Automatica 94 (2018) 409–415

Contents lists available at ScienceDirect

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

Brief paper

Sliding mode predictive control of linear uncertain systems with

delays✩
Humberto Caballero-Barragán a, *, Linda Patricia Osuna-Ibarra a , Alexander G. Loukianov a ,
Franck Plestan b
a
CINVESTAV del IPN, Unidad Guadalajara, Av. del Bosque 1145 CP. 45019, Jalisco, Mexico
b
Ecole Centrale de Nantes, LS2N UMR CNRS 6004, Nantes, France

article info a b s t r a c t
Article history: In this paper, a continuous control strategy for robust stabilization of a class of uncertain multivariable
Received 26 July 2016 linear systems with delays in both the state and control variables is proposed. A predictor is designed
Received in revised form 1 December 2017 to compensate the delay effect in the control input, and then an integral sliding mode control technique
Accepted 22 March 2018
along with super-twisting algorithm is applied to compensate partially the effect of the perturbation term.
Finally, a nominal delay-free part of the control input is designed to stabilize the sliding mode dynamics.
The proposed control scheme is extended to the class of systems modeled in Regular form. For this class
Keywords:
Time-delay systems of perturbed systems with delay in the state, a transformation to the systems with the delay-free state
Sliding mode control is proposed. The stability conditions of the closed-loop uncertain system are derived, and the results
Linear system obtained in this work are compared against previous works. To show the effectiveness of the proposed
Regular form method, simulation results are presented.
© 2018 Elsevier Ltd. All rights reserved.

1. Introduction implementation of these robust control techniques without taking

Many modern industrial processes including systems where
control signals are transmitted through networks, are modeled
by delay differential equations. In these systems, the time delay
can appear in the system state as well as in the control input.
The last case is more dangerous for the closed-loop stability if the
delay is large enough with respect to the plant dynamics rate and
the standard memoryless feedback, i.e. the usual current system
state, is used. Among stability analysis results for this case have
been reported based on the Lyapunov–Krasovskii approach (see
Fridman (2014), Mazenc, Niculescu, and Krstic (2012) and refer-
ences therein). Basically, these results involve the upper bound on
the time delay or the control gain values. On the other hand, the
presence of the plant model uncertainty makes the situation more
complex even for linear time invariant systems. A possible way to
treat this problem is making use of the high gain or Sliding Mode
(SM) control techniques (Utkin, Guldner, & Shi, 1999) which are
effective tools to reject the system uncertainty. However, the direct
✩ The material in this paper was not presented at any conference. This paper was
recommended for publication in revised form by Associate Editor Constantino M.
Lagoa under the direction of Editor Richard Middleton.
*
Corresponding author.
E-mail addresses: hcaballero@gdl.cinvestav.mx (H. Caballero-Barragán),
lposuna@gdl.cinvestav.mx (L.P. Osuna-Ibarra), louk@gdl.cinvestav.mx
(A.G. Loukianov), franck.plestan@ec-nantes.fr (F. Plestan).
into account delays may lead to oscillations and even instability
of the closed-loop system (Fridman, Shustin, & Fridman, 1997). To
overcome this problem, a predictor-based approach that enables
to compensate for the time delay in the control input resulting
in the delay-free closed-loop system, can be applied. First, the
Smith predictor has been proposed in Smith (1957); however,
this frequency domain approach can be implemented for open-
loop stable systems only. To extend this approach to the general
case of MIMO open-loop unstable systems, the Finite Spectrum
Assignment approach based on the solution of LTI system has
been proposed in Kwon and Pearson (1980) and Manitius and
Olbrot (1979). The stability of the closed-loop system with pre-
dictor in absence of disturbances has been analyzed in Furukawa
and Shimemura (1983), and an exhaustive analysis of predictive
control scheme can be found also in Krstic (2009). In the pres-
ence of disturbances, the predictor was successfully applied to
design SM controllers in Edwards and Spurgeon (1998), Polyakov
(2012) and Roh and Oh (1999). However, the matching condi-
tion fulfilled for the uncertainties in the original system does not
hold in the transformed delay-free (prediction) system with the
conventional SM algorithm (Nguang, 2001). To solve this prob-
lem, the integral SM control (Utkin et al., 1999) implementation
has been proposed in Loukianov, Espinosa-Guerra, Castillo-Toledo,
and Utkin (2005, 2006). This approach allows to preserve the
matching condition in the prediction system and, as a result, the
unknown perturbation effect is reduced. On the other hand, in

https://doi.org/10.1016/j.automatica.2018.04.040
0005-1098/© 2018 Elsevier Ltd. All rights reserved.
410 H. Caballero-Barragán et al. / Automatica 94 (2018) 409–415

Léchappé, Moulay, Plestan, Glumineau, and Chriette (2015), a new Assumption 4. All eigenvalues of matrices M1−1 M2 are located
predictive scheme has been designed permitting as well to reduce inside the open unit circle, where M1 and M2 are
the perturbation effect.
M1 = BT (I + eAτ1 )B and M2 = BT eAτ1 B.
In this paper, a class of linear perturbed systems with delay
in the system state and the control input is considered, first, in Assumption 5. The time delays τ0 and τ1 are constant and known.
general case and then in Regular form (Loukianov & Utkin, 1981).
It is assumed that the time delay in the state vector is bigger
3. Predictive control scheme for general case
than in the control input vector and the uncertainties satisfy the
matching condition, as it is common in SM control design. A new 3.1. SM control design
predictive SM control scheme is designed using the advantages of
the predictor (Léchappé et al., 2015) and the integral SM predictive To eliminate the known delayed term D1 x(t − τ0 ) and robustly
controller (Loukianov et al., 2006). This scheme includes a predic- stabilize the system (1) under Assumption 3, the control law is
tor. The predictor is proposed in the form of Léchappé et al. (2015) redefined following the integral SM philosophy as
to stabilize the SM dynamics. To design a stabilizing controller and
u(t) = u0 (t) + u1 (t) + u2 ,
achieve the robustness of the closed-loop predictive system where (2)
the matching condition is preserved contrary to the conventional u2 (t) = −D1 x(t − ∆), ∆ = τ0 − τ1
SM controller (Polyakov, 2012; Roh & Oh, 1999), the integral SM
where u0 (t) ∈ Rm is the nominal control, and u1 (t) ∈ Rm will be
predictive control technique (Loukianov et al., 2006) combined
designed to reject the perturbation term α (·). Substituting (2) into
with super-twisting algorithm (Fridman & Levant, 2002; Moreno
(1), yields
& Osorio, 2008) is used.
The stability analysis shows that in this case, even though the ẋ(t) = Ax(t) + B[u0 (t − τ1 ) + u1 (t − τ1 )
matching condition is preserved, the proposed control scheme
+ α (x(t), x(t − τ0 ), t)]. (3)
cannot totally compensate for the unknown arbitrary perturbation.
m
However, the perturbation effect can be reduced compared to An integral sliding function s1 (t) ∈ R is formulated of the form
Loukianov et al. (2006). Moreover, this scheme enables to totally
compensate the unknown constant perturbation. s1 (t) = Gxp (t) + w (t), w (0) = −Gxp (0), (4)
So, the proposed new predictive SM control scheme which n
with the predictive state xp (t) ∈ R of the system (3) (the predictor
reduces the matched disturbance effect in a linear system with is presented in Léchappé et al. (2015))
delay in the state and input vectors can be considered as the main
xp (t) = ξ (t) + x(t) − ξ (t − τ1 ),
contribution of this paper.
The paper is organized as follows. In Section 2, the problem 0 (5)
∫
statement including assumptions is presented. The predictive con- ξ (t) = eAτ1 x(t) + e−Aθ Bu0 (t + θ )dθ,
−τ1
trol scheme is designed, first, for general class of linear systems
in Section 3, and, then, for systems presented in Regular form in where G ∈ R m×n
is a design matrix, and w (t) is defined by
Section 4, including SM control design (Sections 3.1 and 4.1). The ẇ(t) = − G[Axp (t) + Bu0 (t)]
SM dynamics stability is analyzed in Section 3.2 and compared
with the previous work in Section 3.3. To clarify the proposed + (GB + GeAτ1 B)[u1 (t) − u1 (t − τ1 )] (6)
control scheme, one example is presented in Section 4.2. Aτ1
− Ge B[u1 (t − τ1 ) − u1 (t − 2τ1 )].
2. Problem statement Taking the time derivative of (4) and using (3), (5) and (6), yields
ṡ1 (t) = (GB + GeAτ1 B)[u1 (t)
Consider an uncertain linear system with time delays de-
scribed by + α (x(t), x(t − τ0 ), t)] − GeAτ1 B[u1 (t − τ1 ) (7)
ẋ(t) = Ax(t) + Dx(t − τ0 ) + Bu(t − τ1 ) + α (x(t − τ1 ), t − τ1 , x(t − τ0 − τ1 ), t − τ1 )].
(1)
+ f (x(t), x(t − τ0 ), t), To induce a sliding motion on s1 (t) = 0, the control component
with the initial conditions given by, x(t) = ϕ0 (t), ∀t ∈ [t0 − u1 (t) is selected using super-twisting algorithm (Moreno & Osorio,
τ0 , t0 ], u(t) = ϕ1 (t) ∀t ∈ [t0 − τ1 , t0 ], where x ∈ Rn and u ∈ 2008; Nagesh & Edwards, 2014) as
Rm are the state and control vectors, respectively; the unknown s1 (t)
u1 (t) = M1−1 [−k1 − k2 s1 (t) + v (t)
function f ∈ Rn represents model uncertainties including external 1
∥s1 (t)∥ 2
disturbances; A, D and B are matrices of appropriate dimensions,
rank (B) = m; τ0 and τ1 are time delays, τ0 ≥ τ1 . + M2 u1 (t − τ1 )] (8)
In this paper, the objective is to design a robust controller in an s1 (t)
uncertain scenario (1). Therefore, the following assumptions are v̇ (t) = − k3 − k4 s1 (t),
required: ∥s1 (t)∥
where k1 , k2 , k3 and k4 are the control gains, G = BT ,
Assumption 1. The pair (A, B) is stabilizable, and the state x is
available for the measurement. M1 = BT B + BT eAτ1 B and M2 = BT eAτ1 B. (9)
Substituting control (8) into (7), results in
Assumption 2. The unknown function f (x(t), x(t − τ0 ), t) is locally
s1 (t)
Lipshitz and satisfies the matching condition (Drazenovic, 1969), ṡ1 (t) = − k1 1
− k2 s1 (t) + v (t)
namely, f (x(t), x(t − τ0 ), t) = Bα (x(t), x(t − τ0 ), t); where function ∥s1 (t)∥ 2
α (x(t), x(t − τ0 ), t), α ∈ Rm is bounded.
+ ∆α (x(t), x(t − τ0 ), t) (10)
Assumption 3. There is a matrix D1 ∈ Rm×n such that D = BD1 s1 (t)
holds. v̇ (t) = − k3 − k4 s1 (t),
∥s1 (t)∥
 H. Caballero-Barragán et al. / Automatica 94 (2018) 409–415 411

where ∆α (x(t), x(t − τ0 ), t) = M1 α (x(t), x(t − τ0 ), t) − M2 α (x(t − Remark 1. In this work, the prediction error is
τ1 ), x(t − τ0 ), t − τ1 ). ∆α (x(t), x(t − τ0 ), t) can be separated in ζ (t)−xp (t) =
∆α1 (x(t), x(t − τ0 )) and ∆α2 (t), ∆α1 (x(t), x(t − τ0 )) represents the
system nonlinearity, ∆α2 (t) represents external disturbances and
∫ 0 (19)
it depends only on time t. Under Assumption 3, there exist k1 , k2 , k3 e−Aθ B(∆ατ ,eq (t + θ ) − ∆ατ ,eq (t − τ1 + θ ))dθ
−τ1
and k4 such that s1 (t) converges to zero in finite time (see Nagesh
and Edwards (2014)). while in the previous work (Loukianov et al., 2006), the prediction
error is equal to
∫ 0
3.2. SM dynamics analysis
ζ (t) − ξ (t) = e−Aθ B∆ατ ,eq (t + θ )dθ. (20)
−τ1
Using transformation (5), the delayed system (3) is transformed
Now, to analyze the solution ζ (t) in presence of time-varying
into a system which is delay-free in the nominal part u0 (t) of the
perturbation, the time derivative of (18) is taken, and using (15),
control, namely:
the dynamics for ζ (t) on s1 (t) = 0 becomes
ẋp (t) = Axp (t) + Bu0 (t)
ζ̇ (t) = Aζ (t) + Bu0 (t) + B∆ατ ,eq (t), (21)
+ (B + eAτ1 B)[u1 (t − τ1 ) + α (x(t), x(t − τ0 ), t)]
(11) where ∆ατ ,eq (t) is the predicted value for ∆ατ ,eq (t − τ1 ) defined in
− eAτ1 B[u1 (t − 2τ1 ) (16), i.e., ∆ατ ,eq (t) = α (ζ (t), ζ (t − τ0 ), t + τ1 ) − α (ζ (t − τ1 ), ζ (t −
τ1 − τ0 ), t).
+ α (x(t − τ1 ), x(t − τ0 − τ1 ), t − τ1 )].
From (18), it follows xp (t) = ζ (t) − −τ e−Aθ B∆
∫0
¯ ατ ,eq (t + θ )dθ .
1
It is important to note that both systems (3) and (11) provide Then, substituting it into the nominal control (17) and also into
an equivalent input–output mapping (see Fiagbedzi and Pearson (21), yields
(1986)), this permits to obtain and analyze SM equation using
ζ̇ (t) = Āζ (t) + B[∆ατ ,eq (t)
system (11) instead of system (3).
(22)
∫ 0
Now, imposing ṡ1 (t) = 0 in (7), results
− K0 e−Aθ B∆
¯ ατ ,eq (t + θ )dθ]
0 = M1 [u1 (t) + α (x(t), x(t − τ0 ), t)] − M2 [u1 (t − τ1 ) −τ1
(12)
+α (x(t − τ1 ), t − τ1 , x(t − τ0 − τ1 ), t − τ1 )]. where Ā = (A + BK0 ) is a Hurwitz matrix.
Under Assumption 2, there are constants γ1 > 0 and γ2 > 0
The matrices M1 and M2 defined by (9) ensure Eq. (12) dynamic sta-
such that
bility since M1−1 M2 is Schur matrix (see Assumption 5, (Loukianov,
∆ατ ,eq (t) ≤ γ1 τ1 and ∆ ¯ ατ ,eq (t) ≤ γ2 τ12 .
   
Castillo-Toledo, Hernández, & Núñez, 2003)). (23)
From (12), the equivalent control u1eq (t) is obtained as
Moreover, e−At  ≤ γ3 eλmax (A)t , γ3 > 0, therefore
 

u1eq (t) = −α (x(t), x(t − τ0 ), t) (13)  ∫ 0 

−Aθ ¯
∆ α θ θ  ≤ γ4 τ 2 ,
 
K0 e B τ ,eq (t + )d (24)
1
and 
−τ1

γ2 γ3
u1eq (t − τ1 ) = −α (x(t − τ1 ), x(t − τ1 − τ0 ), t − τ1 ). (14) with γ4 = λmax (A)
∥K0 ∥ (1 + e−λmax (A)τ1 ).

Substituting (13) and (14) into (11), a sliding mode motion on Theorem 1. Suppose Assumptions 1–3 hold. Then, any solution of the
s1 (t) = 0 is described by the following uncertain system: system (21) closed by the controller
ẋp (t) = Axp (t) + Bu0 (t) + B∆ατ ,eq (t − τ1 ) ∫ 0
(15) u0 (t) = K0 xp (t) = K0 [ζ (t) − e−Aθ B∆
¯ ατ ,eq (t + θ )dθ]
+ eAτ1 B∆
ˆ ατ ,eq (t − 2τ1 ), −τ1

where with Ā = (A + BK0 ) a Hurwitz matrix, is ultimately bounded by

√
∆ατ ,eq (t − τ1 ) = −u1eq (t) + u1eq (t − τ1 ), (16) λmax (P)
∥ζ (t)∥ ≤ δ0 , δ0 = µ0 ,
λmin (P)
ˆ ατ ,eq (t − 2τ1 ) = ∆ατ ,eq (t − τ1 ) − ∆ατ ,eq (t − 2τ1 ).
∆ 2γ0 τ1 ∥PB∥
with µ0 = β0
, 0 < β0 < 1, γ0 = γ1 + γ4 τ1 and with P a
The nominal component of the control, u0 (t), in (15), can be now positive definite solution of the Lyapunov equation ĀT P + P Ā = −In ,
selected as where In is the identity matrix. Moreover, if the disturbance term is
constant, then ζ (t) tends asymptotically to zero.
u0 (t) = K0 xp (t), (17)
The proof for Theorem 1 is presented in Appendix A.
where K0 ∈ Rm×n is chosen, by Assumption 1, such that the matrix
(A + BK0 ) is Hurwitz. Corollary 1. For constant disturbances, the asymptotic convergence
At this point, however, it is necessary to analyze the behavior of of xp (t) (see (15)) to zero implies the asymptotic convergence of x(t)
the original predicted variable x(t +τ1 ) instead of merely that from (see (3)).
system (15). Defining ζ (t) = x(t + τ1 ) and using (5), this variable
can be defined on s1 (t) = 0 as Proof. If xp (t) tends to zero, then, from Remark 1, it follows that
ζ (t) = x(t + τ1 ) tends −τ e−Aθ ∆
∫0
∫ 0 ¯ ατ ,eq (t + θ )dθ . For constant
ζ (t) = xp (t) + e−Aθ B∆
¯ ατ ,eq (t + θ )dθ, (18) 1
disturbances, using (13), (14) and (16) results ∆ατ ,eq (t − τ1 ) = 0
−τ1
and ∆ ˆ ατ ,eq (t − 2τ1 ) = 0. Therefore, ζ (t) tends to zero and also x(t)
¯ ατ ,eq (t) = ∆ατ ,eq (t) − ∆ατ ,eq (t − τ1 ).
where ∆ tends to zero.
412 H. Caballero-Barragán et al. / Automatica 94 (2018) 409–415

3.3. Comparative analysis Then, the control u(t) is defined similar to (2)

In this subsection, the obtained results, namely, the ultimate u(t) = u0 (t) + u1 (t) + u2 (t)
(31)
bound is compared with results presented in Loukianov et al. u2 (t) = T X̄ (t − ∆), ∆ = τ0 − τ1 ,
(2006). The closed-loop predictive dynamics on sliding manifold
s1 (t) = 0 obtained in the present paper are described by where X̄ (t) = [x̄1 (t)T x̄2 (t)T x1 (t)T ]T , T = −B−1
2 [E F H ], E =
ẋp (t) = Āxp (t) + B∆ατ ,eq (t − τ1 ) A21 − A22 D̄11 + D̄11 A11 , F = D22 + D̄11 D12 and H = D21 − D22 D̄11 .
(25)
+ eAτ1 B∆
ˆ ατ ,eq (t − 2τ1 ),
Lemma 1. Given the transformation (30) and controller (31), the
while the similar dynamics obtained in Loukianov et al. (2006) are system (28)– (29) is transformed in a system delay-free in the state
expressed as
˙ = Āx̄ + B̄[u0 (t − τ1 ) + u1 (t − τ1 ) + α (x̄(t), x̄(t − τ0 ), t)], (32)
x̄(t)
ξ̇ (t) = Āξ (t) + eAτ1 B∆ατ ,eq (t − τ1 ), (26) [ ] [ ] [ ]
x̄1 (t) A11 D12 0
where x̄(t) = x̄2 (t)
, Ā = 0 A22
, and B̄ = B2
.
where matrix Ā = A + BK0 is Hurwitz. The comparative analysis of
these results is presented in the following theorem.
The proof for Lemma 1 is presented in Appendix C.
Theorem 2. Consider the system (25) resulting in the ultimate bound Thus, the system (28)–(29) is transformed via (30)–(31) in (32)
l1 in (B.5) ∥ζ (t)∥ ≤ l1 , and the system (26) resulting in the ultimate which is similar to system (3). Therefore, the results obtained in
bound l2 in (B.6) ∥ζ (t)∥ ≤ l2 . If the bounds γ1 and γ2 (23) satisfy the Sections 3.1 and 3.2, can be directly implemented for system (32).
relation
γ1
( )
l3
> τ1 1 + , (27) 4.2. Designexample
γ2 l4
then one gets l1 < l2 .
To illustrate the effectiveness of the proposed SM control
The proof for Theorem 2 is presented in Appendix B. Thus, under scheme, we consider the following 3rd order linear uncertain sys-
Theorem 2 condition (27), the proposed controller achieves an tem with time delays in the state and the control
ultimate bound smaller than in Loukianov et al. (2006).
−1 2 3 1.5 1 2. 5
[ ] [ ]
4. Regular form for systems with time delay ż(t) = −2 1 2 z(t) + 1 −2 1 z(t − τ0 )
−1 1 3 1 2 2 (33)
In this section, a class of linear systems (1) which can be trans- ]T
+ 1 0 1 [u(t − τ1 ) + α (z(t), z(t − τ0 ), t)],
[
formed under Assumption 2 into the following Regular form with
delay (see Loukianov and Utkin (1981), Loukianov et al. (2005)) is
considered: where z(t) = [z1 (t) z2 (t) z3 (t)]T , α (z(t), z(t −τ0 ), t) = 0.15(z1 (t)2 +
z2 (t)2 + z3 (t)2 )1/2 − 0.10(z1 (t − τ0 )2 + z2 (t − τ0 )2 + z3 (t − τ0 )2 )1/2 +
ẋ1 (t) = A11 x1 (t) + A12 x2 (t) + D11 x1 (t − τ0 )
3sin(0.5t), τ0 = 0.33 sec and τ1 = 0.11 sec.
+ D12 x2 (t − τ0 ) (28)
Firstly, system (33) is transformed into (see Loukianov and
ẋ2 (t) = A21 x1 (t) + A22 x2 (t) + D21 x1 (t − τ0 ) Utkin (1981), Loukianov et al. (2005)):
+ D22 x2 (t − τ0 )
+ B2 [u(t − τ1 ) + α (x(t), x(t − τ0 ), t)], (29) ẋ1 (t) = A11 x1 (t) + D11 x1 (t − τ0 ) + D12 x2 (t − τ0 ) (34)

where x(t) = [x1 (t)T x2 (t)T ]T is the transformed state vector,

ẋ2 (t) =A21 x1 (t) + A22 x2 (t) + D21 x1 (t − τ0 )
x1 ∈ Rn−m , x2 ∈ Rm , rankB2 = m, n ≥ 2m and the matrix A12
is considered as A12 = 0 (the case when the matrix A12 ̸ = 0 is + D22 x2 (t − τ0 ) + B2 [u(t − τ1 ) (35)
presented in Loukianov and Escoto-Hernandez (2000)).
For this system, Assumption 3 can be relaxed as + α (x(t), x(t − τ0 ), t)],
[ ]T
Assumption 6. There is a matrix D̄11 ∈ Rm×(n−m) such that where x(t) = x1 (t) x2 (t) = Mz(t) (see Loukianov
[ and
] Utkin
0 1
D11 = D12 D̄11 . (1981), Loukianov et al. (2005)), B2 = 1, A11 = −2 [ ]1 , A21 =
0.5
[ ]
−1 1
1 , A22 = 2, D11 = , D12 = , D21 =
[ ]
Assumption 6 is a relaxed version of Assumption 3. This is −1 1 −2 2
because Assumption 3 requires the delayed state to be in the [ ]T
1
space of control, i.e., (B[u(t − τ1 ) + D1 x(t − τ0 )]). Nevertheless, 2
, D22 = 3.
Assumption 6 only needs the delayed state to be matched with the Then, the Regular form (34)–(35) is transformed (see Lemma 1)
pseudo-control, that is, (D12 [x2 (t − τ0 ) + D̄11 x1 (t − τ0 )]). So, it is no via
longer necessary for the complete delayed state x(t − τ0 ) to be in ∫ 0
the space of the control.
x̄1 (t) =eA11 τ0 x1 (t) + e−A11 θ C12 x̄2 (t + θ )dθ
−τ0
4.1. Controller design
x̄2 (t) =x2 (t) + D̄11 x1 (t) (36)
Given systems (28)–(29) and Assumption 6, the following pre-
dictive transformation is proposed: u(t) =u0 (t) + u1 (t) + u2 (t)
0 u2 (t) =T X̄ (t − ∆), ∆ = τ0 − τ1 ,
∫
x̄1 (t) = eA11 τ0 x1 (t) + e−A11 θ D12 x̄2 (t + θ )dθ
−τ0 (30) with X̄ (t) = [x̄1 (t)T x̄2 (t)T x1 (t)T ]T , D̄11 = [0.5 − 1], T =
x̄2 (t) = x2 (t) + D̄11 x1 (t). −[0 2.5 1.5 − 0.5 5], E = [0 2.5], F = 1.5 and H = [−0.5 5]
 H. Caballero-Barragán et al. / Automatica 94 (2018) 409–415 413

Fig. 1. Norm of state z(t). Fig. 2. The signal control u1 (t) and the disturbance −α (z(t), z(t − τ0 ), t).

into the following delay-free state system: vectors is presented. The controller is designed using the mod-
ernized predictor proposed by Léchappé et al. (2015) and the
˙ =Āx̄(t) + B̄[u0 (t − τ1 ) + u1 (t − τ1 )
x̄(t) integral SM predictive control of Loukianov et al. (2006) based
+ α (x̄(t), x̄(t − τ0 ), t)] on super-twisting algorithms (Fridman & Levant, 2002; Moreno &
Osorio, 2008). The stability analysis shows that proposed controller
0 1 1 0
[ ] [ ]
(37) reduces the solution ultimate bound with respect to the previous
= −2 1 2 x̄(t) + 0 [u0 (t − τ1 ) + u1 (t − τ1 ) work (Loukianov et al., 2006). Moreover, if the disturbance is con-
0 0 2 1 stant, then the closed-loop system solution asymptotically tends
to zero. The performed simulations confirm that the steady state
+ α (x̄(t), x̄(t − τ0 ), t)]. error of the proposed controller is smaller than in Loukianov et al.
The complete control law is formulated of the form (2006). This work can be extended for the case when the delay is
time-variant, but it needs a thorough study, also, for systems with
s1 (t)
u(t) =K0 xp (t) + M̄1−1 [−k1 1
− k2 s1 (t) + υ (t) disturbances unmatched and the tracking problem.
∥s1 (t)∥ 2
Appendix A. Proof of Theorem 1
+ M̄2 u1 (t − τ1 )] + T X̄ (t − ∆) (38)

s1 (t) To analyze the stability of the system (22), a Lyapunov function

υ̇ (t) = − k3 − k4 s1 (t) Vζ (t) = 12 ζ T (t)P ζ (t) is used. Differentiating Vζ along the trajecto-
∥s1 (t)∥ ries of (22), we obtain
with M̄1 = (B̄T B̄ + B̄T eĀτ1 B̄), M̄2 = B̄T eAτ1 B̄, where s1 (t) = B̄T xp (t) + V̇ζ (t) ≤ −ζ T (t)ζ (t) + 2ζ T (t)PB[∆ατ ,eq (t)
ω(t), w(0) = −B̄T xp (0), ω̇(t) is computed by ∫ 0
− K0 e−Aθ B∆
¯ ατ ,eq (t + θ )dθ].
ω̇(t) = − B̄ [Ā + B̄K0 ]xp (t) + M1 [u1 (t) − u1 (t − τ1 )]
T
−τ1
− M2 [u1 (t − τ1 ) − u1 (t − 2τ1 )] Thus, using (23) and (24), results in
and the predictive state xp (t) is defined as V̇ζ (t) ≤ − ∥ζ (t)∥2 + 2 ∥ζ (t)∥ ×
xp (t) =ξ (t) + x̄(t) − ξ (t − τ1 ) 0
 [ ∫ ]
PB ∆ατ ,eq (t) − K0 e−Aθ ∆
¯ ατ ,eq (t + θ )dθ 
 
∫ 0  
−τ1
ξ (t) =eĀτ1 x̄(t) + e−Āθ B̄u0 (t + θ )dθ.
−τ1 ≤ −(1 − β0 ) ∥ζ (t)∥ , ∀ ∥ζ (t)∥ ≥ µ0
2

The values of parameters for simulation are K0 = −[148 − 12 21], with

δ1 = 0.50, δ2 = 3.50, k1 = 6, k2 = 2.20, k3 = 56.58 and 2γ0 τ1 ∥PB∥
k4 = 27.33. In Fig. 1, the behavior of the system state solution norm µ0 = , 0 < β0 < 1, (A.1)
(33) ∥z(t)∥ is presented and the comparison with Loukianov et al.
β0
(2006). In Fig. 2, it can be seen that the control component u1 (t) where γ0 = γ1 + γ4 τ1 . Consequently, a solution of the uncertain
tends to the disturbance −α (z(t), z(t − τ0 ), t), reducing its effect. system (22) is ultimately bounded by Khalil (1996)
In the proposed control algorithm, the implementation of the √
integral part (5) is not easy. Therefore, to clarify the simulation λmax (P)
∥ζ (t)∥ ≤ δ0 , δ0 = µ0 . (A.2)
process, the integral presented in the predictor were solved in λmin (P)
this paper using the Simpson’s Rule 1/3. Also, several integral
approximation methods proposed in Assche, Dambrine, Lafay, and Appendix B. Proof of Theorem 2
Richard (1999) can be used.
The corresponding prediction errors are presented in Remark 1.
5. Conclusions and future works To analyze and compare the solutions of the systems (25) and (26),
the following Lyapunov function candidates are used:
In this paper, a new robust predictor-based SM controller for
uncertain linear systems with delay in the state and the input Vxp (t) = xp (t)T P2 xp (t) and Vξ (t) = ξ (t)T P2 ξ (t), (B.1)
414 H. Caballero-Barragán et al. / Automatica 94 (2018) 409–415

where P2 > 0 is the positive definite solution of the Lyapunov gives

equation ĀT P2 + P2 Ā = −I.
Now, using (B.1) the solutions of the systems (25) and (26) can ẋ1 (t) = A11 x1 (t) + D12 x̄2 (t − τ0 ). (C.3)
be estimated, respectively, as follows (Khalil, 1996) Then, a predictor for x1 (t + τ1 ) is designed of the form
xp (t) ≤ xp (0) e− η2 t + bxp , ∀t ≥ t0 , 0
    ∫
(B.2) A11 τ0
x̄1 (t) = e x1 (t) + e−A11 θ D12 x̄2 (t + θ )dθ. (C.4)
η −τ0
∥ξ (t)∥ ≤ ∥ξ (0)∥ e −2t
+ bξ , ∀t ≥ t 0 , (B.3)
Differentiating (C.4) and using (C.3) lead to
where η = (1 − θ )/λmax (P2 ), 0 < θ < 1,
√ x̄˙ 1 (t) = A11 x̄1 (t) + D12 x̄2 (t). (C.5)
2 ∥P2 ∥ λmax (P2 )
∥B∥ [γ1 + eAτ1  γ2 τ1 ]τ1
 
bxp = Now, taking the time derivative of x̄2 (C.2), results in
θ λmin (P2 )
√ (B.4) x̄˙ 2 (t) = E x̄1 (t − τ0 ) + A22 x̄2 (t) + F x̄2 (t − τ0 )
(C.6)
2 ∥P2 ∥ λmax (P2 ) + Hx1 (t − τ0 ) + B2 [u(t − τ1 ) − α (x(t), x(t − τ0 ), t)],
∥B∥ eAτ1  γ1 τ1 .
 
bξ =
θ λmin (P2 )
where E = A21 − A22 D̄11 + D̄11 A11 , F = D22 + D̄11 D12 and H =
Substituting now (B.2) in (19) and using (23), the solution of (21) D21 − D22 D̄11 .
obtained in this paper can be estimated by Formulating the control u(t) as in (2)
 η
∥ζ (t)∥ ≤ xp (0) e− 2 t , ∀ t0 ≤ t ≤ t1

u(t) =u0 (t) + u1 (t) + u2 (t),
(B.5)
∥ζ (t)∥ ≤ l1 , ∀ t > t1 ,
2 [−E x̄1 (t − ∆) − F x̄2 (t − ∆)
1
u2 (t) =B− (C.7)
where l1 = ∥B∥ [(µ1 eAτ1  + ϱ)γ2 τ1 + µ1 γ1 ]τ1 is the ultimate
 
− Hx1 (t − ∆)] = T X̄ (t − ∆), ∆ = τ0 − τ1 ,
bound reached
∫ for this
 work, √
2 ∥ λmax (P2 ) and using (C.7) in (C.6), yields
ϱ =  −τ e−Aθ dθ  and µ1 = 2∥P√
 0
.

1 θ λmin (P2 )
Now, using (B.3), (20) and (23), the solution of (21) which was x̄˙ 2 (t) =A22 x̄2 (t) + B2 [u0 (t − τ1 ) + u1 (t − τ1 )
(C.8)
derived in Loukianov et al. (2006) can be expressed as + α (x(t), x(t − τ0 ), t)].
η
∥ζ (t)∥ ≤ ∥ξ (0)∥ e− 2 t , ∀ t0 ≤ t ≤ t1 Therefore, Regular form (28)–(29) can be transformed via (C.5)–
(B.6)
∥ζ (t)∥ ≤ l2 , ∀ t > t1 , (C.8) in the following system:
˙ = Āx̄ + B̄[u0 (t − τ1 ) + u1 (t − τ1 ) + α (x(t), x(t − τ0 ), t)]
x̄(t)
where l2 = ∥B∥ (µ1 eAτ1  + ϱ)γ1 τ1 is the ultimate bound reached
 
in the work (Loukianov et al., 2006). which is delay-free in the state vector.
If l1 < l2 then
l2 > l1 References
 Aτ 
∥B∥ (µ1 e + ϱ)γ1 τ1 > ∥B∥ [(µ1 eAτ1  + ϱ)γ2 τ1
 
 1 
Assche, V. V., Dambrine, M., Lafay, J., & Richard, J.-P. (1999). Some problems arising in
+ µ1 γ1 ]τ1 the implementation of distributed delay control laws. In Conference on decision
 Aτ  & control, Vol. 5 (pp. 4669–4672).
(µ1 e + ϱ)γ1 > (µ1 eAτ1  + ϱ)γ2 τ1 + µ1 γ1 ,
 
1
Drazenovic, B. (1969). The invariance conditions for variable structure systems.
 
Automatica, 5(3), 287–295.
now, the terms are grouped up Edwards, C., & Spurgeon, S. K. (1998). Sliding mode control: Theory and applications.
(µ1 eAτ1  + ϱ)γ1 − µ1 γ1 > (µ1 eAτ1  + ϱ)γ2 τ1
London, UK: Taylor & Francis.
   
Fiagbedzi, Y. A., & Pearson, A. E. (1986). Feedback stabilization of linear autonomous
µ1 eAτ1  + ϱ − µ1 + µ1
(   )
γ1 time lag systems. IEEE Transactions on Automatic Control, 31(9), 847–855.
> τ1 Fridman, E. (2014). Tutorial on Lyapunov-based methods for time-delay systems.
γ2 µ1 eAτ1  + ϱ − µ1
 
European Journal of Control, 20(6), 271–283.
( ) Fridman, L., & Levant, A. (2002). Sliding mode control in engineering: Chapter 3.-
γ1 µ1 Higher order sliding modes. New York, USA: Marcel Dekker.
> τ1 1 + , Fridman, L. M., Shustin, E. I., & Fridman, E. M. (1997). Steady modes in the relay
γ2 µ1 eAτ1  + ϱ − µ1
 
control systems with time delay and periodic disturbances. In 1st international
conference control of oscillations and chaos, 1997. Vol. 1 (pp. 75–78).
then, results in the condition (27) Furukawa, T., & Shimemura, E. (1983). Predictive control for systems with time
γ1 delay. International Journal of Control, 37(2), 399–412.
( )
l3
> τ1 1 + , Khalil, H. K. (1996). Nonlinear systems. New Jersey, USA: Prentice Hall.
γ2 l4 Krstic, M. (2009). System and control: Foundations and applications. Delay compensa-
√ √ tion for nonlinear, adaptive, and PDE systems. New York, USA: Springer.
with l3 = 2 ∥P2 ∥ λmax (P2 ) and l4 = l3 (eAτ1  − 1) + θϱ λmin (P2 ),
 
Kwon, W., & Pearson, A. (1980). Feedback stabilization of linear systems with
then one gets l1 < l2 . delayed control. IEEE Transactions on Automatic Control, 25(2), 266–269.
This proves Theorem 2. Léchappé, V., Moulay, E., Plestan, F., Glumineau, A., & Chriette, A. (2015). New
predictive scheme for the control of LTI systems with input delay and unknown
disturbances. Automatica, 52, 179–184.
Appendix C. Proof of Lemma 1
Loukianov, A. G., Castillo-Toledo, B., Hernández, J. E., & Núñez, E. (2003). Discontin-
uous regulator for a class of linear time delayed systems. In European control
Under Assumption 6, the subsystem (28) is represented as conference, Cambridge (pp. 65–70).
Loukianov, A. G., & Escoto-Hernandez, J. (2000). Block control of linear time invari-
ẋ1 (t) = A11 x1 (t) + D12 [D̄11 x1 (t − τ0 ) + x2 (t − τ0 )]. (C.1) ant systems with delay. In 39th IEEE conference on decision and control, Vol. 1 (pp.
462–463).
Selecting the virtual control x2 (t) in (C.1) of the form Loukianov, A. G., Espinosa-Guerra, O., Castillo-Toledo, B., & Utkin, V. A. (2005).
Deadtime block compensation sliding mode control of linear system with delay.
x2 (t) = −D̄11 x1 (t) + x̄2 (t), x̄2 ∈ Rm (C.2) In 16th IFAC world congress, Prague, Vol. 16 (1) (pp. 418–518).
 H. Caballero-Barragán et al. / Automatica 94 (2018) 409–415
Loukianov, A. G., Espinosa-Guerra, O., Castillo-Toledo, B., & Utkin, V. A. (2006).
Integral sliding mode control for systems with time delay. In Proceedings of the
IEEE international workshop on variable structure systems (pp. 256–261).
Loukianov, A. G., & Utkin, V. (1981). Methods of reducing equations for dynamic
systems to a regular form. Automation and Remote Control, 42(4), 413–420.
Manitius, A., & Olbrot, A. W. (1979). Finite spectrum assignment problem for
systems with delays. IEEE Transactions on Automatic Control, 24(4), 541–552.
Mazenc, F., Niculescu, S. I., & Krstic, M. (2012). Lyapunov-Krasovskii functionals
and application to input delay compensation for linear time-invariant systems.
Automatica, 48(7), 1317–1327.
Moreno, J. A., & Osorio, M. (2008). A Lyapunov approach to second-order sliding
mode controllers and observers. In 47th conference on decision & control (pp.
2856–2861).
Nagesh, I., & Edwards, C. (2014). A multivariable super-twisting sliding mode
approach. Automatica, 50(3), 984–988.
Nguang, S. K. (2001). Comments on robust stabilization of uncertain input-delay
systems by sliding mode control with delay compensation. Automatica, 37(10),
1380–1396.
Polyakov, A. (2012). Minimization of disturbances effects in time delay predictor-
based sliding mode control systems. Journal of the Franklin Institute, 349(4),
1380–1396.
Roh, Y. H., & Oh, J. H. (1999). Robust stabilization of uncertain input delay systems
by sliding mode with delay compensation. Automatica, 35, 1861–1865.
Smith, O. J. M. (1957). Closer control of loops with dead time. Chemical Engineering
Progress, 53(5), 217–219.
Utkin, V. I., Guldner, J., & Shi, J. (1999). Sliding modes in electromechanical systems.
London, UK: Taylor & Francis.
Humberto Caballero-Barragán was born in Caleta de
Campos, Michoacan, Mexico, in 1990. He received the B.Sc.
degree in electronics engineering from the Technological
Institute of Lazaro Cardenas, Michoacan, Mexico, and M.Sc.
degree in electrical engineering from the CINVESTAV-IPN
(Advanced Studies and Research Center of the National
Polytechnic Institute), Guadalajara Campus, Mexico, in
2013 and 2015 respectively. He is currently working for
the D.Sc. degree in CINVESTAV-IPN, Guadalajara Campus,
Mexico. His main research interests include time delay
systems, robust control and robotic systems.
415
Linda Osuna-Ibarra graduated from UPSIN (Polytechnic
University of Sinaloa), Mazatlan, Mexico, in 2012 receiv-
ing the B.Sc. degree in Mechatronics Engineering. She
received the M.Sc. degree in Electrical Engineering from
CINVESTAV-IPN (Advanced Studies and Research Center of
the National Polytechnic Institute), Guadalajara, Mexico,
in 2015. She is currently working to get the D.Sc. de-
gree from CINVESTAV-IPN Guadalajara, Mexico. Her main
research interests include discrete-time systems, robust
control, and robotics.
Alexander Loukianov graduated from the Polytechnic,
Moscow, Russia, in 1975, and received the Ph.D. degree
in Automatic Control from the Institute of Control Sci-
ences, Russian Academy of Sciences, Moscow, in 1985. He
was with the Institute of Control Sciences during 1978–
1997. Since April 1997, he has been with CINVESTAV-IPN
(Advanced Studies and Research Center of the National
Polytechnic Institute), Guadalajara, Mexico, as a Profes-
sor of electrical engineering graduate programs. He has
published more than 250 technical papers in international
journals, books, and conference proceedings; and he has
served as a Reviewer for various international journals and conferences. His re-
search interests include nonlinear robust control and VSS with sliding modes
applied to robotics, automotive control, electrical drives and power systems control.
Dr. Loukianov is an Associate Editor of the International Journal of Control Theory
and Applications.
Franck Plestan received the Ph.D. in Automatic Control
from the Ecole Centrale de Nantes, France, in 1995. From
September 1996 to August 2000, he was with Louis Pas-
teur University, Strasbourg, France. In September 2000,
he joined the Ecole Centrale de Nantes, Nantes, France,
where he is currently Professor. His research interests
include robust nonlinear control and observer design, and
more recently time-delay systems, with several fields of
applications (electropneumatic actuators, flying systems,
renewable energy systems).