3786 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 63, NO.

6, JUNE 2016

State Estimation in Nonlinear System Using

Sequential Evolutionary Filter
Shen Yin, Senior Member, IEEE, Xiangping Zhu, Student Member, IEEE,
Jianbin Qiu, Senior Member, IEEE, and Huijun Gao, Fellow, IEEE

Abstract—As a commonly encountered problem in the and UKF has been demonstrated in many literatures and lots
particle filters (PFs), the particle impoverishment is caused of the improved versions have been proposed [10], [12], [13].
partially by the reduction of particle diversity after resam- However, these nonlinear filters have the difficulties in dealing
pling. In this paper, a novel particle filtering technique
named sequential evolutionary filter (SEF) is introduced, by with the arbitrary distributions in the considered system.
which the particle impoverishment problem can be effec- The particle filter (PF) provides another effective tool for
tively mitigated. SEF is proposed based on the genetic the state estimation problem of nonlinear systems. Under the
algorithm (GA). A GA-inspired strategy is designed and Bayesian framework, the Monte Carlo simulation method is
incorporated in SEF. With this strategy, the resampling incorporated in PF. The posterior distribution is approximated
used in most of the existing PFs is not necessary, and
the particle diversity can be maintained. The experimental by a number of weighted samples (or particles). Compared
results also demonstrate the effectiveness of SEF. with EKF and UKF, PF offers several superior performances.
For example, it can be used in the system with non-Gaussian
Index Terms—Genetic algorithm (GA), nonlinear system,
distributions. And for the highly nonlinear system, it can gen-
particle filter (PF), sequential evolutionary filter (SEF), state
estimation. erate more accurate state estimations than EKF and UKF. In
the past two decades, PF has found a wide variety of applica-
I. I NTRODUCTION tions in the practical problems. In [14], both the unconstrained
and constrained state estimation problems using PF have been

A CCURATE estimation of the system states is important

for many practical applications, for example, process
monitoring [1]–[4], target tracking [5]–[7], robot navigation [8],
discussed. Das et al. proposed a visual tracking algorithm in
[15] based on PF, and its effectiveness has been demonstrated
in various face and vehicle tracking tasks. The topic of fault
[9], controller design [10], [11], etc. From the Bayesian per- diagnosis and prognosis based on PF can be found in [16] and
spective, the posterior distribution of the system states contains [17]. Besides, in the robotics community, PF has been studied
the information from both the system model and sensor mea- in navigation or localization [8], [9], tracking [5], [7], health
surements. With the obtained posterior distribution, the system management [18], etc.
states can be easily estimated by certain methods, e.g., mini- However, PF still suffers from several problems. Particle
mum mean-squared error (MMSE) and maximum a posteriori impoverishment is one of the commonly encountered problems.
(MAP). Starting from this observation, many nonlinear filtering In PF, the resampling procedure is incorporated to replicate
techniques have been developed under the Bayesian frame- the large-weight particles to replace the small-weight ones,
work to derive the state posterior distribution. The extended which are useless for approximating the posterior distribution.
Kalman filter (EKF) is the most commonly used nonlinear filter. This results in the reduction of particle diversity, which is
It deals with the nonlinearity of the considered system with the one of the reasons for the particle impoverishment problem. A
linear approximation around the operation point at each time simple strategy for mitigating this problem is to rough the over-
instant. The unscented Kalman filter (UKF) is also a nonlinear centralized particles with Gaussian jitter noise [19]. Besides, in
extension of the Kalman filter. Based on the unscented transfor- [20], a risk-sensitive PFs (RSPF) was proposed for mitigating
mation, it approximates the posterior distribution of the system particle impoverishment by constructing explicit risk functions.
states more accurately than EKF. The effectiveness of EKF On the other hand, the filtering problem can also be treated
as the sequential optimization problem. Tomoyuki revealed
Manuscript received August 12, 2015; revised November 13, 2015
and December 28, 2015; accepted January 1, 2016. Date of pub- that PF and genetic algorithm (GA) share the similar struc-
lication January 27, 2016; date of current version May 10, 2016. ture [21]. Based on these observations, many efforts have also
This work was supported in part by the National Natural Science been devoted to mitigating the particle impoverishment prob-
Foundation of China under Grant 61472104 and Grant 61304102, in
part by the Fundamental Research Funds for the Central Universities lem with GA. In [21], Tomoyuki proposed a GA filter (GAF).
under Grant HIT.BRETIV.201507, and in part by the Scientific Research Kwok et al. introduced an evolutionary PF for the prevention of
Foundation for Returned Scholars of Heilongjiang Province under Grant particle impoverishment [22]. Yin et al. proposed an intelligent
LC2015023. (Corresponding author: Xiangping Zhu.)
The authors are with Harbin Institute of Technology, PF (IPF) by incorporating the genetic operators in PF [23]. In
Harbin 150001, China (e-mail: [email protected]; [24], the real-coded GA PFs have been proposed for solving the
[email protected]; [email protected]; [email protected]). high-dimensional state-space problem. However, for most of
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. these existing GA-based PFs, the proposed strategies often act
Digital Object Identifier 10.1109/TIE.2016.2522382 as the additional part to be added in the generic PF. Compared
0278-0046 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 06,2023 at 15:47:57 UTC from IEEE Xplore. Restrictions apply.
YIN et al.: STATE ESTIMATION IN NONLINEAR SYSTEM USING SEF 3787

with the generic PF, although the particle impoverishment {x0 , . . . , xk } and y 1:k = {y 1 , . . . , y k }, respectively. As
problem has been mitigated to some degree, most of these aforementioned, PF approximates the posterior distribution of
improved PFs have the more complex strategies and suffer the state xk with a number of particles, and the approximated
from higher computational burdens. Besides the GA-based PF, form of the joint posterior distribution p(x0:k |y 1:k ) is
there are also some other evolutionary inspired PFs. More N

information can be found in [6] and [25]. p(x0:k |y 1:k ) ≈ wki δ(x0:k − xi0:k ) (2)
In this paper, a novel filter named sequential evolutionary i=1
filter (SEF) is proposed for mitigating the particle impover-
in which δ(·) denotes the Dirac delta function, N is the particle
ishment. It is inspired from the idea of GA. The reduction of
number, and wki is the normalized importance weight (or parti-
particle diversity introduced by resampling can be attributed to
cle weight) for the ith particle. With state xi0:k and weight wki ,
two reasons. The first one is the existence of too many repli-
the ith particle can be formulated as {xi0:k , wki }.
cates of large-weight particles. Second, if no particles located
The definition of the importance weight wk is
in the high probability regions of the posterior distribution, no
resampled particles will be generated in these regions. Based p(x0:k |y 1:k )
wk ∝ (3)
on these two reasons, a GA-inspired strategy is designed in q(x0:k |y 1:k )
SEF. With this strategy, the particle diversity can be maintained. where ∝ means to be proportional to. q(x0:k |y 1:k ) is the impor-
The small-weight particles can be modified into the particles tance distribution. In practice, it is usually difficult to directly
assigned with larger weights. And at the same time, these mod- sample from p(x0:k |y 1:k ) for its attributes such as multivari-
ified small-weight particles will not overlap the large-weight ate and nonstandard. The importance distribution q(x0:k |y 1:k )
ones. Different from most of the existing GA-based PFs, the is introduced to overcome this problem, and compared with
resampling procedure is not necessary in SEF and is replaced by p(x0:k |y 1:k ), it can be easily sampled.
the proposed GA-inspired strategy. The effectiveness of SEF is To derive the iterative formulation of the importance weight
demonstrated in the numerical example and three-tank system. wk , q(x0:k |y 1:k ) is chosen to satisfy
This paper is organized as follows. In Section II, the
brief introductions of the generic PF and GA are presented. q(x0:k |y 1:k ) = q(xk |x0:k−1 , y 1:k )q(x0:k−1 |y 1:k−1 ) (4)
Section III presents the detailed descriptions of the proposed
in which q(xk |x0:k−1 , x1:k ) is regarded as the marginal distri-
SEF. The results of the experiments on the numerical exam-
bution of q(x0:k |x1:k ).
ple and three-tank system are reported in Section IV. Finally,
Under the Bayesian framework, the joint posterior distribu-
Section V concludes this paper.
tion p(x0:k |x1:k ) can be written as
p(xk |x1:k−1 , x0:k )p(xk |x0:k−1 )
II. P RELIMINARIES O N G ENERIC PF (GPF) AND GA p(x0:k |x1:k ) =
p(xk |x1:k−1 )
In this paper, the considered dynamic state-space model is × p(x0:k−1 |y 1:k−1 ). (5)

xk = f(xk−1 , uk , wk ) The denominator part of (5) can be reformulated to
(1) 
y k = h(xk , uk , νk ), k = 1, 2, . . .
p(y k |y 1:k−1 ) = p(y k |xk )p(xk |y 1:k−1 )dxk (6)
where
xk ∈ Rnx and yk ∈ Rny denote the state and measure- in which
ment vectors; 
uk ∈ Rnu is the system input; p(xk |y 1:k−1 ) = p(xk |xk−1 )p(xk−1 |y 1:k−1 )dxk−1 . (7)
wk ∈ Rnw and νk ∈ Rnν denote the white process and
measurement noises. Since p(y k |xk ) and p(xk |xk−1 ) are known from the system
The functions f : Rnx × Rnu × Rnw → Rnx and h : model (1), and p(xk−1 |y 1:k−1 ) is the marginal posterior distri-
R × Rnu × Rnν → Rny are the transition and measure-
nx bution at (k − 1)th time step, p(y k |y 1:k−1 ) can be regarded as a
ment functions. The system model (1) can also be formulated constant factor. Considering the assumptions made on the states
in the statistical form, where the functions f and h, respectively, xk and measurements y k (k = 1, 2, . . .), the joint posterior
correspond to the transition and likelihood distributions, i.e., distribution p(x0:k |y 1:k−1 ) can be further simplified to
p(xk |xk−1 ) and p(y k |xk ). The initial distribution of the state p(x0:k |y 1:k ) ∝ p(y k |xk )p(xk |xk−1 )p(x0:k−1 |y 1:k−1 ). (8)
x0 is known as p(x0 ). In general, the assumptions are made
that xk (k = 1, 2, . . .) subject to the first-order Markov process, By substituting (4) and (8) into (3), a recursive formulation of
and y k (k = 1, 2, . . .) are conditionally independent given the the importance weight wk can be derived as
states. p(y k |xk )p(xk |xk−1 )p(x0:k−1 |y 1:k−1 )
wk ∝ (9)
q(xk |x0:k−1 , y 1:k )q(x0:k−1 |y 1:k−1 )
A. GPF
p(y k |xk )p(xk |xk−1 )
In order to give a brief introduction of GPF, the sequences = wk−1 . (10)
of the states and measurements are denoted as x0:k = q(xk |x0:k−1 , y 1:k )

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 06,2023 at 15:47:57 UTC from IEEE Xplore. Restrictions apply.
3788 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 63, NO. 6, JUNE 2016

TABLE I
M ULTINOMIAL R ESAMPLING

Fig. 1. Intuitively illustration of the particle impoverishment problem

introduced by resampling procedure. The particle weight are presented
as the size of the filled circle.

evolution. Corresponding to the individuals in nature, a num-

ber of potential optimal solutions are created from the solution
In practical applications, the marginal importance distribution space of the optimal problem, and they are updated in every
q(xk |x0:k−1 , y 1:k ) is always chosen as generation. The main parts of GA consist of three genetic
operators, i.e., selection, crossover, and mutation. The evalu-
q(xk |x0:k−1 , y 1:k ) = p(xk |xk−1 ). (11) ation function evaluates the fitness of all potential solutions and
By substituting (11) into (4), the formulation of wk can be assigns the fitness values to them. The selection operator simu-
further simplified to satisfy lates the principle of “survival of the fittest,” and only the highly
fitted potential solutions survive in every generation. These sur-
wk ∝ p(y k |xk )wk−1 . (12) vived solutions can participate the following crossover proce-
dure, in which the survived solutions act as the parent solutions
With (12), the importance weight can be recursively calcu- and exchange part of their information to produce the offspring
lated. If only the marginal posterior distribution p(xk |y 1:k ) is solutions. Just like the individuals in nature, the mutation may
required, the historical sequence state x0:k−1 is not necessary also happen on the offspring solutions by altering part of their
to be stored, and the marginal posterior distribution p(xk |y 1:k ) information. The mutation operator effectively prevents the GA
can be approximated as from tripping in the local optimization.
N
 Due to the robustness and effectiveness in searching global
p(xk |y 1:k ) ≈ wki δ(xk − xik ). (13) solution of the complex optimization problems, the GA has
i=1 been widely applied [27]–[29]. The traditional GAs involve
In the following, only the marginal posterior distribution coding the potential solutions into strings, for example, the
p(xk |y 1:k ) will be considered as the posterior distribution of commonly used binary strings. However, these GAs always
the state. encounter the expensive computation problem, especially when
If the particles {xik , wki } (i = 1, 2, . . .) are directly used the space of the optimization problem is large. The real-
for the (k + 1)th generation, the particle population will suf- coded GAs are another category of GAs, and there is no
fer from serious degeneracy problem. After a few generations, coding procedure needed. Compared with the binary-coded
all but one particles will be assigned with the small weights, GAs, the real-coded GAs possess some superior properties,
which are negligible. The effective sample size Neff [26], is e.g., being more consistent, faster, and more precise. For more
proposed to measure the degree of the particle degeneracy. One detailed descriptions on real-coded GAs, the reader can refer to
commonly used approximation of Neff is [30], [31].
1
Neff ≈ N . (14)
i 2
i=1 (wk ) III. S EQUENTIAL E VOLUTIONARY F ILTER
To solve the particle degeneracy problem, the resampling With the incorporated resampling procedure, the particle
procedure is included in GPF. In each time step, the effec- degeneracy problem is solved in the generic PF. The small-
tive sample size of the particle population is calculated and weight particles are effectively eliminated, and the duplicates
compared with its threshold NT , which can be set manually. of the large-weight particles are used to supplement these elim-
If Neff < NT , the resampling is performed. By resampling, inated particles. However, as mentioned in [26], the resampling
the small-weight particles are replaced by the duplicates of the procedure also introduces new problems for the generic PF.
large-weight particles. For the GPF introduced in [26], the muti- One of the problems is the particle impoverishment, which
nonial resampling is used, and it is briefly formulated in Table I. results from the reduction of particle diversity after resampling.
Fig. 1 intuitively shows the particle impoverishment problem
introduced by the resampling procedure.
In the proposed SEF, the small-weight particles are changed
B. Genetic Algorithm into the large-weight ones. And at the same time, the parti-
GA is a category of population-based optimization algo- cle diversity is also maintained. Compared with the generic
rithms. It is designed by mimicking the process of biological PF, SEF mitigates the particle impoverishment problem. It is
Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 06,2023 at 15:47:57 UTC from IEEE Xplore. Restrictions apply.
YIN et al.: STATE ESTIMATION IN NONLINEAR SYSTEM USING SEF 3789

  
inspired from the idea of GA. An effective strategy, i.e., GA- χ̃L = {x̃1kL , w̃kL
1
}, . . . , {x̃m m
kL , w̃kL }
inspired strategy, is implemented in SEF. In the following, this  . (21)
strategy will be first introduced, and then the full procedure of χ̃S = {x̃1kS , w̃kS
1 N −m
}, . . . , {x̃kS N −m
, w̃kS }
SEF will be presented.
The state of the moved particle is represented by xjkO . Then,
the crossover operator in GA-inspired strategy is
A. GA-Inspired Strategy
x̃jkO = αj x̃jkS + (1 − αj )x̃lkL . (22)
The GA-inspired strategy is designed to maintain the parti-
cle diversity, and finally, mitigate the particle impoverishment Similar to the arithmetic crossover, αj is also a random number,
problem encountered in GPF. Similar to GA, three operators, but it is drawn from the uniform distribution over [0, ᾱ], i.e.,
i.e., separation, crossover, and mutation, constitute the main αj ∼ U (0, ᾱ). ᾱ is the upper bound of αj , and in every step, it
part of the GA-inspired strategy. But, these three operators are can be calculated as
different from the genetic operators in GA. They are developed
according to the attributes of PF. ᾱ = 1 − Neff /N. (23)
Assume that the particles have been obtained and they are
formulated as {xik , wki } (i = 1, 2, . . . , N ). The separation The value of αj determines how much information will be
operator aims to separate the particle population into two cate- transferred from x̃lkL to x̃jkO . The smaller the value of αj is the
gories. One is composed of the small-weight particles and the more information of x̃lkL will be transferred. For each x̃jkS , x̃lkL
other is composed of the large-weight particles. To achieve this is randomly selected from χ̃L . Different from the arithmetic
separation, all the particles are sorted first according to their crossover, the modifications are only made on the small-weight
weights particles, and only one offspring particle, i.e., x̃jkO , is created
  by the crossover in the GA-inspired strategy. For the large-
χ̃ = Sortdesc. {x1k , wk1 }, . . . , {xik , wkN } (15) weight particles, no modifications are made on them, since
 1 1 i
 they are important for approximating the posterior distribution
= {x̃k , w̃k }, . . . , {x̃k , w̃kN } (16)
p(xk |y 1:k ).
where Sortdesc. { · } is the operation used to sort the particles in In order to further improve the diversity of particle popu-
descending order according to their weights wki . {x̃ik , w̃ki } rep- lation, the mutation operator is designed to extend the state-
resents the sorted particle. The separation is performed on these search space of particles. The definition is
sorted particles  l
2x̃kL − x̃jkO , if rj ≤ pM
   j
x̃kM = (24)
χ̃L = {x̃1k , w̃k1 }, . . . , {x̃m m
k , w̃k }  x̃jkO , if rj > pM
(17)
χ̃S = {x̃m+1 k , w̃km+1 }, . . . , {x̃N N
k , w̃k }
where x̃lkL and x̃jkO are the particle states in (23). rj ∼
in which χ̃L is the category including large-weight particles,
U (0, 1), and pM ∈ [0, 1] is the mutation probability. When
and χ̃S is the category including the small-weight particles. m
rj ≤ pM , the mutation will be performed on x̃jkO .
is a integer, and
After performing the GA-inspired strategy, the small-weight
m ≤ γNeff < m + 1. (18) particles will be modified into the ones with the large weights.
In the next time step, all particles will be treated equally. Based
where Neff is the effective sample size calculated by (14). γ ∈ on this idea, the particle weights will be reset to 1/N after the
(0, 1] is a parameter that is set in advance. GA-inspired strategy as
In the real-coded GAs, a commonly used crossover is the 
χ = { x̃1kL , 1/N }, . . . , {x̃m
kL , 1/N }
arithmetic crossover [32]. Suppose that two of the selected par-
(N −m)
ent potential solutions are represented as x1par and x2par , and {x̃1kM , 1/N }, . . . , {x̃kM , 1/N }} (25)
their offsprings are represented as x1off and x2off . The arithmetic  
= { x1k , 1/N }, . . . , {xN
k , 1/N } . (26)
crossover is represented as
To intuitively show the GA-inspired strategy, a diagram under
x1off = ax1par + (1 − a)x2par (19) the one dimension situation is presented in Fig. 2. The sizes
x2off = ax2par + (1 − a)x1par (20) of the circles represent the particle weights. Different colored
circles denote the particles obtained from different stages. The
where a is a random number drawn from the interval [0, 1]. circles filled with navy blue correspond to the parent particles in
The crossover operator used in the GA-inspired strategy is Fig. 1. The circles filled with red and orange denote the small-
modified from the arithmetic crossover. It is designed to mod- and large-weight particles, respectively. The circles filled with
ify the small-weight particles into the particles with the large light blue are the particle created by mutation. And the circle
weights. Denote the state of the large-weight particle as x̃lkL , filled with no color is the final particle, i.e., {xik , 1/N }, created
and the state of the small-weight particle as x̃jkS , in which by GA-inspired strategy.
l = 1, . . . , m and j = 1, . . . , N − m. Hence, the χ̃L and χ̃S Different from the resampling used in GPF, which only repli-
can be reformulated as cates the large-weight particles to replace the small-weight
Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 06,2023 at 15:47:57 UTC from IEEE Xplore. Restrictions apply.
3790 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 63, NO. 6, JUNE 2016

TABLE II
I MPLEMENTATION P ROCEDURE OF SEF

Fig. 2. Diagram of the GA-inspired strategy.

ones, the GA-inspired strategy modifies the the small-weight

particles into the large-weight one. But, these modified small-
weight particles will not overlap the particles with large
the process control [33], [34]. To evaluate the performances of
weights. In this way, the diversity of the particle popula-
SEF, three of other PFs are included for comparison. One is
tion can be maintained, and thus the particle impoverishment
the GPF introduced in Section II-A. The second is the modi-
encountered in GPF is mitigated.
fied PF (RPF) introduced in [19]. Its main idea is to rough the
over-centralized particles with the Gaussian jitter noise. The
B. Sequential Evolutionary Filter roughing procedure can also be recognized as performing the
mutation operator of GA. The last one is called the real-coded
Once the weighted particles {xik , wki } (i = 1, . . . , N ) have GA PF (RGAPF) that is based on the real-coded GA [24].
been obtained, the system state at the kth time can be estimated
by appropriate methods, for example, MMSE and MAP. The
definition of MMSE is A. Numerical Example
N
 The process and measurement equations of the system can
x̂M
k
M SE
= wki xik (27) be formulated as
i=1
1 25xk−1
xk = xk−1 + + 8 cos[1.2(k − 1)] + ωk (30)
and the definition of MAP is 2 1 + x2k−1
1 2
x̂M AP
= arg max wki . (28) yk = x + νk , k = 1, 2, . . . (31)
k
xik 20 k

The choice of MMSE or MAP depends on different estimation where ωk ∼ N (0, σω2 ) and νk ∼ N (0, σν2 ) represent the pro-
problems. In the experimental part of this work, the MMSE is cess and measurement noises, respectively.
√ In the experiment,
used to estimate the system states. the variances are set as σω = 1, σν = 2. xk is the system state
In the GA-inspired strategy, a simpler formulation of the that needs to be estimated. Its initial value is set to x0 = 0.1.
crossover and mutation operators can be derived. Assume that The initial particles {xi0 , 1/N } (i = 1, . . . , N ) are drawn from
we have obtained the separated particles, i.e., {x̃lkL , wkL
l
}∈ a Gaussian distribution, i.e., xi0 ∼ N (0.1, 2).
j j In this experiment, the particle number is set to N = 500.
χ̃L and {x̃kS , wkS } ∈ χ̃S . Then, the simpler formulation can
Both the thresholds of Neff for GPF and RPF are set to
be presented as
NT = 0.5N . The constant tuning parameter K in RGPF is

(αj + 1)x̃lkL − αj x̃jkS , if rj ≤ pM set to K = 0.2 [19]. The parameters in SEF are set as pM =
j
x̃kM = (29) 0.5 and γ = 1. A metric, i.e., the absolute error of state, is
αj x̃jkS + (1 − αj )x̃lkL , if rj > pM defined as
n
which combines the crossover and mutation operators. 1
Compared with the GPF, SEF replaces the resampling with e= |xk − x̂k | (32)
n
k=1
the designed GA-inspired strategy, with which the particle
diversity is maintained and the particle impoverishment in where n is the number of the time steps. The experiment is
GPF is mitigated. A brief implementation procedure of SEF is repeatedly performed for 30 times. To evaluate the performance
presented in Table II. of these four PFs, the averaged absolute error ūe and variance
of the absolute error Var are calculated. To estimate the com-
putational burdens of four PFs, the averaged running time ūt is
IV. E XPERIMENTS recorded.
In this section, two experimental systems are used. The first The experiment results are presented in Table III. As shown
one includes a commonly used numerical example [4], [20]. in the table, four PFs give the comparative estimation accura-
The highly nonlinear and bimodal characteristics of this exam- cies which can be reflected by the average absolute error ūe .
ple bring challenges for state estimation issues. The other one The variances Var of four PFs are all very small. This means
is the three-tank system that always acts as the benchmark in that these four PFs can offer stable state estimation results.
Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 06,2023 at 15:47:57 UTC from IEEE Xplore. Restrictions apply.
YIN et al.: STATE ESTIMATION IN NONLINEAR SYSTEM USING SEF 3791

TABLE III TABLE IV

E XPERIMENTAL R ESULTS OF F OUR PF S IN THE N UMERICAL E XAMPLE PARAMETERS OF THE T HREE -TANK S YSTEM

flow for the ith pipeline, respectively. g is the acceleration due

to gravity, and g = 980 cm/s2 . sgn( · ) is the sign function. Q20
is the flow rate in the pipeline, which outputs the water in tank
2, and it is modeled as

Q20 = a2 s2 2gh2 . (37)

Let x = [x1 , x2 , x3 ]T = [h1 , h2 , h3 ]T and u = [u1 , u2 ]T =

[Q1 , Q2 ]T . In this paper, both the process and measurement
Fig. 3. Three-tank system. noises are considered, and the dynamic state-space model of
the three-tank system can be formulated as
However, the average running time ūt of four PFs are largely
1
different from each other. Compared with GPF, the extra part, ẋ1 = u1 − a1 s1 sgn(x1 − x3 ) 2g|x1 − x3 | + w1
e.g., Gaussian jitter noise or genetic operators, is added in the A
(38)
RPF and RGAPF to mitigate the particle impoverishment. The
added extra part directly results in the higher computational 1
ẋ2 = u2 + a3 s3 sgn(x3 − x2 ) 2g|x3 − x2 |
burden of RPF and RGAPF than GPF’s. From the table, it can A
also be observed that the lowest computational burden among
− a2 s 2 2gx2 + w2 (39)
these four PFs belongs to SEF with ūt = 0.036 s.
1
ẋ3 = a1 s1 sgn(x1 − x3 ) 2g|x1 − x3 |
B. Three-Tank System A
In the practical industrial application, the tank system always − a3 s3 sgn(x3 − x2 ) 2g|x3 − x2 | + w3 (40)
serves as the important part of the plant, especially for the
chemical or biopharmaceutical process. The three-tank system y =x+ν (41)
introduced here is a simplified tank system used in the practi-
cal plant. It acts as a popular nonlinear system for the process in which x = [x1 , x2 , x3 ]T , y = [y1 , y2 , y3 ]T , and ν =
control [33], [34]. The simplified form of the three-tank sys- [ν1 , ν2 , ν3 ]T . wi (i = 1, 2, 3) are the process noises, and ν is
tem is presented in Fig. 3. Three tanks constitute the main part the measurement noise. Here, both wi and ν are assumed to
of the system, and they share the same section area A. These be subject to the Gaussian distribution, i.e., wi ∼ N (0, 2) and
three tanks are connected by pipelines. The water is pumped ν ∼ N (0, Σν ), in which Σν = diag(1, 1, 1). The parameters
into the tanks 1 and 2 with the flow rate Q1 and Q2 , respec- of the system are given in Table IV.
tively. According to the “mass balance” principle, the system In this section, the performance of the introduced SEF will
can be mathematically formulated as be evaluated on the three-tank system. The parameters of SEF
are set as: pM = 0.5 and γ = 0.5. The particle number for
dh1 1
= (Q1 − Q13 ) (33) four discussed filters is set to N = 1000. As in the numeri-
dt A cal example, the thresholds of Neff for GPF and RPF are set
dh2 1 to NT = 0.5N . The initial particles are drawn from p(x̂0 ) =
= (Q2 + Q32 − Q20 ) (34)
dt A N (x̂0 , Σ0 ), in which x̂0 = x0 = [50.93, 23.94, 37.73]T and
dh3 1 Σ0 = diag(2, 2, 2). diag( · ) denotes the diagonal matrix. The
= (Q13 − Q32 ) (35) experiment is performed for 100 s and the data are collected by
dt A
the sampling period of 0.1 s. The system input is
in which h1 , h2 , and h3 are the water levels in three tanks. Qij 
denotes the flow rate in the pipeline from the ith to jth tank, [10, 50]T , t = 20 ∼ 60 s
and it can be modeled as u= (42)
[37, 28]T , others
Qij = ai si sgn(hi − hj ) 2g|hi − hj |. (36)
in which u = [u1 , u2 ]T .
In Fig. 3, the pipelines are numbered 1, 3, and 2 from left to The state estimation results of four filters are presented in
right. si and
Authorized ai represent
licensed the
use limited to: sectionalOF
UNIVERSITY area and coefficient
SOUTHAMPTON. of Figs.
Downloaded 4–7. The
on December black
06,2023 and red
at 15:47:57 lines,
UTC respectively,
from IEEE represent
Xplore. Restrictions the
apply.
3792 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 63, NO. 6, JUNE 2016

Fig. 4. State estimation results of SEF. Fig. 7. State estimation results of RGAPF.

TABLE V
AVERAGED A BSOLUTE E RRORS OF SEF AND GPF

in GPF is mitigated. GPF and RPF offer the similar state esti-
mation results. The worst state estimation results belong to the
RGAPF. The simulations are performed for 10 times, and the
averaged absolute errors of four filters are presented in Table V.
The table also shows that SEF offers the more accurate state
Fig. 5. State estimation results of GPF.
estimation results than GPF, RPF, and RGAPF. In the exper-
iment, we find that the variances of the absolute errors of all
four filters for 10 times experiments approximate to 0. This
means that all these four filters can offer stable state estima-
tion results. The averaged running time ūt for SEF, GPF, RPF,
and RGAPF are 14.573, 11.362, 14.591, and 15.053 s, respec-
tively. As the averaged running time recorded in the numerical
example, the ūt of RPF and RGAPF are larger than GPF’s.
But, for this MIMO system, the ūt of SEF is also greater than
GPF’s.

V. C ONCLUSION
In this paper, an SEF technique, which is based on the PF
and GA, has been introduced. It replaces the resampling in GPF
with a GA-inspired strategy, by which the particle impoverish-
ment encountered in GPF is mitigated. The main idea of the
Fig. 6. State estimation results of RPF.
GA-inspired strategy is to modify the small-weight particles
into the particles with large weights, and at the same time, main-
estimated and real states. From these figures, it can be observed tain the diversity of the particle population. The performances
that SEF provides the best state estimation. This mainly owes to of SEF are evaluated by two experiments. The experimental
the GA-inspired strategy. With this strategy, the particle diver- results show that compared with three other filters, SEF offers
sity is maintained, and the particle impoverishment encountered the better state estimation results.
Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 06,2023 at 15:47:57 UTC from IEEE Xplore. Restrictions apply.
YIN et al.: STATE ESTIMATION IN NONLINEAR SYSTEM USING SEF 3793

R EFERENCES [26] M. S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, “A tutorial

on particle filters for online nonlinear/non-Gaussian Bayesian tracking,”
[1] A. Shenoy, J. Prakash, V. Prasad, S. Shah, and K. McAuley, “Practical IEEE Trans. Signal Process., vol. 50, no. 2, pp. 174–188, Feb. 2002.
issues in state estimation using particle filters: Case studies with polymer [27] M. Paulinas and A. Ušinskas, “A survey of genetic algorithms applica-
reactors,” J. Process Control, vol. 23, no. 2, pp. 120–131, 2013. tions for image enhancement and segmentation,” Inf. Technol. Control,
[2] S. Yin, S. X. Ding, X. Xie, and H. Luo, “A review on basic data- vol. 36, no. 3, pp. 278–284, 2007.
driven approaches for industrial process monitoring,” IEEE Trans. Ind. [28] V. Roberge, M. Tarbouchi, and G. Labonté, “Comparison of parallel
Electron., vol. 61, no. 11, pp. 6418–6428, Nov. 2014. genetic algorithm and particle swarm optimization for real-time UAV path
[3] G. G. Rigatos, “A derivative-free Kalman filtering approach to state planning,” IEEE Trans. Ind. Informat., vol. 9, no. 1, pp. 132–141, Feb.
estimation-based control of nonlinear systems,” IEEE Trans. Ind. 2013.
Electron., vol. 59, no. 10, pp. 3987–3997, Oct. 2012. [29] Y. Yoon and Y.-H. Kim, “An efficient genetic algorithm for maximum
[4] S. Yin, X. Zhu, and O. Kaynak, “Improved PLS focused on coverage deployment in wireless sensor networks,” IEEE Trans. Cybern.,
key-performance-indicator-related fault diagnosis,” IEEE Trans. Ind. vol. 43, no. 5, pp. 1473–1483, Oct. 2013.
Electron., vol. 62, no. 3, pp. 1651–1658, Mar. 2015. [30] K. Deep, K. P. Singh, M. Kansal, and C. Mohan, “A real coded genetic
[5] I. Miller, M. Campbell, and D. Huttenlocher, “Efficient unbiased track- algorithm for solving integer and mixed integer optimization problems,”
ing of multiple dynamic obstacles under large viewpoint changes,” IEEE Appl. Math. Comput., vol. 212, no. 2, pp. 505–518, 2009.
Trans. Robot., vol. 27, no. 1, pp. 29–46, Feb. 2011. [31] B. A. Sawyerr, M. M. Ali, and A. O. Adewumi, “A comparative
[6] X. Zhang, W. Hu, S. Maybank, X. Li, and M. Zhu, “Sequential particle study of some real-coded genetic algorithms for unconstrained global
swarm optimization for visual tracking,” in Proc. IEEE Conf. Comput. optimization,” Optim. Method Softw., vol. 26, no. 6, pp. 945–970, 2011.
Vis. Pattern Recognit. (CVPR), 2008, pp. 1–8. [32] T. Yalcinoz, H. Altun, and M. Uzam, “Economic dispatch solution using
[7] P. Vadakkepat and L. Jing, “Improved particle filter in sensor fusion for a genetic algorithm based on arithmetic crossover,” in Proc. IEEE Proto
tracking randomly moving object,” IEEE Trans. Instrum. Meas., vol. 55, Power Tech. Conf., Proto, Portugal, Sep. 2001, p. 4.
no. 5, pp. 1823–1832, Oct. 2006. [33] S. X. Ding, Model-Based Fault Diagnosis Techniques: Design Schemes,
[8] G. Grisetti, C. Stachniss, and W. Burgard, “Improved techniques for grid Algorithms, and Tools. New York, NY, USA: Springer, 2008.
mapping with Rao-Blackwellized particle filters,” IEEE Trans. Robot., [34] R. Dormido et al., “Development of a web-based control laboratory
vol. 23, no. 1, pp. 34–46, Feb. 2007. for automation technicians: The three-tank system,” IEEE Trans. Educ.,
[9] M. M. Atia, J. Georgy, M. Korenberg, and A. Noureldin, “Real-time vol. 51, no. 1, pp. 35–44, Feb. 2008.
implementation of mixture particle filter for 3D RISS/GPS integrated
navigation solution,” Electron. Lett., vol. 46, no. 15, pp. 1083–1084, Jul.
2010. Shen Yin (M’12–SM’15) received the B.E.
[10] S. Chen, “Kalman filter for robot vision: A survey,” IEEE Trans. Ind. degree in automation from Harbin Institute of
Electron., vol. 59, no. 11, pp. 4409–4420, Nov. 2012. Technology, Harbin, China, in 2004, and the
[11] S. Yin, P. Shi, and H. Yang, “Adaptive fuzzy control of strict-feedback M.Sc. degree in control and information systems
nonlinear time-delay systems with unmodeled dynamics,” IEEE Trans. and the Ph.D. degree in electrical engineering
Cybern., DOI: 10.1109/TCYB.2015.2457894. and information technology from the University
[12] M. A. Khanesar, E. Kayacan, M. Teshnehlab, and O. Kaynak, “Extended of Duisburg-Essen, Essen, Germany, in 2007
Kalman filter based learning algorithm for type-2 fuzzy logic systems and and 2012, respectively.
its experimental evaluation,” IEEE Trans. Ind. Electron., vol. 59, no. 11, Currently, he is a Full Professor in the School
pp. 4443–4455, Nov. 2012. of Astronautics, Harbin Institute of Technology.
[13] A. Germani, C. Manes, and P. Palumbo, “Polynomial extended Kalman His research interests include model-based and
filter,” IEEE Trans. Automat. Control, vol. 50, no. 12, pp. 2059–2064, data-driven fault diagnosis, fault-tolerant control, and big data focused
Dec. 2005. on industrial electronics applications.
[14] J. Prakash, S. C. Patwardhan, and S. L. Shah, “On the choice of impor- Prof. Yin currently serves as the Technical Committee Chair of the
tance distributions for unconstrained and constrained state estimation IEEE Industrial Electronics Society Data Driven Control and Monitoring
using particle filter,” J. Process Control, vol. 21, no. 1, pp. 3–16, 2011. Committee. He is an Associate Editor for the IEEE T RANSACTIONS
[15] S. Das, A. Kale, and N. Vaswani, “Particle filter with a mode tracker ON I NDUSTRIAL E LECTRONICS , Journal of the Franklin Institute,
for visual tracking across illumination changes,” IEEE Trans. Image Neurocomputing, International Journal of Applied Mathematics and
Process., vol. 21, no. 4, pp. 2340–2346, Apr. 2012. Computer Science, IE Technology News, Journal of Intelligent and
[16] A. S. Sarathi Vasan, B. Long, and M. Pecht, “Diagnostics and prognos- Fuzzy Systems, etc.
tics method for analog electronic circuits,” IEEE Trans. Ind. Electron.,
vol. 60, no. 11, pp. 5277–5291, Nov. 2013.
[17] E. Zio and G. Peloni, “Particle filtering prognostic estimation of the
remaining useful life of nonlinear components,” Reliab. Eng. Syst. Safety, Xiangping Zhu (S’14) is currently working
vol. 96, no. 3, pp. 403–409, 2011. toward the Ph.D. degree in control science and
[18] Z. Duan, Z. Cai, and J. Yu, “Fault diagnosis and fault tolerant control for engineering at Harbin Institute of Technology,
wheeled mobile robots under unknown environments: A survey,” in Proc. Harbin, China.
IEEE Int. Conf. Robot. Autom., Apr. 2005, pp. 3428–3433. His research interests include complex
[19] N. J. Gordon, D. J. Salmond, and A. F. Smith, “Novel approach to system modeling, data-driven fault diagno-
nonlinear/non-Gaussian Bayesian state estimation,” Proc. Inst. Elect. sis, nonlinear filters, and their applications in
Eng. Radar Signal Process., vol. 140, no. 2, pp. 107–113, Apr. 1993. industry.
[20] U. Orguner and F. Gustafsson, “Risk-sensitive particle filters for miti-
gating sample impoverishment,” IEEE Trans. Signal Process., vol. 56,
no. 10, pp. 5001–5012, Oct. 2008.
[21] T. Higuchi, “Monte Carlo filter using the genetic algorithm operators,”
J. Stat. Comput. Simul., vol. 59, no. 1, pp. 1–23, 1997. Jianbin Qiu (M’10–SM’15) received the B.Eng.
[22] S. Park, J. P. Hwang, E. Kim, and H.-J. Kang, “A new evolutionary par- and Ph.D. degrees in mechanical and electri-
ticle filter for the prevention of sample impoverishment,” IEEE Trans. cal engineering from the University of Science
Evol. Comput., vol. 13, no. 4, pp. 801–809, Aug. 2009. and Technology of China (USTC), Hefei, China,
[23] S. Yin and X. Zhu, “Intelligent particle filter and its application on fault in 2004 and 2009, respectively, and the Ph.D.
detection of nonlinear system,” IEEE Trans. Ind. Electron., vol. 62, no. 6, degree in mechatronics engineering from the
pp. 3852–3861, Jun. 2015. City University of Hong Kong, Kowloon, Hong
[24] M. S. Hussain, “Real-coded genetic algorithm particle filters for high- Kong, in 2009.
dimensional state spaces,” Dept. Comput. Sci., Ph.D. dissertation, Univ. Since 2009, he has been with the School
College London, London, U.K., 2014. of Astronautics, Harbin Institute of Technology,
[25] J. Zhong and Y.-F. Fung, “Case study and proofs of ant colony optimisa- Harbin, China, where he is currently a Full
tion improved particle filter algorithm,” IET Control Theory Appl., vol. 6, Professor. His research interests include intelligent and hybrid control
no. 5, pp. 689–697, Mar. 2012. systems, signal processing, and robotics.

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 06,2023 at 15:47:57 UTC from IEEE Xplore. Restrictions apply.
3794 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 63, NO. 6, JUNE 2016

Huijun Gao (SM’09–F’14) received the Ph.D.

degree in control science and engineering from
Harbin Institute of Technology, Harbin, China, in
2005. From 2005 to 2007, he carried out his
postdoctoral research with the Department of
Electrical and Computer Engineering, University
of Alberta, Edmonton, AB, Canada.
Since November 2004, he has been with
Harbin Institute of Technology, where he is cur-
rently a Professor and Director of the Research
Institute of Intelligent Control and Systems. His
research interests include network-based control, robust control/filter
theory, time-delay systems, and their engineering applications.
Dr. Gao is a Co-Editor-in-Chief for the IEEE T RANSACTIONS ON
I NDUSTRIAL E LECTRONICS, and an Associate Editor for Automatica,
IEEE T RANSACTIONS ON C YBERNETICS, IEEE T RANSACTIONS ON
F UZZY S YSTEMS, IEEE/ASME T RANSACTIONS ON M ECHATRONICS,
and IEEE T RANSACTIONS ON C ONTROL S YSTEMS T ECHNOLOGY. He is
serving on the Administrative Committee of IEEE Industrial Electronics
Society. He was the recipient of the IEEE J. David Irwin Early Career
Award From IEEE Industrial Electronics Society.

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on December 06,2023 at 15:47:57 UTC from IEEE Xplore. Restrictions apply.