Proceedings of the 19th World Congress

The International Federation of Automatic Control

Cape Town, South Africa. August 24-29, 2014

Event-Triggered Least Squares Fault

Estimation with Stochastic Nonlinearities ⋆
LIU Yang ∗,∗∗∗ WANG Zidong ∗,∗∗ HE Xiao ∗,∗∗∗
ZHOU Dong-Hua ∗,∗∗∗
∗
Department of Automation, Tsinghua University, Beijing, 100084,
P. R. China.
∗∗
Department of Information Systems and Computing, Brunel
University, Uxbridge, Middlesex, UB8 3PH, United Kingdom. (e-mail:
Zidong.Wang@brunel.ac.uk)
∗∗∗
National Laboratory for Information Science and Technology,
Tsinghua University, Beijing, 100084, P. R. China. (e-mail:
zdh@mail.tsinghua.edu.cn)

Abstract: In this paper, the event-triggered least squares state and fault estimation problem
is investigated for a class of systems with stochastic nonlinearities. An event-triggered scheme
is properly proposed whose main idea is to transmit the measurement output to a remote
estimator only when a specified event condition is violated and an event is triggered. A filter is
designed so as to minimize an upper bound of the filtering error covariance with event-triggered
measurement transmissions and additive stochastic nonlinearities. By solving two sets of discrete
matrix equations, the desired filter parameters are calculated recursively and thus the method
is applicable for online computation. Both the state and fault estimation problems are handled
within the same framework using the least squares method. A numerical simulation is exploited
to illustrate the effectiveness of the proposed algorithm.

1. INTRODUCTION posteriori probability density function of states and obtain

the exact minimum mean-squared estimation error. How-
For decades, the state/fault estimation problem has been ever, when the system model is relatively complicated, the
a research focus as one of the fundamental problems in conditional probability density functions may be unavail-
the control and signal processing areas. The traditional able or overcomplex, which would limit the applicabilities
Kalman filter, which serves as an optimal filter for linear of the proposed algorithms. Therefore, there is a practical
systems, minimizes the filtering error covariance at each need to address the event-triggered least squares filtering
time step. To deal with more comprehensive systems problem with less prior knowledge than exact probability
(e.g. systems with nonlinearities or uncertainties), a great density functions (e.g. only the means and covariances of
number of results have been reported to improve the the disturbances).
performance of Kalman filter, such as extended Kalman The analysis problem for nonlinear systems has been ex-
filter (Reif et al. [1999]), mixed H2 /H∞ filter (Xie et al. tensively studied for decades, since nonlinearities are in-
[2004]), etc. evitably encountered in various industrial systems, which
Recently, event-triggered estimation and control problems may lead to undesirable dynamic behaviors. In many
have gained an increasing research interest and much effort real-world systems, nonlinear disturbances may occur in
has been made to address the problems (Tabuada [2007], a random way, because of sudden environment changes,
Wang et al. [2011], Dimarogonas et al. [2012], Hu et al. random failure and repairs of components, intermittent
[2012]). In an event-triggered measurement scheme, the transmission congestion, etc (Shen et al. [2011]). Such
measurement output is transmitted to a remote estima- a kind of nonlinearities can be referred to as stochastic
tor only when a specified event condition is violated in nonlinearities or randomly occurring nonlinearities. There
an event generator. An estimator that could cope with have been some results on filtering problem with stochastic
event-triggered measurement transmission is much needed nonlinearities (Dong et al. [2011], Shen et al. [2011], Ma
to save data transfer and processing power, especially in et al. [2011]), most of which have focused on conservative
wireless networked systems with limited energy resources. robust H∞ filtering schemes without considering the es-
The event-triggered optimal filtering problem in the least timation performances such as the error covariances. On
squares sense has stirred some initial research interests the other hand, the least squares filtering problem with
(Sijs et al. [2009], Wu et al. [2013]). These results have stochastic nonlinearities has not been sufficiently inves-
utilized the probability density functions of states and tigated yet (Hu et al. [2012, 2013]), not to mention the
innovations conditional on measurements to calculate the event-triggered case. This constitutes the main motivation
⋆ This work was supported by National Basic Research Program of our present work.
of China (973 Program) (2010CB731800), National Natural Science
Foundation of China (61210012, 61290324, 61273156).

978-3-902823-62-5/2014 © IFAC 1855

19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

In this article, the joint least squares state and fault Moreover, (2) and (3) can be easily rewritten as follows,
estimation problem is addressed for a class of systems respectively:
{[ ] }
with event-triggered measurement transmissions as well as h̄(x̄k , αk ) 
E x̄ = 0, (5)
additive stochastic nonlinearities. In the proposed event- ḡ(x̄k , βk )  k
triggered scheme, the measurement output is transmitted
and
to a remote estimator only when a specified event con- {[ ][ ]T  }
dition is violated in the event generator. A filter is de- h̄(x̄k , αk ) h̄(x̄j , αj ) 
signed recursively that guarantees the minimization of an E  x̄
ḡ(x̄k , βk ) ḡ(x̄j , βj )  k
upper bound of the filtering error covariance at each time 
step. The main contributions of the paper are outlined as [
T
0, ] if k ̸= j,
follows: 1) a comprehensive system model is put forward = Π̄ x̄ Ω̄
1 k 1 k x̄ 0 (6)
 , if k = j,
that covers event-triggered measurement transmission and 0 Π2 x̄Tk Ω̄2 x̄k
additive stochastic nonlinearities; 2) additive faults and where
system states are simultaneously estimated to facilitate [ ] [ ] [ ]
Π1 0 Ω1 0 Ω2 0
both the fault isolation and the state estimation prob- Π̄1 = , Ω̄1 = , Ω̄2 = .
0 0 0 0 0 0
lems; and 3) the proposed algorithm can be carried out
recursively and thus applicable for online computation. In this paper, the following transmission architecture is
2. PROBLEM FORMULATION considered: the measurement output of (4) is sampled at
each time step by the sampler, and then transmitted to
Consider the following discrete-time nonlinear stochastic the event generator. The current measurement is released
system: by the generator when the current measurement yk+j and
the previously transmitted measurement yk satisfy the
{ following inequality:
xk+1 = Ak xk + Bk uk + h(xk , αk ) + Dk wk + Fk fk , (yk+j − yk )T (yk+j − yk ) > σ, (7)
yk = Ck xk + g(xk , βk ) + Ek vk , x0 = x0 ,
where σ is a predefined positive scalar. If (7) is satisfied,
(1)
the current measurement is forwarded to a Zero-Order
where xk ∈ Rn is the system state; yk ∈ Rm is the Hold (ZOH). Considering the characteristic of ZOH, the
measurement output; uk ∈ Rl is the control input; fk ∈ Rs real estimator input ỹk can be written as
is the additive fault; wk ∈ Rp and vk ∈ Rq are the process ỹk = yki , k ∈ {ki , ki + 1, · · · , ki+1 − 1}, (8)
noise and the measurement noise, respectively. The noise where k0 , k1 , · · · are assumed to be the release times under
sequences are independent zero-mean Gaussian sequences, the strategy (7).
with E{wk wkT } = Wk , E{vk vkT } = Vk . Ak , Bk , Ck , Dk , Ek
and Fk are known matrices with appropriate dimensions. For system (4), an estimator of the following structure is
proposed:
The functions h(xk , αk ) and g(xk , βk ) represent the
x̂k+1|k = Āk x̂k|k , (9)
stochastic nonlinearities. αk ∈ R and βk ∈ R are inde- ( )
pendent zero-mean Gaussian noise sequences. The nonlin- x̂k+1|k+1 = x̂k+1|k + Ξk+1 ỹk+1 − C̄k x̂k+1|k , (10)
earities have the following first moment: where x̂k+1|k ∈ Rn and x̂k|k ∈ Rn are the one-step
prediction and the estimate of x̄k at time step k with
{[ ] } [ ]T
h(xk , αk )  x̂0|0 = xT0 , 0T , respectively, and Ξk+1 is the parameter
E x = 0, (2)
g(xk , βk )  k to be designed.
and the covariance given by The model (4) is put forward in the paper so as to cater for
{[ ][ ]T  } the nonlinearities that occur in a random way. In fact, such
h(xk , αk ) h(xj , αj ) 
E
g(xj , βj )  k
x stochastic nonlinearities could encompass large quantities
g(xk , βk )
of nonlinearities including the state multiplicative noises.

] if k ̸= j,
[ 0, The proposed transmission condition (7), which means
= Π1 xTk Ω1 xk 0 (3) that the current measurement is transmitted only when it
 , if k = j, changes significantly, can lead to lower traffic requirement
0 Π2 xTk Ω2 xk
and more efficient resource utilization. Meanwhile, the
where Ωi and Πi (i ∈ {1, 2}) are known matrices with estimation results would have some robustness to small
appropriate dimensions. variations in the system under such a strategy. In an
Implementing a given state feedback controller uk = Kk xk extreme situation, if σ=0, then all the measurements
[ ]T would be transmitted, and it reduces to traditional time-
to system (1) and defining x̄k = xTk , fkT , we consider the
triggered transmission.
following stochastic nonlinear closed-loop system:
{
x̄k+1 = Āk x̄k + h̄(x̄k , αk ) + D̄k wk , Denote the prediction error and filtering error by ek+1|k =
(4) x̄k+1 − x̂k+1|k , and ek+1|k+1 = x̄k+1 − x̂k+1|k+1 , respec-
yk = C̄k x̄k + ḡ(x̄k , βk ) + Ek vk , { }
where [ ] [ ] tively, and their covariances by Pk+1|k = E ek+1|k eTk+1|k
Ak + Bk Kk Fk Dk { }
Āk =
0 I
, D̄k =
0
, C̄k = [ Ck 0 ] , and Pk+1|k+1 = E ek+1|k+1 eTk+1|k+1 . The goal of the
[ ]T paper is to design a recursive filter in the form of (9)
h̄(x̄k , αk ) = hT (xk , αk ), 0 , ḡ(x̄k , βk ) = g(xk , βk ). and (10) for system (4) such that an upper bound of the

1856
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

filtering error covariance Pk+1|k+1 can be provided and the posteriori probability density function of the states
minimized. based on the probability density functions of states and
innovations conditional on measurements. When the sys-
3. MAIN RESULTS tem dynamics is relatively complicated, for example, the
system contains some stochastic nonlinearities as (4), the
Firstly, by the definitions of Pk+1|k and Pk+1|k+1 , the two conditional probability density functions might be difficult
covariances would be obtained in the sequel, respectively. to calculate, and the exact error covariances might be over-
complicated or even unavailable. A seemingly nature way
Theorem 1. Pk+1|k obeys the following recursion relation: is to find an upper bound of the filtering error covariance,
{ { } }
Pk+1|k =Āk Pk|k ĀTk + Π̄1 tr E x̄k x̄Tk Ω̄1 + D̄k Wk D̄kT . and then minimize the bound by appropriately designing
(11) the filter gain at each time step. This way, the conditional
probability density functions will be no longer needed.
Proof. Based on (4) and (9), we have Before proceeding further, the following lemma is to be
ek+1|k =Āk ek|k + h̄(x̄k , αk ) + D̄k wk . (12) introduced.
Noticing that the stochastic nonlinearity h̄(x̄k , αk ) and the Lemma 3. (Hu et al. [2012]) For any two vectors x, y ∈ Rn ,
additive noise wk are zero-mean, (11) follows directly from the following inequality holds
(6) and (12). xy T + yxT ≤ εxxT + ε−1 yy T , (17)
Theorem 2. Pk+1|k+1 satisfies the following equation: where ε > 0 is a scalar.
Pk+1|k+1 =(I − Ξk+1 C̄k+1 )Pk+1|k (I − Ξk+1 C̄k+1 )T
{ } With Lemma 3, an approach is proposed in the following
+ Ξk+1 E (ỹk+1 − yk+1 )(ỹk+1 − yk+1 )T theorem to determine the filter gain such that an upper
{ { } }
× ΞTk+1 + Ξk+1 Π2 tr E x̄k+1 x̄Tk+1 Ω̄2 ΞTk+1 bound of the filtering error covariance is minimized.
T
+ Ξk+1 Ek+1 Vk+1 Ek+1 ΞTk+1 − Ξk+1 Theorem 4. Let ε be a positive scalar. If the next two
{ } equations:
× E (ỹk+1 − yk+1 )eTk+1|k (I − Ξk+1 C̄k+1 )T { }
{ } Qk+1|k =Āk Qk|k ĀTk + Π1 tr Xk Ω̄1 + D̄k Wk D̄kT , (18)
− (I − Ξk+1 C̄k+1 )E ek+1|k (ỹk+1 − yk+1 )T Qk+1|k+1 =(1 + ε)(I − Ξk+1 C̄k+1 )Qk+1|k (I − Ξk+1
× ΞTk+1 . (13) × C̄k+1 )T + (1 + ε−1 )σΞk+1 ΞTk+1
{ }
Proof. From (4) and (10), it follows that + Ξk+1 Π2 tr Xk+1 Ω̄2 ΞTk+1
( ) T
ek+1|k+1 =ek+1|k − Ξk+1 ỹk+1 − C̄k x̂k+1|k . (14) + Ξk+1 Ek+1 Vk+1 Ek+1 ΞTk+1 , (19)
Adding a zero term Ξk+1 yk+1 −Ξk+1 yk+1 to the right-hand where
side of (14), we have Xk+1 =Āk Xk ĀTk + Π1 tr{Xk Ω̄1 } + D̄k Wk D̄kT , (20)
ek+1|k+1 =(I − Ξk+1 C̄k+1 )ek+1|k − Ξk+1 ḡ(x̄k+1 , βk+1 )
with initial conditions Q0|0 ≥ 0 and X0 = x̄0 x̄T0 ,
have
− Ξk+1 Ek+1 vk+1 − Ξk+1 (ỹk+1 − yk+1 ). (15) positive definite solutions. Then, Qk|k is an upper bound
With zero-mean stochastic nonlinearity ḡ(x̄k+1 , βk+1 ) and of Pk|k . Meanwhile, if the the filter gain is chosen as
additive noise vk+1 , Pk+1|k+1 can be written as: T −1
Ξk+1 = Zk+1 Yk+1 , (21)
Pk+1|k+1 =(I − Ξk+1 C̄k+1 )Pk+1|k (I − Ξk+1 C̄k+1 )T
{ } where
+ Ξk+1 E (ỹk+1 − yk+1 )(ỹk+1 − yk+1 )T
{ } T T
Yk+1 =(1 + ε)C̄k+1 Qk+1|k C̄k+1 + (1 + ε−1 )σI
× ΞTk+1 + Ξk+1 Ek+1 E vk+1 vk+1 T
Ek+1 ΞTk+1
{ } + Π2 tr{Xk+1 Ω̄2 } + Ek+1 Vk+1 Ek+1
T
, (22)
+ Ξk+1 E ḡ(x̄k+1 , βk+1 )ḡ T (x̄k+1 , βk+1 ) ΞTk+1
{ } Zk+1 =(1 + ε)C̄k+1 Qk+1|k , (23)
− (I − Ξk+1 C̄k+1 )E ek+1|k (ỹk+1 − yk+1 )T
{ } then Qk+1|k+1 is minimized.
× ΞTk+1 − Ξk+1 E (ỹk+1 − yk+1 )eTk+1|k
Proof. The conclusions can be obtained by induction. It
× (I − Ξk+1 C̄k+1 )T . (16) is already known that Q0|0 > P0|0 = 0. Then, assuming
Then with (6) and E{vk+1 vk+1 T
} = Vk+1 , (13) can be that for i = 1, 2, . . . , k, Pi|i ≤ Qi|i , it remains to show that
obtained directly. The proof is complete. Pk+1|k+1 ≤ Qk+1|k+1 .
With Lemma 3, we have the following inequality:
In Theorem 1 and Theorem 2, the exact recursion relations { }
of Pk+1|k and Pk+1|k+1 have been obtained. However, the − (I − Ξk+1 C̄k+1 )E ek+1|k (ỹk+1 − yk+1 )T ΞTk+1
{ }
terms which are related to (ỹk+1 − yk+1 ), are very difficult − Ξk+1 E (ỹk+1 − yk+1 )eTk+1|k (I − Ξk+1 C̄k+1 )T
to calculate. This results from the factor that the outputs { }
are restricted not only by the measurement formulation ≤ε(I − Ξk+1 C̄k+1 )E ek+1|k eTk+1|k (I − Ξk+1 C̄k+1 )T
(4) – but also by the transmission strategy (7). In other { }
words, the exact covariances of prediction error and filter- + ε−1 Ξk+1 E (ỹk+1 − yk+1 )(ỹk+1 − yk+1 )T ΞTk+1 .
ing error are dependent on whether the current measure- (24)
ment is transmitted or not. To obtain exact covariances of
prediction error and filtering error, we need to calculate With the definition of Pk+1|k , (24) can be written as

1857
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

{ }
− (I − Ξk+1 C̄k+1 )E ek+1|k (ỹk+1 − yk+1 )T ΞTk+1 The least squares fault and state estimation problem is
{ } solved by Theorem 4 for a class of systems subject to
− Ξk+1 E (ỹk+1 − yk+1 )eTk+1|k (I − Ξk+1 C̄k+1 )T event-triggered measurement transmissions and additive
stochastic nonlinearities. To deal with the event-triggered
≤ε(I − Ξk+1 C̄k+1 )Pk+1|k (I − Ξk+1 C̄k+1 )T
{ } measurement transmissions, special effort has been made
+ ε−1 Ξk+1 E (ỹk+1 − yk+1 )(ỹk+1 − yk+1 )T ΞTk+1 . to calculate an upper bound of the filtering error co-
(25) variance, which is dependent on the covariances of the
stochastic nonlinearities and the threshold in the event
Substituting (25) into (13), we have generator. By doing so, the traditionally required prob-
Pk+1|k+1 ≤(1 + ε)(I − Ξk+1 C̄k+1 )Pk+1|k (I − Ξk+1 ability density functions of states and innovations con-
ditional on measurements, which may be complicated or
× C̄k+1 )T + (1 + ε−1 )Ξk+1 E{(ỹk+1 − yk+1 )
even unavailable for relatively complicated systems, are no
× (ỹk+1 − yk+1 )T }ΞTk+1 + Ξk+1 Ek+1 Vk+1 longer needed. In other words, we do not need to calculate
{ { } }
× Ek+1
T T
Ξk+1 + Ξk+1 Π2 tr E x̄k+1 x̄Tk+1 Ω̄2 the posteriori probability density function of states and
the exact least squares filtering error covariance, thereby
× ΞTk+1 . (26) improving the feasibility and robustness of the algorithms
Based on (7), for any k ∈ N, we have at the cost of sacrificing certain accuracy at an acceptable
level. It is worth mentioning that, when the measurement
(ỹk − yk )(ỹk − yk )T ≤σI. (27) is transmitted at each time step (i.e., σ = 0) and there
is no stochastic nonlinearity, the proposed filter can be
Considering (26) and (27), we have specialized to the classical Kalman filter. Moreover, the
Pk+1|k+1 ≤(1 + ε)(I − Ξk+1 C̄k+1 )Pk+1|k (I − Ξk+1 proposed algorithm is suitable for on-line applications,
× C̄k+1 )T + (1 + ε−1 )σΞk+1 ΞTk+1 since it is carried out by solving discrete matrix equations.
{ { } }
+ Ξk+1 Π2 tr E x̄k+1 x̄Tk+1 Ω̄2 ΞTk+1
T 4. SIMULATION
+ Ξk+1 Ek+1 Vk+1 Ek+1 ΞTk+1 . (28)
{ } Consider system (1) with parameters given as follows:
To proceed further, denote Xk = E x̄k x̄Tk . From (4), it [ ] [ ]
follows directly that 0.1 0 1
Ak = , Fk = , Bk = Dk = I,
Xk+1 =Āk Xk ĀTk + Π1 tr{Xk Ω̄1 } + D̄k Wk D̄kT , −0.5 0.2 0
which is (20). Ck = [ 1 0 ] , Ek = 1, Kk = −0.21I, x0 = [0, 0]T .
Based on our assumption that Pk|k ≤ Qk|k , it can be easily [ ]T
(1) (2)
verified that, Denoting xk = xk , xk , the stochastic nonlinearities
Qk+1|k − Pk+1|k = Āk (Qk|k − Pk|k )ĀTk ≥ 0. (29) are formulated as:
h(xk , αk )
With (28), (29) and the definition of Xk , we have [ ][ ]
Pk+1|k+1 ≤ (1 + ε)(I − Ξk+1 C̄k+1 )Qk+1|k (I − Ξk+1 0.2 (1) (1) (1) (2) (2) (2)
= 0.3sign(xk )xk αk + 0.4xk sign(xk )αk ,
0.1
× C̄k+1 )T + (1 + ε−1 )σΞk+1 ΞTk+1
{ } g(xk , βk )
+ Ξk+1 Π2 tr Xk+1 Ω̄2 ΞTk+1 [ ]
(1) (1) (1) (2) (2) (2)
T =0.3 0.3sign(xk )xk βk + 0.4xk sign(xk )βk ,
+ Ξk+1 Ek+1 Vk+1 Ek+1 ΞTk+1
(i) (i)
= Qk+1|k+1 . where αk and βk (i = 1, 2) stand for zero-mean uncorre-
So far, we have proved that Qk|k is an upper bound of lated Gaussian white noises with unity covariances. Based
Pk|k . on the expressions above, it is straightforward to see that:
Next we are going to show that the filter gain given in (21) {[ ] }
minimizes the upper bound Qk+1|k+1 at each time step. It h(xk , αk ) 
E x = 0,
follows from (22) and (23) that g(xk , βk )  k
Qk+1|k+1 =(1 + ε)Qk+1|k + Ξk+1 Yk+1 ΞTk+1 − Zk+1T
ΞTk+1 and
{[ ][ ]T  }
− Ξk+1 Zk+1 . (30) h(xk , αk ) h(xj , αj ) 
E  xk
T
g(xk , βk ) g(xj , βj ) 
Since Yk+1 = Yk+1 > 0, completing the square with 
respect to Ξk+1 in (30) yields that  0, if k ̸= j,
[ ] [ ]
−1 −1 T 0.04 0.02 0
Qk+1|k+1 =(Ξk+1 − Zk+1
T
Yk+1 )Yk+1 (Ξk+1 − Zk+1
T
Yk+1 ) =
xTk
0.09 0

 0.02 0.01 0 0 0.16 k
x , if k = j.
−1
− Zk+1
T
Yk+1 Zk+1 + (1 + ε1 )Qk+1|k . (31) 0 0 0.09
T −1 Other variables are set as: Q0|0 = 20I, σ = 0.05, and ε = 1.
Thus, it is obvious that when Ξk+1 = Zk+1 Yk+1 , Qk+1|k+1
is minimized and, in such a case, The additive fault is set as
{
Qk+1|k+1 = − Zk+1T −1
Yk+1 Zk+1 + (1 + ε)Qk+1|k . (32) −1, if k ≥ 30,
fk =
That concludes the proof. 0, otherwise.

1858
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

0.2 0.6
yk
ỹk
0 0.5

−0.2 0.4

−0.4 0.3

−0.6 0.2

(2)
−0.8 0.1 Actual state xk
(2)
Estimated state x̂k|k
−1 0

−1.2 −0.1
0 20 40 60 80 100 0 20 40 60 80 100
Time step k Time step k

(2)
Fig. 1. The actual measurement yk and transmitted mea- Fig. 3. The state xk and its estimation
surement ỹk

0.2
0.2 Actual fault
(1) Estimated fault
Actual state xk
(1) 0
0 Estimated state x̂k|k

−0.2
−0.2

−0.4
−0.4

−0.6
−0.6

−0.8
−0.8

−1
−1

−1.2
−1.2 0 20 40 60 80 100
0 20 40 60 80 100 Time step k
Time step k

(1) Fig. 4. The fault and its estimation

Fig. 2. The state xk and its estimation
Fig. 1 depicts the actual measurement and the transmitted mated in a unified framework. A simulation example has
measurement. Figs. 2-4 plot the states/fault and their been presented to show the effectiveness of the proposed
estimations obtained from Theorem 4. It can be seen that, method.
the proposed filter could estimate the states and fault
well with event-triggered measurement transmissions and
additive stochastic nonlinearities. REFERENCES
D.V. Dimarogonas, E. Frazzoli, and K.H. Johansson. Dis-
5. CONCLUSION tributed event-triggered control for multi-agent systems.
IEEE Transactions on Automatic Control, 57(5):1291–
The least squares state and fault estimation problem has 1297, 2012.
been investigated for a class of systems with stochastic H. Dong, Z. Wang, D.W.C. Ho, and H. Gao. Robust H∞
nonlinearities and event-triggered measurement transmis- filtering for Markovian jump systems with randomly oc-
sions. An event-triggered scheme has been properly pro- curring nonlinearities and sensor saturation: The finite-
posed where the measurement output is transmitted to a horizon case. IEEE Transactions on Signal Processing,
remote estimator only when a specified event condition is 59(7):3048–3057, 2011.
violated in an event generator. An appropriate filter gain J. Hu, Z. Wang, H. Gao, and L.K. Stergioulas. Extended
has been determined to minimize an upper bound of the Kalman filtering with stochastic nonlinearities and mul-
filtering error covariance with event-based transmissions tiple missing measurements. Automatica, 48(9):2007–
and additive stochastic nonlinearities at each time step. By 2015, 2012.
solving two sets of discrete matrix equations, the desired J. Hu, Z. Wang, B. Shen, and H. Gao. Gain-constrained re-
filter gain could be calculated recursively, and thus the cursive filtering with stochastic nonlinearities and prob-
method is applicable for online computation. Additive abilistic sensor delays. IEEE Transactions on Signal
faults and system states have been simultaneously esti- Processing, 61(5):1230–1238, 2013.

1859
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

S. Hu, and D. Yue. Event-based H∞ filtering for networked

system with communication delay. Signal Processing,
92(9):2029–2039, 2012.
L. Ma, Z. Wang, Y. Bo, and Z. Guo. A game theory ap-
proach to mixed H2 /H∞ control for a class of stochastic
time-varying systems with randomly occurring nonlin-
earities. Systems and Control Letters, 60(12):1009–1015,
2011.
K. Reif, S. Günther, E. Yaz, and R. Unbehauen. Stochastic
stability of the discrete-time extended Kalman filter.
IEEE Transactions on Automatic Control, 44(4):714–
728, 1999.
B. Shen, Z. Wang, H. Shu, and G. Wei. H∞ filtering for
uncertain time-varying systems with multiple randomly
occurred nonlinearities and successive packet dropouts.
International Journal of Robust and Nonlinear Control,
21(14):1693–1709, 2011.
J. Sijs, and M. Lazar. On event based state estimation.
Hybrid Systems: Computation and Control, 5469:336–
350, 2009.
P. Tabuada. Event-triggered real-time scheduling of sta-
bilizing control tasks. IEEE Transactions on Automatic
Control, 52(9):1680–1685, 2007.
X. Wang, and M.D. Lemmon. Event-triggering in dis-
tributed networked control systems. IEEE Transactions
on Automatic Control, 56(3):586–601, 2011.
J. Wu, Q.S. Jia, K.H. Johansson, and L. Shi. Event-based
sensor data scheduling: Trade-off between communica-
tion rate and estimation quality. IEEE Transactions on
Automatic Control, 58(4):1041–1046, 2013.
L. Xie, L. Lu, D. Zhang, and H. Zhang. Improved robust
H2 and H∞ filtering for uncertain discrete-time systems.
Automatica, 40(5):873–880, 2004.

1860