Professional Documents
Culture Documents
A New Filtering Approach For Continuous Time Linear Sys - 2014 - IFAC Proceeding
A New Filtering Approach For Continuous Time Linear Sys - 2014 - IFAC Proceeding
Abstract: This paper introduces a new filter for linear continuous-time stochastic systems
with delayed measurements. The approach is inspired by an observer designed for deterministic
systems. The obtained solution is suboptimal and does not use distributed integration terms
with advantages in terms of computational load. The relationship between the delay bound and
the variance of the estimation error is formally characterized and confirmed by a numerical
example.
Keywords: Continuous time filters; System state estimation; Filtering techniques; Filtering
theory; Kalman filters; Hilbert spaces.
Cδn denotes the space of the continuous functions that where −1
maps [−δ, 0] into Rn , with the uniform convergence norm, K(t) = P̂ (t)C T GGT
denoted by k · k∞ . W 1,2 indicates the space of absolutely
and P̂ (t) ∈ Rn×n , defined as the covariance matrix of the
continuous functions form [−δ, 0] into Rn . Lδ,n
2 denotes the estimation error ê(t) := x(t) − x̂(t), is the solution of the
space of Lebesgue measurable square integrable functions matrix differential equation
mapping [−δ, 0] into Rn , with the norm denoted by k · k2 . −1
˙
Given a linear operator T which maps a normed space H1 P̂ (t) = AP̂ (t) + P̂ (t)AT − P̂ C T GGT C P̂ (t) + F F T .
into a second normed space H2 , kT k denotes the operator Under the hypothesis that the pair (A, F ) is controllable
norm defined as supx∈H1 kT xkH2 /kxkH1 . and the pair (A, C) is observable, P̂ (t) converges to a
2.2 Problem formulation unique positive definite symmetric matrix P̂∞ , solution of
the steady-state Riccati equation
−1
In this work we deal with the class of linear stochastic AP̂∞ + P̂∞ AT − P̂∞ C T GGT C P̂∞ + F F T = 0
systems with delayed measurements, described by
from which the steady-state Kalman gain K∞ can be
obtained by
ẋ(t) = Ax(t) + Bu(t) + F ω(t) (1) −1
K∞ = P̂∞ C T GGT .
ȳ(t) = Cx(t − δt ) + Gω(t) (2)
Such a gain matrix can be used to replace the time-varying
where x(t) ∈ Rn is the system state, u(t) ∈ Rp is the gain K(t) in (4) for obtaining the steady-steady Kalman-
deterministic input, ω(t) ∈ Rs is the noise term, and Bucy filter
ȳ(t) ∈ Rq is the output signal. The latter is a function of
x̃˙ s (t) = Ax̃s (t) + Bu(t) + K∞ (y(t) − C x̃s (t)) (4)
the state x at the time t − δ, where δ is the measurement
delay. We assume that δ is bounded and known. The which has two key properties listed in the following lemma.
noise term ω is supposed to be a standard white process, Lemma 1. Consider the steady-state Kalman-Bucy filter
such that E[ω(t)ω T (t)] = Is . Independence of the state (4), then:
and observation noises is assumed, that is, F GT = 0.
Finally, we assume that for t ∈ [−δ, 0], x(t) = φ(t), where (i) the estimate x̃s (t) is asymptotically optimal, i.e.
φ ∈ Cδn := C([−δ, 0]; Rn). Note that this also means that limt→∞ E[(x(t) − x̃s (t))(x(t) − x̃s (t))T ] = P̂∞ ;
(ii) σ(A − K∞ C) ∈ C<0 , i.e. the filtering system (4) is
φ ∈ Lδ,n
2 . asymptotically stable.
The principal aim here is to design a filtering system for
estimating the state of (1)-(2). 3. PROPOSED APPROACH
2.3 Preliminary definitions We here propose a filtering system having the form:
Definition 1. (α-exp stability). Consider a linear delay ˙
x̃(t) = Ax̃(t) + Bu(t) + K̃ (ȳ(t) − C x̃(t − δ)) (5)
system with the form with
K̃ = eA∞ δ K∞ , and A∞ = A − K∞ C.
ż(t) = A0 z(t) + A1 (t)z(t − δ) (3)
For t ∈ [−δ, 0], a pre-shape function x̃(t) = φ̃(t) ∈ Cδn
with z(t) = φz (t) for t ∈ [−δ, 0]. For a given real number is assigned to the filter. Note that from property (ii) in
α > 0, system (3) is said to be α-exp stable if there exists Lemma 1 results that σ (A∞ ) ∈ C<0 . The estimation error
a γ > 0 such that ẽ(t) := x(t) − x̃(t) obeys the following delay differential
kz(t)k ≤ e−αt γkφz k2 , ∀t ≥ 0, ∀φ ∈ Cδn . equation
System (3) is said 0-exp stable if it is asymptotically stable. ˙ = Aẽ(t) − K̃C ẽ(t − δ) + F − K̃G ω(t), (6)
ẽ(t)
Definition 2. (Maximal Delay for α-exp stability). For a
given α > 0 the maximal delay for α-exp stability, denoted with ẽ(t) = φẽ (t) := φ(t) − φ̃(t) for t ∈ [−δ, 0], φẽ ∈ Lδ,n
2 .
∆α , is the supremum among all δ > 0 such that system The corresponding error covariance matrix is defined as
(3) is α-exp stable. If system (3) is α-exp stable for any
δ ∈ [0, ∞), then ∆α = ∞. ∆0 indicates the maximal delay P̃ (t) = E ẽ(t)ẽT (t) , (7)
for 0-exp stability. with initial the initial
value T
2.4 Steady-state Kalman-Bucy gain P̃ (0) = E x(0) − φ̃(0) x(0) − φ̃(0) .
3746
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014
(i) all the eigenvalues in σ (A∞ ) are roots of the charac- Theorem 4 states that for any strictly positive α such that
teristic equation of (8); σ(A∞ ) ∈ C<−α , (9) can be used to study the existence
(ii) for any α > 0 a sufficient condition for α-exp stability of a bound for the error covariance matrix Pe(t); whereas
of (8) with delay δ is that Corollary 5 establishes that, under the same hypothesis
Z δ about A∞ , for all δ < δ̄α the error covariance matrix
kCeA∞ t K̃keαt dt < 1; (9) admits a bound, whose expression is given in (13). Using
0 these results we can express the relationship between δ and
(iii) if ȳ(t) is scalar and moreover CeA∞ t K̃ > 0 ∀t ∈ [0, δ], Pe (t) in the following way.
a necessary and sufficient condition for the α-exp Corollary 6. If the pair (A, C) is observable and the pair
stability of (8) with delay δ is that
Z δ (A, F ) is controllable, the error covariance matrix Pe (t) of
the filter (5) for the system (1)–(2) has bounded norm if
CeA∞ t K̃eαt dt < 1. (10)
0 the constant delay δ is such that
Z δ
Lemma 3. Under the hypothesis of Lemma 2, let δ̄α ∈ R+
be such that kCeA∞ t K∞ kdt < 1.
Z δ̄α 0
kCeA∞ t K̃keαt dt = 1 (11) In the conditions of point (ii) of Lemma 3, this condition is
Rδ
0 necessary, i.e. 0 kCeA∞ t K∞ kdt > 1 implies that kPe(t)k
R∞
or δ̄α = ∞, if 0 kCeA∞ t K̃keαt dt = 1. Then, is not bounded.
(i) δ̄α is a lower bound for ∆α ; Proof. Definition (11) implies that δ̄0 > δ̄α for any
(ii) if the output is scalar and CeA∞ t K̃ > 0 ∀t ∈ [0, δ̄α ], α > 0. This means that δ̄0 is the superior element of
then the bound is strict, i.e. δ̄α = ∆α , and it is the the delay values δ for which the boundedness of the
smallest positive solution to the equation error covariance matrix is guaranteed. The delay bound
is therefore obtained by setting α = 0 in (11).
KT Ā−1
α e Āα
− I − 1 = 0, (12) Remark 7. It can be shown that the bound in (13) de-
where KT and Āα are defined as follows creases when α increases. From definition (11) it results
that δ̄α monotonically decreases when α increases. There-
T
Ab − Bb KA = T AT −1, fore, assuming that δ = δ̄α , it follows that the bound of
the error covariance matrix bound (13) decreases with δ,
Ab − Bb K̃ T = T A∞ T −1 , as it could be expected.
KT = K̃ T − KA
T
,
Āα = Ab − Bb K̃ T + αI = T A∞ T −1 + αI. 4. PROOFS
Therefore, δ̄α = ∆α depends only on α, σ(A) and
σ(A∞ ), i.e. it does not depend on B or C and it is Theorem 4 provides a boundedness condition for the co-
the same if A is replaced by a similar matrix. variance matrix Pe (t) of the error system (6), which has the
form of a linear stochastic system with state delay. In order
Note that: 1) by (9), Lemma 2 gives a condition to study to prove this theorem, we need to represent system (6) with
the α-exp stability of the deterministic part of error system the infinite dimensional state-space representation, which
(6); 2) point (i) of Lemma 3, which trivially follows from is often used for managing such a class of systems (e.g.
condition (9), defines a lower bound for the maximum Germani et al. [2000]). In the next section we describe
delay ∆α for which the α-exp stability of the deterministic this representation and the proof of Theorem 4 is given in
part of error system (6) is guaranteed; 3) such a bound can the following Section 4.2.
be computed by solving (12).
In the following theorem these conditions are used to 4.1 Infinite dimensional state-space model of linear delay
derive an upper bound for the error covariance matrix (7). systems
Theorem 4. Let α > 0 be such that σ(A∞ ) ∈ C<−α . If
condition (9) on δ holds true, then for all t > δ it is Consider a linear delay system with the form:
kF0 k2 ż(t) = A0 z(t) + A1 z(t − δ) + F0 ω(t), (14)
λmax Pe (t) ≤ Mα 2 p0 + , (13) n
2α where z(t) ∈ R and ω belongs to the Hilbert space
where p0 is a positive constant that depends on the pre- L2 ([0, tf ]; Rs ) equipped with standard Gaussian cylinder
shape function, Mα is a bounded function of α, and measure (that corresponds to model ω as a standard white-
e process). Variable z in the interval [−δ, 0] is assumed to be
F0 = F − KG. generated as follows:
Z 0
A proof of this theorem is provided in Section 4.
z(θ) = φz (θ) + k(θ, τ )ω̄(τ )dτ (15)
Corollary 5. In the hypothesis of Theorem 4, let δ̄α ∈ R+ −δ
be the solution of (11). Then (13) holds for all δ < δ̄α .
where φz is uniformly bounded and so it belongs to Lδ,n
2 ,
δ,s̄
This corollary trivially follows from Theorem 4 since, by the process ω̄, independent of ω, belongs to L2 equipped
Lemma 3, condition (9) is satisfied for all δ < δ̄α and with the standard Gaussian cylinder measure, and the
therefore Theorem 4 holds. kernel k(θ, τ ) is integrable for τ ∈ [−δ, 0].
3747
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014
System (14) can be rewritten in state-space form in the The representation is now complete and it enables us to
Hilbert space M2 := Rn ×Lδ,n write the solution of (16) as
2 , endowed with the following
inner product: Z t
Z 0 x(t) = T(t)x(0) + T(t − τ )Fω(τ )dτ,
x0 y0 T
, := x0 y0 + xT1 (θ)y1 (θ)dθ. 0
x1 y1 M −δ where T(t) is the semigroup generated by A. Moreover, it
2
In M2 system (14) assumes the form can be proved that the nuclear covariance operator P(t)
evolves in the Hilbert space of Hilbert-Schimdt operators
φz (0) L0
ẋ(t) = Ax(t) + Fω(t), x(0) = + ω̄ (16) with the following equation
φz L1 Z t
where the component x1 (t) ∈ Rn is the current state z(t) P(t) = T(t)P0 T∗ (t) + T(t − τ )FF∗ T∗ (t − τ )dτ. (21)
and x2 (t) ∈ Lδ,n2 is the state trajectory into the interval 0
[t − δ, t]. The operator A : D(A) → M2 is defined as
"A x + A x (−δ)# We finally define the operator Π0n : M2 → Rn as
x 0 0 1 1
A: A 0 = d Π0n x = x0 ,
x1 x1
dθ which simply extracts the first component of the state
with domain x ∈ M2 , and we introduce the useful property stated by
x0 x0 ∈ Rn the following Lemma.
D(A) := | , x0 = x1 (0)
x1 x1 ∈ W 1,2 Lemma 9. Consider the deterministic part of system (16)
and F : Rs → M2 is defined as (i.e. assume F = 0). If there exist α, γ > 0 such that
F ω(t) kΠ0n x(t)k ≤ γe−αt kφz k2 , ∀t ≥ 0, then there exists a
F : Fω(t) = 0 . uniformly bounded Mα > 0 such that kT(t)k ≤ Mα e−αt ,
0
∀t > δ.
The Hilbert-Schmidt operator
L Proof. For any t ≥ δ,
L= 0 ,
L1 x0 (t)
which defines the stochastic initial state x(0), descends x(t) =
x1 (t)
from definition (15) and it is defined as
Z 0 where x0 (t) = Π0n x(t) ∈ Rn and x1 (t) ∈ Lδ,n
2 , i.e. x1 (t) =
L0 : L2δ,s̄ → Rn ; L0 ω = k(0, τ )ω̄(τ )dτ ; x1 (t, θ) with θ ∈ [−δ, 0] . Moreover, by the definition of
−δ D(A), x1 (t, θ) = x0 (t − θ) = Π0n x(t − θ). Therefore,
Z 0 kT(t)x(0)k2M2 kx(t)k2M2
L1 : Lδ,n 1,2 kT(t)k2 = sup = sup
2 →W ; L1 ω(θ) = k(θ, τ )ω̄(τ )dτ. 2
kx(0)kM2 2
x(0)∈D(A) x(0)∈D(A) kx(0)kM2
−δ R 0
The mean value and nuclear covariance of the initial state kΠ0n x(t)k2 + −δ kΠ0n x(t − θ)k2 dθ
x0 are = sup .
x(0)∈D(A) kx(0)k2M2
φz (0)
x̄0 = , P0 = LL∗ . From the hypothesis of the Lemma, it follows that
φz
R0
The operators in the next proposition must be well defined 0 2 0
kΠn x(t)k + 2 kΠn x(t − θ)k dθ
−δ
to be used to our purposes. kT(t)k2 = sup
x(0)∈D(A) kx(0)k2M
Proposition 8. The operators F∗ , FF∗ , and A∗ are as R0 2
3748
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014
Z t
kP(t)k ≤ kT(t)k2 kP0 k + kT(t − τ )Fk2 dτ 550
0 δ=0.5
Z t 500 δ=1.25
2 −2αt
≤ Mα e kP0 k + Mα 2 e−2α(t−τ ) kFk2 dτ 450 δ=2
0 δ=2.25
400
2 −2αt Mα 2
= Mα e kP0 k + kFk2 (1 − e−2αt ) 350
2α
λmax [Π(t)]
300
Mα 2
≤ Mα 2 kP0 k + kFk2 . 250
2α δ=2.25
200
150
We have now all the instruments to prove Theorem 4 as δ=1.25 δ=0.5
100
follows.
50 δ=2
3749
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014
8 6
6 real state
KBP 4
NF
4
KBF
2
2
0 0
x1
1
x
−2 −2
−4
−4
−6
−6
−8
−10 −8
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
time [s] time [s]
Fig. 2. Real and estimated state component x1 for one of the 100 noises realization with δ = 0.5 (left) and δ = 1.25
(right).
Table 1. Numerical results: MSE. observation delays”, Int. J. Innovative Comput. Inform.
and Control, 3(5):1307–1320, October 2007.
δ = 0.5 δ = 1.25 F. Cacace, A. Germani, C. Manes, “Predictor-based con-
x1 x2 x3 x1 x2 x3 trol of linear systems with large and variable measure-
ment delays”, Int. J. of Control, 87(4):704-714, April
NF 0.730 0.774 1.556 1.546 0.864 4.739 2014
KBP 0.698 0.773 1.472 1.064 0.817 2.896 A. Germani, C. Manes, and P. Pepe “A twofold spline ap-
KBF 0.421 0.740 0.812 0.426 0.723 0.823
proximation for finite horizon LQG control of hereditary
systems”, SIAM J. on Control and Optim., 39(4):1233–
performances of the new filter (NF). Table 5 reports the 1295, 2000.
mean square error (MSE) for the three state components of S. Kong, M. Saif, and H Zhang, “Optimal filtering for
the estimation results obtained with δ = 0.5 and δ = 1.25, Itô-stochastic continuous-Time systems with multiple
with the KBP and the NF, and the corresponding results delayed measurements”, IEEE Trans. on Automatic
for the KBF. Obviously, KBF definitely has better perfor- Control, 58(7):1872–1877, July 2013.
mances with respect to both KBP and NF. Moreover, also X. Lu, H. Zhang, W. Wang, and K.-L. Teo, “Kalman filter-
KBP outperforms the proposed filter. This was expected ing for multiple time delay measurements”, Automatica,
because of the optimality of KBP. However, KBP and 41(8):1455–1461, August 2005.
NF clearly have comparable results. Figure 2 reports an V. S. Pugachev, and I. N. Sinitsyn, Stochastic systems:
example of the estimation results for the first component Theory and Applications, Singapore, World Scientic,
of the system for two values of delay, δ = 0.5 and δ = 1.25. 2001.
H. Zhang, X. Lu, and D. Z. Chang, “Optimal estima-
6. CONCLUSIONS tion for continuous-time systems with delayed measure-
ments”, IEEE Trans. on Automatic Control, 51(5):823–
A suboptimal filter for linear stochastic system with de- 827, May 2006.
layed measurement is proposed in this paper. The filer W. Wang, H. Zhang, and L. Xie, “Optimal Filtering
avoids the computational complexities due to the use of for Continuous-Time Linear Systems with Time-Varying
distributed terms of optimal filters. Bound conditions on Delay”, In 6th International Conference on Informa-
the delay for assure the boundedness of the error covari- tion, Communications & Signal Processing, December
ance matrix are provided. A numeric example confirms 2007.
the theoretical results. Future works will be devoted to
the extension of the approach to the time-varying delay
case.
REFERENCES
A. V. Balakrishnan, “Applied Functional Analysis”,
Springer-Verlag, New York, 1981.
M. V. Basin, and R. M. Zuniga, “Optimal filtering
over observations with multiple delays”, Int. J. Robust
Nonlin. Control, 16(8):685–696, May 2004.
M. V. Basin, E. N. Sanchez, and R. M. Zuniga, “Optimal
linear filtering for systems with multiple state and
3750