Communicated by Dr.

Ma Lifeng Ma

Accepted Manuscript

Particle Filtering for Networked Nonlinear Systems Subject to

Random One-step Sensor Delay and Missing Measurements

Long Xu, Kemao Ma, Wenshuo Li, Yurong Liu, Fuad E. Alsaadi

PII: S0925-2312(17)31710-1
DOI: 10.1016/j.neucom.2017.10.059
Reference: NEUCOM 19039

To appear in: Neurocomputing

Received date: 16 August 2017

Revised date: 12 October 2017
Accepted date: 30 October 2017

Please cite this article as: Long Xu, Kemao Ma, Wenshuo Li, Yurong Liu, Fuad E. Alsaadi, Particle
Filtering for Networked Nonlinear Systems Subject to Random One-step Sensor Delay and Missing
Measurements, Neurocomputing (2017), doi: 10.1016/j.neucom.2017.10.059

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Particle Filtering for Networked Nonlinear Systems Subject to

Random One-step Sensor Delay and Missing Measurements

Long Xu a , Kemao Ma a,∗ , Wenshuo Li b , Yurong Liu c,d , Fuad E. Alsaadi d

a
Department of Control Science and Engineering, Control and Simulation Center, Harbin Institute of Technology, Harbin
150080, China

T
b
School of Automation Science and Electrical Engineering, Beihang University, Beijing 100083, China

IP
c
Department of Mathematics, Yangzhou University, Yangzhou 225002, China

CR
d
Communication Systems and Networks (CSN) Research Group, Faculty of Engineering, King Abdulaziz University, Jeddah
21589, Saudi Arabia

Abstract
US
In this paper, the particle filtering problem is investigated for a class of networked nonlinear systems with random one-step
AN
sensor delay and missing measurements. The phenomena of missing measurements and random one-step sensor delay are
modeled by two random variables, both obeying the Bernoulli distribution. Here, we derive an explicit expression for the
likelihood function when the possible occurrence of one-step sensor delay and measurement loss is taken into consideration.
Based on this likelihood function, we propose a novel particle filtering algorithm to treat the nonlinear estimation problem in
M

the simultaneous presence of random sensor delay and measurement loss. Finally, a simulation example is given to illustrate
the feasibility and advantages of the proposed filtering scheme compared with traditional particle filtering algorithm.
ED

Key words: Particle filter; networked systems; random one-step sensor delay; missing measurements; bayesian framework.

1 Introduction pling algorithm was utilized to estimate the nonlinear

PT

system states, where the proposed algorithm takes a set

In recent years, the nonlinear filtering problems were
of samples extracted from the proposed distribution to
extensively considered due to wide applications in many
approximate the distribution of the target and each sam-
CE

fields such as target tracking, location and parameter

ple has a corresponding weight. Due to considering the
estimation [1–6]. It is well known that Kalman first pro-
problem of sample weight degradation in the sequential
posed the classical Kalman filter in [7]. Up to now, the
importance sampling algorithm, the application of se-
Kalman filter is still widely used as a standard frame-
AC

quential importance sampling algorithm in solving non-

work to solve the linear estimation problem. However,
linear filtering problem was very limited at that time. In
in a practical environment, most of the systems have
order to avoid this case, the re-sampling technique was
nonlinear characteristics [8, 9], hence the study of non-
introduced into the Monte Carlo importance sampling
linear filtering problems is of important significance. A
process and a new filter was proposed in [11], which can
well-known method, the Monte Carlo method, was pro-
effectively solve the problem of the sample weight degra-
posed in [10] and used to solve the nonlinear filtering
dation. By summarizing the sequential importance sam-
problems. In addition, the sequential importance sam-

Preprint submitted to Automatica 3 November 2017

 ACCEPTED MANUSCRIPT

pling algorithm, the concept of Sequential Monte Carlo missing measurements have attracted the attention of
method, namely particle filter, was proposed in [12]. The many scholars and a large number of results were re-
basic idea of this method is to approximate the poste- ported [26–30]. Recently, in [31], a linear optimal filtering
rior probability distribution of the system by employing problem was investigated for a class of stochastic time-
a set of samples (or particles) in order to estimate the varying systems subject to state delay and randomly
state of the nonlinear system. Subsequently, the parti- multiple sensor delays based on the innovative analysis
cle filtering algorithm has been widely applied in many approach and recursive projection formula. By using the
practical problems. same method as in [31], the optimal filter was designed in
[32] for a networked stochastic state delay systems with
Due to the wide application of particle filtering ap-
missing measurements. Based on the state augmentation
proach, the scholars paid close attention to the devel-
approach and the minimum mean square error princi-

T
opments of particle filtering technique in many different
ple, the optimal filtering problem was investigated in [33]
research fields, such as modern signal processing, com-

IP
for a class of discrete stochastic systems with finite-
munication, artificial intelligence, autonomous mobile
step autocorrelated process noises, random one-step sen-
robot and target tracking. Accordingly, a large number
sor delay and missing measurements. In [34], the opti-

CR
of application results were published concerning on par-
mal filter was constructed for stochastic systems subject
ticle filtering algorithm. To mention a few, for the situ-
to the random two-step sensor delays and multiplica-
ation of randomly missing measurements in engineering
tive noises. Moreover, an optimal linear filter was con-
practice, a filtering method was proposed in [13] for non-
linear system based on sequential importance sampling
and was applied to online state estimation for nonlin-
ear non-Gauss systems. In [14], a new hybrid filtering
US structed in the sense of the minimum mean-square error
in [35] for discrete-time systems with measurement-delay
and packet dropout. By utilizing the Riccati-like differ-
AN
ence equation approach, new recursive state estimation
algorithm was proposed for a class of nonlinear systems
method was developed in [36] for time-varying complex
with linear state equation and nonlinear measurement
networks subject to missing measurements. However, to
equation based on the approaches of Kalman filter and
the best of authors’ knowledge, these investigations do
particle filter. Based on Bayesian filtering framework,
M

not provide much attention to the problem of the par-

a new particle filter was designed in [15] for nonlinear
ticle filtering algorithm for discrete nonlinear stochastic
systems with random one-step delayed measurements,
systems with random one-step sensor delay and missing
where the probability of the time-delay is unknown and is
ED

measurements.
estimated by maximum likelihood criterion. When tak-
ing the multiple-step time-delay into account, the non- Motivated by the above discussions, in this paper,
linear filtering problem was discussed in [16] for nonlin- we aim to investigate the particle filtering problem for a
PT

ear systems with multiple-step randomly delayed mea- class of networked nonlinear systems with random one-
surements by applying the idea as in [15]. When the step sensor delay and missing measurements. Two typi-
data are measured by multiple sensors, the problem of cal network-induced phenomena, namely the sensor de-
CE

distributed particle filtering was investigated in [17] for lay and measurement loss, are taken into account and in-
nonlinear tracking over wireless sensor networks. More- corporated into the observation equation. The phenom-
over, a distributed particle filter was constructed in [18] ena of random one-step sensor delay and missing mea-
AC

for bearings-only tracking of a target moving on the sur- surements are described by two Bernoulli distributed
face of a sphere and a proof of uniform convergence of random variables with known conditional probabilities.
distributed particle filtering algorithm was given in [19]. Based on the Bayesian filtering framework, the particles
and their weights are updated due to taking the random
In many practical systems, the phenomena of random
one-step sensor delay and missing measurements into ac-
time-delay and missing measurements are inevitably en-
count. Here, by the assumption of first-order Markovian
countered and the accuracy of the filtering algorithm
process, we make the first attempt to propose the par-
is degraded [20–25]. Hence, both the time-delay and
ticle filter for nonlinear systems subject to random one-

2
 ACCEPTED MANUSCRIPT

step sensor delay and missing measurements. It is shown have the statistical properties as follows:
that, based on the new particle filtering algorithm and
Matlab software, we can recursively compute the state Prob {λk = 1} = E {λk } = α,
estimate and the error covariance matrices. Finally, we Prob {λk = 0} = 1 − E {λk } = 1 − α,
give a numerical example to verify the performance of Prob {ξk = 1} = E {ξk } = β,
the proposed filtering algorithm.
Prob {ξk = 0} = 1 − E {ξk } = 1 − β,
Notation. The symbols used in the paper are stan-
dard. Rn denotes the n-dimensional Euclidean space. AT where α, β ∈ [0, 1] are known scalars and λk , ξk are
represents the transpose of a matrix A. E{x} is the ex- uncorrelated with other noise signals.
pectation of the random variable x. The identity matrix Remark 1 In the model (3), if λk = 1, yk = zk , it

T
and the zero matrix are expressed by I and 0 with ap- stands for that the sensor receives the data at time instant
propriate dimensions, respectively. Xk = {xj }kj=1 is the k successfully; if λk = 0 and ξk = 1, yk = zk−1 , it means

IP
set of the state vectors. Zk = {zj }kj=1 stands for the set that there exists one-step delay; if λk = 0 and ξk = 0,
of the actual output. Yk = {yj }kj=1 is the set of the mea- yk = νk , it represents that current measurements are

CR
sured output. N is the number of the particles. δk−j is missing completely.
the Kronecker delta function. If k = j, then δk−j = 1;
R In order to make the subsequent proof process sim-
otherwise, δk−j = 0. is the integral operation. x ∼
p(x) denotes that the variable x obeys a distribution ple, we make the following three assumptions as in the
with probability density function p(x). If a stochastic
process {x}∞ k=0 satisfies first-order Markovian property,
then p(xk |xk−1 , xk−2 , · · · , x0 ) = p(xk |xk−1 ). If the di-
US reference [15].
Assumption 1 The state xk obeys the following first-
order Markovian process:
AN
mensions of the matrices are not definitely stated, they
are considered to be well-matched for algebraic opera- p (xk |xk−1 , H ) = p (xk |xk−1 ) ,
tions.
where H is an arbitrary subset of {x0 , Xk−2 , Yk−1 }.
M

2 Problem Formulation and Preliminaries Assumption 2 The measured output of the sensor yk
and the states xk , xk−1 obey the following relationship
Consider the following discrete nonlinear stochastic
ED

systems with random sensor delay and missing measure- p (yk |xk , xk−1 , H ) = p (yk |xk , xk−1 ) .
ments:
Assumption 3 The previous state Xk−1 and the output
PT

xk+1 = fk (xk ) + ωk (1) of the sensor yk are uncorrelated in terms of proposal

zk = hk (xk ) + νk (2) density function q(·), i.e., q(Xk−1 |Yk ) = q(Xk−1 |Yk−1 ).
CE

yk = λk zk + (1 − λk )ξk zk−1 + (1 − λk )(1 − ξk )νk (3) The purpose of this paper is to design a new particle
filter in Bayesian filtering framework based on the ob-
where xk ∈ Rn is the state vector, zk ∈ Rm is the actual servation sequence {y1 , y2 , · · · , yk } and above three as-
output, and yk ∈ Rm is the measured output of the sen- sumptions.
AC

sor. fk (·) : Rn → Rn and hk (·) : Rn → Rm are known

nonlinear functions. ωk ∈ Rn and νk ∈ Rm are uncor- 3 Design of Estimation Algorithm
related white noises obeying the arbitrary probability
Due to the phenomena of random one-step sensor
density functions.
delay and missing measurements, the selection of im-
The random variables λk and ξk , which describe the portance weights are different from traditional particle
phenomena of random one-step sensor delay and miss- filter. Hence, we simply introduce the importance sam-
ing measurements, satisfy the Bernoulli distribution and pling methods in [37]. Firstly, the following lemma is

3
 ACCEPTED MANUSCRIPT

given in [38] to introduce Baysian theorem and Baysian where the importance weights wki and the normalised
rule. weights w̄ki are formulated as follows:
Lemma 1 Baysian theorem relates joint probability p(Xki , Yk ) wki
density function to conditional and marginal probability wki = , w̄ki = . (7)
q(Xki |Yk ) P
N
density functions, i.e., two events X and Y have the wki
i=1
following relationship:
It is easy to see that we can approximate posterior dis-
p(X, Y ) = p(X|Y )p(Y ) = p(Y |X)p(X). tribution p(Xk |Yk ) from (3) and (6) as follows:

Baysian rule is expressed as: N

X  
p(Xk |Yk ) ≈ p̂(Xk |Yk ) = w̄ki δ Xk − Xki . (8)
p(Y |X)p(X)

T
i=1
p(X|Y ) = .
p(Y )

IP
By applying the importance sampling approach, a
Based on the Lemma 1 and the proposal density func- recursive computation formula of importance weights wki
tion q(Xk |Yk ), the following optimal estimation of arbi-

CR
can be obtained for nonlinear discrete stochastic systems
trary function g(Xk ) in the mean square sense can be with time-delay and missing measurements.
derived easily:
Theorem 1 Based on the assumptions 1-3, the impor-
ĝ(Xk ) = E [g(Xk )|Yk ]
Z
= g(Xk )p(Xk |Yk )dXk
Z
p(Xk |Yk )
US tance weights wki can be computed recursively as follows:

wki =
p(yk |xik , xik−1 )p(xik |xik−1 ) i
q(xik |Xk−1
i , Yk )
wk−1 , (9)
AN
= g(Xk ) q(Xk |Yk )dXk . (4)
q(Xk |Yk )
where
From the Lemma 1, we can obtain
p(yk |xik , xik−1 ) = αpνk (yk − hk (xik ))
p(Xk , Yk )
M

p(Xk |Yk ) = +(1 − α)βpνk−1 (yk − hk−1 (xik−1 ))

p(Yk )
p(Xk , Yk ) +(1 − α)(1 − β)pνk−1 (yk ), (10)
=R
p(Xk , Yk )dXk
ED

p(Xk , Yk ) xik is obtained from the proposal density function

= R p(X ,Y ) . (5)
q(xk |Xk−1 , Yk ), pνk (·) and pνk−1 (·) stand for the prob-
q(Xk |Yk ) q(Xk |Yk )dXk
k k

ability density functions of observation noises νk and

PT

The following equation can be derived by substituting νk−1 .

(5) into (4):
Proof : By applying the assumption 3 and the
Z p(Xk ,Yk )
Bayesian theorem, one has
CE

q(Xk |Yk )
ĝ(Xk ) = g(Xk ) R p(Xk ,Yk )
q(Xk |Yk )dXk .
q(Xk |Yk ) q(Xk |Yk )dXk
q(Xk |Yk ) = q(xk |Xk−1 , Yk )q(Xk−1 |Yk )
= q(xk |Xk−1 , Yk )q(Xk−1 |Yk−1 ). (11)
Let wk = p(X k ,Yk )
q(Xk |Yk ) , where wk is the importance
AC

weight. Then, by choosing independent and identically Hence, we can get xik and Xk−1
i
from q(xk |Xk−1 , Yk ) and
 N
distributed sample Xki i=1 from the proposal density q(Xk−1 |Yk−1 ), respectively. From the Bayesian theorem,
function q(Xk |Yk ), we have: the following equation can be obtained:

1
P
N
N wki g(Xki ) N
X p(Xk , Yk ) = p(xk , Xk−1 , yk , Yk−1 )
i=1
ĝ(Xk ) ≈ = w̄ki g(Xki ), (6) = p(yk |xk , Xk−1 , Yk−1 )p(xk , Xk−1 , Yk−1 )
1
P
N
wki i=1
N
i=1 = p(yk |xk , Xk−1 , Yk−1 )

4
 ACCEPTED MANUSCRIPT

×p(xk |Xk−1 , Yk−1 )p(Xk−1 , Yk−1 ). (12) Thirdly, also similar to the above two cases, the fol-
lowing equation is computed:
By using the assumptions 1 and 2, the equation (12) can
be equivalently rewritten as follows: p(yk |λk = 0, ξk = 0, xk , xk−1 ) = pνk−1 (yk ). (19)

p(Xk , Yk ) = p(yk |xk , xk−1 ) Based on the similar derivation in [15], the likelihood
density function p(yk |xk , xk−1 ) can be calculated by us-
×p(xk |xk−1 )p(Xk−1 , Yk−1 ). (13)
ing equations (16), (18) and (19) as follows:

Based on the equations (11) and (13), the importance p(yk |xk , xk−1 )
weights wk can be expressed in the following form: Z
= p(yk , λk |ξk , xk , xk−1 )dλk

T
Z
p(Xk , Yk )
wk = = p(yk |λk , ξk , xk , xk−1 )p(λk |ξk , xk , xk−1 )dλk

IP
q(Xk |Yk )
Z
p(yk |xk , xk−1 )p(xk |xk−1 ) p(Xk−1 , Yk−1 )
= = p(yk |λk , ξk , xk , xk−1 )p(λk )dλk
q(xk |Xk−1 , Yk ) q(Xk−1 |Yk−1 )

CR
p(yk |xk , xk−1 )p(xk |xk−1 ) = αp(yk |λk = 1, xk , xk−1 )
= wk−1 . (14) Z
q(xk |Xk−1 , Yk ) +(1 − α) p(yk , ξk |λk = 0, xk , xk−1 )dξk
= αp(yk |λk = 1, xk , xk−1 )
Since the output of the sensor yk may be zk−1 or
νk if the one-step time-delay or missing measurements
happen, the update of importance weights wk in (14)
US Z
+(1 − α) p(yk |ξk , λk = 0, xk , xk−1 )
×p(ξk |λk = 0, xk , xk−1 )dξk
AN
is different from that in the traditional particle filter. = αp(yk |λk = 1, xk , xk−1 )
Also, the measurement yk depends on the value of the
+(1 − α)βp(yk |λk = 0, ξk = 1, xk , xk−1 )
Bernoulli random variables λk and ξk . Since λk and ξk
+(1 − α)(1 − β)p(yk |λk = 0, ξk = 0, xk , xk−1 )
are independent with other signals, we can evaluate the
M

likelihood density as follows: = αpνk (yk − hk (xk ))

+(1 − α)βpνk−1 (yk − hk−1 (xk−1 ))
Firstly, if λk = 1, we can obtain the following equa-
+(1 − α)(1 − β)pνk−1 (yk ). (20)
tion by (2) and (3):
ED

From the equations (14) and (20), we have the recursions

yk = hk (xk ) + νk . (15) of importance weights wki as the equations (9) and (10).
The proof of this theorem is complete.
PT

Hence, it is easy to see that yk − hk (xk ) and νk have the

According to Theorem 1, a new recursive algorithm
same distribution functions if xk is known, i.e.
can be established to obtain the nonlinear filter for the
CE

addressed discrete stochastic nonlinear systems with

p(yk |λk = 1, xk , xk−1 ) = pνk (yk − hk (xk )). (16)
random one-step sensor delay and missing measure-
ments. We can design the particle filter in Theorem 1
Next, if λk = 0 and ξk = 1, it holds that as the following algorithm.
AC

Remark 2 In the above algorithm, the first attempt is

yk = hk (xk−1 ) + νk−1 . (17) to develop the particle filtering algorithm for a class of
networked nonlinear discrete stochastic systems subject
Then, similar to (16), we have to random one-step sensor delay and missing measure-
ments. Due to the phenomena of time delay and miss-
p(yk |λk = 0, ξk = 1, xk , xk−1 ) ing measurements, the particles and their weights are up-
= pνk−1 (yk − hk−1 (xk−1 )). (18) dated as Theorem 1 to reduce the conservativeness and

5
 ACCEPTED MANUSCRIPT

Algorithm 1 The steps of the particle filtering algo- dom one-step sensor delay and missing measurements.
rithm are shown as follows:: Hence, the measurement yk depends on the value of the
Step 1. Initialisation: Bernoulli random variables λk and ξk . Then, the deriva-
The particles are extracted from the prior density func- tion of the likelihood density function p(yk |xk , xk−1 ) is
tion p(x0 ) to obtain a new particle set {xi0 }N
i=1 . The divided into two parts. When the random variable λk = 1,
weight of all particles is N1 . the sensor output yk depends on xk ; otherwise, the out-
For k = 1, 2, · · · put yk depends on ξk , xk and xk−1 . Next, we discuss
Step 2. Sequential importance re-sampling: the random variable ξk taking the value of 0 or 1. If
1 . Choose the prior density function as the proposal ξk = 1, yk depends on xk−1 ; otherwise, ys depends on νk .
density function, i.e. q(xik |Xk−1
i
, Yk ) = p(xik |xik−1 ). Therefore, the key of calculating the importance weights
i N
{xk }i=1 were the samples which extracted from the im- wk is the derivation of the likelihood density function

T
portance distribution density function. p(yk |xk , xk−1 ) and we made great efforts to address this

IP
2 . Calculate the weights of particles by Theorem 1. issue.
Due to 1 and the equations (9),(10), the weights can be
Remark 4 It’s well known that the extended Kalman
computed as follows:

CR
filter is designed by using the first-order linearization for
nonlinear stochastic systems, where the error will lead to
wki = αpνk (yk − hk (xik ))
the divergence of the proposed filtering algorithm. Hence,
+(1 − α)βpνk−1 (yk − hk−1 (xik−1 ))

i the unscented Kalman filtering is proposed based on ap-
+(1 − α)(1 − β)pνk−1 (yk ) wk−1

3 . Normalize the weights:

.
US plying the unscented transformation approach to reduce
the error of the the extended Kalman filtering algorithm.
However, in many practice systems, the non-Gaussian
AN
wki problems can not be solved by employing the unscented
w̄ki = . Kalman filtering. Therefore, the particle filtering algo-
P
N
i
wk
i=1
rithm, also called the Bayesian estimation algorithm,
has been received considerable attention. Simply speak-
M

Step 3. Re-sample: ing, the particle filtering algorithm is to approximate the

Based on the technology of the sequential importance probability density function of the system by employing a
sampling, the particles are re-sampled. set of random samples which spread in the sate space to
ED

Step 4. Compute the state estimation: obtain the minimum variance distribution of the state.
N Here, the samples or particles, when the number of par-
X
x̂k = w̄ki xik , ticles N → ∞, can approximate any form of probability
PT

i=1 density distribution. What’s more, the particle filtering

N
X algorithm isn’t subject to the Gaussian distribution and
Pk = w̄ki (xik − x̂k )(xik − x̂k )T .
i=1
it can approximate the more wide distribution. Due to
CE

the advantage of the particle filtering algorithm for the

Step 5. Return to the Step 2. nonlinear non-Gaussian systems, the particle filtering al-
The end. gorithm proposed in the paper has extensive application
AC

than the traditional extended Kalman filter and unscented

ensure the accuracy of the state estimation algorithm. Kalman filter.
Remark 3 In this paper, we consider the phenomena of
missing measurements and random one-step sensor delay
4 An Illustrative Example
together when designing the particle filtering algorithm.
To be more specific, two Bernoulli distributed random We give a simulation example to illustrate the per-
variables with known conditional probabilities (i.e. λk formance of the developed filtering algorithm in this sec-
and ξk ) are employed to describe the phenomena of ran- tion.

6
 ACCEPTED MANUSCRIPT

Consider the following nonlinear discrete stochastic

1.4
systems: e1k in this paper
1.2 e1k in [15]
e1k in [37]
xk+1 = 3.5xk (1 − xk ) + ωk 1

zk = cos(xk π) + x2k + νk 0.8

yk = λk zk + (1 − λk )ξk zk−1 + (1 − λk )(1 − ξk )νk 0.6

 T 0.4
where xk = x1k x2k x3k . The process noises ωk and
0.2
observation noises νk are uncorrelated white noises and
0
satisfy ωk ∼ N (0, Qk ) and νk ∼ N (0, Rk ) with the pos-
−0.2

T
itive definite covariance matrices Qk = 0.1I3 , Rk =
0.01I3 . −0.4
0 10 20 30 40 50

IP
k/time step
Let
 T  T Fig. 1. The error of state x1k

CR
x0 = 0.1 0.15 0.2 ; x̂0|0 = −0.9 −0.85 −0.8 ;
P0|0 = I3 ; α = 0.8; β = 0.85
1.4
e2k in this paper
and eik denote the estimation error of xik , i.e., eik = xik −
x̂ik|k , where i = 1, 2, 3.
According to Theorem 1, the particle filtering al-
US 1.2

0.8
e2k in [15]
e2k in [37]
AN
gorithm can be designed based on the Bayesian filtering
0.6
framework and the assumption of first-order Markovian
0.4
process. The values of states and their estimates at every
time step can be recursively computed by utilizing the 0.2
M

given algorithm and Matlab software. The state estima- 0

tion errors eik (i = 1, 2, 3) of the proposed filtering algo- −0.2

rithm are plotted in Figs. 1-3. Figs. 4-6 plot the actual
ED

−0.4
system states xik and their estimates x̂ik|k (i = 1, 2, 3). 0 10 20 30 40 50
k/time step
From the simulations, we can see that the range of er-
ror fluctuation in our paper is relatively small compared Fig. 2. The error of state x2k
PT

with the error in the references [15, 37]. It is shown that

the proposed filter can estimate the system state well.
compute the weights of particles according to the equa-
In our paper, we analyze and compare the different
CE

tions (9) and (10) in the particle filtering algorithm to

rates of time-delay and missing measurements in order
reduce the errors in the sampling process.
to investigate the effects from the considered phenomena
in depth (i.e., Case I: α = 0.8, β = 0.85, 0.55, 0.25; Case
5 Conclusions
AC

II: α = 0.8, 0.5, 0.2, β = 0.85 ). The corresponding sim-

ulation results are given in Figs. 7-12. Based on the sim- The problem of particle filtering problem has been
ulation results, we can conclude that the different prob- investigated for a class of networked nonlinear stochas-
abilities of sensor delay or missing measurements have a tic systems subject to random one-step sensor delay and
significant influence on the accuracy of the filter. Com- missing measurements. Based on the Bayesian filtering
paratively speaking, the proposed filter can estimate the framework and the assumption of first-order Markovian
system states effectively compared with traditional par- process, the particles and their weights have been up-
ticle filtering algorithm in [37]. The reason is that we dated and a new particle filtering algorithm has been

7
 ACCEPTED MANUSCRIPT

1.1
e3k in this paper Actual state x2k
1.2 e3k in [15] filter in this paper
1
e3k in [37]

1 0.9

0.8
0.8

0.6
0.7

0.4
0.6

0.2
0.5
0
0.4

T
−0.2
0.3
0 10 20 30 40 50 0 10 20 30 40 50

IP
k/time step k/time step

Fig. 3. The error of state x3k Fig. 5. The trajectories of x2k and x̂2k|k

CR
1.1
Actual state x3k

1.1

1
Actual state x1k
filter in this paper US 1

0.9

0.8
filter in this paper
AN
0.9

0.8 0.7

0.7 0.6
M

0.6 0.5

0.5 0.4
ED

0.4 0.3
0 10 20 30 40 50
k/time step
0.3
0 10 20 30 40 50
k/time step
Fig. 6. The trajectories of x3k and x̂3k|k
PT

Fig. 4. The trajectories of x1k and x̂1k|k

Acknowledgments
CE

This work was supported in part by the National Nat-

ural Science Foundation of China (NSFC) under Grants
proposed. Accordingly, a recursive algorithm has been
61174001, 61321062, 11301118 and 11271103.
given to design the nonlinear filter. Moreover, a simula-
AC

tion example has been used to show the feasibility and References
usefulness of the proposed approach compared with the
[1] E. Özkan, N. Wahlström, and S. Godsill, Rao-blackwellised
traditional particle filtering algorithm. Further research
particle filter for star-convex extended target tracking
topics include the extensions of the obtained results to models, 19th International Conference on Information
deal with the particle filtering problems for networked Fusion, (2016) 1193–1199.
systems with multi-step sensor delay and disturbances [2] F. Wang, Y. Zhen, B. Zhong, and R. Ji, Robust infrared
as in [39]. The corresponding results will appear in the target tracking based on particle filter with embedded
near future. saliency detection, Information Sciences, 301 (2015) 215–226.

8
 ACCEPTED MANUSCRIPT

Actual state x1k Actual state x3k

1.2 β=0.85 1.2 β=0.85
β=0.55 β=0.55
β=0.25 β=0.25

1 1

0.8 0.8

0.6 0.6

0.4 0.4

T
0.2 0.2
0 10 20 30 40 50 0 10 20 30 40 50

IP
k/time step k/time step

Fig. 7. The trajectories of x1k (α = 0.8) Fig. 9. The trajectories of x3k (α = 0.8)

CR
1.4
Actual state x2k Actual state x1k

US
1.2 β=0.85 α=0.8
α=0.5
β=0.55 1.2 α=0.2
β=0.25

1 1
AN
0.8
0.8

0.6
0.6
0.4
M

0.4
0.2
ED

0.2 0
0 10 20 30 40 50 0 10 20 30 40 50
k/time step k/time step

Fig. 8. The trajectories of x2k (α = 0.8) Fig. 10. The trajectories of x1k (β = 0.85)
PT

[3] S. Chattaraj and A. Mukherjee, Efficient estimation of system likelihoods, Scandinavian Journal of Statistic, 41 (4) (2014)
states of a poorly modeled 2-D target tracking system 970–987.
CE

using evolutionary strategy based particle filter algorithm, [7] R. E. Kalman, A new approach to linear filtering and
2016 IEEE First International Conference on Control, prediction problems, Transactions of the ASME Journal of
Measurement and Instrumentation, (2016) 346–350. the Basic Engineering,, 82 (1960) 35–45.
AC

[4] N. Kantas, A. Doucet, S. Singh, J. Maciejowski, and [8] J. Hu, D. Chen, and J. Du, State estimation for a class
N. Chopin, On particle methods for parameter estimation in of discrete nonlinear systems with randomly occurring
state-space models, Statistical Science, 30 (3) (2015) 328–351. uncertainties and distributed sensor delays, International
[5] X. Liu, Z. Chen, C. Zhang, and J. Wu, A novel temperature- Journal of General Systems, 43 (3-4) (2014) 387–401.
compensated model for power Li-ion batteries with dual- [9] J. Hu, S. Liu, D. Ji, and S. Li, On co-design of filter and
particle-filter state of charge estimation, Applied Energy, 123 fault estimator against randomly occurring nonlinearities and
(2014) 263–272. randomly occurring deception attacks, International Journal
[6] T. Dean, S. Singh, A. Jasra, and G. Peters, Parameter of General Systems, 45 (5) (2016) 619–632.
estimation for hidden Markov models with intractable [10] W. Chan, Theory and Applications of Monte Carlo

9
 ACCEPTED MANUSCRIPT

[15] Y. Zhang, Y. Huang, N. Li, and L. Zhao, Particle

1.4
Actual state x2k filter with one-step randomly delayed measurements and
α=0.8
1.2
α=0.5 unknown latency probability, International Journal of
α=0.2
Systems Science, 47 (1) (2015) 209–221.
1 [16] Y. Zhang, Y. Huang, N. Li, and L. Zhao, Particle filter
for nonlinear systems with multiple step randomly delayed
0.8
measurements, Eletronics Letters, 51 (23) (2015) 1859–1861.

0.6 [17] J. Read, K. Achutegui, and J. Mı́guez, A distributed particle

filter for nonlinear tracking in wireless sensor networks, Signal
0.4 Processing, 98 (2014) 121–134.
[18] J. Yu, M. Rabbat, S. Blouin, and M. Coates, A distributed
0.2
particle filter for bearings-only tracking on spherical surfaces,

T
0 IEEE Signal Processing Letters, 23 (3) (2016) 326–330.
0 10 20 30 40 50

IP
k/time step [19] J. Miguez and M. Vázquez, A proof of uniform convergence
over time for a distributed particle filter, Signal Processing,
Fig. 11. The trajectories of x2k (β = 0.85) 122 (2016) 152–163.

CR
[20] J. Hu, Z. Wang, B. Shen, and H. Gao, Quantised recursive
filtering for a class of nonlinear systems with multiplicative
1.4 noises and missing measurements, International Journal of
Actual state x3k

US
α=0.8 Control, 86 (4) (2013) 650–663.
α=0.5
1.2 α=0.2
[21] Z. Wang, H. Dong, B. Shen, and H. Gao, Finite-horizon
1 H∞ filtering with missing measurements and quantization
effects, IEEE Transactions on Automatic Control, 58 (7)
AN
0.8 (2013) 1707–1718.
[22] Z. Wang, B. Shen, and X. Liu, H∞ filtering with randomly
0.6
occurring sensor saturations and missing measurements,
0.4
Automatica, 48 (3) (2012) 556–562.
M

[23] J. Hu, Z. Wang, and H. Gao, Recursive filtering with

0.2 random parameter matrices, multiple fading measurements
and correlated noises, Automatica, 49 (11) (2013) 3441–3448.
ED

0
0 10 20 30 40 50
[24] Y. Yuan, F. Sun, and H. Liu, Resilient control of cyber-
k/time step
physical systems against intelligent attacker: a hierarchal
stackelberg game approach, International Journal of Systems
Fig. 12. The trajectories of x3k (β = 0.85)
Science, 47 (9) (2016) 2067–2077.
PT

[25] J. Hu, Z. Wang, F. E. Alsaadi, and T. Hayat, Event-

Simulations, Rijeka, Croatia: InTech, (2013). based filtering for time-varying nonlinear systems subject
[11] N. Gordon, D. Salmond, and A. Smith, Novel approach to multiple missing measurements with uncertain missing
CE

to nonlinear/non-Gaussian Bayesian state estimation, IEE probabilities, Information Fusion, 38 (2017) 74-83.
Proceedings F-Radar and Signal Processing, 140 (2) (1993) [26] N. Hou, H. Dong, Z. Wang, W. Ren, and F. E. Alsaadi,
107–113. Non-fragile state estimation for discrete Markovian jumping
AC

[12] J. Carpenter, P. Clifford, and P. Fearnhead, Improved particle neural networks, Neurocomputing, 179 (2016) 238–245.
filter for nonlinear problems, IEE Proceedings F-Radar, [27] Y. Yuan, H. Yuan, L. Gao, and H. Yang, Resilient control of
Sonar and Navigation, 146 (1) (1999) 2–7. networked control system under DoS attacks: A unified game
[13] J. Li and H. Zhou, A particle filter algorithm in the approach, IEEE Transactions on Industrial Informatics, 12
presence of missing measurements for a nonlinear system, (5) (2016) 1786–1794.
Acta Armamentarii, 30 (10) (2009) 1405–1408. [28] Y. Yuan, H. Yuan, Z. Wang, L. Guo, and H. Yang, Optimal
[14] N. Xia, S. Qiu, C. Li, and F. Li, A nonlinear filtering control for networked control systems with disturbances: a
algorithm combining the Kalman filter and the particle filter, delta operator approach, IET Control Theory & Applications,
Acta Electronica Sinica, 41 (1) (2013) 148–152. 11 (9) (2017) 1325–1332.

10
 ACCEPTED MANUSCRIPT

[29] Y. Song, J. Hu, D. Chen, D. Ji, and F. Liu, Recursive

approach to networked fault estimation with packet dropouts
and randomly occurring uncertainties, Neurocomputing, 214
(2016) 340–349.

[30] Q. Liu, Z. Wang, X. He, G. Ghinea, and F. E. Alsaadi,

A resilient approach to distributed filter design for time-
varying systems under stochastic nonlinearities and sensor
degradation, IEEE Transactions on Signal Processing, 65 (5)
(2017) 1300–1309.

[31] D. Chen and L. Xu, Optimal Kalman filtering for a class of

T
state delay systems with randomly multiple sensor delays,
Abstract and Applied Analysis, (2014) Article ID 716716.

IP
Long Xu received the B.Sc. degree in Applied Mathe-
[32] J. Du, L. Xu, Y. Liu, Y. Song, and X. Fan, Linear matics from Harbin University, Harbin, China, in 2011,
optimal filtering for time-delay networked systems subject
and M.Sc. degree in Applied Mathematics from Harbin

CR
to missing measurements with individual occurrence
probability, Neurocomputing, 214 (2016) 767–774.
University of Science and Technology, Harbin, China,
in 2015. He is currently a Ph.D. candidate with Control
[33] D. Chen, L. Xu, and J. Du, Optimal filtering for systems with Science and Engineering, Harbin Institute of Tech-
finite-step autocorrelated process noises, random one-step
sensor delay and missing measurements, Communications in
Nonlinear Science and Numerical Simulation, 32 (2016) 211–
224.
US nology, Harbin, China. His current research interests
include nonlinear control and filtering, time-varying
systems and stochastic systems.
AN
[34] D. Chen, Y. Yu, L. Xu, and X. Liu, Kalman filtering for
discrete stochastic systems with multiplicative noises and
random two-step sensor delays, Discrete Dynamics in Nature
M

and Society, (2015) Article ID 809734.

[35] X. Song, Z. Duan, and J. Park, Linear optimal estimation

ED

for discrete-time systems with measurement-delay and packet

dropping, Applied Mathematics and Computation, 284 (2016)
115–1242.
PT

[36] J. Hu, Z. Wang, S. Liu, and H. Gao, A variance-constrained

approach to recursive state estimation for time-varying
complex networks with missing measurements, Automatica,
64 (2016) 155–162.
CE

Kemao Ma received his Bachelor degree in engineering

[37] M. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, A from the Department of Information and Control Engi-
tutorial on particle filters for online nonlinear/non-Gaussian
neering, Xi’an Jiaotong University, China, in 1992, and
Bayesian tracking, IEEE Transactions on Signal Processing,
AC

his Ph.D. degree from Harbin Institute of Technology,

50 (2) (2002) 174–188.
China, in 1998. Since 2006, he has been a professor at
[38] S. M. Kay, Fundamentals of statistical signal processing: the Control and Simulation Center, School of Astronau-
estimation theory, Upper Saddle River, New Jersey: Prentice tics, Harbin Institute of Technology, China. His current
Hall, (1993). research interest includes nonlinear control, non-smooth
analysis and control, robust control and their applica-
[39] Y. Yuan, L. Guo, and Z. Wang, Composite control of linear
tions to guidance and control of flight vehicles.
quadratic games in delta domain with disturbance observers,
Journal of the Franklin Institute, 354 (4) (2017) 1673–1695.

11
 ACCEPTED MANUSCRIPT

interests include stochastic control,neural networks,

complex networks, nonlinear dynamics, time-delay sys-
tems, multi-agent systems, and chaotic dynamics.

T
IP
Wenshuo Lireceived the B.Eng. degree from Shan-
dong University, Jinan, China, in 2012. He is currently Fuad E. Alsaadi received the BSc and MSc degrees
working towards the Ph.D. degree in Beihang Univer- in electronic and communication from King AbdulAziz

CR
sity, Beijing, China. From Jan. 2016 to Jan. 2017, he University, Jeddah, Saudi Arabia, in 1996 and 2002.
was an exchange Ph.D. student in Brunel University He then received the PhD degree in optical wireless
London, UK. His research interests include nonlinear communication systems from the University of Leeds,
filtering, Monte Carlo methods, anti-disturbance con-
trol and filtering, and their applications in aerospace
and networked systems.
US Leeds, UK, in 2011. Between 1996 and 2005, he worked
in Jeddah as a communication instructor in the College
of Electronics & Communication. He is currently an
AN
assistant professor of the Electrical and Computer En-
gineering Department within the Faculty of Engineer-
ing, King Abdulaziz University, Jeddah, Saudi Arabia.
He published widely in the top IEEE communications
M

conferences and journals and has received the Carter

award, University of Leeds for the best PhD. He has re-
search interests in optical systems and networks, signal
ED

processing, synchronization and systems design.

PT
CE

Yurong Liu was born in China in 1964. He received his

B.Sc. degree in Mathematics from Suzhou University,
Suzhou, China, in 1986, the M.Sc. degree in Applied
AC

Mathematics from Nanjing University of Science and

Technology, Nanjing, China, in 1989, and the Ph.D. de-
gree in Applied Mathematics from Suzhou University,
Suzhou, China, in 2001. Dr. Liu is currently a profes-
sor with the Department of Mathematics at Yangzhou
University, China. He also serves as an Associate Editor
of Neurocomputing. So far, he has published more than
70 papers in refereed international journals. His current

12