International Journal of Automation and Computing 12(4), August 2015, 343-367

DOI: 10.1007/s11633-015-0894-x

Recent Progress in Networked Control Systems — A

Survey
Yuan-Qing Xia Yu-Long Gao Li-Ping Yan Meng-Yin Fu
School of Automation, Beijing Institute of Technology, Beijing 100081, China

Abstract: For the past decades, networked control systems (NCSs), as an interdisciplinary subject, have been one of the main
research highlights and many fruitful results from different aspects have been achieved. With these growing research trends, it is
significant to consolidate the latest knowledge and information to keep up with the research needs. In this paper, the results of different
aspects of NCSs, such as quantization, estimation, fault detection and networked predictive control, are summarized. In addition, with
the development of cloud technique, cloud control systems are proposed for the further development of NCSs.

Keywords: Networked control systems, quantization, filter, data fusion, fault detection, networked predictive control, cloud control
systems.

1 Introduction the most important issues to maintain QoS and QoC at the
same time in the research of NCSs.
For the past decades, as a research highlight, networked Due to the existence of the network in NCSs, some com-
control systems (NCSs) not only have attracted the atten- munication problems may degrade the performance of the
tion of many scholars but also have generated huge prof- system. They are presented as follows:
its in the industry. Compared with the traditional control 1) Network-induced delay: The network-induced delay
systems, NCSs have obvious advantages: low cost, flexibil- can be formulated as
ity, easy re-configurability, natural reliability, robustness to
failure, and adaptation capability. Up to now, networked τ (tk ) = τsc (tk ) + τca (tk ) (1)
control techniques have been applied into many fields, such where τ (tk ) is the total transmission delay at sampling time
as transportation networks, power grids, water distribution tk , τsc (tk ) and τca (tk ) are the network delays from the sen-
networks, telephone networks, global financial network, ge- sor to the controller and from the controller to the actuator
netic expression networks, etc. nodes, respectively. For NCSs, there were other delays such
In general, the research of NCSs can be categorized into as τs (tk ), τc (tk ) and τa (tk )[1] , which are the computational
two parts: control of network and control over network. delays in the sensor, controller, and actuator nodes, respec-
The former is to study the problem about network, such as tively. In the existing literatures, different delay models are
routing control, congestion reduction, efficient data commu- introduced. In [2], four major delay models, namely con-
nication, networking protocol, etc. The latter is to achieve stant delay model, mutually independent stochastic delay
the performance of the control system with network being model, Markov chain model, and hidden Markov model,
the transmission media. In this survey, we mainly discuss were summarized.
the latter. With respect to control over network, shared- 2) Packet dropout: Due to several factors such as data
network control systems and remote control systems are two traffic congestion, data collision or interference, packet
important types of NCSs. The research of NCSs is focused loss is an inherent problem in most communication net-
on two aspects: the quality of service (QoS) and the quality works. Generally, the dropout process is often modeled as
of control (QoC). QoS is related to the performance of the a Bernoulli process or a Markov process.
network, such as transmission rates and error rates. QoC 3) Limited capacity of channel: The limited capacity of
generally refers to the stability of the system. It is one of communication channel refers to the limited bandwidth and
the limited transmission energy, which both influence the
Survey Paper transmission quality.
Manuscript received October 1, 2014; accepted January 26, 2015
This work was supported by National Basic Research Program of 4) Network security: Any network medium, particularly
China (973 Program) (No. 2012CB720000), National Natural Sci- wireless medium, is susceptible to easy intercepting. Net-
ence Foundation of China (Nos. 61225015 and 60974011), Founda-
tion for Innovative Research Groups of the National Natural Science work security is receiving more and more concern.
Foundation of China (No. 61321002), Beijing Municipal Natural Sci- Generally speaking, the research of NCSs is mainly to
ence Foundation (Nos. 4102053 and 4101001), Beijing Natural Sci- solve the aforementioned problems. Zhang et al.[3] surveyed
ence Foundation (Nos. 4132042) and Beijing Higher Education Young
Elite Teacher Project (No. YETP1212). the main methodologies which can cope with those typical
Recommended by Editor-in-Chief Guo-Ping Liu network-induced constraints, namely time delays, packet
c Institute of Automation, Chinese Academy of Science and
Springer-Verlag Berlin Heidelberg 2015 losses and disorder, time-varying transmission intervals,
344 International Journal of Automation and Computing 12(4), August 2015

Fig. 1 The framework of NCS

competition of multiple nodes accessing networks, and data constant one with a finite set of values. When a network
quantization. From the framework of a general NCS shown exists in the system, quantization is motivated by the com-
in Fig. 1, the network-induced problems may influence al- munication constraints. To achieve better performance of
most every stage. The whole closed-loop can be described the considered systems, the effect of data quantization on
as follows: The sensors of the system sense the states or the system should be taken into consideration.
the outputs. Due to the uncertainty or the disturbance, A continuous variable is converted into a discretized vari-
these measurements are transmitted through the communi- able by quantizer. Generally, this will inevitably introduce
cation network to estimator, such as filter or data fusion, to information loss due to the existence of quantization error.
estimate the exact states. Then, these processed data are A fundamental problem in NCSs is how the quantization
used to detect the fault or/and compute the control input. error will affect the quality of the system operation. It is
Finally, the control input is fed to the actual control object. easy to see that the less number of bits we use for quan-
Under the effect of the network, most of the existing re- tization, the larger quantization error will be induced. If
sults about NCSs can consider only a single stage in Fig. 1. the quantization error is too large, the controller may not
With respect to the theoretical side, many important im- be able to generate a stabilizing control input. Thus, one
provements and advances have been obtained in the past branch of the study of the quantization is focused on how
decades. The motivation of this survey is to summarize to design a proper quantizer and achieve the stability of the
these important results about NCSs and present a frame- system.
work of NCSs to newcomers in this field. The existence of the quantizer will result in two phe-
The remainder of this paper is organized as follows. Sec- nomenons, i.e., saturation and performance deterioration
tion 2 presents the results about the quantization of NCSs, around the original point. If the signal exceeds the quan-
specially the logarithmic quantizer and the uniform quan- tization range of the quantizer, a large quantization er-
tizer. In Section 3, the estimation stage in NCSs, mainly ror may occur and result in that the closed-loop system
filtering and data fusion, is summarized. Then, the results is unstable. This is the influence of the saturation. If the
about the fault detection for NCSs are simply shown in Sec- state approaches to the original point, the limitation of the
tion 4. In Section 5, since networked predictive control is quantizer s accuracy will lead to that the signal cannot be
a significant approach to NCSs, the corresponding results exactly quantized. In this case, the closed-loop system in-
about predictive control are categorized into model predic- evitably cannot achieve the asymptotic stability.
tive control (MPC), packet based predictive control and The early work on quantized feedback control is mainly
data driven predictive control. Moreover, the distributed motivated by digital computers as instrument for imple-
predictive control is also summarized in this section. Sec- menting control systems. While the network is introduced
tion 6 proposes the cloud control system as one of the fur- as the transmission medium, new techniques need to be
ther developments of NCSs and describes its basic idea. studied. In [4], it is pointed out that when ordinary “lin-
Conclusions are given in Section 7. ear” feedback of quantized state measurements is applied,
the closed-loop system will behave chaotically. When the
2 Quantization state is one-dimensional, a quantitative statistical analysis
of the resulting closed-loop dynamics reveals that the exis-
Quantization is an significant problem in NCSs. Ow- tence of an invariant probability measure on the state space,
ing to the limited transmission capacity of the network, which is absolutely continuous with respect to the Lebesgue
signals must be quantized before they are sent to the net- measure and with respect to which the closed-loop system
work. Quantization is implemented by a quantizer, which is ergodic. In addition, as a hot topic, the event-triggering
is a device to convert a real-valued signal into a piecewise control with quantization was explored[5−8] , Hu and Yue[6]
 Y. Q. Xia et al. / Recent Progress in Networked Control Systems – A Survey 345

considered both state and control input quantization and lization of single-input linear systems with quantized con-
proposed an asymptotical stability criteria in terms of lin- trol feedback. Fu and Xie[10] extended the results of [9] to
ear matrix inequalities (LMIs). multiple input multiple output (MIMO) systems and pre-
Nowadays, the logarithmic quantizer and the uniform sented that the quantization feedback design problem can
quantizer are two popular quantizers. When one compares be converted into the robust control problem. This is cru-
the logarithmic quantizer with the uniform one, it is not cial for the research of the logarithmic quantizer. In [11],
hard to obtain that the former yields better performance quantization issues were studied in a remote control sys-
around the origin and the latter can be operated more eas- tem. Hayakama et al.[12, 13] studied the adaptive quantized
ily. Each kind of quantizer has both advantages and disad- control for the linear uncertain discrete-time systems and
vantages. In the following, we will discuss the logarithmic nonlinear uncertain systems, respectively. You et al.[14] in-
quantizer and the uniform quantizer, respectively. vestigated the attainability of the minimum average data
rate for stabilization of linear systems via logarithmic quan-
2.1 Logarithmic quantizer tization. By exploring some geometric properties of the
The logarithmic quantizer is a kind of the static quan- logarithmic quantizer and using the fact that the logarith-
tizer. It is formulated as mic quantizer is sector bounded and nondecreasing, Zhou
⎧ vi vi et al.[15] presented a new approach to the stability analysis
⎪
⎨ vi , if 1 + δ < y < 1 − δ , y > 0
⎪ of quantized feedback control systems.
q(y) = 0, if y = 0 (2) For NCSs, the time delay and the packet loss are two
⎪
⎪
⎩ −q(−y), if y < 0 usual issues. Due to their existence, most of the traditional
logarithmic quantizers are not suitable. Up to now, some
where U = {±vi : vi = ρi v0 , i = ±1, ±2, · · · } ∪ {±v0 } ∪ quantization results for the NCSs have been obtained. In
{0}, 0 < ρ < 1, v0 > 0 is the set of the quantization values. [16], the guaranteed cost control design for linear systems
This quantizer is of infinite quantization level. A quantizer with parameter uncertainty was considered. The logarith-
with limited quantitative level is mic quantizers were used in two sides and the effects of both
⎧ vi vi the quantization levels and the network conditions are also
⎪
⎪ vi , if <y< ,0 < i ≤ N −1
⎪
⎪ 1 + δ 1 −δ considered. Zhang et al.[17, 18] discussed a class of H2 fil-
⎨ 0, if 0 ≤ y ≤ vN−1
⎪
ters and H∞ filters for nonlinear systems with logarithmic
q(y) = 1+δ (3)
⎪ v0 quantizer and packet loss, respectively. Li et al.[19] consid-
⎪ v0 , if y >
⎪
⎪
⎪ 1 + δ ered the robust H∞ control problem for uncertain discrete-
⎩
−q(−y), if y < 0 time Takagi and Sugeno (T-S) fuzzy networked control sys-
tems with state quantization, where a logarithmic quan-
where U = {±vi : vi = ρi v0 , i = ±1, ±2, · · · , ±(N − 1)} ∪
tizer was used and new model of network-based control with
{±v0 } ∪ {0}, 0 < ρ < 1, v0 > 0 is the set of the quanti-
simultaneous consideration of network-induced delays and
zation values. Fig. 2 shows an example of the logarithmic
packet dropouts was constructed. A less conservative delay-
quantizer.
dependent stability condition was thus derived. In [20], the
input quantization and packet dropouts were considered si-
multaneously. The packet dropout process of the channel
was modelled as an independent identically distributed case
and a logarithmic quantizer was adopted. It turns out that
as long as the dropout rate satisfies α < 1 + ln δ−llnln(1+δ)
Λ
,
the system is globally exponentially stabile, i.e.,

 
(t)∞ ] < e−λt X
E [x(t) − x

where t > 0, E represents the mean, Λ = eF T , T is the

sample interval, F is one positive real number, δ > ρN−1 ,
N ≥ 3, 0 < ρ < 1, x(0)∞ < X, λ = lnTR and
Fig. 2 An example of logarithmic quantizer R = Λ( 1+δ δ
)1−α < 1. Xia et al.[21] considered a quan-
tized system with finite-level quantized input computed
from quantized measurements (QIQM). The problem of
The logarithmic quantizer was firstly proposed in [9]. globally asymptotic stability of QIQM system is converted
Elia and Mitter[9] turned out that the logarithmic quantizer into the one of an equivalent system depending on a mul-
is the coarsest quantizer which has the minimum quantiza- tiplier which is nonnegative and bounded. If the open-loop
tion density. This work also discussed the quadratic stabi- system is stable and the closed transfer function satisfies
346 International Journal of Automation and Computing 12(4), August 2015

that inf w∈R Re{G(ejw )} > ε, where

⎧ 1 1
⎪
⎪ Λ, if 0 < δ ≤
⎪
⎪ (δ − 1)(δ + 1) 2 2
⎨
(2δ − 1)δ 1
q(y) = Λ+ Λ,
⎪
⎪ (1 − δ) 2 (δ + 1)2 (δ − 1)(δ + 1)2
⎪
⎪
⎩ 1
if ≤ δ < 1
2
Fq δFq 1
Λ= −
Q(Fq ) Q(Fq )
Fq = inf F̃ = inf F η|η ∈ Zη = ±ρi v0 , i ∈ {0, i, · · · , N },
ρ+1
Fη ≥ v0 (4)
2ρ
Fig. 3 An example of uniform quantizer
and F is the feedback matrix, Q is the quantizer, ρ > 1 and
v0 > 0, then the closed-loop system is globally asymptoti- Brokett et al.[22, 23] proposed the uniform quantizers with
cally stable about origin. an arbitrarily shaped quantitative area. Based on these
As we can see, the existing results involving logarith- quantizers, the zoom theory is used for linear systems, and
mic quantizer are mainly related to the linear systems with the nonlinear systems, and the sufficient condition for the
infinite quantization levels. Then, how to analyze the sta- asymptotical stability is given. The stabilization problem
bility of QIQM nonlinear systems, which are affected by for discrete time linear systems with multidimensional state
finite logarithmic quantization levels, is a hard problem to and one-dimensional input using quantized feedbacks with
be solved. Moreover, the problems mentioned above, which a memory structure was investigated in [24]. Bullo and
are affected by other networked induced issues such as time- Liberzon[25] discussed the local optimization of the quan-
delay and packet dropout simultaneously, are beyond doubt tized control. In order to obtain a control strategy which
the focal and difficult points of future study. yields arbitrarily small values of lnTC (a weak form of the
LN
pole assignability property), ln has to be big enough,
2.2 Uniform quantizer C
where T is the time that shrinks the state of the plant from
The uniform quantizers with an arbitrarily shaped quan- a starting set to a target set, C is the contraction rate, L
titative area have some following conditions: is the number of the controller states and N denotes the
1) if y2 ≤ M μ, then q(y) − y2 ≤ Δμ number of the possible control values. Reference [26] is the
2) if y2 > M μ, then q(y)2 > M μ − Δμ extension of the previous results. Reference [27] presented
where M is the saturation value and Δ is the sensitivity. a general framework of the design and analysis approach for
The first condition gives the upper bound of the quanti- the quantized systems.
zation error when the quantization is not saturated. The Meanwhile, considering the delay and the packet loss
second condition provides an approach to test whether the which the network induces, some improvements for the uni-
quantization is saturated. Specially, as shown in Fig. 3, an form quantizer have been obtained. For the linear discrete-
uniform quantizer whose quantitative area can be rectangle time system with quantization and packet dropout, You
is formulated as and Xie[28, 29] proposed the necessary and a sufficient con-
⎧ dition for the mean square stability through quantization
⎪ 1 1
⎪
⎪ 0, if − < y < feedback. The dropout process of [28] is a Bernoulli process
⎪
⎪ 2 2
⎪
⎪ 2i − 1 2i + 1 and the dropout process of [29] is a Markov chain. However,
⎨ i, if <y< , i = 1, 2, · · · , K − 1
2 2 in these works, the relationship between the dropout rate
q(y) = 2K − 1
⎪
⎪ K, if y ≥ and the eigenvalues of the system matrix influences the per-
⎪
⎪
⎪
⎪ 2 formance of the quantized system. The quantitative area of
⎪
⎩ 1
−q(−y), if y ≤ − . the uniform quantizer is a rectangle. Using uniform quan-
2
(5) tizer, Tian et al.[30] explored quantised stabilisation and
quantised H∞ control design for NCSs. The effects of both
When a uniform quantizer is used, the zoom strategy is network-induced delay and quantization levels were consid-
a useful control policy, which is divided into two stages: ered. The design was shown in terms of LMIs. For the
“zoom-in” and “zoom-out”. If the initial state to be quan- nonlinear system with quantization, Persis[31] turned out
tized is saturated, “zoom-out” stage is adopted to increase that if the nonlinear system can be stabilised by “standard”
the sensitivity Δ until the state gets unsaturated. In the state feedback, then it can be stabilised by encoded state
“zoom-in” stage, the sensitivity Δ is decreased to drive the feedback. In [32], considering data rate constraints, Persis
state to zero. Otherwise, if the initial state to be quantized studied the problem of stabilisation of nonlinear systems
is unsaturated, the “zoom-out” stage can be omitted and using output measurements, and an “embedded-observer”
the “zoom-out” stage is implemented directly. decoder and a controller were designed to guarantee the ro-
 Y. Q. Xia et al. / Recent Progress in Networked Control Systems – A Survey 347

bust practical stability of the system. Yan et al.[33] consid- the performance of the system. For example, in order to re-
ered the mean square stability based on the zoom strategy duce the number of measurements to be transmitted from
and Lyapunov theory. Given a fixed packet dropout rate, sensor to estimator, the design of transmission scheduler
the sufficient condition for the stability of the closed-loop and estimator for linear discrete-time stochastic systems
system was given in the form of LMI. It is independent of was considered[38] . However, NCSs are mostly nonlinear
the eigenvalues of the system matrix. Moreover, the quan- systems. Filters, which are fit for nonlinear systems, are ex-
tizer is of arbitrarily shaped quantitative area. When the tremely desired. In general, there are two kinds of methods
dropout process is a Markov process for the given failure when designing the filtering for nonlinear dynamic systems.
rate and recovery rate, sufficient conditions were given for One is to linearize the nonlinear systems. The extended
the closed-loop fuzzy systems to be mean square stable, Kalman filter (EKF) proposed by Sunahara and Bucy was
and the feedback controllers were designed to ensure the motivated by this thought. The other one is to adopt the
mean square stabilities of fuzzy systems[34] . Xia et al.[35] sample of nonlinear distribution. Typical filters are the par-
explored the design and stability analysis of networked con- ticle filter and the unscented Kalman filter (UKF) based on
trol systems with quantization and noise. Kalman filter unscented transformation (UT). Considering the characters
(KF) is used to obtain the estimated states and the quan- of the NCSs such as the delay, the packet dropout and the
tization regions of uniform quantizer were of rectangular channel attenuation, the filtering for NCSs is researched by
shapes. Based on Lyapunov theory and invariable set the- many scholars. In the following, we will present the recent
ory, a sufficient condition for the stability of the closed- research results of the filtering for linear systems and non-
loop system was given. Under the effect of the quantization linear systems, respectively.
and the unknown disturbance, for the linear system, Sharon 3.1.1 Filtering for linear systems
and Liberzon[36] studied the input-to-state stability of the Up to now, there are numerous filtering approaches for
closed-loop system. The proposed controller switches re- linear systems. In general, the continuous-time linear sys-
peatedly between “zooming-out” and “zooming-in”. Two tem is formulated as
modes to implement the “zooming-in” phases, which at-
ẋ(t) = Ac (t)x(t) + w(t)
tenuate an unknown disturbance while using the minimal
number of quantization regions, were designed. z(t) = Cc (t)x(t) + v(t) (6)
Similar to logarithmic quantizer, the problem of stability and the discrete-time vision is
of nonlinear systems, especially for QIQM ones, which are
affected by uniform quantizer with zoom stagey, time-delay x(k + 1) = Ad (k)x(k) + w(k)
and packet dropout, is the main and hard direction in the z(k) = Cd (k)x(k) + v(k) (7)
future study. Moreover, the quantized value of zoom stagey
in each step is undertrained. Therefore, it may be the inter- where x is the state, z is the measurement, Ac or Ad is the
esting problem to combine the zoom stagey and logarithmic system matrix, and Cc or Cd is the output matrix, w is the
quantizer to confirm each step quantized value to improve process noise, and v is the measurement noise.
the closed-loop system performance. In the last few years, filtering with random packet
dropouts in NCSs was the focus of several research
3 Filter and data fusion studies[39−41] . As in [40], the H∞ filtering for NCSs with
the packet loss is of the following form
Due to the complexity of the network, state estimation
x̂(k + 1) = af x̂(k) + bf u(k) + cf y(k) (8)
has become a crucial and significant research field in NCSs.
In general, there are two kinds of approaches for state esti- ẑ(k) = Lf x̂(k) (9)
mation: filtering and data fusion. A filter is usually used to where x̂(k + 1) is the estimate of the state, u(k) is the
estimate the states of the system with measurement noise. system input, y(k) is the output, ẑt is the estimated
Data fusion is the technique in which data from multiple signal, and af , bf , cf , Lf are the filter parameters to be
sensors and related information are combined to improve designed. A set of linear matrix inequalities is used
accuracy. Mostly, many scholars use filter-based fusion to to solve the corresponding optimal filter design problem,
get a more accurate state estimation. In the following, we which is convex. Zhang et al.[41] studied the H∞ filter-
will discuss the filtering and data fusion in NCSs, respec- ing problem for networked discrete-time general multiple-
tively. input-multiple-output (MIMO) system with random packet
losses. Each measurement loss process is described by a
3.1 Filtering
two-state Markov chain. The desired filters with minimized
The filter is to estimate the states or a linear combi- H∞ noise attenuation level bound is designed by solving a
nation of them by using the measured system inputs and convex optimization problem. The presented necessary and
outputs. KF was first proposed by Kalman[37] in 1960 and sufficient condition describes a relation between the packet
it is mainly used for the linear systems. KF is a kind of loss probability and two parameters (the exponential decay
optimal filter and can be easily and properly implemented rate of the filtering error system and the H∞ noise attenu-
by computer. Some modified KFs are proposed to improve ation level), to make the system mean-square exponentially
348 International Journal of Automation and Computing 12(4), August 2015

stable and achieve a prescribed H∞ noise attenuation per- systems with random sensor delay, multiple packet dropout
formance. Sinopoli et al.[42] considered the discrete-time and uncertain observation was investigated. The stochastic
linear time-invariant system similar to (7) and designed the H2 -norm of the estimation error was used as a criterion for
modified KF with the random arrival: the filter design and the filter was designed by solving a
set of LMIs. Reference [48] is concerned with the filtering
x̂(k + 1|k) = Ax̂(k|k) (10)
problem for a class of discrete-time stochastic nonlinear net-
T
P (k + 1|k) = AP (k|k)A + Q (11) worked control systems with network-induced incomplete
T measurements (the multiple random communication delays
x̂(k + 1|k + 1) = x̂(k + 1|k) + P (k + 1|k)C ×
   and random packet losses). By using the LMI method and
CP (k + 1|k)C T + γ(k + 1)R + 1 − γ(k + 1) × delay-dependent technique, sufficient conditions are derived
−1   for the exponentially mean-square stability of the filtering-
σ2 I × z(k + 1) − C x̂(k + 1|k) (12) error dynamics. Shi and Fang[49] considered the problem
of parameter estimation and output estimation for systems
P (k + 1|k + 1) = P (k + 1|k) − P (k + 1|k)C T × in a transmission control protocol (TCP). The input and
   output missing data is modeled as two separate Bernoulli
CP (k + 1|k)C T + γ(k + 1)R + 1 − γ(k + 1) processes characterised by probabilities of missing data, and
−1 a recursive algorithm for parameter estimation by modify-
σ2I × CP (k + 1|k) (13)
ing the KF-based algorithm is developed.
where Q and R are the covariance matrices of w and v, From the existing results above, we have that the fil-
γ(k) is the drop loss rate, σ 2 I is the variance of the obser- tering problem is often formulated by probabilistic ap-
vation if the packet is loss. And in addition to the special proaches based on the probabilities of the uncertainties oc-
instructions, T represents the transposition of a matrix in curring between the sensor and the filter. However, a non-
the survey. Taking the limit as σ → ∞, the update equa- probabilistic approach by time-stamping the measurement
tions can be obtained as packets was proposed in [50]. The optimal state estimator
was dependent on the possible control input and this ap-
x̂(k + 1|k + 1) = x̂(k + 1|k) + γ(k + 1)× proach was suitable for single-measurement packets, multi-
K(k + 1)(z(k + 1) − C x̂(k + 1|k)) (14) ple measurement packets, the case of burst arrivals (where
P (k + 1|k + 1) = P (k + 1|k) − γ(k + 1)× more than one packet may arrive between the receivers pre-
vious and current sampling times), and the scenario where
K(k + 1)CP (k + 1|k) (15)
the control input is non-zero and subject to delays and
where K(k + 1) = P (k + 1|k)C T (CP (k + 1|k)C T + R)−1 is packet dropouts. Yue and Han[51] not only considered the
the filtering gain matrix and γk is a random binary variable transmission delay and the packet dropout, but also stud-
which represents that whether the observation arrives. ied the parameter uncertainty. The proposed criteria for
The state delay is another issue when designing the fil- H∞ performance analysis of the filtering-error system are
ter. He et al.[43] designed a new robust H∞ filter for a expressed as a set of linear matrix inequalities. The fil-
class of networked systems with multiple state-delays. Cer- ter design can be obtained by using convex optimization
tain LMIs, which depend on the occurrence probability of method. Similarly, Xia and Han[52] also considered the sys-
both the random sensor delay and missing measurement, tem uncertainty and gave the sufficient condition for the
are established to ensure the existence of the desired filters. design of the robust KF. Yang et al.[53] considered a delta
Hu and Yue[44] explored event-based H∞ filtering for net- operator KF and applied it to an inverted pendulum model
worked systems with communication delay. The sufficient in experiment. A new sufficient condition of convergence
conditions for the exponential stability was given in terms analysis for the designed delta operator KF based on Lya-
of LMIs. punov functional was given. In [54], a data-driven subspace
Mostly, many other results considered both the packed identification method combined with the Kalman on-line
dropout and the state delay at the same time[42, 45−49] . In filtering algorithm was proposed to solve the state estima-
[45], the adaptive filtering schemes were proposed for state tion problem for a class of dynamical systems where the
estimation in NCSs with mixed uncertainties of random exact models can not be established. In addition, consider-
measurement delays, packet dropouts and missing measure- ing constrained communication energy and bandwidth, an
ments. The filter gains can be obtained by solving a set of event-based sensor data scheduler for linear systems was
recursive discrete-time Riccati equations. In [46], for linear designed in [55] and the corresponding minimum squared
systems with both random delay and packet drop, Schen- error estimator was also derived.
ato presented two alternative estimator architectures which In the following, we will introduce some filtering applica-
are more computationally efficient, and provided upper and tions in nonlinear NCSs.
lower bounds for the performance of the time-varying esti- 3.1.2 Filtering for nonlinear systems
mator. The stability of these estimators does not depend For nonlinear systems, the stability analysis of nonlin-
on packet delay but only on the overall packet loss probabil- ear filter with random packet dropouts remains open. Up
ity. In [47], the problem of optimal filtering of discrete-time to now, the dropouts have been modeled as the Bernoulli
 Y. Q. Xia et al. / Recent Progress in Networked Control Systems – A Survey 349

process[42] or a Markov chain[56] . In [57], the H∞ filter- f and h are continuously differentiable. wk ∈ Rw and
ing problem was investigated for a new class of discrete- vk ∈ Rv are the process and the measurement noise signal,
time networked nonlinear systems with both random delays respectively, which are assumed to be uncorrelated zero-
(discrete delays and infinite distributed delay) and packet mean Gaussian white noise processes. E(wl wkT ) = Qk δkj
dropouts (Bernoulli distribution). Dong et al.[58] studied and E(vl vkT ) = Rk δkj . Qk and Rk are positive matrices. x0
the robust H∞ filtering problem for a class of uncertain is uncorrelated with wk and vk .
nonlinear networked systems with both multiple stochas- Case 1. The packet dropout is modeled as a Bernoulli
tic time-varying communication delays and multiple packet process with parameter γ ∈ [0, 1].
dropouts. This work aimed to design a linear full-order Define a binary stochastic variable γk as
filter such that the estimation error converged to zero ex-
ponentially in the mean square while the disturbance re-
1, a measurement arrives after the k-th step
jection attenuation was constrained to a given level by γk =
0, otherwise
means of the H∞ performance index. The communica-
tion delays were regarded as a sequence of random vari-
ables which are mutually independent and obey Bernoulli and denote the variance of the output noise at time k as
distribution. The packet dropout phenomenon occurs with Rk if γk = 1, and σ 2 I if γk = 0 for some σ 2 . Sinpoli et
a certain probabilistic distribution in the interval [0, 1]. The al.[42] pointed out that the absence of an observation corre-
EKF is the most popular approach to recursive nonlinear sponds to the limiting case of σ → ∞. The modified UKF
estimation[59−61] . In some simulations, it overweighs other is described as[67] :
sophisticated filters[61] . However, a large linearization er- Step 1. Select sigma points. Assume that xk has mean
ror, the derivation of Jacobian matrices and limitation of xk and covariance P k .
Gaussian noise are three well-known drawbacks of the EKF.
When the systems are strongly nonlinear and the noise is χ0,k = x̄k (18)
non-Gaussian, desired performance characteristics may not √
χi,k = x̄k + (a nP̄k )i , i = 1, · · · , n (19)
be obtained. In these cases, the particle filtering (PF) meth- √
ods proposed by Gorden in 1993 based on Monte Carlo χi,k = x̄k − (a nP̄k )i , i = n + 1, · · · , 2n (20)
simulation are available[62] . Recently, the PF has been ex-
√
tensively used in the fields of simultaneous localization and where a is a proportion parameter and ( nP̄k )i is the vec-
mapping (SLAM) for robots[63] and visual tracking[64] . In tor of the i-th column of the matrix square root.
[65], a continuous-discrete version of particle filter (CD-PF) Step 2. Prediction.
for continuous-discrete nonlinear systems was proposed and
some comparisons between CD-PF and CD-EKF were pre-
χi,k+1|k = f (χi,k ), i = 1, · · · , 2n (21)
sented.
2n
UKF is also a filtering method used for nonlinear sys-
x̂k+1|k = i χi,k+1|k (22)
tems. Kluge et al.[66] analyzed the error behavior of the i=0
EKF, and models the arrival of the observation as a random 2n
process and a maximum dropout interval. Li and Xia[67] ex- P̂k+1|k = T
i χ̃i,k+1|k χ̃i,k+1|k + Qk (23)
tended the idea of [66] to the case of UKF, and presents an i=0
analysis of UKF for general nonlinear stochastic systems
with intermittent observations. Under independent identi- where χ̃i,k+1|k = χi,k+1|k − x̂k+1|k , 0 = 1 − a12 , i = 1
2na2
cal distribution packet dropouts, stability of the estimator (i = 1, · · · , 2n), and 2n
i=0 i = 1.
may be analyzed by a modified discrete-time Riccati recur- Step 3. Update.
sion effectively[66, 67] . In contrast, this approach is not fea-
sible for the Markovian packet dropouts model any more
zi,k+1|k = h(χi,k+1|k ), i = 0, 1, · · · , 2n (24)
since the channel is described by several independent pa-
2n
rameters, and the stability analysis faces more challenging.
ẑk+1|k = i ẑi,k+1|k (25)
Next, we will give the detailed descriptions of some results
i=0
about UKF. 2n
Consider the discrete nonlinear system which is formu- T
P̂zz,k+1|k = z̃i,k+1|k z̃i,k+1|k + γk+1 Rk+1 +
lated as i=0

(1 − γk+1 )σ 2 I (26)
xk+1 = f (xk , uk ) + wk (16)
2n
zk+1 = h(xk+1 ) + vk+1 (17) P̂xz,k+1|k = T
χ̃i,k+1|k z̃i,k+1|k (27)
i=0
where k ∈ N is the discrete time, and N = {0, 1, · · · }. −1
x̂k+1 = x̂k+1|k + P̂xz,k+1|k P̂zz,k+1|k (zk+1|k − ẑk+1|k ) (28)
xk ∈ Rn is the state, uk ∈ Rm is the control input, and −1 T
zk ∈ Rq is the measured output. The nonlinear functions P̂k+1 = P̂k+1|k − P̂xz,k+1|k P̂zz,k+1|k P̂xz,k+1|k (29)
350 International Journal of Automation and Computing 12(4), August 2015

where z̃i,k+1|k = zi,k+1|k − ẑi,k+1|k . As σ → ∞, estimation error and the boundedness of the estimation er-
ror covariance matrices were obtained in [68].
x̂k+1 = x̂k+1|k + γk+1 Kk+1 (zk+1|k − ẑk+1|k ) (30) Under some conditions (see Assumption 1 in [68]), the
T
P̂k+1 = P̂k+1|k − γk+1 Kk+1 P̂xz,k+1|k (31) mean of the estimation error E(x̃k 2 ) is bounded by
2n ηE(x̃0 2 )(1 − λ)n + ν, where η, ν > 0.
T
Kk+1 = P̂xz,k+1|k [ z̃i,k+1|k z̃i,k+1|k + Rk+1 ]−1 . (32) Under the condition that (Fk , Hk ) is uniform observable,
i=0 the stability in stopping times and the sampling times is
Step 4. Repeat Steps 1–3 for the next sample. given respectively as
From Theorem 1 in [67], the boundedness of the estima- 1) Stability in stopping times: If the recovery rate q sat-
tion error for UKF is obtained as follows: isfies that ᾱ2 f¯2 (1−q) < 1, then supk∈N E[P̂τk |τk−1 ] < ∞.
Theorem 1[67] . If the following conditions hold: 2) Stability in sampling times: If the conditions ᾱ2 f¯2 +
1) there exist real positive constants f , h, β, α, ᾱ, q, σ̄ ᾱ2 f¯2 + 2σ̄ h̄2 ḡ 2 < 1 and P1|0 > 0 are satisfied, then
q̄, q ∗ , q̂,
¯ r̄, r ∗ , p, p, for every k > 0, such that supk∈N E[P̂k|k−1 ] < ∞.
The details can be referred to Theorems 1 and 2 in [68].
f 2 I ≤ Fk FkT , h2 ≤ Hk HkT , βI ≤ βk (33) In fact, there exist fruitful results about filtering for the
∗ general system. When network is adopted as the communi-
α ≤ αk ≤ ᾱI, qI ≤ Qk ≤ q̄I, q I≤ Q∗k (34)
cation medium, the previous filtering methods are improved
∗
Rk ≤ r̄I, r ≤ Rk∗ , pI ≤ P̂k ≤ p̄I (35) to deal with the network-induced problems, such as delay,
packet loss, etc. In the future, in the filter design process,
2) there exist real positive constants f¯, φ̄, h̄, β̄ such that
more practical factors, such as bandwidth, network load
γk Fk  ≤ f¯, φk  ≤ φ̄ (36) and sensor saturation, should be taken into consideration.
γk Hk  ≤ h̄, βk  ≤ β̄ (37) 3.2 Data fusion
then the estimation error is exponentially bounded in mean Data fusion is a broad research area, which includes de-
square and bounded with probability 1. tection fusion, image fusion, data association, estimation
In addition, the error covariance matrices are also fusion, etc. In this section, data fusion mainly refers to
bounded for the modified UKF. estimation fusion. In addition to the filtering, estima-
Theorem 2[67] . Assume that the linearized dynamic sat- tion fusion is another important approach to estimation.
isfies a modified uniform observability condition as in [66], In general, filtering often deals with the data from sin-
and that the following conditions hold: gle sensor. In contrarily, estimation fusion is the estima-
1) P̂1|0 > 0 and Hk−1 exists and satisfies Hk−1  ≤ h−1 tion which involves multi-sensor data. Multi-sensor data
2) there exists real positive constants h, β, r̄ ∗ , φ̄, for fusion has been one of the key technologies of networked
every k > 0, such that control systems, which have been served as an interesting
benchmark in the past decades. The aim of multi-sensor
h2 I ≤ Hk HkT , βI ≤ βk , Rk∗ ≤ r̄ ∗ I data fusion is to increase the accuracy of the estimation.
Qk ≤ q̄I, φk  ≤ φ̄. It has been applied to both military field and nonmilitary
field. Military applications include automated target recog-
ᾱ2 f¯2 −1
If 1 > γ > ᾱ2 f¯2 φ̄2
, then there exists a upper boundedness nition, guidance for autonomous vehicles, remote sensing,
p̄ > 0 such that E(P̂k+1 ) ≤ p̄I. battlefield surveillance, and automated threat-recognition
Case 2. The packet dropout is modeled as a Markov systems, such as identification-friend-foe-neutral systems.
chain with a transition probability matrix as Nonmilitary applications include monitoring of manufac-
  turing processes, condition-based maintenance of complex
1−q q machinery, robotics, and medical applications.
Π = (P {γk+1 = j|γk = i})i,j∈{0,1} =
p 1−p Multi-sensor data fusions can be divided into three cate-
gories:
where γk = 1 represents that the packet has been suc- 1) Centralized data fusion. It is also called measure-
cessfully transmitted to the filter, otherwise γk = 0, and ment fusion. All the measurements from different sensors
0 < p, q < 1. are transmitted to the fusion center and dealt within the
Li and Xia[68] considered the nonlinear system (16) with fusion center. Due to the full involvement of all the mea-
linear observation as surements, this method is optimal. The drawback is strin-
zk = Hk xk + vk . (38) gent requirements for bandwidth of the channels, enormous
energy consumption and processor with high performance
The modified UKF in [68] is similar to [67], in which in the fusion center.
the packet dropout is a Bernoulli process. Due to the lin- 2) Distributed data fusion. It is also called state fu-
T
ear observation equation, P̂zz,k+1 = Hk+1 P̂k+1|k Hk+1 + sion. After locally preprocessing the measurement of each
2
γk+1 Rk+1 + (1 − γk+1 )σ I and as σ → ∞, Kk+1 = sensor, the results are sent to the fusion center. This
T
P̂xa,k+1 [Hk+1 P̂k+1|k Hk+1 + Rk+1 ]. The boundedness of the method not only decreases the demand for the channel, but
 Y. Q. Xia et al. / Recent Progress in Networked Control Systems – A Survey 351

also has a strong fault tolerance ability. node which can obtain global knowledge of network and
3) Hybrid data fusion. This method is the combina- designed a decentralized data fusion scheme with the in-
tion of the centralized data fusion and the distributed data formation of each s nearest communication neighbor. For
fusion. However, the communication cost and the compu- the past decade, distributed data fusion has attracted con-
tation cost are high. siderable attention. In [80], an architectural framework of
In the following, we will mainly present the recent results distributed data fusion was developed and it consisted of a
of the centralized data fusion and the distributed data fu- data fusion application programming interface (API) and
sion, respectively. Sequently, considering the unreliability a distributed algorithm for energy-aware role assignment.
of the network, data fusion in NCSs is also introduced in Mahmoud and Xia[81] studied the distributed filtering and
the end of this subsection. fusion in the wireless sensor network. In [82], support vector
3.2.1 Centralized data fusion machines were used to compress the probabilistic informa-
The early results about centralized data fusion were rel- tion available in the form of independent identical distribu-
atively simple. Singer and Kanyuck[69] firstly studied the tion samples and to solve the Bayesian data fusion problem.
problem of the multi-sensor state estimation. However, it Distributed data fusion has also been applied to many
was assumed that the estimation error of each sensor is mu- fields. Regazzoni and Tesei[83] described an approach to
tually independent. Willner et al.[70] proposed a centralized estimating the number of pedestrians present in a real-life
multi-sensor KF algorithm. Afterwards, in general, more complex scene by processing a sequence of images. The pro-
scholars studied the systems in which the noise of each sen- posed estimation approach exploits temporal information
sor is correlated. The centralized information integration by means of a distributed KF network. In [84], distributed
architecture proposed in [71] is based on the output aug- estimation algorithms based on the information graph were
mented fusion method. The usual approach in the classical presented for arbitrary fusion architectures and related to
EKF implementation is based on the assumption of con- linear and nonlinear distributed estimation results and the
stant diagonal matrices for both process and measurement proposed approaches were applied to tracking.
covariances. A new adaptive modified EKF (AMEKF) al- 3.2.3 Data fusion in NCSs
gorithm has been developed to prevent the filter divergence. Due to the existence of the network-induced delay, the
The centralized data fusion has been widely applied. packet dropout and other characteristics in the network,
Munz et al.[72] constructed a multi-sensor fusion system the traditional data fusion techniques are not proper and
which was implemented to be independent of both sensor some new data fusion approaches for NCSs were studied.
hardware properties and application requirements in the In the following, we will introduce some advances for the
driver assistance system. This supports changes in sensor unreliable networks.
combination or application requirements. Moreover, the en- Besada-Protas et al.[85] proposed a set of new central-
vironmental model can be used at the same time. A prob- ized algorithms for estimating the state of linear dynamic
abilistic approach for this generic centralized fusion system multiple-input multiple-output control systems with asyn-
was presented and discussed. In [73], a centralized fusion chronous, non-systematically delayed and corrupted mea-
scheme was adopted for obstacle and road tracking. A KF surements provided by a set of sensors. A streamlined
based curvilinear model predicted drivers behavior and the memory and computational efficient reorganization of the
system was able to assess the level of threat for all moving basic operations of the Kalman and information filters (KF
obstacles and decide the warning policy. The proposed ap- & IF) were designed. In [86], the optimal distributed fu-
proach has been tested with real data to show the perfor- sion update algorithm with multiple local asynchronous
mance of the tracking system. (1-step-lag) out-of-sequence-measurement (OOSM) update
3.2.2 Distributed data fusion was firstly deduced, which was proved, under some regular-
As we know, the centralized data fusion usually has the ity conditions, to be equivalent to the corresponding opti-
optimal performance for making most of all the informa- mal centralized update algorithm with all-sensor 1-step-lag
tion. Relatively, the distributed algorithm sacrifices the OOSMs. Reference [87] was concerned with the distributed
optimal performance to increase the flexibility and decrease KF problem for a class of networked multi-sensor fusion
the communication cost. If the distributed data fusion is systems (NMFSs) with missing sensor measurements, ran-
said to be optimal, compared to the centralized data fu- dom transmission delays and packet dropouts. An optimal
sion, it often has the approximate optimal performance distributed fusion KF (DFKF) was designed based on the
under some assumptions or conditions. For early results optimal fusion criterion weighted by matrices. In [88], a
about distributed data fusion, please refer to [74–77]. Bar- new networked multi-sensor data fusion method which was
Shalom[78] proposed that there exits the same process noise based on federated filter is proposed.
in the dynamic of different sensors. Thus, the estimation In [89], a networked data-fusion approach with packet
of each sensor is mutually dependent. Based on this, Bar- losses and variable delays was considered and a federated fil-
Shalom[78] presented a distributed state fusion algorithm ter was employed to fuse the data transmitted over the net-
and discussed the influence of process noise on the vari- work, which plays an important role in the data-processing
ance of the estimation error. Grime et al.[79] considered center. The federated filters are of peak covariance stabil-
that there is no central fusion center and no single sensor ity if the delayed data is processed by the proposed algo-
352 International Journal of Automation and Computing 12(4), August 2015

rithm. Zhu et al.[90] proposed a centralized architecture for linear systems when the noises of different sensors
and a distributed architecture with buffer for the problem were cross-correlated and also coupled with the system
of networked data. A modified KF was adopted to estimate noise of the previous step. Bian et al.[97] considered the
the state with measurement delay and loss. In [91], a cen- medium access constraint. By introducing a stochastic
tralised data fusion approach was presented for the system, process, a new model was developed that can truly reflect
in which the measurements of the local sensors are time- this contradiction (the number of transmitted sensors and
stamped, and then transmitted through the network to the delay steps). The upper bounds of the expected estima-
fusion centre. A buffer was designed to deal with the mea- 
tion error covariance and estimation error covariance with
surement delay or loss. A probabilistic metric to evaluate one-step delay are E [Pt|t−1 ] ≤ N̄P r and Pt|t−1 < M̄ ,
the performance of the system was employed. respectively, where N̄P r = A(C1T R1−1 C1 )−1 AT + Q and
Multi-rate framework has caused the interests of many M̄ = A(C0T R0−1 C0 )−1 AT + Q. In [98], energy efficient
scholars because of the decrease of the communication state estimation was considered. The packet loss proba-
cost. Fig. 4 shows an example of multiple sampling rates. bility depended upon time-varying channel gains, packet
Zhang et al.[92] presented a distributed fusion estimation lengths and transmitted power of sensors. The sufficient
method for estimating states of a dynamical process ob- conditions, imposing constraint on the packet loss prob-
served by wireless sensor networks (WSNs) with random ability of each channel were established by taking into
packet losses. It was assumed that the dynamical process is account each observable subsystem structure to guaran-
not changing too rapidly, and a multi-rate scheme by which tee that the expectation of the trace of estimation error
the sensors estimate states at a faster time scale and ex- covariance matrices is exponentially bounded, and the up-
2
change information with neighbors at a slower time scale per bound is E(tr(Pt|t )) ≤ S1 σmax (Bo−1 ) S t i
i=1 [ρi tr(P0|0 ) +
is proposed to reduce communication costs. Liang et al.[93] ηi
], where Bo = 1 1
col(Bo , Bo , · · · , Bo ), ηi
S
=
ρi
presented a multi-rate filtering problem for a class of net- tr(Aio (Coi )−1 Ri (Coi )−1 (Aio ) )P ro(γti = 1) + tr(Boi Q(Boi ) ),
worked multi-sensor fusion systems with packet dropouts. ρi = (1−P ro(γti = 1))αi , αi = σmax 2
(Aio ) for i = 1, 2, · · · , S,
Different from the single-rate estimator design with packet Bo and Ao are the matrices related to the sensors, Q and
i i
dropouts which are treated as stochastic parameters, an un- Ri are the covariance matrices, σmax is the largest singular
known input observer was proposed, where packet dropouts value and γt is the transmission outcome of the sensor.
are represented as zero-mean white input noises of the lin- Up to now, filtering for NCSs with single channels has
ear time-variant estimation error system. Yan et al.[94] con- been researched widely. However, data fusion is still a novel
sidered the state estimation for a kind of nonlinear multi- research branch. The existing approaches mainly involve
rate multisensor asynchronous sampling dynamic system. the linear systems, in which some priori statistical charac-
N sensors observed a single target independently at multi- teristics of measurement are given and mostly the distur-
ple sampling rates, and the dynamic system was formulated bances in different sensors are assumed to be uncorrelated.
at the highest sampling rate. The multiscale system the- In the future, the optimal fusion method for nonlinear sys-
ory and the modified sigma point KF were used to design tem needs to be designed in the presence of parameter un-
the fusion method. Measurements randomly missing with certainty. Moreover, the correlation between the distur-
Bernoulli distribution could also be allowed in [94]. bances should also be taken into account.

4 Fault diagnosis of NCSs

In this section, we mainly summarize the progress on

fault diagnosis of NCSs. Essentially, fault diagnosis is the
process in which the available information in the diagno-
sis system is processed to generate the running state and
the fault condition of the system. Different from the tradi-
tional fault diagnosis, the fault diagnosis of NCSs has the
Fig. 4 An example of multiple sampling rate following characters:
1) Because the data transmitting in the network channel
There are also some significant results which are con- is digital quantity, the model of the system is of the discrete
cerned with the communication constraints and energy form.
consumption. In [95], the realistic sensing/reporting chan- 2) The transmission delay of the system is non-ignorable.
nels and the correlated log-normal shadow-fading in all 3) Since there are different node drive forms and different
wireless links of the cooperative network were considered. data transmission methods in the network, the correspond-
It was shown that shadowing correlation on the report- ing models of the system have different forms.
ing channel can yield performance degradations similar to 4) The model of the controlled object is always nonlinear.
the shadowing correlation on the sensing channel. In [96], 5) The packet dropout in NCSs brings a negative effect
the problem of optimal state estimation was concerned to the fault diagnosis.
 Y. Q. Xia et al. / Recent Progress in Networked Control Systems – A Survey 353

Some recent progress about fault diagnosis will be shown sidered as external disturbances. In [116], fault detection
in the following. For the fault diagnosis system, two neg- for Takagi-Sugeno fuzzy discrete systems in finite-frequency
ative factors need to be considered: the uncertainty or domain was studied. Faults were considered in a middle
the disturbance and the known input induced by the de- frequency domain, while disturbances were considered in
lay. In [99], the influence of network-induced delay on con- another certain finite-frequency domain interval. The filter
ventional observer based fault detection systems was first design was formulated as a two-objective optimization al-
investigated. However, this method cannot meet the re- gorithm. Liu and Xia[117] presented a fault detection and
quirements for effective residual information. To solve this compensation scheme based on likelihood ratios (LRs) for
problem, Zhang et al.[100] presented that if the unknown networked predictive control systems with random network-
network-induced delay is less than the sampling period, an induced time delays and clock asynchronism. The compen-
NCS with disturbance and fault can be regarded as the NCS sator eliminated the effect of the random delay. In order
without delay. Moreover, the effect of the network-induced to reduce the influence of the asynchronism, the observer
delay can be seen as an unknown input. The fault diagno- updated based on the time schedule of a remote node clock.
sis system was robust to this input. In [101], a relationship Two frameworks were designed to update the LRs of fault.
between the parity relation and the H2 optimal residual The convergence analysis of a generalized LR test with in-
generators was established, and the optimal parity vector termittent observations was given. Xia and Liu[118] con-
H2 converged to the H2 -optimal post-filter with s → ∞. sidered the detection of abrupt changes for MIMO linear
The effect of the unknown input was decreased. When the systems based on frequency domain data. The measured
delay is larger than the sampling period, the traditional inputs and outputs were maped from the time domain to
method is not effective. For this problem, in [102], the prob- the frequency domain by using the real discrete-time Fourier
lem of robust fault estimation for a class of uncertain NCSs transform. The latent state change was modeled as an ini-
with random communication network-induced delays was tial state disturbance to be estimated on the basis of fre-
investigated. Based on the Lyapunov-Razumikhin method, quency domain samples. In this work, the occurrence time
Huang and Nguang[102] gave a delay-dependent fault esti- was estimated by maximizing a likelihood ratio function.
mator in terms of the solvability of bilinear matrix inequal- In [119], the detection of abrupt change was used in elec-
ities (BMIs) and proposed an iterative algorithm to change trocardiogram. An linear time varying (LTV) model with
this non-convex BMI problem into quasi-convex optimiza- Gaussian white noise was used to describe the real electro-
tion problems. In this work, delay was not required to be cardiogram signal. Based on the estimated system param-
less than the sampling period, but bounded. eters and tuned covariances of noise, the generalized like-
In order to rationally utilize network resources and re- lihood ratio (GLR) tests for electrocardiogram signal were
duce the network congestion and delay, Xie et al.[103] con- developed for abrupt change detection.
structed a simple information-scheduling scheme to describe To sum up, there are relatively fewer results for fault
the scheduling approach of system signals in NCSs. For the diagnosis of nonlinear NCSs. How to extend the achieve-
sake of the optimization, the scheme made the NCSs share ments for linear systems to nonlinear systems, especially
the network resource periodically. Considering the short considering the network-induced delay and the uncertainty
delay in the network, Ye et al.[99, 101, 104−106] utilized the simultaneously, is still a difficulty at the current stage. In
state estimation and the observer theory to reduce the ef- addition, the affects of network protocol for the fault diag-
fect of the delay. A lowpass filter and the Cayley-Hamilton nosis should also be emphasized in the future.
based approach were proposed in these works. Huo and
Fang[107−110] studied the fault diagnosis systems with ran- 5 Networked predictive control
dom delays and communication constraints. In addition, for
the NCSs with large delay, Fang et al.[111−113] used quasi Due to the existence of the network, the traditional con-
Takagi-Sugeno fuzzy model to design the fault diagnosis ap- trol strategies are not suitable for the novel networked
proach. These works not only considered the influence of control systems. Numerous control approaches for NCSs
the network-induced delay, but also investigated the effect have attracted the interests of the researchers. These ap-
of the packet loss. Furthermore, Mao et al.[114] studied the proaches aim to deal with the time delay and the packet
problem of protocol and fault detection for nonlinear NCSs. dropout. Chou and Tipsuwan[120] divided the recent net-
In addition, Liu et al.[115] researched the fault detec- worked control techniques into five classes (augmented de-
tion of linear systems over networks with bounded packet terministic discrete-time model method, queuing method,
loss. The packet loss process is assumed to be arbitrary optimal stochastic control method, perturbation method
or Markovian. For a bounded packet loss rate, a switched and sampling time scheduling method). In the following,
system was constructed to be the monitored system by re- we will present the main results about networked predic-
sampling it at each time instant when the measurements tive control (NPC). First, MPC for NCSs is summarized
arrive at the fault detection node. This work designed a fil- and the framework to deal with delay and dropout using
ter for the switched system, which updates only at the time MPC is constructed. Then, the predictor, based on a se-
instant when new measurements arrived at the fault detec- quence of past measurements, is called data package based
tion node. Similarly, the input data packets lost were con- predictor here. The corresponding predictive method is
354 International Journal of Automation and Computing 12(4), August 2015

called data package based predictive control and the re- for all l = 0, · · · , N − 1
sults of this method are shown. Thirdly, when the sys-  
tem is extremely complex, the model of the system cannot x(k + l + 1|k) = f x(k + l|k), u(x + l|k)
be obtained. What can be used are only the input and
output. Consequently, the data-driven predictive control is x(k + l|k) ∈ X
proposed. Up to now, there are few achievements about this u(k + l|k) ∈ U
aspect. The main idea is presented in this survey. With the x(k + N |k) ∈ Ω
increase of the complexity of the task, distributed coopera-
tive control for NCSs is studied and a considerable number where X is admissible state set, U is the admissible control
of distributed control techniques were studied in the past set and Ω is the terminal set.
years. Due to the space limit, in the end of this section, In general, the cost function has a quadric form, i.e.,
we only discuss some important results of the distributed L(x(k + l|k), u(k + l|k)) = x(k + 1|k)2Q + u(k + l|k)2R
predictive control for NCSs. and F (x(k + N |k)) = x(k + N |k)2P . In order to ensure
the stability and the feasibility, Q, R and P need to satisfy
5.1 Model predictive control for NCSs some conditions. More details about them can be found in
[121, 122].
MPC, which is also called receding horizon control
For NCSs, MPC has been used as a significant tool to deal
(RHC), is a form of control in which the current control
with the negative effect which the delay and the dropout in-
action is obtained by solving on-line, at each sampling in-
duce. In the following, we will summarize MPC for NCSs
stant, a finite horizon open-loop optimal control problem,
without constraints and with constraints, respectively.
using the current state of the plant as the initial state. The
When only considering the delay and the dropout, many
optimization yields an optimal control sequence and the
works about MPC can be found. Irwin et al.[123] presented a
first control in this sequence is applied to the plant[121] .
co-design approach to wireless network control systems and
For the past decades, it has gained notable attention in
Lyapunov stability was derived using LMIs for the multi-
control of dynamic systems and has played an important
observer linear quadratic (LQ) control. Liu et al.[124] pro-
role in control practice. One of the significant advantages
posed a modified MPC, which uses the future control se-
for MPC is the capability of handling the constraints. Due
quence to compensate for the forward communication time
to the existence of the constraints, the traditional determi-
delay, to deal with networked control systems with ran-
nate controller may result in constraint violations. How-
dom communication time delay. For both fixed and ran-
ever, in the MPC framework, the constraints are regarded
dom communication time delays, if there exists P such that
as one part of the optimization problem and can be handled
ΛT (τ )P Λ(τ ) < P where Λ(τ ) is the closed-loop system
properly. Another merit of MPC is the optimality of the
matrix, the closed-loop is stable. In [125], the predictive
solution in some meaning. The objective function of MPC
control and the integral control were used to compensate
often represents the cost of the state, the control input and
for the network link effects and eliminate the tracking er-
other items with different senses. In most cases, the analyt-
ror. Zhao et al.[126, 127] designed an improved predictive
ical solution cannot be provided and only the local optimal
controller using delayed sensing data and proposed a com-
numerical solution can be gained.
pensation scheme to overcome the negative effects of the
Consider a nonlinear discrete-time system
network-induced delays and data packet dropouts in both
  forward and backward channels.
x(k + 1) = f x(k), u(k) . (39) In addition, some achievements for NCSs with con-
straints can be found. In [128], the input and the out-
MPC can be formulated as put were constrained by the nonlinear formulations. MPC
based on the mixed-integer programming was used to de-
rive the controller. It turned out that the controller design
J ∗ (k, x∗ , u∗ ) = min J(k, x, u) (40)
u(·|k) was feasible due to the special nature of the data transmis-
sion strategy, only a small number of logic values was in-
with volved. In order to minimize bandwidth utilization, Good-
win et al.[129] imposed a communication constraint which
N−1   restricted all transmitted data to belong to a finite set and
J(k, x, u) = L(x(k + l|k), u(k + l|k)) + only permited one plant to be addressed at a time. The
l=0 moving horizon approach was used to deal with both con-
 
F x(k + N |k) (41) trol and measurement quantization issues. Varutti et al.[130]
proposed an event-based model predictive control approach
to nonlinear continuous time systems under state and in-
subject to
put constraints. The method can counteract bounded de-
lays, information losses, as well as deal with event trigger-
x(k|k) = x(k) ing due to sensors and actuators. Quevedo and Nešić[131]
 Y. Q. Xia et al. / Recent Progress in Networked Control Systems – A Survey 355

studied an MPC scheme in which packets containing opti-

mizing sequences of control inputs were sent over an unre-
liable communication network with packet loss. MPC for
nonlinear systems subject to packet losses was proposed
in [132]. The proposed controller allowed for an explicit
characterization of the stability region and it turned out
that this region was an invariant set for the closed-loop
system with packet losses. In [133], a robust MPC with
a network delay compensation strategy was proposed for
uncertain discrete-time constrained nonlinear systems with
time-varying transmission delays and packet dropouts. Xia
et al.[134] considered robust constrained MPC of systems
with polytopic uncertainties. Sufficient conditions for the
existence of parameter-dependent Lyapunov functions were
proposed in terms of LMIs, which would reduce the con-
servativeness resulted from using a single Lyapunov func- Fig. 5 Date packing based predictive control for NCSs
tion. The combination of event-triggering control and MPC
was also used for NCSs[135−138] . A novel event-triggered where N1 is the finite KF gain[154] . The processed states
strategies was proposed in [137] for the control of uncertain are then sent to the controller node.
nonlinear systems with additive disturbances under robust In general, there inevitably exist the unknown network
MPC. The main function of the event-triggered scheme was transmission delays in both the forward (from controller
to trigger the solution of the optimal control problem of the to actuator) channel (CAC) and feedback (from sensor to
robust MPC. controller) channel (SCC). A networked predictive control
Other results can be found in [139−152]. scheme mainly consists of the control prediction generator
and the network delay compensator. The former is designed
5.2 Data package based predictive control here to generate a set of future control predictions, while
for NCSs the later is used to compensate for the unknown random
network delays. If the network can transmit a set of data
The fundamental idea and the basic framework of the at the same time, it is assumed that a predictive control
data package based predictive control will be shown firstly. sequence at time k is packed and then sent to the plant side
The discrete dynamic system is considered as through a network. In the compensator node, the latest
control value from these control sequences will be chosen
  on the plant side.
x(k + 1) = f x(k), u(k), w(k) (42) For example, when the time delays in SCC and in CAC
  are respectively 0 and ki , given that the following control
y(k) = g x(k), u(k), v(k) (43)
predictive sequences are received on the plant side:

[uT T T T
t−k1 |t−k1 , ut−k1 +1|t−k1 , · · · , ut|t−k1 , · · · , ut+N−k1 |t−k1 ]
T

where x(k) is the system state, u(k) is the system input,

y(k) is the system output, f (·) and g(·) are the system equa- [uT T T T
t−k2 |t−k2 , ut−k2 +1|t−k2 , · · · , ut|t−k2 , · · · , ut+N−k2 |t−k2 ]
T

tions, which can be linear or nonlinear, w(k) is the unknown ···

process disturbance, and v(k) is the unknown measurement [uT T
, uT T T
t−kt |t−kt , ut−kt +1|t−kt , · · · t|t−kt , · · · , ut+N−kt |t−kt ]
noise. (47)
Package based predictive control has been proved to be
an effective method for networked control systems with where the control sequences ut|t−ki , i = 1, 2, · · · , t, are
network-induced time delay and data dropouts[153, 154] . The available to be chosen as the control input of the plant at
framework of package based predictive control system is time t. The output of the network delay compensator, i.e.,
shown in Fig. 5. It s worth noting that, in order to satisfy the input of the actuator will be
that the measurements are processed in sequence, a buffer
is set at the controller node. The estimated states and pre- ut = ut|t−min{k1 ,k2 ,··· ,kt } . (48)
dictive states can be obtained by using KKF:
In fact, by using the networked predictive control scheme
presented in this section, the control performance of the
 
closed-loop system with network delay is very similar to
x̂(k|k) = KF S, û(k − 1|k − 1), y(k) (44)
  that of the closed-loop system without network delay.
x̂(k + i|k) = KF S, û(k|k), y(k) , i = 1, 2, · · · , N1 (45) Next, we will show some results about package based pre-
dictive control[155−171] . Xia et al.[155] studied the stochastic
û(k + i|k) = K(k + i)x̂(k + i|k), i = 1, 2, · · · , N1 (46) stability analysis of networked control systems with random
356 International Journal of Automation and Computing 12(4), August 2015

network delay. A scheme, including control prediction gen- it is obvious that this framework is different from the data
erator and network-delay compensator, was proposed. The package based predictive method in [155].
control prediction generator, which makes the closed-loop
system achieve the desired control performance, was used to
generate a set of future control predictions and the network-
delay compensator removed the effects of the network trans-
mission delay and data dropout. In [161], random network
delay in the feedback channel was considered and stabil-
ity criteria of closed-loop networked predictive control sys-
tems was given. Liu et al.[167] considered both constant and
random network delays in the forward and feedback chan-
nels. A networked predictive control scheme was proposed
to overcome the effects of network delay and data dropout.
The necessary and sufficient conditions for the stability of
closed-loop NCS with constant time delay were given. Xia
et al.[170] proposed the switched system s method, which
solves the stochastic problem in a deterministic way.

5.3 Data-driven predictive control for Fig. 6 Data driven predictive control for NCSs
NCSs
Up to now, there are some improvements of data-driven
According to traditional control design frameworks, con- predictive control[172−175] . Laurı́ et al.[173] proposed model-
trol and monitoring need support from dynamic models. based predictive control methodology in the space of the
In general, the state or the output measurements from the latent variables for continuous processes. Xia et al.[175]
plant are the most significant elements to build the system proposed data-driven predictive control for networked con-
model. When finishing the design of a controller or a mon- trol systems, which was designed by applying the sub-
itor, the model will no longer exist. However, while using space matrices technique and obtained directly from the
models in this framework, the modeling error and the com- input/output data transferred from networks. Implement-
plexity in building the model cannot be neglected. Specif- ing identification and control in the latent variable space is
ically, with Internet of things (IOT), big data needs to be easier than in the case of correlation in the data set. It can
processed urgently. And inevitably, the complexity of the act as a prefilter to reduce the effect of noisy data, and can
traditional control frameworks will continue to increase. To finally reduce computational complexity.
address these issues, the data-driven scheme was developed, 5.4 Networked distributed predictive con-
which has been applied in the industrial process control field
trol
as well as in complex systems. Further more, if the digital
data flows through the network, which connects the con- In this subsection, we mainly emphasize the distributed
troller and the actuator, then the data-driven method is model predictive control. The considerable increase of the
proper for NCSs, especially when the system model is hard computational power of the hardware and the rapid de-
to build. velopment of the communication technology bring novel
In the following, the notions of data-driven predictive opportunity to the decentralized or distributed MPC
networked control system are introduced. The typical struc- (DMPC)[176−185] . Maestre and Negenborn[186] mainly in-
ture of data-driven predictive networked control system is troduced a comprehensive overview distributed MPC and
shown in Fig. 6. In such a control system, the predictive contributed to answering the question what kind of dis-
control signals are generated by the subspace projection tributed model-based control technique could be appropri-
method. And data-driven predictive networked control sys- ate for control of large-scale intermodal networks, or parts
tems and data package based predictive control systems dif- thereof. Negenborn and Maestre[187] presented the main
fer only in the controller. In data-driven networked control features of distributed MPC in three aspects (process fea-
systems, when the past input and past output from the tures, control architecture features and theoretical features)
sensors are transmitted to the controller, the data-driven and provided the future development of DMPC. Event-
predictive control algorithm will be used to generate a se- based DMPC was explored in [188, 189] and it was ap-
quence of predictive control inputs. And then those control plied to a multi-agent system. In general, for DMPC, there
sequences will be transmitted to a buffer at the actuator are three update approaches: synchronous DMPC, sequen-
site over the network. At last, according to the predic- tial DMPC and iterative DMPC. In synchronous DMPC
tive networked control scheme described in (47) and (48), scheme, all the agents solve the optimization problem at
the compensator will select the proper control input. From the same time[184, 185, 190−192] . In the sequential DMPC
the description above, the data driven networked control framework, all the agents solve the optimization problem in
scheme can be obtained directly without any model. And, the specified order[180−182] . The advantage of the sequen-
 Y. Q. Xia et al. / Recent Progress in Networked Control Systems – A Survey 357

tial scheme is that less communication and computation approach to NCSs. Currently, most existing researches de-
is needed, as compared with the synchronous one. In the pend on the system model, especially the linear one. When
iterative DMPC, each agent solves its optimization prob- using the nonlinear predictive control, the stability of the
lems and communicates the solution with its neighbors it- NCSs is still a challenge. Data-based control is a novel
eratively until the local convergence of the solution. research direction, which can decrease the dependence on
The stability and feasibility of the DMPC are the sig- the accurate model. The research for data-driven predictive
nificantly challenging issues. When the goal region of the control is at its initial stage and there are many theoretical
system is a set of certain setpoints, many theoretical and problems to be handled. In addition, distributed predictive
applied results were proposed[182−184, 190−192] . Richards schemes have been adopted in NCSs, such as the multi-
and How[182] proposed a formulation of robust DMPC for agent systems. However, when the delay and the packet
the systems with coupled constraints and unknown but loss are considered, many novel approaches need to be ex-
bounded additive disturbances. However, if the set to con- plored.
verge is a dynamic region such as consensus or synchro-
nization and tracking, there are some issues in DMPC
framework. In [183], a general DMPC framework for
cooperative control was considered and an algorithm to 6 Further development of networked
test the satisfaction of the terminal region was formu- control systems—cloud control sys-
lated. However, all the subsystems share a public termi-
nal region and the terminal region is possibly conserva-
tems
tive. Wang and Ding[184] proposed a positively invariant
time-varying terminal state region with the corresponding
auxiliary controller to achieve the multi-vehicle formation Nowadays, due to its extremely powerful computation
control. Dunbar and Murray[190, 191] divided the central- power and storage capacity, the cloud technique has been
ized MPC (CMPC) scheme into DMPC one. Then, lower one key tool in the industry. This also brings a new oppor-
bounds on the update rate and upper bounds on the mis- tunity to the control technique. Cloud control systems have
match between actual and assumed trajectories were estab- become one of the most promising directions[201] . As the
lished to achieve multi-agent formation stabilization. increase of the system scale, huge requirements for com-
Considering the communication delay and the noise, putation and communication cannot be ignored. In gen-
there are only a few works[193−195] . Li and Shi[194] studied eral, most of complex systems can not be controlled prop-
the robust DMPC problem for a group of nonlinear agents erly in the absence of powerful tools and adequate system
(subsystems) subject to control input constraints and ex- information. However, the development of new technolo-
ternal disturbances. A robustness constraint was designed gies, including magical innovations in software and hard-
to deal with external disturbances. The feasibility of the ware, provides a necessary platform for computability. To
robust DMPC scheme and the robust stability of the over- some extent, cloud computing has come into our life. In
all agent system were analyzed, respectively. Li and Shi[195] other word, cloud computing is not only the original prod-
proposed the robust DMPC for large-scale continuous-time uct concept but also a kind of service. In general, it is a
nonlinear systems subject to communication delays and ex- byproduct of the ease-of-access to remote computing sites
ternal disturbances. They proved that 1) the bounds of provided by the Internet. A new supplement, consumption,
external disturbances, the sampling period and the bound and delivery model for IT services with the foundation of
of communication delays all affect the feasibility; 2) the sta- Internet protocols was described in [202] . It often includes
bility is related to the bounds of external disturbances, the the provisioning of dynamically scalable and the virtualized
sampling period, the bound of communication delays and resources[203, 204] . In practical systems, the cloud comput-
the minimum eigenvalues of the cooperation matrices; 3) ing system offers a shared pool of configurable resources,
the closed-loop system is stabilized into a robust invariant which include computation, software, data access, storage
set under the proposed conditions. services, etc. Therefore, even if without the physical loca-
In addition to distributed networked predictive control, tion and configuration of the service provider[205] , the end-
other distributed networked control approaches were also users can consume power automatically. This is to say, no
proposed for the cooperation of multiple NCSs. Take the matter the computers or other devices in such a system just
application of multiple rigid spacecrafts as an example. The play an utility role over the network, we can share resources,
compound control methodology for flight vehicles was pro- software and information, etc. For example, even if the pro-
posed in [196]. Zhou et al.[197−200] mainly studied the syn- grams are installed locally on others computers, the users
chronization and tracking control of multiple rigid space- can access and use them through a web browser[206] .
craft with uncertainty and disturbance. Finite-time termi- The cloud control systems merge the merits of cloud com-
nal sliding mode technique was adopted in these works. Due puting, advanced theory of NCSs and other recent devel-
to the limitation of space, other distributed approaches will oped related results together. To some extent, it will exceed
not be presented here. our expectation and have an extremely promising perspec-
In summary, networked predictive control is an important tive. The scheme of cloud control systems is shown in Fig. 7.
358 International Journal of Automation and Computing 12(4), August 2015

given. However, there are few theoretical results for the

cloud control systems. To sum up, as the vital technique of
IOT, there are still fruitful results about NCSs that need
to be investigated in the future.

References
[1] A. Onat, T. Naskali, E. Parlakay, O. Mutluer. Control over
imperfect networks: Model-based predictive networked con-
trol systems. IEEE Transactions on Industrial Electronics,
vol. 58, no. 3, pp. 905–913, 2011.
Fig. 7 Cloud control systems

In a large-scale control system, the captured infor- [2] Y. Ge, Q. G. Chen, M. Jiang, Y. Q. Huang. Modeling of
mation consists of information captured by ubiquitous random delays in networked control systems. Journal of
information-sensing mobile devices, aerial sensory technolo- Control Science and Engineering, vol. 2013, Article number
383415, 2013.
gies (remote sensing), software logs, cameras, microphones,
radio-frequency identification readers, and wireless sensor
networks[207] . That is the origin of “Big data”[208] , which [3] L. X. Zhang, H. J. Gao, O. Kaynak. Network-induced
constraints in networked control systems—A survey.
is a collection of data sets. However, the set is so large and IEEE Transactions on Industrial Informatics, vol. 9, no. 1,
complex that on-hand traditional database management or pp. 403–416, 2013.
computation tools cannot process it. As we know, the law of
large numbers proposes that some useful deterministic con- [4] D. F. Delchamps. Stabilizing a linear system with quantized
clusions can be obtained with the help of sufficient data. state feedback. IEEE Transactions on Automatic Control,
In addition, big data is confronted with a quite number vol. 35, no. 8, pp. 916–924, 1990.
of challenges including capturing, storage, search, sharing,
transfer, analysis, visualization, etc. In the cloud control [5] L. Bao, M. Skoglund, K. H. Johansson. Encoder-decoder de-
systems, big data will be sent to the centers of cloud com- sign for event-triggered feedback control over bandlimited
puting to be processed firstly. After that, control signal, channels. In Proceedings of the American Control Confer-
ence, IEEE, Minneapolis, USA, pp. 4183–4188, 2006.
such as scheduling schemes, predictive control sequences
and other useful information will be generated instantly
for cloud control systems. To sum up, cloud control sys- [6] S. L. Hu, D. Yue. Event-triggered control design of linear
networked systems with quantizations. ISA Transactions,
tems will provide us powerful tools to control the complex
vol. 51, no. 1, pp. 153–162, 2012.
system, which we can not image before. More information
about cloud control can be seen in [209].
[7] E. Garcia, P. J. Antsaklis. Model-based event-triggered con-
trol for systems with quantization and time-varying network
7 Conclusions delays. IEEE Transactions on Automatic Control, vol. 58,
no. 2, pp. 422–434, 2013.
The paper surveys some resent progresses in different
components of NCSs. Quantization in NCSs, especially log-
[8] L. C. Li, X. F. Wang, M. Lemmon. Stabilizing bit-rates
arithmic and uniform quantizers, is discussed. Estimation in quantized event triggered control systems. In Proceed-
is very important to systems with uncertainty. As an useful ings of the 15th ACM International Conference on Hybrid
estimation tool, filtering, in which delay and drop-out are Systems: Computation and Control, ACM, Beijing, China,
considered, is presented for linear and nonlinear systems, pp. 245–254, 2012.
respectively. Data fusion in NCSs is significantly related
to filtering. In this paper, both centralized and distributed [9] N. Elia, S. K. Mitter. Stabilization of linear systems with
data fusions are summarized in detail. Particularly, con- limited information. IEEE Transactions on Automatic Con-
trol, vol. 46, no. 9, pp. 1384–1400, 2001.
sidering time delay and packet loss, fusion for NCSs is also
discussed. In addition, some results for fault diagnosis are
given. Specially, the uncertainty and the known input in- [10] M. Y. Fu, L. H. Xie. The sector bound approach to quan-
tized feedback control. IEEE Transactions on Automatic
duced by the delay are taken into consideration. As an Control, vol. 50, no. 11, pp. 1698–1711, 2005.
important control technique, predictive control is widely
used for NCSs. In this paper, MPC, data packaged based
[11] H. Ishii, T. Başar. Remote control of LTI systems over net-
predictive control and data-driven predictive control are de- works with state quantization. Systems & Control Letters,
tailedly discussed. Some progresses in distributed MPC for vol. 54, no. 1, pp. 15–31, 2005.
NCSs are also presented. In the end, cloud control systems
are introduced to deal with the system, in which there ex- [12] T. Hayakawa, H. Ishii, K. Tsumura. Adaptive quantized
ist big data and the requirement for rapid computation, control for linear uncertain discrete-time systems. Automat-
and some basic concepts of the cloud control systems are ica, vol. 45, no. 3, pp. 692–700, 2009.
 Y. Q. Xia et al. / Recent Progress in Networked Control Systems – A Survey 359

[13] T. Hayakawa, H. Ishii, K. Tsumura. Adaptive quantized [26] D. Liberzon, D. Nesic. Input-to-state stabilization of linear
control for nonlinear uncertain systems. Systems & Control systems with quantized state measurements. IEEE Trans-
Letters, vol. 58, no. 9, pp. 625–632, 2009. actions on Automatic Control, vol. 52, no. 5, pp. 767–781,
2007.

[14] K. Y. You, W. Z. Su, M. Y. Fu, L. H. Xie. Attainabil-

ity of the minimum data rate for stabilization of linear [27] D. Nešić, D. Liberzon. A unified framework for design and
systems via logarithmic quantization. Automatica, vol. 47, analysis of networked and quantized control systems. IEEE
no. 1, pp. 170–176, 2011. Transactions on Automatic Control, vol. 54, no. 4, pp. 732–
747, 2009.
[15] B. Zhou, G. R. Duan, J. Lam. On the absolute stability ap-
proach to quantized feedback control. Automatica, vol. 46,
[28] K. Y. You, L. H. Xie. Minimum data rate for mean square
no. 2, pp. 337–346, 2010.
stabilization of discrete LTI systems over lossy channels.
IEEE Transactions on Automatic Control, vol. 55, no. 10,
[16] D. Yue, C. Peng, G. Y. Tang. Guaranteed cost control of pp. 2373–2378, 2010.
linear systems over networks with state and input quantisa-
tions. IET Control Theory & Applications, vol. 153, no. 6,
pp. 658–664, 2006. [29] K. Y. You, L. H. Xie. Minimum data rate for mean square
stabilizability of linear systems with Markovian packet
losses. IEEE Transactions on Automatic Control, vol. 56,
[17] C. Z. Zhang, G. Feng, H. J. Gao, J. B. Qiu. Generalized H2 no. 4, pp. 772–785, 2011.
filter design for T-S fuzzy systems with quantization and
packet loss. In Proceedings of IEEE Symposium on Com-
putational Intelligence in Control and Automation, IEEE, [30] E. Tian, D. Yue, X. Zhao. Quantised control design for
Paris, France, pp. 52–59, 2011. networked control systems. IET Control Theory & Appli-
cations, vol. 1, no. 6, pp. 1693–1699, 2007.

[18] C. Z. Zhang, G. Feng, H. J. Gao, J. B. Qiu. H∞ filtering

for nonlinear discrete-time systems subject to quantization [31] C. De Persis. On stabilization of nonlinear systems under
and packet dropouts. IEEE Transactions on Fuzzy Systems, data rate constraints using output measurements. Interna-
vol. 19, no. 2, pp. 353–365, 2011. tional Journal of Robust and Nonlinear Control, vol. 16,
no. 6, pp. 315–332, 2006.
[19] L. Li, Y. Q. Xia, J. Q. Qiu, H. J. Yang. Robust H∞
networked control for discrete-time fuzzy systems with
[32] C. De Persis. Robust stabilization of nonlinear systems by
state quantisation. International Journal of Systems Sci-
quantized and ternary control. Systems & Control Letters,
ence, vol. 43, no. 12, pp. 2249–2260, 2012.
vol. 58, no. 8, pp. 602–608, 2009.

[20] J. J. Yan, Y. Q. Xia. Stabilisation of non-linear continuous

system with input quantisation and packet dropout. IET [33] J. J. Yan, Y. Q. Xia, B. Liu, M. Y. Fu. Stabilisation of
Control Theory & Applications, vol. 6, no. 15, pp. 2426– quantised linear systems with packet dropout. IET Control
2433, 2012. Theory & Applications, vol. 5, no. 8, pp. 982–989, 2011.

[21] Y. Q. Xia, J. J. Yan, P. Shi, M. Y. Fu. Stability analy- [34] J. J. Yan, Y. Q. Xia, L. Li. Stabilization of fuzzy sys-
sis of discrete-time systems with quantized feedback and tems with quantization and packet dropout. International
measurements. IEEE Transactions on Industrial Informat- Journal of Robust and Nonlinear Control, vol. 24, no. 10,
ics, vol. 9, no. 1, pp. 313–324, 2013. pp. 1563–1583, 2014.

[22] R. W. Brockett, D. Liberzon. Quantized feedback stabiliza- [35] Y. Q. Xia, J. J. Yan, J. Z. Shang, M. Y. Fu, B. Liu. Stabi-
tion of linear systems. IEEE Transactions on Automatic lization of quantized systems based on Kalman filter. Con-
Control, vol. 45, no. 7, pp. 1279–1289, 2000. trol Engineering Practice, vol. 20, no. 10, pp. 954–962, 2012.

[23] D. Liberzon. Hybrid feedback stabilization of systems with

[36] Y. Sharon, D. Liberzon. Input to state stabilizing controller
quantized signals. Automatica, vol. 39, no. 9, pp. 1543–1554,
for systems with coarse quantization. IEEE Transactions on
2003.
Automatic Control, vol. 57, no. 4, pp. 830–844, 2012.

[24] F. Fagnani, S. Zampieri. Quantized stabilization of linear

systems: Complexity versus performance. IEEE Transac- [37] R. E. Kalman. A new approach to linear filtering and pre-
tions on Automatic Control, vol. 49, no. 9, pp. 1534–1548, diction problems. Journal of Fluids Engineering, vol. 82,
2004. no. 1, pp. 35–45, 1960.

[25] F. Bullo, D. Liberzon. Quantized control via locational opti- [38] K. Y. You, L. H. Xie. Kalman filtering with scheduled mea-
mization. IEEE Transactions on Automatic Control, vol. 51, surements. IEEE Transactions on Signal Processing, vol. 61,
no. 1, pp. 2–13, 2006. no. 6, pp. 1520–1530, 2013.
360 International Journal of Automation and Computing 12(4), August 2015

[39] M. Sahebsara, T. W. Chen, S. L. Shah. Optimal H2 fil- [51] D. Yue, Q. L. Han. Network-based robust H∞ filtering for
tering in networked control systems with multiple packet uncertain linear systems. IEEE Transactions on Signal Pro-
dropout. IEEE Transactions on Automatic Control, vol. 52, cessing, vol. 54, no. 11, pp. 4293–4301, 2006.
no. 8, pp. 1508–1513, 2007.
[52] Y. Q. Xia, J. Q. Han. Robust Kalman filtering for systems
under norm bounded uncertainties in all system matrices
[40] M. Sahebsara, T. W. Chen, S. L. Shah. Optimal H∞ fil-
tering in networked control systems with multiple packet and error covariance constraints. Journal of Systems Sci-
dropouts. Systems & Control Letters, vol. 57, no. 9, pp. 696– ence and Complexity, vol. 18, no. 4, pp. 439–445, 2005.
702, 2008.
[53] H. J. Yang, Y. Q. Xia, P. Shi, M. Y. Fu. A novel delta oper-
ator Kalman filter design and convergence analysis. IEEE
[41] W. A. Zhang, L. Yu, H. B. Song. H∞ filtering of networked Transactions on Circuits and Systems I: Regular Papers,
discrete-time systems with random packet losses. Informa- vol. 58, no. 10, pp. 2458–2468, 2011.
tion Sciences, vol. 179, no. 22, pp. 3944–3955, 2009.

[54] W. Xie, Y. Q. Xia. Data-driven method for Kalman fil-

[42] B. Sinopoli, L. Schenato, M. Franceschetti, K. Poolla, M. tering. In Proceedings of the 2nd International Conference
I. Jordan, S. S. Sastry. Kalman filtering with intermit- on Intelligent Control and Information Processing, IEEE,
tent observations. IEEE Transactions on Automatic Con- Harbin, China, pp. 830–835, 2011.
trol, vol. 49, no. 9, pp. 1453–1464, 2004.
[55] J. F. Wu, Q. S. Jia, K. H. Johansson, L. Shi. Event-based
sensor data scheduling: Trade-off between communication
[43] X. He, Z. D. Wang, D. Zhou. Robust H∞ filtering for net- rate and estimation quality. IEEE Transactions on Auto-
worked systems with multiple state delays. International matic Control, vol. 58, no. 4, pp. 1041–1046, 2013.
Journal of Control, vol. 80, no. 8, pp. 1217–1232, 2007.

[56] D. E. Quevedo, D. Nešić. Robust stability of packetized

[44] S. L. Hu, D. Yue. Event-based H∞ filtering for net- predictive control of nonlinear systems with disturbances
worked system with communication delay. Signal Process- and Markovian packet losses. Automatica, vol. 48, no. 8,
ing, vol. 92, no. 9, pp. 2029–2039, 2012. pp. 1803–1811, 2012.

[57] R. N. Yang, P. Shi, G. P. Liu. Filtering for discrete-time

[45] M. Moayedi, Y. K. Foo, Y. C. Soh. Adaptive Kalman fil-
networked nonlinear systems with mixed random delays and
tering in networked systems with random sensor delays,
packet dropouts. IEEE Transactions on Automatic Control,
multiple packet dropouts and missing measurements. IEEE
vol. 56, no. 11, pp. 2655–2660, 2011.
Transactions on Signal Processing, vol. 58, no. 3, pp. 1577–
1588, 2010.
[58] H. L. Dong, Z. D. Wang, H. J. Gao. Robust H∞ filter-
ing for a class of nonlinear networked systems with multi-
[46] L. Schenato. Optimal estimation in networked control sys- ple stochastic communication delays and packet dropouts.
tems subject to random delay and packet drop. IEEE Trans- IEEE Transactions on Signal Processing, vol. 58, no. 4,
actions on Automatic Control, vol. 53, no. 5, pp. 1311–1317, pp. 1957–1966, 2010.
2008.

[59] M. Hilairet, F. Auger, E. Berthelot. Speed and rotor flux es-

[47] M. Sahebsara, T. Chen, S. L. Shah. Optimal filtering with timation of induction machines using a two-stage extended
random sensor delay, multiple packet dropout and uncer- Kalman filter. Automatica, vol. 45, no. 8, pp. 1819–1827,
tain observations. International Journal of Control, vol. 80, 2009.
no. 2, pp. 292–301, 2007.
[60] H. W. Sorenson. Kalman Filtering: Theory and Applica-
tion, Piscataway, New Jersey, USA: IEEE, 1985.
[48] G. L. Wei, Z. D. Wang, X. He, H. S. Shu. Filtering for
networked stochastic time-delay systems with sector non-
linearity. IEEE Transactions on Circuits and Systems II: [61] J. J. LaViola. A comparison of unscented and extended
Express Briefs, vol. 56, no. 1, pp. 71–75, 2009. Kalman filtering for estimating quaternion motion. In Pro-
ceedings of the American Control Conference, IEEE, Col-
orado, USA, pp. 2435–2440, 2003.
[49] Y. Shi, H. Z. Fang. Kalman filter-based identification for
systems with randomly missing measurements in a net- [62] N. J. Gordon, D. J. Salmond, A. F. M. Smith. Novel ap-
work environment. International Journal of Control, vol. 83, proach to nonlinear/non-Gaussian Bayesian state estima-
no. 3, pp. 538–551, 2010. tion. IEE Proceedings F on Radar and Signal Processing,
vol. 140, no. 2, pp. 107–113, 1993.
[50] M. Moayedi, Y. K. Foo, Y. C. Soh. Filtering for networked
control systems with single/multiple measurement packets [63] S. Guadarrama, A. Ruiz-Mayor. Approximate robotic
subject to multiple-step measurement delays and multiple mapping from sonar data by modeling perceptions with
packet dropouts. International Journal of Systems Science, antonyms. Information Sciences, vol. 180, no. 21, pp. 4164–
vol. 42, no. 3, pp. 335–348, 2011. 4188, 2010.
 Y. Q. Xia et al. / Recent Progress in Networked Control Systems – A Survey 361

[64] H. P. Liu, F. C. Sun. Efficient visual tracking using particle [77] R. Lobbia, M. Kent. Data fusion of decentralized local
filter with incremental likelihood calculation. Information tracker outputs. IEEE Transactions on Aerospace and Elec-
Sciences, vol. 195, pp. 141–153, 2012. tronic Systems, vol. 30, no. 3, pp. 787–799, 1994.

[65] Y. Q. Xia, Z. H. Deng, L. Li, X. M. Geng. A new continuous- [78] Y. Bar-Shalom. On the track-to-track correlation problem.
discrete particle filter for continuous-discrete nonlinear sys- IEEE Transactions on Automatic Control, vol. 26, no. 2,
tems. Information Sciences, vol. 242, pp. 64–75, 2013. pp. 571–572, 1981.

[66] S. Kluge, K. Reif, M. Brokate. Stochastic stability of the ex- [79] S. Grime, H. F. Durrant-Whyte, P. Ho. Communication
tended Kalman filter with intermittent observations. IEEE in decentralized data-fusion systems. In Proceedings of
Transactions on Automatic Control, vol. 55, no. 2, pp. 514– the American Control Conference, IEEE, Chicago, USA,
518, 2010. pp. 3299–3303, 1992.

[67] L. Li, Y. Q. Xia. Stochastic stability of the unscented [80] R. Kumar, M. Wolenetz, B. Agarwalla, J. S. Shin, P. W.
Kalman filter with intermittent observations. Automatica, Hutto, A. Paul, U. Ramachandran. DFuse: A framework for
vol. 48, no. 5, pp. 978–981, 2012. distributed data fusion. In Proceedings of the 1st Interna-
tional Conference on Embedded Networked Sensor Systems,
ACM, New York, USA, pp. 114–125, 2003.
[68] L. Li, Y. Q. Xia. Unscented Kalman filter over unreliable
communication networks with Markovian packet dropouts.
IEEE Transactions on Automatic Control, vol. 58, no. 12, [81] M. S. Mahmoud, Y. Q. Xia. Networked Filtering and Fu-
pp. 3224–3230, 2013. sion in Wireless Sensor Networks, New York: CRC Press,
2014.

[69] R. A. Singer, A. J. Kanyuck. Computer control of multiple

site track correlation. Automatica, vol. 7, no. 4, pp. 455–463, [82] S. Challa, M. Palaniswami, A. Shilton. Distributed data fu-
1971. sion using support vector machines. In Proceedings of the
5th International Conference on Information Fusion, IEEE,
Annapolis, USA, pp. 881–885, 2002.
[70] D. Willner, C. B. Chang, K. P. Dunn. Kalman filter
algorithms for a multi-sensor system. In Proceedings of
IEEE Conference on Decision and Control including the [83] C. S. Regazzoni, A. Tesei. Distributed data fusion for real-
15th Symposium on Adaptive Processes, IEEE, Clearwa- time crowding estimation. Signal Processing, vol. 53, no. 1,
ter, USA, pp. 570–574, 1976. pp. 47–63, 1996.

[71] K. Salahshoor, M. Mosallaei, M. Bayat. Centralized and de- [84] M. E. Liggins, C. Y. Chong, I. Kadar, M. G. Alford, V.
centralized process and sensor fault monitoring using data Vannicola, S. Thomopoulos. Distributed fusion architec-
fusion based on adaptive extended Kalman filter algorithm. tures and algorithms for target tracking. Proceedings of the
Measurement, vol. 41, no. 10, pp. 1059–1076, 2008. IEEE, vol. 85, no. 1, pp. 95–107, 1997.

[72] M. Munz, M. Mählisch, K. Dietmayer. Generic centralized [85] E. Besada-Portas, J. A. Lopez-Orozco, J. Besada, J. M. De
multi sensor data fusion based on probabilistic sensor and la Cruz. Multisensor fusion for linear control systems with
environment models for driver assistance systems. IEEE asynchronous, out-of-sequence and erroneous data. Auto-
Intelligent Transportation Systems Magazine, vol. 2, no. 1, matica, vol. 47, no. 7, pp. 1399–1408, 2011.
pp. 6–17, 2010.

[86] X. J. Shen, E. B. Song, Y. M. Zhu, Y. T. Luo. Globally op-

[73] A. Polychronopoulos, U. Scheunert, F. Tango. Centralized timal distributed Kalman fusion with local out-of-sequence-
data fusion for obstacle and road borders tracking in a col- measurement updates. IEEE Transactions on Automatic
lision warning system. In Proceedings of the 7th Interna- Control, vol. 54, no. 8, pp. 1928–1934, 2009.
tional Conference on Information Fusion, 2004.

[87] B. Chen, W. A. Zhang, L. Yu. Distributed fusion estima-

[74] G. J. Bierman, M. R. Belzer. A decentralized square root in- tion with missing measurements, random transmission de-
formation filter/smoother. In Proceedings of the 24th IEEE lays and packet dropouts. IEEE Transactions on Automatic
Conference on Decision and Control, IEEE, Fort Laud- Control, vol. 59, no. 7, pp. 1961–1967, 2014.
erdale, USA, pp. 1902–1905, 1985.

[88] Y. Q. Xia, J. Z. Shang, J. Chen, G. P. Liu. Data fusion

[75] N. A. Carlson. Federated square root filter for decentral- over network. In Proceedings of the 27th Chinese Control
ized parallel processors. IEEE Transactions on Aerospace Conference, IEEE, Kunming, China, pp. 452–456, 2008.
and Electronic Systems, vol. 26, no. 3, pp. 517–525, 1990.

[89] Y. Q. Xia, J. Z. Shang, J. Chen, G. P. Liu. Networked

[76] D. A. Castanon, D. Teneketzis. Distributed estimation al- data fusion with packet losses and variable delays. IEEE
gorithms for nonlinear systems. IEEE Transactions on Au- Transactions on Systems, Man, and Cybernetics – Part B:
tomatic Control, vol. 30, no. 5, pp. 418–425, 1985. Cybernetics, vol. 39, no. 5, pp. 1107–1120, 2009.
362 International Journal of Automation and Computing 12(4), August 2015

[90] C. Zhu, Y. Q. Xia, L. P. Yan, M. Y. Fu. Multi-channel net- [103] L. B. Xie, H. J. Fang, Y. Zheng. Guaranteed cost control
worked data fusion with intermittent observations. In Pro- for networked control systems. Journal of Control Theory
ceedings of the 29th Chinese Control Conference, IEEE, and Applications, vol. 2, no. 2, pp. 143–148, 2004.
Beijing, China, pp. 4317–4322, 2010.

[91] C. Zhu, Y. Q. Xia, L. P. Yan, M. Y. Fu. Centralised fusion [104] H. Ye, Y. Q. Wang. Application of parity relation and sta-
over unreliable networks. International Journal of Control, tionary wavelet transform to fault detection of networked
vol. 85, no. 4, pp. 409–418, 2012. control systems. In Proceedings of the 1st IEEE Confer-
ence on Industrial Electronics and Application, IEEE, Sin-
gapore, pp. 1–6, 2006.
[92] W. A. Zhang, G. Feng, L. Yu. Multi-rate distributed fusion
estimation for sensor networks with packet losses. Automat-
ica, vol. 48, no. 9, pp. 2016–2028, 2012. [105] H. Ye, G. Z. Wang, S. X. Ding. A new parity space ap-
proach for fault detection based on stationary wavelet trans-
[93] Y. Liang, T. W. Chen, Q. Pan. Multi-rate stochastic H∞ form. IEEE Transactions on Automatic Control, vol. 49,
filtering for networked multi-sensor fusion. Automatica, no. 2, pp. 281–287, 2004.
vol. 46, no. 2, pp. 437–444, 2010.

[106] H. Ye, R. He, H. Liu, G. Z. Wang. A new approach for fault

[94] L. P. Yan, B. Xiao, Y. Q. Xia, M. Y. Fu. State estimation detection of networked control systems. In Proceedings of
for asynchronous multirate multisensor nonlinear dynamic the 14th IFAC Symposium on System Identification, IFAC,
systems with missing measurements. International Journal New castle, Australia, pp. 654–659, 2006.
of Adaptive Control and Signal Processing, vol. 26, no. 6,
pp. 516–529, 2012.
[107] Z. H. Huo, H. J. Fang. Robust H∞ filter design for net-
[95] M. Renzo, L. Imbriglio, F. Graziosi, F. Santucci. Dis- worked control system with random time delays. In Pro-
tributed data fusion over correlated log-normal sensing and ceedings of the 10th IEEE International Conference on En-
reporting channels: Application to cognitive radio net- gineering of Complex Computer Systems, IEEE, Shanghai,
works. IEEE Transactions on Wireless Communications, China, pp. 333–340, 2005.
vol. 8, no. 12, pp. 5813–5821, 2009.

[108] Z. H. Huo, H. J. Fang. Research on robust fault detection

[96] L. P. Yan, X. R. Li, Y. Q. Xia, M. Y. Fu. Optimal se-
for NCS with long delay based on LMI. International Jour-
quential and distributed fusion for state estimation in cross-
nal of Advances in Systems Science and Applications, vol. 5,
correlated noise. Automatica, vol. 49, no. 12, pp. 3607–3612,
no. 3, pp. 404–410, 2005.
2013.

[97] X. L. Bian, Y. Q. Xia, Z. H. Deng, M. Y. Fu. One- [109] Z. H. Huo, H. J. Fang. Fault-tolerant control research
channel networked data fusion with communication con- for networked control system under communication con-
straint. Journal of the Franklin Institute, vol. 351, no. 1, straints. Acta Automatica Sinica, vol. 32, no. 5, pp. 659–666,
pp. 156–173, 2014. 2006.

[98] X. L. Bian, Y. Q. Xia. Energy efficient data fusion over

wireless channels with power control. IET Signal Process- [110] Z. H. Huo, H. J. Fang. Fault-tolerant control of networked
ing, vol. 9, no. 3, pp. 206–217, 2015. control systems with random time-delays. Information and
Control, vol. 35, no. 5, pp. 584–587, 2006.

[99] H. Ye, S. X. Ding. Fault detection of networked control sys-

tems with network-induced delay. In Proceedings of the 8th [111] H. J. Fang, H. Zhang, Y. W. Fang, F. Yang. Quasi T-S
International Conference on Control, Automation, Robotics fuzzy models and stable controllers for networked control
and Vision, IEEE, Kunming, China, pp. 294–297, 2004. systems. In Proceedings of the 6th World Congress on In-
telligent Control and Automation, IEEE, Dalian, China,
[100] W. Zhang, M. S. Branicky, S. M. Phillips. Stability of net- pp. 220–223, 2006.
worked control systems. IEEE Control Systems Magazine,
vol. 21, no. 1, pp. 84–99, 2001.
[112] Y. Zheng, H. J. Fang, H. O. Wang. Takagi-sugeno fuzzy-
model-based fault detection for networked control sys-
[101] H. Ye, P. Zhang, S. X. Ding, G. Z. Wang. A time-frequency tems with Markov delays. IEEE Transactions on Systems,
domain fault detection approach based on parity relation Man, and Cybernetics – Part B: Cybernetics, vol. 36, no. 4,
and wavelet transform. In Proceedings of the 39th IEEE pp. 924–929, 2006.
Conference on Control and Decision, IEEE, Sydney, Aus-
tralia, pp. 4156–4161, 2000.
[113] H. J. Fang, F. Yang, Y. Zheng, H. Zhang. Fuzzy model-
[102] D. Huang, S. K. Nguang. Robust fault estimator design ing and fault detection for networked control systems. In
for uncertain networked control systems with random time Proceedings of the 6th IFAC Symposium on Fault Detec-
delays: An ILMI approach. Information Sciences, vol. 180, tion, Supervision and Safety of Technical Processes, IFAC,
no. 3, pp. 465–480, 2010. Beijing, China, pp. 1091–1096, 2006.
 Y. Q. Xia et al. / Recent Progress in Networked Control Systems – A Survey 363

[114] Z. H. Mao, B. Jiang, P. Shi. Protocol and fault detec- [126] Y. B. Zhao, G. P. Liu, D. Rees. Improved predictive con-
tion design for nonlinear networked control systems. IEEE trol approach to networked control systems. IET Control
Transactions on Circuits and Systems II: Express Briefs, Theory & Applications, vol. 2, no. 8, pp. 675–681, 2008.
vol. 56, no. 3, pp. 255–259, 2009.

[127] Y. B. Zhao, G. P. Liu, D. Rees. Integrated predictive con-

[115] B. Liu, Y. Q. Xia, Y. Yang, M. Y. Fu. Robust fault trol and scheduling co-design for networked control systems.
detection of linear systems over networks with bounded IET Control Theory & Applications, vol. 2, no. 1, pp. 7–15,
packet loss. Journal of the Franklin Institute, vol. 349, no. 7, 2008.
pp. 2480–2499, 2012.
[128] G. F. Zhang, X. Chen, T. W. Chen. A model predictive
control approach to networked systems. In Proceedings of
[116] H. J. Yang, Y. Q. Xia, B. Liu. Fault detection for TCS
the 46th IEEE Conference on Decision and Control, IEEE,
fuzzy discrete systems in finite-frequency domain. IEEE
New Orleans, USA, pp. 3339–3344, 2007.
Transactions on Systems, Man, and Cybernetics – Part B:
Cybernetics, vol. 41, no. 4, pp. 911–920, 2011.
[129] G. C. Goodwin, H. Haimovich, D. E. Quevedo, J. S. Welsh.
A moving horizon approach to networked control system
[117] B. Liu, Y. Q. Xia. Fault detection and compensation for design. IEEE Transactions on Automatic Control, vol. 49,
linear systems over networks with random delays and clock no. 9, pp. 1427–1445, 2004.
asynchronism. IEEE Transactions on Industrial Electronics,
vol. 58, no. 9, pp. 4396–4406, 2011.
[130] P. Varutti, B. Kern, T. Faulwasser, R. Findeisen. Event-
based model predictive control for networked control sys-
[118] Y. Q. Xia, B. Liu. Maximum likelihood ratio detection of tems. In Proceedings of the 28th Chinese Control Confer-
abrupt state change for MIMO linear systems based on fre- ence and Decision and the 48th IEEE Conference on Con-
quency domain data. International Journal of Robust and trol Conference, IEEE, Shanghai, China, pp. 567–572, 2009.
Nonlinear Control, vol. 23, no. 8, pp. 858–877, 2013.
[131] D. E. Quevedo, D. Nešić. Input-to-state stability of packe-
[119] Y. Q. Xia, A. Amann, B. Liu. Detection of abrupt changes tized predictive control over unreliable networks affected by
in electrocardiogram with generalised likelihood ratio al- packet-dropouts. IEEE Transactions on Automatic Control,
gorithm. IET Signal Processing, vol. 4, no. 6, pp. 650–657, vol. 56, no. 2, pp. 370–375, 2011.
2010.
[132] D. Muñoz de la Peña, P. D. Christofides. Lyapunov-
based model predictive control of nonlinear systems subject
[120] M. Y. Chow, Y. Tipsuwan. Network-based control systems: to data losses. IEEE Transactions on Automatic Control,
A tutorial. In Proceedings of the 27th Annual Conference vol. 53, no. 9, pp. 2076–2089, 2008.
of the IEEE Industrial Electronics Society, IEEE, Denver,
USA, pp. 1593–1602, 2001.
[133] G. Pin, T. Parisini. Networked predictive control of un-
certain constrained nonlinear systems: Recursive feasibility
[121] D. Q. Mayne, J. B. Rawlings, C. V. Rao, P. O. M. Scokaert. and input-to-state stability analysis. IEEE Transactions on
Constrained model predictive control: Stability and opti- Automatic Control, vol. 56, no. 1, pp. 72–87, 2011.
mality. Automatica, vol. 36, no. 6, pp. 789–814, 2000.

[134] Y. Q. Xia, G. P. Liu, P. Shi, J. Chen, D. Rees. Robust

[122] H. Chen, F. Allgöwer. A quasi-infinite horizon nonlinear constrained model predictive control based on parameter-
model predictive control scheme with guaranteed stability. dependent Lyapunov functions. Circuits, Systems and Sig-
Automatica, vol. 34, no. 10, pp. 1205–1217, 1998. nal Processing, vol. 27, no. 4, pp. 429–446, 2008.

[135] Y. Iino, T. Hatanaka, M. Fujita. Event-predictive control

[123] G. W. Irwin, J. Chen, A. McKernan, W. G. Scanlon. Co-
for energy saving of wireless networked control system. In
design of predictive controllers for wireless network control.
Proceedings of the American Control Conference, IEEE,
IET Control Theory & Applications, vol. 4, no. 2, pp. 186–
Saint Louis, USA, pp. 2236–2242, 2009.
196, 2010.

[136] E. Garcia, P. J. Antsaklis. Model-based event-triggered

[124] G. P. Liu, J. X. Mu, D. Rees, S. C. Chai. Design and control with time-varying network delays. Proceedings of
stability analysis of networked control systems with ran- the 50th IEEE Conference on Decision and Control and Eu-
dom communication time delay using the modified MPC. ropean Control Conference, IEEE, Orlando, USA, pp. 1650–
International Journal of Control, vol. 79, no. 4, pp. 288–297, 1655, 2011.
2006.

[137] A. Eqtami, D. V. Dimarogonas, K. J. Kyriakopoulos. Novel

[125] H. Zhang, Y. Shi, M. X. Liu. H∞ step tracking control event-triggered strategies for model predictive controllers.
for networked discrete-time nonlinear systems with integral In Proceedings of the 50th IEEE Conference on Decision
and predictive actions. IEEE Transactions on Industrial In- and Control and European Control Conference, IEEE, Or-
formatics, vol. 9, no. 1, pp. 337–345, 2013. lando, USA, pp. 3392–3397, 2011.
364 International Journal of Automation and Computing 12(4), August 2015

[138] E. Henriksson, D. E. Quevedo, H. Sandberg, K. H. Jo- [150] B. Xue, N. Li, S. Li, Q. Zhu. Robust model predictive con-
hansson. Self-triggered model predictive control for network trol for networked control systems with quantisation. IET
scheduling and control. In Proceedings of the 8th IFAC Control Theory & Applications, vol. 4, no. 12, pp. 2896–
Symposium on Advanced Control of Chemical Processes, 2906, 2010.
IFAC, Furama Riverfront, Singapore, pp. 432–438, 2012.

[151] D. Srinivasagupta, H. Schättler, B. Joseph. Time-stamped

[139] J. Wu, L. Q. Zhang, T. W. Chen. Model predictive control model predictive control: An algorithm for control of pro-
for networked control systems. International Journal of Ro- cesses with random delays. Computers & Chemical Engi-
bust and Nonlinear Control, vol. 19, no. 9, pp. 1016–1035, neering, vol. 28, no. 8, pp. 1337–1346, 2004.
2009.

[152] Y. Q. Xia, H. J. Yang, P. Shi, M. Y. Fu. Constrained

[140] J. F. Liu, D. Muñoz de la Peña, P. D. Christofides, J. F. infinite-horizon model predictive control for fuzzy-discrete-
Davis. Lyapunov-based model predictive control of nonlin- time systems. IEEE Transactions on Fuzzy Systems, vol. 18,
ear systems subject to time-varying measurement delays. no. 2, pp. 429–436, 2010.
International Journal of Adaptive Control and Signal Pro-
cessing, vol. 23, no. 8, pp. 788–807, 2009.
[153] Y. Q. Xia, M. Y. Fu, P. Shi. Analysis and Synthesis of Dy-
namical Systems with Time-delays, Berlin, Germany, Hei-
[141] L. C. Jin, R. Kumar, N. Elia. Model predictive control- delberg: Springer, 2009.
based real-time power system protection schemes. IEEE
Transactions on Power Systems, vol. 25, no. 2, pp. 988–998,
2010. [154] Y. Q. Xia, M. Y. Fu, G. P. Liu. Analysis and Synthesis of
Networked Control Systems, Berlin, Germany, Heidelberg:
Springer, 2011.
[142] B. Yu, Y. Shi, J. Huang. Modified generalized predictive
control of networked systems with application to a hy-
draulic position control system. Journal of Dynamic Sys- [155] Y. Q. Xia, G. P. Liu, M. Y. Fu, D. Rees. Predictive con-
tems, Measurement, and Control, vol. 133, no. 3, Article trol of networked systems with random delay and data
number 031009, 2011. dropout. IET Control Theory & Applications, vol. 3, no. 11,
pp. 1476–1486, 2009.
[143] Y. Zou, T. Chen, S. Li. Network-based predictive control
of multirate systems. IET Control Theory & Applications, [156] Y. Q. Xia, F. M. Fu, B. Liu, G. P. Liu. Design and per-
vol. 4, no. 7, pp. 1145–1156, 2010. formance analysis of networked control systems with ran-
dom delay. Journal of Systems Engineering and Electronics,
vol. 20, no. 4, pp. 807–822, 2009.
[144] Y. Zhang, S. Y. Li. Networked model predictive control
based on neighbourhood optimization for serially connected
large-scale processes. Journal of Process Control, vol. 17, [157] S. C. Chai, G. P. Liu, D. Rees. Design and implementation
no. 1, pp. 37–50, 2007. of networked predictive control systems. In Proceedings of
the 16th IFAC World Congress, IFAC, Prague, Czech Re-
[145] G. P. Liu, J. X. Mu, D. Rees. Networked predictive control public, 2005.
of systems with random communication delay. In Proceed-
ings of UKACC International Conference on Control, Bath, [158] W. S. Hu, G. P. Liu, D. Rees. Design and implementation
UK, 2004. of networked predictive control systems based on round trip
time delay measurement. In Proceedings of the American
[146] Y. B. Zhao, G. P. Liu, D. Rees. Networked predictive con- Control Conference, IEEE, Minneapolis, USA, pp. 674–679,
trol systems based on the Hammerstein model. IEEE Trans- 2006.
actions on Circuits and Systems II: Express Briefs, vol. 55,
no. 5, pp. 469–473, 2008.
[159] S. C. Chai, G. P. Liu, D. Rees, Y. Q. Xia. Design and
practical implementation of internet-based predictive con-
[147] D. E. Quevedo, J. Ostergaard, D. Nešić. Packetized pre- trol of a servo system. IEEE Transactions on Control Sys-
dictive control of stochastic systems over bit-rate limited tems Technology, vol. 16, no. 1, pp. 158–168, 2008.
channels with packet loss. IEEE Transactions on Automatic
Control, vol. 56, no. 12, pp. 2854–2868, 2011.
[160] G. P. Liu, D. Rees, S. C. Chai. Design and practical im-
plementation of networked predictive control systems. In
[148] D. E. Quevedo, E. I. Silva, G. C. Goodwin. Packetized Proceedings of the 12th IEEE International Conference on
predictive control over erasure channels. In Proceedings of Networking, Sensing and Control, IEEE, Tucson, USA,
the American Control Conference, IEEE, New York, USA, pp. 336–341, 2005.
pp. 1003–1008, 2007.

[161] G. P. Liu, Y. Q. Xia, D. Rees, W. S. Hu. Design and sta-

[149] P. Mhaskar, N. H. El-Farra, P. D. Christofides. Predic- bility criteria of networked predictive control systems with
tive control of switched nonlinear systems with scheduled random network delay in the feedback channel. IEEE Trans-
mode transitions. IEEE Transactions on Automatic Con- actions on Systems, Man, and Cybernetics – Part C: Ap-
trol, vol. 50, no. 11, pp. 1670–1680, 2005. plications and Reviews, vol. 37, no. 2, pp. 173–184, 2007.
 Y. Q. Xia et al. / Recent Progress in Networked Control Systems – A Survey 365

[162] W. S. Hu, G. P. Liu, D. Rees. Event-driven networked [174] J. X. Xu, Z. S. Hou. Notes on data-driven system ap-
predictive control. IEEE Transactions on Industrial Elec- proaches. Acta Automatica Sinica, vol. 35, no. 6, pp. 668–
tronics, vol. 54, no. 3, pp. 1603–1613, 2007. 675, 2009.

[163] Y. Q. Xia, L. Li, G. P. Liu, P. Shi. H∞ predictive con- [175] Y. Q. Xia, W. Xie, B. Liu, X. Y. Wang. Data-driven pre-
trol of networked control systems. International Journal of dictive control for networked control systems. Information
Control, vol. 84, no. 6, pp. 1080–1097, 2011. Sciences, vol. 235, pp. 45–54, 2013.

[164] R. Wang, G. P. Liu, W. Wang, D. Rees, Y. B. Zhao. [176] G. Chaloulos, P. Hokayem, J. Lygeros. Distributed hier-
H∞ control for networked predictive control systems based archical MPC for conflict resolution in air traffic control.
on the switched Lyapunov function method. IEEE Trans- In Proceedings of the American Control Conference, IEEE,
actions on Industrial Electronics, vol. 57, no. 10, pp. 3565– Baltimore, USA, pp. 3945–3950, 2010.
3571, 2010.

[177] T. Keviczky, F. Borrelli, K. Fregene, D. Godbole, G. J.

[165] B. Liu, Y. Q. Xia, M. S. Mahmoud, H. Wu, S. S. Cui. Balas. Decentralized receding horizon control and coordina-
New predictive control scheme for networked control sys- tion of autonomous vehicle formations. IEEE Transactions
tems. Circuits, Systems, and Signal Processing, vol. 31, on Control Systems Technology, vol. 16, no. 1, pp. 19–33,
no. 3, pp. 945–960, 2012. 2008.

[166] G. P. Liu, S. C. Chai, J. X. Mu, D. Rees. Networked predic- [178] B. T. Stewart, S. J. Wright, J. B. Rawlings. Cooperative
tive control of systems with random delay in signal trans- distributed model predictive control for nonlinear systems.
mission channels. International Journal of Systems Science, Journal of Process Control, vol. 21, no. 5, pp. 698–704, 2011.
vol. 39, no. 11, pp. 1055–1064, 2008.

[179] A. N. Venkat, I. A. Hiskens, J. B. Rawlings, S. J. Wright.

[167] G. P. Liu, Y. Q. Xia, J. Chen, D. Rees, W. S. Hu. Net- Distributed MPC strategies with application to power sys-
worked predictive control of systems with random network tem automatic generation control. IEEE Transactions on
delays in both forward and feedback channels. IEEE Trans- Control Systems Technology, vol. 16, no. 6, pp. 1192–1206,
actions on Industrial Electronics, vol. 54, no. 3, pp. 1282– 2008.
1297, 2007.

[180] Y. Kuwata, A. Richards, T. Schouwenaars, J. P. How. Dis-

[168] G. P. Liu. Predictive controller design of networked sys- tributed robust receding horizon control for multivehicle
tems with communication delays and data loss. IEEE guidance. IEEE Transactions on Control Systems Technol-
Transactions on Circuits and Systems II: Express Briefs, ogy, vol. 15, no. 4, pp. 627–641, 2007.
vol. 57, no. 6, pp. 481–485, 2010.

[181] P. Trodden, A. Richards. Cooperative distributed MPC

[169] R. N. Yang, G. P. Liu, P. Shi, C. Thomas, M. V. Basin. Pre- of linear systems with coupled constraints. Automatica,
dictive output feedback control for networked control sys- vol. 49, no. 2, pp. 479–487, 2013.
tems. IEEE Transactions on Industrial Electronics, vol. 61,
no. 1, pp. 512–520, 2014.
[182] A. Richards, J. P. How. Robust distributed model predic-
tive control. International Journal of Control, vol. 80, no. 9,
[170] Y. Q. Xia, J. Chen, G. P. Liu, D. Rees. Stability analysis pp. 1517–1531, 2007.
of networked predictive control systems with random net-
work delay. In Proceedings of IEEE International Confer-
ence on Networking, Sensing and Control, IEEE, London, [183] M. A. Müller, M. Reble, F. Allgöwer. A general distributed
UK, pp. 815–820, 2007. MPC framework for cooperative control. In Proceedings
of the 18th IFAC World Congress, IFAC, Milano, Italy,
pp. 7987–7992, 2011.
[171] R. Wang, B. Wang, G. P. Liu, W. Wang, D. Rees. H∞
controller design for networked predictive control systems
based on the average dwell-time approach. IEEE Transac- [184] P. Wang, B. C. Ding. Distributed RHC for tracking and
tions on Circuits and Systems II: Express Briefs, vol. 57, formation of nonholonomic multi-vehicle systems. IEEE
no. 4, pp. 310–314, 2010. Transactions on Automatic Control, vol. 59, no. 6, pp. 1439–
1453, 2014.

[172] R. Kadali, B. Huang, A. Rossiter. A data driven subspace

approach to predictive controller design. Control Engineer- [185] W. B. Dunbar. Distributed receding horizon control of dy-
ing Practice, vol. 11, no. 3, pp. 261–278, 2003. namically coupled nonlinear systems. IEEE Transactions on
Automatic Control, vol. 52, no. 7, pp. 1249–1263, 2007.

[173] D. Laurı́, J. A. Rossiter, J. Sanchis, M. Martı́nez. Data-

driven latent-variable model-based predictive control for [186] J. M. Maestre, R. R. Negenborn. Distributed Model Pre-
continuous processes. Journal of Process Control, vol. 20, dictive Control Made Easy, The Netherlands: Springer,
no. 10, pp. 1207–1219, 2010. 2014.
366 International Journal of Automation and Computing 12(4), August 2015

[187] R. R. Negenborn, J. M. Maestre. Distributed model predic- [200] Y. Q. Xia, N. Zhou, K. F. Lu, Y. Li. Attitude control of
tive control: An overview and roadmap of future research multiple rigid bodies with uncertainties and disturbances.
opportunities. IEEE Control Systems Magazine, vol. 34, IEEE/CAA Journal of Automatica Sinica, vol. 2, no. 1,
no. 4, pp. 87–97, 2014. pp. 2–10, 2015.

[188] A. Eqtami, D. V. Dimarogonas, K. J. Kyriakopoulos. [201] Y. Q. Xia. From networked control systems to cloud con-
Event-based model predictive control for the cooperation trol systems. In Proceedings of the 31st Chinese Control
of distributed agents. In Proceedings of the American Con- Conference, IEEE, Hefei, China, pp. 5878–5883, 2012.
trol Conference, IEEE, Montreal, Canada, pp. 6473–6478,
2012.
[202] The economist, cloud computing: Clash
of the clouds, [Online], Available:
[189] A. Eqtami, D. V. Dimarogonas, K. J. Kyriakopoulos. http://www.economist.com/node/14637206, 3 Novem-
Event-triggered strategies for decentralized model predic- ber, 2009.
tive controllers. In Proceedings of the 18th IFAC World
Congress, IFAC, Milano, Italy, pp. 10068–10073, 2011.
[203] Gartner says cloud computing will be as in-
fluential as e-business, [Online], Availabe:
[190] W. B. Dunbar. Distributed Receding Horizon Control of http://www.gartner.com/newsroom/id/707508, August
Multiagent Systems, Ph. D dissertation, California Institute 22, 2010.
of Technology, USA, 2004.

[204] G. Galen. What cloud computing really means. InfoWorld,

[191] W. B. Dunbar, R. M. Murray. Distributed receding horizon [Online], Available: http://www.infoworld.com/d/cloud-
control for multi-vehicle formation stabilization. Automat- computing/what-cloud-computing-really-means-031, June
ica, vol. 42, no. 4, pp. 549–558. 2006. 6, 2009.

[192] T. Keviczky, F. Borrelli, G. J. Balas. Decentralized reced-

[205] Cloud computing, [Online], Available: http://en.
ing horizon control for large scale dynamically decoupled
wikipedia.org/wiki/Cloud computing, November 27, 2011.
systems. Automatica, vol. 42, no. 12, pp. 2105–2115, 2006.

[206] Cloud computing, [Online], Available:

[193] H. P. Li, Y. Shi. Distributed model predictive control of
http://www.cloudcomputingdefined.com/, July 17,
constrained nonlinear systems with communication delays.
2010.
Systems & Control Letters, vol. 62, no. 10, pp. 819–826,
2013.
[207] T. Segaran, J. Hammerbacher. Beautiful Data: The Sto-
[194] H. P. Li, Y. Shi. Robust distributed model predictive con- ries Behind Elegant Data Solutions, Sebastopol: O Reilly
trol of constrained continuous-time nonlinear systems: A Media, 2009.
robustness constraint approach. IEEE Transactions on Au-
tomatic Control, vol. 59, no. 6, pp. 1673–1678, 2014. [208] T. White. Hadoop: The Definitive Guide, 3rd ed., Se-
bastopol: O Reilly Media, 2012.
[195] H. P. Li, Y. Shi. Distributed receding horizon control of
large-scale nonlinear systems: Handling communication de- [209] Y. Q. Xia. Cloud control systems. IEEE/CAA Journal of
lays and disturbances. Automatica, vol. 50, no. 4, pp. 1264– Automatica Sinica, vol. 2, no. 2, pp. 134−142, 2015.
1271, 2014.

[196] Y. Q. Xia, M. Y. Fu. Compound Control Methodology Yuan-Qing Xia received the M. Sc. de-
for Flight Vehicles, Berlin, Germany, Heidelberg: Springer, gree in fundamental mathematics from An-
2013. hui University, China in 1998, and the
Ph. D. degree in control theory and con-
trol engineering from Beijing University of
[197] N. Zhou, Y. Q. Xia, K. F. Lu, Y. Li. Decentralised finite- Aeronautics and Astronautics, China in
time attitude synchronisation and tracking control for rigid 2001. From 2002 to 2003, he was a post-
spacecraft. International Journal of Systems Science, to be doctoral research associate in the Institute
published. of Systems Science, Academy of Mathemat-
ics and System Sciences, Chinese Academy of Sciences China.
[198] N. Zhou, Y. Q. Xia, M. L. Wang, M. Y. Fu. Finite-time From 2003 to 2006, he was with the National University of Sin-
attitude control of multiple rigid spacecraft using terminal gapore and the University of Glamorgan, UK, as a research fel-
sliding mode. International Journal of Robust and Nonlin- low, respectively. From 2007 to 2008, he was a guest professor
ear Control, to be published. with Innsbruck Medical University, Austria. In 2012, he was ap-
pointed as Xu Teli Distinguished Professor at Beijing Institute
of Technology and obtained the National Science Foundation for
[199] N. Zhou, Y. Q. Xia, K. F. Lu. Attitude synchronization Distinguished Young Scholars of China. Since 2008, he has been
of rigid spacecraft using terminal sliding mode. In Proceed- working for Beijing Institute of Technology as a professor. He
ings of the 32nd Chinese Control Conference, IEEE, Xi an, has published eight monographs in Springer and John Wiley,
China, pp. 706–711, 2013. and more than 100 papers in journals. He obtained the Second
 Y. Q. Xia et al. / Recent Progress in Networked Control Systems – A Survey 367

Award of the Beijing Municipal Science and Technology (No. 1) Beijing Institute of Technology (BIT), China. From March 2012
in 2010, the Second National Award for Science and Technology to March 2013, supported by China Scholarship Council (CSC),
(No. 2) in 2011, and the Second Natural Science Award of The she was a visiting scholar in the University of New Orleans,
Ministry of Education (No. 1) in 2012. USA. Currently, she is an associate professor in BIT. She has
His research interests include networked control systems, ro- co-authored two books and more than 40 journal and conference
bust control, active disturbance rejection control and flight con- papers.
trol. Her research interests include multisensor data fusion, state
E-mail: xia [email protected] (Corresponding author) estimation, image registration, intelligent navigation, and inte-
ORCID iD: 0000-0002-5977-4911 grated navigation.
E-mail: [email protected]
Yu-Long Gao received the B. Sc. degree
in automation from Beijing Institute of Meng-Yin Fu received the B. Sc. degree
Technology, China. He is now pursuing the in radio electronics from Liaoning Univer-
M. Sc. degree in control science and engi- sity, China in 1987, the M. Sc. degree in
neering in Beijing Institute of Technology. control theory from Beijing Institute of
His research interests include networked Technology, China in 1992, and the Ph. D.
control system, model predictive control, degree in geodetic engineering from Insti-
and multi-agent system. tute of Geodesy and Geophysics, Chinese
E-mail: [email protected] Academy of Sciences, China in 2000. From
August 1987 to September 1989, he worked
Li-Ping Yan received the B. Sc. degree as an assistant engineer at the Weather Bureau of Inner Mon-
and M. Sc. degree both in mathematics golia. Since March 1992, he has been with Beijing Institute of
from Henan University, China in 2000 and Technology as a professor. He has coauthored three books and
2003, respectively, and received the Ph. D. published more than 50 papers in technical journals and con-
degree in control science and engineering ferences. He is an executive director of the Chinese Society of
from Tsinghua University, China in 2007. Inertial Technology, fellow of Beijing Association for Science and
From January 2007 to July 2009, she was Technology.
a postdoctoral research associate in the His research interests include inertial navigation and intelli-
Equipment Academy of Airforce, China. gent navigation, guidance and control.
Since July 2009, she has been with the School of Automation, E-mail: [email protected]