Automatica 76 (2017) 243–250

Contents lists available at ScienceDirect

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

Brief paper

Switching LPV controller design under uncertain scheduling

parameters✩
Pan Zhao, Ryozo Nagamune
Department of Mechanical Engineering, The University of British Columbia, Vancouver, BC V6T 1Z4, Canada

article info abstract

Article history: This paper addresses the problem of designing output-feedback switching linear parameter-varying (LPV)
Received 19 November 2015 controllers under inexact measurement of scheduling parameters. The switching LPV controllers are
Received in revised form robustly designed so that the stability and L2 -gain performance of the switched closed-loop system
25 April 2016
can be guaranteed even under controller switching determined by measured (not actual) scheduling
Accepted 21 September 2016
Available online 8 December 2016
parameters. As for the switching rules, both hysteresis switching and average-dwell-time (ADT) switching
are considered. Solvability conditions for the controller design problem are expressed in terms of
Keywords:
parameter-dependent linear matrix inequalities and nonconvex switching surface conditions, with a line
Gain-scheduling search parameter. The nonconvex conditions can be convexified for hysteresis switching controller design
LPV systems by imposing practical validity constraints, and for ADT switching controller design by adding equality
Switching control constraints on some of Lyapunov variables on the switching surfaces. The effectiveness of the proposed
Parameter-dependent LMIs method is demonstrated with a numerical example.
Uncertainty © 2016 Elsevier Ltd. All rights reserved.

1. Introduction separated by switching surfaces, and one local LPV controller is

designed for each subset. Switching among local controllers is gen-
Since its advent over three decades ago (Shamma, 1988), linear erally implemented according to parameter-dependent switch-
parameter-varying (LPV) control has now been well-established ing rules such as hysteresis switching, average-dwell-time (ADT)
theoretically, and applied to various engineering problems (Ap- switching and smooth switching (Hanifzadegan & Nagamune,
karian & Adams, 1998; Apkarian, Gahinet, & Becker, 1995; Lu, 2014). State-dependent switching without online measurement
2012; Sato, 2011; Scherer, 2001; Shamma & Athans, 1991; Wu, of varying parameters is considered in Deaecto, Geromel, and
Yang, Packard, & Becker, 1996). In a conventional LPV control sys- Daafouz (2011). There is also some work for simultaneous design
tem, a single LPV controller is generally used to control an LPV of switching surfaces and a switching LPV controller (Javadian &
plant. However, when an LPV plant has a large range of param- Nagamune, 2014; Zhao & Nagamune, 2015).
eter variation, a single LPV controller often leads to conserva- In LPV control applications, exact values of scheduling parame-
tive control performance, or even worse, the problem of finding a ters may not be available for adaptation of controller parameters,
due to sensor drift and noise, or parameter estimation errors. Un-
stabilizing LPV controller becomes infeasible. To resolve the con-
certainties on scheduling parameters have been considered in LPV
servatism or infeasibility issue, one approach is to utilize the
controller design for both state-feedback cases (Heemels, Daafouz,
switching LPV controller (Lu & Wu, 2004; Lu, Wu, & Kim, 2006).
& Millerioux, 2010; Sato, Ebihara, & Peaucelle, 2010) and output-
In this approach, the parameter set is divided into several subsets,
feedback cases (Daafouz, Bernussou, & Geromel, 2008; Sato &
Peaucelle, 2013). Daafouz et al. (2008) dealt with multiplicative
uncertainties. However, the method cannot handle bias errors that
✩ Supported by the Natural Sciences and Engineering Research Council of are common in many measurement systems. A more general ap-
Canada Discovery Grant (NSERC RGPIN 341886-12) and Vanier Canada Graduate proach was proposed in Sato and Peaucelle (2013), which consid-
Scholarship, Canada Research Chairs, and the Institute for Computing, Information ered additive uncertainties. Coexistence of additive and multiplica-
and Cognitive Systems at the University of British Columbia. The material in this tive uncertainties were addressed in Agulhari, Tognetti, Oliveira,
paper was not presented at any conference. This paper was recommended for
and Peres (2013) and Sato (2015).
publication in revised form by Associate Editor Sophie Tarbouriech under the
direction of Editor Richard Middleton. Despite the aforementioned research results, there is no work
E-mail addresses: [email protected] (P. Zhao), [email protected] reported on design of switching LPV controllers incorporating
(R. Nagamune). measurement uncertainties in scheduling parameters. This work
http://dx.doi.org/10.1016/j.automatica.2016.10.026
0005-1098/© 2016 Elsevier Ltd. All rights reserved.
244 P. Zhao, R. Nagamune / Automatica 76 (2017) 243–250

is important because large parameter variation can exist in ap-

plications where significant uncertainties in scheduling parameter
measurement are unavoidable. The main purpose of this paper is
to accomplish this practically-important but still-missing work.
The main contribution of this paper is to propose a system-
atic approach to the design of output-feedback switching LPV con-
trollers with guaranteed L2 -gain performance, under measurement
uncertainties in scheduling parameters. In our scheme, both local
controller adaptation and switching among local controllers de-
pend on measured (not actual) scheduling parameters (denoted as
θ̂ hereafter). For robustness guarantee, the admissible regions of Fig. 1. Partition of Θ for a two-dimensional set, N = 6.
the actual scheduling parameters (denoted as θ hereafter) and θ̂
for local controllers are determined. Using the admissible regions, Remark 1. The assumption of constant B2 and C2 is needed
the design problem is formulated as an optimization problem in- for on-line calculation of controller matrices without the actual
volving parameter-dependent linear matrix inequalities (PDLMIs) scheduling parameter vector θ (see (20)). This assumption can
with a line search parameter. Other minor contributions of this pa- always be satisfied by pre-filtering of the control inputs u and/or
per include: post-filtering of the measured outputs y. See Apkarian et al. (1995,
• Convexification of originally nonconvex switching surfaces p. 1255) for details. As long as the bandwidth of filters is much
conditions for hysteresis switching by considering practical higher than the plant, the effects of the filters on the performance
validity constraints (i.e., the controller does not depend on the of the closed-loop system can be neglected.
derivative of scheduling parameters).
• Modification of ADT switching rule for the inexact measure- Additionally, the parameter vector θ and its rate of variation θ̇
ment case, and acquisition of convex switching surface condi- are supposed to be bounded by hyper-rectangular sets as:
tions for ADT switching that is not given in Lu and Wu (2004).
θ ∈ Θ, Θ := {θ ∈ Rs : θ i ≤ θi ≤ θ i , i ∈ Zs },
This paper is organized as follows. Section 2 gives preliminaries (2)
and formulates the switching LPV controller design problem
θ̇ ∈ Ωθ , Ωθ := {θ̇ ∈ Rs : ωi ≤ θ̇i ≤ ωi , i ∈ Zs }.
with uncertain scheduling parameters. Section 3 presents the For the main results of this paper, we consider additive uncertain-
constraints on switching surfaces for ADT switching. To restate ties in scheduling parameter measurement while necessary modi-
the problem in terms of PDLMIs with respect to θ and θ̂ , we fications of the results for multiplicative uncertainties are given in
explicitly show the admissible regions of (θ , θ̂ ) in Section 4. The the Appendix. With additive uncertainties, the actual scheduling
main theorems for design of switching LPV controllers against parameter θi is measured as θ̂i = θi +δi , where δi represents the un-
measurement uncertainties are given in Section 5. Section 6 certainty in the measurement of θi . The vectors θ̂ := [θ̂1 , . . . , θ̂s ]T
presents a numerical example to illustrate the efficacy of the and δ := [δ1 , . . . , δs ]T denote the measured scheduling param-
proposed method. eters and the uncertainties in the scheduling parameters, respec-
Notations in this paper are as follows. Zj denotes the integer tively. Another assumption is that the uncertainty δi is independent
set {1, 2, . . . , j}. Rn and S n denote the sets of n-dimensional real of each other, and its bound is independent of the actual scheduling
vectors, and n × n real symmetric matrices, respectively. In and 0 parameter θi . The bounds of uncertainties and their rate of varia-
denote an n × n identity matrix, and a zero matrix of appropriate tion can generally be estimated a priori, and are assumed to satisfy
dimension. For a square matrix X , ⟨X ⟩ is the shorthand notation
of X + X T , and diag(X1 , X2 ) denotes a matrix with matrices X1 δ ∈ ∆, ∆ := {δ ∈ Rs : δ i ≤ δi ≤ δ i , i ∈ Zs },
and X2 as diagonal blocks. For real symmetric matrices M and N, (3)
M > N (M ≥ N ) means that M − N is positive definite (positive δ̇ ∈ Ωδ , Ωδ := {δ̇ ∈ Rs : ν i ≤ δ̇i ≤ ν i , i ∈ Zs },
semidefinite). In symmetric matrices, ⋆ denotes an off-diagonal
where δ i < 0 < δ i is presumed without loss of generality. From
block induced by symmetry. ∅ denotes an empty set and int(A)
˙
(2) and (3), we have θ̂ = θ̇ + δ̇ ∈ Ωθ̂ , where
means the interior of a set A. For two sets A and B, A \ B and A ∩ B
denote the difference {x : x ∈ A and x ̸∈ B} and the intersection
{x : x ∈ A and x ∈ B}, respectively. For a polytope Θ , ver(Θ ) Ωθ̂ := {v ∈ Rs : ωi + ν i ≤ vi ≤ ωi + ν i , i ∈ Zs }. (4)
represents the set of vertices, and Θi represents the projected set
Suppose that the parameter set Θ is partitioned into N closed
of Θ onto the ith coordinate.
hyper-rectangular subsets {Θ (j) }j∈ZN by dividing Θi into Ni inter-
vals with overlapped segments between any adjacent intervals,
2. Problem setting
where N = Πis=1 Ni and Θ = ∪{Θ (j) }j∈ZN . See Fig. 1 for an
illustration. The subsets are separated by a family of switching
We consider the following LPV plant:
surfaces. Let S (j,k) denote the switching surfaces specifying the
ẋ = A(θ )x + B1 (θ )w + B2 u, one-directional move of θ̂ from subset Θ (j) to subset Θ (k) . Here we

z = C1 (θ )x + D11 (θ )w + D12 (θ )u, (1) define that two sets are adjacent if the interior of their intersection
y = C2 x + D21 (θ )w, is non-empty. Moreover, two sets Θ (j) and Θ (k) are defined to be
adjacent in θκ direction (for some κ ∈ Zs ) if they are adjacent and
where θ := [θ1 , . . . , θs ]T is the vector of scheduling parameters
their projected sets onto θκ are different, i.e.
which are measurable in real time, x(t ) ∈ Rn is the state vector,
u(t ) ∈ Rnu is the control input vector, w(t ) ∈ Rnw is the vector int(Θ (j) ∩ Θ (k) ) ̸= ∅, Θκ(j) ̸= Θκ(k) . (5)
of exogenous inputs, y(t ) ∈ Rny is the measured output vector,
and z (t ) ∈ Rnz is the vector of output signals related to the By this definition, in Fig. 1, Θ (1) is adjacent to Θ (2) in θ1 direction,
performance of the control system. The matrices in (1) are assumed to Θ (6) in θ2 direction and to Θ (5) in both θ1 and θ2 directions.
to have compatible dimensions and to be dependent on θ except For the LPV system (1), we consider a switching LPV controller
B2 and C2 . consisting of a family of full-order output-feedback LPV controllers
 P. Zhao, R. Nagamune / Automatica 76 (2017) 243–250 245

in the form of
(j) (j)
  
AK (θ̂ ) BK (θ̂ )
 
ẋK xK
= (j) (j)
, j ∈ ZN , (6)
u CK (θ̂ ) DK (θ̂ ) y
  
=:K (j) (θ̂)

where xK (t ) ∈ Rn is the controller state vector, and a local LPV

controller K (j) (θ̂ ) ∈ R(n+nu )×(n+ny ) takes charge of a specific pa- Fig. 2. Modification of nominal switching surfaces for ADT switching to
rameter subset Θ (j) . Note that controller matrices are functions of accommodate measurement uncertainties.
measured scheduling parameters θ̂ , not of actual scheduling pa-
rameters θ . Local controllers are switched according to switch- where Sκnom is the projection onto θκ of the nominal switching
ing rules such as hysteresis switching and ADT switching, when surface S nom between Θ (j) and Θ (k) , and Sκ(k,j) < Sκ(j,k) is assumed.
the trajectory of θ̂ hits one of the switching surfaces. A switch- See Fig. 2 for an illustration. Hereafter, σ denotes a switching signal
ing signal σ designates the active local controller, and is generated based on θ̂ and modified switching surfaces, while σ nom denotes a
based on θ̂ . Hysteresis switching occurs as follows: Let σ (0) = j switching signal based on θ and nominal switching surfaces. With
if θ̂(0) ∈ Θ (j) (pick an arbitrary j if there exist multiple j). For any the modified switching surfaces, we have the following lemma and
t > 0, if σ (t − ) = j (where σ (t − ) is the left limit of σ at time t) theorem.
and θ̂ (t ) ∈ Θ (j) , keep σ (t ) = j. On the other hand, if σ (t − ) = j, Lemma 3. Considering switchings between two adjacent subsets, for
θ̂(t − ) ∈ S (j,k) , and θ̂ (t ) ∈ Θ (k) \ Θ (j) , then let σ (t ) = k. any trajectories θ (·) and θ̂ (·) satisfying (3), there must be at least one
Under switching LPV control, the closed-loop system becomes switching that happens based on σ nom between any two successive
a switched LPV system, which can be described as switchings based on σ .
(σ ) (σ ) Proof. We prove this lemma by considering switchings between
ẋcl = Acl (θ , θ̂ )xcl + Bcl (θ , θ̂ )w,

(σ ) (σ ) (7) two subsets that are adjacent in only one direction, while the proof
z = Ccl (θ , θ̂ )xcl + Dcl (θ , θ̂ )w,
for the case of two subsets adjacent in multiple directions can
(σ ) (σ ) (σ )
where xcl := [xT xTK ]T ∈ R2n , and the expressions of Acl , Bcl , Ccl be derived analogously. We use the notations in Fig. 2. Assume
(σ ) a switching from σ = j to σ = k happens at t1 , which means
and Dcl in terms of plant and controller matrices can be readily
obtained. θ̂κ (t1 ) > Sκ(j,k) . Then we have
The controller design problem tackled in this paper is stated
θκ (t1 ) ≥ θ̂κ (t1 ) − δ κ > Sκ(j,k) − δ κ ≥ Sκnom ,
below.
where the first and third inequalities are from (3) and (9),
Problem 2. Given an LPV plant (1) with scheduling parameters respectively. This indicates that σ nom (t1 ) = k. Suppose next
θ satisfying (2) and the parameter subsets {Θ (j) }j∈ZN , as well as switching from σ = k to σ = j happens at t2 . Then we have
measurement uncertainties δ satisfying (3), find a switching LPV
controller (6) with a switching rule that enforces stability and
θ̂κ (t2 ) < Sκ(k,j) , and thus
minimizes the L2 -gain bound γ of the closed-loop system (7) for θκ (t2 ) ≤ θ̂κ (t1 ) − δ κ < Sκ(k,j) − δ κ ≤ Sκnom ,
all admissible trajectories θ (·) and δ(·).
where the assumptions in (3) and (9) are used again. This implies
3. Constraints on switching surfaces for ADT switching that σ nom (t2 ) = j. Comparing σ nom value at t1 and t2 , we conclude
that at least one switching from σ nom = k to σ nom = j happens in
For ADT switching without considering measurement uncer- the interval (t1 , t2 ]. 
tainties (Lu & Wu, 2004), there is only one switching surface, which
Theorem 4. For any trajectories θ (·) and θ̂ (·) satisfying (3), the ADT
we refer to as nominal switching surface, introduced for any two
τ of a switching signal σ is not less than the ADT τ nom of a switching
adjacent subsets. In this case, overlapped area of the two adjacent
signal σ nom , i.e. τ ≥ τ nom .
subsets is the switching surface, and has no interior in contrast to
our assumption (5). A switching signal σ is defined to have average Proof. Let Nσ nom (T , t ) and Nσ (T , t ) denote the number of switch-
dwell time (ADT) τ which is not less than τ0 > 0 if there exists a ings at time interval [t , T ] based on σ nom and σ , respectively. Ac-
finite positive number N0 such that cording to (8), there exists N0nom > 0 satisfying
T −t T −t
Nσ (T , t ) ≤ N0 + , ∀T ≥ t ≥ 0, (8) Nσ nom (T , t ) ≤ N0nom + , ∀T ≥ t ≥ 0. (10)
τ0 τ nom
where Nσ (T , t ) is the number of switchings on time interval [t , T ], If switching happens between only one pair of (i.e. two) adjacent
and N0 is called the chatter bound. subsets, then we have Nσ (T , t ) ≤ Nσ nom (T , t ) + 1 from Lemma 3. If
Without measurement uncertainties, the ADT (denoted by switching happens between m(≥ 1) pairs of adjacent subsets, then
τ nom ) of switching signals is estimated based on the trajectories we have Nσ (T , t ) ≤ Nσ nom (T , t ) + m, which with (10) will further
of actual scheduling parameters θ and nominal switching surfaces, yields
and used in controller design. However, when there are uncertain- T −t T −t
ties in scheduling parameter measurement, the uncertainties may Nσ (T , t ) ≤ N0nom + m + ≤ N0 + , ∀T ≥ t ≥ 0,
cause extra switchings, thereby making the ADT τ of the switch- τ nom τ nom
ing signals based on θ̂ smaller than the nominal value τ nom . As a where N0 := N0nom + m. Based on the definition of ADT, the
result, the controller designed based on τ nom may not work. To re- switching signal σ has ADT that is not less than τ nom . 
solve this issue, for subsets Θ (j) and Θ (k) that are adjacent in θκ
Remark 5. Theorem 4 implies that the ADT estimated based on
direction, we will introduce two switching surfaces S (j,k) and S (k,j)
nominal switching surfaces and actual scheduling parameters
satisfying
can be utilized as it is in switching LPV controller design when
Sκ(k,j) − δ κ ≤ Sκnom ≤ Sκ(j,k) − δ κ , (9) scheduling parameters are not exactly measurable.
246 P. Zhao, R. Nagamune / Automatica 76 (2017) 243–250

(j)
To explicitly show the shape of Φi , let us consider four specific
(j)
cases of Θi , illustrated in Fig. 3, and listed as follows:
(j)
Case 1: Θi = [θ i , Si(j,p) ] for some p ̸= j,
(j)
Case 2: Θi = [Si(j,p) , Si(j,q) ] for some p, q ̸= j,
(j)
Case 3: Θi = [Si(j,q) , θ i ] for some q ̸= j,
(j)
Case 4: Θi = [θ i , θ i ], i.e. Θi is not partitioned.
(j)
Fig. 3. Θi and admissible value of θi at switching. For Cases 1–3 shown in Fig. 3, the value of θi right before
switching from σ = j should be inside the shaded region. An
With two switching surfaces introduced for any two adjacent example of θi and θ̂i at switching is also given in Fig. 3. The
subsets, switching signal σ for ADT switching is generated in (j)
admissible region Φi is obtained, by utilizing (12), as polytopes
the same way as for hysteresis switching. However, as will be
depicted in Fig. 4.
explained in Section 5, the constraint on Lyapunov functions
at switching surfaces for ADT switching is different from that
for hysteresis switching. Under hysteresis switching, Lyapunov 5. Switching LPV controller design under uncertain scheduling
functions need to be monotonically decreasing at switching, while parameters
such requirement can be relaxed for ADT switching.
5.1. Hysteresis switching
4. Admissible region determination
We first give a lemma that can be obtained from Lu and Wu
To express the design problem in terms of LMIs with respect to (2004, Section 3.1). The proof is omitted for brevity.
θ and θ̂ , we need to determine the admissible region of (θ , θ̂ ) for
σ = j, which is denoted as Φ (j) hereafter. Since there is no essential Lemma 6. Suppose that there exists a family of parameter-dependent
difference in partition of Θ and σ generation between hysteresis positive-definite matrices Z (j) (θ̂ ) ∈ S 2n such that (13) holds for any
switching and ADT switching, we will not distinguish them for j ∈ ZN and (14) holds for any adjacent Θ (j) and Θ (k) . Then, with
determination of Φ (j) . The region Φ (j) consists of all possible pairs the switching LPV controller (6) with hysteresis switching rule, the
of (θ, θ̂) when the local controller K (j) (θ̂ ) is active. Note that when closed-loop system (7) is exponentially stable and its performance
K (j) (θ̂) is active, θ̂ should be in subset Θ (j) , or in other words, θ̂i ∥z ∥2 < γ ∥w∥2 is achieved with γ = maxj∈ZN {γ (j) } for all admissible
(j) trajectories θ (·) and δ(·).
should stay in Θi for all i ∈ Zs . Therefore Φ (j) can be obtained as
the Cartesian product: 
(j)

(j)

Acl (θ , θ̂ )Z (j) (θ̂ ) − Ż (j) (θ̂ ) ⋆ Bcl (θ , θ̂ )
(j)
Φ (j) = Φ1 × · · · × Φs(j) ,
Dcl (θ , θ̂ ) < 0,
(11) (j) (j)
Ccl (θ , θ̂ )Z (j) (θ̂ ) −γ (j) Inz
 

where Φi
(j)
is the projected set of Φ (j) onto the ith coordinate, and ⋆ ⋆ −γ (j) Inw
(j)
includes all admissible pairs of (θi , θ̂i ) to make θ̂ lie in subset Θ .
(j) ∀(θ , θ̂ , θ̂˙ ) ∈ Φ (j) × ver (Ωθ̂ ), (13)
The expression of Φi is
(j) (j)
Z (j) (θ̂ ) ≤ Z (k) (θ̂ ), ∀θ̂ ∈ S (j,k) . (14)
Φi := {(θi , θ̂i ) : θi + δi = θ̂i ∈ Θi , δi ∈ ∆i , θi ∈ Θi }. (12)

(a) Case 1. (b) Case 2.

(c) Case 3. (d) Case 4.

Fig. 4. Admissible region of (θi , θ̂i ) for σ = j.

 P. Zhao, R. Nagamune / Automatica 76 (2017) 243–250 247

The following theorem is obtained based on Lemma 6 and Sato and Table 1
Selection of Lyapunov variables and switching surface conditions for hysteresis
Peaucelle (2013, Theorem 1). For brevity, hereafter, we omit the
switching with practical validity.
dependence of plant matrices on θ , and the dependence of other
Variables X (j) , Y (j) Variables M (j) , N (j)
matrices on θ̂ . Switching surf. cond.

X (j) , Y0 X (j) − Y0 −1 , −Y0 X (j) ≤ X (k)

Theorem 7. Suppose that there exist a family of parameter- X0 , Y (j) −X0 , Y (j) − X0 −1 Y (j) ≥ Y (k)
dependent matrices X (j) ∈ S n and Y (j) ∈ S n , parameter-dependent X0 , Y0 any M (j) , N (j) from (21) –
(j) (j) (j) (j)
matrices ÂK , B̂K , ĈK and D̂K , and a positive scalar ε such that condi-
tions (15) and (16) hold for any j ∈ ZN , and one of (18) and (19) holds parameterized as
for any adjacent Θ (j) and Θ (k) . Then, with the switching LPV con-
Y (j) N (j) X (j) M (j)
   
−1
troller (6) obtained by (20), the closed-loop system (7) is exponen- P (j) = T , P (j) = T , (23)
tially stable, and its performance ∥z ∥2 < γ ∥w∥2 is achieved with N (j) ? M (j) ?
γ = maxj∈ZN {γ (j) } for all admissible trajectories θ (·) and δ(·).
where ‘?’ denotes irrelevant matrix blocks. The matrix inequality
[Subset conditions] (15) guarantees that Lyapunov matrices {P (j) }j∈ZN are positive
definite in their subsets (see Scherer, Gahinet, and Chilali (1997)
X (j)
 
In −1
> 0, ∀θ̂ ∈ Θ (j) , (15) for details). Setting Z (j) := P (j) , the inequality (16) leads to
In Y (j) (13) according to Sato and Peaucelle (2013), and (18) is equivalent
H (θ , θ̂ ) ⋆ to (14) through congruence transformation, when M (j) and N (j)
 
 (j) < 0, satisfying (21) are chosen to be X (j) − Y (j)
−1
and −Y (j) , respectively.
εX (A(θ ) − A(θ̂ ))T Y (j)

0 0 −ε In
(j) −1
If we choose M = −X and N = Y (j) − X (j) , (19) can
(j) (j)

∀(θ, θ̂ , θ̂˙ ) ∈ Φ (j) × ver (Ωθ̂ ), (16) be obtained as another equivalent condition for (14). Thus, from
Lemma 6, the statements are proved. 
where
Note that the constraints (18) and (19) are nonconvex. How-
⋆ diag(−Ẋ (j) , Ẏ (j) )
   
M11 0 ever, if we impose practical validity constraint, these nonconvex
H (θ, θ̂ ) := + , (17)
M21 M22 0 0 conditions can be simplified and become convex, as explained next.
To be practically valid, the switching LPV controller should not de-
⟨AX (j) + B2 ĈK(j) ⟩ ⋆
 
M11 := (j) (j) (j) , pend on the derivative of θ̂ . Therefore, as suggested in Apkarian
ÂK + [A + B2 D̂K C2 ]T ⟨Y (j) A + B̂K C2 ⟩ and Adams (1998), at least one of X (j) and Y (j) must be constant.
(j) (j) (j)
T T
When X (j) (Y (j) ) are selected to be constant for each j, all
 
[B1 + B2 D̂K D21 ] [Y B1 + B̂K D21 ]
M21 := ,
(j) (j) (j) X (j) (Y (j) ), j ∈ ZN , must be the same in order to satisfy condition
 C1 X + (j)
D12 ĈK C1 + D12 D̂K C2
(18) or (19). Take X (j) for example. Given constant matrices X (j) and
⋆

−γ Inw
M22 := (j) . X (k) for some j, k, either (18) or (19) indicates that X (j) = X (k) when
D11 + D12 D̂K D21 −γ (j) Inz
we consider the constraints at both S (j,k) and S (k,j) . After examining
[Switching surface conditions] the switching between any two adjacent subsets, we conclude that
all constant X (j) , j ∈ ZN have to be the same.
Y (j) ≥ Y (k) ,

Using the constant matrix, denoted by X0 , (18) and (19) reduce
−1 −1 ∀θ̂ ∈ S (j,k) , (18)
X (j) − Y (j) ≤ X (k) − Y (k) , to convex equality constraints

X (j) ≤ X (k) , Y (j) = Y (k) , ∀θ̂ ∈ S (j,k) ,


(24)
(j) (j) −1 (k) (k) −1 ∀θ̂ ∈ S (j,k) . (19)
Y −X ≥Y −X , and convex inequality constraints
[Controller reconstruction] Y (j) ≥ Y (k) , ∀θ̂ ∈ S (j,k) , (25)
 (j) (j) −1
 T respectively. Of course, (25) is less conservative than (24), and thus
 AK = N (j) ÂK − Ẏ (j) X (j) − Ṅ (j) M (j) should be chosen. Analogously, using a constant matrix Y0 for all

− Y (j) (A(θ̂ ) − B2 D̂K(j) C2 )X(j) Y (j) , we obtain convex switching surface conditions X (j) ≤ X (k) for




any θ̂ ∈ S (j,k) from (18).

− B̂(Kj) C2 X (j) − Y (j) B2 ĈK(j) M (j) ,

 −T
All allowable choices of X (j) and Y (j) for practical validity,

(j) −1

(j) (j)
 (20)

 BK = N (j) B̂K − Y (j) B2 D̂K , with associated switching surface conditions, are summarized in
Table 1.


(j) (j) (j)
  −T
C = ĈK − D̂K C2 X (j) M (j) ,



 K(j)


(j) 5.2. ADT switching
DK = D̂K ,

where N (j) and M (j) are from the factorization When hysteresis switching is used for switching controller
design, the stability and performance are guaranteed no matter
T
N (j) M (j) = I − Y (j) X (j) . (21) how frequently the switching events occur. However, if we
know that only a small number of switchings can happen in
average over a finite time interval, then ADT switching may yield
Proof. The parameter-dependent Lyapunov function (PDLF) for better performance than hysteresis switching. This is because
the closed-loop system (7) is selected to be a piecewise-continuous the Lyapunov function is no longer needed to be monotonically
function depending on θ̂ , defined as decreasing during switching under ADT switching logic. The
following theorem is obtained based on the results in Lu and Wu
V (σ ) (xcl , θ̂ ) := xTcl P (σ ) (θ̂ )xcl , (22) (2004, Theorem 3) and Sato and Peaucelle (2013, Theorem 8).

where {P (j) }j∈ZN is a family of positive definite parameter- Theorem 8. Given scalars λ0 > 0 and µ > 1, suppose there exist
dependent matrices. The matrix P (j) and its inverse can be a family of parameter-dependent matrices X (j) ∈ S n and Y (j) ∈ S n ,
248 P. Zhao, R. Nagamune / Automatica 76 (2017) 243–250

(j) (j) (j) (j)

parameter-dependent matrices ÂK , B̂K , ĈK and D̂K , and a positive becomes more restrictive compared to its counterpart (16) for
scalar ε such that convex conditions (15) and (27) hold for any j ∈ ZN , hysteresis switching due to addition of the extra block containing
and one of (28) and (29) holds for any adjacent Θ (j) and Θ (k) . Then, λ0 . The extra block is introduced to compensate possible increase
with the switching controller (6) obtained by (20), and by replacing of the Lyapunov function. From the perspective of (27), small λ0 is
(j) (j)
AK with AK + (λ0 /2)In , the closed-loop system (7) is exponentially desired, corresponding to small µ, while large µ is desired when
stable, and its performance ∥z ∥2 < γ ∥w∥2 is achieved with γ = considering switching surface conditions (28) and (29). Hence,
maxj∈ZN {γ (j) } for all admissible trajectories θ (·) and δ(·) and for there is a tradeoff between small λ0 and large µ. For fixed λ0 , as
every switching signal σ with ADT τ becomes larger, so does µ. Therefore, ADT switching is most
suitable for systems with large ADT.
τ > (ln µ)/λ0 . (26)
5.3. Solving PDLMIs
[Subset conditions] (15) and

λ0 X (j) ⋆ PDLMIs appear in Theorems 7 and 8, which will lead to

    
H (θ, θ̂ ) + diag λ0 , 0, 0 ⋆  an infinite-dimensional and infinitely constrained problem for
 In λ0 Y (j)  < 0, general parameter dependence of the plant. A standard approach
2
 (j) to the selection of finite-dimensional optimization functions is to
εX (A(θ ) − A(θ̂ ))T Y (j) 0 0

−ε In
imitate the parameter dependence of the plant in the optimization
(j) (j)
∀(θ, θ̂ , θ̂˙ ) ∈ Φ (j) × ver (Ωθ̂ ), (27) variables X (j) , Y (j) , ÂK , . . . , D̂K (Apkarian & Adams, 1998). In order
to obtain a finite number of LMIs, gridding technique generally
where H (θ , θ̂ ) is defined in (17). has to be employed for general parameter dependence of the plant
(Apkarian & Adams, 1998). Under special circumstances like affine
[Switching surface conditions]
 (j) or polynomial parameter dependence of the plant, techniques such
Y = Y (k) , as multiconvexity concepts (Apkarian & Tuan, 2000), coefficient
µX (k) − X (j) ⋆ ∀θ̂ ∈ S (j,k) , check and SOS relaxations (Henrion & Garulli, 2005), can be

(28)
(j) ≥ 0, employed to obtain a finite number of sufficient LMI conditions.
µ − 1In
 
Y
Since a scalar optimization variable ε is introduced in Theo-
rems 7 and 8, and coupled with matrix optimization variables X (j) ,
 (j)
X (k) ,
X =
µY (j) − Y (k) ⋆ ∀θ̂ ∈ S (j,k) . a line search for ε has to be implemented. Actually, to reduce the

(29)
(j) ≥ 0, conservatism, one can use different ε (j) for different subset Θ (j) . In
µ − 1In

X
this case, the line search will be time-consuming due to increased
computational complexity. Alternatively, noting that ε (j) are only
Proof. Here we just prove that both (28) and (29) are sufficient coupled with X (j) , one can alternately optimize, via an LMI solver,
−1 −1
conditions for µP (j) ≥ P (k) , or equivalently P (j) ≤ µP (k) , ε (j) or X (j) while fixing the other. For ADT switching controller de-
with the parameterization of P and P (j)
given in (23). The proof (j) −1 sign, due to coupling between µ and X (j) in (28) or Y (j) in (29), a
of (27) is straightforward from Lu and Wu (2004) and Sato and line search or alternative optimization is also needed for µ.
Peaucelle (2013). Schur complement to the inequality condition in
−1 6. A numerical example
(28) gives µX (k) − X (j) − (µ − 1)Y (j) ≥ 0, which, together with
the equality condition in (28) and µ > 1, leads to
In this section, we apply the proposed switching LPV control
synthesis technique to a numerical example1 with the following
   
(j) −1 (j) (j) −1 (k) −1 (k) (k) −1
diag Y ,X −Y ≤ µ diag Y ,X −Y .
state-space matrices (see (1)):
(30)
1 + θ12
 
0 0 0
≤ µP (j) −1 (k) −1 −10 (θ + 0.5θ1 + 1.25)
4 2
θ2 − 10 0 0
This inequality is equivalent to P through congru- A(θ ) =  1 ,

ence transformation, when M (j) and N (j) satisfying (21) are chosen −1 0 −1 0
to be X (j) − Y (j)
−1
and −Y (j) , respectively. If we choose M (j) = −X (j)
−8 0 0 −28.723
−1 T
and N (j) = Y (j) − X (j) , another equivalent nonconvex condition
T
, 106 ,
 
B1 = 0 0 1 8 B2 = 0 0 0
for µP (j) ≥ P (k) can be obtained analogously, which is guaranteed
−0.1 6.822 0.1
   
0 0
by the convex condition (29).  C1 = , D11 = ,
0 0 0 0 0
One can see that the nonconvex switching surface condition  
0
, 0 , D21 = 1,
 
(30) cannot be convexified by imposing practical validity con- D12 = C2 = −1 0 0
0.1
straints because constant matrices X (j) or Y (j) , j ∈ ZN do not need to
be the same due to the existence of µ. Therefore, the equality con- where θi ∈ Θi = [−2, 2] and |θ̇i | ≤ 2 for i = 1, 2. Assume that
straints X (j) = X (k) or Y (j) = Y (k) are made in order to attain convex θi is measured with uncertainty δi ∈ ∆i = [−ξ , ξ ], and its rate
conditions (28) or (29). Note that convex switching surface condi- |δ̇i | ≤ 20.
tions for ADT switching controller design are not available in Lu For designing switching LPV controllers, we partitioned Θi into
and Wu (2004). Depending on selection of constant X (j) or Y (j) , and (1) (2)
Θi = [−2, 0.2] and Θi = [−0.2, 2] for i = 1, 2. For simplicity,
(28) or (29), there are totally four conditions, each of which is suffi- (j) (j)
all the parameter-dependent matrices X (j) , Y (j) , ÂK , . . . , D̂K are
cient to guarantee the stability of the switched closed-loop system.
However, it is unclear which condition among the four conditions parameterized to be affine with respect to θ̂ . For instance, when
provides the least conservative controller before actual controller X (j) is chosen to be parameter-dependent, then it is parameterized
design and performance comparisons. as
(j) (j) (j)
Unlike hysteresis switching, the Lyapunov function does not X (j) (θ̂ ) = X0 + θ̂1 X1 + θ̂2 X2 .
need to be monotonically decreasing in ADT switching case.
Instead, it can increase by µ times compared to its value before
switching. However, the subset condition (27) for ADT switching 1 Matlab code is available at http://cel.mech.ubc.ca/software-files/.
 P. Zhao, R. Nagamune / Automatica 76 (2017) 243–250 249

Table 2
γ obtained using non-switching and hysteresis switching controllers (NS: non-
switching, S: switching, aeb := a × 10b ).

ξ X0 , Y0 X (j) (θ̂), Y0 X0 , Y (j) (θ̂)

NS S NS S NS S

0 0.1643 0.1642 0.1641 0.1622 0.1512 0.1464

0.01 5.376e4 4.235e3 2.207 1.819 2.095 1.716

(3.162) (10−5 ) (10−5 )

0.1 2.007e6 5.177e3 50.717 13.748 46.406 10.751

(3.162) (3.162 · 10−5 ) (3.162 · 10−5 ) (a) Scheduling parameter.

0.2 2.567e6 7.161e3 202.530 76.449 182.605 26.191

(1) (3.162 · 10−5 ) (3.162 · 10−5 )

Table 3
γ given by ADT switching controllers with X0 , Y (j) (θ̂) and ε = 3.162 · 10−5 .
ξ τ = 100 τ =1 τ = 0.01
0 0.1451 (1.618) 0.1451 (1.023) Infeasible
0.2 25.228 (63.241) 25.908 (2.559) Infeasible (b) Switching signal.

The software package LMI Lab (Gahinet, Nemirovskii, Laub, & Chi-
lali, 1994) and gridding technique were used to solve PDLMIs.
For gridding of the involved polytopes, we used a rather coarse
gridding that includes their vertices and centers. The results of γ
minimization for different uncertain bounds ξ and parameteriza-
tion of Lyapunov matrices are shown in Table 2, where X0 (Y0 )
means that X (j) (Y (j) ) are constant and the same for all subsets (see
Table 1). The line search for ε was conducted with 13 logarithmi-
cally equi-spaced points between 10−5 and 102 for non-switching
LPV controller design, and the obtained ε was used for switch-
ing LPV controller design as well. The ε obtained by line searches
for different cases is given in parentheses in Table 2. One can see
that the switching LPV controller always gives better γ -value than
the non-switching LPV controller, and the improvement brought
by switching became more significant when the uncertainties in
scheduling parameter measurement increased.
(c) Disturbance input and Channel 1 of z (t ).
All the above computations were conducted in Matlab 2015b on
a PC with Intel i5-3470 CPU and 16 GB RAM running Win7 64-bit
OS. After ε is fixed, the computation time for designing a switching
LPV controller was about 60 min when using a PDLF, and 6 min
when using constant Lyapunov functions. If we fix the scheduling
parameter θ2 to be zero, i.e., we drop the term θ2 in A matrix, then
the computational time for designing a switching LPV controller
decreased significantly to around 20 s when using a PDLF and 6 s
when using a constant Lyapunov function. These sharp decreases
of computation time are due to reduction of both optimization
variables and grid points for solving PDLMIs.
We also tried design of ADT switching controllers using X0 and
Y (j) (θ̂ ) as well as switching surface condition (29). The line search
for µ was implemented with 11 logarithmically equi-spaced points
between 100.01 and 102 . The γ -value obtained for different ADT
τ and uncertainty bounds ξ is displayed in Table 3, where the
corresponding µ value is given in the parentheses. Note that when
ξ = 0, the subsets Θi(1) = [−2, 0] and Θi(2) = [0, 2] for i = 1, 2 (d) Control input.
were used. In comparison to the last column of Table 2, one can
see that, for large τ , ADT switching control yielded a better γ - Fig. 5. Simulation results.
value than hysteresis switching control, while we could not find
a feasible ADT switching controller for small τ . K2 : ADT switching controller considering δ designed with τ = 1
using the proposed approach, and
Time-domain simulations were performed for |δi | < 0.2. For
K3 : Hysteresis switching controller ignoring δ designed using the
comparisons, we tested three controllers designed with X0 and
approach in Lu and Wu (2004).
Y (j) (θ̂ ), which are
Trajectories of the scheduling parameters are depicted in
K1 : Hysteresis switching controller considering δ designed using Fig. 5(a), where the actual scheduling parameters are measured
the proposed approach, with bias error and random noise. Fig. 5(b) illustrates the switching
250 P. Zhao, R. Nagamune / Automatica 76 (2017) 243–250

signal. Fig. 5(c) shows the trajectories of w and z1 while control Apkarian, P., & Tuan, H. D. (2000). Parameterized LMIs in control theory. SIAM
input is given in Fig. 5(d). It can be seen that K3 could not Journal on Control and Optimization, 38(4), 1241–1264.
Daafouz, J., Bernussou, J., & Geromel, J. C. (2008). On inexact LPV control design
even stabilize the plant, due to ignorance of uncertainties in the of continuous-time polytopic systems. IEEE Transactions on Automatic Control,
scheduling parameter measurement. On the other hand, both K1 53(7), 1674–1678.
and K2 achieved the closed-loop stability, and furthermore, yielded Deaecto, G. S., Geromel, J. C., & Daafouz, J. (2011). Switched state-feedback control
for continuous time-varying polytopic systems. International Journal of Control,
smaller z1 -value than K3 even before the signal z1 started diverging 84(9), 1500–1508.
at around 4 s. In Fig. 5(d), we can also see that there were jumps in Gahinet, P., Nemirovskii, A., Laub, A.J., & Chilali, M. (1994). The LMI control toolbox.
control input during switching instants. In Proceedings of IEEE conference on decision and control (pp. 2038–2041).
Hanifzadegan, M., & Nagamune, R. (2014). Smooth switching LPV controller design
for LPV systems. Automatica, 50(5), 1481–1488.
7. Conclusion Heemels, W. P. M. H., Daafouz, J., & Millerioux, G. (2010). Observer-based control
of discrete-time LPV systems with uncertain parameters. IEEE Transactions on
Automatic Control, 55(9), 2130–2135.
This paper presented an approach to switching LPV controller Henrion, D., & Garulli, A. (Eds.) (2005). Positive polynomials in control. Berlin:
design considering uncertainties in measurement of scheduling Springer Verlag.
Javadian, M., & Nagamune, R. (2014). A numerical optimization approach to
parameters. As a key step in our approach, admissible regions switching surface design for switching LPV control. In Proceedings of American
of actual and measured scheduling parameters for all local control conference (pp. 3784–3789).
controllers were obtained according to the switching rules Lu, B., & Wu, F. (2004). Switching LPV control designs using multiple parameter-
dependent Lyapunov functions. Automatica, 40(11), 1973–1980.
based on measured scheduling parameters. Convex conditions Lu, B., Wu, F., & Kim, S. (2006). Switching LPV control of an F-16 aircraft via
on switchings surfaces were gained for hysteresis switching as controller state reset. IEEE Transactions on Control Systems Technology, 14(2),
a result of practical validity constraint, and for ADT switching 267–277.
Mohammadpour, J., & Scherer, C. (Eds.) (2012). Control of linear parameter varying
owing to equality constraints on some of Lyapunov variables systems with applications. New York: Springer Verlag.
at switching surfaces. A numerical example demonstrated the Sato, M. (2011). Gain-scheduled output-feedback controllers depending solely
advantages of the proposed controller over a non-switching LPV on scheduling parameters via parameter-dependent Lyapunov functions.
Automatica, 47(12), 2786–2790.
controller considering measurement uncertainties and a switching Sato, M. (2015). Gain-scheduled output feedback controllers for discrete-time
LPV controller ignoring measurement uncertainties. It was also LPV systems using bounded inexact scheduling parameters. In Proceedings of
illustrated that ADT switching may give better performance than American control conference (pp. 730–735).
Sato, M., Ebihara, Y., & Peaucelle, D. (2010). Gain-scheduled state-feedback
hysteresis switching when ADT is large. Our future work includes controllers using inexactly measured scheduling parameters: H2 and H∞
the application of the proposed approach to practical examples. problems. In Proceedings of American control conference (pp. 3094–3099).
Sato, M., & Peaucelle, D. (2013). Gain-scheduled output-feedback controllers using
inexact scheduling parameters for continuous-time LPV systems. Automatica,
Appendix. Modifications for multiplicative uncertainty cases 49(4), 1019–1025.
Scherer, C. (2001). LPV control and full block multipliers. Automatica, 37(3),
When the scheduling parameter θi is large, multiplicative 361–375.
Scherer, C., Gahinet, P., & Chilali, M. (1997). Multiobjective output-feedback control
uncertainties may be a more realistic representation of the real via LMI optimization. IEEE Transactions on Automatic Control, 42(7), 896–911.
uncertainties than additive uncertainties. With multiplicative Shamma, J. S. (1988). Analysis and design of gain scheduled control systems. (Ph.D.
thesis), Massachusetts Institute of Technology.
uncertainty δim in θi , the measured scheduling parameter will be Shamma, J. S., & Athans, M. (1991). Guaranteed properties of gain scheduled control
m
θ̂i = (1 + δim )θi , where δim ∈ ∆m
i := [δ i , δ i ] with −1 < δ i ≤
m m for linear parameter-varying plants. Automatica, 27(3), 559–564.
m Wu, F., Yang, X. H., Packard, A., & Becker, G. (1996). Induced L2 -norm control for LPV
0 ≤ δ i < 1 presumed. The bounds for rate of variation of θ̂i can be system with bounded parameter variation rates. International Journal of Robust
estimated from sensor measurements or through the bounds of θ̇i and Nonlinear Control, 6(9–10), 983–998.
Zhao, P., & Nagamune, R. (2015). Optimal switching surface design for state-
and δ̇im if the latter is known. feedback switching LPV control. In Proceedings of American control conference
Comparing with additive uncertainties, the differences for (pp. 817–822).
dealing with multiplicative uncertainties are the constraints on
modified switching surfaces for ADT switching and the expressions
Pan Zhao is a Ph.D. candidate and a Vanier Scholar in the
of admissible regions for local controllers, while the LMIs in Department of Mechanical Engineering at the University
Theorems 7 and 8 are still the same. Under multiplicative of British Columbia, Vancouver, BC, Canada. He received
uncertainties, for ADT switching controller design, the conditions the B.S. and M.S. degrees from Beihang University, Beijing,
China, in 2009 and 2012, respectively. From 2012 to
(9) need to be modified to
2013, he was a modeling & simulation engineer in Hirain
Technologies, Beijing, China. His current research interests
   
Sκ(j,k) Sκ(j,k) Sκ(k,j) Sκ(k,j) include robust and gain-scheduling control theory, LMIs
max , nom
≤ Sκ ≤ min , .
1 + δm 1 + δm
m m and their applications to mechanical systems.
κ 1 + δκ κ 1 + δκ
(31)
By imposing the constraints (31), we can derive counterparts
of Lemma 3 and Theorem 4 for multiplicative uncertainties. Ryozo Nagamune received the B.S. and M.S. degrees
(j)
The expression of admissible region Φi in (12) for additive from the Department of Control Engineering, Osaka
University, Osaka, Japan, in 1995 and 1997, respectively,
uncertainties needs to be replaced by and the Ph.D. degree from the Division of Optimization
and Systems Theory, the Royal Institute of Technology,
(j) (j)
Φi := {(θi , θ̂i ) : (1 + δim )θi = θ̂i ∈ Θi , δim ∈ ∆m
i , θi ∈ Θi }. (32)
Stockholm, Sweden, in 2002. From 2003 to 2006, he was
a Postdoctoral Researcher at the Mittag-Leffler Institute,
the University of California at Berkeley, and the Royal
Institute of Technology. He has been with the Department
References of Mechanical Engineering, the University of British
Columbia, Vancouver, BC, Canada since August 2006, and
Agulhari, C.M., Tognetti, E.S., Oliveira, R.C.L.F., & Peres, P.L.D. (2013). H∞ dynamic he is currently an Associate Professor. His research interests include robust control
output feedback for LPV systems subject to inexactly measured scheduling theory and applications to mechanical systems. Dr. Nagamune was a recipient of
parameters. In Proceedings of American control conference (pp. 6060–6065). the IEEE CDC Best Student Paper Award in 2001 and of the Best Paper Award in 23rd
Apkarian, P., & Adams, R. J. (1998). Advanced gain-scheduling techniques for Canadian Congress of Applied Mechanics in 2011. He is a Chair of IEEE Vancouver
uncertain systems. IEEE Transactions on Control Systems Technology, 6(1), 21–32. Joint Chapter of Control Systems, Robotics and Automation, and Systems, Man, and
Apkarian, P., Gahinet, P., & Becker, G. (1995). Self-scheduled H∞ control of linear Cybernetics Societies. Dr. Nagamune is the recipient of the Canada Research Chair
parameter-varying systems: a design example. Automatica, 31(9), 1251–1261. (CRC), Tier 2, in Control Engineering.