Professional Documents
Culture Documents
0022 PDF
0022 PDF
3
Hyatt Regency Riverfront, St. Louis, MO, USA
June 10-12, 2009
Abstract— Sliding-mode observer design is considered for be constructed for systems that do not satisfy the observer
linear systems with unknown inputs when the so-called observer matching condition.
matching condition is not satisfied. To circumvent the restriction In this paper, we design the sliding-mode observer pre-
imposed by the observer matching condition, the method of
utilizing auxiliary outputs generated by high-order sliding- sented in [8] for systems that do not satisfy the observer
mode exact differentiators in the sliding-mode observer design matching condition. We adopt the idea of auxiliary outputs
has been proposed in the literature. In this paper, an alternative used in [14], but propose an alternative approach for the
approach is proposed to use high-gain approximate differen- generation of auxiliary outputs. We use high-gain observers
tiators of simpler architecture instead of high-order sliding- rather than high-order sliding-mode observers to obtain the
mode exact differentiators. The capability of reconstructing
the unknown inputs using the proposed high-gain approximate estimates of auxiliary outputs. The high-gain observer is
differentiator based sliding-mode observer is also discussed and often referred to as approximate differentiator [15]. The
then illustrated with a numerical example. proposed high-gain approximate differentiator based sliding-
mode observer can achieve good state estimation perfor-
I. I NTRODUCTION mance. The advantage of our developed technique is that
the overall observer architecture is simpler than the high-
Unknown input observer (UIO) has been developed to order sliding-mode exact differentiator based sliding-mode
estimate the states of the system with inputs that are unknown observer proposed in [14]. We also discuss the capability
or partially known. Linear UIO architectures that have been of the proposed high-gain approximate differentiator based
developed for linear system are presented in [1]–[5]. UIO ar- sliding-mode observer in the unknown input reconstruction,
chitectures for non-linear systems with unknown inputs have which is then illustrated with a numerical example.
been reported in [6], [7]. Motivated by the design of sliding-
mode controllers, first-order sliding mode based UIOs have II. S YSTEM D ESCRIPTION AND P ROBLEM S TATEMENT
been discussed in [5], [8]–[10]. The main advantage of We consider the following class of linear time-invariant
sliding-mode observers over their linear counterparts is that systems with unknown inputs
while in sliding, they are insensitive to the unknown inputs,
and, moreover, they can be used to reconstruct unknown ẋ = Ax + B 1 u1 + B 2 u2
(1)
inputs which could be a combination of system disturbances, y = Cx,
faults or non-linearities. The reconstruction of unknown where x ∈ Rn , y ∈ Rp , u1 ∈ Rm1 and u2 ∈ Rm2 are
inputs has found impressive applications in fault-detection the state, output, known and unknown input vectors, and
and isolation [4], [9], [10]. B 1 ∈ Rn×m1 , B 2 ∈ Rn×m2 and C ∈ Rp×n are known
The necessary and sufficient conditions for the existence constant matrices. For the above system, we assume that
of most of the unknown input observers proposed thus far 1) B 2 and C are of full rank, that is, rank B 2 = m2 and
are that the observer matching condition is satisfied and the rank C = p, and m2 ≤ p;
invariant zeros of the system involving unknown input are 2) there is ρ > 0 such that ku2 (t)k ≤ ρ for all t, where
in the open left half complex plane. However, the observer k·k denotes the standard Euclidean norm;
matching condition seriously restrict the applicability of slid- 3) the invariant zeros of the system model given by the
ing mode observers. Recently, high-order sliding mode based triple (A, B 2 , C) are in the open left-hand complex
unknown input observers [11]–[14] have been developed plane, or equivalently,
to overcome this restrictive condition. In [13], a change
of coordinates is performed using a constructive algorithm sI n − A B 2
rank = n + m2 . (2)
to transform the system into a quasi-block triangular ob- C O
servable form. Then a step-by-step second order sliding-
for all s such that ℜ(s) ≥ 0.
mode observer is constructed for the transformed system.
In [14], auxiliary outputs are defined so that the conventional It follows from [8] that, if the so-called observer matching
unknown input sliding-mode observer proposed in [9] can condition [13] is also satisfied for the system modeled by (1),
that is,
Karanjit Kalsi, Jianming Lian and Stanislaw H. Żak are with the School rank B 2 = rank(CB 2 ) = m2 , (3)
of Electrical and Computer Engineering, Purdue University, IN 47907,
USA. {kkalsi,jlian,zak}@purdue.edu. Stefen Hui is with we can construct the Walcott-Żak sliding-mode observer,
the Department of Mathematical Sciences, San Diego State University, San
Diego, CA 92182, USA. hui@saturn.sdsu.edu. ˙ = Ax̂ + B 1 u1 + L (y − ŷ) − B 2 E(y, ŷ, η)
x̂ (4)
III. H IGH -G AIN A PPROXIMATE D IFFERENTIATOR b̄i2 = [ci B 1 · · · ci Aγi −1 B 1 ]⊤ and c̄i = [1 0 · · · 0]. We
assume, as in [14], that x and ẋ are bounded and |yij | ≤ dij ,
In this section, we propose a high-gain approximate dif- which implies that u1 is bounded. If γi > 1, we construct
ferentiator based sliding-mode observer for the systems that the following high-gain observers,
do not satisfy the observer matching condition.
ŷ˙ i1 = ŷi2 + αǫi1 (yi1 − ŷi1 ) + ci B 1 u1
..
A. Auxiliary Output Signals
.
α i −1)
We first define as in [14] the auxiliary outputs that are then ŷ˙ i(γi −1) = ŷiγi + ǫi(γ γi −1 (yi1 − ŷi1 ) + ci Aγi −2 B 1 u1
α
used to construct the sliding-mode observer. Let ci be the i- ˙ŷiγ = iγ γ
i
(y i1 − ŷ i1 ) + ci Aγi −1
B 1 u 1 ,
i ǫ i
th row of the output matrix C. Recall that the relative degree (8)
of the i-th output yi with respect to the unknown input u2 where ǫ ∈ (0, 1) is a design parameter and αij , j = 1, . . . , γi ,
is defined to be the smallest positive integer ri such that are selected so that the roots of the equation, sγi +αi1 sγi −1 +
· · · + αi(γi −1) s + αiγi = 0, have negative real parts. Let
ci Ak B 2 = 0, k = 0, . . . , ri − 2 y hi = [ŷi1 · · · ŷiγi ]⊤ and li = [αi1 /ǫ · · · αiγi /ǫγi ]⊤ . We
ci Ari −1 B 2 6= 0. can rewrite (8) as
1190
Note that the induced Euclidean norm of D is 1, that is, e(t) is bounded for t0 ≤ t ≤ t0 + T (ǫ). For t ≥ t0 + T (ǫ),
kDk = 1. Let βi = 0 and Ti (ǫ) = 0 if γi = 1. Thus, because y sh (t) = y h (t) and y h = y a − Dζ, the dynamics
itPfollows from the proposition that kζk ≤ βǫ, where β = of the state estimation error (15) become
p 1
( i=1 βi2 ) 2 , after a finite time T (ǫ) = max1≤i≤p Ti (ǫ), and
limǫ→0 T (ǫ) = 0. ė = Ae + La (y h − ŷ a ) − B 2 u2 − B 2 E a (y h , ŷ a , η)
= (A − La C a ) e − La Dζ − B 2 u2
IV. S TATE E STIMATION P ERFORMANCE A NALYSIS
− B 2 E a (y h , ŷ a , η). (16)
In order to eliminate the peaking phenomena that accom-
panies the operation of the above high-gain observer [17], Consider the Lyapunov function candidate V = 21 e⊤ P a e
we introduce the saturation of the signal y h such that y sh = for t ≥ t0 + T (ǫ). Evaluating the time derivative of V on
[y sh1 ⊤ · · · y shp ⊤ ]⊤ , where y shi = y ai = yi1 if γi = 1 and the solutions of (16), we obtain
ŷi1
ŷiγi
⊤ V̇ = e⊤ P (A − La C a ) e − e⊤ P a La Dζ
s
y hi = Si1 sat · · · Siγi sat − e⊤ P a B 2 u2 − e⊤ P a B 2 E a (y h , ŷ a , η)
Si1 Siγi
with Sij > dij if γi > 1. Then we construct the following = −e⊤ Qa e − e⊤ P a La Dζ
sliding-mode observer, − (F a C a e)⊤ u2 − (F a C a e)⊤ E a (y h , ŷ a , η)
˙ = Ax̂+B 1 u1 +La (y s − ŷ )−B 2 E a (y s , ŷ , η), (13) = −e⊤ Qa e − e⊤ P a La Dζ − (F a C a e + F a Dζ)⊤ u2
x̂ h a h a
− (F a C a e + F a Dζ)⊤ E a (y h , ŷ a , η)
where ŷ a = C a x̂ and
+ (F a Dζ)⊤ u2 + (F a Dζ)⊤ E a (y h , ŷ a , η).
F a (ŷ a −y sh )
s η F ŷ −y s if F a (ŷ a − y sh ) 6= 0 If F a (C a e + Dζ) = 0, then
E a (y h , ŷ a , η) = k a ( a h )k
0 if F a (ŷ a − y sh ) = 0. −(F a C a e + F a Dζ)⊤ u2
where La ∈ Rn×γ and F a ∈ Rm2 ×γ are matrices such that − (F a C a e + F a C a e)⊤ E a = 0. (17)
⊤
(A − La C a ) P a + P a (A − La C a ) = −2Qa < 0 On the other hand, if F a (C a e + Dζ) 6= 0, then
and −(F a C a e + F a Dζ)⊤ u2 − (F a C a e + F a C a e)⊤ E a
F aC a = B⊤
2 Pa (14) = −(F a C a e + F a Dζ)⊤ u2
for some symmetric positive definite P a ∈ Rn×n and Qa ∈ F a C a e + F a Dζ
− η(F a C a e + F a Dζ)⊤
Rn×n . It follows from (1) and (13) that kF a C a e + F a Dζk
≤ −(η − ρ)kF a C a e + F a Dζk ≤ 0. (18)
ė = Ae + La (y sh − ŷ a ) − B 2 u2 − B 2 E a (y sh , ŷ a , η). (15)
It follows from (17) and (18) that in both cases we have
In the following, we analyze the performance of the proposed
high-gain approximate differentiator based sliding-mode ob- V̇ ≤ −e⊤ Qa e − e⊤ P a La Dζ
server given by (13). + (F a Dζ)⊤ u2 + (F a Dζ)⊤ E a (y h , ŷ a , η)
Theorem 1: For the dynamical system (1) and the associ- ≤ −λmin (Qa )kek2 + βǫkP a La kkek
ated sliding-mode observer (13) with high-gain approximate
differentiators (9), there exists a constant ǫ∗ ∈ (0, 1) such + (η + ρ)βǫkF a k
√
that if ǫ ∈ (0, ǫ∗ ) and η ≥ ρ, then the state estimation error = −2µa V + κ1 ǫ V + κ2 ǫ, (19)
e(t) is uniformly ultimately bounded. Specifically, after a √ q
finite time Tf (ǫ), we have ke(t)k ≤ κ(ǫ), where where κ1 = 2βkP a La k/ λmax(P a ) and κ2 = (η +
p s ρ)βkF a k. It follows from (19) that
κ1 ǫ + κ21 ǫ2 + 4µa κ1 ǫ 2 √
κ(ǫ) = V̇ ≤ −µa V − µa V + κ1 ǫ V + κ2 ǫ
2µa λmin (P a ) √ √
= −µa V − V − R− V − R+ , (20)
for positive constants µa , κ1 and κ2 .
Proof: It follows from the proposition that kζ(t)k ≤ βǫ where p
for t ≥ t0 + T (ǫ). Then, it follows from (12) that kya (t) − κ1 ǫ − κ21 ǫ2 + 4µa κ2 ǫ
R− = <0
y h (t)k ≤ βǫ for t ≥ t0 +T (ǫ). There exists a constant ǭ such 2µa
that if kya (t) − y h (t)k ≤ βǭ, then y h (t) is not saturated,
and
that is, y sh (t) = y h (t). Thus, we can choose ǫ∗ = min{ǭ, 1} κ1 ǫ +
p
κ21 ǫ2 + 4µa κ1 ǫ
such that if ǫ ∈ (0, ǫ∗ ), then kζ(t)k ≤ βǫ and y sh (t) = y h (t) R+ = > 0.
2µa
after a finite time T (ǫ).
For t0 ≤ t ≤ t0 + T (ǫ), it is guaranteed that the observer We conclude from (20) that V̇ < 0 when kek > R+ . Thus,
state vector x̂(t) in (13) is bounded because u1 , y sh and the state estimation error e is uniformly ultimately bounded
E a (y sh , ŷ a , η) are bounded and A−La C a is Hurwitz. Thus, with respect to any closed ball of radius greater than R+ .
1191
√ 2
Hence,
√ as long √ as V > R+ , that is, V > R+ , we It follows from (22) that if we choose η such that
have ( V − R− )( V − R+ ) < 0. Therefore, if V (t0 + κ3 + κ4 + κ5 + κ6 + κ7
T (ǫ)) = V (e(t0 + T (ǫ))) > R+ 2
and V (t) > R+ 2
for η≥ + ε,
κ8
t ≥ t0 + Tf (ǫ), then V̇ ≤ −µa V , which implies that
where ε is a small positive constant, then
V (t) ≤ exp (−µa (t − t0 − T (ǫ))) V (t0 + T (ǫ)). Thus, we
can find a finite time Tf (ǫ) such that V (t) ≤ R+ 2
for σ ⊤ σ̇ ≤ −εkσk, (23)
t ≥ t0 + T (ǫ), where Tf (ǫ) is the solution to the equation
2 which implies the above hyperplane is invariant. Let Ts
V (t0 + T (ǫ)) exp(−µa (Tf (ǫ) − T (ǫ))) = R+ as
denote the time the sliding surface is reached. Using the
1 V (t0 + T (ǫ)) same arguments as in [9, p. 53], we obtain
Tf (ǫ) = T (ǫ) + ln 2 .
µa R+ kσ(t0 + Tf (R))k
2 2 Ts ≤ t0 + Tf (R) + .
On the other hand, if V (t0 + T (ǫ)) ≤ R+ , then V (t) ≤ R+ ε
for t ≥ t0 + T (ǫ). In such a case, we can choose Tf (ǫ) = Thus, the proof of the theorem is complete.
T (ǫ). Therefore, there exists a finite time Tf (ǫ) such that
V (t) ≤ R+ 2
for t ≥ t0 + Tf (ǫ), which implies that ke(t)k ≤ V. U NKNOWN I NPUT R ECONSTRUCTION
κ(ǫ). The proof of the theorem is complete. It follows from Theorem 2 that the manifold {(e, ζ) : σ =
Remark: It follows from Theorem 1 that the state esti- F a (C a e+Dζ) = 0} is invariant and is reached after a finite
mation error enters the closed ball {e : kek ≤ κ(ǫ)} after a time. Therefore, we have
finite time Tf (ǫ). It is easy to verify that
σ̇ = F a C a (A − La C a )e − F a C a La Dζ − F a C a B 2 u2
∞ if V (t0 ) 6= 0
lim Tf (ǫ) = − F a C a B 2 E a (y sh , ŷ a , η) + F a D ζ̇ = 0. (24)
ǫ→0+ 0 if V (t0 ) = 0,
Substituting (14) into (24) and performing simple manipula-
because limǫ→0+ T (ǫ) = 0 and limǫ→0+ R+ = 0. Moreover,
tions, we obtain
the radius of the above closed ball can be adjusted by the −1
design parameter ǫ and because limǫ→0+ κ(ǫ) = 0, the state
u2 = B ⊤ 2 P aB2 F a C a (A − La C a )e + F a D ζ̇
estimation error e converges to the origin as ǫ goes to zero.
Theorem 2: For sufficiently large η, the sliding surface, F a C a La Dζ − E a (y sh , ŷ a , η).. (25)
{(e, ζ) : σ = F a (C a e + Dζ) = 0} is invariant in the
(e, ζ)-space and is reached in finite time. By the proposition, we have kζ(t)k ≤ βǫ for t ≥ t0 + T (ǫ).
Proof: Let ζ = [ζ ⊤ ⊤ ⊤
1 · · · ζ p ] . Using (11) for γi > 1
By Theorem 1, we have ke(t)k ≤ κ(ǫ) for t ≥ t0 + Tf (ǫ),
and the fact that ζ i = 0 if γi = 1, we obtain where limǫ→0+ κ(ǫ) = 0. Therefore, for sufficiently small ǫ,
kζ(t)k and ke(t)k becomes negligible after a finite time. If,
ǫζ̇ = Āc ζ + ǫB̄ 1 f (x, u2 ), (21)
in addition, kζ̇(t)k becomes negligible for sufficiently small
where Āc = diag[Āc1 · · · Ācp ], B̄ 1 = diag[b̄11 · · · b̄p1 ] ǫ, it follows from (25) that after a finite time,
with Āci = O and b̄i1 = 0 if γi = 1 and f (x, u2 ) =
u2 ≈ −E a (y sh , ŷ a , η). (26)
[f1 (x, u2 ) · · · fp (x, u2 )]⊤ . Because x and u2 are bounded,
we have kf (x, u2 )k ≤ β1 for some β1 > 0. For t ≥ t0 + That is, we can use the proposed architecture to estimate the
Tf (ǫ), it follows from (16) and (21) that unknown input u2 for sufficiently small ǫ.
Recall that ζ = [ζ ⊤ ⊤ ⊤
1 · · · ζ p ] and ζ i = ζ̇ i = 0 if γi = 1.
σ ⊤ σ̇ = σ ⊤ F a C a ė + F a D ζ̇
Thus, in order to obtain (26), it remains to show that if γi >
⊤ 1, then kζ̇ i (t)k becomes negligible for sufficiently small ǫ.
=σ F a C a (A − La C a )e − F a C a La Dζ
We first rewrite (11) as
− F a C a B 2 u2 − F a C a B 2 E a 1
ζ̇ i (t) = Āci ζ i (t) + v i (t), (27)
1 ǫ
+ F a D Āc ζ + F a D B̄ 1 f (x, u2 )
ǫ where v i (t) = b̄i1 fi (x(t), u2 (t)). Because x(t) and u2 (t)
≤ κ(ǫ)kF a C a (A − La C a )kkσk are bounded, it follows from (7) that fi (x(t), u2 (t)) is
+ βǫkF a C a La kkσk + β1 kF a kkB̄1 kkσk bounded. Thus, v i (t) is bounded. To proceed, we define two
+ λmax (B ⊤ notions regarding the function v i (t).
2 P a B 2 )ku2 kkσk
Definition 1: A function v i (t) is left-continuous if
− ηλmin (B ⊤
2 P a B 2 )kσk + βkF
a kkĀc kkσk limǫ→0+ v i (t − ǫ) = v i (t) for all t.
κ3 + κ4 + κ5 + κ6 + κ7 Definition 2: A function v i (t) defined on S ⊂ R is
=− η− κ8 kσk, (22)
κ8 weakly uniformly continuous if for every ν > 0, there exists
where a δ > 0 such that for each interval Ω ⊂ S with length less
than δ, kv i (s) − v i (t)k < ν for s, t ∈ Ω.
κ3 = κ(ǫ)kF a C a (A − La C a )k, κ4 = βǫkF a C a La k,
In the following, we use S1 + S2 , where S1 , S2 ⊂ R, to
κ5 = ρλmax (B ⊤
2 P a B 2 ), κ6 = βkF a kkĀc k, denote the set {s1 + s2 : s1 ∈ S1 , s2 ∈ S2 }. If S1 or S2 is
κ7 = kF a kkB̄ 1 kkf(x, u2 )k, κ8 = λmin (B ⊤
2 P a B 2 ). empty, then S1 + S2 is defined to be empty.
1192
Let J denote the set of points at which v i (t) is discontin- 1.5
y
uous and let τ > t0 > 0. It can be shown, as in [18], that 1
12
y12 estimate
if v i (t) is left-continuous, then limǫ→0+ ζ̇ i (t) = 0 for each
t > t0 ≥ 0. Moreover, if v i (t) is also weakly uniformly 0.5
1.5
VI. N UMERICAL E XAMPLE y
13
y13 estimate
1
In this section, we illustrate the effectiveness of our
0.5
proposed high-gain approximate differentiator based sliding-
mode observer with a numerical example. Our simulations 0
1193
3
1.5 u
21
x u21 estimate
1
x1 estimate 2
1 1
0
0.5
−1
0 −2
−3
−0.5 0 2 4 6 8 10
0 2 4 6 8 10 time (sec)
time (sec)
4
1 u
22
x 3 u22 estimate
2
x2 estimate
2
0.5 1
−1
0 −2
−3
−4
−0.5 0 2 4 6 8 10
0 2 4 6 8 10 time (sec)
time (sec)
1194