Professional Documents
Culture Documents
s12555-021-1060-6
s12555-021-1060-6
s12555-021-1060-6
http://dx.doi.org/10.1007/s12555-021-1060-6 http://www.springer.com/12555
Abstract: In this study, a novel distributed Kalman filter based on a possibilistic framework was proposed to mit-
igate fuzzy noisein nonlinear multiagent systems. To describe fuzzy uncertainty, noises were modeled as fuzzy
random variables with trapezoidal probability distributions instead of Gaussian distributions. A fuzzy information
fusion (FIF) algorithm was proposed to fuse fuzzy state estimations from neighboring nodes. The nonlinear problem
was solved by using local linearization. A distributed extended fuzzy information filter was designed by combining
the FIF algorithm and local linearization in distributed sensor networks. The stability of this filter was analyzed.
Finally, a target tracking simulation was performed to detail the effectiveness of the proposed filter algorithm.
Manuscript received December 15, 2021; revised May 11, 2022 and June 15, 2022; accepted June 16, 2022. Recommended by Associate
Editor Sung Hyun Kim under the direction of Senior Editor Choon Ki Ahn.
Xiaobo Zhang, Haoshen Lin, Gang Liu, and Bing He are with the Department of Electronic Engineering, PLA Rocket Force University of
Engineering, Xi’an, Shaanxi 710025, China (e-mails: {zhangxiaobo102800, linhaoshen1, liug029}@163.com, hb830513@126.com).
* Corresponding author.
The area, center of gravity, and uncertainty of the dis- In this study, a suitable distributed fuzzy filtering algo-
tribution are given as follows [15]: rithm was formulated for a nonlinear system with fuzzy
Z noise to ensure that at any time t ∈ Z+ , each agent i can ob-
χx = πX (x)dx, (5) (1) (2) (3)
tain its local posterior estimation E[x̂i,t ] ∼ Π(x̂i,t ; x̂i,t ; x̂i,t ;
X
R (4) (1) (2) (3)
xπX (x)dx x̂i,t ) by fusing {y j,t , R j,t } and E[x̄ j,t ] ∼ Π(x̄ j,t ; x̄ j,t ; x̄ j,t ;
C[x] = x̃ = , (6) (4)
χx from neighbors j ∈ Ni,in (t) based on local measure-
x̄ j,t )
P̂x = U[xC[(x − x̃)2 ]. (7) ment (2) and the global equation of motion (1). Specifi-
The expectation, central gradient, and uncertainty of wt cally, we focused on solving the following problems. First,
and vi,t are expressed as follows [15]: how can the TPDs obtained from different neighbors be
fused? Second, how should the nonlinear problem be han-
(1) (2) (3) (4)
E[wt ] ∼ Π(wt ; wt ; wt ; wt ),
dled and how should the local estimation be updated for
w̃t = C[wt ] = 0, (8) each agent such that the proposed filter is fully distributed?
Third, how can the boundedness of the estimation errors
U[wt ] = Qt .
of the proposed algorithm be analyzed?
(1) (2) (3) (4)
E[vi,t ] ∼ Π(vi,t ; vi,t ; vi,t ; vi,t ),
The following common DKF assumptions are made
ṽi,t = C[vi,t ] = 0, (9) [28].
Assumption 1: The undirected graph G is connected,
U[vi,t ] = Ri,t .
and the system is observable.
Remark 1: Studies have assumed that noises wk and
Assumption 2: The true state xt is independent of vi,t ,
vi,k follow some probability distribution (e.g., Gaus-
and xt 0 is independent of wt [2], ∀t,t 0 ∈ Z+ and t 0 > t, that
sian distribution). However, in practice, in many sys-
is
tems,noiseis modeled as FRVS. For example, Matía mod-
eled the data from the LIDAR sensor on the Doris robot C[xt v> >
i,t ] = 0, C[xt 0 wt ] = 0; (10)
platform as a TPD because the complexity and inaccuracy
of the calibration method cause distrust[34]. Furthermore, and for each agent
if the noise does not follow a “bell-curve” shape, is unbal-
C[vi,t v>j,t ] = 0, ∀i 6= j, (11)
anced, or some distrust exists about the subjective value, a
possible region can be used to cover changing noise char- C[(x̄i,t − xt )v>j,t ] = 0, ∀i, j. (12)
acteristics. Moreover, noises in some physical systems can
Assumption 3: The functions f and hi are twice con-
be modeled as FRVs rather than random variables [23].
tinuously differentiable on Rn , where n = dim(x).
For example, when using measurement tools that directly
detail the accuracy interval (e.g., a radar sensor with an
accuracy of +/ − 2 mm [24], a thermocouple with an ac- 3. DEFIF FRAMEWORK
curacy of +/ − 2.2◦ C [25]), the distribution of the interval To address the problem of distributed estimation for
values within the system will be accurate. nonlinear system with fuzzy noise, we first briefly review
Remark 2: FRVs exhibit many distribution shapes. We the extended FKF (EFKF) in [15] and FIF algorithm in
restricted the analysis to the trapezoidal distribution be- [21]. Next, we provide the DEFIF algorithm.
cause in practice, a criterion to define the distribution Consider nonlinear dynamics (1). The update steps of
shape between the possible and the impossible regions the EFKF proposed in [15] are as follows:
is yet to be established. Triangular [26], rectangular, and (l) (l)
singleton distributions are particular cases of trapezoidal
x̂t = [I −Wt Ht ]x̂t +Wt yt ,
distributions. For example, Serrurier and Prade [27] used > > −1
Wt = P̂t−1 Ht [Ht P̂t−1 Ht + Rt ] ,
the maximum possibilistic informational distance princi-
P̂t = [I −Wt Ht ]P̂t−1 , (13)
ple to develop a TPD from a set of data that follows
(l) (l)
unknown multimodal nonsymmetric probability distribu- x̂t+1 = Ft x̂t ,
tions and build a triangular possibility distribution from a
P̂t+1 = Ft P̂t Ft> + Qt ,
set of data that follows a skewed normal probability dis-
tribution. (l) (l)
where Ft = (∂ f /∂ x)|C(x̄t ), Ht = (∂ h/∂ x) | C(x̄t ), yt is
The state xt inherits the uncertainty of the noise, as de- the real measurement, and the updating gain matrix Wt is
fined in (8) and (9). Thus, the problem considered in this obtained by minimizing the uncertainty matrix P̂t .
study differs from the DSE problem based on the proba- To fuse fuzzy random variables, Lin et al. [21] proposed
bilistic framework. The predictive estimation set and pos- the FIF algorithm,which can be obtained by using a linear
teriori estimation set of xt by agent i are denoted as {x̄i,t , combination of TPDs and uncertainty matrix estimates for
P̄i,t } and {x̂i,t , P̂i,t }, respectively. any possible unknown correlation as follows:
Distributed Fuzzy Extended Kalman Filter for Multiagent Systems 1695
(l)
Lemma 1 (Fuzzy information fusion): Suppose {x1 , The local estimate of agent i is expressed in the infor-
(l) (l) (l)
P11 }, {x2 , P22 }, ..., {xn , Pnn } are n estimate sets of the mation form by using the fuzzy information vector ϕ̄i,t ,
same fuzzy variable x defined by E[x] ∼ Π(x(1) ; x(2) ; x(3) ; (l)
P̄i,t−1 x̄i,t and uncertainty information matrix Φ̄i,t , P̄i,t−1 .
x(4) ). Next, the general form of the FIF can be expressed Furthermore, the novel observation of agent i is defined
as follows: in the form of an information pair as follows:
n
(l) (l) (l) (l)
xc(l) = Pcc ∑ ωi Pii−1 xi , (14) ζi,t = Hi,t> R−1
i,t (yi,t − h(x̄i,t ) + Hi,t x̄i,t ), (21)
i=1
n Ξi,t = Hi,t> R−1
i,t Hi,t , (22)
Pcc−1 = ∑ ωi Pii−1 , (15)
(l)
i=1 where Hi,t = (∂ h/∂ x)|x̄i,t . Agent i ∈ V collects the local
where ωi ∈ [0, 1] and ∑ni=1 ωi = 1. Further, for all choices novel observation with the local information fuzzy infor-
(l) (l)
of ωi and unknown correlation mation vector ϕ̄i,t , P̄i,t−1 x̄i,t and uncertainty information
matrix Φ̄i,t , P̄i,t−1 to obtain the local posterior information
P̃i j = C[(x − C[xi ])(x − C[x j ])> ], (16) pair.
(l)
The estimate set {xc , Pcc } satisfies the following proper-
(l) (l) (l)
ties: ϕ̂i,t = P̄i,t−1 x̄i,t + ζi,t , (23)
1) xc is unbiased as long as each xi is unbiased, i.e., Φ̂i,t = P̄i,t−1 + Ξi,t . (24)
Table 1. DEFIF for agent i at time k. for an order of convergence greater than one under suitable
assumptions.
(l)
At time t, obtain prior information P̄i,t , x̄i,t , The relationship among the estimation errors ẽi,t = xt −
Perform the local linearization step: (l)
C(x̄i,t ) can be derived from Lemma 1 as follows:
Obtain the novel information pair:
(l) (l)
ζi,t = Hi,t> R−1 i,t (yi,t − h(x̄i,t ) + Hi,t ),
(l)
ẽi,t = xt − C(x̄i,t ) = C(xt − x̄i,t ) = C(ei,t ),
(l) (l)
(29)
> −1
Ξi,t = Hi,t Ri,t Hi,t ,
(l) (l)
Obtain the local posterior information pair: where ei,t = xt − x̄i,t . In this case, the boundedness of ẽi,t
(l) (l) (l)
ϕ̄i,t = P̄i,t−1 x̄i,t + ζi,t , (l)
can be converted into that of C(ei,t ). The boundedness of
Φ̄i,t = P̄i,t−1 + Ξi,t . (l)
C(ei,t ) can be proved as follows:
Perform the consensus iteration step By investigating the aforementioned Taylor expansions,
(l)
For L=1 to S C(ei,t ) can be expressed in terms of linearized and nonlin-
(l) (l)
ϕ̂i,t (L + 1) = ∑ j∈Ni ω i j ϕ̂i,t (L) ear terms as follows:
Φ̂i,t (L + 1) = ∑ j∈Ni ω i j Φ̂i,t (L) Proposition 1: The DEFIF algorithm can be initialized
End For at time t = 1 with the positive definite information matri-
Update the local estimation and uncertainty matrix ces Φ̄i,1 . For any i and any t, the matrices Φ̂i,t are invertible
(l) (l)
P̂i,t = [Φ̂i,t (S)]−1 , x̂i,t = [Φ̂i,t (S)]−1 ϕ̂i,t (S); (l) (l)
and C(ei,t ) = C(xt − x̄i,t ) obey the recursion
(1) (2) (3) (4)
E[x̂i,t ] ∼ Π(x̂i,t ; x̂i,t ; x̂i,t ; x̂i,t );
x̃i,t = C[x̂i,t ], U[x̂i,t ] = P̂i,t ; (l)
C(ei,t+1 )
(l) (l)
Prediction:
(l) (l)
= C( ∑ ωLi j Fi,t (Φ̂i,t )−1 Φ̄ j,t e j,t + ϕ(xt , x̂i,t )
x̄i,t+1 = f (x̂i,t ), Fi,t = (∂ f /∂ x)|C(x̂i,t ) j∈Ni
(l)
Φ̂i,t+1 = Qt − Qt Fi,t (Φ̄i,t + (Fi,t )> Qt Fi,t )−1 (Fi,t )> Qt (l)
+ ∑ ωLi j Fi,t (Φ̂i,t )−1 (H j,t )> R−1 j
j,t χ (xt , x̄ j,t )). (30)
j∈Ni
where ε j , tr(P̄j,t ).
Proof: The predicted and estimated uncertainty metrics
Ω denotes the consensus matrix, whose elements are
can be expressed as follows:
the consensus weights ω i j for any i, j ∈ N . Furthermore,
ωLi j represents the (i, j)-th element of ΩL , that is, the Lth Φ̄i,t+1 = (Fi,t (Φ̂i,t )−1 (Fi,t )> +W −1 )−1 , (31)
power of the consensus matrix Ω. The consensus matrix Ω
is row stochastic and primitive provided that the network and
is strongly connected.
The recursive form of the DEFIF algorithm has been Φ̂i,t = ∑ ωLi j Φ̄ j,t + ∑ ωLi j (H j,t )> R−1
j,t H j,t . (32)
j∈Ni j∈Ni
summarized in Table 1. Notably, the DEFIF algorithm is
fully distributed and can overcome the nonlinear problem In this case, Φ̄i,1 > 0, i ∈ N implies that Φ̂i,t > 0, i ∈ N
of systems with fuzzy noises. for any finite t. Moreover,
Remark 4: In this study, the consistency of DEFIF al-
(l) (l)
gorithms is not tenable under nonlinear conditions, as de- ei,t+1 = f (xt ) − f (x̂i,t ) + wt (33)
tailed from the proof of Theorem 2 in [21]. However, Sec- (l) (l)
= Fi,t (xt − x̂i,t ) + ϕ(xt , x̂i,t ) + wt . (34)
tion 4 reveals that the DEFIF algorithms enjoy local stabil-
ity provided that, similarly to the linear case, suitable con- (l)
Furthermore, the estimate x̂i,t can be expressed as follows:
nectivity and collective observability assumptions hold.
h
(l) (l)
x̂i,t = (Φ̂i,t )−1 ∑ ωLi j Φ̄ j,t x̄ j,t
4. STABILITY ANALYSIS j∈Ni
i
(l)
The stability characteristics of the DEFIF are described + ∑ ωLi j (H j,t )> R−1
j,t ȳ( j,t) , (35)
in this section. Because the DEFIF is based on the lin- j∈Ni
Because S1 S2 S3
H j,t xt − ytj = H j,t (xt − x̄ j,t ) − ytj + h j (x̄ j,t )
= H j,t (xt − x̄ j,t ) + h j (x̄ j,t ) − h j (xt ) − v j,t
= χ j (xt , x̄ j,t ) − v j,t , (37)
(l) (l)
xt − x̂i,t = (Φ̂i,t )−1 ( ∑ ωLi j Φ̄ j,t (xt − x̄i,t )
j∈Ni Fig. 2. Communication topology graph among sensors.
(l)
+ ∑ ωLi j (H j,t )> R−1 j
j,t (χ (xt , x̄i,t ) − v j,t )).
j∈Ni
follows:
(38)
sin(Ωt T ) tT)
1 Ωt 0 − 1−cos(Ω
Ωt 0
Subsequently, the following formula can be derived 0 − sin(Ωt T )
0 cos(Ωt T ) 0
Fct = 0 1−cos(Ω tT) sin(Ωt T )
, (42)
(l) 1 0
C(ei,t+1 ) Ωt
0 sin(Ωt T )
Ωt
0 cos(Ωt T ) 0
(l) (l)
= C( ∑ ωLi j Fi,t (Φ̂i,t )−1 Φ̄ j,t e j,t + ϕ(xt , x̂i,t ) 0 0 0 0 1
j∈Ni T2
(l) 2 0 0
+ ∑ ωLi j Fi,t (Φ̂i,t )−1 (H j,t )> R−1 j
j,t χ (xt , x̄ j,t ) T 0 0
j∈Ni
T2
Gct = 0 2 0 , (43)
∑ ωLi, j Fi,t (Φ̂i,t )−1 (H j,t )> R−1
+ wt − j,t v j,t ). (39) 0 T 0
j∈Ni
0 0 1
According to (8) and (9), the last two terms of (39) are as reported in [29]. wt = [wt,x , wt,y , wt,z ]T , wt,x , wt,y , wt,z
zero. Subsequently, (34) can be obtained. denote the noise terms for the acceleration in the x and y
(l)
The boundedness of C(ei,t ) can be established using directions and for the turn rate.
Proposition 1. N (N = 6, in this study) nonlinear sensors were used
Theorem 1: Assumptions 1-3 are considered to be sat- to measure the angle and distance of the noncooperative
isfied in this study. Assume that nonnegative reals ā and target. The measurement by sensor i (i = 1, 2, ..., N) can
c̄ and positive reals ω and ω̄ exist such that kFi,t k ≤ ā, be defined as follows:
(l)
kHi,t k ≤ c̄, 0 < ωI ≤ Φ̂i,t ≤ ω̄I. In this case, C(ei,t ) is h ϑ −ϑi
i,t
bounded; that is yi,t = arctan ,
i,t − i
q i
(l)
lim sup C(ei,t ) ≤ ∞. (40) (i,t − i )2 + (ϑi,t − ϑ i )2 + vi,t , (44)
t→∞
1 1 1 1
0 0 0 0
-10 0 10 -2 0 2 -0.02 0 0.02 -0.02 0 0.02
(a) TPD for ∆i,0 . (b) TPD for ∆˙i,0 . (a) TPD for wtx . (b) TPD for wty .
1 1 1
1
0.5 0.5 0.5
0.5
0 0 0
-10 0 10 0
-2 0 2
-2 0 2 -0.1 0 0.1
10-3
(c) TPD for ∆ϑi,0 . (d) TPD for ∆ϑ̇i,0 . (c) TPD for wtz . (d) TPD for vi,tx .
1 1
0.5 0.5
0 0
-0.1 0 0.1 -5 0 5
10-3
(e) TPD for ∆Ωi,0 . (e) TPD for vi,ty .
Fig. 3. TPDs for the initial state errors ∆xi,0 . Fig. 4. TPDs for the nosies wt and vi,t .
60
ces Qw and Rv , respectively. In this study, the noise terms
wt and vi,t are modeled using the possibility distribution 40
displayed in Fig. 4. 20
By using these simulation settings, a random trajectory
0
of the target was created. Subsequently, the trajectories es-
timated by each sensor are generated using the DEFIF al- -20
-160 -140 -120 -100 -80 -60
gorithms, as displayed in Fig. 5. x (m)
Figs. 6-9 illustrate the estimates of t , ˙t , ϑt , and ϑ̇t of
the target state quantity by sensor 1. The solid lines in- Fig. 5. The true trajectory and estimated trajectories by
dicate the TPD, and the dashed line indicates the center each sensor.
of gravity (i.e., the final estimate of the target state). The
estimate always is within the possible region. Moreover,
as time elapses, the coverage area of the TPD rapidly de- of the possible regions is nearly invariant, which satisfies
creases. The average width of the TPD is smaller than that the requirement of the coherence of the size of regions
of the distribution defined for the noise models and initial [15].
error. This finding revealed that the fuzziness of the es- Overall, 200 independent Monte Carlo trials were per-
timation decreased with the increase in the algorithm re- formed, and the mean square estimation errors of posi-
cursions. Furthermore, the evolutionary trajectories of the tion (MSEP) and mean square estimation errors of veloc-
four vertices of the TPD tend to be parallel. Thus, the size ity (MSEV) were computed as the performance indices.
Distributed Fuzzy Extended Kalman Filter for Multiagent Systems 1699
-50 150
100
-100 50
-130
-140
0
-150
0 0.2 0.4
-150 -50
0 1 2 3 4 5 0 1 2 3 4 5
Time (s) Time (s)
Fig. 6. The estimates of the target state quantity by sensor Fig. 8. The estimates of the target state quantity by sensor
1, possibility distribution evolution of t . 1, possibility distribution evolution of ϑt .
60 250
40 200
20 150
0 100
-20 50
-40 0
0 1 2 3 4 5 0 1 2 3 4 5
Time (s) Time (s)
Fig. 7. The estimates of the target state quantity by sensor Fig. 9. The estimates of the target state quantity by sensor
1, possibility distribution evolution of ˙t . 1, possibility distribution evolution of ϑ̇t .
The MSEP and MSEV of agent i can be defined as fol- MSEV between DEFIF, FEKF, and CFEKF. Because of
lows: inadequate research on the CFIF, this study caused the
Nment DFIF degenerate to the CFIF by assuming that graph G
1
MSEPti =
Nment ∑ ( p̂i,tj − pt )T ( p̂i,tj − pt ), (47) is strongly connected. Notably, the weight ωi j,k in CFEKF
j=1 is set as ωi j,k = N1 instead of calculating the suboptimal
Nment solution according to (26). The results revealed that the
1
MSEVti =
Nment ∑ (v̂i,tj − vt )T (v̂i,tj − vt ), (48) estimation error of DEFIF is close to that of the CFEKF,
j=1
although the DEFIF is inferior to the centralized version
where pt , [px,t , py,t ]T = [xt (1), xt (3)]T represents the in the average sense. The results also revealed that the es-
true position vector, and vt , [vx,t , vy,t ]T = [xt (2), xt (4)]T timation error of the DEFIF is superior to the FEKF in
denotes the true velocity vector. Furthermore, p̂i,t , [i,t , [19].
ϑi,t ]T = [x̂i,t (1), x̂i,t (3)]T and v̂i,t , [˙i,t , ϑ̇i,t ]T = [x̂i,t (2), Remark 5: The results revealed that combining the FIF
x̂i,t (4)]T are the corresponding estimated position vector algorithm and local linearization in distributed sensor net-
and velocity vector of agent i, respectively. works is effective, and the local estimation error is uni-
The performances of the proposed DEFIF are compared formly bounded. Furthermore, the agents pass fuzzy in-
with those of the fuzzy extended Kalman filter (FEKF) formation to other agents in the DEFIF algorithm; conse-
in [19] and the centralized fuzzy extended Kalman filter quently, the estimation accuracy of the entire multiagent
(CFEKF). Fig. 10 present a comparison of the MSEP and system can be increased.
1700 Xiaobo Zhang, Haoshen Lin, Gang Liu, and Bing He
10
8
10
6
5
5
4
0
4.4 4.5 4.6 4.7 4.8 4.9
0
0 1 2 3 4 5 3 3.5 4 4.5
Time (s) Time (s)
MSE of velocity MSE of velocity
200 35
Sensor1 of DEFIF
Sensor2 of DEFIF L=1
Sensor3 of DEFIF
Sensor4 of DEFIF
Sensor5 of DEFIF 30 L=5
Sensor6 of DEFIF
MSE of velocity (m/s)
20
100 15
10
15
5
50 10
0
4.4 4.5 4.6 4.7 4.8 4.9
0
0 1 2 3 4 5 3.4 3.6 3.8 4 4.2 4.4 4.6
Time (s) Time (s)
Fig. 10. A comparison of the MSEP and MSEV be- Fig. 11. A comparison of the MSEP and MSEV of the DE-
tween the proposed DEFIF algorithm, FEKF, and FIF algorithm with diffident consensus iterations.
CFEKF.
estimation problem using other linearization paradigms, where the final inequality is based on the fact that Φ̂i,t ≥
ij
such as volume Kalman filtering. ∑ j∈Ni ωL Φ̄ j,t and Lemma 2 of [30]. Consequently,
[7] S. Chen and C. Chen, “Probabilistic fuzzy system for un- [21] H. Lin, C. Hu, Z. Deng, and G. Liu, “Distributed kalman
certain localization and map building of mobile robots,” filter with fuzzy noises over multi-agent systems,” IEEE
IEEE Transacrtions on Instrumentation and Measurement, Transacrtions on Fuzzy Systems, vol. 30, no. 7, pp. 2550-
vol. 61, no. 6, pp. 1546-1560, June 2012. 2562 2022.
[8] L. A. Zadeh, “Fuzzy sets as a basis for a theory of possi- [22] X. Zhang, H. Lin, G. Liu, and B. Liu, “Distributed Kalman
bility,” Fuzzy sets and Systems, vol. 1, no. 1, pp. 3-28, June filter with fuzzy noises over multiagent systems under
1978. switching topology,” Digital Signal Processing, vol. 122,
103326, April 2022.
[9] S. Aja-Fernandez, C. Alberola-Lopez, and J. Ruiz-Alzola.
“A fuzzy-controlled Kalman filter applied to stereo-visual [23] H. Kwakernaak, “Fuzzy random variables–II. Algorithms
tracking schemes,” Signal Processing, vol. 83, no. 1, pp. and examples for the discrete case,” Information Sciences,
101-120, January 2003. vol. 17, no. 3, pp. 253-278, April 1979.
[10] Z. Hu, P. Shi, L. Wu, and C. K. Ahn, “Event-based dis- [24] T. Hies, S. Parasuraman, Y. Wang, R. Duester, H. Eikaas,
tributed filtering approach to nonlinear stochastic systems and K. Tan, “Enhanced water-level detection by image pro-
over sensor networks,” International Journal of Control, cessing,” Proc. of 10th International Conference on Hy-
Automation, and Systems, vol. 17 no. 4, pp. 896-906, droinformatics, Hamburg, 2012.
March 2019. [25] S. Kumar and B. Mitra, “Design and fabrication of a smart
[11] D. Longo, G. Muscato, and V. Sacco, “Localization us- temperature transmitter for thermocouple,” i-Manager’s
ing fuzzy and kalman filtering data fusion,” Proceedings of Journal on Instrumentation Control Engineering, vol. 7,
the 5th International Conference on Climbing and Walking no. 3, pp. 21-32, May 2019.
Robots, pp. 25-27, Catania, 2002. [26] F. Tatari, M.-R. Akbarzadeh-T, M. Mazouchi, and G. Javid,
[12] S. Shi, Z. Fei, P. Shi, and C. K. Ahn, “Asynchronous fil- “Agent-based centralized fuzzy kalman filtering for un-
tering for discrete-time switched T-S fuzzy systems,” IEEE certain stochastic estimation,” Proc. of Fifth International
Transacrtions on Fuzzy Systems, vol. 28, no. 8, pp. 1531- Conference on Soft Computing, Computing with Words
1541, August 2020. and Perceptions in System Analysis, Decision and Control,
Famagusta, 2009.
[13] M.-J. Yu, “INS/GPS integration system using adaptive
filter for estimating measurement noise variance,” IEEE [27] M. Serrurier and H. Prade, “An informational distance
Transacrtions on Aerospace and Electronic Systems, vol. for estimating the faithfulness of a possibility distribution,
48, no. 2, pp. 1786-1792, April 2012. viewed as a family of probability distributions, with respect
to data,” International Journal of Approximate Reasoning,
[14] G. Agamennoni, J. I. Nieto, and E. M. Nebot, “Approx- vol. 54, no. 7, pp. 919-933, September 2013.
imate inference in state space models with heavy-tailed
noise,” IEEE Transacrtions on Signal Processing, vol. 60, [28] D. W. Casbeer and R. Beard, “Distributed information fil-
no. 10, pp. 5024-5037, August 2012. tering using consensus filters,” American Control Confer-
ence, St. Louis, 2009.
[15] F. Mata, A. Jimenez, B. M. Al-Hadithi, D. Rodrguez-
Losada, and R. Galan, “The fuzzy kalman filter: State esti- [29] W. Niehsen, “Information fusion based on fast covariance
mation using possibilistic techniques,” Fuzzy Sets and Sys- intersection filtering,” Proc. of International Conference on
tems, vol,157, no. 16, pp. 2145-2170, August 2006. Information Fusion, Annapolis, 2002.
[16] D. Ye, H. Lin, X. Yang, B. He, and D. Pan, “Spatial tar- [30] G. Battistelli and L. Chisci, “Kullback-leibler average, con-
get localization using fuzzy square-root cubature kalman sensus on probability densities, and distributed state esti-
filter,” Proc. of 31st Youth Academic Annual Conference mation with guaranteed stability,” Automatica, vol. 50, no.
of Chinese Association of Automation, pp. 73-80, China, 3, pp. 707-718, March 2014.
2016.
[17] X. Yang, G. Liu, J. Guo, H. Wang, and B. He, “The robust
passive location algorithm for maneuvering target track-
ing,”Mathematical Problems in Engineering, vol. 2015, pp.
1-8, June 2015.
Xiaobo Zhang received his B.S. and M.S.
[18] Z. Zhou, C. Hu, H. Fan, and J. Li, “Fault prediction of
degrees from the Space Engineering Uni-
the nonlinear systems with uncertainty,” Simulation Mod-
versity, China, in 2007 and 2011, respec-
elling Practice and Theory, vol. 16, no. 6. pp. 690-703,
tively. He is currently pursuing a Ph.D.
June 2008.
degree in control science and engineering
[19] F. Mata, V. Jimenez, B. P. Alvarado, and R. Haber, “The with the Rocket Force University of Engi-
fuzzy Kalman filter: Improving its implementation by re- neering (RFUE), China. His research inter-
formulating uncertainty representation,” Fuzzy Sets and ests include information processing, con-
Systems, vol. 402, pp. 78-104, January 2021. trol theory, and distributed estimation.
[20] V. Jimenez, B. P. Alvarado, and F. Mata, “A set of practical
experiments to validate the fuzzy kalman filter,” Fuzzy Sets
and Systems, vol. 417, pp. 152-170, August 2021.
Distributed Fuzzy Extended Kalman Filter for Multiagent Systems 1703