Anda di halaman 1dari 14

This article was downloaded by: [North Carolina State University]

On: 25 November 2012, At: 10:02


Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Control


Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/tcon20

New methodologies for adaptive sliding mode control


a b a c
F. Plestan , Y. Shtessel , V. Brgeault & A. Poznyak
a
Ecole Centrale de Nantes-IRCCyN, Nantes, France
b
The University of Alabama in Huntsville, Huntsville, USA
c
CINVESTAV-IPN, Mexico DC, Mexico
Version of record first published: 21 Jul 2010.

To cite this article: F. Plestan, Y. Shtessel, V. Brgeault & A. Poznyak (2010): New methodologies for adaptive sliding mode
control, International Journal of Control, 83:9, 1907-1919

To link to this article: http://dx.doi.org/10.1080/00207179.2010.501385

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to
anyone is expressly forbidden.

The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug doses should
be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims,
proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly or indirectly in
connection with or arising out of the use of this material.
International Journal of Control
Vol. 83, No. 9, September 2010, 19071919

New methodologies for adaptive sliding mode control


F. Plestana*, Y. Shtesselb, V. Bregeaulta and A. Poznyakc
a
Ecole Centrale de Nantes-IRCCyN, Nantes, France; bThe University of Alabama in Huntsville,
Huntsville, USA; cCINVESTAV-IPN, Mexico DC, Mexico
(Received 28 May 2009; final version received 11 June 2010)

This article proposes new methodologies for the design of adaptive sliding mode control. The goal is to obtain a
robust sliding mode adaptive-gain control law with respect to uncertainties and perturbations without the
knowledge of uncertainties/perturbations bound (only the boundness feature is known). The proposed
Downloaded by [North Carolina State University] at 10:02 25 November 2012

approaches consist in having a dynamical adaptive control gain that establishes a sliding mode in finite time.
Gain dynamics also ensures that there is no overestimation of the gain with respect to the real a priori unknown
value of uncertainties. The efficacy of both proposed algorithms is confirmed on a tutorial example and while
controlling an electropneumatic actuator.
Keywords: sliding mode; adaptive control; pneumatic actuator

1. Introduction as possible whereas sufficient to counteract the


The sliding mode control is a very popular strategy for uncertainties/perturbations. As recalled previously,
control of nonlinear uncertain systems, with a very this problem is an exciting challenge for applications
large frame of applications fields (Slotine and Li 1991; given that, in many cases, gains are also overestimated,
Utkin, Guldner, and Shi 1999). Due to the use of which gives larger control magnitude and larger
discontinuous function and high control gain, its main chattering. In order to adapt the gain, many controllers
features are the robustness of closed-loop system and based on fuzzy tools (Munoz and Sbarbaro 2000; Tao,
the finite-time convergence. However, its design Chan, and Lee 2003) have been published; however,
requires the knowledge of uncertainties bound, which these papers do not guarantee the tracking perfor-
could be, from a practical point of view, a hard task; it mances. In Huang, Kuo, and Chang (2008) gain
often follows that this bound is overestimated, which dynamics directly depends on the tracking error
yields to excessive gain. Then, the main drawback of (sliding variable): the control gain is increasing since
the sliding mode control, the well-known chattering sliding mode is not established. Once this is the case,
phenomenon (for its analysis, see Boiko and Fridman gain dynamics equals 0. The main drawback of this
(2005); Boiko, Fridman, Pisano, and Usai (2007)), is approach is the gain over-estimation with respect to
important and could damage actuators and systems. uncertainties bound (see e.g. in Huang et al. (2008,
A first way to reduce the chattering is the use of a Section III)). Furthermore, this approach is not
boundary layer: in this case, many approaches have directly applicable, but requires modifications for its
proposed adequate controller gains tuning (Slotine and application to real systems: thus, the sign function is
Sastry 1983). A second way to decrease the chattering replaced by a saturation function where the boundary
phenomenon is the use of higher order sliding mode layer width affects accuracy and robustness.
controller (Levant 1993, 2001, 2007; Bartolini, Ferrara, Furthermore, no boundary layer width tuning metho-
Usai, and Utkin 2000; Laghrouche, Plestan, and dology is provided. A method proposed in Lee and
Glumineau 2007; Plestan, Glumineau, and Utkin (2007) in order to limit the switching gain must
Laghrouche 2008). However, in both these control be mentioned. The idea is based on use of equivalent
approaches, knowledge of uncertainties bound is control: once sliding mode occurs, disturbance magni-
required. tude is evaluable and allows an adequate tuning of
As the objective is the non-requirement of the control gain. However, this approach requires the
uncertainties bound, another way consists in using knowledge of uncertainties/perturbations bounds and
adaptive sliding mode, the goal being to ensure a the use of low-pass filter, which introduces signal
dynamical adaptation of the control gain to be as small magnitude attenuation, delay and transient behaviour

*Corresponding author. Email: Franck.Plestan@irccyn.ec-nantes.fr

ISSN 00207179 print/ISSN 13665820 online


 2010 Taylor & Francis
DOI: 10.1080/00207179.2010.501385
http://www.informaworld.com
1908 F. Plestan et al.

when disturbances are acting. In Hall and Shtessel for x 2 X . It is assumed that M, m and M exist but
(2006) a gain-adaptation algorithm is proposed by are not known. The objective of this article is to
using sliding mode disturbance observer. The main propose a new sliding mode controller u(, t) with the
drawback is that the knowledge of uncertainties same features as classical SMC (robustness, finite-time
bounds is required to design observer-based controller. convergence) but without any information on uncer-
The objective of the current article is to propose tainties and perturbations (appearing in f(x)); this
new methodologies for control of a class of uncertain latter is only known to be bounded. Furthermore, this
nonlinear systems for which uncertainties are bounded, objective allows to ensure a global stability of closed-
but this bound (which is finite) is not known. The first loop system whereas the classical way (with knowledge
controller is based on a gain-adaptation law derived of uncertainties bounds) only ensures its semi-global
from coupling of Lee and Utkin (2007) and Huang stability. In the sequel, definitions of ideal and real
et al. (2008) methods, whereas the second one is using sliding mode are recalled.
an original gain-adaptation law. In both the control
Definition 1 (Levant 1993): Consider the nonlinear
algorithms, the gain is dynamically tuned in order to
system (1), and let the system be closed by some
ensure the establishment of a sliding mode; once this
Downloaded by [North Carolina State University] at 10:02 25 November 2012

possibly dynamical discontinuous feedback. Variable 


sliding mode is achieved, the gain is adjusted in order is a continuous function, and the set
to get a sufficient value in order to counteract the
perturbations and uncertainties. S fx 2 X j x, t 0g,
Section 2 states the problem, displays some recalls
on sliding mode and presents a review of Lee and called sliding surface, is non-empty and is locally an
Utkin (2007) and Huang et al. (2008) by addressing integral set in the Filippov sense (Filippov 1988), i.e. it
their deficiencies. Section 3 displays both the proposed consists of Filippovs trajectories of the discontinuous
adaptive sliding mode control algorithms and Section 4 dynamical system. The motion on S is called sliding
details -tuning. In order to prove the feasibility of the mode with respect to the sliding variable .
approaches, both simulations (control of tutorial In real applications, an ideal sliding mode as
system and position control of electropneumatic defined in Definition 1 cannot be established. Then, it
actuator model) are presented in Section 5. Section 6 is necessary to introduce the concept of real sliding
concludes this article. mode.
Definition 2 (Levant 1993): Given the sliding variable
(x, t), the real sliding surface associated with (1) is
2. Preliminaries defined as (with  4 0)
2.1 Problem statement
S  fx 2 X j jj 5 g: 4
Consider the nonlinear uncertain system

x_ f x gx  u 1
Definition 3 (Levant 1993): Consider the non-empty
with x 2 X  IRn the state vector and u 2 IR the control real sliding surface S  given by (4), and assume that it
input. Functions f(x) and g(x) are smooth uncertain is locally an integral set in the Filippov sense. The
functions and are bounded for x 2 X ; furthermore, f(x) corresponding behaviour of system (1) on (4) is called
contains unmeasured perturbations term and g(x) 6 0 real sliding mode with respect to the sliding variable
for x 2 X (i.e. system (1) is controllable for all x 2 X ). (x, t).
The control objective consists in forcing the continuous
function (x, t), named sliding variable, to 0. Supposing
that  admits a relative degree equal to 1 with respect 2.2 Discussion of adaptive sliding mode control
to u, one gets solutions and motivations
@ @ @ @ @ Two adaptive sliding mode controllers have been given
_ x_  f x  gx u
@x @t @t @x @x
|{z} |{z} in Lee and Utkin (2007) and Huang et al. (2008) and
x,t x,t are discussed now. These two approaches follow two
different ways: the first does not require the knowledge
x, t x, t  u: 2
of uncertainties/perturbations bound and consists in
Functions (x, t) and (x, t) are supposed to be increasing the control gain since a sliding mode is
such that established whereas the second one requires the
knowledge of the bound and is using the equivalent
jj  M , 0 5 m    M 3 control concept in order to evaluate and to minimise
International Journal of Control 1909

the control gain. Then, the same objective can be Discussion: The main feature of this controller is the
claimed, which is the dynamical adaptation of the gain adjustment of the control gain by using the equivalent
in order to counteract perturbations and uncertainties. control concept. It implies that chattering is decreas-
ing. However, K-adaptation law requires the
Theorem 1 (Huang et al. 2008): Given the nonlinear
knowledge of uncertainties/perturbations bounds.
uncertain system (1) with the sliding variable (x, t)
Furthermore, the use of low-pass filter introduces in
dynamics (2) controlled by
the closed-loop system dynamics and transient phe-
u Kt  signx, t nomena in case of perturbations. Also note that
methodologies for the tuning of parameters  and 
with the adaptation law K(t) given by have not been detailed except their positivity and the
fact that the time constant of the low-pass filter  must
K_ K  jx, tj 5
be sufficiently small. However, their tuning could
with K 4 0 and K(0) 4 0, then there exists a finite time engender very different behaviours of closed-loop
tF  0 so that a sliding mode is established in system for system.
Downloaded by [North Carolina State University] at 10:02 25 November 2012

all t  tF, i.e. (x, t) 0 for t  tF. Motivations: The main objective of the current article
Discussion: The main feature of this approach is that consists in providing new adaptive sliding mode
it does not require a priori the knowledge of control controllers such that:
gain. However, from K-dynamics, it yields that when . Uncertainties/perturbation bounds exist but
 0, K_ 0. In this case, the gain K is clearly are not known.
overestimated with respect to uncertainties, which . The gain-adaptation law does not overesti-
induces larger chattering (see simulations results in mate uncertainties/perturbations magnitude
Huang et al. (2008, Section III). Furthermore, this and then, the obtained control magnitude is
methodology is applicable only for ideal sliding mode, reasonable.
the objective  0 being reachable. In the case of real
sliding mode, this latter objective is not reachable which A first solution to the problem under interest in this
induces that K gain is always increasing. In Huang et al. article consists in combining the two previous gain-
(2008), the authors propose to modify K-dynamics by adaptation laws: the first one would be used since a
introducing a boundary layer neighbouring the sliding sliding mode is established, whereas the second allows
surface  0. It means that accuracy has to be to decrease the control gain once sliding mode is
sacrificed in order to apply the previous controller established. A second solution is an original adaptive
and that the control gain is still overestimated. Also, real sliding mode control law that does not allow to
with discrete-time measurements, adaptation law (5) overestimate the gain value without using concept of
inevitably implies the tending of K to infinity, if the equivalent control.
sampling interval is separated from zero.
Theorem 2 (Lee and Utkin 2007): Given the nonlinear
uncertain system (1) with the sliding variable (x, t) 3. Main results
dynamics (2) controlled by
In this section, two adaptive sliding mode control
u Kt  signx, t algorithms that establish the sliding modes in SISO
systems with bounded perturbations with unknown
with the adaptation law K(t) given by bounds, are presented. The first algorithm combines
Kt K  jj  the adaptive schemes formulated in Theorems 1 and 2,
Section 2, while mitigating the deficiencies of the
with K 4 0,  4 0 and  the average of sign() obtained combined gain-adaptation schemes. The second adap-
through a low pass-filter tive sliding mode control algorithm is completely
  _  signx, t original. It does not require the estimation of the
perturbations via equivalent control as in Lee and
with  4 0, if Utkin (2007) (Theorem 2) and does not overestimate
  the control gain as in Huang et al. (2008) (Theorem 1).

K   ,
 6 Furthermore, the second adaptive-gain sliding mode

control algorithm requires smaller amount of tuning
then there exists a finite time tF 4 0 so that the sliding parameters than the first algorithm, and is developed
mode is established for all t  tF. in the real sliding mode context.
1910 F. Plestan et al.

3.1 First adaptive sliding mode control law so that


Consider the following controller
Kt  K  8t 4 0:
u K  signx, t 7

with the gain K(t) defined by ( being a positive Proof of Lemma 1 is given in the Appendix.
parameter)
Theorem 3: Given the nonlinear uncertain system (1)
. If jj 6 0, K(t) is the solution of with the sliding variable (x, t) dynamics (2) controlled
  by (7)(9), there exists a finite time tF 4 0 so that a
K_ K 1  x, t 8
sliding mode is established for all t  tF, i.e. (x, t) 0
with K 1 4 0 and K(0) 4 0. for t  tF.
. If  0, K(t) reads as Proof: The proof is composed of two steps. The first
step concerns  6 0 which yields K-dynamics described
Kt K 2  jj K 3 by (8), whereas the second one concerns  0 with
9
Downloaded by [North Carolina State University] at 10:02 25 November 2012

 _  signx, t K-dynamics described by (9).

with K 2 Kt , K 3 4 0 and  4 0. t is the . Suppose that (0) 6 0: in this case, K(t)-
largest time value such that, by denoting t  dynamics read as (8). Consider the following
the time just before t , (x(t ), t ) 6 0 and Lyapunov candidate function, with K a
(x(t ), t ) 0. positive constant
1 1
Discussion: By supposing that j(x(0), 0)j 6 0, the V  2 K  K  2 : 10
adaptive sliding mode control law (7)(9) works as 2 2
follows: One has
. The gain K(t) is increasing due to the 1
adaptation law (8) up to a value large V_         K  sign K  K   K 1  jj

enough to counteract the bounded uncertainty 1
with unknown bounds in (1) until the  M  jj  m  K  jj K  K   K 1  jj

sliding mode starts. Denote the time instant
when the sliding mode starts for the first M  jj  m  K  jj m  K   jj  m  K 
time as t1. 1
 jj K  K   K 1  jj
. As sliding mode has started, i.e. (x(t), t) 0, 
from t t1, K(t) follows the gain-adaptation M  m  K   jj K  K 
law (9). Then, gain K(t) is adapted through (9)  
K
with K 2 Kt1 . Note that this strategy will  m  jj  jj :
allow to decrease the gain and then to adjust it 
with respect to the current uncertainties/ From Lemma 1, there always exists K 4 0
perturbations. such that K(t)  K 5 0 for all t 4 0. It yields
. However, if the varying uncertainty/perturba-
tion exceeds the value K 2 Kt1 , then the V_  m  K   M jj
|{z}
sliding mode will be destroyed and (x(t), t) 
will not be equal to zero. Next, the gain-  
K
adaptation will happen in accordance with (8).  m  jj  jj  jK  K  j: 11

The gain K(t) will be increasing until the |{z}
sliding mode occurs again at the reaching time K
instant t2. There always exists K and  such that
. As sliding mode has occurred and (x(t), t) 0 
K  4 Mm and  5 Km , which yields  4 0
from t t2, K(t) now follows the and K 4 0. Then, one gets
gain-adaptation law (9) with K 2 Kt2 . And
p jj p jK  K  j
so on.1 V_    2 p  K  2 p
2 2
Lemma 1: Given the nonlinear uncertain system (1) n p  
po jj jK  K  j
with the sliding variable (x, t) dynamics (2)  min  2, K 2 p p    V1=2
controlled by (7)(9), the gain K(t) has an 2 2
upper-bound, i.e. there exists a positive constant K 12
International Journal of Control 1911
p p
with 2 minf  , K  g. Therefore, finite- based on Lyapunov analysis when jj 4 , and on
time convergence to a domain  0 is Theorem 1 of Lee and Utkin (2007) when jj  .
guaranteed from any initial condition
j(0)j 4 0, and the reaching time tr can be
easily estimated as
3.2 Second adaptive sliding mode control law
2V01=2 The first adaptive sliding mode control law uses
tr  :
concept of equivalent control which introduces low-
pass filter dynamics with  parameter that is not easy
. Suppose now that (0) 0: Theorem 1 of Lee
to tune (Utkin et al. 1999). The controller displayed
and Utkin (2007) claims that, if K is large
in this section does not estimate the boundary of
enough with respect to uncertainties/perturba-
perturbation and uncertainties. But, there is an
tions effects, then sliding mode control (7)
eminent price to do that: the new strategy guarantees
with the gain-adaptation algorithm (9) allows
a real sliding mode only. Consider the following
keeping trajectories of system (1) on the
controller
Downloaded by [North Carolina State University] at 10:02 25 November 2012

sliding surface  0. Given that trajectories


of system (1) reach  0, it means that K is u K  signx, t 15
large enough as required by Theorem 1 of Lee
with the gain K(t) defined such that
and Utkin (2007). Then, the sliding mode is (
established in system (1) for all t  tr. Theorem K  jx, tj  signjx, tj   if K 4

3 is proven. K_ 16

if K 

Implementation issues. The proposed adaptive-gain


sliding mode control algorithm in (7)(9) is not ready with K(0) 4 0, K 4 0,  4 0 and
4 0 very small. The
for the practical implementations. In fact, it is not parameter
is introduced in order to get only positive
possible to reach the objective  0 due to sampled values for K. In the sequel, for discussion and proof,
computation, noisy measurements or other non- and without loss of generality but for a sake of clarity,
idealities. That is why it is important considering the one supposes that K(t) 4
for all t 4 0.
implementation of the previous controller in a real Discussion: Once sliding mode with respect to (x, t)
sliding mode context. The following modifications of is established, the proposed gain-adaptation law (16)
the gain-adaptation algorithm (7)(9) are proposed2 allows the gain K declining (while j(x, t)j 5 ). In
. If j(x, t)j 4  4 0, K(t) is the solution of other words, the gain K will be kept at the smallest
  level that allows a given accuracy of -stabilisation.
K_ K 1  x, t 13 Of course, as described in the sequel, this adaptation
law allows to get an adequate gain with respect to
with K 1 4 0 and K(0) 4 0. uncertainties/perturbations magnitude.
. If j(x, t)j  , K(t) reads as
Lemma 2: Given the nonlinear uncertain system (1)
Kt K 2  jj K 3 with the sliding variable (x, t) dynamics (2) controlled
14
 _  signx, t by (15) and (16), the gain K(t) has an upper-bound, i.e.
there exists a positive constant K so that
with K 2 Kt , K 3 4 0 and  4 0. t is the
largest time value such that, by denoting t  Kt  K  8t 4 0:
the time just before t , j(x(t ), t )j 4  and
j(x(t ), t )j  .
Proof of Lemma 2 is given in the Appendix.
Corollary 1: Given the nonlinear uncertain system (1)
with the sliding variable (x, t)) dynamics (2) controlled Theorem 4: Given the nonlinear uncertain system (1)
by (7), (13) and (14), there exists a finite time tF 4 0 so with the sliding variable (x, t) dynamics (2) controlled
that a real sliding mode is established for all t  tF, i.e. by (15) and (16), there exists a finite time tF 4 0 so that
j(x, t)j 5  for t  tF. a real sliding mode is established for all t  tF, i.e.
j(x, t)j 5  for t  tF, with
Proof: The proof is composed of two steps. The first s
step concerns jj 4  which yields K-dynamics 2
described by (13), whereas the second one concerns  2 M : 17
 m
K
jj   with K-dynamics described by (14). Then, the
proof follows the same way as the proof of Theorem 3
1912 F. Plestan et al.
p p
Proof: The proof is based on Lyapunovs approach with 2 minf  , K  g.
and shows that, when jj 4 , then control strategy
. Case 1 Suppose that jj 4 . is positive if
ensures that jj 5  in a finite time. Furthermore, it is
proven that as soon as  reaches the domain jj  , it K
stays in the domain jj   defined by (17) for all m  jj  jj  K 4 0

consecutive time. Therefore, the proof shows that the
real sliding mode is established in finite time in the K  
) 5 : 22
domain jj  . m   K
Consider the following Lyapunov candidate From (21), one gets
function
1 1 V_    V1=2     V1=2 : 23
V  2 K  K  2 : 18
2 2 It is always possible to choose  such that the
One has previous inequality fulfils. Therefore, finite-
time convergence to a domain jj   is
Downloaded by [North Carolina State University] at 10:02 25 November 2012

1
V_         K  sign K  K  guaranteed from any initial condition

j(0)j 4 .
 K  jj  signjj   . Case 2 Suppose now that jj 5 . Function
1 in (20) can be negative. It means that V_ would
 M  jj  m  K  jj K  K 
 be sign indefinite, and it is not possible to

 K  jj  signjj   conclude on the closed-loop system stability.
M  jj  m  K  jj m  K   jj Therefore, jj can increase over . As soon as
jj becomes greater than , V_    V1=2 and
1
 m  K   jj K  K   K  jj  signjj   V starts decreasing. Apparently, decrease of V
 can be achieved via increase of K allowing jj
M  m  K   jj K  K  to increase before it starts decreasing down
 
K to jj  . Without loss of generality, let us
 m  jj  jj  signjj   : estimate the overshoot when  0 (0) 

and K0 K(0) K(0) 4 0: considering the
Introduce parameter K 4 0 as worst case (with respect to uncertainties/
V_ M  m  K   jj K  K  perturbations), one has
 
K _ M  K  m
 m  jj  jj  signjj  
 24
 
K_ K  jj:
K  jK  K j  K  jK  K j: 19
Then, it yields
From Lemma 2, there always exists K 4 0 such that
K(t)  K 5 0 for all t 4 0. It yields q  q 
t 0 cos K  m t Mp  K0  m
 sin K  mt ,
V_  M m  K  jj  K  jK  K  j K m
|{z}
s
 40 q   
  K  M
K Kt 0
m
sin Km t K0 
m
 m  jj  jj  signjj    K  jK  K  j:
 q 
|{z}
 cos K  m t M : 25
m
20
Then, one gets
Then, one gets
s  
q
V_    jj  K  jK  K  j  2 M  K0  m 2 
t 0 sin Km t  ,
p jj p jK  K  j K m
   2 p  K  2 p  s
2 2   q 
n p   K M 2
po jj jK  K  j Kt 0 2 K0  
sin Km t K
M
:
 min  2, K 2 p p  m m m
2 2
   V1=2  21 26
International Journal of Control 1913

It appears from (26) that, when  0  ! , the From (28), one has
maximum value  of  M reads as
s jj  jj  K:
2 It yields
 2 M : 27
 m
K
jt1  t1  Kt1  signt1 j  2M  Kt1 :

_
Furthermore, as gain K(t) fulfils (28), -sign is the
In conclusion,  converges to the domain jj   in a opposite of  sign. It means that, if 0  (t1)   (resp.
finite time, but could be sustained in the bigger domain     0), function  will decrease (resp. increase).
jj  . Therefore, the real sliding mode exists in the Then, in order to guarantee that j(t1 Te)j will not
domain jj  . exceed , the worst case is (t1) 0. From (29), it
yields
jt1 Te j  2M  Kt1 Te : 30
4. On -tuning
Downloaded by [North Carolina State University] at 10:02 25 November 2012

Given that it must be ensured j(t1 Te)j  , and as


The objective of this section consists in providing a the best accuracy and the closed-loop system are the
methodology for the tuning of parameter  for both the both objectives,  has to be chosen as a time-varying
algorithms in the case of sampled controllers. In fact, function equal to the smallest admissible value
the choice of parameter  has to be made by an
adequate way because a bad tuning could provide t 2M  Kt1 Te : 31
either instability and control gain increasing to infinity,
or bad accuracy for closed-loop system as described in Discussion: Equation (31) means that the knowledge
the sequel. of bounds of  is required for the choice of  in the case
of sampling controller. However, note that this latter
. If parameter  is too small, and due to large equation gives the best tuning. As said previously,
gain K and sampling period Te, system  can be tuned to a too large value which guarantees
trajectories are such that jj never stays the stability but with a worst accuracy. Furthermore,
lower than . From K-dynamics (13) or (16), Equation (31) gives at least a methodology for
it yields that gain K is increasing, which -computation: no such information is given in
induces larger oscillation, and so on . . . . Huang et al. (2008) whereas such a parameter is
. If parameter  is too large, system trajectories required for the boundary layer in case of practical
are such that, in spite of large gain K and implementation of the controller.
sampling period Te, jj is evolving around , it However, in the case of many practical applica-
follows that controller accuracy is not as good tions, the knowledge of  bounds is not really required.
as possible. In fact, when applied, the controller design procedure
Note that  should rather be too large than too small is very often the following: uncertain functions read
because, in the first case, closed-loop system trajec- as  Nom D and  Nom D where Nom
tories are globally uniformly ultimately bounded even and Nom are the nominal known functions and D
if accuracy is bad. and D their unknown parts. Furthermore, in most of
Parameter  has to be tuned such that, as long as the cases, one has jNomj  jDj and jNomj  jDj.
K(t) is greater than j/j, jj 5 . In this case, given Then, by supposing that Nom is invertible, the
that gain K(t) is sufficient to counteract the perturba- control law3
tions, there is no reason that  increases over . Nom v
Suppose that, at t t1, j(t1)j   and u
Nom
 
t1 
Kt1   : 28 is applied to system (2) which gives
t1     
Nom D
It is clear that control gain K is sufficient to ensure _ D D 1   v: 32
Nom Nom
jt1 Te j  :
As jNomj  jDj, one gets j1  D
Nom
j  2. From (31), it
From Eulers formula, one has yields
t1 Te t1 t1  t1  Kt1  signt1
 Te : t 4KtTe :
29
1914 F. Plestan et al.

5. Simulations Control algorithm 2 in Equations (15) and (16). Control


5.1 Tutorial input u and gain-adaptation law K(t) respectively read
as in Equations (15) and (16). Gain is initialised at
Consider the following uncertain system
K(0) 10, its dynamics being tuned with K 1000 and
_ t u: 33
0.1.
Figures 2 and 3 display simulations results of both
Function (t) is an uncertain bounded function previous controllers applied to system (33). It appears
described by Figure 1. Simulations have been made that both control laws yield very similar results. Based
by supposing that (0) 10 and Te 0.0001 s. From on simulation plots, one can conclude that both
(31), parameter  reads as (t) 2K(t)  Te. adaptation algorithms provide the controller gain
K(t) to follow closely the perturbation (t) whose
Control algorithm 1 in Equations (7), (13) and (14):
profile and boundary are not known a priori (see
Control input u and gain-adaptation law K(t), respec-
bottom-right plots in Figures 2 and 3). Therefore, the
tively, read as in Equations (7), (13) and (14). Gain is
control gain K(t) is not overestimated and control
initialised at K(0) 10, its dynamics being tuned with
chattering is minimal. Also, the sliding variable  is
Downloaded by [North Carolina State University] at 10:02 25 November 2012

K 1 1000, K 3 1 and  0.1 s.


robustly constrainted to jj 5 (t) with exception of
sparse spikes.
70
60
50
5.2 Electropneumatic actuator
40
30 The electropneumatic system under interest is a double
20 acting actuator (Figure 4) composed of two chambers,
10 denoted P (as positive) and N (as negative). The air
0 mass flow rates entering the two chambers are
10 modulated by two three-way servodistributors con-
20 trolled with two electrical inputs of opposite signs
0 50 100 150 200 250 300 (u and u). The pneumatic jack horizontally moves a
Figure 1. Unknown term  versus time (s). load carriage of mass M.

10
0.06
8
0.04
6 0.02
4 0

2 0.02
0.04
0
0.06
2
0 50 100 150 200 250 300 0 50 100 150 200 250 300

400 350
300 300
200 250
100 200
0 150
100 100
200 50
300 0
400 50
0 50 100 150 200 250 300 0 50 100 150 200 250 300

Figure 2. Control algorithm 1 in Equations (7), (13) and (14). Top-left. Sliding variable  versus time (s). Top-right. Zoom on
sliding variable  (), (t) and (t) (  ) versus time (s). Bottom-left. Control input u versus time (s). Bottom-right. Gain K(t)
() and perturbation (t) (  ) versus time (s).
International Journal of Control 1915

5.2.1 Model  
krT SN
Following standard assumptions on the pneumatic part p_N qmP u, pN pN v
VN y rT
of the electropneumatic system (Shearer 1956; Mac Cloy
1 34
1968; Brun, Belgharbi, Sesmat, Thomasset, and v_ SP pP  SN pN  bv  Ff  Fext

Scavarda 1999; Girin, Plestan, Brun, and Glumineau M


2009; Laghrouche, Smaoui, Plestan, and Brun 2006) one y_ v
gets a nonlinear dynamic model for the whole system with y the load carriage position, v its velocity and pP
  and pN the pressures of P and N chambers. VX (X P
krT SP
p_ P qmN u, pp  pP v or N ) is the volume in the chamber X, qmX(uX, pX)
VP y rT

10
0.06
8
0.04
Downloaded by [North Carolina State University] at 10:02 25 November 2012

6 0.02
4 0

2 0.02
0.04
0
0.06
2
0 50 100 150 200 250 300 50 100 150 200 250 300

400 350
300 300
200 250
100 200
0 150
100 100
200 50
300 0
400 50
0 50 100 150 200 250 300 0 50 100 150 200 250 300

Figure 3. Control algorithm 2 in Equations (15), (16). Top-left. Sliding variable  versus time (s). Top-right. Zoom on sliding
variable  (), (t) and (t) (  ) versus time (s). Bottom-left. Control input u versus time (s). Bottom-right. Gain K(t) () and
perturbation (t) (  ) versus time (s).

Figure 4. Electropneumatic system.


1916 F. Plestan et al.

the mass flow rate provided from the servodistributor 0.25


X, k the polytropic constant, r the perfect gas constant, 0.2
0.15
T the supply temperature and SX the piston area in the 0.1
chamber X. The term Ff represents all the dry friction 0.05
forces which act on the moving part in the presence of 0
viscous friction (b  v) and an external force only due to 0.05
atmospheric pressure (Fext). 0.1
The model of mass flow rate delivered by each 0.15
0.2
servodistributor can be reduced to a static function
0.25
described by two relationships qmP(u, pP) and 0 1 2 3 4 5 6
qmN(u, pN). The first two equations of (34) concern
Figure 5. Desired position trajectory yd(t) (m) versus time (s).
the pneumatic part of the system and are obtained
from the state equation of perfect gases, the mass
conservation law and the polytropic law under the
defined in Figure 5. Following the previous section,
assumptions given above. The last two equations
Downloaded by [North Carolina State University] at 10:02 25 November 2012

consider the sliding variable


describe the mechanical part and are derived from
the fundamental mechanical equation applied to the  2  y  yd t 2  y_  y_d t y  y d t
moving part. In order to get an affine nonlinear state
36
model, the mass flow rate static characteristic issued
from measurements (Sesmat and Scavarda 1996) reads with 4 0. As system (34) and (35) admits a relative
as (Belgharbi, Thomasset, Scavarda, and Sesmat 1999) degree equal to 1 for  with respect to u, one gets4
qmP u, pP P pP P pP , sgnu  u, _    u: 37
35
qmN u, pN N pN  N pN , sgnu  u
As described in Laghrouche et al. (2006) and Girin
with functions P and P (resp. N and N) fifth-order et al. (2009) functions () and () can be written as
polynomial functions with respect to pP (resp. pN). the sum of a nominal part (Nom and Nom) and an
Denoting x [ pP pN y v]T, system (34) can be written uncertain one (D and D), i.e.
as nonlinear system (1). Let us define X as the physical
domain  Nom  D,  Nom  D: 38
From Laghrouche et al. (2006) and Girin et al. (2009)
X x j 1 bar  pP  7 bar, 1 bar  pN  7 bar,

functions  and  are bounded for all x 2 X ;
200 mm  y  200 mm, jvj  1 m s1 : furthermore, for all x 2 X , Nom 4 0. Then,  and 
It yields that, for x 2 X , system dynamics are bounded fulfil (3). It yields that the control law reads as5
under a bounded control input u. 1
Two kinds of uncertainties are taken into account: u Nom v 39
Nom
uncertainties due to the identification of physical
parameters, and perturbations. Viscous and dry fric- with control input v reading as previous control
tion coefficients have been identified, but the determi- algorithms. Parameter  reads as (t) 4K(t)  Te.
nation of their variations around their nominal values Parameters have been tuned as
is a hard task. For example, the dry friction coefficient
is difficult to identify depending on the track surface Te 103 s, K 250, K 1 50, K0 1, 33:
quality (thus the piston position), the seal wear, the Furthermore, for Algorithm 1, one has
working conditions (temperature, pressure, quality of
air) etc. The mass flow rate delivered by each K 3 1,  0:1 s:
servodistributor has been approximated by polynomial
functions (35); the uncertainties on () and () have For Algorithm 2, one has
0.1. Following simula-
been evaluated to 15% and 5%, respectively. tions have been made by making variations of load
Finally, during the load moving, the total mass in mass (20%)6 and uncertainties on and (20%).
displacement can evolve from 17 up to 47 kg. Figures 6 and 7 display, for both the control laws, the
actuator position with respect to desired trajectory
(top), the control input u (centre) and the gain K
5.2.2 Control design and simulations (bottom). It appears that both the strategies yield quite
The aim of the control law is to get a good accuracy in similar results. Note that, with the proposed gain
terms of position tracking for the desired trajectory tuning, Algorithm 1 induces lower magnitudes of
International Journal of Control 1917

0.3
0.2
0.1
0
0.1
0.2
0.3
0 1 2 3 4 5 6
10
5
0
5
10
0 1 2 3 4 5 6
400
300
200
100
Downloaded by [North Carolina State University] at 10:02 25 November 2012

0
0 1 2 3 4 5 6

Figure 6. Control algorithm 1: Top. Current () and desired (- - -) position trajectories (m) versus time (s). Centre. Control input
u (V ) versus time (s). Bottom. Control gain K versus time (s).

0.3
0.2
0.1
0
0.1
0.2
0.3
0 1 2 3 4 5 6
10
5
0
5
10
0 1 2 3 4 5 6
1000
800
600
400
200
0
0 1 2 3 4 5 6

Figure 7. Control algorithm 2: Top. Current () and desired (- - -) position trajectories (m) versus time (s). Centre. Control input
u (V ) versus time (s). Bottom. Control gain K versus time (s).

control and gain in this time interval; in fact, not overestimated. The first algorithm is based on
Algorithm 2 gain is increasing to a larger value and evaluation of uncertainties/perturbation by using
takes time to decrease to similar values of Algorithm 1. equivalent control concept that requires employment
However, in the authors experience, implementation of low-pass filter. The second adaptive control law
of Algorithm 2 is clearly easier. does not estimate the boundary of perturbations/
uncertainties and yields establishing a real sliding
mode. The efficacy of these new strategies has been
6. Conclusion confirmed on a tutorial example as well as by
This article proposes two new methodologies for controlling the electropneumatic actuator. Future
adaptive sliding mode controller design. Both the works will concentrate on extending the developed
algorithms allow establishing of the sliding mode via methodologies to both MIMO uncertain nonlinear
the sliding mode control laws with gain-adaptation systems and systems with higher order adaptive-gain
without a priori knowing uncertainties/perturbations sliding mode control, as well as on experimental
bounds while both the adaptive-gain values are validating the advanced algorithms.
1918 F. Plestan et al.

Acknowledgements Actuator: Application to an Aeronautical Benchmark,


This work has been made while Franck Plestan was on IEEE Transactions on Control Systems Technology, 17,
sabbatical at UAH. This stay has been financially supported 633645.
in part by Ecole Centrale de Nantes, IRCCyN, and Centrale Hall, C.E., and Shtessel, Y.B. (2006), Sliding Mode
Initiatives Foundation, Lille, France. Disturbance Observer-based Control for a Reusable
Launch Vehicle, AIAA Journal of Guidance, Control, and
Dynamics, 29, 13151328.
Huang, Y.-J., Kuo, T.-C., and Chang, S.-H. (2008),
Notes Adaptive Sliding-mode Control for Nonlinear Systems
1. It yields that t defined just after (9) equals first-of-all t1, with Uncertain Parameters, IEEE Transactions on System,
then t2. Furthermore, in implementation case, t is Man, and Cybernetics Part B: Cybernetics, 38, 534539.
on-line determined by comparing, at each time t, Laghrouche, S., Smaoui, M., Plestan, F., and Brun, X.
(x(t), t) and (x(t), t) 0.
(2006), Higher Order Sliding Mode Control Based on
2. The parameter  can be time-varying, as in the following
Optimal Approach of an Electropneumatic Actuator,
section. Its tuning is detailed in Section 4.
3. In Castro-Linares, Glumineau, Laghrouche, and Plestan International Journal of Control, 79, 119131.
(2004) it has been shown that, for such class of Laghrouche, S., Plestan, F., and Glumineau, A. (2007),
Downloaded by [North Carolina State University] at 10:02 25 November 2012

uncertainties, this control law allows to decrease the Higher Order Sliding Mode Control Based on Integral
influence of perturbations and uncertainties, and then to Sliding Surface, Automatica, 43, 531537.
reduce the control gain. Lee, H., and Utkin, V.I. (2007), Chattering Suppression
4. In Brun et al. (1999) stability of zero-dynamics has been Methods in Sliding Mode Control Systems, Annual
numerically established. Reviews in Control, 31, 179188.
5. It can be shown that Nom is invertible in the physical Levant, A. (1993), Sliding Order and Sliding Accuracy in
domain of the application (Laghrouche et al. 2006; Girin Sliding Mode Control, International Journal of Control,
et al. 2009).
58, 12471263.
6. This value and the following represent parameters
variations with respect to their nominal values. Levant, A. (2001), Universal SISO Sliding-mode Controllers
with Finite-time Convergence, IEEE Transactions on
Automatic Control, 49, 14471451.
Levant, A. (2007), Principles of 2-sliding Mode Design,
References Automatica, 43, 576586.
Mac Cloy, D. (1968), Discharge Characteristics of Servo
Bartolini, G., Ferrara, A., Usai, E., and Utkin, V.I. (2000), Valve Orifices, in Fluid Power International Conference,
On Multi-input Chattering-free Second-order Sliding Olympia, Greece, pp. 4350.
Mode Control, IEEE Transactions on Automatic Control, Munoz, D., and Sbarbaro, D. (2000), An Adaptive Sliding
45, 17111717. Mode Controller for Discrete Nonlinear Systems, IEEE
Belgharbi, M., Thomasset, D., Scavarda, S., and Sesmat, S. Transactions on Industrial Electronics, 47, 574581.
(1999), Analytical Model of the Flow Stage of a Pneumatic Plestan, F., Glumineau, A., and Laghrouche, S. (2008),
Servo-distributor for Simulation and Nonlinear Control, A New Algorithm for High-order Sliding Mode Control,
in Scandinavian International Conference on Fluid Power International Journal of Robust and Nonlinear Control, 18,
SICFP99, Tampere, Finland, pp. 847860. 441453.
Boiko, I., and Fridman, L. (2005), Analysis of Chattering in Sesmat, S., and Scavarda, S. (1996), Static Characteristics of
Continuous Sliding-mode Controllers, IEEE Transaction a Three Way Servovalve, in Conference on Fluid Power
on Automatic Control, 50, 14421446. Technology, Aachen, Germany, pp. 643652.
Boiko, I., Fridman, L., Pisano, A., and Usai, E. (2007), Shearer, J.L. (1956), Study of Pneumatic Processes in the
Analysis of Chattering in Systems with Second Order Continuous Control of Motion with Compressed Air,
Sliding Modes, IEEE Transactions on Automatic Control, Transactions of the American Society Mechanical
52, 20852102. Engineers, 78, 233249.
Brun, X., Belgharbi, M., Sesmat, S., Thomasset, D., and Slotine, J.-J., and Li, W. (1991), Applied Nonlinear Control,
Scavarda, S. (1999), Control of an Electropneumatic Englewood Cliffs, NJ: Prentice Hall.
Actuator, Comparison between some Linear and Slotine, J.-J., and Sastry, S.S. (1983), Tracking Control of
Nonlinear Control Laws, Journal of Systems and Control Nonlinear System using Sliding Surfaces, with Application
Engineering, 213, 387406. to Robot Manipulators, International Journal of Control,
Castro-Linares, R., Glumineau, A., Laghrouche, S., and 38, 465492.
Plestan, F. (2004), High Order Sliding Mode Observer- Tao, C.W., Chan, M.L., and Lee, T.T. (2003), Adaptive
based Control, 2nd IFAC Symposium on System, Structure Fuzzy Sliding Mode Controller for Linear Systems with
and Control, Oaxaca, Mexico, pp. 517522. Mismatched Time-varying Uncertainties, IEEE
Filippov, A.F. (1988), Differential Equations with Transactions on System, Man, and Cybernetics Part B:
Discontinuous Right-hand Side, Dordrecht, The Cybernetics, 33, 283294.
Netherlands: Kluwer Academic Publishers. Utkin, V.I., Guldner, J., and Shi, J. (1999), Sliding Mode in
Girin, A., Plestan, F., Brun, X., and Glumineau, A. Control in Electromechanical Systems, London: Taylor &
(2009), Robust Control of an Electropneumatic Francis.
International Journal of Control 1919

Appendix s(t )

Proof of Lemma 1
Suppose that (x, t) 6 0. From K-dynamics (8), and given
e
that functions  and  are supposed bounded, it follows that t
K is increasing and there exists a time t1 such that
M |y(t )G (t )|
Kt1 4 :
m
Note that this gain has a finite value by the absolute
continuity property of K(t). From t t1, given K-dynamics, K (t ) t
gain K is large enough to make the sliding variable  t1 t2 t3 t4
decreasing. Then, it yields that, in a finite time t2,  0 and
K(t2) admits a bounded value. It yields that there always Figure A1. Scheme describing the behaviour of  (top) and
exists a positive constant K such that K 4 K(t), for K (bottom) versus time.
all t  0.
Downloaded by [North Carolina State University] at 10:02 25 November 2012

From t t3, gain K is not large enough to counteract


Proof of Lemma 2 perturbations and uncertainties as it is decreasing. It yields
Suppose that j(x, t)j 4 . From K-dynamics, and given that that there exists a time instant t4 4 t3 such that j(t4)j 4 .
functions  and  are supposed bounded, it follows that K is The process then restarts from the beginning. By the
increasing and there exists a time t1 (see Figure A1) such that assumptions (3), the gains K(ti) remain bounded uniformly
  on ti. In fact,
t1 
Kt1  :  
t1  ti  M
Kti   : K
From t t1, given K-dynamics, gain K is large enough to ti  m
make the sliding variable  decreasing. Then, it yields that, in
a finite time t2 (Figure A1), jj 5 . It yields that gain K is and, hence, there always exists a finite constant K
decreasing from t2, gain K being at a maximum value at such that
t t2. From K-dynamics, it yields that there exists a time
instant t3 4 t2 (Figure A1) such that K   K  ,
 
t3  which proves the desired result.
Kt3  :
t3