Professional Documents
Culture Documents
2013 Fuzz Bezzaoucha 2
2013 Fuzz Bezzaoucha 2
net/publication/236162695
CITATIONS READS
9 207
4 authors:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by José Ragot on 27 May 2014.
e-mail:firstname.name@univ-lorraine.fr.
Abstract—This paper addresses the stabilization of nonlinear such that the system stability is ensured and the control gains
systems described by Takagi-Sugeno models affected by input are calculated depending on the saturation level.
actuator saturation. A parallel distributed compensation design The authors would like to precise that even if the expressions
is used for the state feedback controller. Stabilization conditions of the T-S saturation has some similarity to the polytopic
in the sense of the Lyapunov method are derived and expressed as one used in [8] and [9], the development, control strategy
a linear matrix inequality problem. The obtained gains depend
on the actuator saturation limits. A cart-pendulum example is
and objectives are completely different since in the proposed
presented to illustrate the effectiveness of the proposed approach. approach the invariant sets are never considered; moreover the
objective is the stabilization of nonlinear systems represented
with the Takagi-Sugeno (T-S) models and the synthesis of a
Keywords—Takagi-Sugeno systems, actuator saturation, parallel state feedback controller by parallel distributed compensation
distributed compensation control law (PDC) with control gains explicitly depending on the satu-
ration level. The controller is subject to actuator saturation
I. INTRODUCTION and the input saturation is directly taken into account in the
controller design. Stabilization conditions are derived from the
Actuator saturation or control input saturation is probably Lyapunov method and expressed as linear matrix inequalities
one of the most usual nonlinearity encountered in control (LMI). A cart-pendulum example is considered to illustrate
engineering due to the physical impossibility of applying the effectiveness of the proposed approach. The nonlinear
unlimited control signals and/or safety constraints. equations of the cart-pendulum are given and, using the Sector
In general, there are two main design strategies to deal with Nonlinearity Transformation (SNT) [10], [11], a T-S model
actuator saturations. The first strategy is a two-step approach of the system is deduced and used. The system is subject
in which a nominal linear controller is first constructed by to actuator saturations, it will be shown that these input
ignoring the actuator saturation. Then, after this controller has constraints may cause the instability of the nonlinear system,
been designed, usually using standard linear design tools, a but with the proposed approach the stability of the closed-loop
so called anti-windup compensator is designed to handle the system is ensured.
saturation constraints [1], [2], [3], [4]. A typical anti-windup The rest of this paper is organized as follows. Section 2
scheme consists in augmenting a nominal pre-designed linear introduces the Takagi-Sugeno structure for modeling and some
controller with a compensator based on the discrepancy be- preliminary results, mathematical notations and a brief descrip-
tween unsaturated and saturated control signals fed to the plant tion of the saturation. It is followed by the representation of the
[5]. The second strategy considers the saturation constraint nonlinear saturation by a T-S structure in section 3. In section
from the beginning of the design task. Several approaches were 4 is designed a state feedback control law depending on the
developed, one of them is the invariant sets framework, which saturation bounds. A numerical example and some simulation
has been significantly developed in control engineering over results are given in section 5. Conclusions and future works
the last decades, see [6], [7]. This framework ensures that are exposed in section 6.
every state trajectory that starts inside the invariant set will
not exceed it, i.e. the system states will be bounded inside this
set.
An interesting approach for the determination of the invariant
sets framework was applied in [8] and [9]. This approach
II. P RELIMINARIES
consists in the so-called polytopic rewriting of the saturation
constraint. This polytopic representation is then used to deter-
minate (maximize) the largest invariant set in which the system A. Takagi-Sugeno structure for modeling
states remain bounded.
In the contrary of the previous method, that most likely ensures Initially introduced in [12], the T-S models represent a
to never reach the saturation level in order to preserve the simple and accurate method to study the nonlinear behaviors.
closed loop performances, in this paper, we explicitly consider The T-S modeling allows to represent the behavior of nonlinear
the saturation constraint, and even admit its occurrence. In fact, systems by the interpolation of a set of linear submodels [10],
the authors use the Takagi-Sugeno (T-S) representation of the [13], [14]. Each submodel contributes to the global behavior
saturation (as well known as the polytopic representation) to of the nonlinear system through a weighting function µi (ξ(t)).
integrate the limitation constraints into the control synthesis, The T-S structure is given by
where ujsat = sat(uj (t)), for j = 1, . . . , nu .
n
X Considering the three parts of the saturated signal (4), each
component of the vector usat (t) is written as
ẋ(t) = µi (ξ(t))(Ai x(t) + Bi u(t))
i=1 3
n (1) X
ujsat (t) = µji (uj (t)) (λji uj (t) + γij ), j = 1, . . . , nu (7)
X
y(t) = µi (ξ(t))(Ci x(t) + Di u(t))
i=1
i=1
with
where x(t) ∈ R is the system state variable, u(t) ∈ Rnu is
nx
λj1 = 0 λj2 = 1 λj3 = 0 (8)
the control input and y(t) ∈ Rm is the system output. ξ(t) ∈
Rq is the decision variable vector assumed to be measurable γ1j = ujmin γ2j = 0 γ3j = ujmax (9)
(as the system output) or known (as the system input). The
weighting functions µi (ξ(t)) of the n submodels satisfy the and the weighting functions defined by
convex sum property
1−sign(uj (t)−ujmin )
n
µ1 (uj (t)) = 2
X
sign(uj (t)−ujmin )−sign(uj (t)−ujmax ) (10)
µi (ξ(t)) = 1 µ2 (uj (t)) = 2
(2)
i=1 1+sign(uj (t)−ujmax )
µ3 (uj (t)) =
0 ≤ µ (ξ(t)) ≤ 1, i = 1, . . . , n
i 2
In the remaining of the paper, the following lemmas are used: Based on the convex sum property of the weighting functions
(2), the control input vector u(t) ∈ Rnu subject to actuator
Lemma 1. Consider two matrices X and Y with appropriate saturation can be written in order to have the same activation
dimensions and G a symmetric positive definite matrix. The functions for all the input vector components as:
following property is verified
usat (t) =
X T Y + Y T X ≤ X T GX + Y T G−1 Y (3) !
3
X nu X
Y 3
P −1
P1 =
(36) The pendulum rotates in a vertical plan around an axis
Rj = Kj P1
located on a cart. The cart can move along a horizontal rail,
condition (35) is linearized as lying in the rotation plane. The state of the system is a vector
T
x = ( x1 x2 x3 x4 ) where x1 = z(t) is the cart
S(P1 ATi − RjT ΛTk BiT ) + Σ−1k <0 position, x2 = ż(t) the cart velocity, x3 = θ(t) is the angle
i = 1, . . . , n; j = 1, . . . , n; k = 1, . . . , 3nu (37) between the downward direction and the pendulum (x2 = 0
Applying Schur’s complement to (37), it becomes for the upright position of the pendulum), and x4 = θ̇(t) is the
pendulum angular velocity. A control force F (t) parallel to the
S(P1 ATi − RjT ΛTk BiT )
I rail is applied to the cart. The pendulum mass and cart mass
Qijk = <0 (38) are denoted m = 1Kg and M = 5Kg respectively. l = 0.1m
I −Σk
is the pendulum length and I the moment of inertia of the
As the weighting functions satisfy (2) and Σk > 0, if (38) pendulum with respect to its axis on the cart. The cart friction
is satisfied for i = 1, . . . , n, j = 1, . . . , n, k = 1, . . . , 3nu , is composed of the viscous friction proportional to the cart
and k x k2 > δε , then V̇ (x(t)) < 0, which implies q that x(t) velocity f ż(t) (f = 100N/m/s), a friction proportional to
converges to an origin centered ball of radius δ
. the cart position ks z(t) (ks = 0.001N/rad/s) and a friction
qε torque in the angular motion of the pendulum proportional to
δ
The objective is now to minimize the radius ε. Firstly δ the angular velocity k θ̇(t) (k = 0.003N/rad/s). The state
equations are as follows: with
zi1 = max zi (t), zi2 = min zi (t), i = 1, . . . , 4 (47)
(m + M )z̈(t) + ks z(t) + f ż(t) − mlθ̈(t)cos(θ(t))
(
+mlθ̇2 (t)sin(θ(t)) = F (t) Fi1 (t) = zzii1
(t)−zi2
−zi2
(48)
zi1 −zi (t)
Fi2 (t) = zi1 −zi2
−mlz̈(t)cos(θ(t)) + (ml2 + I)θ̈(t)
The two partitions Fi1 (t) and Fi2 (t) of each premise variable
+k θ̇(t) + mglsin(θ(t)) = 0
zi (t), i = 1, 2, 3, 4 will contribute to the reconstruction of each
(40) weighting function of the T-S model. The sub-model number
In order to ensure the nonlinear system stability, the system depends on the non-linearities, such that r = 24 = 16.
equations (40) is written under a T-S from by applying the The T-S model is then given by the following equations:
sector nonlinearity approach which allows to exactly represent
16
the nonlinear system without any loss of information. Since X
the transition from the nonlinear system to the T-S one is not ẋ(t) = µi (t)(Ai x(t) + Bi u(t)) (49)
the main subject of this paper and for place limitation, only i=1
the main steps are given. with 16
First, the quasi-LPV form (quasi-linear with parameter vary- Y
ing) is derived
µi (t) = Fj,σj (zj (t))
i
j=1 (50)
ẋ(t) = A(x(t))x(t) + B(x(t))u(t) (41)
Ai = A(z1,σi1 , z2,σi2 , z3,σi3 , z4,σi4 )
Bi = B(z1,σi1 , z2,σi2 , z3,σi3 , z4,σi4 )
with :
T σij may take the value 1 or 2 and represents the number of the
x(t) = z(t) ż(t) θ(t) θ̇(t) , u(t) = F (t) (42) j th partition.
In order to simplify the LMI implementation, it is assumed that
0 1 0 0
the control gains Kj are equal for all the submodels. Then, two
−ks /(m + M ) a1 (t) a2 (t) a3 (t) control laws are proposed:
A(x(t)) =
0 0 0 1 u(t) = −Kn x(t) and u(t) = −KT S x(t) (51)
a7 (t) a4 (t) a5 (t) a6 (t)
T for the nominal controller (without input saturation) and the
B(t) = ( 0 b1 (t) 0 b2 (t) )
(43) proposed T-S controller respectivelly.
and the parameters ai (t) and bi (t) are defined as follows: The nominal control gain calculated from (21) is equal to:
mlf z1 (t)z22 (t)
Kn = ( 3.4068 −95.7038 −29.8001 3.8235 ) (52)
f
a1 (t) = − m+M −
m+M
Considering the saturation levels equal umin = 0 and umax =
a2 (t) = − mlzm+M 3 (t)z4 (t)
− 3, The control gain KT S is then calculated by solving the LMI
m2 l2 z22 (t)z1 (t)z3 (t)z4 (t)
− mlgz1 (t)z2 (t)z3 (t) given by theorem 1 and is equal to:
m+M
a (t) = kz1 (t) KT S = ( 0.0003 5.9835 0.6139 −0.0024 ) (53)
3
a4 (t) = −f z1 (t)z2 (t) T
For an initial position x0 = ( 0 0 π/12 0 ) , figure 2
a5 (t) = −mlz1 (t)z2 (t)z3 (t)z4 (t) − (m + M )gz1 (t)z3 (t)
shows the system states for the following three cases: xn for
a6 (t) = − k(m+M ) the nominal case without saturation, xnsat is the system states
ml z1 (t)
a (t) = −k z (t)z (t) with saturated nominal control, it is clear from the depicted
7 s 1 2
1 ml
figure that the saturation has a destabilizing effect on the
b (t) = m+M + m+M z22 (t)z1 (t)
1 system. And finally xsatT S denotes the system state obtained
b2 (t) = z1 (t)z2(t) with the proposed T-S approach.
(44) From the depicted figures, it is clear that with the proposed
with the following non-linearity : T-S approach the system stability is ensured as well as the
1
convergence to an origin centered-ball. However, since we
z1 (t) = −mlcos2 (x3 (t))+(l+I/ml)(m+M ) considered the same gain for all the 16 sub-models, the
z2 (t) = cos(x3 (t))
obtained solution is slightly conservative, which may justify
(45) the oscillatory response for the third and fourth state.
z3 (t) = sin(x 3 (t))
x3
In order to relax these constraints, the nonlinear system may
z4 (t) = x24 (t)
be simplified by considering the small angles approximations:
The reader will note that the nonlinearities choice cos(θ) ≈ 1 and sin(θ) ≈ θ. Hence, instead of the four previous
(z1 (t), z2 (t), z1 (t) and z4 (t)) is not unique and that the non-linearities, only one is considered, namely: z4 (t) = x24 (t)
premise variables are bounded and given by the angle and and a T-S system with two sub-models is then defined with
speed limitation. Now, the convex polytopic transformation is the T-S controller is then given by
applied 2
X
u(t) = − µi (ξ(t))KiT S x(t) (54)
zi (t) = Fi1 (t)zi1 + Fi2 (t)zi2 , i = 1, . . . , 4 (46) i=1
saturation, these saturations are expressed in terms of the con-
0.2
x
1n
0.4
x
2n
trol variables and of the saturation limits. As a consequence,
0.15
x
x
1nsat
0.3
x
x
2nsat these limits are taken into account when solving the LMI
1satTS 2satTS
optimization problem and when computing the gains of the
0.1 0.2
controller. Moreover, the proposed Takagi-Sugeno approach
0.05 0.1 allows to extend the use of linear tools, namely the LMI
0 0
formalism, to a nonlinear control problem.
As an illustrative example, the state feedback controller design
−0.05
0 2 4 6 8 10
−0.1
0 2 4 6 8 10 was applied to a cart-pendulum nonlinear T-S system. The ob-
tained results show the effectiveness of the proposed approach
for the studied example.
0.3 2
x3n x4n
0.2 x3nsat x4nsat
0.1
x3satTS
1
x4satTS R EFERENCES
0 0 [1] A. Syaichu-Rohman and R. Middleton, “Anti-windup schemes for
−0.1 discrete time systems: An LMI-based design,” in 5th Asian Control
−1
Conference, Australia, 20-23 July 2004.
−0.2
−0.3
−2 [2] L. Zaccarian and A. Teel, “A common framework for anti-windup,
bumpless transfer and reliable design,” Automatica, vol. 38, no. 10,
−0.4 −3
0 2 4 6 8 10 0 2 4 6 8 10 pp. 1735–1744, 2002.
[3] E. Mulder, M. Kothare, and M. Morari, “Multivariable antiwindup con-
troller synthesis using linear matrix inequalities,” Automatica, vol. 37,
no. 9, pp. 1407–1416, 2001.
Fig. 2. System states [4] G. Grimm, J. Hatfield, I. Postlethwaite, A. Teel, M. Turner, and L. Za-
ccarian, “Antiwindup for stable linear systems with input saturation:
an LMI-based synthesis,” IEEE Transactions on Automatic Control,
with vol. 48, no. 9, pp. 1509–1525, 2003.
[5] A. Syaichu-Rohman, “Optimisation based feedback control of input
K1T S = ( 0.0127 −15.0442 15.8757 0.7854 ) constrained linear systems,” Ph.D. dissertation, The University of New-
(55) castle Callaghan, Australia, 2005.
K2T S = ( 0.0079 −19.0341 8.7667 0.5326 )
[6] A. Alessandri, M. Baglietto, and G. Battistelli, “On estimation error
As seen on Figure 3, displaying the obtained state responses, bounds for receding-horizon filters using quadratic boundedness,” IEEE
Transactions on Automatic Control, vol. 49, no. 8, pp. 1350–1355, 2004.
[7] F. Blanchini and S. Miani, Set-theoretic methods in control. Systems
0.2 0.4
and Control: Foundations and Applications. Boston-Basel-Berlin:
x1n x2n
Birkhauser, 2008.
x1nsat x2nsat
0.15 0.3
x1satTS x2satTS [8] Y. Cao and Z. Lin, “Robust stability analysis and fuzzy-scheduling
control for nonlinear systems subject to actuator saturation,” IEEE Fuzzy
0.1 0.2
Sets and Systems, vol. 11, no. 1, pp. 57–67, 2003.
0.05 0.1 [9] S. Tarbouriech, G. Garcia, J. Gomes da Silva Jr., and I. Queinnec,
Stability and Stabilization of Linear Systems with Saturating Actuators.
0 0
London: Springer-Verlag, 2011.
−0.05
0 2 4 6 8 10
−0.1
0 2 4 6 8 10
[10] K. Tanaka and H. Wang, Fuzzy Control Systems Design and Analysis:
A Linear Matrix Inequality Approach. New York: John Wiley & Sons,
Inc., 2001.
0.3 2 [11] A. Nagy, G. Mourot, B. Marx, J. Ragot, and G. Schutz, “Systematic
x3n x4n
0.2 x3nsat x4nsat
multimodeling methodology applied to an activated sludge reactor
1
x3satTS x4satTS model,” Industrial & Engineering Chemestry Research, vol. 49, no. 6,
0.1
0 pp. 2790–2799, 2010.
0
−1 [12] T. Takagi and M. Sugeno, “Fuzzy identification of systems and its
−0.1 applications to modeling and control,” IEEE Transactions on Systems,
−2
−0.2 Man, and Cybernetics, vol. 15, no. 1, pp. 116–132, 1985.
−0.3 −3 [13] K. Tanaka, T. Hori, and H. Wang, “A multiple Lyapunov function
−0.4 −4 approach to stabilization of fuzzy control systems,” IEEE Transactions
0 2 4 6 8 10 0 2 4 6 8 10
on Fuzzy Systems, vol. 11, no. 4, pp. 582–589, 2003.
[14] K. Tanaka, H. Ohtake, and H. Wang, “A descriptor system approach
to fuzzy control system design via fuzzy Lyapunov functions,” IEEE
Transactions on Fuzzy Systems, vol. 15, no. 3, pp. 333–341, 2007.
Fig. 3. System states
[15] K. Zhang, B. Jiang, and P. Shi, “A new approach to observer-based
fault-tolerant controller design for Takagi-Sugeno fuzzy systems with
it is clear that the system states are stable with less oscillations state delay,” Circuits, systems, and signal processing, vol. 28, no. 5, pp.
and converge to an origin centered ball. 679–697, 2009.
[16] S. Bezzaoucha, B. Marx, D. Maquin, and J. Ragot, “Contribution to the
constrained output feedback control,” in American Control Conference,
VI. C ONCLUSIONS ACC, Washington, DC, USA, June 2013.
Considering the saturation nonlinearity, a nonlinear system
can be represented in an augmented T-S form. It is important
to note that, using the proposed representation of the actuator