Original article

Proc IMechE Part I:

J Systems and Control Engineering

An L1 fuzzy adaptive controller for a 2016, Vol. 230(8) 736–748

Ó IMechE 2016
Reprints and permissions:
class of SISO nonaffine nonlinear sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/0959651816652023

systems: application to the control of pii.sagepub.com

an electropneumatic actuator

Ahsene Boubakir1, Toufik Souanef2, Salim Labiod1, Franck Plestan3 and

Fares Boudjema4

Abstract
This paper proposes an L1 fuzzy adaptive controller for a class of uncertain continuous-time single-input single-output
nonaffine nonlinear systems. The structure of this controller is derived based on L1 adaptive control design methodol-
ogy and integrates a fuzzy system. The latter is used to approximate as best as possible a function of an unknown ideal
implicit controller, which provides good results and improves the performance significantly. The L1 fuzzy adaptive con-
troller consists of a predictor, a control law and its adaptive laws. The major advantage of the proposed control scheme
is its ability to guarantee uniformly bounded transient and tracking performance for the controlled system. These perfor-
mance bounds can be rendered arbitrarily small by the systematic choice of design parameters. The effectiveness and
feasibility of the proposed L1 fuzzy adaptive controller are examined experimentally in the position control of a pneu-
matic actuator system.

Keywords
L1 adaptive control, nonaffine nonlinear systems, fuzzy system, transient and tracking performance, pneumatic actuator
system.

Date received: 26 November 2015; accepted: 11 April 2016

Introduction uncertainty and with time-varying unknown para-

meters and bounded disturbances. They also considered
The L1 adaptive control method1 has drawn much an output feedback formulation for a single-input sin-
attention in recent years because of its robust character gle-output (SISO) linear system with time-varying dis-
and ability to provide high control performance, both turbances.7 Cao and Hovakimyan8 discussed the
in transient and steady states, in spite of the presence of stability margin of the L1 architecture. An interesting
uncertainties, unmodeled dynamics and disturbances. comparative study between the L1 adaptive controller
In fact, L1 adaptive control is an improvement of and the direct and indirect MRAC, using frequency
Model Reference Adaptive Control (MRAC)2 with domain tools, is given by Kharisov et al.9 Some of Cao
state predictor architecture by introducing a low-pass and Hovakimyan’s research papers3,5,7,8 culminated in
filtering mechanism in the control law, which decouples a book,1 published in 2010, which details the principle
adaptation speed from robustness. Hence, unlike the
MRAC control scheme, L1 adaptive control allows fast
1
adaptation without sacrificing robustness, and provides LAJ, Faculty of Science and Technology, University of Jijel, Algeria
2
Institut of Flight Mechanics and Flight Control, University of Stuttgart,
a uniformly bounded transient response, in addition to
Germany
stable tracking for a system’s input and output signals 3
LUNAM Université, Ecole Centrale de Nantes, IRCCyN, France
simultaneously. Its principle was first introduced in 4
LCP, Department of Automatic Control, National Polytechnic School,
20063,4 by Cao and Hovakimyan. Since then, develop- Algeria
ment of controllers based on L1 adaptive control meth-
Corresponding author:
odology has been the topic of numerous publications. Ahsene Boubakir, LAJ, Faculty of Science and Technology, University of
Cao and Hovakimyan5,6 designed two state feedback Jijel, BP. 98, Ouled Aissa, 18000 Jijel, Algeria.
L1 adaptive controllers for linear systems, with constant Email: [email protected]
Boubakir et al. 737

of L1 adaptive control and its application to classes of class of SISO nonaffine nonlinear systems. Firstly, the
linear and nonlinear systems with performance studies. existence of unknown ideal implicit control law which
The choice of structure and the coefficients of the feed- can achieve control objectives is demonstrated. Then, a
back filter plays an important role in the efficiency of a fuzzy approximation of an unknown nonlinear func-
designed L1 adaptive controller; this is widely discussed tion of the ideal implicit controller is considered in a
in the literature.1,10 Cooper et al.11 discussed the advan- controller design based on an L1 adaptive control
tages of using neural networks as an approximator of method. Even though conventional L1 control, using
nonlinearity in L1 adaptive control design for a SISO fast projection-type adaptive laws, can handle
affine nonlinear system. Motivated by the performance unknown nonlinear functions,22–24 it is of both theore-
results of the L1 adaptive controller both in transient tical and practical interest to resort to learning tech-
and steady state, various applications of this kind of niques such as fuzzy systems. The fact of incorporating
controller have been devised. The performance and a fuzzy approximator in the L1 control architecture, in
robustness characteristics of an L1 adaptive flight con- order to estimate unknown nonlinearity, can relax
trol system have been studied and evaluated experimen- some design parameters that depend on the choice of
tally by Dobrokhodov et al.12 based on Rohr’s adaptation rate while still achieving high performance.
counterexample. This latter is a benchmark problem, Generally, in L1 adaptive control design, the state pre-
presented in the early 1980s, to show the limitations of dictor used is a function of the uncertainty approxima-
the adaptive controllers developed at that time. Michini tion. Since a fuzzy system can be designed to be a
and How13 developed and implemented an L1 adaptive universal approximator for continuous functions,26,27
output feedback controller for an indoor autonomous an approximation based on a fuzzy system is more effi-
quadrotor helicopter with a systematic design of the cient, which leads to a small state prediction and thus
control parameters. Jin et al.14 designed and implemen- better control performance. Contrary to L1 adaptive
ted an L1 augmented dynamic inversion controller for control approachs using only fast projection-type adap-
hover attitude control of a tail-sitter vertical take-off tative laws and requiring a small integration step size
and landing micro aerial vehicle. These results12,13,14 to handle uncertainties, which imposes stringent con-
confirm the effectiveness and practicality of the L1 straints on hardware CPU speed for its implementa-
adaptive control scheme in flight control systems. tion,11 the presented L1 fuzzy adaptive controller can
Bennehar et al.15 proposed a L1 adaptive controller for achieve the same performance with a lower adaptation
a parallel kinematic manipulator, with experimental rate or enlarged integration step size. The presented
validation. Malinga and Buckner studied real-time control scheme guarantees the stability of the closed-
implementation of the L1 adaptive control scheme to loop system and ensures arbitrarily close tracking of
regulate the bending angle of a flexible beam, actuated the input and output signals of a stable virtual refer-
by a single shape memory alloy tendon.16 The L1 adap- ence system. The ability of the proposed controller is
tive control method has also been used in other sys- examined experimentally in the position control of a
tems: for example, spacecraft17 and underwater pneumatic actuator system.
vehicles,18,19 to name a couple. The main advantages of the proposed control
For nonlinear systems, most of the existing L1 adap- scheme and the contributions of this research work can
tive architectures are developed for affine nonlinear be stated as follows.
systems20,21: few research works are available for non-
affine nonlinear systems. The idea of using an L1 adap-  The L1 fuzzy adaptive controller is developed,
tive control method for nonaffine nonlinear systems under certain acceptable assumptions, for a general
was first introduced by Luo et al.,22 under restrictive class of SISO nonaffine nonlinear systems with
assumptions on nonlinearity. An L1 adaptive controller internal dynamics, and is able to guarantee uni-
was designed by Zou et al.23 for a class of uncertain formly bounded transient and tracking perfor-
systems in the presence of time-varying unknown non- mance for the controlled system.
linear hysteresis in its input. Recently, an improvement  The presented controller does not rely on a physical
of the controllers devised by Luo et al.22 and Zou et model of the controlled system and requires only
al.23 was suggested by Choe et al.,24 who relaxed the the knowledge of the relative degree.
restrictive assumptions made on the plant model,  The control law, its adaptation law and the
including unmodeled dynamics in the control design, designed predictor are relatively easy to implement.
and considered a more general filtering structure. In a The fuzzy approximator used significantly
later work, Luo et al.25 developed an L1 adaptive con- improves estimation of the unknown nonlinearity
trol law for a class of unknown non-affine multi-input and state prediction, which leads to good control
multi-output (MIMO) time-varying nonlinear systems performance.
having internal dynamics with unmeasurable states.
In this paper, an L1 adaptive controller that uses a Throughout this paper, kkL1 and kkL‘ denote,
fuzzy approximator as a supplementary means to respectively, the L1 and LN norms of a function. For a
achieve a good estimation of uncertainty and improves signal jðtÞ : ½0, ‘Þ ! <n , the truncated LN norm at the
D
the resulting performance, is introduced for a general time instant t is defined as kjt kLN ¼ sup0 ł t ł t kjðtÞk‘
738 Proc IMechE Part I: J Systems and Control Engineering 230(8)

D  r1
and the LN-norm as kjkLN ¼ supt ø 0 kjðtÞk‘ . Given a d
sðtÞ = +l eð t Þ l . 0 ð3Þ
stable proper SISO R ‘system FðsÞ, its L1-norm is defined dt
to be kFðsÞkL1 = 0 jFðtÞjdt, where FðtÞ is the impulse
response of FðsÞ. Further, for a strictly proper stable From (3), it is easy to verify that sðtÞ = 0 is a linear
system yðsÞ = FðsÞuðsÞ, the inequality kyt kL‘ ł differential equation whose solution implies that the
kFðsÞkL1 kut kL‘ holds. tracking error eðtÞ of the closed-loop system converges
to zero with a time constant ðr  1Þ=l.28 In addition,
the derivatives of the tracking error eðtÞ up to the r  1
L1 fuzzy adaptive controller design order also converge to zero. Furthermore, when the fil-
Problem formulation tered tracking error s is bounded – for example,
jsðtÞj ł F whereF is a positive constant – we can con-
This paper considers an nth order single-input clude that28 eðiÞ ðtÞ ł 2i lðir + 1Þ F, i = 0, . . . , r  1.
nonaffine-in-control system of relative degree r, repre- Hence, the control objective becomes the design of a
sented in normal form control law to stabilize sðtÞ.
The time derivative of the filtered error (3) can be
x_ 1 = x2 , x_ 2 = x3 , . . . , x_ r1 = xr
evaluated as
x_ r = fðx, z, uÞ
ð1Þ ðrÞ
z_ = Qðx, z, uÞ s_ = yd ðtÞ + br1 eðr1Þ +    + b1 e_  fðx, z, uÞ ð4Þ
y = x1
where bi = ðriðr1 Þ! ri
Þ!ði1Þ! l , i = 1, . . . , r  1: Let d be a
 T
where xT , zT 2 <n is the overall state vector signal defined as
of the system, which is assumed to be available
D ðrÞ
for
 measurement,
 such that x = ½x1 ,    , xr T = dðt, sÞ ¼ yd ðtÞ + br1 eðr1Þ +    + b1 e_ + as ð5Þ
ðr1Þ T
y,    , y 2 < and z 2 <nr , u 2 < is the scalar
r

control input, y 2 < is the scalar control output, and with a . 0. From (2) and (4), and by adding and sub-
f : <r 3 <nr 3 < ! < and Q : <r 3 <nr 3 < ! <nr tracting as to the right-hand side of (4), one gets the
are smooth and unknown nonlinear functions. following dynamics
The input–output form of the nonlinear system (1) s_ = as  ðfðx, z, uÞ  dÞ
can be rewritten as follows ð6Þ
z_ = Qðx, z, uÞ
yðrÞ = fðx, z, uÞ
ð2Þ Now, considering Assumption 1 (and using the fact
z_ = Qðx, z, uÞ that the signal d, introduced in (5), does not explicitly
depend on the input u), it follows that the partial deri-
The control objective is to design a control law uðtÞ
vative of fðx, z, uÞ  d = 0 with respect to u satisfies
for system (1) such that the system output yðtÞ tracks a
bounded reference signal yr ðtÞ, both in transient and ∂fðx, z, uÞ  d ∂fðx, z, uÞ
steady state, while all other signals remain bounded. = .0 ð7Þ
∂u ∂u
Before proceeding to L1 fuzzy adaptive controller
design, throughout this paper we consider the following Based on the use of the implicit function theorem, we
assumptions regarding system (1) and the reference tra- conclude that the nonlinear algebraic equation
jectory yr ðtÞ. fðx, z, uÞ  d = 0 is locally solvable for the input u for
each ðx, z, dÞ. Thus, there
 exists an ideal
 control input
Assumption 1. The function fu ðx, z, uÞ = ∂fðx,∂uz, uÞ is uH ðx, z, dÞ fulfilling f x, z, uH ðx, z, dÞ  d = 0 for all
strictly positive and bounded for all ðx, z, dÞ 2 Dx 3 Dz 3 <. It results that when we chose
ðx, z, uÞ 2 Dx 3 Dz 3 < with Dx  <r and Dz  <nr the control input u as the ideal control law uH ðx, z, dÞ,
being domains, i.e. 0 \ x 0 \ fu ðx, z, uÞ \ x 1 where x0 the closed-loop dynamics (6) becomes
and x 1 are some positive constants. Note that fu ðx, z, uÞ
can be assumed to be strictly negative and bounded, s_ = as
ð8Þ
i.e. x1 \ fu ðx, z, uÞ \  x 0 \ 0, and the proposed z_ = Qðx; z, uÞ
controller can be similarly derived. h
which implies that sðtÞ ! 0 as t ! ‘ and, therefore,
Assumption 2. The zdynamics in (1) is bounded- the tracking error eðtÞ and all its derivatives up to r  1
input-bounded-output stable. h converge to zero. Applying now the mean value theo-
Assumption 3. The reference trajectory yr ðtÞ and its rem to fðx, z, uÞ,29,30,31 this latter can be expressed
ðiÞ
time derivatives yr ðtÞ, i = 1, . . . , r, are smooth and around the ideal control law uH as follows
bounded. h ∂fðx, z, ua Þ  
In the first step, let us define the tracking error as fðx, z, uÞ = fðx, z, uÞ + u  uH ð9Þ
∂ua
eðtÞ = yr ðtÞ  y(t), and the filtered tracking error s as
the following where ua = au + ð1  aÞuH and 0 \ a \ 1. Substituting
(9) into (6) yields
Boubakir et al. 739

    
s_ = as  f x, z, uH + fua u  uH  d Assumption 5. Let uH ðtÞ be a continuous function with
ð10Þ a uniformly bounded derivative
z_ = Qðx, z, uÞ
  H
Using the fact that f x, z, uH  d = 0, the dynamics u_ ðtÞ ł duH \ ‘, 8t ø 0 ð17Þ
(10) can be written as the following
where duH can be arbitrarily large. h
∂fðx, z, ua Þ ∂fðx, z, ua Þ H In this paper, we use a zero-order Takagi–Sugeno
s_ = as  u+ u
∂ua ∂ua ð11Þ fuzzy system27 that establishes a mapping from an input
z_ = Qðx, z, uÞ vector p = ½p1 , . . . , pm T 2 Dp  <m to a scalar output
variable yf 2 <, such that Dp = Dp1 3    3 Dpm and
For the unknown nonlinear
 term ∂fðx,∂uz,a ua Þ u in (11) Dp1  <. For each input pi , let us define Mi fuzzy sets
we can always find O and D x, z, u, uH such that Fij , j = 1, . . . , Mi , then the fuzzy system consists of a
sequence of N rules of the form if–then given as
∂fðx, z, ua Þ  
u = Ou  D x, z, u, uH ð12Þ
∂ua Rk : If p1 is Gk1 and . . . and pm isGkm , then
ð18Þ
where 0 \ -0 \ O \ -1 is an unknown constant gain, yf is ykf ðk = 1, . . . , NÞ
-0 and -1 are some positive constants, and D represents
where Gki 2 F1i , . . . , FM i
, i = 1, . . . , n, ykf denotes the
the mismatch part between ∂u∂fa u and Ou. i
crisp output of the kth rule. The use of the singleton
Substituting
 (12) in  (11) and denoting
fuzzifier and the product inference engine, gives (19) as
F x, z, u, uH = D x, z, u, uH + ∂fðx,∂uz,a ua Þ uH , the first
the final output of the fuzzy system
equation in (11) reads as
PN
  k = 1 mk ðp Þyf
k
s_ =  as  Ou + F x, z, u, uH ð13Þ yf ðpÞ = PN ð19Þ
k = 1 mk ðp Þ
Before proceeding further,  let us consider an n o
assumption about the term F x, z, u, uH . Q
m
where mk ðpÞ = mGki ðpi Þ with Gki 2 mF1i , . . . , mFMi
 i=1 i
Assumption 4. The nonlinear function F x, z, u, uH and mF j ðpi Þ is the membership function of the fuzzy set
ðx, z, dðt, sÞÞÞ in (13) is continuous in its arguments, i

Lipschitz in s and bounded at s = 0, such that Fij . The output given by (19) can also be expressed in a
  compact form, such as
    
F x, z, u, uH s2  F x, z, u, uH s1  ł Ljs2  s1 j ð14Þ
yf ðpÞ = wT ðpÞu ð20Þ
   
 
F x, z, u, uH s = 0  ł L0 ð15Þ where u is a vector containing all consequent para-
meters and wðpÞ = ½w1 ðpÞ, . . . , wN ðpÞT is a set of
where L and L0 are positive constants.   h fuzzy basis functions given as w k ðp Þ =
Note that the nonlinear function F x, z, u, uH in mk ðpÞ
PN , k = 1, . . . , N. In addition, the fuzzy system
(13) is unknown. As we will see shortly, a fuzzy approx- k=1
mk ðpÞ
imation of this unknown function is considered in the (20) is assumed to be well-defined by the designer so
controller design. P
that N k = 1 mk ðpÞ 6¼ 0 for all p 2 Op . It is important to
remember that a fuzzy system in the form of (20) can
Fuzzy system approximator design approximate a continuous nonlinear function over a
domain of interest to any desired accuracy.26
In order to derive the L1 fuzzy adaptive
 controller, we Now, substituting (16) into (13), we get
assume that the term F x, z, u, uH can be approxi-
mated over a compact set Dx 3 Dz 3 < by a fuzzy sys- s_ =  as  Ou + wT ðpÞuH + eðpÞ ð21Þ
tem up to a desired accuracy as
Next, we introduce the elements of the proposed L1
  fuzzy adaptive controller with its performance analysis.
F x, z, u, uH ðx, z, dðt, sÞÞ = vT ðpÞuH + eðpÞ,
jeðpÞj ł eH ð16Þ
L1 fuzzy adaptive control structure
The architecture of the fuzzy system used in this work
Based on equation (13), we consider the following
will be introduced shortly. In (16) p = ½x, z, u, dT
predictor
denotes the input vector of the fuzzy system, vðpÞ is a
fuzzy basis function vector fixed by the designer, eðpÞ ^_ =  a^
s ^ + wT ðpÞ^u
s  Ou ð22Þ
is the fuzzy approximation error, and uH is an optimal
unknown vector that minimizes the approximation where s ^ and ^u are the
^ 2 < is the prediction of s, O
error over an operating compact set, uH 2 Y, and eH is adaptive estimates that are governed by the following
a uniform bound for e. adaptation laws
740 Proc IMechE Part I: J Systems and Control Engineering 230(8)

^_ = GProj O,
O ^ s~u ð23Þ Analysis of L1 fuzzy adaptive controller
Closed-loop reference system
 
u^_ = GProj ^u,  s
~ wðpÞ ð24Þ We now consider the ideal non-adaptive version of the
D controller introduced in (22), (23), (24) and (25) and
where s ~¼s ^  s is the prediction error, G . 0 is the define the L1 reference system as
adaptation gain and Projð, Þ denotes the projection
p p
operator defined, for two vectors O 2 < and q 2 < , s_ ref = asref  Ouref + vT ðpref ÞuH + eðpref Þ
1
as follows CðsÞ ð29Þ
uref = h ðsÞ
O ref
ProjðO , qÞ =
( where href ðsÞ is the Laplace transform of
hT ðO Þ
q  rhkðr
O Þr
qhðO Þ ifhðO Þ . 0 and rhT ðO Þq . 0 D
hðOÞk2 href ðtÞ ¼ vT ðpref ÞuH .
q otherwise
Lemma 1. For the closed-loop reference system in (29),
ðe + 1ÞO T
O O 2 if the filter CðsÞ is designed such that the L1 norm con-
where hðO Þ = O e O2 with Omax 2 <+ being
max
dition in (28) is fulfilled, then
O max
the norm bound imposed on the vector O and eO 2 <+
represents the projection tolerance bound of our ksref kL1 ł rr ð30Þ
choice. By a suitable choice of Omax and eO , the projec-
kGðsÞkL ðe H
+ L0 Þ + kHðsÞkL e + ksin ðsÞkL‘
H
tion operator Projð, Þ ensures that the adaptive esti- where rr = 1
1kGðsÞkL L
1
:
mates ^u and O ^ remain inside the compact sets Y and 1

D
C ¼ ½-l , -u , respectively, such that 0 \ -l \ -0 \ Proof. From the definition of the closed-loop reference
-1 \ -u . system given in (29), one gets
The control law is chosen to be
sref ðsÞ = GðsÞhref ðsÞ + HðsÞeref ðsÞ + sin ðsÞ ð31Þ
u = KDðsÞhðsÞ ð25Þ
where sin ðsÞ = HðsÞs0 and eref ðsÞ is the Laplace trans-
where hðsÞ is the Laplace transform of the signal form of eðpref Þ. Since HðsÞ, CðsÞ and GðsÞ are proper
D
hðtÞ ¼ Ou^ + vT ðpÞ^u, K is a positive feedback gain, BIBO-stable transfer functions, and sin ðtÞ is uniformly
and DðsÞ is a strictly-proper transfer function which bounded, it follows from (31) and (16) that for all
leads to the stable closed-loop filter t 2 ½0, t the following upper bound holds

OKDðsÞ ksref t kL‘ ł kGðsÞkL1 khref t kL‘

CðsÞ = ð26Þ ð32Þ
1 + OKDðsÞ + kHðsÞkL1 eH + ksin t kL‘

with DC gain Cð0Þ = 1. One simple choice is DðsÞ = 1s From the definition of href ðtÞ, and based on (16), (14)
which yields a first order strictly proper CðsÞ in the fol- and (15), we have
lowing form  
khref t kL‘ = F x, z, u, uH  eðpÞ L‘
   
OK  
CðsÞ = ð27Þ ł eH + F x; y, u, uH s = 0  + Lksref t kL‘ ð33Þ
s + OK
ł eH + Lksref t kL‘ + L0
The proposed L1 fuzzy adaptive controller consists
of the predictor (22), the adaptive laws (23) and (24), Substituting (33) into (32) yields
and the controller (25), subject to the following L1-
norm condition ksref t kL‘ ł kGðsÞkL1 eH + Lksref t kL‘ + L0
ð34Þ
kGðsÞkL1 L \ 1 ð28Þ + kHðsÞkL1 eH + ksin kL‘

where GðsÞ = HðsÞð1  CðsÞÞ, HðsÞ = ðs + aÞ1 and L Solving (34) for ksref t kL‘ , one obtains
is given in (14).
ksref t kL‘
Remark 1. The positive design parameters l and a  
kGðsÞkL1 eH + L0 + kHðsÞkL1 eH + ksin kL‘
introduced in (3) and (5), respectively, are used to spe- ł
1  kGðsÞkL1 L
cify the desired closed-loop dynamics. The bandwidth-
limited filter C(s) in (26), defined by the feedback gain ð35Þ
K and the strictly proper transfer function DðsÞ, is
If the L1-norm condition in (28) is verified, then
designed to achieve good tracking of reference signals
ksref t kL‘ is uniformly bounded for all t . 0. Therefore
and to compensate the undesirable effects of the uncer-
the relationship in (30) holds and this completes the
tainties within a prespecified range of frequencies. h
proof.
Boubakir et al. 741

Transient and steady-state performance 2 2 ~T _ H 

V_ ł  L~
s2 + eH + u u  ð43Þ
G
We now present and study the transient and steady-
state performance bound between the L1 fuzzy adaptive The projection operator guarantees that ^uðtÞ 2 Y and
controller, defined via (22), (23), (24) and (25) and the ^ ðtÞ 2 C for t ø 0 and therefore
O
D
closed-loop reference system (29). Letting s ~¼s ^  s be  
1 ~T ~ 1 ~2
the prediction error, in order to define its dynamics, we max u u+ O
tø0 G G
subtract (21) from (22), which results in   ð44Þ
1 H 2
~ + v T ðp Þu
~  eðp Þ ł 4 max u + ð-u  -l Þ
~_ =  a~
s s  Ou ð36Þ G uH 2Y

~ ¼O D
^  O and u ~ = ^u  uH are the estimation
where O  T considering
Moreover,  the upper bounds in (17), the
~ _ H 
errors. Furthermore, the error dynamics in (36) can also term G2 u u  is bounded as
be represented in the frequency domain as  
2 ~T _ H 4 H
~ ðsÞ = HðsÞðh
s ~ ðsÞ  eðsÞÞ, ð37Þ u u ł max u duH ð45Þ
G G uH 2Y
where h ~ ðsÞ is the Laplace transform of After substituting (44) and (45) into (43), we get
D ~ + vT ðpÞu.
~
~ ðtÞ ¼ Ou
h
V_ ðtÞ ł  LVðtÞ + Lc0 ð46Þ
Lemma 2. Considering the system (1) and the L1 fuzzy
adaptive controller defined via (22), (23), (24) and (25), Hence, if at any time t1 . 0, one has Vðt1 Þ . c0 ,
subject to (28), we have the inequality (46) implies V_ ðt1 Þ \ 0. In addition,
pffiffiffiffiffi since s^ ð
0Þ = sð0Þ, we can easily  verify that
ks
~ k L ‘ ł c0 ð38Þ
Vð0Þ \ 1
G 4 max uH + ð-u  -l Þ2 \ c0 . Thus,
where uH 2Y

  VðtÞ \ c0 for all t . 0. Now, using s ~ 2 ł VðtÞ and the

pffiffiffiffiffi
1 fact that kk‘ ł kk, we obtain ks ~ kL‘ ł c0 , which
c0 = 4 max uH + ð-u  -l Þ2
G uH 2Y concludes the proof. j
1   ð39Þ We notice that the bound on the prediction error
4L 2
+ max uH duH + L1 eH (39) can be arbitrarily improved by increasing the adap-
G uH 2Y
tation gain G and the design parameter a. Now we can
D
and L ¼ ð2a  1Þ with a . 12. formulate our main result in the following theorem.
Proof. Let us consider the Lyapunov-like function Theorem 1. Given the system in (1) and (21), the refer-
ence system in (29) and the L1 fuzzy adaptive controller
1 ~T ~ 1 ~2 defined via (22), (23), (24) and (25), subject to the L1-
~2 +
V=s u u+ O ð40Þ
G G norm condition in (28), we have
from (36), (23) and (24), the time derivative V_ of (40) ksref  skL‘ ł g 1
verifies ð47Þ
kuref  ukL‘ ł g2
 
V_ = 2~
s a~ ~ + vT ðpÞu
s  Ou ~  eðpÞ + 2 u ~_ + 2 O
~T u ~O~_ where
   G G 
~ ~ 1 ^_ ~T s 1 ^_ 3kGðsÞkL1 + kHðsÞkL1
= 2a~ s 2 + 2O su + O + 2u ~ vðpÞ+ ½u kCðsÞkL1 pffiffiffiffiffi
G G g1 = c0 + eH
1  kGðsÞkL1 1  kGðsÞkL1
2 ~T _ H
2~seðpÞ  u u ð48Þ
G
ł 2a~ 2
s + 2O ~ Proj O, ^ s
~u  s ~u CðsÞ  H
 H 1 ðsÞ pffiffiffiffiffi
g2 = Lg1 + 3e + c0 ð49Þ
O L1 O L1
+ 2u ~T Proj u,
^ s ~ wðpÞ + s
~ wðpÞ
CðsÞ
 
2 ~T _ H  such that H1 ðsÞ = H ðsÞ is propre and stable transfer
~ jeH + u
+ 2js u  function.
G D
Letting hðtÞ ¼ vT ðpÞuH ðtÞ, the control law in (25)
ð41Þ
can be written as
Based on the properties of the projection operator, one
u = KDðsÞðOu + hðsÞ + h
~ ðsÞÞ
obtains
KDðsÞ
2 ~T _ H  or equivalently u= ðhðsÞ + h~ ðsÞÞ:
V_ ł  2a~
s2 + 2js
~ jeH + u u  ð42Þ 1 + OKDðsÞ
G
Moreover, it follows from the definition of CðsÞ in (26)
or that
742 Proc IMechE Part I: J Systems and Control Engineering 230(8)

CðsÞ we note that since CðsÞ is strictly proper, the system

u= ð hð s Þ + h
~ ðsÞÞ ð50Þ
O H1 ðsÞ is proper and stable. Therefore, the expression in
(57) can be upper bounded, for t 2 ½0, t, as follows
Substitution of the control law (50) into (21) yields

s = GðsÞhðsÞ  HðsÞCðsÞ~
hðsÞ + HðsÞeðsÞ + sin ðsÞ ð51Þ ðuref ðsÞ  uðsÞÞ t L‘
CðsÞ
which can also be written as ł L ðs  sref Þ t L‘
+ 3eH
O L1 ð59Þ
s = GðsÞhðsÞ  HðsÞCðsÞðh~ ðsÞ  eðsÞÞ H 1 ðsÞ pffiffiffiffiffi
ð52Þ + c0 :
+ GðsÞeðsÞ + sin ðsÞ O L1

On the other hand, note that sref ðsÞ is defined in (31). Finally, one obtains kðuref ðsÞ uðsÞÞ t kL‘ ł COðsÞ
Subtracting sðsÞ from sref ðsÞ gives L1
 H
 H1 ðsÞ pffiffiffiffiffi
Lg 1 + 3e + O c0 , which holds uniformly
L1
sref  s = GðsÞhe ðsÞ + CðsÞ~
sðsÞ
ð53Þ for all t . 0, leading to the second bound in (47).
 GðsÞeðsÞ + HðsÞeref ðsÞ
D   Corollary 1. Considering the system in (1) and (21), the
where he ðtÞ ¼ vT ðpref Þ  vT ðpÞ uH ðtÞ. Now, from reference system in (29) and the L1 fuzzy adaptive con-
(53), we can derive the following upper bound for troller defined via (22), (23), (24) and (25) subject to the
t 2 ½0, t L1-norm condition in (28), we have
ðsref  sÞ t L‘
ł kGðsÞkL1 khe t kL‘ lim ðsref ðtÞ  sðtÞÞ = 0 ð60Þ
G!‘, eH !0
~ t kL‘ + kGðsÞkL1 + kHðsÞkL1 eH
+ kCðsÞkL1 ks
lim ðuref ðtÞ  uðtÞÞ = 0 ð61Þ
ð54Þ G!‘, eH !0

From (16) and (14), it follows that an upper bound on Proof. It is easy to verify that, for the upper bounds c0 ,
khe t kL‘ is given by g 1 and g2 , introduced respectively in (39), (48) and
 
(49), we have lim c0 = 0, lim g 1 = 0 and
G!‘, eH !0 G!‘, eH !0
k h e t k L‘ = vT ðpref ÞuH  vT ðpÞuH t L
  
‘
 lim g 2 = 0. Recalling that ksref  skL‘ ł g 1 and
ł F x, z, u, u jsref  F x, z, u, uH js
H
+ 2eH G!‘, eH !0
t L‘
kuref  ukL‘ ł g 2 , it follows that lim
ł L ðsref  sÞ t L‘
+ 2eH G!‘, eH !0
ðsref ðtÞ  sðtÞÞ = 0 and lim ðuref ðtÞ  uðtÞÞ = 0.
ð55Þ G!‘, eH !0
This completes the proof.
Substituting (55) into (54) and solving for
ðsref  sÞ t L‘
leads to Remark 2. It is important to point out that the control
law uref ðtÞ in the closed-loop reference system is not
kCðsÞkL1 implementable, since it depends on the unknown quan-
ðsref  sÞ t ł ks
~ t kL ‘
L‘ 1  kGðsÞkL1 L tities O and uH , and it is used only for analysis pur-
ð56Þ poses. From Theorem 1 and Corollary 1, we conclude
3kGðsÞkL1 + kHðsÞkL1
+ eH that, if the approximation error eH of the fuzzy system
1  kGðsÞkL1 L is small enough and the adaptive gain G is chosen suffi-
ciently large, the L1 fuzzy adaptive controller can force
kCðsÞk pffiffiffiffiffi
consequently ðsref  sÞ t L‘ ł 1kCðsÞkL1 L c0 + the signals sðtÞ and uðtÞ to track sref ðtÞ and uref ðtÞ of
L1
  the stable reference system (29) both in transient and
3kGðsÞkL + kHðsÞkL
1 1
eH which holds uniformly for all steady-state. Thus, the control design is reduced to the
1kCðsÞk L L1
selection of l, a and CðsÞ such that the closed-loop ref-
t . 0 and proves the first bound in (47).
erence system has the desired response. h
Next, to prove the second bound in (47), we derive
the following relationship from (50) and (29) Remark 3. It is worth noting that the bandwidth-limited
filter CðsÞ plays a key role in the proposed control
C ðsÞ CðsÞ
uref ðsÞ  uðsÞ = he ð s Þ  ~ ðsÞ
h ð57Þ scheme. Notice that if we set CðsÞ = 1, the norm of
O O H1 ðsÞ in (58) is reduced to H1 ðsÞ = H1ðsÞ which is not
We refer to Lemma A.12.1 in1 and rewrite CðsÞ
~ ðsÞ as bounded, since HðsÞ is strictly proper. Consequently,
O h
follows without the bandwidth-limited filter, i.e., CðsÞ = 1, one
cannot achieve a uniform performance bound for the
C ðsÞ 1 CðsÞ control signal. h
~ ðsÞ = H1 ðsÞ~
h sðsÞ + eðsÞ ð58Þ
O O O
Boubakir et al. 743

_ y€T , the physical

actuator. Denoting H = ½pP , pN , y, y,
domain Z of the used pneumatic actuator system is
defined as

Z = f H j1 bar ł pP ł 7 bar, 1bar ł pN ł 7bar,

jyj ł 72mm, jy_ j ł 1m=s, jy€j ł 15m=s2 g ð64Þ

Let us denote x = ½x1 , x2 , x3 T = ½y, y, _ y€T and

T T
z = ½z1 , z2  = ½pP , pN  , from (63) and (62), the full
dynamic model of the actuator reads as the following
nonaffine nonlinear system
8
Figure 1. Photo of the used pneumatic experimental set-up > x_ 1 = x2 , x_ 2 = x3
>
>
located at IRCCyN. < x_ 3 = fðx, z, uÞ
ð65Þ
>
> z_ = Qðx, z, uÞ
>
:
Application to an electropneumatic y = x1
actuator The functions fðx, z, uÞ and Qðx, z, uÞ in (65) are given
System description by
The electropneumatic system considered for the experi-  0 T qm ðu, z1 Þ qm ðu, z2 Þ
Skr
mental validation is composed of two actuators which fðx, z, uÞ = 
M VP ðx1 Þ VN ðx1 Þ
are controlled by two servodistributors, as represented  2  z1
kS z2

1  
in Figures 1 and 2. The first is the main actuator and  + x2  bv x2 + F_ ext
the second the perturbation actuator, which is used to M VP ðx1 Þ VN ðx1 Þ M
ð66Þ
     0T   T
kr0 T S kr S
Qðx, z, uÞ = qm ðu, z1 Þ  z1 x 2 , qm ðu, z2 Þ + z2 x2 ð67Þ
VP ðx1 Þ r0 T VN ðx1 Þ r0 T
Consequently, the input–output representation of (65)
generate a dynamical load force on the main actuator.32 can be written as
The performances of the proposed controller have been 
examined in the control of the main actuator position. y(3) = fðx, z, uÞ
ð68Þ
This latter consists of two chambers, denoted P and N z_ = Qðx, z, uÞ
and equipped with a pneumatic jack able to provide a which is in the general input–output form given by (2),
horizontal movement for a load carriage of mass M. with r = 3 in this case.
Under some assumptions, as detailed by Taleb et
al.,33 the pressure dynamics in both chambers N and P, Remark 4. The serious challenge for the dynamic model
denoted respectively by pN and pP , read as (65) is the knowledge of the mass flow rates qm ðu, pP Þ
8 and qm ðu, pN Þ. To overcome this drawback, in the lit-
>  0T 
kr S

erature these unknown quantities have been experimen-
>
>
< p_ P = V ðyÞ qm ðu, pP Þ  r T pP v tally identified by an affine model:33,34 qm ðu, pP Þ =
P 0
  ð62Þ up + cp u and qm ðu, pN Þ = uN  cN u with uj . 0 and
>
> 
kr T S
>
: p_ N =
0
qm ðu, pN Þ + p v cj . 0, j 2 ðP, NÞ, which are fifth-order polynomials
VN ðyÞ r0 T N
with respect to pj . It is worth pointing out that the
On the other hand, the motion dynamics of the actua- experimental identification of the flow rate qm ðÞ is not
tor is governed by needed when we use the proposed control scheme, both
( in control design and implementation phases. h
y_ = v
1 ð63Þ Remark 5. Since the mass flow rates qm ðu, pP Þ and
v_ = ðSðpP  pN Þ  bv v  Fext Þ qm ðu, pN Þ are unknown we cannot prove that
M
Assumption 1 is satisfied for system (65). For this rea-
where VP ðyÞ = V0 + Sy and VN ðyÞ = V0  Sy with V0 son, we only assume that fu ðx, z, uÞ is strictly positive
is the half-cylinder volume, S denotes the piston sur- and bounded without proving it. However, it is worth
face, y the actuator position (piston position), v is the noting that, when we use the experimental identifica-
piston velocity, u the control input voltage bounded as tion of the mass flow rates, it results that
juj ł 10V, k represents the polytropic constant, T the SkrT cP cN
chamber temperature, r0 the perfect gas constant, bv fu ðx, z, uÞ = M V P ð yÞ + VN ðyÞ .0 and bounded,
the viscous friction coefficient, and Fext denotes the which implies that Assumption 1 is satisfied in this
external force produced by the ‘‘perturbation’’ case. h
744 Proc IMechE Part I: J Systems and Control Engineering 230(8)

Figure 2. L1 fuzzy adaptive controller applied to the pneumatic actuator system.

Experiment results adaptation gain, DðsÞ in (26) is chosen as DðsÞ = 1s , the

The experimental results relative to position control of projection bounds ^ ðtÞ 2 ½0:01, 20
O and
the pneumatic actuator system, using the designed L1 ^uðtÞ 2 ½20, 20127 where 127 2 R27 is the vector with all
fuzzy adaptive controller, are presented in this subsec- elements 1, the initial values of u^ðtÞ are set equal to zero
tion. The experimental study was done with the experi- and O^ ð0Þ = 0:01 is the initial value of O ^ ðtÞ. To test the
mental set-up located at IRCCyN’s laboratory in robustness of the controller, the perturbation actuator
Nantes, France and equipped with a dSpace DS1104 produces and applies an external force to the controlled
controller board on which the control law was imple- actuator, as illustrated in Figure 3(bottom).
mented. The control objective was to force the piston The experiment results are reported in Figures 4–10.
position y to track the reference trajectory The piston position y and its reference trajectory yr are
yr ðtÞ = 50 sinð0:4ptÞ mm. given in Figure 4 and the resulting error curve
Since the relative degree of the piston position is e = yr  y in Figure 3(top). From these figures, we can
r = 3, we chose the filtered tracking error s as see that the piston position tracks the reference signal
 2
sðtÞ = dtd + l eðtÞ. The control law was selected in and the tracking error remains approximately zero. We
the form of (25), its adaptive laws as in (23) and (24), note also that, in spite of the effects of the perturbation
and the predictor as in (22). A fuzzy system in the form force and the impact of measurement noise, the control
 T
ðp + p Þ input u, described by Figure 5, is bounded. Figures 9
of (20) that has p = u, P 2 N , d as input is used and 10 show the time evolution of ^uðtÞ and O ^ ðtÞ. It is
in this experiment. For each variable pi , i = 1, 2, 3, clear that these estimated parameters are bounded dur-
three Gaussian ing the control process. Figure 6 indicates that the pre-
 membership functions in the form diction s^ ðtÞ converges to the actual filtered tracking
p1 + jij 2
mF j ðpi Þ = exp 0:5 ki , j = 1, 2, 3, are used error sðtÞ. Figures 7 and 8 illustrate that the pressures
i
pN and pP in the two chambers P and N remain
with centers j1j = ½2, 0, 2T and a variance k1 = 0:5
bounded.
for the variable p1 , j 2j = ½2,  4,  6T , k2 = 0:5 for To highlight the performances and benefits of the
the variable p2 , j 3j = ½0:04, 0, 0:04T and k3 = 0:01 proposed L1 fuzzy adaptive controller, an experimental
for the variable p3 . The L1 fuzzy adaptive controller is comparative study has been conducted under the same
implemented with sampling time set to Ts = 0:001 s conditions between the L1 fuzzy adaptive controller
and the design parameters are selected as follows: and a PID and PI controller. The gains of the PID and
l = 2:5, a = 10, the feedback gain K = 65, the PI conventional controllers considered in this
Boubakir et al. 745

Figure 3. (top) Tracking position error (mm) versus time (s). Figure 6. Actual filtered tracking error sðtÞ (solid line) and its
(bottom) Perturbation force (N) versus time (s). ^ ðtÞ (dotted line) versus time (s).
prediction s

Figure 4. Actual position y (solid line) and its reference yr ðtÞ

(dotted line) (mm) versus time (s). Figure 7. Pressure in chamber N (bar) versus time (s).

Figure 5. Control input u(V) versus time (s). Figure 8. Pressure in chamber P (bar) versus time (s).
746 Proc IMechE Part I: J Systems and Control Engineering 230(8)

Table 1. Comparisons of position error e when yr ðtÞ = 50 sinð0:2ptÞ mm.

Control strategy Avg (abs(e)) Range of e Standard deviation of e

L1 fuzzy adaptive controller 4.18673104 ð1:9, 2:2Þ3103 5.37723104

PID Controller 343104 ð8:5, 8Þ3103 403104
PI Controller 343104 ð8:4, 7:4Þ3103 393104

Table 2. Comparisons of position error e when yr ðtÞ = 50 sinð0:4ptÞ mm.

Control strategy Avg(abs(e)) Range of e Standard deviation of e

L1 fuzzy adaptive controller 7.22393104 ð2:4, 3Þ3103 0.904803103

PID Controller 573104 ð11:5, 12:2Þ3103 6.53103
PI Controller 593104 ð11:8, 13:9Þ3103 6.83103

^ ðtÞ versus time (s).

Figure 9. Estimate parameter O Figure 10. Evolution of the estimate parameters Euclidian
norm ^uðtÞ versus time (s).

comparative study are selected as those which give the

best performance and have been obtained through trial for the pneumatic actuator system, contrary to the tra-
and error. We ditional PID controller.
 implemented the PID controller
 with the
parameters Kp= 250, KI = 40, Kd= 15 and the PI
controller with Kp = 250, KI = 50 . The performance
of the three controllers was examined with two differ-
Conclusion
ent trajectories and based on the characteristics of the This paper proposes an L1 fuzzy adaptive controller to
resulting position tracking error e in the time interval provide satisfactory transient response and guarantee
t 2 ½5 s, 15s. Tables 1 and 2 contain the average of the stable tracking for a class of uncertain SISO nonaffine
absolute value of the position tracking error e, the nonlinear systems. The presented controller has been
range of e and its standard deviation for the desired developed based on L1 adaptive control design philoso-
trajectories yr ðtÞ = 50 sinð0:2ptÞ mm and phy and incorporates a fuzzy system, which is used to
yr ðtÞ = 50 sinð0:4ptÞ mm, respectively. estimate a function of an unknown ideal implicit con-
Based on the characteristics of the resulting position troller. It has been demonstrated that when we use the
tracking error e reported in Tables 1 and 2 for each con- proposed control method, the control signal and the
trol method, it can be said that the proposed L1 fuzzy system’s response can arbitrarily approximate the cor-
adaptive controller yields a much better tracking per- responding signals of a stable closed-loop reference sys-
formance than the used PID and PI controllers. The tem, which is introduced for analysis purposes and to
superiority of the proposed L1 fuzzy adaptive controller achieve the desired specifications. The proposed L1
over the traditional PID controller is mainly due to the fuzzy adaptive controller represents an efficient and
ability of the presented control scheme to guarantee stable model free controller for the general class of
uniformly bounded transient and tracking performance SISO nonaffine nonlinear systems considered in this
Boubakir et al. 747

research work. The obtained experimental results con- In: Proceedings of the AIAA guidance, navigation, and con-
firm the performance of the proposed L1 fuzzy adap- trol conference, Chicago, IL, USA, 10–13 August 2009,
tive controller. paper no. AIAA 2009-5754, pp.1–15.
14. Jin W, Bifeng S, Liguang W, et al. L1 adaptive dynamic
inversion controller for an X-wing tail- sitter MAV in
Declaration of conflicting interests hover flight. Procedia Eng 2015; 99: 969–974.
The author(s) declared no potential conflicts of interest 15. Bennehar M, Chemori A and Pierrot F. L1 adaptive con-
with respect to the research, authorship, and/or publi- trol of parallel kinematic manipulators: Design and real-
cation of this article. time experiments. In: IEEE international conference on
robotics and automation (ICRA), Seattle, USA, 26–30
May 2015, pp.1587–1592.
Funding 16. Malinga B and Buckner GD. L1 adaptive control of a
shape memory alloy actuated flexible beam. Syst Sci Con-
The author(s) received no financial support for the
trol Eng 2015; 3(1): 460–471.
research, authorship, and/or publication of this article.
17. Roshanian J, Zareh M, Afshari HH, et al. An L1 adaptive
closed-loop guidance law for an orbital injection problem.
References Proc IMechE I: J Syst Control Eng 2009; 223(6): 753–761.
18. Lee KW and Singh NS. Multi-input submarine control
1. Hovakimyan N and Cao C. Adaptive control theory: via L1 adaptive feedback despite uncertainties. Proc
guaranteed robustness with fast adaptation. Philadelphia, IMechE I: J Syst Control Eng 2014; 228(5): 330–347.
PA: Siam - Society for Industrial and Applied Mathe- 19. Maalouf D, Chemori A and Creuze V. L1 Adaptive depth
matics, 2010. and pitch control of an underwater vehicle with real-time
2. Ioannou PA and Sun J. Robust adaptive control. Engle- experiments. Ocean Eng 2015; 98: 66–77.
wood Cliffs, NJ: Prentice Hall, 1996. 20. Xiaofeng W and Hovakimyan N. L1 adaptive controller
3. Cao C and Hovakimyan N. Design and analysis of a for nonlinear reference systems. In: American control con-
novel L1 adaptive controller, Part I: control signal and ference, San Francisco, CA, USA, 29 June–1 July 2011,
asymptotic stability. In: American control conference, pp.594–599. New York: IEEE.
Minneapolis, MN, USA, 14–16 June 2006, pp.3397– 21. Xiaofeng W and Hovakimyan N. L1 adaptive controller
3402. New York: IEEE. for nonlinear time-varying reference systems. Syst Control
4. Cao C and Hovakimyan N. Design and analysis of a Lett 2012; 61(4): 455–463.
novel L1 adaptive control architecture, Part II: guaran- 22. Luo J, Cao C and Hovakimyan N. L1 adaptive controller
teed transient performance. In: American control confer- for a class of systems with unknown nonlinearities. In:
ence, Minneapolis, MN, USA, 14–16 June 2006, American control conference, Baltimore, MD, USA, 30
pp.3403–3408. New York: IEEE. June–2 July 2010, pp.1659–1664. New York: IEEE.
5. Cao C and Hovakimyan N. Design and analysis of a 23. Zou X, Cao C and Hovakimyan N. L1 adaptive control-
novel L1 adaptive control architeture with guaranteed ler for systems with hysteresis uncertainties. In: American
transient performance. IEEE Trans Autom Control 2008; control conference, Baltimore, MD, USA, 30 June–2 July
53(2): 586–591. 2010. pp.6662–6667. New York: IEEE.
6. Cao C and Hovakimyan N. Adaptive controller for sys- 24. Choe R, Xargay E and Hovakimyan N. L1 adaptive con-
tems with unknown time-varying parameters and distur- trol for a class of uncertain nonaffine-in-control non-
bances in the presence of non-zero trajectory linear systems. IEEE Trans Autom Control 2015; 61(3):
initialization error. Int J Control 2008; 81(7): 1147–1161. 840–846.
7. Cao C and Hovakimyan N. L1 adaptive output feedback 25. Luo J, Cao C and Yang Q. L1 adaptive controller for a
controller for systems of unknown dimension. IEEE class of non–affine multi–input multi–output nonlinear
Trans Autom Control 2008; 53(3): 815–821. systems. Int J Control 2013; 86(2): 348–359.
8. Cao C and Hovakimyan N. Stability margins of L1 26. Wang LX. Adaptive fuzzy systems and control: design and
adaptive control architecture. IEEE Trans Autom Control stability analysis. Englewood Cliffs, NJ: Prentice Hall,
2010; 55(2): 480–487. 1994.
9. Kharisov E, Hovakimyan N and Astrom KJ. Compari- 27. Labiod S and Guerra TM. Direct adaptive fuzzy control
son of architectures and robustness of model reference for a class of MIMO nonlinear systems. Int J Syst Sci
adaptive controllers and L1 adaptive controllers. Int J 2007; 38(8): 665–675.
Adapt Control Signal Process 2014; 28(7–8): 633–663. 28. Slotine JJ and Li W. Applied nonlinear control. Engle-
10. Hashim HA, El-Ferik S and Abido MA. A fuzzy logic wood Cliffs, NJ: Prentice Hall, 1991.
feedback filter design tuned with PSO for adaptive con- 29. Khalil HK. Nonlinear systems. 2nd ed. Englewood Cliffs,
troller. Expert Syst Appl 2015; 42(23): 9077–9085. NJ: Prentice-Hall, 1996.
11. Cooper J, Che J and Cao C. The use of learning in fast 30. Malik SC. Mathematical analysis. New York: John Wiley
adaptation algorithms. Int J Adapt Control Signal Pro- & Sons, 1984.
cess 2014; 28(3–5): 325–340. 31. Grossman SI and Derrick WR. Advanced engineering
12. Dobrokhodov V, Kaminer I, Kitsios I, et al. Experimen- mathematics. New York: Happer & Row, 1998.
tal validation of L1 adaptive control: The Rohrs counter- 32. Boubakir A, Plestan F, Labiod S, et al. A stable linear
example in flight. J Guid control Dyn 2011; 34(5): 1311– adaptive controller applied to a pneumatic actuatorsys-
1328. tem. In: IEEE conference on decision and control (CDC),
13. Michini B and How JP. L1 adaptive control for indoor Florence, Italy, 10–13 December 2013, pp.6112–6117.
autonomous vehicles: design process and flight testing. New York: IEEE.
748 Proc IMechE Part I: J Systems and Control Engineering 230(8)

33. Taleb M, Levant A and Plestan F. Pneumatic actuator pN Pressure in the chambers N
control: Solution based on adaptive twisting and experi- pP Pressure in the chambers P
mentation. Control Eng Pract 2013; 21(5): 727–736. qm ðÞ Mass flow rate provided by the servo-
34. Plestan F, Shtessel Y, Bregeault V, et al. New methodolo- distributor to cylinder chamber
gies for adaptive sliding mode control. Int J Control 2010; r Relative degree
83(9): 1907–1919. u, y Control input, regulated output
uH Ideal control law
Appendix 1 x, z State vectors
^, s
s, s ~ Filtered tracking error, prediction of s,
Notation
prediction error
C ðsÞ Low-pass filter transfer function p, vðpÞ Input vector of the fuzzy system, basis
D ðsÞ Design factor in low-pass filter transfer function vector
function u, uH Vector containing consequent parameters
f u ð Þ Control effectiveness of the fuzzy system, optimal unknown
Fext External force produced by the vector
perturbation actuator ^ ^u
O, Adaptive estimates
K Low-pass filter bandwidth parameter G Adaptive gain