Upload
rui-wang
View
213
Download
0
Embed Size (px)
Citation preview
Journal of Control Theory and Applications 1 (2006) 32–37
Non-fragile hybrid guaranteed cost control
for a class of uncertain switched linear systems
Rui WANG, Jun ZHAO
(School of Information Science and Engineering, Northeastern University, Shenyang Liaoning 110004, China)
Abstract: This paper focuses on the problem of non-fragile hybrid guaranteed cost control for a class of uncer-
tain switched linear systems. The controller gain to be designed is assumed to have additive gain variations. Based on
multiple-Lyapunov function technique, the design of non-fragile hybrid guaranteed cost controllers is derived to make the
corresponding closed-loop system asymptotically stable for all admissible uncertainties. Furthermore, a convex optimiza-
tion approach with LMIs constraints is introduced to select the optimal non-fragile guaranteed cost controllers. Finally, a
simulation example illustrates the effectiveness of the proposed approach.Keywords: Switched systems; Guaranteed cost control; Non-fragile control; Multiple-Lyapunov function
1 Introduction
In recent years, the switched control systems have beenattracting considerable attention in the control commu-nity [1∼7]. Basically, a switched system belongs to a spe-cial class of hybrid systems, which consist of a family ofcontinuous-time or discrete-time subsystems and a switch-ing law that specifies the switching between them. Suchcontrol systems appear in many applications, such as com-munication networks, switching power converters and manyother fields.
On the other hand, when controlling a real plant, it isalso desirable to design a control system which is not onlyasymptotically stable but also guarantees an adequate levelof performance index. One way to address the robust per-formance problem is the so-called guaranteed cost controlapproach first introduced by Chang and Peng [8]. Basedon the GCC design approach, many significant results havebeen proposed [9∼11]. Although these methods yield con-trollers that are robust with regard to system uncertainty,their robustness with regard to controller uncertainty has notbeen considered. That is to say, these design methods are allbased on a basic assumption that the control will be im-plemented exactly. However, in practice, controllers have acertain degree of errors due to finite word length in any dig-
ital system, the imprecision inherent in analogue systems,actuator degradations, and the requirement for additionaltuning of parameters in the final controller implementation,etc. These errors could destabilize the closed-loop system.In [12], Keel and Bhattacharyya have shown by a numberof examples that the controllers designed by using weightedH∞, μ and l1 synthesis techniques may be very fragile withrespect to errors in the controller coefficients. This brings aproblem: how to design a controller for a given plant suchthat the controller is insensitive to some amount of errorwith respect to its gain, i.e. the controller is non-fragileor resilient. Recently, there have been some efforts to dealwith the non-fragile controller design problem [13,14]. Asto switched systems, because continuous-time and discrete-time dynamics exist simultaneously, the study of the non-fragile control becomes more complicated and difficult. Upto now, to the best of our knowledge, no results of stabi-lization and GCC based on non-fragile control method forswitched systems have been available.
This paper studies the problem of non-fragile hybridguaranteed cost control (NHGCC) for a class of uncertainswitched linear systems. Additive controller gain perturba-tion is considered. By employing multiple-Lyapunov func-tion technique, the design of non-fragile hybrid guaranteed
Received 30 September 2005; revised 14 December 2005.
This work was supported by the National Natural Science Foundation of China (No.60274009, 60574013), and the Natural Science Foundation of
Liaoning Province(No.20032020).
R.WANG et al. / Journal of Control Theory and Applications 1 (2006) 32–37 33
cost controllers is given in terms of linear matrix inequal-ities (LMIs). The upper bound on the guaranteed cost isminimized by solving a convex optimization problem withLMIs constraints.
In this paper, ∗ denotes the symmetric block in one sym-
metric matrix. I denotes the identity matrix of appropriate
dimension. The trace of a matrix is denoted by tr(·).
2 Problem formulation
Consider uncertain switched linear systems described bythe state-space model of the form:
x(t) = (Aσ + ΔAσ)x(t) + (Bσ + ΔBσ)uσ(t)
x(0) = x0, (1)
where x(t) ∈ Rn is the state, σ(t) : [0, +∞) → M =
{1, 2, · · · , m} is a piecewise constant switching signal tobe designed, which usually depends on time t or state x,Aσ, Bσ are known real constant matrices of appropriatedimensions, ui ∈ R
qi is the control input, ΔAσ, ΔBσ arereal-valued matrices representing time-varying parameteruncertainties, and have the following form
[ΔAi, ΔBi] = NiF1i(t)[Ci, Di],
where Ci, Di, Ni are known constant matrices of appropri-ate dimensions, F1i(t) is an unknown matrix function satis-fying
FT1i(t)F1i(t) � I.
For convenience, we adopt the following notations [1] forswitched system (1). Let
Σ = { x0; (i0, t0), (i1, t1), · · · , (in, tn), · · · , |ik ∈ M, k ∈ N} (2)
denote a switching sequence with the initial state x0 and theinitial time t0 , where (ik, tk) means that the ikth subsystemis activated for t ∈ [tk, tk+1).
The cost function associated with switched system (1) isgiven by
J =� +∞
0[x(t)TQx(t) + uT
σ(t)Ruσ(t)]dt, (3)
where Q and R are positive definite weighted matrices.
The difficulty in dealing with fragility is that the con-troller parameters are part of the design. Several mod-ellings of controller uncertainties have been considered tocope with the difficulty. Either the additive uncertainty[13] or multiplicative form [14] enters the system. For thegiven uncertain switched system (1), the controllers actuallyimplemented is
ui(t) = (Ki + ΔKi)x(t), (4)
where Ki is controller gain to be designed, and ΔKi repre-sents controller gain variation. In this paper, the followinggain variation is considered:
ΔKi is of the additive variation:
ΔKi = HiF2iEi, (5)
where FT2i(t)F2i(t) � I, i ∈ M .
Definition 1 For the uncertain switched system (1), ifthere exist a state feedback control law u∗
i for each subsys-tem and a switching law σ(t), and a positive scalar J∗ suchthat for all admissible uncertainties, the closed-loop systemis asymptotically stable and the value of the cost function(3) satisfies J � J∗, then J∗ is said to be a non-fragilehybrid guaranteed cost and u∗
σ(t) is said to be a NHGCClaw.
Remark 1 When σ(t) ≡ i, the switched linear system(1) degenerates into a regular linear system and the NHGCCproblem becomes the standard non-fragile guaranteed costcontrol problem [13].
Remark 2 The NHGCC problem with controller gainvariation is different from the standard guaranteed cost con-trol problem with system uncertainty, because the controllergain perturbation ΔKi not only affects system matrices butalso enters the cost function J as defined in (3), namely,
J =� +∞
0[x(t)TQx(t) + (Ki + ΔKi)TR(Ki + ΔKi)]dt.
The objective of this paper is to design NHGCC law(4) such that switched system (1) satisfies NHGCC for alladmissible uncertainties.
3 Main results
In this section, we present solvability conditions and
design method for NHGCC problem.
Theorem 1 If there exist constants βij (either all non-
negative or all nonpositive), symmetric positive definite
matrices Pi and positive constants ε1i, ε2i, λi, (i ∈ M) such
that the following matrix inequalities⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣
Ωi ∗ ∗ ∗ ∗−λiB
Ti Pi −R−1 ∗ ∗ ∗
NTi Pi 0 −ε−1
1i I ∗ ∗Ci − λiDiB
Ti Pi ε2iDiHiH
Ti 0 −ε1iI ∗
HTi BT
i Pi HTi 0 HT
i DTi −ε−1
2i I
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦
< 0 (6)
hold, where
34 R.WANG et al. / Journal of Control Theory and Applications 1 (2006) 32–37
Ωi = Ξi +m∑
j=1
βij(Pj − Pi),
Ξi = Q + PiAi + ATi Pi − 2λiPiBiB
Ti Pi + ε−1
2i ETi Ei,
then there exists NHGCC law (4) with control gain vari-
ation ΔKi as given in (5), such that the system (1) satis-
fies NHGCC, where the control gain is Ki = −λiBTi Pi.
When βij � 0, the switching law is given by σ(t) =
arg min{xTPix, i ∈ M}, and the performance upper
bound is
J∗ = min{xT0 Pix0, i ∈ M}; (7)
when βij � 0, the switching law is designed as σ(t) =
arg max{xTPix, i ∈ M}, and the performance upper
bound is
J∗ = max{xT0 Pix0, i ∈ M}. (8)
Proof Without loss of generality, suppose that βij � 0.
Obviously, ∀x ∈ Rn\ {0}, there exists an i ∈ M such that
xT(Pj − Pi)x � 0,∀i ∈ M . Thus, according to the matrix
inequalities (6), we have⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣
Ξi ∗ ∗ ∗ ∗−λiB
Ti Pi −R−1 ∗ ∗ ∗
NTi Pi 0 −ε−1
1i I ∗ ∗Ci − λiDiB
Ti Pi ε2iDiHiH
Ti 0 −ε1iI ∗
HTi BT
i Pi HTi 0 HT
i DTi −ε−1
2i I
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦
< 0. (9)
For any i ∈ M , let
Ωi = {x ∈ Rn|xT(Pj − Pi)x � 0,∀j ∈ M},
thenm⋃
i=1
Ωi = Rn\{0}. Construct the sets
Ω1 = Ω1, · · · , Ωi = Ωi −i−1⋃j=1
Ωj , · · ·, Ωm = Ωm −m−1⋃j=1
Ωj .
Obviously, we havem⋃
i=1
Ωi = Rn\{0} and Ωi
⋂Ωj =
φ, i = j.
The matrix inequalities (9)are equivalent to the followinginequalities. (Proof see Appendix)⎡
⎣ Σi (Ki + ΔKi)T
Ki + ΔKi −R−1
⎤⎦ < 0, (10)
where
Σi = [(Ai + ΔAi) + (Bi + ΔBi)(Ki + ΔKi)]TPi
+Pi[(Ai + ΔAi) + (Bi + ΔBi)(Ki + ΔKi)] + Q.
By the Schur complement Lemma, (10) ⇔
[(Ai + ΔAi) + (Bi + ΔBi)(Ki + ΔKi)]TPi
+Pi[(Ai + ΔAi) + (Bi + ΔBi)(Ki + ΔKi)] + Q
+(Ki + ΔKi)TR(Ki + ΔKi) < 0. (11)
Define the Lyapunov function candidate as
Vi(x(t)) = xTPix,
where Pi is positive definite matrix satisfying (6). Switchinglaw is designed as follows
σ(t) = arg min{xTPix, i ∈ M}.
When x(t) ∈ Ωi, from (11), the derivative of Vi(x) alongthe trajectory of the closed-loop system (1) is
Vi(x) < −xT[Q + (Ki + ΔKi)TR(Ki + ΔKi)]x
< 0. (12)
According to multiple-Lyapunov function technique, thesystem (1) is asymptotically stable.
In the following, we show that the closed-loop systemsatisfies the performance upper bound. Without loss of gen-erality, suppose xT
0 Pi0x0 = minik∈M
{xT0 Pik
x0}. According to
the switching sequence (2), we obtain
J =� +∞
0[x(t)TQx(t) + uT
σ(t)Ruσ(t)]dt
=m∑
i=1
∞∑j=1
� tij+1
tij[x(t)TQx(t) + uT
i Rui + Vi(x(t))]dt
− ∞∑k=0
� tk+1
tkVik
(x(t))dt
<− ∞∑k=0
� tk+1
tkVik
(x(t))dt
=− (Vi0(x(t1)) − Vi0(x(t0)) + Vi1(x(t2)) + · · ·)= Vi0(x(t0)) = xT
0 Pi0x0 = mini∈M
{xT0 Pix0}.
From Definition 1, we know (7) is a performance upperbound of system (1).
When βij � 0, similar proof can be given with the per-formance upper bound (8).
Remark 3 Pre- and post-multiplying both sides ofinequalities (6) by diag{Xi, I, I, I, I} with Xi = P−1
i ,and applying Schur complement Lemma yield the follow-ing LMIs ⎡
⎣Γi ΨTi
Ψi Υi
⎤⎦ < 0, (13)
R.WANG et al. / Journal of Control Theory and Applications 1 (2006) 32–37 35
where
Γi =⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣
Λi ∗ ∗ ∗ ∗−λiB
Ti −R−1 ∗ ∗ ∗
NTi 0 −ε−1
1i I ∗ ∗CiXi − λiDiB
Ti ε2iDiHiH
Ti 0 −ε1iI ∗
HTi BT
i HTi 0 HT
i DTi −ε−1
2i I
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦
Λi = AiXi + XiATi − 2λiBiB
Ti −
m∑j=1
βijXi ,
Ψi =[Xi EiXi Xi Xi Xi
]T [I 0 0 · · · 0
],
Υi = diag{−Q−1,−ε2iI,−β−1i1 P−1
1 , · · · − β−1i,i−1P
−1i−1,
− β−1i,i+1P
−1i+1, · · · ,−β−1
imP−1m }.
The cost upper bound (7) depends on feasible solution ofLMIs (13). The different feasible solutions of LMIs (13) canresult in different cost upper bounds. Theorem 1 provides asufficient condition for the solution to the NHGCC prob-lem. But it remains unclear as how to optimize the matrixPi in order to achieve the minimal guaranteed cost of theclosed-loop system. This problem is solved in the followingtheorem.
Theorem 2 For the given switched system (1) and per-formance index (3), if the following optimization problems
minX1,X2,··· ,Xm,Mi
tr(Mi)
s. t. i) (14),
ii)
⎡⎣ Mi xT
0
x0 Xi
⎤⎦ > 0, i ∈ M (14)
have solutions (X1, M1), (X2, M2), · · · , (Xm, Mm). Thenthere exists optimal NHGCC law u∗
i (t) = (Ki +ΔKi)x(t),and the minimal cost upper bound is
J∗ = mini∈M
xT0 X−1
i x0,
where Ki = −λiBTi X−1
i .
Proof If (X1, M1), (X2, M2), · · · , (Xm, Mm) aresolutions of the optimization problems (14), then they arealso feasible solutions of constraint condition i). Accordingto Theorem 1 and Remark 3, u∗
i (t) = (Ki + ΔKi)x(t) is aNHGCC law for switched system (1).
In addition, by Schur complement Lemma, constraint
condition ii) is equivalent to Mi > xT0 X−1
i x0. Thus, the
minimization of tr(Mi) implies the minimization of the
tr(xT0 X−1
i x0). Therefore, the minimal cost upper bound is
J∗ = mini∈M
xT0 X−1
i x0. The optimality of the solution of the
optimization problem (14) follows from the convexity of the
objective function and of the constraints. This completes the
proof.
4 Numerical example
Consider the following uncertain switched linear system
x(t) = (Ai + ΔAi)x(t) + (Bi + ΔBi)ui,
x(0) =[−1 2
]T
, i = 1, 2, (15)
where
A1 =
⎡⎣−5 0
−3 1
⎤⎦ , A2 =
⎡⎣ 1 0
2 −2
⎤⎦ , B1 =
⎡⎣ 1 1
−2 −1
⎤⎦ ,
B2 =
⎡⎣−1 2
1 1
⎤⎦ , C1 =
⎡⎣ 0.2 0.5
0.1 0
⎤⎦ , C2 =
⎡⎣ 0 0.2
0.4 0
⎤⎦ ,
Ni =
⎡⎣ 0 0.3
0.5 0
⎤⎦ , Di =
⎡⎣ 0.1 0
0 0.1
⎤⎦ .
We assume that controller gain has additive variation,
namely, ΔKi = HiF2iEi, where Hi = 0.5,
Ei =
⎡⎣ 0.1 0
0 0.1
⎤⎦ , F1i(t) = F2i(t) = sin t, (i = 1, 2).
Consider positive definite weighted matrices
Q = R =
⎡⎣ 0.5 0
0 0.5
⎤⎦ .
It is easy to see that (Ai, Bi) are not controllable, namely,
the two subsystems of system (15) are unstable. Let β12 =
1.5, β21 = 1, λi = 0.5, ε1i = ε2i = 1, (i = 1, 2). Now,
we apply Theorem 1 and Theorem 2 to design the optimal
state feedback controllers. By solving optimization problem
(14), we have
P1opt =
⎡⎣ 1.4780 0.4057
0.4057 1.2206
⎤⎦ , P2opt =
⎡⎣ 19.0822 9.3902
9.3902 5.2374
⎤⎦ .
Let
Ω1 = {x ∈ Rn|xT(P2opt − P1opt)x � 0, x = 0},
Ω2 = {x ∈ Rn|xT(P1opt − P2opt)x � 0, x = 0},
then Ω1
⋃Ω2 = R
n\ {0}. The switching law is designed
by
σ(t) =
{1, x(t) ∈ Ω1,
2, x(t) ∈ Ω2\Ω1.
36 R.WANG et al. / Journal of Control Theory and Applications 1 (2006) 32–37
The optimal state feedback controllers are designed as
u∗i (t) = (Ki+ΔKi)x(t), Ki = −λiB
Ti Piopt, i = 1, 2. It is
easy to see from Fig. 1 that the system (15) is asymptotically
stable under the optimal state feedback controllers. The
performance upper bound is J∗ = 4.7373 without using
optimization problem (14), while it is J∗ = xT0 P2optx0 =
2.4710 after optimization problem (14) is applied.
Fig. 1 The state responses of switched system (15).
5 Conclusions
In this article, based on the multiple-Lyapunov function
technique, we have investigated a design method to the
optimal NHGCC problem for a class of switched linear sys-
tems via state feedback controllers with additive variation.
Furthermore, a convex optimization problem with LMIs
constraints is formulated to design the optimal guaranteed
cost controller which minimizes the guaranteed cost of the
closed-loop uncertain system.
References
[1] M. S. Branicky. Multiple Lyapunov functions and other analysis
tools for switched and hybrid systems[J]. IEEE Trans. on Automatic
Control, 1998, 43(4): 475-482.
[2] Z. D. Sun, S. S. Ge. Switched Linear Systems-Control and Design[M].
New York: Springer-Verlag, 2004.
[3] J. Zhao, M. W. Spong. Hybrid control for global stabilization of the
cart-pendulum system[J]. Automatica, 2001, 37(12): 1941-1951.
[4] J. Zhao, G. M. Dimirovski. Quadratic stability of a class of switched
nonlinear systems[J]. IEEE Trans. on Automatic Control, 2004, 49(4):
574-578.
[5] G. M. Xie, L. Wang. Necessary and sufficient conditions for
controllability and observability of switched impulsive control
systems[J]. IEEE Trans. on Automatic Control, 2004, 49(6): 960-966.
[6] D. Z. Cheng. Controllability of switched bilinear systems[J]. IEEE
Trans. on Automatic Control, 2005, 50(4): 511-515.
[7] S. S. L. Chang, T. K. C. Peng. Adaptive guaranteed cost control
of systems with uncertain parameters[J]. IEEE Trans. on Automatic
Control, 1972, 17(4): 474-483.
[8] L. Yu, J. Chu. An LMI approach to guaranteed cost control of linear
uncertain time-delay systems[J]. Automatica, 1999, 35(6): 1155-
1159.
[9] W. H. Chen, J. X. Xu, Z. H. Guan. Guaranteed cost control
for uncertain markovian jump systems with mode-dependent time-
delays[J]. IEEE Trans. on Automatic Control, 2003, 48(12): 2270-
2277.
[10] L. Yu, F. Gao. Optimal guaranteed cost control of discrete-time
uncertain systems with both state and input delays[J]. J. of the
Franklin Institute, 2001, 338(1): 101-110.
[11] Keel, Bhattacharyya. Robust, fragile, or optimal[J]. IEEE Trans. on
Robotics Automatic, 1997, 9(4): 423-431.
[12] J. S. Yee, G. H. Yang, J. L. Wang. Non-fragile guaranteed cost for
discrete-time uncertain linear systems[J]. Int. J. of Systems Science,
2001, 32(7): 845-853.
[13] W. M. Haddad, J. R. Corrado. Robust resilient dynamic controllers for
systems with parametric uncertainty and controller gain variations[C]
// In Proc. of the American Control Conf.. Philadephia, PA, 1998,
2837-2841.
[14] J. H. Park, H. Y. Jung. On the design of non-fragile guaranteed
cost controller for a class of uncertain dynamic systems with state
delays[J]. Applied Mathematics and Computation, 2004, 150(1): 245-
257.
Appendix
Equivalence of (9) and (10).
Proof (9)⇔26666666666664
Θi + ε−12i ET
i Ei ∗ ∗ ∗ ∗
Ki −R−1 ∗ ∗ ∗
NTi Pi 0 −ε−1
1i I ∗ ∗
Ci + DiKi ε2iDiHiHTi 0 −ε1iI ∗
HTi BT
i Pi HTi 0 HT
i DTi −ε−1
2i I
37777777777775
< 0, (a1)
where
Θi = (Ai + BiKi)TPi + Pi(Ai + BiKi) + Q.
R.WANG et al. / Journal of Control Theory and Applications 1 (2006) 32–37 37
By Schur complement Lemma, (a1) ⇔26666666664
Φi ∗ ∗ ∗
Ki + ΔKi −R−1 ∗ ∗
NTi Pi 0 −ε−1
1i I ∗
Ci + Di(Ki + ΔKi) 0 0 −ε1iI
37777777775
< 0, (a2)
where
Φi = [Ai + Bi(Ki + ΔKi)]TPi + Pi[Ai + Bi(Ki + ΔKi)]
+Q.
According to Schur complement Lemma, (a2) ⇔264
Wi ∗
Ki + ΔKi −R−1
375 < 0 ⇔
264
Vi ∗
Ki + ΔKi −R−1
375 < 0 ⇔
264
Σi (Ki + ΔKi)T
Ki + ΔKi −R−1
375 < 0.
where
Wi = Φi + ε1iPiNiNTi Pi
+ε−11i [Ci + Di(Ki + ΔKi)]
T[Ci + Di(Ki + ΔKi)],
Vi = Φi + ΔATi Pi + PiΔAi
+(Ki + ΔKi)T ΔBT
i Pi + PiΔBi(Ki + ΔKi).
Rui WANG received the B.E. and
M.E. degrees in mathematics in 2001
and 2004, respectively, both from
Bohai University. She is currently a
Ph.D. candidate in Control Theory and
Applications at Northeastern Univer-
sity. Her main research interests include
switched systems, fault-tolerant con-
trol. E-mail: [email protected].
Jun ZHAO received the Ph.D in Con-
trol Theory and Applications in 1991 at
Northeastern University, China. From
1992 to 1993 he was a postdoctoral fel-
low at the same University. Since 1994
he has been with School of Informa-
tion Science and Engineering, North-
eastern University, China, where he is
currently a professor. From February
1998 to February 1999, he was a visiting scholar at the Coordinated Sci-
ence Laboratory, University of Illinois at Urbana-Champaign. He has held
a Research Fellow position at Department of Electronic Engineering, City
University of Hong Kong. His main research interests include switched
systems, nonlinear systems, geometric control theory, and robust control.
E-mail: [email protected].