2582 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO.

5, MAY 2022

A Distributed Optimization Scheme for State Estimation of Nonlinear

Networks With Norm-Bounded Uncertainties
Peihu Duan , Qishao Wang , Zhisheng Duan , and Guanrong Chen

Abstract—This article investigates state estimation for a class of challenge. Generally, the existing robust estimation strategies can be
complex networks, in which the dynamics of each node is subject divided into two categories: the linear matrix inequalities (LMIs)-based
to Gaussian noise, system uncertainties, and nonlinearities. Based
on a regularized least-squares approach, the estimation problem is approach and the variance-constrained approach.
reformulated as an optimization problem, solving for a solution in Derived from the H∞ filtering for a single uncertain system, a large
a distributed way by utilizing a decoupling technique. Then, based number of LMIs-based techniques have been developed for complex
on this solution, a class of estimators is designed to handle the networks with various constraints, achieving an H∞ estimation per-
system dynamics and constraints. A novel feature of this design
lies in the unified modeling of uncertainties and nonlinearities, the
formance [7]–[12]. Earlier, Ding et al. [7] proposed an estimator for
decoupling of nodes, and the construction of recursive approxi- complex networks with randomly occurring sensor saturations, where
mate covariance matrices for the optimization problem. Further- the exponential mean-square stability of the estimation was ensured
more, the feasibility of the proposed estimators and the bound- under an augmented LMIs-based condition. Then, to reduce the data
edness of mean-square estimation errors are ensured under a transmission frequency, Wang et al. [9] embedded an event-triggered
developed criterion, which is easier to check than some typical
estimation strategies including the linear matrix inequalities-based algorithm into the estimator for complex networks in the presence of
and the variance-constrained ones. Finally, the effectiveness of the mixed delays and Gaussian noise. Later, randomly varying coupling
theoretical results is verified by a numerical simulation. and communication constraints in complex networks were handled by
Index Terms—Distributed state estimation, regularized least-
an asynchronous estimator in [10], with gains derived utilizing the
squares approach, stochastic complex network, uncertainty and LMIs method. Note that the approaches in [7]–[10] are suitable for
nonlinearity. time-invariant systems (or those with time-invariant estimation error
dynamics). Recently, to address the state estimation problem for time-
varying complex dynamical networks, recursive LMIs approaches were
I. INTRODUCTION designed on a finite horizon in [12], where several LMIs are needed to
Over the past two decades, complex networks have drawn increas- be solved to obtain the estimator gains at every step. In general, although
ing attention since they can model many physical systems, such as the LMIs-based approaches can handle the H∞ state estimation problem
cooperative unmanned aerial vehicles, networked manipulators, and for complex networks, they are limited by the following drawbacks:
multisatellite systems [1], [2]. One significant research issue is to 1) the LMIs are usually solved in an augmented form, which pays a
estimate the system states of nodes in complex networks by utilizing high computational price;
system models and measurements [3]–[6]. The difference between the 2) the LMIs have to be solved at every step for a time-varying system,
state estimation for a single system and that for a complex network which limits the online operations of the estimation algorithms.
lies in the coupling features between nodes, consequently, the latter is The variance-constrained approach is another efficient way to handle
much more complicated than the former. Moreover, there always exist the robust state estimation problem for complex networks [13]–[18].
disturbances, nonlinearities, and uncertainties in system models, thus In their pioneering work, Hu et al. [13] designed an augmented es-
how to design a robust estimator for such a complex network is a great timator for a class of time-varying coupled complex networks with
fading measurements by employing the Riccati-like difference equa-
tions technique. In this article, the estimation error covariance is upper
Manuscript received November 29, 2020; revised March 14, 2021;
bounded by a sequence of recursive matrices. Later, this scheme was
accepted May 30, 2021. Date of publication June 22, 2021; date of
current version April 26, 2022. This work was supported in part by the extended to cases with time-varying complex networks in [14] and
National Key R&D Program of China under Grant 2018AAA0102703, in uncertain inner coupling in [15], respectively. Note that the estimator
part by the National Natural Science Foundation of China under Grant gain matrices in [14] and [15] are determined in the compact form,
U1713223 and Grant 62003013, in part by the China Postdoctoral Sci- which are solved by utilizing the global information. However, in terms
ence Foundation under Grant BX20190025 and Grant 2019M660405,
and in part by the Hong Kong Research Grants Council under the GRF of computational complexity and operation efficiency, it is preferable to
Grant CityU11206320. Recommended by Associate Editor Prashant G. resolve the gains in a distributed way for each node [16]–[18]. For this
Mehta. (Corresponding author: Zhisheng Duan.) purpose, Li et al. [16] designed a nonaugmented estimator by solving
Peihu Duan and Zhisheng Duan are with the State Key Laboratory two Riccati-like difference equations, without the need of calculating
for Turbulence and Complex Systems, Department of Mechanics and
cross-covariance matrices between coupling nodes. In particular, two
Engineering Science, College of Engineering, Peking University, Beijing
100871, China (e-mail: [email protected]; [email protected]). inequality conditions on the Riccati-like equations had to be satisfied
Qishao Wang is with the Department of Mechanical and Engi- at every step, which limits the estimator’s applicability. Recently, a
neering Science, Beihang University, Beijing 100191, China (e-mail: boundedness analysis of the estimation error dynamics was presented
[email protected]). in [18], where a sufficient condition concerning system matrices was
Guanrong Chen is with the Department of Electrical Engineering, City
University of Hong Kong, Hong Kong (e-mail: [email protected]). introduced to ensure the estimator feasibility. However, this condition
Color versions of one or more figures in this article are available at was ineffective since it contains several parameters implicitly in the
https://doi.org/10.1109/TAC.2021.3091182. estimators to be carefully designed. Therefore, the estimators designed
Digital Object Identifier 10.1109/TAC.2021.3091182 by variance-constrained methods are subject to strict feasibility criteria.

0018-9286 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:30:56 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 5, MAY 2022 2583

Besides, the estimators in [14]–[18] failed to handle the case of system Lemma 2: For any given matrices P ∈ Rn×m and Q ∈ Rn×m , and
parameters with uncertainties, which usually exist in physical systems, a positive scalar β, one has
such as manipulators and artificial satellites. As a remedy, a penalization
approach was introduced to handle the parametric errors for the state (P + Q)T (P + Q) ≤ (1 + β)P T P + (1 + β −1 )QT Q.
estimation of networked systems in [19]. However, it was assumed that
Lemma 3: For any given matrices Pi ∈ Rn×m , i = 1, 2, . . ., N , one
every system matrix was first-order differentiable with respect to each
has
parametric error, and a central unit was needed to derive all estimator
N T  N 
gains. In this sense, the algorithm in [19] is not fully distributed.   N
This article focuses on the state estimation problem for complex Pi Pi ≤ N PiT Pi .
coupled networks with deterministic matrix uncertainties, Gaussian i=1 i=1 i=1

noise, and nonlinearities in each node’s dynamics. Following the idea

of regularized least-squares for a time-invariant single system [20] and B. Problem Statement
using a decoupling technique, the problem is reformulated as an opti- Consider a class of discrete-time uncertain nonlinear networks con-
mization problem such that the disturbances and nonlinearities can be sisting of N nodes (agents) as follows:
simultaneously modeled under a unified framework. Then, based on the

measurements of each node, a novel class of estimators is presented with xk+1,i = f (xk,i ) + a πij g(xk,j ) + ωk,i
gains solved in a distributed way. Moreover, a boundedness analysis of j∈Ni
the estimator parameters and mean-square errors (MSEs) is presented.
Compared to [20], this article faces two main technical challenges: yk,i = hi (xk,i ) + νk,i , i = 1, . . . , N
1) the introduction of the system nonlinearity will greatly increase
where xk,i ∈ Rn is the system state of the node i at step k, yk,i ∈ Rr
the difficulty of the robust state estimation problem since the
is the measurement at the step k, a is a positive constant, πij is the
parametric errors and the linearization errors are heterogeneous;
coupling strength between nodes with πij = 0 if j ∈ Ni (the set of node
2) the coupling in the networked system makes the derivation of the
i’s neighbors, including the node i itself), otherwise πij = 0, ωk,i ∈ Rn
estimator gains more difficult, where the information of neighbors
is the process white Gaussian noise with covariance Qk,i > 0, while
needs to be evaluated accurately and utilized effectively.
νk,i ∈ Rr is the measurement white Gaussian noise with covariance
Moreover, the analysis of the estimation performance, together with
Rk,i > 0. Moreover, ωk,i and νk,i are uncorrelated for all i = 1, · · · , N ,
the feasibility of the estimator, is particularly challenging.
k = 1, 2, . . . , and f (·), g(·) and hi (·) are nonlinear system functions.
The main contributions of this article are threefold. First, the pro-
In some physical systems, such as cooperative unmanned aerial ve-
posed estimators can deal with time-varying nonlinear system dynamics
hicles and multimanipulator systems, the exact processing dynamics is
with norm-bounded system matrix uncertainties, more general than
difficult or even impossible to know due to imprecise system parameters
those in the literature. Second, by using a decoupling technique, the
and structures. This likely gives rise to the inaccuracy in f (·), g(·),
estimator gains are solved for each node individually, i.e., in a dis-
and hi (·) with the modeling errors being bounded by positive definite
tributed way rather than in an augmented form as the LMIs-based
matrices [22], [23]. Under such situations, taking f (·) for example, it
approaches [6]–[12]. Thus, the proposed estimators have the superiority
is more feasibly described by
in terms of the computational complexity. Third, based on a rigorous
boundedness analysis of MSE, one criterion is derived to ensure the f (xk,i ) = fˆ(xk,i ) + Δfk,i xk,i
feasibility of the estimators, which is easier to check than those from
variance-constrained approaches [14]–[17]. where fˆ(·) is the known nominal system function, Δfk,i xk,i is the
The rest of this article is organized as follows. In Section II, some structured uncertain term with Δfk,i = E1,k,i Δ1,k,i , in which E1,k,i
notations, useful lemmas, and the problem statement are presented. In is a known matrix and Δ1,k,i 2 ≤ 1, representing the structure and
Section III, a class of robust state estimators for complex networks is amplitude of the uncertainty, respectively. In the following, E1,k,i plays
designed in a distributed fashion, respectively. Both the feasibility of the a key role in restraining system perturbation and ensuring estimation
estimators and the estimation performance are analyzed. In Section IV, performance. Although such a model cannot characterize all uncertain
a numerical simulation is provided to verify the effectiveness of the nonlinear systems, it can describe a large number of physical plants,
theoretical results. Section V concludes this article. such as Lur’e systems and all those satisfying the Lipschitz condition.
In this article, the developed method for dealing with the nonlinearity
II. PRELIMINARIES AND PROBLEM STATEMENT and uncertainty of f (·) can be directly extended for g(·) and hi (·).
Hence, for simplicity, g(xk,i ) and hi (xk,i ) are assumed to be linear
A. Notations and Useful Lemmas
as Gxk,i and Hi xk,i , respectively, with known matrices G and Hi .
Notation: P T , P 2 , and P −1 stand for the transpose, 2-norm and On the basis of the aforementioned discussion, the system dynamics is
inverse of the matrix P , respectively. P > Q means that P − Q is a modified to be
positive definite matrix. For a matrix P ∈ Rn×n and a vector q ∈ Rn×1 , 
q2P denotes q T P q. diag{·} represents a block-diagonal matrix. E{·} xk+1,i = fˆ(xk,i ) + Δfk,i xk,i + a πij Gxk,j + ωk,i (1)
is the mathematical expectation of a random variable. Tr(·) denotes the j∈Ni

trace of a matrix. In is the n × n identity matrix. 0n is the n × n zero yk,i = Hi xk,i + νk,i , i = 1, . . . , N. (2)
matrix. σmax {·} denotes the maximum singular value of a matrix. SλRe (·)
Re
is the set of eigenvalues’ real parts of a square matrix, and S λ (·) is the To deal with the state estimation problem for the aforementioned
set of the absolute values of elements in SλRe (·). nonlinear networks and measurements containing noise, uncertainty,
Lemma 1 [21]: For any given matrices P ∈ Rn×n , Q ∈ Rm×m , and nonlinearity, some methods have been developed [14], [16], [24],
and S ∈ Rm×n , if P −1 and Q−1 exist, then which are summarized as follows.
First, denote the priori estimate and the posteriori estimate by xk,i
(P + S T QS)−1 = P −1 − P −1 S T (Q−1 + SP −1 S T )−1 SP −1 . and x̂k,i , respectively. Here, by utilizing previous estimates and current

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:30:56 UTC from IEEE Xplore. Restrictions apply.
2584 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 5, MAY 2022

measurements, the structure for estimates at the current step is designed III. MAIN RESULTS
as
In this section, a novel state estimation framework is introduced for

xk+1,i = fˆ(x̂k,i ) + a πij Gx̂k,j (3) dynamic networks (1) subject to Gaussian noise, nonlinearities, and
j∈Ni uncertainties. The boundedness of the estimator gain is analyzed. The
estimation performance is theoretically evaluated.
x̂k+1,i = xk+1,i + Kk+1,i (yk+1,i − Hi xk+1,i ) (4)

where Kk+1,i is the estimator gain to be optimized. This structure A. Design of Distributed State Estimators
is proposed for good reasons. First, the priori estimate is designed In this subsection, inspired by the regularized least-squares problem
by following the same structure as that of the plant. Then, this value with uncertainties discussed in [20], the state estimation problem for
is modified to obtain the posteriori estimate by using the innovation the uncertain nonlinear networks (1) is investigated in the following
error of the node i. If innovation errors of neighbors are adopted to distributed form.
develop the estimator, then neighbors’ information of every neighbor is First, introduce an approximate covariance matrix for the node i at
needed to derive the estimator gain of the node i, which consumes step k + 1 as
much more communication resources. Besides, the introduction of
such global information leads to complex coupled terms, making it −1
Pk+1,i = (1 + a)(P̆k+1,i + HiT R̂k+1,i
−1
Hi + 2λk,i In )−1 + (a + a2 )
practically impossible to solve the estimator gain in a distributed way. 
Then, define the estimator error at step k as ek,i = xk,i − x̂k,i . Note ×(N − 1) 2
πij −1
G(Pk,j 2
+ πij GT HiT Zk,ij Hi G)−1 GT (7)
that the normal nonlinear function satisfies j∈Ni ,j=i

fˆ(xk,i ) − fˆ(x̂k,i ) = Fk,i ek,i + O(ek,i ) (5) with

ˆ
where Fk,i = ∂ f∂x
(x)
|x̂k,i and O(ek,i ) represents the high-order terms. P̆k+1,i = (Fk,i + aπii G)Pk,i (Fk,i + aπii G)T + Qk,i
Furthermore, O(ek,i ) can be described by [14], [17] −1
R̂k+1,i = (1 + a)−1 Rk+1,i − λk,i Hi Ek,i Ek,i
T
HiT
O(ek,i ) = E2,k,i Δ2,k,i ek,i (6)
2
Zk,ij = πij (a + a2 )(N − 1)Rk+1,i
−1
(8)
where E2,k,i is a known matrix and Δ2,k,i is the uncertain matrix that
satisfies Δ2,k,i 2 ≤ 1. where Ek,i = [E1,k,i , E2,k,i ], and parameters λk,i and αk,i ∈ [0, 1]
Based on the aforementioned process for dealing with nonlinearities, are two scalars to be further designed in the last part of this subsection.
the dynamics of the estimation error at step k + 1 can be derived as In this article, Pk,i and P̆k,i are not real estimation posterior and
  priori error covariance matrices, respectively. Instead, they are iterated
ek+1,i = I − Kk+1,i Hi Δfk,i xk,i + (Fk,i + E2,k,i Δ2,k,i )ek,i similarly to those of the standard Kalman filter, in order to act as
  the real covariance matrices to evaluate the posterior and priori error
+ ωk,i + a πij Gek,j − Kk+1,i νk+1,i . statistical characteristics. In the following subsections, it will be proved
j∈Ni that they indeed guarantee the boundedness of the estimator gains and
the estimation performance.
Now, define MSE at step k + 1 for the node i as M (ek+1,i ) =
Then, design the following distributed estimator:
E{eTk+1,i ek+1,i } and an auxiliary matrix as P (ek+1,i ) =
E{ek+1,i eTk+1,i }. It is worth noting that M (ek+1,i ) = Tr(P (ek+1,i )). x̂k+1,i = (1 − αk,i )[xuk+1,i + K1,k+1,i buk+1,i ] + αk,i xk+1,i
Then, the estimator gains Kk+1,i , i = 1, . . . , N , are optimized by ⎡ ⎤
 
{Kk+1,i }opt = arg min Tr(P (ek+1,i )) ∀i ∈ N + ⎣αk,i K2,k+1,ij bk+1,i πij ⎦ (9)
Kk+1,i
j∈Ni ,j=i j∈Ni ,j=i

Since P (ek+1,i ) contains the error-estimate coupling term ek,i x̂Tk,i , the
linearization error E2,k,i Δ2,k,i ek,i , and the uncertain term Δfk,i xk,i , where
some matrix inequality techniques have to be employed. For example, 
in [16] and [17], Young’s inequality and a quadratic matrix inequality xk+1,i = fˆ(x̂k,i ) + a πij Gx̂k,j
j∈Ni
are introduced to provide a deterministic upper bound of the MSE,
which is generally too conservative. Thus, the estimation performance is xuk+1,i = xk+1,i − λk,i (Fk,i + aπii G)Pk,i x̂k,i /(1 − αk,i )
degraded. Moreover, in those methods, some complex conditions have
to be checked to ensure the estimator feasibility at every estimation step. bk+1,i = yk+1,i − Hi xk+1,i
In [8] and [12], P (ek+1,i ), i = 1, . . . , N , are developed in compact
buk+1,i = yk+1,i − Hi xuk+1,i
forms, and the estimation gains Kk+1,i , i = 1, · · · , N , are optimized
−1
by solving several higher order LMIs. In this sense, these algorithms K1,k+1,i = (P̆k+1,i + HiT R̂k+1,i
−1
Hi + 2λk,i In )−1 HiT R̂k+1,i
−1

are centralized and the computational complexity is relatively high.

−1
To address the aforementioned limitations, this article focuses on the K2,k+1,ij = aπij G(Pk,j + GT HiT Zk,ij Hi G)−1 GT HiT Zk,ij .
following two problems. (10)
Problem 1: How to design a distributed estimator of low com-
putational complexity in the simultaneous presence of noise, system With the appropriate parameters λk,i and αk,i , the proposed state es-
uncertainties, and nonlinearities? timation algorithm for the uncertain nonlinear networks in a distributed
Problem 2: How to guarantee the feasibility of the proposed estimator sense is summarized as Algorithm 1. Due to the space limitation, the
and the estimation performance? derivation of the Algorithm 1 is given in [25].

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:30:56 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 5, MAY 2022 2585

system uncertainties, rigorous proofs for the feasibility of the estimators

Algorithm 1: State Estimation for the Node i at Step k + 1 by the
and the boundedness of MSE are provided in this article.
Distributed Method.
Proposition 1: The computational complexity of Algorithm 1 is
Initialization: The parameters λ0,i P̆0,i and P0,i are set as: O(|Ni |n3 + r3 ), where Ni is the neighbor set of the node i.
T T −1 Proof: Note that Algorithm 1 consists of three recursive equations
λ0,i = (1 + β)E0,i H0,i R0,i H0,i E0,i 2 , β > 0,
(7) and (9) with several auxiliary matrices. According to [26], the
0 < P̆0,i ≤ Q0 , computational complexity of (7) can be evaluated as O(|Ni |n3 + r3 ).
Similarly, the one of (9) is O(|Ni |nr). Besides, regarding the aux-
−1 T −1
P0,i = (P̆0,i + H0,i R̂0,i H0,i )−1 . iliary matrices in (10), the associated computational complexity is
O(|Ni |n3 + r3 ). Hence, the total computational complexity of Algo-
Input: yk+1,i , x̂k,j and Pk,j , j ∈ Ni ; rithm 1 is O(|Ni |n3 + r3 ). 
Output: x̂k+1,i and Pk+1,i ; Remark 3: Compared with the centralized algorithms in [14]
1: Exchange the package (x̂k,i , Pk,i ) with its neighbors; and [15], whose computational complexity is O(N 3 n3 + N 3 r3 ), Al-
2: Calculate the recursive matrix Pk+1,i by (7); gorithm 1 in this article shows great superiority of saving computational
3: Get the priori state estimates xk+1,i and xuk+1,i by (10); resources.
4: Compute the posterior state estimate x̂k+1,i by (9); In the following, an optimization approach is presented for deter-
mining parameters λk,i and αk,i in (9).
Theorem 1: Consider networks (1) with measurements (2). The
Actually, Algorithm 1 is designed based on the following optimiza- parameters λk,i and αk,i in estimator (9) can be designed by
tion problem:
Jˆ2,i (λk,i , αk,i )
opt opt
{αk,i , λk,i } = arg min (13)
αk,i ∈[0, 1]
min J1,i (xk,j , wk,i , xk+1,i , j ∈ Ni ) λk,i ≥H T T H2
xk,j
wk,i
 where
s.t. xk+1,i = fˆ(xk,i ) + Δfk,i xk,i + a πij Gxk,j + wk,i (11) T Tˆ
j∈Ni Jˆ2,i (λk,i , αk,i ) = λk,i E b,k,i E b,k,i + η opt opt
k,i S k,i η k,i

T
with − 2λk,i E b,k,i E a,k,i η opt opt 2
k,i + Ak,i η k,i − (1 − αk,i )bk+1,i  ˆ T k,i

J1,i = xk,i − x̂k,i 2P −1 + wk,i 2Q−1 + yk+1,i − Hi xk+1,i 2R−1 

2
k,i k,i k+1,i + (eopt
k,ij P −1 + Hi Geopt
k,ij − α̂k,i bk+1,i 2Zk,ij ),
k,j
 j∈Ni ,j=i
+ xk,j − x̂k,j 2P −1 (12)
k,j with
j∈Ni ,j=i

η opt ˆ T ˆ −1 T ˆ
where Pk,i is designed in (7), and Qk,i and Rk+1,i are given in k,i = [S k,i + Ak,i T k,i Ak,i ] [(1 − αk,i )Ak,i T k,i bk+1,i
Section II-B. It is worth mentioning that xk,j and wk,i in (11) are T
optimization variables of the problem, instead of the real state and the + λk,i E a,k,i E b,k,i ],
process noise. The structure of J1,i in (12) can be interpreted as follows. eopt −1 T T −1 T T
k,ij = α̂k,i [Pk,j + G Hi Zk,ij Hi G] G Hi Zk,ij bk+1,i
The first three parts of J1,i are optimized to obtain the priori estimate
according to the priori information, such as the system model and the and
noise covariance. Then, by adding the last term, the local measurement ⎛ ⎞
information is utilized efficiently to correct the priori estimate. Hence, 
α̂k,i = αk,i / ⎝ πij ⎠ , Ak,i = Hi [aπii G + Fk,i , In ]
the cost function attempts to balance the model prediction process and
j∈Ni ,j=i
the measurement feedback process.
Remark 1: The parameter αk,i is introduced to act as a regulatory −1
S k,i = diag[Pk,i , Q−1
k,i ], Tˆ k,i = R̂k+1,i
−1
indicator, balancing the weights of the innovation errors on the node
i and its neighbors. Specifically, the priori estimate of the node i is E a,k,i = [In In ]T [In 0n ], E b,k,i = −[In 0n ]T x̂k,i
corrected by (1 − αk,i )bk+1,i , while the ones of node i’s neighbors are
ˆ = S + λ ET E
S
corrected by αk,i bk+1,i /( j∈Ni ,j=i πij ). Compared with the existing k,i k,i k,i a,k,i a,k,i .
methods putting node i’s own and neighbors’ information together,
this decoupling design simplifies the solution process. Meanwhile, by Due to the space limitation, the proof of Theorem 1 is given in [25].
this technique, the impact of the whole network on the subsystems is Remark 4: Although Jˆ2,i (λk,i , αk,i ) in (13) is not convex with re-
decoupled onto the neighbors at each estimation step. Here, how to spect to λk,i and αk,i , some existing algorithms can be used to solve the
obtain the optimized αk,i is the key to guarantee the decoupling effect, problem, such as the quasi-Newton methods [27], [28]. Generally, the
which will be studied in the following part. solving methods for (13) are mature and their computational complexity
Remark 2: Compared with the estimation algorithm in [19], Algo- is acceptable.
rithm 1 possesses the following differences and advantages. First, the As shown in [29] and [30], one can set λk,i = (1 + β)λ̆k,i for
estimator gain of each node is developed based only on neighbors’ computational simplification, where β ∈ (0, 1), and
information, different from that in [19], where a central unit is needed λ̆k,i = (1 + a)−1 Ek,i
T
HiT Rk+1,i
−1
Hi Ek,i 2 . (14)
to collect global information. Second, every system matrix is required to
be first-order differentiable with respect to each parametric error in [19], For example, in [30], β = 0.5 usually generates a desirable estimation
while the system matrix uncertainties in this article are described by performance. For linear time-invariant systems, when the estimator
bounded parameters based on the graph theory. Third, in the presence of gains converge, λk,i tends to a fixed value. Thus, in such situations, λk,i

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:30:56 UTC from IEEE Xplore. Restrictions apply.
2586 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 5, MAY 2022

can be precomputed to achieve a satisfactory suboptimal estimation. Assumption 2: (Fk,i + aπii G, Hi ) is uniformly observable, i.e.,
opt
Further, the analytical expression of αk,i is given as follows. there exists a positive scalar κ and a positive integer N such that
opt
Corollary 1: If λk,i = (1 + β)λ̆k,i , then αk,i can be designed by

N −1

⎧ κI ≤ ΨTk+h−1,k HiT Rk+1,i

−1
Hi Ψk+h−1,k
⎪
⎨0 if Φ̂k,i < 0 h=0
opt
αk,i = Φ̂k,i if Φ̂k,i ∈ [0, 1] (15) where
⎪
⎩
1 if Φ̂k,i > 1 
(Fk+h−1,i + aπii G) · · · (Fk,i + aπii G) if h ≥ 1
Ψk+h−1,k =
with I if h = 0 .
⎛ ⎞
 For linear time-invariant systems, Assumption 2 is equivalent to that
Φ̂k,i = (Φ1k,i + 2Φ2k,i )/ ⎝2Φ2k,i +2 Φ3k,ij ⎠ (H, Fi + aπii G) is observable, where H = [H1T , . . ., HN T T
] .
j∈Ni ,j=i Assumption 3: [17], [32] The parameters in (1) are uniformly
bounded, i.e., there exist positive constants κF,1 , κF,2 , κE,1 , κE,2 ,
where κH , κω,1 , κω,2 , κν,1 , and κν,2 such that κF,1 ≤ Fk,i + aπii G2 ≤
κF,2 , G2 ≤ κG , E1,k,i 2 ≤ κE,1 , E2,k,i 2 ≤ κE,2 , Hi 2 ≤
T
Φ1k,i = − 3λk,i E b,k,i E a,k,i [Sˆ + AT Tˆ A ]−1 AT Tˆ b
k,i k,i k,i k,i k,i k,i k+1,i κH , κω,1 ≤ Qk 2 ≤ κω,2 , and κν,1 ≤ Rk,i 2 ≤ κν,2 ∀k = 1, 2,
−1 . . ., and ∀i = 1, . . ., N .
ˆ −1 AT )−1 b
Φ2k,i = bTk+1,i (Tˆ k,i + Ak,i S k,i k,i k+1,i Theorem 2: Consider networks (1) with measurements (2). If As-
⎛ ⎞ sumptions 1–3 hold and λk,i = (1 + β)λ̆k,i with λ̆k,i defined in (14),
 the recursive matrix Pk,i is uniformly upper and lower bounded, pro-
Φ3k,ij = bTk+1,i (Zk,ij
−1
+ Hi GPk,j GT HiT )−1 bk+1,i ⎝ πij ⎠ .
vided that
j∈Ni ,j=i
Re 1
max{S λ (G)} < (a + a2 )− 2 (N − 1)−1 πm
−1

Due to the space limitation, the proof of Corollary 1 is given in [25].

From (15), Φ1k,i is related to the level of system perturbations, while where πm = max{πij , i, j ∈ [1, . . . , N ]}.
Φ3k,ij is related to the confidence of node i’s neighbors’ information. If The proof of Theorem 2 is given in [25].
system perturbations are larger, which leads to a larger Φ1k,i , then αk,i opt
is Remark 7: Theorem 2 provides a sufficient condition for the stability
larger. Similarly, if the neighbors’ information is less confident, which of the proposed estimation algorithm. Note that, in [17] and [16], the
leads to a larger Φ3k,ij , then αk,i opt
is smaller. In this sense, αk,i acts as corresponding condition can be summarized as that one bound of MSE
a regulatory indicator to determine the weight of the residual error for must be smaller than a given (or chosen) positive number at each step.
each node’s own and its neighbors’ information. In particular, if there Instead, in this article, one only needs to check a condition at the
are no system uncertainties and neighbors, i.e., Δfk,i = 0 and a = 0, beginning of the estimation process, which shows its superiority in
then the estimation problem becomes the standard Kalman filtering terms of online solving the state estimation problem for applications.
problem. In such case, one can show that Φ1k,i = 0 and Φ3k,ij = 0, Assumption 4: The system (1) reduces to a class of uncertain lin-
opt ear time-invariant complex network, i.e., Fk,i = Fi , E1,k,i = E1,i ,
thus, αk,i = 1, which is consistent with the standard Kalman filter.
E2,k,i = E2,i , Qk,i = Qi , and Rk,i = Ri .
Remark 5: In this article, the homogeneity of f (·) and g(·) for each
Generally, for slowly varying systems with relatively fast-sampling
node is not needed for the derivation of Algorithm 1. Instead, it is
sensors, this assumption naturally holds.
required to guarantee the estimation performance by Algorithm 1 under
Corollary 2: Consider networks (1) with measurements (2). If
a much easier-to-check condition compared to the literature [14], [16],
Assumptions 1–4 hold, αk,i = αi , and λk,i = (1 + β)λ̆k,i , then the
and [17], which will be clarified in the following subsections. Even
recursive matrix Pk,i is convergent.
though the two functions are same for all nodes, their Jacobian matrices,
ˆ(x) The proof of Corollary 2 is given in [25]. Built on Theorem 2,
e.g., Fk,i = ∂ f∂x |x̂k,i , i = 1, . . . , N , are not identical due to different Corollary 2 reveals that the convergence of the estimator gains in (9) is
x̂k,i . In this sense, the model considered in this article is heterogeneous, ensured for time-invariant systems.
which is as general as the time-varying linear models adopted in [13]
and [19] and the nonlinear models adopted in [16] and [17].
C. Boundedness Analysis of the MSE
In this subsection, the boundedness of the MSE by the proposed dis-
B. Boundedness Analysis of Estimator Gains
tributed estimator is analyzed for the following two cases: 1) Δfk,i = 0;
In this subsection, the boundedness of the distributed estimator gains and 2) Δfk,i = 0.
in (9) is studied. The key issue is to analyze the uniform boundedness First, consider the case of Δfk,i = 0, i.e., the uncertainty in the
of Pk,i , P̆k,i, and P̂k,i . Without losing generality, it suffices to study system function is omitted. The corresponding result on the estimation
the one of Pk,i for any k. performance by the distributed estimator (9) is summarized as follows.
Assumption 1: The matrix Fk,i + aπii G is nonsingular. Theorem 3: Consider networks (1) with Δfk,i = 0 and measure-
Remark 6: This assumption is automatically satisfied if the model ments (2). Then, the distributed estimator (9) is unbiased. Furthermore,
and its Jacobian matrix Fk,i + aπii G are obtained by the discretization suppose that Assumptions 1–4 hold and λk,i = (1 + β)λ̆k,i . Then, the
of a continuous-time system [31]. Besides, this assumption can be MSE for the node i is ultimately upper bounded if
0 0
relaxed to be that the matrix Fk,i is nonsingular, where Fk,i is a   −1
matrix in the “neighborhood” of Fk,i + aπii G, i.e., Fk,i + aπii G − HF 2 < 1 + κ−1 −1
F,1 κE,2 + aπm κF,1 κG N (N − 1) (16)
0
Fk,i 2 ≤ 1 with 1 being a small positive number. In Assumption
0
1, Fk,i + aπii G instead of Fk,i is chosen just for readability and where HF = (IN n + (P̆ −1 + λ)−1 H T R̂−1 H)−1 F , with P̆ = (1 +
simplicity. a)F (P̆ −1 + H T R̂−1 H + λ)−1 F T + Q, F = diag[F1 , . . ., FN ], H =

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:30:56 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 5, MAY 2022 2587

diag[H1 , . . ., HN ], R̂ = diag[R̂1 , . . ., R̂N ], Q = diag[Q1 , . . ., QN ],

and λ = diag[λ1 In , . . ., λN In ].
The proof of Theorem 3 is given in [25].
Remark 8: Theorem 3 provides a condition for guaranteeing the
estimation performance. Actually, by some calculations, this condition
can be further relaxed as κF,2 + aπm κG ≤ 2 with 2 being a positive
scalar in terms of N , κE,2 , etc. This equivalently requires that the
observed system should evolve slowly or the sensors should have the
ability of fast sampling, which is consistent with the intuition. Com-
pared with the recursive methods used in [14] and [16], the condition
for the aforementioned algorithm only needs to be justified once, while
the ones in [14] and [16] contain some parameters that must be carefully
chosen at every step, e.g., αk,i and βk,i in [16].
When it comes to the case of Δfk,i = 0, the following assumption
is needed.
Assumption 5: Δfk,i xk,i is uniformly upper bounded, i.e., there
exists a positive scalar κu such that Δfk,i xk,i 2 ≤ κu .
Corollary 3: Consider networks (1) with measurements (2). If the
conditions of Theorem 3 and Assumption 5 hold, the MSE for the node
i is ultimately upper bounded.
The proof of Corollary 3 can be performed by simply combining the
one in the aforementioned Theorem 3 with the one in [29, Th. 3].
Fig. 1. States and estimates of each node. (a) node 1. (b) node 2. (c)
node 3. (d) node 4.
IV. SIMULATION EXAMPLE
In this section, the theoretical results are visualized by a comparative
simulation, where a discrete time-varying complex coupled network of
four nodes is considered. Similarly to the simulation model considered
in [14], the system nonlinear function in (1) is set as
 
0.9xk,i (1) + sin(0.5xk,i (2))
f (xk,i ) = (17)
0.9xk,i (2) − sin(0.5xk,i (1))

with xk,i = [xk,i (1), xk,i (2)]T , and the measurement matrices are
given as
Hk,1 = [0.90 0.25], Hk,2 = [0.95 0.65]
Hk,3 = [0.90 0.35], Hk,4 = [0.85 0.20].
Besides, the coefficient a = 0.1, the coupling strength πij is chosen as

−0.3, i = j
πij =
0.1, i = j

and the innercoupling matrix G = 0.2I2 . The processing and measure-

ment noise covariances for each node are given as Qk,i = 0.001I2
and Rk,i = 0.01, i = 1, . . ., 4, k = 1, 2, . . .. The initial values of
the system states and estimates are chosen as x0,1 = [2.0 − 2.8]T ,
x0,2 = [2.5 − 2.5]T , x0,3 = [2.5 − 2.0]T , x0,4 = [2.0 − 2.0]T , and Fig. 2. Estimation errors of each node. (a) node 1. (b) node 2. (c) node
x̂0,i = x0,i + 0.2i × [δi , δi ]T , i = 1, . . ., 4, where δi is zero-mean 3. (d) node 4.
Gaussian noise with unit covariance. Other parameters are given as
β = 0.5, P̆0,i = 0.001, and αk,i = 0.1 ∀k = 1, 2, . . ..
The simulation results are illustrated from the following three as- node in the presence of disturbances and nonlinearities, which indicates
pects: that the estimation performance is ensured by the proposed distributed
1) the estimation performance of each node by Algorithm 1; estimation algorithm.
2) a comparative study with several typical estimation algorithms;
3) a comparative study with different network structures. B. Part 2
To better illustrate the performance of the proposed distributed
A. Part 1
robust estimation algorithm (DREA), a comparative simulation for the
By using Algorithm 1 based on the aforementioned parameters, the network (1) with several typical estimation algorithms is presented, in-
numerical simulation results are shown in Figs. 1 and 2, where the cluding the augmented recursive estimation algorithm (AREA) in [14]
states, the estimates, and the estimation errors of each node are plotted. and [15], the nonaugmented variance-constrained estimation algorithm
It can be seen that the estimates well follow the real states of each (NVEA) in [16]–[18]. Before proceeding to the comparative results,

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:30:56 UTC from IEEE Xplore. Restrictions apply.
2588 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 5, MAY 2022

Fig. 5. Different network topologies. (a) star. (b) ring. (c) chain.

Fig. 3. Mean-square errors (MSEs) of node 2 without uncertainties.

Fig. 6. Mean-square errors (MSEs) of node 2 with different topologies.

π14 = 0.1, π21 = 0.1, π22 = −0.1, π32 = 0.1, π33 = −0.2, π34 =
0.1, π43 = 0.1, and π44 = −0.1. For the chain graph, π21 = 0.1,
Fig. 4. Mean-square errors (MSEs) of node 2 with uncertainties. π22 = −0.1, π32 = 0.1, π33 = −0.1, π43 = 0.1, and π44 = −0.1.
Then, without loss of generality, take node 2, for example, of which
the average MSEs (dB: −10 × lg(MSEk,i )) obtained by the proposed
define the average MSEs for the node i as follows:
algorithm is shown in Fig. 6. It can be seen that the errors of node 2 with
1  l
1000 different topologies are similar, which indicates that the proposed algo-
MSEk,i = x̂ − xlk,i 22 rithm can ensure the estimation performance under multiple network
1000 l=1 k,i
topologies.
where l denotes the lth trial. First, consider the dynamical network
without uncertainties, i.e., Δfk,i = 0 in (1). Without loss of generality, V. CONCLUSION
the average MSEs (dB: −10 × lg(MSEk,i )) for node 2 by the afore-
In this article, a state estimator has been proposed to estimate the
mentioned algorithms are shown in Fig. 3. Then, consider the case with
states of a class of uncertain nonlinear networks. The uncertainties, dis-
uncertainties, where the unknown real system nonlinear function is
turbances and nonlinearities in system dynamics were modeled under a
 
0.9xk,i (1) + sin(0.525xk,i (2)) unified framework. By using the regularized least-squares method and
f (xk,i ) =
0.9xk,i (2) − sin(0.525xk,i (1)) a decoupling technique, the estimator gains were designed by solving
a distributed optimization problem. Differing from the LMIs-based
and the known nominal system nonlinear function is given in (17). The approach and the variance-constrained approach, the feasibility of the
average MSEs for node 2 of this uncertain case are provided in Fig. 4. It estimators and the boundedness of the MSE have been guaranteed
can be found that in either case, the estimation performance of DREA with an easily checked condition. The good estimation performance
is better than that of AREA and NVEA. This further shows the supe- of the proposed estimation algorithm was illustrated by comparative
riority of the proposed estimators in terms of handling uncertainties, simulations.
disturbances, and nonlinearities.

REFERENCES
C. Part 3
[1] A. Arenas, A. Daz-Guilera, J. Kurths, Y. Moreno, and C. Zhou, “Synchro-
To better illustrate the effectiveness of the proposed algorithm in nization in complex networks,” Phys. Rep., vol. 469, no. 3, pp. 93–153,
dealing with general network structures, the simulation for the network 2008.
(1) with different topologies is presented, including a star graph, a chain [2] G. Wen, P. Wang, X. Yu, W. Yu, and J. Cao, “Pinning synchronization
graph, and a ring graph, as shown in Fig. 5. In this part, the system of complex switching networks with a leader of nonzero control inputs,”
IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 66, no. 8, pp. 3100–3112,
model is chosen to be the same as that in Part 1, expect the coupling Aug. 2019.
strengths πij . For the star graph, π11 = −0.1, π12 = 0.1, π21 = 0.1, [3] P. Duan, Z. Duan, Y. Lv, and G. Chen, “Distributed finite-horizon extended
π22 = −0.3, π23 = 0.1, π24 = 0.1, π32 = 0.1, π33 = −0.1, π42 = Kalman filtering for uncertain nonlinear systems,” IEEE Trans. Cybern.,
0.1, and π44 = −0.1. For the ring graph, π11 = −0.2, π12 = 0.1, vol. 51, no. 2, pp. 512–520, Feb. 2021.

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:30:56 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 67, NO. 5, MAY 2022 2589

[4] D. Shi, T. Chen, and M. Darouach, “Event-based state estimation of linear [17] W. Li, Y. Jia, and J. Du, “State estimation for stochastic complex networks
dynamic systems with unknown exogenous inputs,” Automatica, vol. 69, with switching topology,” IEEE Trans. Autom. Control, vol. 62, no. 12,
pp. 275–288, 2016. pp. 6377–6384, Dec. 2017.
[5] P. Duan, G. Lv, Z. Duan, and Y. Lv, “Resilient state estimation for complex [18] W. Li, Y. Jia, and J. Du, “Resilient filtering for nonlinear complex networks
dynamic networks with system model perturbation,” IEEE Trans. Control with multiplicative noise,” IEEE Trans. Autom. Control, vol. 64, no. 6,
Netw. Syst., vol. 8, no. 1, pp. 135–146, Mar. 2021. pp. 2522–2528, Jun. 2019.
[6] B. Shen, Z. Wang, and X. Liu, “Bounded H∞ synchronization and state [19] T. Zhou, “Coordinated one-step optimal distributed state prediction for
estimation for discrete time-varying stochastic complex networks over a a networked dynamical system,” IEEE Trans. Autom. Control, vol. 58,
finite horizon,” IEEE Trans. Neural Netw., vol. 22, no. 1, pp. 145–157, no. 11, pp. 2756–2771, Nov. 2013.
Jan. 2011. [20] A. H. Sayed, V. H. Nascimento, and F. A. M. Cipparrone, “A regularized
[7] D. Ding, Z. Wang, B. Shen, and H. Shu, “H∞ state estimation for discrete- robust design criterion for uncertain data,” SIAM J. Matrix Anal. Appl.,
time complex networks with randomly occurring sensor saturations and vol. 23, no. 4, pp. 1120–1142, 2002.
randomly varying sensor delays,” IEEE Trans. Neural Netw. Learn. Syst., [21] K. Zhou et al., Robust and Optimal Control. Englewood Cliffs, NJ, USA:
vol. 23, no. 5, pp. 725–736, May 2012. Prentice-Hall, 1996.
[8] B. Shen, Z. Wang, D. Ding, and H. Shu, “H∞ state estimation for complex [22] K. Xiong, C. Wei, and L. Liu, “Robust Kalman filtering for discrete-time
networks with uncertain inner coupling and incomplete measurements,” nonlinear systems with parameter uncertainties,” Aerosp. Sci. Technol.,
IEEE Trans. Neural Netw. Learn. Syst., vol. 24, no. 12, pp. 2027–2037, vol. 18, no. 1, pp. 15–24, 2012.
Dec. 2013. [23] B. Xiao, S. Yin, and O. Kaynak, “Tracking control of robotic manipulators
[9] L. Wang, Z. Wang, T. Huang, and G. Wei, “An event-triggered approach with uncertain kinematics and dynamics,” IEEE Trans. Ind. Electron.,
to state estimation for a class of complex networks with mixed time delays vol. 63, no. 10, pp. 6439–6449, Oct. 2016.
and nonlinearities,” IEEE Trans. Cybern., vol. 46, no. 11, pp. 2497–2508, [24] Y. Theodor and U. Shaked, “Robust discrete-time minimum-variance
Nov. 2016. filtering,” IEEE Trans. Signal Process., vol. 44, no. 2, pp. 181–189,
[10] Y. Xu, R. Lu, H. Peng, K. Xie, and A. Xue, “Asynchronous dissipative Feb. 1996.
state estimation for stochastic complex networks with quantized jumping [25] P. Duan, Q. Wang, Z. Duan, and G. Chen, “A distributed optimization
coupling and uncertain measurements,” IEEE Trans. Neural Netw. Learn. scheme for state estimation of nonlinear networks with norm-bounded un-
Syst., vol. 28, no. 2, pp. 268–277, Feb. 2017. certainties,” 2020. [Online]. Available: https://arxiv.org/abs/2011.13206.
[11] X. Wu, G.-P. Jiang, and X. Wang, “State estimation for general complex [26] S. Arora and B. Barak, Computational Complexity: A Modern Approach.
dynamical networks with packet loss,” IEEE Trans. Circuits Syst. II, Exp. Cambridge, U.K.: Cambridge Univ. Press, 2009.
Briefs, vol. 65, no. 11, pp. 1753–1757, Nov. 2018. [27] M. Bartholomew-Biggs, Nonlinear Optimization With Engineering Appli-
[12] H. Dong, N. Hou, Z. Wang, and W. Ren, “Variance-constrained state cations. New York, NY, USA: Springer, 2008.
estimation for complex networks with randomly varying topologies,” IEEE [28] S. P. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
Trans. Neural Netw. Learn. Syst., vol. 29, no. 7, pp. 2757–2768, Jul. 2018. Cambridge Univ. Press, 2004.
[13] J. Hu, Z. Wang, and H. Gao, “Recursive filtering with random parameter [29] A. H. Sayed, “A framework for state-space estimation with uncertain
matrices, multiple fading measurements and correlated noises,” Automat- models,” IEEE Trans. Autom. Control, vol. 46, no. 7, pp. 998–1013,
ica, vol. 49, no. 11, pp. 3440–3448, 2013. Jul. 2001.
[14] J. Hu, Z. Wang, S. Liu, and H. Gao, “A variance-constrained approach to [30] T. Zhou, “On the convergence and stability of a robust state estimator,”
recursive state estimation for time-varying complex networks with missing IEEE Trans. Autom. Control, vol. 55, no. 3, pp. 708–714, Mar. 2010.
measurements,” Automatica, vol. 64, pp. 155–162, 2016. [31] G. Battistelli, L. Chisci, and D. Selvi, “A distributed Kalman filter with
[15] J. Hu, Z. Wang, G. Liu, and H. Zhang, “Variance-constrained recursive event-triggered communication and guaranteed stability,” Automatica,
state estimation for time-varying complex networks with quantized mea- vol. 93, pp. 75–82, 2018.
surements and uncertain inner coupling,” IEEE Trans. Neural Netw. Learn. [32] S. Bonnabel and J. Slotine, “A contraction theory-based analysis of the
Syst., vol. 31, no. 6, pp. 1955–1967, Jun. 2020. stability of the deterministic extended Kalman filter,” IEEE Trans. Autom.
[16] W. Li, Y. Jia, and J. Du, “Non-augmented state estimation for nonlinear Control, vol. 60, no. 2, pp. 565–569, Feb. 2015.
stochastic coupling networks,” Automatica, vol. 78, pp. 119–122, 2017.

Authorized licensed use limited to: The University of Hong Kong Libraries. Downloaded on October 28,2024 at 03:30:56 UTC from IEEE Xplore. Restrictions apply.