83)–(97). All the states (X (t),
B̂ (x, t) = b̂i (x, t)Bi (82)
i=1
u(x, t), D̂(t), b̂i (x, t)) of the closed-loop system are globally bounded
 Y. Zhu, M. Krstic and H. Su / Automatica 116 (2020) 108902 7

and the regulation of X (t) and U(t) such that limt →∞ X (t) = where
limt →∞ U(t) = 0 is achieved. ■ (
˙
)
ψ D̂(t), B̂˙ (x, t)
Proof. Taking the time-derivative of (85) along (74)–(76), and
( n
)
∂ K (β̂ (t)) ˙ ∂ K (β̂ (t)) ˙
b̂i (x, t) eAcl (β̂ (t))D̂(t)(x−1)
∑
using the integration by parts in x, we obtain = D̂(t) +
∫ 1
∂ D̂(t) i=1
∂ b̂ i (x , t)
˙
( )
B (x)u(x, t)dx + φ D̂(t), B̂˙ (x, t)
[(
Ż (t) = AX (t) + ∂ Acl (β̂ (t)) ˙
Acl (β̂ (t))D̂(t)(x−1)
0 × Z (t) + K (β̂ (t))e D̂(t)
D̂(t)
∫ 1 ∫ x ∂ D̂(t)
+ e−AD̂(t)(x−y) B̂ (y, t)dyux (x, t)dx n
) ]
D 0 0 ∑ ∂ Acl (β̂ (t)) ˙ ˙
∫ 1 + b̂i (x, t) D̂(t) + Acl (β̂ (t))D̂(t) (x − 1)Z (t)
˙ ∂ b̂i (x, t)
( )
= AX (t) + B (x, t)u(x, t)dx + φ D̂(t), B̂˙ (x, t) i=1
˙
0
( )
∫ 1 + K (β̂ (t))eAcl (β̂ (t))D̂(t)(x−1) φ D̂(t), B̂˙ (x, t) (103)
D̂(t)
+ e−AD̂(t)(1−y) B̂ (y, t)dyu(1, t)
D 0 Multiplying (102) by D and minus (101), employing the control
D̂(t)
∫ 1 law (83), we get
− B̂ (x, t)u(x, t)dx (
˙
)
D 0 Dwt (x, t) = wx (x, t) − Dψ D̂(t), B̂˙ (x, t) − D̃(t)hD (x, t)
∫ 1 ∫ x
D̂(t) n
AD̂(t)e−AD̂(t)(x−y) B̂ (y, t)dyu(x, t)dx 1
∫
+ ∑
D 0 0 − Dhbi (x, t) b̃i (y, t)fbi (y, t)dy (104)
∫ 1 0 i=1
˙
( )
= AX (t) + B (x, t)u(x, t)dx + φ D̂(t), B̂˙ (x, t)
0
w (1, t) = 0, x ∈ [0, 1] (105)
( )∫
D̃(t) 1
where hD (x, t) and hbi (x, t) have been defined in (93) and (95).
+ 1− e−AD̂(t)(1−x) B̂ (x, t)dxu(1, t) ∫1
D 0
Please note that D̂(t) and 0 b̂i (x, t)dx are bounded as the pro-
( )∫ jector operators ensure them to stay in the interval (77). Making
1
D̃(t) use of Young’s and Cauchy–Schwarz inequalities, it is evident that
− 1− B̂ (x, t)u(x, t)dx
D 0
the inverse transformation of (90) implies

∥u(x, t)∥2 ≤ Mu |Z (t)|2 + ∥w(x, t)∥2

( ) ( )
D̃(t)
∫ 1 ∫ x (106)
+ 1− D̂(t) Ae−AD̂(t)(x−y)
D 0 0 where Mu > 0 is a constant.
ε2
Utilizing (106) and inequalities 0 < < 1 and 0 < <
|ε|
× B̂ (y, t)dyu(x, t)dx (98)
1, it is easy to show that
1+ε 2 1+ε 2

where ⏐ ⏐ |Z (t)|2 + ∥w(x, t)∥2

⏐˙ ⏐
(
˙
) ⏐D̂(t)⏐ ≤ γD MD ≤ γD M̄D (107)
φ D̂(t), B̂˙ (x, t) 1 + Ξ (t)
|Z (t)|2 + ∥w(x, t)∥2
⏐∫ 1 ⏐
∫ 1∫ x ˙
b̂i (x, t)dx⏐⏐ ≤ γb Mbi ≤ γb M̄bi
[( ⏐ ⏐
˙ ˙ (108)
)
−AD̂(t)(x−y)
⏐
= e D̂(t)I − D̂(t)AD̂(t)(x − y) ⏐
0 1 + Ξ (t)
0 0
n ] where MD , Mbi , M̄D , M̄bi are positive constants.
∑ ˙
× B̂ (y, t) + D̂(t) b̂i (y, t)Bi dyu(x, t)dx (99) Build the Lyapunov candidate such that
i=1 n ∫ 1
g ∑ gD
V (t) = D log 1 + Ξ (t) ++ 2
b̃i (x, t)2 dx (109)
( )
Under the control law (83), the formula (98) becomes D̃(t) +
γD 0 γb
i=1
(
˙
) D̃(t)
Ż (t) = Acl (β̂ (t))Z (t) + φ D̂(t), B̂˙ (x, t) + fD (t) where Ξ (t) has been defined in (91).
D Taking the time-derivative of (109) along the target closed-
∫ 1 n
loop system (100), (104)–(105), employing (106)–(108), following
∑
+ b̃i (x, t)fbi (x, t)dx (100)
0 a similar procedure of proof of Theorem 1, we get
i=1 [ ]
N
where Acl (β̂ (t)) is given in (84), fD (t) and fbi (x, t) have been V̇ (t) ≤ −|Z (t)|2 − w(0, t)2 − ∥w(x, t)∥2
1 + Ξ (t)
defined in (92) and (94).
Taking partial derivatives of (90) with respect to x and t, 2g (
˙
)
− D̃(t) D̂(t) − γD τD (t)
respectively, we get γD
n ∫ 1
2gD
wx (x, t) = ux (x, t) − K (β̂ (t))eAcl (β̂ (t))D̂(t)(x−1) ˙
∑ ( )
− b̃i (x, t) b̂i (x, t) − γb τbi (x, t) dx
0 γb
× Acl (β̂ (t))D̂(t)Z (t) (101) i=1
[ ]
(
˙
) N
wt (x, t) = ut (x, t) − ψ D̂(t), B̂˙ (x, t) −|Z (t)| − w(0, t) − ∥w(x, t)∥
2 2 2
≤ (110)
1 + Ξ (t)
(
− K (β̂ (t))e Acl (β̂ (t))D̂(t)(x−1)
Acl (β̂ (t))Z (t) where N > 0 is a constant, and τD (t) and τbi (x, t) are given in
(87) and (89). Thus Z (t) and w (x, t) are bounded and converge to
D̃(t)
∫ 1 n
∑ ) zero. By inverse conversions of (85) and (90), the original states
+ fD (t) + b̃i (y, t)fbi (y, t)dy (102) X (t) and u(x, t) are bounded and converge to zero. Thus the proof
D 0 i=1 of Theorem 3 is proved.
8 Y. Zhu, M. Krstic and H. Su / Automatica 116 (2020) 108902

Remark 4. A special case of (71) is that the input vector B(Dx) for is based on the certainty-equivalence principle and the update
x ∈ [0, 1] is a continuous function of Dx and is parameterizable laws are on the basis of the construction of a Lyapunov func-
in the delay D such that tion with normalization. Two different cases of input dynamics
p are taken into account: (1) the input vector is a constant vec-
tor independent of delay, (2) the input vector is a continuous
∑
B(Dx) = B0 + ρi (D)Bi (x) (111)
vector-valued function dependent upon delay.
i=1

where B0 ∈ Rn is a known constant vector, Bi (x) : [0, 1] → Rn References

for i = 1, . . . , p are known continuous vector functions of x,
and ρi (D) for i = 1, . . . , p are unknown constant parameters Anfinsen, H., & Aamo, O.-M. (2019). Adaptive control of hyperbolic PDEs. Springer.
dependent upon D. A three-dimensional example of (111) is given Artstein, Z. (1982). Linear systems with delayed controls: A reduction. IEEE
below. Transactions on Automatic Control, 27(4), 869–879.
Bekiaris-Liberis, N., Jankovic, M., & Krstic, M. (2013). Adaptive stabilization of LTI
systems with distributed input delay. International Journal of Adaptive Control
Example 2. and Signal Processing, 27, 47–65.
⎡ 1
⎤ 1
⎡ ⎤ Bekiaris-Liberis, N., & Krstic, M. (2011). Lyapunov stability of linear predictor
Dx + 0 x
[ ] [ ]
√ Dx+D 1 x+1 feedback for distributed input delays. IEEE Transactions on Automatic Control,
B(Dx) = ⎣ Dx + 1 ⎦ = 1 +D 0 + ⎣ 0 ⎦ 56, 655–660.
2 0 D
D 2 x2 + 2 0 Bresch-Pietri, D., Chauvin, J., & Petit, N. (2012). Adaptive control scheme for
⎡ ⎤ ⎡ ⎤ uncertain time-delay systems. Automatica, 48, 1536–1552.
√ √0 0 Bresch-Pietri, D., & Krstic, M. (2009). Adaptive trajectory tracking despite
+ D ⎣ x⎦ + D2 ⎣ 0 ⎦ unknown input delay and plant parameters. Automatica, 45, 2074–2081.
0 x2 Goodwin, G. C., & Sin, K. S. (2014). Adaptive filtering, prediction and control.
Courier Corporation.
= B0 + ρ1 (D)B1 (x) + ρ2 (D)B2 (x) Niculescu, S.-I., & Annaswamy, A. M. (2003). An adaptive smith-controller for
time-delay systems with relative degree n∗ ≤ 2. Systems & Control Letters,
+ ρ3 (D)B3 (x) + ρ4 (D)B4 (x) (112) 49, 347–358.
Ortega, R., & Lozano, R. (1988). Globally stable adaptive controller for systems
On the basis of (111), we further denote
with delay. International Journal of Control, 47, 17–23.
p p Smith, O. J. M. (1959). A controller to overcome dead time. ISA, 6, 28–33.
∑ ∑
B (x) = DB(Dx) = DB0 + Dρi (D)Bi (x) = bi Bi (x) (113) Smyshlyaev, A., & Krstic, M. (2010). Adaptive control of parabolic PDEs. Princeton
University Press.
i=1 i=0
Yildiz, Y., Annaswamy, A., Kolmanovsky, I. V., & Yanakiev, D. (2010). Adaptive
where b0 = D and bi = Dρi (D) for i = 1, . . . , p are unknown posicast controller for time-delay systems with relative degree n∗ ≤ 2.
constant parameters with respect to D. Automatica, 46, 279–289.
We assume there exist known constants bi , bi such that 0 < Zhou, J., Wen, C., & Wang, W. (2009). Adaptive backstepping control of uncertain
systems with unknown input time-delay. Automatica, 45(6), 1415–1422.
bi ≤ bi ≤ bi . Instead of functional estimators in (81), we use Zhu, Y., Krstic, M., & Su, H. (2017). Adaptive output feedback control for
scalar estimators b̂i (t) for i = 0, . . . , p as the estimates of bi with uncertain linear time-delay systems. IEEE Transactions on Automatic Control,
estimation errors being 62, 545–560.
Zhu, Y., Krstic, M., & Su, H. (2018a). Adaptive global stabilization of uncertain
b̃i (t) = bi − b̂i (t) (114) multi-input linear time-delay systems by PDE full-state feedback. Automatica,
96, 270–279.
and Zhu, Y., Krstic, M., & Su, H. (2018b). PDE boundary control of multi-input
p LTI systems with distinct and uncertain input delays. IEEE Transactions on
Automatic Control, 63, 4270–4277.
∑
B̂ (x, t) = b̂i (t)Bi (x) (115) Zhu, Y., Krstic, M., & Su, H. (2019). PDE output feedback control of LTI systems
i=0 with uncertain multi-input delays, plant parameters and ODE state. Systems
The update law of functional adaptation (88)–(89) is reduced to and Control Letters, 123, 1–7.
Zhu, Y., Su, H., & Krstic, M. (2015). Adaptive backstepping control of uncertain
the scalar adaptation below linear systems under unknown actuator delay. Automatica, 54, 256–265.
˙
γb > 0
b̂i (t) = γb Proj[b ,bi ] {τbi (t)}, (116)
i
∫1 Yang Zhu received his Bachelor degree from Hangzhou
1/gZ (t)P(β̂ (t))fbi (t) − 0 (1 + x)w (x, t)hbi (x, t)dx
T Dianzi University, China, in Jul. 2010, and Ph.D. degree
τbi (t) = (117) at Zhejiang University, China, in Jun. 2015. Supported
1 + Ξ (t) by China Scholarship Council, he was a Visiting Gradu-
ate Student at University of California, San Diego, USA
where everything else are defined the same as those in (89)
from Sep. 2013 to Oct. 2014. Between Aug. 2015 and
except for Sep. 2016, he was a Post-doc at University of Texas
∫ 1 at Dallas, USA. From Feb. 2017 to Oct. 2018, he was a
fbi (t) = Bi (x)u(x, t)dx (118) Research Fellow at Nanyang Technological University,
0 Singapore. Since Oct. 2018, he has been a Post-doc at
Tel Aviv University, Israel. His research involves control
hbi (x, t) = K (β̂ (t))eAcl (β̂ (t))D̂(t)(x−1) fbi (t) (119) theory with its application to transportation engineering.

and
Miroslav Krstic serves as Sr. Assoc. Vice Chancellor
⎨0, b̂i (t) = bi and τ < 0
⎧
for Research at UC San Diego, overseeing Organized
Proj[b ,bi ] {τ } = 0, b̂i (t) = bi and τ > 0 ■ (120) Research Units, the research scientist and postdoc-
i toral scholar affairs, proposal development services,
τ , else
⎩
and research initiatives. He is Distinguished Professor
of Mechanical and Aerospace Engineering, holds the
4. Conclusion Alspach endowed chair, and is the founding director
of the Cymer Center for Control Systems and Dynam-
ics. Krstic has been elected Fellow of seven scientific
This paper presents an adaptive approach for stabilizing linear societies - IEEE, IFAC, ASME, SIAM, AAAS, IET (UK), and
systems with unknown distributed input delays. The control law AIAA (Assoc. Fellow) - and as a foreign member of the
 Y. Zhu, M. Krstic and H. Su / Automatica 116 (2020) 108902
Serbian Academy of Sciences and Arts and of the Academy of Engineering of
Serbia. He has received the SIAM Reid Prize, ASME Oldenburger Medal, Nyquist
Lecture Prize, Paynter Outstanding Investigator Award, Ragazzini Education
Award, IFAC Nonlinear Control Systems Award, Chestnut textbook prize, Control
Systems Society Distinguished Member Award, the PECASE, NSF Career, and ONR
Young Investigator awards, the Schuck (’96 and’19) and Axelby paper prizes,
and the first UCSD Research Award given to an engineer. Krstic has also been
awarded the Springer Visiting Professorship at UC Berkeley, the Distinguished
Visiting Fellowship of the Royal Academy of Engineering, and the Invitation
Fellowship of the Japan Society for the Promotion of Science. He serves as
Editor-in-Chief of Systems & Control Letters and has been serving as Senior
Editor in Automatica and IEEE Transactions on Automatic Control, as editor
of two Springer book series, and has served as Vice President for Technical
Activities of the IEEE Control Systems Society and as chair of the IEEE CSS Fellow
Committee. Krstic has coauthored thirteen books on adaptive, nonlinear, and
stochastic control, extremum seeking, control of PDE systems including turbulent
flows, and control of delay systems.
9
Hongye Su received his Ph.D. degree from Zhejiang
University, China, in 1995. He has been promoted to
the full Professor at the same university since 2000.
Since 2014, he has held the positions of Director of
the Institute of Cyber-Systems and Control, as well
as Director of the State Key Laboratory of Industrial
Control Technology at Zhejiang University. He was the
Dean of the College of Control Science and Engineering
at Zhejiang University between 2011 and 2013. He
received the National Science Fund for Distinguished
Young Scholars in 2000. Then he has been promoted as
Chang Jiang Chair Professor since 2013. He has taken charge of a large number
of significant research projects including National Basic Research Program of
China (973 Program) and National High-technology Research and Development
Program of China (863 Program). His main academic activities are in the areas
of robust control, nonlinear adaptive control and chemical industrial control.