Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 21, NO.

3, JUNE 2016 1659

H∞ Observer Design for LPV Systems With


Uncertain Measurements on Scheduling
Variables: Application to an Electric
Ground Vehicle
Hui Zhang, Member, IEEE, Guoguang Zhang, and Junmin Wang, Senior Member, IEEE

Abstract—In this paper, we aim to study the observer de- exogenous parameters are available, an LPV system can be re-
sign problem for polytopic linear-parameter-varying (LPV) formulated into a convex combination of linear-time-invariant
systems with uncertain measurements on scheduling vari- (LTI) systems and the well-developed results on LTI systems
ables. Due to the uncertain measurements, the uncertain-
ties are considered in the weighting factors. It is assumed can be extended to the LPV systems. If the nonlinearities are
that the vertices of polytope are the same when the determined with exogenous nonstationary parameters, the non-
measurements on scheduling variables are uncertain and linear system can be transferred into an LPV system and the
perfect. Then, an LPV system with the uncertain weight- analysis and synthesis of the nonlinear system are available to
ing factors can be transferred to an LPV system with un- be carried out under the framework of linear systems. Due to
certainties. To deal with the uncertainties and unknown
disturbance in the observer design problem, we propose the significant advantages, the research on LPV systems have
a gain-scheduling sliding mode observer. Defining the esti- gained numerous attention from both theoretical development
mation error as the state vector minus the estimated state and application sides. The authors in [7] investigated the model
vector, the estimation error dynamics is established. The predictive control (MPC) for LPV systems with bounded rates
sliding mode observer design method is developed based of parameter changes. The theory of MPC control was extended
on analysis results of the established estimation error sys-
tem. The proposed observer design method is then applied to LPV systems. The dynamic output feedback controller de-
to an electric ground vehicle (EGV) in which the measure- sign for discrete-time LPV systems was presented in [8]. The
ment of longitudinal velocity is assumed to be uncertain. uncertain issue of LPV systems was discussed in [9] and the
Experimental tests and comparisons are given to show the observer-based controller was designed for discrete-time LPV
advantages of the proposed design method and the de-
system. The finite-time stability for LPV systems was discussed
signed observer.
in [10]. At the application side, the LPV techniques have been
Index Terms—Electric ground vehicle, H∞ filtering, successfully applied to a number of practical setups, to name
linear-parameter-varying system, sliding mode observer, a few, such as polymer electrolyte membrane (PEM) fuel cell
uncertain measurements.
systems in [11], Diesel engines in [12], inverted pendulums in
[13], and wind turbine systems in [14].
I. INTRODUCTION When the exogenous nonstationary parameters are measur-
able online, it is desired that the controllers and observers for
HE last two decades have witnessed the increasing at-
T tention to linear-parameter-varying (LPV) systems; see
[1]–[6] and the references therein. The main reason arises from
LPV systems are also dependent on the measurable parameters.
In this way, the conservativeness of the design can be reduced.
This is the idea of gain-scheduling approach. In most of gain-
the fact the LPV technique is very useful to deal with smooth
scheduling controller or observer design work, it is generally
nonlinearities. In LPV systems, the state-space models depend
assumed that the measurements on the scheduling variables are
on exogenous nonstationary parameters. When the bounds of
perfect and the controller or observer directly employs the mea-
sured scheduling variables. However, in many practical appli-
Manuscript received April 5, 2015; revised December 28, 2015; ac-
cepted January 25, 2016. Date of publication January 27, 2016; date
cations, the measurements on the scheduling variables cannot
of current version April 28, 2016. Recommended by Technical Editor M. be so ideal. For example, the vehicle longitudinal velocity is
Basin. The work of H. Zhang was supported in part by The National quite important for vehicle control and diagnosis. For engine-
Science Foundation of China under Grant 61403252.
H. Zhang is with the School of Transportation Science and
powered vehicles, the longitudinal velocity is estimated with
Engineering, Beihang University, Beijing 100191 China (e-mail: the engine speed. Without the vehicle slip, this approach may
huizhang285@gmail.com). be accurate. However, the vehicle should maintain the slip ratio
G. Zhang and J. Wang are with the Department of Mechani-
cal and Aerospace Engineering, The Ohio State University, Colum-
at a certain level to fulfill a good road cohesive condition. There-
bus, OH 43210 USA (e-mail: zhang.3682@buckeyemail.osu.edu; fore, the estimated vehicle longitudinal velocity is inaccurate.
wang.1381@osu.edu). If the inaccurate or uncertain measurements of the scheduling
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
variables are directly taken as the ideal measurements, the per-
Digital Object Identifier 10.1109/TMECH.2016.2522759 formance of the designed controllers and observers cannot be

1083-4435 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
1660 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 21, NO. 3, JUNE 2016

guaranteed. Recently, there are some research works on the un- scheduling variables are known and the scheduling variables are
certain measurements of LPV scheduling variables. In [9], the measurable, the number of polytope vertices to describe the sys-
uncertain-measurement-induced issue was regarded as a distur- tem is 2r and the system matrix set S = (A(α), B1 (α), B2 (α))
bance. However, when the scheduling variables lie in the system expressed as
matrix and the uncertainties are large, definitely, the stability  
 
2r 
2r
will be degraded and the disturbance method cannot be applied. 
S = ΩΩ = αi Ωi ; 0 ≤ αi ≤ 1; αi = 1 (2)
In [15] and [16], the difference between the weighting factors i=1 i=1
calculated with ideal measurements and the estimated values
satisfies with a Lipschitz condition. where Ωi = (Ai , B1,i , B2,i ) and the value of matrix set for each
The main principle is to design an observer to estimate the vertex is known. Since the scheduling variables are measurable
unavailable states [17]–[22]. In the literature, there are a lot of online, the value of weighting factor αi for each vertex can
well-known types of observer such as the Kalman filter [23], be determined online. To formulate the problem, we make the
[24], the Luenberger observer [25], [26], the H∞ filter [27], following assumptions.
[28], the H2 filter [29], [30], the mixed H2 /H∞ filter [31], [32],
and the sliding mode observer [33], [34]. Each type of observer Assumption 1: The matrix Ai is a Hurwitz matrix and each
has its own advantages and application fields. The H∞ filters subsystem is observable.
are robust to system uncertainties. However, the uncertainties Assumption 2: The symbol u denotes the control input that is
should be well studied or modeled before the filters design. The known and the symbol ω is the disturbance that is unknown
sliding model observers or controllers are also capable to deal but bounded as
with system uncertainties and only the bounds of uncertainties ω ≤ δ1 (3)
are necessary [35], [36].
In this study, we aim to study the observer design problem for where δ1 is a known positive scalar.
LPV systems with uncertain measurements on scheduling vari- Assumption 3: Though the measurements on the scheduling
ables. The contributions of this study can be summarized as: 1) variables may not be accurate, the vertices of the describ-
we employ a different method to describe the LPV systems with ing polytope do not change, that is, we use one polytope to
uncertain measurements on scheduling variables; 2) we propose describe the system.
robust H∞ sliding mode observer for LPV systems subject to the
exogenous disturbances, unknown inputs, and uncertainties on In this study, we dedicate to develop a gain-scheduling sliding
scheduling variables; 3) slack matrices are introduced to reduce mode observer for the LPV system. If the scheduling variables
the conservativeness of obtained condition; and 4) the proposed are perfectly measurable, the determined weighting factors can
design method is successfully applied to an EGV to estimate the be directly used in the gain-scheduling observer design. How-
sideslip angle. ever, each physical sensor has its own resolution and the mea-
Notation: The notations used in this paper are standard. surement cannot be totally perfect. If the weighting factors of the
Superscript “T” and “−1” indicate matrix transposition and original system in (1) are directly used in the gain-scheduling
matrix inverse; in symmetric block matrices or long ma- observer, the mismatched weighting factors may deteriorate
trix expressions, we use ∗ as an ellipsis for the terms that the designed observer. Therefore, it is necessary to consider
are introduced by symmetry; and X represents the Eu- the asynchronous weighting factors. Consider a generalized case
clidean norm for vectors and the spectral norm for matri- for the scheduling variables with uncertain measurements as
ces. Matrices, if their dimensions are not explicitly stated, k i α̂i ≤ αi = ki α̂i ≤ k̄i α̂i ∀ i = 1, . . . , 2r (4)
are assumed to be compatible for algebraic operations. For
a real symmetric matrix P , the notations P > 0, and P < where α̂i denotes the weighting factor for the ith vertex calcu-
0, mean that P is positive definite, and negative definite, lated with the imprecisely measured scheduling variables, ki is
respectively. the uncertain factor, and k i and k̄i are known bounds of ki . With
the expression in (4), one gets
II. PROBLEM FORMULATION AND PRELIMINARY
 
r r
2 2
Consider a class of the LPV system with the following ex- A(α) = αi Ai = α̂i (Ai + (ki − 1)Ai )
pression: i=1 i=1

 
r r
2 2
ẋ = A(α)x + B1 (α)u + B2 (α)ω
= α̂i (Ai + ΔAi ) , α̂ = 1
y = Cx + Dv (1) i=1 i=1

 
r r
where x is the system state vector, u is the input vector, ω is 2 2

the disturbance, y is the measured output vector, and v denotes B1 (α) = αi B1,i = α̂i (B1,i + (ki − 1)B1,i )
the measurement noise. A(α), B1 (α), B2 (α), C, and D are i=1 i=1


r
matrices with compatible dimensions. Suppose that the number 2
of scheduling variables ρl (l = 1, . . . , r) is r and the scheduling = α̂i (B1,i + ΔB1,i )
variable is independent with each other. Then, if the bounds of i=1
ZHANG et al.: H∞ OBSERVER DESIGN FOR LPV SYSTEMS WITH UNCERTAIN MEASUREMENTS ON SCHEDULING VARIABLES 1661

 
r r
2 2
B2 (α)ω = α̂i ki B2,i ω = α̂i B2,i ω1
i=1 i=1

ΔAi = (ki − 1)Ai , ΔB1,i = (ki − 1)B1,i , ω1 = ki ω.


(5)
Here, X(α̂) represents a matrix determined with the measured
Fig. 1. Schematic diagram of the observer design problem.
scheduling variables. To facilitate the observer design, we make
the following practical assumptions:
O2: The effect from the external input ν to the signal z is
ΔAi  ≤ ζi , and ω1  ≤ δ2 (6) constrained as
with positive scalars ζi and δ2 . The assumptions are used to z2 < γν2 (11)
constrain the induced uncertainty and the external noise, which
would benefit the sliding mode observer design. Since ki has where X2 denotes the 2 norm of L2 -bounded signal
a lower bound and a upper bound, the matrix ΔB1,i can be X and γ is the H∞ performance index.
rewritten as O3: The discontinuous functions ϕ1 (α̂) and ϕ2 (α̂) can com-
pensate for the effects of ΔA(α̂)x and B2 (α̂)ω1 , re-
ΔB1,i = δ3,i M B1,i (7) spectively.
where δ3,i is a positive scalar with the value of δ3,i = max{|k̄i − Remark 1: The schematic diagram of the observer design
1|, |k i − 1|} and M  ≤ 1. problem is illustrated in Fig. 1. The primary objective is to de-
For the LPV system in (1), we propose the following sliding sign the sliding model observer to estimate the signal z for the
mode observer: LPV system. The signal z can be the states or the combina-
tion of the states. The available signals for the sliding mode
x̂˙ = A(α̂)x̂ + B1 (α̂)u + L(α̂)(y − ŷ) + ϕ1 (α̂) + ϕ2 (α̂) observer include the output of the LPV system and the schedul-
ŷ = C x̂ (8) ing variables. Different from most of the existing works on
LPV systems, we consider the uncertain measurements on the
where x̂ is the estimated state vector, L(α̂) is the observer gain scheduling variables, that is, the scheduling variables are not
to be tuned, and ϕ1 (α̂) and ϕ2 (α̂) are two discontinuous func- precisely known. Though we discuss and formulate the prob-
tions to compensate for the induced uncertainty ΔA(α̂) and the lem for polytopic LPV systems with uncertain measurements
disturbance ω1 , respectively. The main objective of observer de- on scheduling variables, the case with precise measurements on
sign is to make the estimated state vector x̂ to follow the state scheduling variables is a special one of the uncertain case. If
vector x quickly. Defining the estimation error as e = x − x̂, ki = 1, ∀ i = 1, . . . , 2r , ΔAi and ΔB1,i become zero matrices.
the estimation error dynamics is expressed as The formulated problem is reduced to the case with precise mea-
 surements. For the matrices C and D in the output y, they are
ė = (A(α̂) − L(α̂)C)e + ΔA(α̂)x + B̂1 (α̂)ν not assumed to depend on the scheduling variables. The main
 reason is that, for most of practical applications, the outputs are
+ B2 (α̂)ω1 − ϕ1 (α̂) − ϕ2 (α̂) (9) only the states or the combination of states besides the mea-
surement noises. The weighting factors are not involved in the
where output matrices.
 Remark 2: The proposed observer in (8) is different from
B̂1 (α̂) = B̄1 (α̂) + Δ3 (α̂)M B̃1 (α̂), B̄1 (α̂) = 0 −L(α̂)D the Luenberger-like observer since there are two discontinuous

terms ϕ1 (α̂) and ϕ2 (α̂). The discontinuous terms are used to
 u
Δ3 (α̂) = δ3 (α̂)I, B̃1 (α̂) = B1 (α̂) 0 ,ν = . compensate for the uncertainties and the unknown inputs. This
v type of sliding mode observers were original developed in [37].
Due to the good performance, the idea was then extended to
We can see from (9) that the gains to be tuned, and state vector
various setups and applications; see [38]–[40]. Motivated by the
x and the unknown external input, ϕ1 (α̂) and ϕ2 (α̂) are also
aforementioned works, we develop the sliding mode observer
involved in the estimation error system. Since the main criterion
for LPV systems. It is not a straightforward extension of the
of selecting the gain L(α̂) is to make the estimation error system
existing works. The significant differences include that: 1) the
stable such that the estimation error will converge to zero, we
studied setups are LPV systems with uncertain measurements;
define a new variable as
2) not only the stability but also the H∞ performance [41],
z = Ge (10) [42] will be discussed in this study; and 3) the proposed design
method will be applied to a practical application.
with a constant matrix G. Then, the challenges and objectives
are to tune the observer gain L(α̂) and discontinuous functions
III. OBSERVER DESIGN
ϕ1 (α̂) and ϕ2 (α̂) such that
O1: The estimation error system in (9) is asymptotically In this section, we propose the observer design method based
stable when the external input is zero. on the analysis conditions for the estimation error system. Since
1662 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 21, NO. 3, JUNE 2016

the observer gain is unknown and to be determined, we assume 


2r
x̂T x̂ −1 T
that the gain L(α̂) is given during the stability and H∞ perfor- = α̂i ψ1 (1 + ψ4 )ζi2 P C ey , if ey = 0
mance analysis. Before proceeding, we introduce the following i=1
2eTy ey
useful lemmas.

r
2
Lemma 1: [43] If there exist real matrices Ξ = ΞT , Ê and F̂ ϕ1 (α̂) = α̂i ϕ1,i = 0, if ey = 0
with compatible dimensions, and M satisfying M  ≤ 1, then, i=1
the following condition

r
2
ϕ2 (α̂) = α̂i ϕ2,i = 0, if ey = 0
Ξ + ÊM F̂ + (ÊM F̂ )T < 0 (12)
i=1

is satisfied if and only if there exists a positive scalar ψ > 0 such ψ1


ψ4 = . (16)
that the following condition ψ2 − ψ1
⎡ ⎤ Proof: It infers from [44] that the asymptotically stable and
Ξ Ê ψ F̂ T
⎢ ⎥ the H∞ performance of the estimation error system are both
⎣∗ −ψI 0 ⎦<0 (13)
achieved if the following statement is satisfied:
∗ ∗ −ψI
J = V̇ + z T z − γ 2 ν T ν < 0 (17)
is satisfied.
Lemma 2: For any matrices Ê and F̂ with compatible di- where V is the Lyapunov function and selected as V = eT P e
mensions and a positive scalar ψ, the following condition holds: with a positive-definite matrix P . Considering the trajectories of
the estimation error system, the expression J can be evaluated
Ê T F̂ + (Ê T F̂ )T ≤ ψ Ê T Ê + ψ −1 F̂ T F̂ . (14) as
Observe that the uncertain term M in the condition of (12) J = ėT P e + eT P ė + z T z − γ 2 ν T ν
is eliminated in the condition of (13). Therefore, Lemma 1 can
be used to deal with the bounded uncertainties. The following = eT ((A(α̂) − L(α̂)C)T P + P (A(α̂) − L(α̂)C))e
theorem gives the conditions that guarantee the asymptotically
stable and the H∞ performance of the estimation error system + xT ΔAT (α̂)P e + (xT ΔAT (α̂)P e)T
in (9). + 2eT P B̂1 (α̂)ν + 2eT P B2 (α̂)ω1
Theorem 1: Given a positive scalar γ, the estimation error
system in (9) is asymptotically stable and the H∞ performance − 2ϕT1 (α̂)P e − 2eT P ϕ2 (α̂)
is guaranteed if there exists a positive-definite matrix P and
+ eT GT (α̂)G(α̂)e − γ 2 ν T ν. (18)
positive scalars ψ1 , ψ2 , and ψ3 , such that the following condition
is satisfied:
In terms of Lemma 2, one gets
⎡ ⎤
Λ1 P B̂1 (α̂) GT P
⎢ ⎥ xT ΔAT (α̂)P e + (xT ΔAT (α̂)P e)T
⎢ ∗ −γ 2 I 0 0 ⎥
⎢ ⎥
⎢ ⎥<0 (15) ≤ ψ1−1 eT P 2 e + ψ1 xT ΔAT (α̂)ΔA(α̂)x
⎢ ∗ ∗ −I 0 ⎥
⎣ ⎦
∗ ∗ ∗ −ψ1 I = ψ1−1 eT P 2 e + ψ1 (x̂ + e)T ΔAT (α̂)ΔA(α̂)(x̂ + e)
≤ ψ1−1 eT P 2 e + ψ1 ζ 2 (α̂)(x̂ + e)T (x̂ + e)
where
= ψ1−1 eT P 2 e + ψ1 ζ 2 (α̂)(x̂T x̂ + eT e + x̂T e + eT x̂)
Λ1 = P (A(α̂) − L(α̂)C) + (P (A(α̂) − L(α̂)C))T
= ψ1−1 eT P 2 e + ψ1 ζ 2 (α̂)(x̂T x̂ + eT e + x̂T e + eT x̂)
2
+ ψ2 ζ (α̂)I + ψ3 I
≤ ψ1−1 eT P 2 e + ψ2 ζ 2 (α̂)eT e + ψ1 (1 + ψ4 )ζ 2 (α̂)x̂T x̂
ey = y − ŷ (19)

r
2
ϕ1 (α̂) = α̂i ϕ1,i where ψ2 is defined as ψ2 = ψ 1 (ψψ44 +1) . If ey is zero, since each
i=1 subsystem is observable, the estimation error is zero. If ey is
nonzero, we have the following expressions:

r
2
P B2,i 2 −1 T
= α̂i δ22 ψ3−1 P C ey , if ey = 0
i=1
2eTy ey −2ϕT1 (α̂)P e = −δ22 ψ3−1 P B2 (α̂)2 . (20)


2r
2eT P B2 (α̂)w1 ≤ ψ3 eT e + δ22 ψ3−1 P B2 (α̂)2 (21)
ϕ2 (α̂) = α̂i ϕ2,i
i=1 −2ϕT2 (α̂)P e = −ψ1 (1 + ψ4 )ζ 2 (α̂)x̂T x̂. (22)
ZHANG et al.: H∞ OBSERVER DESIGN FOR LPV SYSTEMS WITH UNCERTAIN MEASUREMENTS ON SCHEDULING VARIABLES 1663

Then, the expression J equals to The main disadvantage of polytopic LPV systems is that the
number of vertices would exponentially increase with respect
J = e Γe + 2e P B̂1 (α̂)ν − γ ν ν
T T 2 T
to the number of scheduling variables. When the number is

T

large, the computational load becomes heavy and it is difficult
e Γ P B̂1 (α̂) e
= (23) to find the feasible solution. Thus, it is necessary to reduce the
ν ∗ −γ I 2 ν conservativeness of obtained condition. The following theorem
offers less conservative results.
where Γ = P (A(α̂) − L(α̂)C) + (P (A(α̂) − L(α̂)C))T +
Theorem 3: Given a positive scalar γ, the estimation error
ψ1−1 P 2 + ψ2 ζ 2 (α̂)I + ψ3 I + GT G. Performing the Schur
system in (9) is asymptotically stable and the H∞ performance
compliment to (15), we can get

is guaranteed if there exists a positive-definite matrix P , random
Γ P B̂1 (α̂) matrices E(α̂), F (α̂), and positive scalars ψ1 , ψ2 , ψ3 , and ψ5 (α̂)
< 0. (24) such that the following conditions are satisfied:
∗ −γ 2 I

Obviously, the condition in (17) is satisfied when the condition Λ2 Λ3 F (α̂)B̄1 (α̂) GT

(15) holds. Therefore, both the stability and the H∞ perfor- ⎢ ∗ −E(α̂) − E T (α̂) E T (α̂)B̄1 (α̂) 0

mance are achieved when the condition (15) holds. The proof is ⎢ ∗ ∗ −γ 2 I
⎢ 0
completed.  ⎢
Ψ= ⎢ ∗⎢ ∗ ∗ −I
Though Theorem 1 provides the condition that can guarantee

the stability and the H∞ performance, the varying matrix M is ⎢ ∗ ∗ ∗ ∗

involved. Since the value of matrix M is uncertain, the condi- ⎢ ∗ ∗ ∗ ∗

tion in (15) has infinite dimensions. In the following theorem,
Lemma 1 will be used to eliminate the matrix M . ∗ ∗ ∗ ∗
Theorem 2: Given a positive scalar γ, the estimation error ⎤
P F (α̂)Δ3 (α̂) 0
system in (9) is asymptotically stable and the H∞ performance ⎥
0 T
E (α̂)Δ3 (α̂) 0 ⎥
is guaranteed if there exists a positive-definite matrix P and ⎥
positive scalars ψ1 , ψ2 , ψ3 , and ψ5 (α̂) such that the following 0 0 ψ5 (α̂)B̃1 (α̂) ⎥
T


condition is satisfied: 0 0 0 ⎥<0 (27)
⎡ ⎤ ⎥
GT ⎥
Λ1 P B̄1 (α̂) P P Δ3 (α̂) 0 −ψ1 I 0 0 ⎥
⎢ ⎥ ⎥
⎢ ∗ −γ 2 I 0 0 0 ψ5 (α̂)B̃1T (α̂) ⎥ ⎥
⎢ ⎥ ∗ −ψ5 (α̂)I 0 ⎦
⎢ ⎥
⎢ ∗ ∗ −I 0 0 0 ⎥



⎥ < 0.
∗ ∗ −ψ5 (α̂)I
⎢ ∗ ∗ ∗ −ψ1 I 0 0 ⎥
⎢ ⎥
⎢ ⎥ where Λ2 = F (α̂)(A(α̂) − L(α̂)C) + (F (α̂)(A(α̂) − L(α̂)
⎢ ∗ ∗ ∗ ∗ −ψ5 (α̂)I 0 ⎥
⎣ ⎦ C))T + ψ2 ζ 2 (α̂) + ψ3 I and Λ3 = P − F (α̂) + (A(α̂) −
∗ ∗ ∗ ∗ ∗ −ψ5 (α̂)I L(α̂)C)T E(α̂).
(25) Proof: The gap between the condition (27) and the condition
(25) can be bridged by multiplying Φ on the left of Ψ and the
Proof: The condition in (15) can be rewritten as transpose on the right of Ψ with
⎡ ⎤
Λ1 P B̄1 (α̂) GT P ⎡ ⎤
⎢ ⎥ I (A(α̂) − L(α̂)C)T 0 0 0 0 0
⎢ ∗ −γ 2 I 0 0 ⎥ ⎢ ⎥
⎢ ⎥ ⎢0 B̄1T (α̂) I 0 0 0 0⎥
⎢ ∗ ∗ −I 0 ⎥ ⎢ ⎥
⎣ ⎦ ⎢0 0 0 I 0 0 0⎥
⎢ ⎥
∗ ∗ ∗ −ψ1 Φ=⎢ ⎥ . (28)
⎢0 0 0 0 I 0 0 ⎥
⎡ ⎤ ⎢ ⎥
P Δ3 (α̂) ⎢ T ⎥
⎢ ⎥ ⎣ 0 Δ 3 (α̂) 0 0 0 I 0 ⎦
⎢ 0 ⎥  
+⎢

⎥M 0
⎥ B̃1 (α̂) 0 0 0 0 0 0 0 0 I
⎣ 0 ⎦

0 Since the weighting factor vector α̂ is involved in the condi-
⎛⎡ ⎤ ⎞T tion of Theorem 3, the condition has infinite dimensions. The
P Δ3 (α̂)
⎜⎢ ⎥ ⎟ following theorem is going to project the condition to the ver-
⎜⎢ 0 ⎥  ⎟ tices of the polytope such that the conditions can be used to
+⎜⎢
⎜⎢
⎥ M 0 B̃1 (α̂)
⎥ 0 0 ⎟
⎟ < 0. (26)
⎝⎣ 0 ⎦ ⎠ calculate the observer gain.
Theorem 4: Given a positive scalar γ and scalars λi , the
0
estimation error system in (9) is asymptotically stable and
In terms of Lemma 1, the conditions (15) and (25) are equivalent the H∞ performance is guaranteed if there exists a positive-
with each other.  definite matrix P , random matrices Fi , L̂i , and positive scalars
1664 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 21, NO. 3, JUNE 2016

ψ1 , ψ2 , ψ3 , and ψ5,i such that the following conditions are is nonsingular. Therefore, F (α̂) is nonsingular and the inverse
satisfied: exists. With the definition of L̂(α̂), the observer gain can be
calculated in terms of (30). 
Ψij + Ψj i < 0 (29)
It is necessary to mention that the H∞ performance index γ
where indicates the effect of the disturbance to the signal z. It is desired
⎡   that the performance index γ is as small as possible. The optimal
Λ2,ij Λ3,ij 0 −L̂i D GT
⎢   observer is defined as the one with the gain which is obtained
⎢ ∗ −λi Fi − λi FiT 0 −λi L̂i D 0 by minimizing the value for γ. The minimization can be done

⎢ in terms of the following corollary.
⎢ ∗ ∗ −γ I2
0
⎢ Corollary 1: The minimum H∞ performance index γ ∗ in

Ψij = ⎢ ∗ ∗ ∗ −I Theorem 4 can be found by solving the following convex opti-


⎢ ∗ ∗ ∗ ∗ mization problem:

⎢ ∗ ∗ ∗ ∗

⎣ γ = min γ
∗ ∗ ∗ ∗ s. t. (29).

P Fi Δ3,j 0 Remark 3: It was reported in [37] that the sliding mode ob-

0 λi Fi Δ3,j 0 ⎥ servers are different from Luenberger observers since there are

T ⎥ nonlinear discontinuous terms incorporated in the observer ex-
0 0 ψ5,i B̃1,j ⎥
⎥ pressions. The discontinuous terms are dependent on the out-
0 0 0 ⎥<0
⎥ put estimation error. It can be seen from (8) and (16) that

−ψ1 I 0 0 ⎥ there are two output-estimation-error-dependent discontinuous

∗ −ψ5,i I ⎥ terms. Therefore, the designed observer in (8) is a sliding mode
0 ⎦ observer. The sliding surface is the estimation error space.
∗ ∗ −ψ5,i I Moreover, we can see from Theorem 1 that the discontinuous
functions ϕ1 (α̂) and ϕ2 (α̂) may have too large values when
Λ2,ij = Fi Aj − L̂i C + (Fi Aj − L̂i C)T + ψ2 ζi2 I + ψ3 I, the residual ey is quite small. These large values of discontinu-
Λ3,ij = P − Fi + λi ATj Fi − λi C T L̂Ti , and 1 ≤ i ≤ j ≤ 2r . ous functions are harmful for the design observer. To avoid this
In addition, the observer gain L(α̂) can be calculated via the phenomenon, the discontinuous functions in Theorem 1 can be
following expression: reformulated as
L(α̂) = F −1 (α̂)L̂(α̂). (30) 
2r

ϕ1 (α̂) = α̂i ϕ1,i


Proof: Note that, in Theorem 3, both E(α̂) and F (α̂) are i=1
coupled with the observer gain L(α̂). Moreover, E(α̂), F (α̂),

r
2
and L(α̂) are all unknown. In order to eliminate the coupled P B2,i 2 −1 T
= α̂i δ22 ψ3−1 P C ey , if ey  > φ
terms, we make the following assumptions:
i=1
2eTy ey

r
2

r
2
ψ5 (α̂) = α̂i ψi ϕ2 (α̂) = α̂i ϕ2,i
i=1
i=1

r
2

r
2
F (α̂) = α̂i Fi x̂T x̂ −1 T
= α̂i ψ1 (1 + ψ4 )ζi2 P C ey , if ey  > φ
i=1
i=1
2eTy ey

2r

2r

E T (α̂) = λi α̂i Fi . (31) ϕ1 (α̂) = α̂i ϕ1,i = 0, if ey  ≤ φ


i=1
i=1

Define a new variable as L̂(α̂) = F (α̂)L(α̂) and assume that 


2r

ϕ2 (α̂) = α̂i ϕ2,i = 0, if ey  ≤ φ



r
2
i=1
L̂(α̂) = α̂i L̂i . (32)
i=1 where φ is a positive scalar that can be obtained by using the
Then, it is obvious that trial-and-error method. Then, the residual ey would be bounded
to φ.

2 r
2
r
−1 
2r
Remark 4: In robust control and filtering problems, conser-
Ψ= α̂i2 Ψii + α̂i α̂j (Ψij + Ψj i ). (33) vativeness is one of the most frequently discussed issues since
i=1 i=1 j =i+1
conservative results may lead to infeasible solutions. Generally,
Since α̂i α̂j is nonnegative, Ψ is negative definite if the condi- there are two main approaches to derive less conservative re-
tions in (29) are satisfied. In addition, it infers from (29) that Fi sults. One is to introduce slack matrices and the other one is
ZHANG et al.: H∞ OBSERVER DESIGN FOR LPV SYSTEMS WITH UNCERTAIN MEASUREMENTS ON SCHEDULING VARIABLES 1665

the fact that tracking/braking forces are difficult to measure or


estimate as the forces are related to the variable road condition,
it is reasonable to assume that the direct yaw moment Mz is
unknown.
In this application, the EGV which is developed in The Ohio
State University [49] is driven on good road condition and the
tire sideslip angles are not too large. Then, we can assume that
Fig. 2. Schematic diagram of a simplified lateral dynamics model. the tires are working in the linear areas and the corresponding
state-space model of the lateral dynamics is governed as

ẋ = Ax + B1 δ + B2 Mz (34)
to develop parameter-dependent conditions. In this study, we
introduce two slack matrices F (α̂) and E(α̂) in Theorem 3. In where
Theorem 1 and Theorem 2, the positive-definite P is coupled ⎡ c̄ ⎤
with the system matrix A(α̂) − L(α̂)C. However, in Theorem
f ⎡ ⎤
⎢ mvx ⎥ 0
3, a nonsingular matrix F (α̂) is coupled with the system matrix β
x= , B1 = ⎢
⎣ l c̄
⎥ , B2 = ⎣ 1 ⎦

instead of the positive-definite matrix. Therefore, the derived η f f
condition in Theorem 3 is relaxed. Moreover, both F (α̂) and Iz
Iz
E(α̂) are dependent on the scheduling variables. ⎡ ⎤
−c̄f − c̄r lr c̄r − lf c̄f
Remark 5: Observe from (30) that the calculation of observer
⎢ − 1 ⎥
gain consists of online computation and offline computation. ⎢ mvx mvx2 ⎥
A= ⎢ ⎥
The offline computation is to solve the conditions in (29) and ⎣ lr c̄r − lf c̄f −lf2 c̄f − lr2 c̄r ⎦
derive the values for Fi and L̂i ∀i = 1, . . . , 2r . The online Iz Iz vx
calculation includes the computation of weighting factors α̂i
and the matrix operation in (30). Moreover, there are some where β is the vehicle sideslip angle, vx is the longitudinal
prescribed scalars λi ∀i = 1, . . . , 2r in Theorem 4. In order to velocity, and the approximated values of corning stiffness are
simply the observer design, the scalars can be selected as one. denoted by c̄f and c̄r , respectively.
However, in order to derive smaller H∞ performance index, Since the longitudinal velocity vx is not always constant,
the nonlinear optimization algorithms such as the MATLAB the model in (34) is not an LTI system but a nonlinear system
functions “fmincon” and “fminsearch” can be employed. [50]. It can be converted to an LPV representative. Suppose
that the longitudinal velocity vx is within the range [ v x v̄x ] in
IV. APPLICATION TO AN EGV which v x is the lower bound of the velocity and v̄x is the upper
bound of the velocity. The scheduling variables are chosen as
The lateral stability is of importance for ground vehicles and
ρ1 = v1x and ρ2 = v12 . Then, the nonlinear model in (34) can be
the vehicle sideslip angle is a critical index for the lateral stabil- x
represented by
ity. It is well-known that the vehicle sideslip angle is necessary
for electronic stability program systems. Unfortunately, cur- ẋ = A(α)x + B1 (α)δ + B2 Mz
rently, the sideslip angle is not measurable by using affordable
physical sensors [45]–[48]. In order to deal with the conflict, 
4

an alternative approach is to estimate the sideslip angle with = αi (Ai x + B1,i δ + B2 Mz ) (35)
i=1
relatively cheap physical sensors. We can see from the existing
work that the yaw rate measurement is a good choice for the where
sideslip angle estimation since the commercialized gyroscope ⎡ ⎤
−c̄f − c̄r lr c̄r − lf c̄f ⎡ c̄ ⎤
can get the accurate yaw rate. − 1 f
To analyze the lateral dynamics and establish the model, we ⎢ mv x mv x2 ⎥ ⎢ mv ⎥
⎢ ⎥
⎥ , B1,1 = ⎢ ⎥
x
employ a two-degree-of-freedom bicycle model as shown in A1 = ⎢ ⎣ l c̄ ⎦
⎣ lr c̄r − lf c̄f −lf2 c̄f − lr2 c̄r ⎦ f f
Fig. 2. The importation notations are given as follows: the total
Iz Iz v x Iz
ground vehicle mass is represented by m; the inertia moment ⎡ ⎤
about the yaw axis through its center of gravity (CG) is repre- −c̄f − c̄r lr c̄r − lf c̄f ⎡ c̄f ⎤
⎢ − 1⎥
sented by Iz ; lf and lr are used to denote distances of the front ⎢ mv̄x mv 2x ⎥ ⎢ mv̄x ⎥
axis and the rear axis from the CG, respectively; the front-wheel A2 = ⎢ ⎥ , B1,2 = ⎢
⎣ l c̄


⎣ lr c̄r − lf c̄f −lf c̄f − lr c̄r
2 2 ⎦ f f
steering angle is δ, which is controlled by the driver and can be
Iz Iz v̄x Iz
measured online via the encoder mounted on the steering col-
⎡ ⎤ ⎡ ⎤
umn; αf and αr stand for the wheel slip angles of the front −c̄f − c̄r lr c̄r − lf c̄f c̄f
⎢ − 1⎥
and rear tires, which are decided the front and rear lateral tire mv x mv̄x2
⎢ mv x ⎥
⎢ ⎥
forces Fy f and Fy r , respectively. In addition, Mz denotes the A3 = ⎢ ⎥ , B1,3 = ⎢
⎣ l c̄


⎣ lr c̄r − lf c̄f −lf c̄f − lr c̄r ⎦
2 2
f f
direct yaw moment about the yaw axis, which is induced by
the unequal tracking/braking forces on the force tyres. Due to Iz Iz v x Iz
1666 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 21, NO. 3, JUNE 2016

⎡ ⎤ ⎡ c̄ ⎤
−c̄f − c̄r lr c̄r − lf c̄f f
⎢ − 1 ⎥
⎢ mv̄x mv̄x2
⎥ ⎢ mv̄ ⎥
⎥ , B1,4 = ⎢ ⎥
x
A4 = ⎢ ⎣ ⎦
⎣ lr c̄r − lf c̄f −lf c̄f − lr c̄r ⎦
2 2 lf c̄f
Iz Iz v̄x Iz
     
1   1 1  1 1 
 v x − v̄1x   v1x2 − v̄12   v − vx   v 2 − v̄ x 
2
α1 =    , α2 =  x  x 
x

  1 1  1 1 
 v1x − v̄1x   v12x − v̄12 
x
 vx − v̄ x   vx 2 − v̄ x 
2

     
1   1 1  1 1 
 v x − v̄1x   v1x2 − v12   v − vx   v 2 − vx 
2
α3 =   x
 , α4 =  x  x .
  1 1  1 1 
 v1x − v̄1x   v12x − v̄12 
x
 vx − v̄ x   v 2x − v̄ x2 

It is necessary to mention that in traditional vehicles the longi-


tudinal velocity vx is generally estimated with the engine speed
and in electric vehicles the longitudinal velocity vx is estimated
with the motor speed. Considering the variable vehicle slip in
the longitudinal direction, the estimated longitudinal velocity
Fig. 3. Acquisition system and corresponding sensors.
cannot be accurate, that is, the calculated weighting factors α1 ,
α2 , α3 , and α4 with the estimated longitudinal velocity are
inaccurate. It is reasonable to assume that
 
ki ∈ 0.9 1.1 ∀ i = 1, 2, 3, 4. (36)
Then, ΔAi  can be calculated. In addition, the upper bound
of Mz is the maximal longitudinal force multiplied by one half
of vehicle width. The proposed algorithm of observer design is
applied to estimate the sideslip angle with the yaw rate mea-
surements, that is, the output is
y = Cx + v (37)
with C = [ 0 1 ]. In this application, we design two observers
for comparisons.
 
Observer 1 : ki ∈ 0.9 1.1 ∀ i = 1, 2, 3, 4.
Observer 2 : ki = 1∀ i = 1, 2, 3, 4.
Since the main objective is to estimate the sideslip angle,
the value of matrix G is selected as G = [ 1 0 ]. For a fair
Fig. 4. Front-wheel steering angle in the first test.
comparison, all the conditions are identical when designing the
two observers.
In order to evaluate the performance of the designed ob-
servers, the EGV is equipped with an acquisition system. The equip to vehicles in the market. In order to simulate the inac-
system is consisted of a steering angle encoder and a high-end curate longitudinal velocity measurements, the GPS-measured
RT3003 navigation system from Oxford Technical Solutions. It longitudinal speed is multiplied by a factor within the range
is necessary to mention that the navigation system can measure [ 0.95 1.05 ] and the factor randomly changes with respect to
the longitudinal, the lateral, the yaw, and the pitch motions with time.
a high accuracy. The encoder placed on the steering column We carry out the experimental tests at Transportation Re-
can measure the hand-wheel steering angle. The front-wheel search Center (TRC) in East Liberty, OH, USA. Fig. 4 depicts
steering angle is available since the hand-wheel steering angle the front-wheel steering angle during the first test. The maximal
and the front-wheel steering angle have a fixed relationship. steering angle is around 16◦ and the front wheels are turned from
The acquisition system is described in Fig. 3. The utilized out- left to right frequently. Fig. 5 shows the longitudinal velocity
puts of the GPS include the longitudinal speed (vx ), the lateral vx in the test. For the safety issue, the EGV is not driven too
speed (vy ), and the yaw rate (η). During the experiments, the fast. The red-dash curve is the actual velocity, which is measured
measured signals are collected by a dSPACE MicroAutoBox with the accurate GPS. Since the revolution is high, the red curve
and the data are then transmitted and stored in a Host PC. It is can be taken as the actual value. The blue-solid curve illustrates
necessary to mention that the longitudinal velocity (vx ) can be the simulated uncertain measurements and the uncertain factor
precisely measured in this application since the equipped GPS is realized with the Simulink block Uniform Random Num-
has a high performance. However, the expensive GPS cannot ber (minimum 0.95 and maximal 1.05). It is obvious that the
ZHANG et al.: H∞ OBSERVER DESIGN FOR LPV SYSTEMS WITH UNCERTAIN MEASUREMENTS ON SCHEDULING VARIABLES 1667

Fig. 7. Sideslip angle comparison for the designed observers in the


Fig. 5. Longitudinal velocity in the first test. first test.

TABLE I
COMPARISON OF SIDESLIP ANGLE ESTIMATION ERROR eβ FOR TWO
OBSERVERS IN THE FIRST TEST

Method 2 Norm of e β ∞ Norm of e β

Observer 2 0.7673 0.0742


Observer 1 0.6121 0.0642
Relative improvement 21.94% 13.48%

is defined as the measured value minus the estimate value. The


errors are sampled with the sampling period of 0.1 s. Table I
lists the comparison of sideslip angle estimation error eβ for two
observers. Observe that, in terms of 2 norm and ∞ norm of eβ ,
observer 1 can lead to better performance than observer 2 and
the relative improvements are 21.94% and 13.48%, respectively.
In the first test, the data are recorded when the vehicle starts
Fig. 6. Yaw rate comparison for the designed observers in the first test.
up. Therefore, the measured initial sideslip angle is the same
with the estimated initial sideslip angle. In the second test, the
measured initial conditions are nonzero and the initial condi-
weighting factors calculated with the measured velocity are not tions of the designed observers are zeros. Fig. 8 depicts the
the same with the ones calculated by using the actual velocity. front-wheel steering angle in the second test. The measured lon-
Therefore, the uncertain weighting factors should be considered gitudinal velocity is also multiplied by a random factor similar
from the application perspective. to the first test for the design observers.
With the experimental data, both observers are used to esti- With the experimental data in the second test, both observers
mate the sideslip angle during the test. Since the longitudinal are used to estimate the sideslip angle and show the transient
speed vx and the lateral speed vy are available from the accurate response. Fig. 9 shows the yaw rate comparison between the
GPS, the measured sideslip angle is the ratio of vy and vx when measured signal and the estimated values with the designed two
vx is nonzero. When vx is close to zero, the sideslip angle is observers. Similarly, we can see from the zoom-in figure that
defined as zero. Fig. 6 shows the yaw rate comparison between both observers can estimate the yaw rate well and there is al-
the measured signal and the estimated values with the designed most no estimation error. Moreover, the estimated yaw rate can
two observers. We can see from the zoom-in figure that both increase from zero to the measured value quickly. Fig. 10 il-
observers can estimate the yaw rate well and there is almost no lustrates the sideslip angle comparison between the measured
estimation error. Fig. 7 illustrates the sideslip angle comparison value and the estimated values. Again, observe that the observer
between the measured value and the estimated values. It infers 1 can achieve better performance than observer 2. The estimated
from the figure that the observer 1 can achieve better perfor- value can follow the actual value in the first two seconds, that is,
mance than observer 2. The sideslip angle estimation error eβ the transient response is excellent. Table II lists the comparison
1668 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 21, NO. 3, JUNE 2016

TABLE II
COMPARISON OF SIDESLIP ANGLE ESTIMATION ERROR eβ FOR TWO
OBSERVERS IN THE SECOND TEST

Method 2 Norm of e β ∞ Norm of e β

Observer 2 0.6533 0.0456


Observer 1 0.5295 0.0418
Relative improvement 18.95% 8.33%

of sideslip angle estimation error eβ for two observers. We can


seen that, in terms of 2 norm and ∞ norm of eβ , observer 1
can lead to better performance than observer 2 and the rela-
tive improvements are 18.95% and 8.33%, respectively. We can
conclude that, with the consideration of uncertain scheduling
parameters during the design, the designed observer is more
reasonable and can achieve better performance.
Fig. 8. Front-wheel steering angle in the second test.
V. CONCLUSION
In this study, we have investigated the gain-scheduling ob-
server design for LPV systems with uncertain scheduling vari-
ables. Due to the uncertain measurements on scheduling vari-
ables, the corresponding weighting factors were not the same
with the actual ones and assumed to be uncertain times of the ac-
tual values. Then, the system was converted to an LPV system
with uncertainties. A gain-scheduling sliding mode observer
was proposed to estimate the system state vector. Both the sta-
bility and the H∞ performance were studied for the estimation
error system. Based on the analytic results, the observer design
method was developed. The developed method was applied to
an EGV. Experimental tests were carried out and the compari-
son results validated the advantages of the designed observer. In
the future research, we will focus on the sliding mode controller
design [51], [52] and mode predictive control [53], [54].
REFERENCES
Fig. 9. Yaw rate comparison for the designed observers in the second
test. [1] M. Butcher and A. Karimi, “Linear parameter-varying iterative learning
control with application to a linear motor system,” IEEE/ASME Trans.
Mechatronics, vol. 15, no. 3, pp. 412–420, Jun. 2010.
[2] S. Zhang, J.-J. Yang, and G. Zhu, “LPV modeling and mixed constrained
H2 /H∞ control of an electronic throttle,” IEEE/ASME Trans. Mechatron-
ics, vol. 20, no. 5, pp. 2120–2032, Oct. 2015.
[3] P. Apkarian, P. Gahinet, and G. Becker, “Self-scheduled H∞ control
of linear parameter-varying systems: a design example,” Automatica,
vol. 31, no. 9, pp. 1251–1261, 1995.
[4] J. S. Shamma and M. Athans, “Guaranteed properties of gain scheduled
control for linear parameter-varying plants,” Automatica, vol. 27, no. 3,
pp. 559–564, 1991.
[5] G. Becker and A. Packard, “Robust performance of linear parametrically
varying systems using parametrically-dependent linear feedback,” Syst.
Control Lett., vol. 23, no. 3, pp. 205–215, 1994.
[6] L. Wu, X. Yang, and F. Li, “Nonfragile output tracking control of hy-
personic air-breathing vehicles with an LPV model,” IEEE/ASME Trans.
Mechatronics, vol. 18, no. 4, pp. 1280–1288, Aug. 2013.
[7] D. Li and Y. Xi, “The feedback robust MPC for LPV systems with bounded
rates of parameter changes,” IEEE Trans. Autom. Control, vol. 55, no. 2,
pp. 503–507, Feb. 2010.
[8] J. De Caigny, J. F. Camino, R. C. L. F. Oliveira, P. L. D. Peres, and
J. Swevers, “Gain-scheduled dynamic output feedback control for discrete-
time LPV systems,” Int. J. Robust Nonlinear Control, vol. 22, no. 5,
pp. 535–558, 2012.
Fig. 10. Sideslip angle comparison for the designed observers in the [9] W. Heemels, J. Daafouz, and G. Millerioux, “Observer-based control of
second test. discrete-time LPV systems with uncertain parameters,” IEEE Trans. Au-
tom. Control, vol. 55, no. 9, pp. 2130–2135, Sep. 2010.
ZHANG et al.: H∞ OBSERVER DESIGN FOR LPV SYSTEMS WITH UNCERTAIN MEASUREMENTS ON SCHEDULING VARIABLES 1669

[10] F. Amato, M. Ariola, and C. Cosentino, “Finite-time stability of linear [33] M.-F. Hsieh and J. Wang, “Sliding-mode observer for urea-selective cat-
time-varying systems: Analysis and controller design,” IEEE Trans. Au- alytic reduction (SCR) mid-catalyst ammonia concentration estimation,”
tom. Control, vol. 55, no. 4, pp. 1003–1008, Apr. 2010. Int. J. Auto. Technol., vol. 12, no. 3, pp. 321–329, 2011.
[11] F. Bianchi, C. Kunusch, C. Ocampo-Martinez, and R. Sanchez-Pena, “A [34] L. Zhao, J. Huang, H. Liu, B. Li, and W. Kong, “Second-order
gain-scheduled LPV control for oxygen stoichiometry regulation in PEM sliding-mode observer with online parameter identification for sensorless
fuel cell systems,” IEEE Trans. Control Syst. Technol., vol. 22, no. 5, induction motor drives,” IEEE Trans. Ind. Electron., vol. 61, no. 10,
pp. 1837–1844, Sep. 2014. pp. 5280–5289, Oct. 2014.
[12] X. Wei and L. del Re, “Gain scheduled H∞ control for air path systems of [35] Y. Lin, Y. Shi, and R. Burton, “Modeling and robust discrete-time sliding
diesel engines using LPV techniques,” IEEE Trans. Control Syst. Technol., mode contril design for a fluid power electrohydraulic actuator (EHA)
vol. 15, no. 3, pp. 406–415, May 2007. system,” IEEE/ASME Trans. Mechatronics, vol. 18, no. 1, pp. 1–10,
[13] D. Robert, O. Sename, and D. Simon, “An H∞ LPV design for sampling Feb. 2013.
varying controllers: Experimentation with a T-inverted pendulum,” IEEE [36] F. Li, L. Wu, P. Shi, and C. C. Lim, “State estimation and sliding mode con-
Trans. Control Syst. Technol., vol. 18, no. 3, pp. 741–749, May 2010. trol for semi-Markovian jump systems with mismatched uncertainties,”
[14] P. Gebraad, J. van Wingerden, P. Fleming, and A. Wright, “LPV identi- Automatica, vol. 51, no. 1, pp. 385–393, 2015.
fication of wind turbine rotor vibrational dynamics using periodic distur- [37] C. P. Tan and C. Edwards, “An LMI approach for designing sliding mode
bance basis functions,” IEEE Trans. Control Syst. Technol., vol. 21, no. 4, observers,” Int. J. Control, vol. 74, no. 6, pp. 1559–1568, 2001.
pp. 1183–1190, Jul. 2013. [38] P. Bergsten, R. Palm, and D. Driankov, “Observers for Takagi-Sugeno
[15] D. Ichalal, B. Marx, J. Ragot, and D. Maquin, “State estimation of Takagi- fuzzy systems,” IEEE Trans. Syst., Man, Cybern. B, vol. 32, no. 1,
Sugeno systems with unmeasurable premise variables,” IET Control The- pp. 114–121, Feb. 2002.
ory Appl., vol. 4, no. 5, pp. 897–908, 2010. [39] C. P. Tan and C. Edwards, “Sliding mode observers for detection and
[16] K.-Y. Lian, C.-H. Chiang, and H.-W. Tu, “LMI-based sensorless control reconstruction of sensor faults,” Automatica, vol. 38, no. 10, pp. 1815–
of permanent-magnet synchronous motors,” IEEE Trans. Ind. Electron., 1821, 2002.
vol. 54, no. 5, pp. 2769–2778, Oct. 2007. [40] A. Akhenak, M. Chadli, J. Ragot, and D. Maquin, “Design of sliding
[17] K. Cho, J. Kim, S. B. Choi, and S. Oh, “A high-precision motion con- mode unknown input observer for uncertain Takagi-Sugeno model,” in
trol based on a periodic adaptive disturbance observer in a PMLSM,” Proc. Mediterranean Conf. Control Autom., Athens, Greece, Jun. 2007,
IEEE/ASME Trans. Mechatronics, vol. 20, no. 5, pp. 2158–2171, pp. 1–6.
Oct. 2015. [41] M. Fallah, R. Bhat, and W. F. Xie, “Optimized control of semiactive
[18] G. Phanomchoeng and R. Rajamani, “Real-time estimation of rollover in- suspension systems using H∞ robust control theory and current signal
dex for tripped rollovers with a novel unknown input nonlinear observer,” estimation,” IEEE/ASME Trans. Mechatronics, vol. 17, no. 4, pp. 767–
IEEE/ASME Trans. Mechatronics, vol. 19, no. 2, pp. 743–754, Apr. 2014. 778, Aug. 2012.
[19] W. He and S. S. Ge, “Vibration control of a nonuniform wind tur- [42] S. Ibaraki, S. Suryanarayanan, and M. Tomizuka, “Design of Luenberger
bine tower via disturbance observer,” IEEE/ASME Trans. Mechatronics, state observers using fixed-structure H∞ optimization and its applica-
vol. 20, no. 1, pp. 237–244, Feb. 2015. tion to fault detection in lane-keeping control of automated vehicles,”
[20] H. Zhang, J. Wang, and Y.-Y. Wang, “Nonlinear observer design of diesel IEEE/ASME Trans. Mechatronics, vol. 10, no. 1, pp. 34–42, Feb. 2005.
engine selective catalytic reduction systems with NO x sensor measure- [43] L. Xie and Y. C. Soh, “Robust control of linear systems with generalized
ments,” IEEE/ASME Trans. Mechatronics, vol. 20, no. 4, pp. 1585–1594, positive real uncertainty,” Automatica, vol. 33, no. 5, pp. 963–967, 1997.
Aug. 2015. [44] H. Zhang, Y. Shi, and A. Saadat Mehr, “Robust weighted H∞ filtering for
[21] H. Hur and H.-S. Ahn, “Unknown input H∞ observer-based localization networked systems with intermitted measurements of multiple sensors,”
of a mobile robot with sensor failure,” IEEE/ASME Trans. Mechatronics, Int. J. Adapt. Control Signal Process., vol. 25, no. 4, pp. 313–330, 2011.
vol. 19, no. 6, pp. 1830–1838, Dec. 2014. [45] R. Wang, H. Zhang, and J. Wang, “Linear parameter-varying controller
[22] A. Mujumdar, B. Tamhane, and S. Kurode, “Observer-based sliding design for four wheel independently-actuated electric ground vehicles
mode control for a class of noncommensurate fractional-order sys- with active steering systems,” IEEE Trans. Control Syst. Technol., vol. 22,
tems,” IEEE/ASME Trans. Mechatronics, vol. 20, no. 5, pp. 2504–2512, no. 4, pp. 1281–1296, Jul. 2014.
Oct. 2015. [46] D. Piyabongkarn, R. Rajamani, J. A. Grogg, and J. Y. Lew, “Development
[23] R. E. Kalman, “A new approach to linear filtering and prediction prob- and experimental evaluation of a slip angle estimator for vehicle stability
lems,” J. Basic Eng., vol. 82, pp. 35–45, 1960. control,” IEEE Trans. Control Syst. Technol., vol. 17, no. 1, pp. 78–88,
[24] S. Helm, M. Kozek, and S. Jakubek, “Combustion torque estimation and Jan. 2009.
misfire detection for calibration of combustion engines by parametric [47] Y. H. Hsu, S. M. Laws, and J. C. Gerdes, “Estimation of tire slip angle and
Kalman filtering,” IEEE Trans. Ind. Electron., vol. 59, no. 11, pp. 4326– friction limits using steering torque,” IEEE Trans. Control Syst. Technol.,
4337, Nov. 2012. vol. 18, no. 4, pp. 896–907, Jul. 2010.
[25] D. Luenberger, “Observing the state of a linear system,” IEEE Trans. Mil. [48] S. H. You, J. O. Hahn, and H. Lee, “New adaptive approaches to real-time
Electron., vol. ME-8, no. 2, pp. 74–80, Apr. 1964. estimation of vehicle sideslip angle,” Control Eng. Pract., vol. 17, no. 12,
[26] K. Erazo and E. M. Hernandez, “A model-based observer for state and pp. 1367–1379, 2009.
stress estimation in structural and mechanical systems: Experimental val- [49] R. Wang, Y. Chen, D. Feng, X. Huang, and J. Wang, “Development and
idation,” Mech. Syst. Signal. Process., vol. 43, no. 1-2, pp. 141–152, performance characterization of an electric ground vehicle with inde-
Feb. 2014. pendently actuated in-wheel motors,” J. Power Sources, vol. 196, no. 8,
[27] H. Gao and T. Chen, “H∞ estimation for uncertain systems with limited pp. 3962–3971, 2011.
communication capacity,” IEEE Trans. Autom. Control, vol. 52, no. 11, [50] H. Zhang, X. Zhang, and J. Wang, “Robust gain-scheduling energy-to-
pp. 2070–2084, Nov. 2007. peak control of vehicle lateral dynamics stabilisation,” Vehicle Syst. Dy-
[28] M. Sahebsara, T. Chen, and S. L. Shah, “Optimal H∞ filtering in net- nam., vol. 52, no. 3, pp. 309–340, 2014.
worked control systems with multiple packet dropouts,” Syst. Control [51] H. Li, P. Shi, D. Yao, and L. Wu, “Observer-based adaptive sliding
Lett., vol. 57, no. 9, pp. 696–702, 2008. mode control of nonlinear Markovian jump systems,” Automatica, vol. 64,
[29] V. Dragan and A.-M. Stoica, “Optimal H2 filtering for a class of linear no. 1, pp. 133–142, 2016.
stochastic systems with sampling,” Automatica, vol. 48, no. 10, pp. 2494– [52] H. Li and J. Wang and P. Shi, “Output-feedback based sliding mode control
2501, 2012. for fuzzy systems with actuator saturation,” IEEE Trans. Fuzzy Syst., to
[30] L. Wu and W. X. Zheng, “Reduced-order H2 filtering for discrete linear be published.
repetitive processes,” Signal Process., vol. 91, no. 7, pp. 1636–1644, 2011. [53] H. Li and Y. Shi, “Robust distributed model predictive control of con-
[31] W. M. Haddad, D. S. Bernstein, and D. Mustafa, “Mixed-norm H2 /H∞ strained continuous-time nonlinear systems: A robustness constraint ap-
regulation and estimation: The discrete-time case,” Syst. Control Lett., proach,” IEEE Trans. Autom. Control, vol. 59, no. 6, pp. 1673–1678,
vol. 16, no. 4, pp. 235–247, 1991. Jun. 2014.
[32] D. J. N. Limebeer, B. D. O. Anderson, and B. Hendel, “A Nash game [54] H. Li and Y. Shi, “Robust distributed receding horizon control of
approach to the mixed H2 /H∞ control problem,” IEEE Trans. Autom. continuous-time nonlinear systems: Handling communication delays and
Control, vol. 39, no. 1, pp. 69–82, Jan. 1994. disturbances,” Automatica, vol. 50, no. 41, pp. 1264–1271, 2014.
1670 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 21, NO. 3, JUNE 2016

Hui Zhang (M’15) received the B.Sc. degree Junmin Wang (SM’14) received the B.E. de-
in mechanical design manufacturing and au- gree in automotive engineering and the first
tomation from the Harbin Institute of Technol- M.S. degree in power machinery and engineer-
ogy, Weihai, China, in 2006, the M.Sc. degree ing from the Tsinghua University, Beijing, China
in automotive engineering from Jilin University, in 1997 and 2000, respectively, the second and
Changchun, China, in 2008, and the Ph.D. de- third M.S. degrees in electrical engineering and
gree in mechanical engineering from the Univer- mechanical engineering from the University of
sity of Victoria, Victoria, Canada in 2012. Minnesota, Twin Cities, MN, USA, in 2003, and
He was a Research Associate at the Depart- the Ph.D. degree in mechanical engineering
ment of Mechanical and Aerospace Engineer- from the University of Texas at Austin, TX, USA,
ing, Ohio State University, Columbus, OH, USA. in 2007.
His research interests include Diesel engine aftertreatment systems, He has five years of full-time industrial research experience (May
vehicle dynamics and control, Mechatronics, robust control and filter- 2003–August 2008) at Southwest Research Institute, San Antonio, TX,
ing, networked control systems, and signal processing. He is an au- USA. In September 2008, he joined Ohio State University and founded
thor/coauthor of more than 50 peer-reviewed papers on journals and the Vehicle Systems and Control Laboratory, whose major research in-
conference proceedings. terests include control, modeling, estimation, optimization, and diagno-
Dr. Zhang has served on the IFAC Technical Committee on Automo- sis of dynamical systems, specifically for automotive engine, powertrain,
tive Control, ASME Automotive and Transportation Systems Technical aftertreatment, hybrid, flexible fuel, renewable energy, (electric) ground
Committee, SAE Commercial Vehicle Committee, and International Pro- vehicle, transportation, sustainable mobility, energy storage, and mecha-
gram Committee for the IASTED International Conference on Control tronic systems. He is the author or coauthor of more than 230 peer-
and Applications. He serves as an Associate Editor for Neurocomputing; reviewed publications including 110 journal articles and 11 U.S. patents.
Board member of International Journal of Hybrid and Electric Vehicles, Prof. Wang received the Ohio State University Harrison Faculty Award
Mechanical Systems and Signal Processing; Guest Editor of Mechatron- for Excellence in 2015, IEEE Distinguished Lecturer in 2015, National
ics, Journal of the Franklin Institute, and International Journal of Vehicle Science Foundation CAREER Award, SAE International Ralph R. Tee-
Deisgn; and Conference Editorial Board of the ASME Dynamic Systems tor Educational Award in 2012, SAE International Vincent Bendix Au-
and Control Division, American Control Conference, and ASME Dynamic tomotive Electronics Engineering Award in 2011, and Office of Naval
Systems and Control Conference. Research Young Investigator Program (ONR-YIP) Award in 2009. He is
a Fellow of the Society of Automotive Engineers and serves as an As-
sociate Editor for various international journals.

Guoguang Zhang received the B.S. degree in


vehicle engineering from the Shandong Uni-
versity of Technology, Zibo, China, in 2010
and the M.S. degree in mechanical engineering
from Shanghai Jiao Tong University, Shanghai,
China, in 2013. He is currently working toward
the Ph.D. degree in the Department of Mechani-
cal and Aerospace Engineering at the Ohio State
University, Columbus, OH, USA.
His research interests include modeling, esti-
mation, and control systems for ground vehicles.

You might also like