Automatica ( ) –

Contents lists available at ScienceDirect

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

Technical Communique

Barrier function-based adaptive sliding mode control✩

Hussein Obeid a , Leonid M. Fridman b, *, Salah Laghrouche a , Mohamed Harmouche c
a
Femto-ST UMR CNRS, Univ. Bourgogne Franche-Comté/UTBM, 90010, Belfort, France
b
Departement of Robotics and Control, Engineering Faculty, Universidad Nacional Autónoma de México (UNAM), D.F 04510, Mexico
c
Actility, Paris, France

article info a b s t r a c t
Article history: In this paper, a new barrier function-based adaptive strategy is proposed for first order sliding mode
Received 19 August 2017 controller. This strategy is applied to a class of first order disturbed systems whose disturbance is bounded
Received in revised form 2 December 2017 with unknown boundary. The proposed barrier strategy can ensure the convergence of the output variable
Accepted 9 February 2018
and maintain it in a predefined neighborhood of zero independent of the upper bound of the disturbance,
Available online xxxx
without overestimating the control gain.
Keywords: © 2018 Elsevier Ltd. All rights reserved.
Sliding mode
Adaptive control
Barrier functions

1. Introduction
For systems with matching disturbances, the sliding mode con-
trol has proven its high efficiency (Utkin, 1992). Indeed, it provides
a closed-loop insensitivity to these disturbances and guarantees
the finite-time convergence. However, the implementation of the
first order sliding mode controllers (FOSMCs) requires the knowl-
edge of the upper bound of disturbances. In practice, this bound is
not constant and, moreover, frequently it is unknown. This means
that the gains of the FOSMCs are overestimated. This is a main
obstacle in the FOSMCs implementation growing the undesired
chattering effect (Boiko, 2008).
Recently, two different strategies to create adaptive sliding
mode (SM) controllers have been considered in the case where the
upper bound of the disturbance exists but it is unknown.
The first strategy of adaptation consists in increasing the gain
until the moment when the SM is reached, and then the gain is
fixed at this value, ensuring an ideal SM for some interval of time.
When the disturbance grows, the SM can be lost, therefore the
gain increases to reach it again (Incremona, Cucuzzella, & Ferrara,
✩ The authors are grateful for the financial support from projects RECH-
MOB15000008 of Franche-Comté Regional Council (France), CONACyT(Consejo
Nacional de Ciencia y Tecnologia): 282013; PAPIIT-UNAM (Programa de Apoyo a
Proyectos de Investigación e Innovación Tecnológica) IN 113216 and PASPA-UNAM
(PROGRAMA DE APOYOS PARA LA SUPERACIÓN DEL PERSONAL ACADÉMICO DE LA
UNAM). The material in this paper was not presented at any conference. This paper
was recommended for publication in revised form by Associate Editor Maria Letizia
Corradini under the direction of Editor André L. Tits.
*
Corresponding author.
E-mail addresses: hussein.obeid@utbm.fr (H. Obeid), Lfridman@unam.mx
(L.M. Fridman), salah.laghrouche@utbm.fr (S. Laghrouche),
mohamed.harmouche@actility.com (M. Harmouche).
2016; Negrete-Chvez & Moreno, 2016). However, the FOSMC gain
in this strategy is overestimated and one cannot be sure that the
SM will not be lost in the future. To overcome this problem, an
approach based on increasing and decreasing the gain has been
developed (Bartolini, Levant, Plestan, Taleb, & Punta, 2012; Incre-
mona et al., 2016; Plestan, Shtessel, Bregeault, & Poznyak, 2010;
Shtessel, Taleb, & Plestan, 2012). This approach ensures the finite-
time convergence of the sliding variable to some neighborhood of
zero without big overestimation of the gain. The main drawback of
this approach is that the size of the above mentioned neighborhood
and the time of convergence depend on the unknown upper bound
of disturbance, i.e. they are unknown a priori and one can never be
sure that SM will never be lost for bigger values of time.
The second strategy of adaptation is based on the usage of
the equivalent control value as an estimation of the disturbance
(Bartolini, Ferrara, Pisano, & Usai, 1998; Edwards & Shtessel, 2016;
Oliveira, Cunha, & Hsu, 2016; Utkin & Poznyak, 2013). In Oliveira
et al. (2016) a model based approach is presented. To realize
this strategy a low-pass filtered approximation of the equivalent
control were proposed. However, during the realization, the filter
constant should be chosen much less than the inverse of the upper
bound of the first derivative of disturbance.
The aim of this paper is to propose an adaptive strategy that
can achieve the convergence of the output variable to a predefined
neighborhood of zero, with a control gain that is not overestimated,
and without using any information about the upper bound of the
disturbance, nor the use of the low pass filter.
This paper proposes the use of Barrier Functions (BFs) as an
adaptive strategy for FOSMC in order to reach above mentioned
goal. In this current paper, two different classes of BFs are used:
the positive semi-definite BF and a positive definite BF.

https://doi.org/10.1016/j.automatica.2018.03.078
0005-1098/© 2018 Elsevier Ltd. All rights reserved.

Please cite this article in press as: Obeid, H., et al., Barrier function-based adaptive sliding mode control. Automatica (2018),
https://doi.org/10.1016/j.automatica.2018.03.078.
2 H. Obeid et al. / Automatica (
The main advantages of the proposed barrier adaptive SM con-
trol are:
• The output variable converges in a finite time to a predefined
neighborhood of zero, independently of the bound of the
disturbance, and cannot exceed it.
• The gain provided by the proposed strategy is not overesti-
mated, as it can only achieve the convergence of the output
variable to a predefined neighborhood of zero.
• The proposed strategy theoretically does not require neither
the bounds of the disturbance nor the use of the low-pass filter.
This paper is organized as follows. In Section 2 the problem for-
mulation is given. Section 3 presents the barrier adaptive FOSMC
algorithm. Finally, some conclusions are drawn in Section 4.
2. Problem formulation
Consider the first order system
ṡ(t) = u(t) + δ (t),
where s(t) ∈ R is the output variable, u(t) is the FOSMC and δ (t) is
a disturbance. Here δ (t) is bounded function with unknown bound,
i.e. |δ (t)| ≤ δmax . The bound δmax > 0 exists but is not known.
In this context, the gain of the FOSMC is to be adapted in ac-
cordance with the adaptive strategy defined later. The idea behind
the proposed adaptive strategy is to first increase the adaptive gain
until the output variable reaches a small neighborhood of zero 2ε at
time t̄ by using a constant derivative gain as in Plestan et al. (2010).
Secondly, for t > t̄, the adaptive gain switches to a BF that can
maintain the output variable in the predefined neighborhood of
zero |s(t)| < ε .
2.1. Preliminaries
2.1.1. Barrier functions (BFs)
Definition 1. Let us suppose that some ε > 0 is given and fixed,
the BF can be defined as an even continuous function Kb : x ∈
]−ε, ε [ → Kb (x) ∈ [b, ∞[ strictly increasing on [0, ε [ .
• lim|x|→ε Kb (x) = +∞.
• Kb (x) has a unique minimum at zero and Kb (0) = b ≥ 0.
In this paper, the following two different classes of BFs are
considered;
ε F̄
• Positive definite BFs (PBFs): Kpb (x) = ε−|x|
, i.e. Kpb (0)
0.
• Positive Semi-definite BFs (PSBFs): Kpsb (x) =
Kpsb (0) = 0.
The PBF Kpb (x) and the PSBF Kpsb (x) are illustrated in Fig. 1.
3. Barrier adaptive FOSMC algorithm
The following theorem is true for both possible FOSMC gains
design: using KB (s(t)) = Kpb (s(t)) and KB (s(t)) = Kpsb (s(t)).
Theorem 2. Consider system (1) with bounded disturbance δ (t) with
the controller
u(t) = −K (t , s(t))sign(s(t)),
and with the adaptive control gain K (t , s)
Ka (t), K̇a (t) = K̄ |s(t)|, if 0 < t ≤ t̄
{
K (t , s(t)) =
KB (s(t)), if t > t̄
where K̄ to be arbitrary positive constant.
Please cite this article in press as: Obeid, H.,
https://doi.org/10.1016/j.automatica.2018.03.078.
) –
(a) Kpb (x). (b) Kpsb (x).
Fig. 1. Schematic illustrations of Kpb (x) and Kpsb (x).
Then for any s(0) and ε > 0, there exists t̄ the smallest root of
equation |s(t)| ≤ 2ε such that for all t ≥ t̄, the inequality |s(t)| < ε
(1) holds.
The proof of Theorem 2 is given in the Appendix.
Remark 3. Note that this strategy allows the adaptive gain to
increase and decrease based on the current value of the output
variable. When the output variable is going to zero, the adaptive
gain decreases till the value which allows to compensate the dis-
turbance.
On the other hand, when the disturbance grows and the control
gain is less than the absolute value of disturbance, the output
variable grows and the control gain can grow if it is necessary till
the level ensuring that the system solution will never leave the ε
vicinity of zero.
Remark 4. Theoretically, the a priori knowledge of actuator capac-
ity P is not required, but it should be supposed that the actuator
is capable to compensate the disturbance. However, in practice, an
actuator is used and its capacity P is known. In this case for discrete
implementation of the proposed algorithm, the sampling step τ
should be chosen as τ ≪ ε/P . Otherwise, the attractive feature of
the BF will be lost, and the output variable will leave the predefined
neighborhood of zero.
The behavior of each barrier function PBF and PSBF, and the
achievement of real or ideal SM in finite time, together with the
continuity or discontinuity of the control signal are discussed in
Sections 3.1 and 3.2.
= F̄ >
3.1. Adaptation with PBF
|x|
ε−|x|
, i.e.
Consider the adaptation with PBF. In this case, Kpb (s(t)) has a
lower bound F̄ when s(t) = 0. Therefore, when |δ (t)| < F̄ the
adaptive gain is overestimated. In this case, this strategy provides
an ideal SM. In order to attenuate this overestimation, F̄ can be
chosen small enough. The usage of PBF when the bound of the
disturbance is less than F̄ will provide a discontinuous control
signal leading to the chattering whose amplitude is proportional
to the choice of F̄ .
3.2. Adaptation with PSBF
Consider now the adaptation with PSBF. In this case, Kpsb (s(t))
(2) tends to zero when s(t) → 0. Hence, Kpsb (s(t)) has the same
|s(t)| |s(t)|
behavior as ε in the neighborhood of zero, i.e. ε ≪ 1 →
Kpsb (s(t)) = ε−|s(t)| ≈ ε .
|s(t)| |s(t)|
(3) This means that if δ (t) and s(t) tend monotonically to zero,
consequently the adaptive gain Kpsb (s(t)) will go to zero. The dis-
continuity of the control signal can appear only once at time t̄,
et al., Barrier function-based adaptive sliding mode control. Automatica (2018),
 H. Obeid et al. / Automatica ( ) – 3

Table 1
Parameter values of the PBF, PSBF and APS.
PBF PSBF APS
Parameter values K̄ = 1000, ε = 0.02, K̄ = 1000, ε = 0.02, K̄ = 1000, ε = 0.02,
F̄ = 0.1, Ka (0) = 10 Ka (0) = 10 µ = 0.1 , K (0) = 10

(a) (b)

(c) (d)

Fig. 2. (a) Disturbance δ (t), and the zoomed-in plot of s(t) with: (b) PBF, (c) PSBF, (d) APS.

when the adaptive gain switches to PSBF. It is necessary to note
that starting from time t̄, the control signal becomes continuous.
3.3. Simulation results
Two cases are considered carrying out simulations in order to
show two main advantages of the proposed adaptive strategy over
the adaptive algorithm presented in Plestan et al. (2010).
For the barrier strategy, the adaptive gain is given by Eq. (3) with
KB (s(t)) once defined as PBF and then as PSBF. In contrast, for the
adaptive algorithm in Plestan et al. (2010), referred to as (APS), the
adaptive gain is computed through
K̄ |s(t)|sign(|s(t)| − ε ), if K > µ
{
K̇ (t) =
µ, if K ≤ µ
where K̄ , ε and µ are positive constants to be selected.
Consider the first order system (1) with the initial value as
s(0) = 0.04. Table 1 contains the parameter values of PBF, PSBF
and APS, where the parameter ε is selected as ε = 0.02, while all
the others are set according to (Plestan et al., 2010). The attention
will be focused on the behavior of each adaptation for t > t̄, since
starting from this moment the adaptive gain switches to the BFs.
3.3.1. Case of increasing disturbance amplitude
In this subsection, the disturbance is given by
0.3sin(t), if t ≤ 100 s
{
δ (t) = 1.4sin(t), if 100 s < t ≤ 200 s
6.2sin(t), if t > 200 s.
This means that the disturbance amplitude is δmax = 0.3 for
t ≤ 100 s, δmax = 1.4 for 100 < t ≤ 200 s, and δmax = 6.2
for t > 200 s (see Fig. 2(a)). The zoomed-in plots of the output
variable s(t) with PBF, PSBF and APS are depicted in Figs. 2(b)–
2(c)–2(d). It can be noticed in Figs. 2(b)–2(c) that for both barrier
functions PBF and PSBF, the output variable does not exceed the
predefined neighborhood of zero ε = 0.02. On the other hand, it
can be seen in Fig. 2(d) that the size of the neighborhood of zero
to which converges s(t) with APS is changing together with the
amplitude of disturbance δmax (for t ≤ 100 s, |s(t)| ≤ 0.024, for
100 < t ≤ 200 s, |s(t)| < 0.029, and for t > 200 s |s(t)| ≤ 0.031).
Therefore, it cannot be predefined. Also, it is worth noting that
when the amplitude of disturbance suddenly increases to a big
value, the output variable will also jump to a big value (t = 200 s,
(4)
|s(t)| > 0.04).
3.3.2. Case of monotonically decreasing disturbance
The following monotonically decreasing disturbance is consid-
ered (Fig. 3(a))
8.2
δ (t) = + 0.002.
t +1
Figs. 3(b)–3(c)–3(d) show the behavior of the control signal
u(t) with PBF, PSBF and APS. For t > 83 s the amplitude of the
disturbance becomes less than F̄ = µ = 0.1 (see Fig. 3(a)). That is
why starting for this moment, the chattering will appear for both
PBF and APS with the amplitude equal to 0.1 (see Figs. 3(b)–3(d)).
On the other hand, the control signal for PSBF is continuous and
will go to zero without chattering (see Fig. 3(c)).

Please cite this article in press as: Obeid, H., et al., Barrier function-based adaptive sliding mode control. Automatica (2018),
https://doi.org/10.1016/j.automatica.2018.03.078.
4 H. Obeid et al. / Automatica ( ) –

(a) (b)

(c) (d)

Fig. 3. (a) Disturbance δ (t), and the control u(t) with: (b) PBF, (c) PSBF, (d) APS.

( )
4. Conclusion Proof. The time derivative of V s(t), KB (s(t)) is given as follows

This paper presents a barrier strategy to adapt the gain of V̇ = sδ − sKB (s)sign(s) + (KB (s) − KB (0))K̇B (s)
FOSMC for the case of first order disturbed system with the up- (A.2)
per bound of the disturbance is unknown. This strategy ensures ≤ δmax |s| − KB (s)|s| + (KB (s) − KB (0))K̇B (s),
the convergence of the output variable and prevents its violation since KB (s(t)) = Kpb (s(t)), one gets
outside a predefined neighborhood of zero. Moreover, the adaptive
gain in this strategy is not overestimated. V̇ ≤ − (−δmax + Kpb (s))|s|
  
βs
Appendix. Proof of Theorem 2
εF̄
− (−δmax + Kpb (s))|Kpb (s) − F̄ | (A.3)
Proof. According to Plestan et al. (2010) the system solution will (ε − |s(t)|)2 
reach 2ε in finite time.
 
   βs
ζ >0
Denote t = t̄ the smallest root of equation |s(t)| ≤ 2ε , and
consider the Lyapunov candidate function containing both the = −βs |s| − ζ βs |Kpb (s) − F̄ |.
output variable and the adaptive gain dynamics
• Suppose that F̄ < δmax , and consider the case when |s(t)| >
( ) 1 1( )2
s1 . Then, taking into account that Kpb (s(t)) is an increasing
V s(t), KB (s(t)) = s2 (t) + KB (s(t)) − KB (0) .
2 2 function on [0, ε[, it yields Kpb (s(t)) > Kpb (s1 ) = δmax . This
leads to βs > 0. Therefore
A.1. Adaptation with PBF
V̇ ≤ −βs |s| − ζ βs |Kpb (s) − F̄ |
Consider the case when KB (s(t)) = Kpb (s(t)). It is shown that if √ ( |s| |Kpb (s) − F̄ | )
|s(t̄)| > s1 where = −βs 2 √ + ζ √
⎧ 2 2
F̄ (A.4)
⎨
ε (1 − ), if F̄ < δmax √ ( |s| |Kpb (s) − F̄ | )
s1 = δmax (A.1) ≤ −βs 2 min{1, ζ } √ + √
⎩
0, if F̄ ≥ δmax 2 2
1 √
then PBF ensures that |s(t)| ≤ s1 in a finite time period τ1 . ≤ −β1 V , 2 with β1 = βs 2 min{1, ζ }.
Furthermore, it is proven that for all t ≥ t̄ + τ1 , the output variable
will remain inside |s(t)| ≤ s1 < ε . Note that τ1 = 0 if |s(t̄)| ≤ s1 . • Suppose that F̄ ≥ δmax and |s(t)| > s1 = 0. Thus, βs > 0 since
Kpb (s(t)) > Kpb (0) = F̄ ≥ δmax (see Eq. (A.3)). It yields that
1
Lemma 5. Given the output variable dynamics (1) controlled by (2) V̇ ≤ −β1 V 2 .
and (3) with KB (s(t)) = Kpb (s(t)). Then, for all t ≥ t̄, and for all
Therefore, finite time convergence to the domain |s(t)| ≤ s1 is
|s(t)| > s1 , the time derivative of the Lyapunov function satisfies the
ensured, and the reaching time τ1 can be estimated as
following inequality
( ) ( ) ( ( ) 21 ( ) 12 )
1
V̇ s(t), Kpb (s(t)) ≤ −β1 V 2 s(t), Kpb (s(t)) , with β1 > 0 2 −V s1 , Kpb (s1 ) + V s(t̄), Kpb (s(t̄))
τ1 ≤ . □
which yields a finite time convergence to the domain |s(t)| ≤ s1 . β1

Please cite this article in press as: Obeid, H., et al., Barrier function-based adaptive sliding mode control. Automatica (2018),
https://doi.org/10.1016/j.automatica.2018.03.078.
 H. Obeid et al. / Automatica ( ) – 5

Now, the case when t ≥ t̄ + τ1 , and |s(t)| < s1 is studied. In V̇ ≤ −βs |s| − ζ βs |Kpsb (s)|
this case, V̇ would be sign indefinite (see Eq. (A.3)), and s(t) can go
toward s1 . Please note that at the time instant when |s(t)| reaches
√ ( | s| |Kpsb (s)| )
= −βs 2 √ + ζ √
s1 , V̇ ≤ 0 since βs = 0, and V would be constant or decreasing. This 2 2
means that for all t ≥ t̄ + τ1 , |s(t)| ≤ s1 . √ ( |s| (A.7)
|Kpsb (s)| )
Finally, as |s(t̄)| ≤ 2ε , it becomes evident that for all t ≥ t̄ the ≤ −βs 2 min{1, ζ } √ + √
inequality |s(t)| < ε holds. 2 2
1 √
≤ −β2 V , 2 with β2 = βs 2 min{1, ζ }.
A.2. Adaptation with PSBF Therefore, finite time convergence to the domain |s(t)| ≤ s2 is
ensured, and the reaching time τ2 can be estimated as
Consider now the case when KB (s(t)) = Kpsb (s(t)). Following the ( ( ) 12 ( ) 21 )
same proof and replacing Lemma 5 by the following Lemma 6, it is 2 −V s2 , Kpsb (s2 ) + V s(t̄), Kpsb (s(t̄))
easy to demonstrate that if |s(t̄)| > s2 where
τ2 ≤ . □
δmax β2
s2 = ε ( ) (A.5)
δmax + 1 Theorem 2 is proven. □
then PSBF ensures that |s(t)| ≤ s2 in a finite time period τ2 .
Moreover, for all t ≥ t̄ + τ2 the output variable will remain inside References
|s(t)| ≤ s2 < ε. Note that τ2 = 0 if |s(t̄)| ≤ s2 .
Bartolini, G., Ferrara, A., Pisano, A., & Usai, E. (1998). Adaptive reduction of the
control effort in chattering-free sliding-mode control of uncertain nonlinear
systems. Applied Mathematics and Computer Science, 8(1), 51–71.
Lemma 6. Given the output variable dynamics (1) controlled by (2) Bartolini, Giorgio, Levant, Arie, Plestan, Franck, Taleb, Mohammed, & Punta,
and (3) with KB (s(t)) = Kpsb (s(t)). Then, for all t ≥ t̄, and for all Elisabetta (2012). Adaptation of sliding modes. IMA Journal of Mathematical
|s(t)| > s2 , the time derivative of the Lyapunov function satisfies the Control and Information, 30(3), 285–300.
Boiko, Igor (2008). Discontinuous control systems: frequency-domain analysis and
following inequality
design. Springer Science & Business Media.
( ) ( )
1 Edwards, Christopher, & Shtessel, Yuri B. (2016). Adaptive continuous higher order
V̇ s(t), Kpsb (s(t)) ≤ −β2 V 2 s(t), Kpsb (s(t)) , with β2 > 0 sliding mode control. Automatica, 65, 183–190.
Incremona, Gian Paolo, Cucuzzella, Michele, & Ferrara, Antonella (2016). Adaptive
which yields a finite time convergence to the domain |s(t)| ≤ s2 . suboptimal second-order sliding mode control for microgrids. International
Journal of Control, 89(9), 1849–1867.
Negrete-Chvez, Daniel Y., & Moreno, Jaime A. (2016). Second-order sliding mode
Proof. Taking into account that KB (s(t)) = Kpsb (s(t)), the upper output feedback controller with adaptation. International Journal of Adaptive
bound of Eq. (A.2) can be derived as Control and Signal Processing, 30(8–10), 1523–1543.
Oliveira, Tiago Roux, Cunha, José Paulo VS, & Hsu, Liu (2016). Adaptive sliding
V̇ ≤ − (−δmax + Kpsb (s))|s| mode control for disturbances with unknown bounds. In 2016 14th international
workshop on variable structure systems (pp. 59–64). IEEE.
  
βs Plestan, Franck, Shtessel, Yuri, Bregeault, Vincent, & Poznyak, Alexander (2010).
New methodologies for adaptive sliding mode control. International Journal of
ε Control, 83(9), 1907–1919.
− (−δmax + Kpsb (s))|Kpsb (s)| (A.6)
(ε − |s(t)|)2    Shtessel, Yuri, Taleb, Mohammed, & Plestan, Franck (2012). A novel adaptive-
   βs gain supertwisting sliding mode controller: methodology and application.
ζ >0 Automatica, 48(5), 759–769.
Utkin, V. I. (1992). Sliding modes in optimization and control problems. New York:
= −βs |s| − ζ βs |Kpsb (s)|. Springer Verlag.
Utkin, Vadim I., & Poznyak, Alex S. (2013). Adaptive sliding mode control with
Consider the case when |s(t)| > s2 . Then, Kpsb (s(t)) > Kpsb (s2 ) = application to super-twist algorithm: Equivalent control method. Automatica,
δmax . This leads to βs > 0. Therefore 49(1), 39–47.

Please cite this article in press as: Obeid, H., et al., Barrier function-based adaptive sliding mode control. Automatica (2018),
https://doi.org/10.1016/j.automatica.2018.03.078.