2022 IEEE 61st Conference on Decision and Control (CDC)

December 6-9, 2022. Cancún, Mexico

On Distributed Robust Optimal Filter Design With Bounded-Power

Disturbances and White Gaussian Noises
Hongda Sun, Jiahao Li, Yu Feng
2022 IEEE 61st Conference on Decision and Control (CDC) | 978-1-6654-6761-2/22/$31.00 ©2022 IEEE | DOI: 10.1109/CDC51059.2022.9992321

Abstract— This paper is concerned with the distributed great attention in recent research literature [10]–[14]. In [15],
robust optimal filtering problem for discrete-time systems sub- Olfati-Saber converts the problem of distributed Kalman
ject to both bounded-power disturbances and white Gaussian filtering to two separate dynamic consensus problems. Dif-
noises. Relied on the system level synthesis, an upper bound
of the estimation performance is characterized in terms of fusion strategies are adopted to handle distributed Kalman
system responses of the error dynamics and parameters of filtering, fixed-lag smoothing and fixed-point smoothing
the disturbances and noises. Based on such characterization, in [16]. In [17], distributed information Kalman filter is
a numerically tractable algorithm is presented for distributed proposed for the estimation of unstable systems observed
multi-objective filter design. Moreover, the relative estimation by a sparsely connected network of sensors to obtain un-
performance loss incurred by imperfect modeling uncertainty
and the finite truncation is explicitly established. A numerical biased distributed estimates. Moreover, a sensor scheduling
example is also included to show the effectiveness of the current problem of distributed estimation subject to power constraint
results. is studied in [18]. In [19], authors investigate the distributed
Index Terms— Distributed filtering, multi-objective design, Kalman filter design for sensor networks subject to unknown
relative estimation performance loss, system level synthesis. random link failures by introducing a failure detection strat-
egy. Driven by the residual signal related to the modeling
I. I NTRODUCTION mismatch, a cascaded scheme with enhanced robustness is
A desirable estimation scheme for engineering problems proposed for the distributed remote state estimation and the
is often requested to reach prescribed estimation accuracy fundamental limits of data arrival rate is also derived in [20].
with the presence of external disturbances and noises, system Such scheme is further extended to the case of consensus
parameter uncertainty, and variations of operating conditions. filtering for target tracking through lossy wireless sensor
Under such circumstance, the design objective in general networks in [21]. Furthermore, distributed filtering with H∞
cannot be fulfilled by using either Kalman filter or H∞ filter performance for sensor networks has been also reported in
alone, since the former is highly model dependent and the the literature, for instance [22], [23].
latter may not be suitable to tackle stochastic noises [1]. In this paper, we attempt to address the problem of
Therefore, ongoing research efforts on multi-objective (or distributed filtering for discrete-time systems subject to both
mixed performance) filtering have been witnessed during bounded-power disturbances and white Gaussian noises.
the last decades. For example, in [2] a frequency domain With the recently developed system level synthesis ap-
approach is adopted for achieving a mixed H2 /H∞ estima- proach [24], [25], an upper bound of the estimation per-
tion trade-off. The ‘H∞ Gaussian filter’ is proposed within formance is first characterized though system responses of
the Nash game framework by solving two coupled Riccati the error dynamics and parameters of disturbances/noises.
equations in [3]. Such filter is shown to be efficient for A computationally tractable algorithm is then presented to
disturbance torque estimation of a motor drive system with deal with the distributed filter’s gain design. Moreover, the
parameter variations and noises in [4], and is extended to the relative estimation performance loss induced by imperfect
network environment subject to random package dropouts modeling and finite approximation is explicitly conducted.
in [5]. In [6], a new unbiased finite impulse response (FIR) The rest of this paper is organized as follows. Section II
filter with improved robustness is proposed in state space formulates the problem. The distributed design method,
model. Moreover, other approaches have been appealed to together with an algorithm, is included in Section III and
address such multi-objective filtering problems. For example the relative performance loss is established in Section IV. A
see [7]–[9] and the references therein. numerical example and conclusions are given in Section V
On the other hand, compared with traditional centralized and Section VI, respectively.
solutions, distributed strategies are known as an attractive Notation: The superscripts ‘0 ’ and ‘∗’ represent the trans-
alternative for sensor networks due to scalability, flexibility, pose and complex conjugate transpose, respectively. Rm rep-
fewer communications, and parallel processing. The dis- resents m-dimension Euclidean space. The notations E{·},
tributed filtering problem has thus received increased and k·k, Tr{·}, λmin (·), and σ̄{·} denote the standard expectation
operator, Euclidean norm, trace, the minimum eigenvalue,
This work was supported by the Natural Science Foundation of and the largest singular value, respectively. For a matrix A ∈
China under Grant 61973276. Authors are with the Information Engi- Rm×n , its 2 norm and pFrobenius norm are defined as kAk2 =
neering College, Zhejiang University of Technology, 288 Liuhe Road,
Hangzhou, 310023, Zhejiang, China. Corresponding author: Yu Feng; E- σ̄{A} and kAkF = Tr{A0 A}, respectively. Moreover, we
mail: [email protected] denote Diag{·}, Vec{·} and Col{·} as the diagonalization,

978-1-6654-6761-2/22/$31.00 ©2022 IEEE 4086

Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on September 27,2023 at 07:41:19 UTC from IEEE Xplore. Restrictions apply.
vectorization and columnization operations, respectively. edges excluding self-loops. Moreover, the set of neighbors of
RH∞ and RH2 stand for the spaces of proper and strictly the node i is defined by Vei = {j ∈ V : (i, j) ∈ E}. Denote Vi
proper real rational transfer matrices, respectively. Given a as the set of the ith node and its neighbors, i.e., Vi = V
ei ∪{i}.
transfer function G(z) ∈ RH∞ , its H∞ norm is defined Each node, in addition to the local measurement, is also able
as kGk∞ = sup σ̄{G(ejω )}. If G(z) ∈ RH2 , its H2 norm to receive observations from its neighborhood. For simplicity,
ω s we denote
1 Rπ
is defined as kGk2 = Tr{G(ejω )G∗ (ejω )}dω. For yVi (k) = Vec{yj (k)}j∈Vi , CVi (k) = Col{Cj }j∈Vi
2π −π
simplicity, arguments of a function will be omitted when no vVi (k) = Vec{vj (k)}j∈Vi , RVi (k) = Diag{Rj }j∈Vi .
confusion can arise.
Consider the local filter Fi associated with node i ∈ V as
II. P ROBLEM F ORMULATION
Fi : x̂i (k + 1) = Ax̂i (k) + Li (q)[yVi (k) − CVi x̂i (k)], (3)
where x̂i (0) = x̂i,0 , Li (q) ∈ Rn×qi with qi =
P
stochastic signal u(k) = pj is the
 Given a real discrete-time
0 j∈Vi
u1 (k) · · · um (k) , with ui (k), i = 1, . . . , m, being local dynamic estimation gain to be determined, and q−1 is
a stationary random process, we define the autocorrelation the shift operator. Combining the plant (1), the sensors (2)
matrix of u(k) as follows and the filter (3), the local estimation error, denoted as
N −1
1 X ei (k) = x(k) − x̂i (k), associated with the ith node is thus
Ru (τ ) = lim E {u(k + τ )u0 (k)} , given by
N →∞ N
k=0
ei (k + 1) = (A − Li (q)CVi )ei (k) + d(k) (4)
where τ ∈ Z. The discrete-time Fourier transform of the au-
tocorrelation matrix Ru (τ ), called the power spectral density + w(k) − Li (q)vVi (k).
(PSD) of u(k), is defined as Here, the plant (1) is assumed observable with regard to
∞
X each sensor and its neighbors, i.e. (A, CVi ) is observable, for
Φu (ω) = Ru (τ )e−jωτ . i = 1, . . . , M . Then, we define the cost functional associated
τ =−∞ with each node as
Let P be the space of all bounded power signals. Then N
1 X 0
the mean power norm can be defined on P as Ji = lim ei (k)ei (k) = kei (k)k2P . (5)
N →∞ N
v k=1
u N −1
u 1 X p
ku(k)kP = t lim E {ku(k)k2 } = Tr{Ru (0)}. Problem 1 Consider the plant (1), the sensors (2), and the
N →∞ N
k=0 local filter Fi (3) with i ∈ V. The distributed robust optimal
The above seminorm has been widely adopted in robust filtering problem under consideration is to find the dynamic
control and filtering theory [1], [3], [20]. See [1] for a gain Li minimizing the cost functional Ji .
comprehensive interpretation of this seminorm. III. D ISTRIBUTED F ILTER D ESIGN
Moreover, we consider the following discrete-time linear
In this section, we present a way to address Problem 1
plant
defined previously. For the error dynamics (4), the transfer
x(k + 1) = Ax(k) + d(k) + w(k), x(0) = x0 , (1) matrices from the process noise w(k) to the local estimation
error ei (k) and from the measurement noise vVi (k) to the
where x(k) ∈ Rn , d(k) ∈ Rn and w(k) ∈ Rn are the local estimation error ei (k) are defined, respectively, as
state vector, bounded-power disturbance related to modeling
uncertainty/error, and process noise described by a white Si (z) = (zI − A + Li (z)CVi )−1 , (6)
Gaussian noise with mean zero and covariance Q ≥ 0, Hi (z) = −(zI − A + Li (z)CVi ) −1
Li (z). (7)
respectively. The matrix A ∈ Rn×n is constant and the initial
condition x(0) is assumed without loss of generality to be Here, Si and Hi are referred to as the system responses [24]
deterministic. The plant (1) is observed by M sensors with of the error dynamics.
the following model By the definitions (6) and (7), it is obvious that Si and
Hi are strictly proper. Moreover, it is easy to see that the
yi (k) = Ci x(k) + vi (k), i = 1, . . . , M, (2) system responses Si and Hi satisfy
where yi (k) ∈ Rpi is the local measurement and vi (k) ∈ Rpi Si (z)(zI − A + Li (z)CVi ) = I,
is the measurement noise, described by a white Gaussian
Si (z)(zI − A) − Hi (z)CVi = I.
noise with zero mean and covariance Ri > 0. The matrix
Ci ∈ Rpi ×n is constant and we also assume that the signals Then, since the estimation error needs to be finite, Si and
d(k), vi (k) and w(k) are mutually uncorrelated. Hi must be stable. Therefore, the following lemma can be
For the sensor network consisting of M nodes, let V = stated.
{1, . . . , M } be the set of nodes and E ⊂ V × V be the set of

4087
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on September 27,2023 at 07:41:19 UTC from IEEE Xplore. Restrictions apply.
Lemma 1 The transfer matrices (Si , Hi ) with finite error |z| > 1. Then φi (k) can be alternatively expressed as
k
can be induced by a filter with the structure (3) if and only P
φi (k) = si (j)d(k − j), which yields
if the following conditions hold j=1
 
  zI − A
Si Hi
−CVi
= I, Si , Hi ∈ RH2 . (8) kSi (q)d(k)k2P
N k
1 XX 0
Moreover, the desired local filter can be implemented as = lim d (k − j)s0i (j)si (j)d(k − j)
N →∞ N
x̂i (k) = −Hi (q)yVi (k). j=1
k=1
Proof: It is straightforward to apply the result in [25] N k
1 XX
to the distributed case. = lim Tr {s0i (j)si (j)d(k − j)d0 (k − j)}
N →∞ N
Lemma 1 gives a system level parameterization for dis- j=1
k=1
tributed filtering, which provides an alternative to the well- N k
1 XX
known Youla-Kucěra parameterization and characterizes the ≤ lim Tr {s0i (j)si (j)} Tr {d(k − j)d0 (k − j)}
N →∞ N
broadest known class of constrained linear control/filtering j=1
k=1
problem with a convex formulation. Such approach has been N k
1 XX
appealed for the study of connection of machine learning and = lim Tr {s0i (j)si (j)} kd(k − j)k22
N →∞ N
control [26], [27]. Interested reader may refer to [24] for a j=1
k=1
comprehensive introduction of the system level approach. N
1 XX
k
With Lemma 1, the cost functional (5) can be further ≤kdk2∞ lim Tr {s0i (j)si (j)}
N →∞ N
developed as follows. j=1k=1
∞
X
Theorem 1 Given the transfer matrices Si and Hi satisfy- =kdk2∞ Tr {s0i (j)si (j)}
ing (8), then the cost functional (5) is bounded by j=1

 1  2 1 2 ≤δ 2 kSi k22 ,
Ji ≤ Si Q 2 δSi + Hi RV2 , (9)
2 i 2 where the first inequality follows by the fact that Tr{ΦΨ} ≤
where δ is such that kdk∞ ≤ δ. Tr{Φ}Tr{Ψ} if Φ and Ψ are positive semi-definite with
appropriate dimensions [28]. Therefore, the condition (9)
Proof: According to Lemma 1 and the definition of holds, immediately
Si , Hi , the local error ei (k) in (4) can be recast as By Theorem 1, it is easy to see that Problem 1 can be
ei (k) = Si (q)w(k) + Hi (q)vVi (k) + Si (q)d(k). transformed into the following optimization
 2 1
 1
Since d(k), vVi (k) and w(k) are mutually uncorrelated, min Si Q 2 δSi 2 + kHi RV2i k22 ,
{Si ,Hi }
taking mean power norm on both sides leads to 
 zI − A

s.t.

Si Hi = I, (11)
kei (k)k2P = kSi (q)w(k)k2P + kHi (q)vVi (k)k2P (10) −CVi
Si , Hi ∈ RH2 .
+ kSi (q)d(k)k2P .
Note that the above optimization involves a semi-infinite
Denote the PSD of the signal Si (q)w(k) as Φei w (ejω ). There programming. In order to solve this problem efficiently, an
holds FIR approximation is adopted, and such truncation leads to
Φei w (ejω ) = Si (ejω )Φw (ejω )Si0 (ejω ), a polynomial solvability time [26]. The relative performance
loss induced by such approximation will be discussed in the
where Φw (ejω ) is the PSD of w(k). Hence, next section. The following algorithm gives a way to find
Z π
1 the solution to the optimization (11).
kSi (q)w(k)k2P = Tr Si (ejω )Φw (ejω )Si0 (ejω ) dω.

2π −π Algorithm 1 Distributed Robust Optimal Filter Design
Since w(k) is white Gaussian noise, Φw (ejω ) = Q. Thus, 1: Initialize (X, y, Z), ϕ0 , , set θ = θ0 > 0
1
Z π 2: Compute Infeasibility Measure ϕ
kSi (q)w(k)k2P = Tr Si (ejω )QSi0 (ejω ) dω

2π −π 3: while ϕ > ϕ0 and Tr{X 0 Z} >  do
1 4: Compute Search Direction (∆X, ∆y, ∆Z)
=kSi Q 2 k22 . 5: α = min{1, λmin (X−θ−1 ∆X) },
Similarly, the second term on the right side of (10) can be β = min{1, λmin (Z−θ−1 ∆Z) }
written as 6: Update X = X + α∆X, y = y + β∆y,
1 Z = Z + β∆Z
kHi (q)vVi (k)k2P = kHi R k22 . 2
Vi 7: Update θ = θ0 + 0.09 min{α, β}
Moreover, let φi (k) = Si (q)d(k) and si (k) be the kth 8: Update Infeasibility Measure ϕ
∞
9: end while
si (k)z −k for
P
spectral components of Si , i.e. Si (z) =
k=0 10: for l = 1, . . . , n do

4088
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on September 27,2023 at 07:41:19 UTC from IEEE Xplore. Restrictions apply.
11: D(:, l) = X(2+Ti (l −1)(qi +n) : lTi (qi +n)+1, 1) also denote si? (k) and hi? (k) as the kth spectral components
12: end for of Si? and Hi? , respectively. The following theorem presents
k = 1,. . . , Ti do
13: for  an optimality gap between the problems (12) and (13).
si (k)
14: = D(1 + (k − 1)(qi + n) : k(qi + n), 1 : n)
hi (k) Theorem 2 Consider Υi? ≥ 1 and 0 ≤ ρi? < 1 such
15: end for
that max{ksi? (k)kF , khi? (k)kF } ≤ Υi? ρki? . For each sensor
In Algorithm 1, Ti is the FIR approximation length asso- node i, i = 1, . . . , M , if
ciate with the node i, θ is the step length parameter, µ and
(kAk2 + kCVi k2 )Υi?
 
1
 represent sufficiently small positive numbers, and y and Z Ti ≥ lg , (14)
are the variables of the dual problem of the optimization (11). 1 − ρi? αi
The infeasibility measure (from Line 2 to Line 8) can be with αi being arbitrary such that 0 < αi < 1, then the
obtained from the system data A, CVi , Q, RVi , Ti , δ, and a relative estimation performance loss satisfies
specific solvability method for infeasibility measure is found
in [29]. In Line 6, if the minimum eigenvalue in either 
fi − fi? 1
expression is positive, then we take α = 1 (β = 1). The ≤ (kAk2 + kCVi k2 )Υi? ρTi?i (15)
implementation of the local filter can be realized as follows fi? 1 − αi
δ 2 Υ2i? ρ2i?

Ti + .
fi? (1 − ρ2i? )
X
x̂i (k) = − hi (j)yVi (k − j).
j=1 Ti Ti
si? (k)z −k , Hi = hi? (k)z −k ,
P P
IV. R ELATIVE E STIMATION P ERFORMANCE L OSS Proof: Set Si =
k=1 k=1
Ti
It is obvious that both the bounded-power disturbance
v̂i (k)z −k , where
P
and V̂i =
related to modeling error and the finite truncation intro- k=0
duced in Algorithm 1 for improving efficiency of problem- 
solving affects the estimation performance. In this section, −si? (Ti )A − hi? (Ti )CVi , if k = Ti ;
v̂i (k) =
the difference between the proposed optimization (11) and 0, Otherwise.
its counterpart without modeling error and approximation
It is easy to verifythat Si , Hi and V̂i satisfy the condition
(equivalently, the conventional distributed Kalman filtering)   zI − A
is explored, and an optimality gap is explicitly derived in Si Hi = I +V̂i , which indicates that the choice
CVi
terms of the size of imperfect modeling bound and the FIR of Si and Hi satisfies the first constraint of (12). Note that
truncation length. si? (k) and hi? (k) are the kth spectral components of Si?
To this end, for i = 1, . . . , M , we turn the optimization and Hi? , which are the optimal solution of the problem (13),
problem (11) into hence it is straightforward to see Si , Hi ∈ RH2 . Thus, the
second constraint of (12) is also satisfied. Moreover, we have
(
1  1  2
min min Si Q 2 δSi 2
γi ∈[0 1) {Si ,Hi } 1 − γi Ti
X
) kV̂i k∞ ≤ kv̂i (k)k2 ,
1 1
2
+ kHi RVi k2
2 k=0
1 − γi (12) ≤ (ksi? (Ti )k2 kAk2 + khi? (Ti )k2 kCVi k2 ),
 
  zI − A
s.t. Si Hi = I + V̂i , ≤ (ksi? (Ti )kF kAk2 + khi? (Ti )kF kCVi k2 ),
CVi
Si , Hi ∈ RH2 , ≤ (kAk2 + kCVi k2 )Υi? ρTi?i ,
kV̂i k∞ ≤ γi , where the third inequality follows by the relation between
where V̂i is a relaxation variable to capture the gap induced Frobenius norm and 2 norm. Then, by (14) and the fact that
by finite truncation [27]. For comparison convenience, we lg(ρi? ) ≤ ρi? − 1, there holds
adopt the similar form as (11) to obtain an optimal distributed 
(kAk2 + kCVi k2 )Υi?

filtering problem without bounded-power disturbance and −Ti lg(ρi? ) ≥ lg ,
αi
finite truncation as follows
1 1 which further implies
min kSi Q 2 k22 + kHi RV2i k22 ,
{Si ,Hi } 
 zI − A
 (kAk2 + kCVi k2 )Υi? ρTi?i ≤ αi < 1.

s.t. Si Hi = I, (13)
C Vi Hence, Si , Hi , and V̂i satisfy the third constraint of (12),
Si , Hi ∈ RH2 . with γi being chosen as γi = (kAk2 + kCVi k2 )Υi? ρTi?i .
Moreover, let fi be the optimal cost of the problem (12), Therefore, the choice of Si , Hi and V̂i verifies all constraints
fi? be the optimal cost of the problem (13), and {Si? , Hi? } of the problem (12), and the resulting objective function is
be the optimal solution to the problem (13), respectively. We finite and not smaller than fi , since fi is the optimal cost

4089
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on September 27,2023 at 07:41:19 UTC from IEEE Xplore. Restrictions apply.
 3
2 DKF
3
5
DHGF
Algorithm 1
1 7 2.5

6 4 2

AMSD
8
9
1.5

10
1

Fig. 1. Sensor Network With 10 Nodes.

0.5
0 50 100 150
step
of this problem. Hence,
Fig. 2. AMSDs under different filtering schemes
1  1 1

fi ≤ kSi Q 2 k22 + kHi RV2 k22 + kSi δk22 TABLE I
1 − γi i

1 AVERAGE AND S TANDARD D EVIATION FOR D IFFERENT F ILTERS

≤ (fi? + kSi k22 δ 2 ) Average Standard Deviation
1 − γi
Ti
DKF 1.6497 0.5491
1  X  DHGF 1.5539 0.5064
= fi? + ksi (k)k2F δ 2 Algorithm 1 1.3885 0.4075
1 − γi
k=1
Ti
1  X 
CDF
≤ fi? + δ 2 Υ2i? ρ2k
i?
1
1 − γi DKF
k=1 0.9 DHGF
2 2  Algorithm 1
1  Υ ρ 0.8
≤ fi? + δ 2 i? i? ,
1 − γi 1 − ρ2i? 0.7

where the second inequality holds by truncation, and the 0.6

F(AMSD)

equality follows by the Parseval’s theorem. Hence, 0.5

fi − fi? 1 1  δ 2 Υ2 ρ2  0.4
i? i?
≤ −1+ , 0.3
fi? 1 − γi fi? (1 − γi ) 1 − ρ2i?
 δ 2 Υ2 ρ2  0.2
γi 1 i? i?
≤ + , 0.1
1 − αi fi? (1 − αi ) 1 − ρ2i?
0
0 0.5 1 1.5 2 2.5 3 3.5
since γi ≤ αi . Therefore, straightforward computation leads AMSD
to the condition (15).
Theorem 2 explicitly characterizes the impacts of Fig. 3. Empirical CDFs of AMSD for different filtering schemes
bounded-power disturbance (modeling uncertainty/error), re-
lated to the terms with δ, and the finite truncation, related
to the terms with Ti , on the relative estimation performance with zero initial condition, Q = Diag{0.1, 0.7}, and d(k) =
loss. When the modeling uncertainty is removed, i.e., d(k) = 0.5|sin(0.2k)|. The variances of measurement noises vi (k)
0, the following neater expression of the relative estimation are set to be σi2 with σ1 = 0.9, σ2 = 1.2, σ3 = 0.5, σ4 = 1.3,
performance loss can be further obtained σ5 = 0.6, σ6 = 1.5, σ7 = 0.8, σ8 = 2.7, σ9 = 1, and
 
fi − fi? 1 Ti σ10 = 0.7. Let δ = 0.5 and the FIR length be Ti = 16, for
≤ (kAk2 + kCVi k2 )Υi? ρi? ,
fi? 1 − αi i = 1, . . . , 10.
which implies that that the relative performance loss induced To evaluate the estimation performance, we adopt the
by the finite truncation decreases exponentially with the following average mean square deviation (AMSD)
increase of truncation size Ti . L N
1 X X (l) (l)
V. N UMERICAL I LLUSTRATION AMSD(k) := kx (k) − x̂i (k)k2 ,
LN i=1 l=1
In this section, we present an example to show the
effectiveness of the proposed method. A sensor network with where l denotes the lth Monte Carlo run.
10 nodes is shown in Fig. 1. Consider the following system Set the initial condition of the local filters as x̂i (0) =
data for the plant (1) and sensors (2) [0, 1.5]0 and L = 150. Figure 2 exhibits a comparison
     among AMSDs obtained by distributed Kalman filter (DKF),
0.8 −0.1 1 0 i is odd; distributed H∞ Gaussian filter (DHGF) [11] and Algorithm
A= , Ci = 
0.12 0.25 0 1 i is even, 1, and the corresponding averages and standard deviations

4090
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on September 27,2023 at 07:41:19 UTC from IEEE Xplore. Restrictions apply.
 TABLE II
[7] H. Gao, J. Lam, L. Xie, and C. Wang, “New approach to mixed
E FFECTS OF D ISTURBANCE A MPLITUDE ON R ELATIVE E STIMATION H2 /H∞ filtering for ploytopic discrete-time systems,” IEEE Trans-
P ERFORMANCE L OSS actions on Signal Processing, vol. 53, pp. 3183–3191, 2005.
Disturbance Amplitude 0.5 0.55 0.6 [8] H. Huang and G. Feng, “Delay-dependent H∞ and generalized H2
filtering for delayed neural networks,” IEEE Transactions on Circuits
Relative Loss Bound 80.32% 97.12% 115.65% and Systems I: Regular Papers, vol. 56, no. 4, pp. 846–857, 2009.
(f1 − f1? )/f1? 65.42% 78.23% 92.09% [9] Y. Feng, X. Chen, and Y. Tan, “An improved paradigm for robust
optimal filtering over lossy networks,” in Proceedings of the 59th IEEE
Conference on Decision and Control, Jeju Island, Republic of Korea,
2020, pp. 4336–4341.
are listed in Table I. It is observed that the present method [10] R. Olfati-Saber, “Kalman-consensus filter: Optimality, stability, and
performance,” in Proceedings of the 48th IEEE Conference on Deci-
outperforms DKF and DHGF. We also plot the empirical sion and Control, Shanghai, China, 2009, pp. 7036–7042.
cumulative distribution functions (CDFs) of the AMSD for [11] Y. Feng and Z. Chen, “A game approach to distributed robust optimal
the aforementioned filters in Fig. 3, where F (AMSD) := filtering with consensus through multiple sensors: Theory and appli-
cation,” IEEE Transactions on Industrial Electronics, vol. 67, no. 11,
Prob{AMSD ≤ a} with a ≥ 0. It is also shown that the pp. 9693–9702, 2020.
proposed result achieves the best estimation performance. [12] Y. Feng, Z. Chen, and X. Chen, “Distributed H∞ Gaussian consensus
We also verify the relative estimation performance loss. filtering for discrete-time systems over lossy sensor networks,” SIAM
Journal on Control and Optimization, vol. 58, no. 1, pp. 34–58, 2020.
Take Node 1 as example. Choose Υ1? = 2.5, and ρ1? = 0.57. [13] S. Lzumi and X. Xin, “Distributed spatial filtering over networked
Table II shows the influence of the amplitude of the bounded- systems,” IEEE Control Systems Letters, vol. 5, no. 2, pp. 617–622,
power disturbance on the relative estimation performance 2021.
[14] S. Battilotti, F. Cacace, and M. d’Angelo, “A stability with optimality
loss, where the second row stands for the upper bound of analysis of consensus-based distributed filters for discrete-time linear
the loss, i.e. the right side of (15), and the third row is systems,” Automatica, vol. 129, 109598, 2021.
the true value of the loss. It is seen intuitively that bigger [15] R. Olfati-Saber, “Distributed Kalman filter with embedded consensus
filters,” in Proceedings of the 44th IEEE Conference on Decision and
the disturbance amplitude is, higher the relative estimation Control, Seville, Spain, 2005, pp. 8179–8184.
performance loss is. Similarly, a smaller truncation length [16] F. Cattivelli and A. Sayed, “Diffusion strategies for distributed Kalman
leads to a larger estimation performance loss. For example, filtering and smoothing,” IEEE Transactions on Automatic Control,
vol. 55, pp. 2069–2084, 2010.
when the truncation length is reduced by half, with the [17] S. Das and J. Moura, “Distributed Kalman filtering with dynamic
amplitude of the bounded-power disturbance 0.5, the upper observations consensus,” IEEE Transactions on Signal Processing,
bound of the relative estimation performance loss increases vol. 63, no. 17, pp. 4458–4473, 2015.
[18] W. Yang, C. Yang, H. Shi, L. Shi, and G. Chen, “Stochastic link
to 88.12%. activation for distributed filtering under sensor power constraint,”
Automatica, vol. 75, pp. 109–118, 2017.
VI. C ONCLUSION [19] S. Battilotti, F. Cacace, M. d’Angelo, and A. Germani, “Distributed
Kalman filtering over sensor networks with unknown random link
The problem of distributed filter design for discrete-time failures,” IEEE Control Systems Letters, vol. 2, pp. 587–592, 2018.
systems with both bounded-power disturbances and white [20] Y. Feng, Y. Tan, G. Gu, and X. Chen, “Remote state estimation with
noises is discussed in this paper. The estimation performance enhanced robustness in the presence of data packet dropouts,” IEEE
Transactions on Automatic Control, DOI: 10.1109/TAC.2021.3130886.
is characterized within the system level synthesis framework [21] Y. Feng, K. Dai, and X. Chen, “Enhanced distributed consensus
by system responses of the error dynamics and parameters filtering over lossy wireless sensor networks with application to
of disturbances/noises. A computationally tractable algorithm target speed tracking,” IEEE Transactions on Industrial Electronics,
accepted.
is proposed for filter gains design. The relative estimation [22] D. Ding, Z. Wang, H. Dong, and H. Shu, “Distributed H∞ state es-
performance loss is also derived in terms of modeling timation with stochastic parameters and nonlinearities through sensor
uncertainty and length of FIR approximation. One of our networks: The finite-horizon case,” Automatica, vol. 48, no. 8, pp.
1575–1585, 2012.
future research topics is to extend the current results into the [23] L. Zhang, Z. Ning, and Z. Wang, “Distributed filtering for fuzzy time-
case of unreliable communication networks. delay systems with packet dropouts and redundant channels,” IEEE
Transactions on Systems, Man, and Cybernetics: Systems, vol. 46,
R EFERENCES no. 4, pp. 559–572, 2016.
[24] J. Anderson, J. Doyle, S. Low, and N. Matni, “System level synthesis,”
[1] K. Zhou, J. Doyle, and K. Glover, Robust and Optimal Control. Annual Reviews in Control, vol. 47, pp. 364–393, 2019.
Prentice-Hall, 1996. [25] Y. Wang, S. You, and N. Matni, “Localized distributed Kalman filters
[2] P. Khargonekar, M. Rotea, and E. Baeyens, “Mixed H2 /H∞ filtering,” for large-scale systems,” IFAC-PapersOnLine, vol. 48, no. 22, pp. 52–
International Journal of Robust and Nonlinear Control, vol. 3, pp. 57, 2015.
313–330, 1996. [26] S. Dean, H. Mania, N. Matni, B. Recht, and S. Tu, “On the sample
[3] X. Chen and K. Zhou, “H∞ Gaussian filter on infinite time horizon,” complexity of the linear quadratic regulator,” Foundations of Compu-
IEEE Transactions on Circuits and Systems-Part I: Fundamental tational Mathematics, vol. 20, no. 4, pp. 633–679, 2020.
Theory and Applications, vol. 49, no. 5, pp. 674–679, May 2002. [27] S. Fattahi, N. Matni, and S. Sojoudi, “Efficient learning of distributed
[4] D. Grignion, X. Chen, N. Kar, and H. Qian, “Estimation of load linear-quadratic control policies,” SIAM Journal on Control and Op-
disturbance torque for DC motor drive systems under robustness and timization, vol. 58, no. 5, pp. 2927–2951, 2020.
sensitivity consideration,” IEEE Transactions on Industrial Electron- [28] R. Horn and C. Johnson, Matrix Analysis. Cambridge University
ics, vol. 61, no. 2, pp. 930–942, 2014. Press, 2012.
[5] Y. Feng, X. Nie, and X. Chen, “Robust optimal filtering over lossy [29] K. Toh, M. Todd, and R. Tütüncü, “SDPT3—a MATLAB software
networks,” IEEE Transactions on Automatic Control, vol. 65, no. 5, package for semidefinite programming, version 1.3,” Optimization
pp. 2272–2277, 2020. Methods and Software, vol. 11, no. 1-4, pp. 545–581, 1999.
[6] C. K. Ahn, Y. S. Shmaliy, and S. Zhao, “A new unbiased FIR filter
with improved robustness based on frobenius norm with exponential
weight,” IEEE Transactions on Circuits and Systems II: Express Briefs,
vol. 65, no. 4, pp. 521–525, 2018.

4091
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on September 27,2023 at 07:41:19 UTC from IEEE Xplore. Restrictions apply.