Sei sulla pagina 1di 5

Control Systems Designed for Wireless Sensor and

Actuator Networks*
Xianghui Cao
State Key Lab. of Industrial
Control Technology
Zhejiang University
Hangzhou, 310027, China
xh.cao@ieee.org
Jiming Chen
State Key Lab. of Industrial
Control Technology
Zhejiang University
Hangzhou, 310027, China
jmchen@ieee.org
Yang Xiao
Dept. of Computer Science
University of Alabama
Tuscaloosa, AL, 35487, USA
yangxiao@ieee.org
Youxian Sun
State Key Lab. of Industrial
Control Technology
Zhejiang University
Hangzhou, 310027, China
yxsun@iipc.zju.edu.cn
AbstractThis paper presents a theoretical model of control
and communication over wireless sensor and actuator networks
(WSANs). We propose two control schemes, a centralized control
scheme (CC) in which decisions are made based on global
information, and a distributed control scheme (DC) that en-
ables distributed actuators to make decisions locally. Because
of global information, CC can obtain optimal control at each
step. However, when that information is delivered over lossy
wireless channels, it could become unstable. It is demonstrated
by simulations that DC could also stabilize the control system
analogously with the CC, though with more steps.
I. INTRODUCTION
Wireless sensor networks (WSNs) can be widely applied
in environmental monitoring, biological and chemical attack
detection [1], home automation [2], agricultural usages, etc.
The sensors are low-cost and low power devices with limited
computation and communication capabilities, and they can
only work as passive devices that collect data from the
environment without interaction. WSANs comprise of groups
of sensors and actuators that are connected with wireless
medium. Sensors and actuators collaborate together to monitor
and affect physical world, in which sensors serve to collect
environmental information, while actuators can make deci-
sions, interact with the environment, and carry out appropriate
actions according to the sensory information. The actuators
are resource rich devices with more energy, higher power,
and better processing capabilities than sensors. In many cases,
single or multiple controllers or agents are involved in WSANs
to collect sensory information, schedule tasks, make decision,
and drive actuators. From a viewpoint of control theory ,
traditional WSNs are unidirectional or open-loop systems
that only detect the physical world, whereas, WSANs are
bidirectional or closed loop systems that can further interact
with it automatically.
*This work is supported by the Nature Science Foundation of China (NSFC)
under Grant 60604029, Nature Science Foundation of Zhejiang Province under
Grant Y106384, Joint Funds of NSFC-Guangdong under Grant U0735003,
the Specialized Research Fund for the Doctoral Program of Higher Education
(SRFDP) under Grant 20050335020, the R&D Program of Zhejiang Province
under Grant No.2007C31038 and 111 Projects under Grant B07031. Prof.
Xiaos work was partially supported by the US National Science Foundation
(NSF) under the grant CNS-0716211.
WSANs have many applications. The following example
is a typical example of WASNs and has attracted a lot of
research attentions [3]. In an intelligent lighting system for
ofce automation, a number of sensors (photo-sensors) and
actuators (lights) are used to satisfy users preferences. The
goal of such a system is to drive the lights (dimmable lighting
ballasts mounted on a ceiling) to provide lighting comfort
levels while save electrical energy and expenditure. Sensors
sense their surroundings and provide real-time illuminance
measurements which are going to be validated and aggregated.
A control unit analyzes sensor readings, evaluates current
status such as user preferences, electricity rates, if required,
and then makes decisions to adjust the actuators. To run this
system, a multi-agent approach is considered, which is able to
handle preferences of multiple users in a workspace, beneting
from supervised distributed learning method. The learner treats
the sensors readings as input and user actions as target values
[3]. The paper [4] presents another stationary lighting control
system in a large cubicle farm, which focuses on strategies
of making control decisions. In order to meet a pre-required
light intensity at certain user cubicles, a number of sensors are
deployed at those cubicles to calibrate manner of the lights.
An HVAC (Humidity, Ventilation, Air Conditioning) con-
trol system for buildings is another application of WSANs.
A main purpose of such system is to reduce energy cost.
Different from light control, temperature, airow, humidity
and other environmental variables are not very easy to be
controlled. Similarly, in the area of precise farming, agriculture
maintenance, actuators (irrigators, heating and cooling devices,
automatic fertilizing equipments, etc.) function to maintain the
living environment of crops and livestock.
The introduction of intelligent actuators to WSNs lever-
ages these monitoring-oriented systems to enable interaction
between human and the physical world, and control. In
return, nding control strategy is a new challenging issue
that should be addressed. Although there have been some
researches on WSANs especially on communication protocols
and eld tests such as [5][6], few of them focus their efforts
on control system design. Sinopoli et. al [7] introduced a
control problem involved in WSANs motivated by distributed
PEG, and presented the research challenges within a dis-
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the ICC 2008 proceedings.
978-1-4244-2075-9/08/$25.00 2008 Crown Copyright
tributed control system, e.g., no global information. They
also proposed a hierarchical control model including a lower
continuous component and a higher discrete component. A
distributed actuation scheme was discussed in [8]. Because of
the superiority of energy saving of the distributed actuation
scheme over a centralized one, the paper emphasized on
the distributed system and developed several useful actuation
strategies based on simulations. However, this paper did not
give theoretical presentations of the actuation problem. In
[4], a distributed control problem was addressed by a new
algorithm and network communication details. Though, the
paper does not make enough consideration of the control
strategy to be applied. There are also some other papers on the
control issues within the area of WSANs, e.g., the paper [9]
describes a design of control strategies based on distributed
data using Q-leaning framework; the paper [10] models a
control system based on frame theory; Gerkey et. al have built
up a market-like competitive system in which actuators biding
for tasks based on their abilities in [11]. Our paper focuses on
the control strategy for running over WASNs.
One big challenge in designing the control strategy is to
use a small number of actuators to control the environment
parameters and match a large number of user preferences.
In the view of save energy, its reasonable to employ fewer
actuators to perform user required tasks. While low cost
sensors are always in large amount to provide more informa-
tion. As a result, some actuators are responsible for multiple
user preferences; at the same time, some user preferences
are going to be realized by multiple actuators. In this case,
the performance of the WSAN in terms of satisfying user
requirements is inapparent, and will be claried in the paper.
We will propose two distinct control approaches as shown
in Fig. 1: (1) the central controlled (CC) WSAN in which
theres only one centralized controller that generates control
instructions based on the entire network information; (2) the
distribute controlled (DC) WSAN in which control algorithms
are performed locally by each of the actuators. They will be
discussed in detail in Section III and Section IV.
Actuator
Sensor
Centralized
Controller ra
rs
Sensorypackets
Controlleroutputs
(a) Centralized control scheme
Distributed
Controller
&Actuator
Sensor
ra
rs
Sensorypackets
transmittingin
multi-hopway
(b) Distributed control scheme
Fig. 1. Topology view of two distinct control system schemes designed for
the WSAN.
II. SYSTEM MODELING
We focus on problems involving only one type of the
user requirement (e.g., the illuminance, the environmental
temperature, the water inux to farms, etc.) that needs to
be controlled. Sensors are deployed at places where user
preferences have been specied whoes magnitudes are denoted
by a vector s
p
. Basic notations we used are listed in Table. I.
The automatical control process starts when the user require-
ments has been settled or some abnormal events take place.
Performance of the actuators will be measured and fed back
by the sensors to the controller for evaluation and making
decision of the next step. Suppose the sensors and actuators
both work synchronously.
TABLE I
NOTATIONS DEFINITION
variable Denition
ns total number of sensors deployed in the area of con-
cern
na total number of actuators
S
j
the jth sensor, j {1, , ns}
A
i
the ith actuator, i {1, , na}
environmental variable
sp
j
user desired at S
j
s location
y
j
the actual at S
j
ym
j
S
j
s measurement about
x
i
A
i
s output that could inuence , x
i
[X
i
, X
i
]
u
i
the controller(s) output towards A
i
sp, x, y, ym, u vector forms of {sp
j
}, {x
i
}, {y
j
}, {ym
j
}, {e
j
},
{u
i
}, respectively. sp, y, ym R
ns
, x, u R
na
E
SR
system residual error dened in Eq. (5), as a metric
of the system performance
rs the communication range of each sensor and actuator
ra the inuence range of each actuator
k the kth control step, k = 0, 1, 2,
n
[k]
, n
[k]
, n
[k]
system noises
The change of each actuator actuation is assumed linearly
proportional to the controllers input if received by this actu-
ator, which is described by the rst line of Eq. (1). In cases
when the controllers input is not received by the actuator, it
remains unchanged. Due to the actuators physical restraints,
their actuation always drifts from the expected value, thus
additional noise
[k]
is added. We also assume the current
value of environment variable is related to both its previous
value and the actuators actuation, which is described by the
second line of Eq. (1). For instance, the temperature of an
ofce cannot maintain unchanged even without actuators (air
conditioners) since thermal exchanging with outside open air
probably takes place all the time, i.e., temperature of the next
relates but may not equal to the previous. This is modeled by
the term contains C. The elements of C are less than 1 and
possibly negative. In cases of lighting control, C = 0 since
the illuminance is irrelevant to its historical values. We call
the terms x
[k1]
and Cy
[k1]
the inertial parts of the system.
Noise
[k]
is similarly added because of environmental noise.
Above all, the kth control step can be described by

x
[k]
= x
[k1]
+V
ca[k]
Bu
[k]
+n
[k]
y
[k]
= Cy
[k1]
+Dx
[k]
+n
[k]
(1)
where B, C, D are matrices of appropriate dimensions. Matrix
B, indicating the coefcient of the controllers effect on
the actuators, is assumed inversable (i.e., B
1
). V
ca[k]
=
Diag{v
i[k]
}, where each v
i[k]
is i.i.d bernoulli random vari-
able with P(v
i[k]
= 1) = v
i
, indicateing the effect of
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the ICC 2008 proceedings.
Centralized
Controller
Actuators Plant
Sensors
Wireless communication
Loop2
Loop1
Fig. 2. Control ow chart of the centralized scheme
controllers packet loss on actuators states. For instance, if
v
i[k]
= 0, x
i[k]
= x
i[k1]
i.e., A
i
will keep its actuation if
no instruction packet received from the controllers.
The state feedback process provides the controller for esti-
mating actuators states with
f
[k]
= V
ac[k]
x
[k]
(2)
where V
ac[k]
R
n

s
ns
depicts the packet losses during state
feedback with n

s
be the number of sensory packets arriving the
controller. Every nonzero entry of V
ac[k]
is an i.i.d bernoulli
random variable. Note that the controller can know V
ac[k]
by checking feedback packets original IDs. Theres another
feedback process to the controller, the output feedback by the
sensor nodes for evaluating actuators performance, as shown
in Eq. (3).
y
m[k]
= V
sc[k]
(y
[k]
+n
[k]
) (3)
n
[k]
is the measurement noise. V
sc[k]
is similarly dened as
V
ac[k]
.
III. CENTRALIZED CONTROL SCHEME (CC)
The ow chart of CC is shown in Fig. 2, where the Plant
stands for the environment. The centralized controller makes
decisions based on the feedback from actuators as well as
sensors. u is the output of the controller which drives the
actuators. Because of the unreliability of wireless communi-
cations, the fed back information to the controller is always
observed intermittent and noisy. The controller should rst
estimate present state of the whole network, before deciding
its output.
After the (k 1)th step and before making the
kth decision, the controller already knows I
[k]
=
{u
k1
, f
k1
, y
k1
m
, V
k1
ac
, V
k1
sc
}, ()
k
= {()
[0]
, . . . , ()
[k]
}.
Via the two feedback processes, the controller could obtain an
estimation of the actuators outputs and also the system outputs,
i.e., x
[k1]
= E{x
[k1]
|I
[k]
} and y
[k1]
= E{y
[k1]
|I
[k]
}.
These estimations can be obtained by running the Extended
Kalman Filter (EKF) [12] rst on the subsystems described
by Loop1 as in Fig. 2, and then on the subsystem of Loop2.
If the controller takes u
[k]
as its output, the error between
the set points and the system outputs is
e
[k]
= s
p
y
[k]
= r
[k]
DV
ca[k]
Bu
[k]
(4)
where r
[k]
= s
p
Cy
[k1]
Dx
[k1]
Dn
[k]
n
[k]
. Dene
the system residual error as
E
SR[k]
= e
T
[k]
e
[k]
(5)
The estimated E
SR
base on I
[k]
is

E
SR[k|k]
= E

E
SR[k]

I
k

= E

r
T
[k]
r
[k]
2u
T
[k]
B
T
V
ca[k]
D
T
r
[k]

I
k

+ E

u
T
[k]
B
T
V
T
ca[k]
D
T
DV
ca[k]
Bu
[k]

I
k

= 2u
T
[k]
B
T
z
[k]
+u
T
[k]
B
T
QBu
[k]
(6)
where = E

r
T
[k]
r
[k]

, Q = V
ca
D
T
DV
ca
+ (V
ca

V
2
ca
)G{D
T
D}, z
[k]
= V
ca
D
T
E

r
[k]

I
k

= V
ca
D
T
(s
p

C y
[k1]
D x
[k1]
). Now, by applying the quadratic pro-
gramming method, the controller could nd the optimal u
[k]
,
i.e., u
opt[k]
, that minimizes the estimated system residual error

E
SR[k|k]
. The method is described in Eq. (7).
MIN

E
SR[k|k]
FIND u
opt[k]
S.T. x
[k]
= x
[k1]
+Bu
[k]
[X, X]

(7)
where the condition line considers the worst case.
IV. DISTRIBUTED CONTROL SCHEME (DC)
The distribute control scheme breaks up the centralized
control and distribute it each of the actuators so that they can
decide by themselves. However, one of the arising problem lies
in the difculty to address the system coupling. On one hand,
to reach each user requirement, perhaps a group of actuators
will be aroused up; on the other hand, each actuator may
inuence the measurement of a subgroup of all the sensors. An
actuator doesnt have a global view of the whole system such
as the sensors measurements and the other actuators outputs
which could disturbance its anticipated local performance. One
approach that let all the actuators communicate with each
other, especially between those inuence the measurement of
the same sensors, before making any decision, could certainly
alleviate above situation. However, this may probably cause
the Domino Effect that every actuator may have to wait for a
long time to decide its output. As a result, such approach is
somewhat unfeasible especially for large scaled WSANs. In
this paper in stead, a distributed controller makes decisions
only based on local sensory information, but at the expense of
multiple coordination steps. The network-wide synchroniza-
tion is achieved in [4] by presetting two T
timeout
s, one for
sensors to synchronously report messages and the other for
actuators for synchronous actuation.
In DC, the system formula is quite the same as Eq. (1)
except for the term V
ca[k]
, which equals to I in the distributed
case. The optimal controller design for CC does not work
any more, since the actuators in DC scheme do not have
global view of the whole system and no knowledge of global
parameters such as Q (Eq. (6)) as a result. Instead of pursuing
global optimal solution, we introduce the concept of the BP
(Back Propagation) Articial Neural Network (ANN), due to
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the ICC 2008 proceedings.
Input Layer
Middle Layer
Output Layer
Fig. 3. Architecture of the designed BP neural network.
its parallel computing ability, in the purpose of discovering
distributed control strategies for the actuators. The gradient
descending method which is commonly used in BP ANNs is
going to be applied.
Basically, three layered BP ANN is built for DC, namely,
the input layer, the middle layer (or the so-called hidden layer)
and the output layer, as shown in Fig. 3. The input for nerve
cells in the middle layer and the output for nerve cells in
the output layer can be formulated using Eq. (1). In order to
apply the gradient descending method, we need to compensate
the inertial parts of the system (see Eq. (1)). Therefore, let
u
[k1]
= u

[k1]
B
1
x
[k2]
. Given the desired output s
p
and the learning error E
L
(Eq. (8)), we add a momentum
term of u

[k1]
to the control decision to compensate the
inertial part of y
[k]
(eq. (9)).
E
L[k]
=
1
2
(s
p
y
[k1]
)
T
(s
p
y
[k1]
) (8)
By using Eq. (1), we then get
u

[k]
=

u

[k1]
E
L[k]
+ u

[k1]
= B
T
D
T
e
[k1]
+ u

[k1]
(9)
where e
[k1]
= s
p
y
[k1]
, = Diag{
1
, . . . ,
na
} is called
the learning step length, = Diag{
1
, . . . ,
na
}. Thus,
u
[k]
= u

[k1]
+ u

[k]
B
1
x
[k1]
= (I
nana
+ )u
[k1]
u
[k2]
+ B
T
D
T
e
[k1]
B
1

x
[k1]
(I
nana
+ )x
[k2]
+ x
[k3]

(10)
Then, for each actuator A
i
, its output is obtained by
u
i[k]
= (1 +
i
)u
i[k1]

i
u
i[k2]
+
i
b
i
d
T
i
e
[k1]

1
b
i

x
i[k1]
(1 +
i
)x
i[k2]
+
i
x
i[k3]

(11)
where d
i
is the ith column vector of D. All the above items,
except for d
T
i
e
[k1]
are local and known by A
i
. For A
i
,
since only the sensors measurements that are inuenced by
A
i
will be collected, each locally unknown e
i[k1]
will be
multiplied by 0 in d
T
i[k1]
in above equation. Besides, e
j[k1]
that is temporarily not received by A
i
will be viewed as 0.
Consequently, each actuator is able to calculate its output
according to Eq. (11) based on only local information.
V. PERFORMANCE EVALUATION
In this section, our proposed control schemes (CC and DC)
are simulated and evaluated. We conducted our simulations
within Matlab 7.3 environment. The sensor area of concern
was 1612, where we uniformly deployed 48 sensors and 20
actuators, as shown in Fig. 4 which also gives the distribution
of the user requirements (s
p
) at the sensor places. The coef-
cient of actuator A
j
s inuence on sensor S
j
s measurement
(entity d
ij
of the matrix D) is set inversely proportional to
the square distance between them. It does not matter whether
the inverse proportion is exact, its just an example of dening
matrix D. At the initial testing stage, the actuators function in
turn such that they can get the knowledge about D indicating
the inuencing coefcients judging from sensors feedback.
During the simulations, B = I
nana
, C = 0.8I
nsns
.
sensor actuator
26
21
16
11
r
a
s
p
Fig. 4. Deployment of the WSAN and the spatial distribution of the user
requirements (sp) at the sensor places. The circle of radium ra indicates the
inuence range of the actuators.
In CC, the centralized controller is xed at the geometric
center of the sensor eld. In the coordination stage, each sensor
sends its ID as well as measurement of to the controller
which will then calculate the optimal actions for each of the
actuators according to Eq. (7). The sensor-controller communi-
cation uses multi-hopping since the communication range of a
sensor is constrained by a small r
s
. While in DC, each actuator
only need to know local inuencing coefcients according to
Eq. (11). Both CC and DC are assumed keep the sensors and
actuators (controllers) working synchronously.
Fig. 5 illustrates the simulation results of the performances
of the CC and DC with the deployment plotted in Fig. 4
when = 0.8, = 0.5, x [30, 30], packet loss ratio is
20%, and all system noises power equal 0. Note, the vertical
axes, labeled by E
SR
(%), represents the percentage of
E
SR[k]
over its initial value for each coordination step. This
is also true for the following gures. It can be seen the
WSAN under centralized control converges much quicker than
under distributed control scheme, because the former runs the
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the ICC 2008 proceedings.
0 5 10 15 20 25 30 35 40 45 50
0
20
40
60
80
100
k
E
S
R
(%
)
CC
DC
CC: 8.2%
DC: 8.9%
Fig. 5. Performances of the control strategies of CC and DC, in ideal case
(theres no system noise, packet loss, or actuation bounds). The residual error
obtained by both control strategies after 50 steps are annotated within this
gure. The +s represent the results of 10 independent simulations, while the
solid line represents the mean of those results.
optimal control strategy with the controllers global knowledge
at each step, while the latter pursuits the optimal step by step.
Due to the unreliability of wireless communications, losing
packets is a common phenomenon. The packet lost here also
take into account the packets which are delayed so much that
out of the synchronized period for each control step amongst
the sensors, actuators and controller. We use a constant to
model the average packet loss rate of the wireless channels.
The simulation results are shown in Fig. 6. = 1 meaning no
feedback packets are received by the controller or actuators,
and therefore nothing will the control systems do. Easily seen,
CC manifests much sensitivity towards . Other simulations
reveal that it will even become unstable when > 50%.
Things are much favorable to us for DC since it even works
under much bigger , which demonstrates the great advantages
of the distributed control strategy.
VI. CONCLUSION
In this paper, we addressed the control problem using
wireless sensor and actuator networks for applications of
environment control, building automation, etc. Sensors sense
their surroundings, their measurements are analyzed, and ac-
tuators are controlled to satisfy the network users preferences.
We developed a theoretical model for such systems, based
on which two general control schemes including the central
control (CC) and the distribute control (DC) are designed
respectively. Although, CC could optimally, on the whole,
utilize the actuators to reach user requirements, it suffers
several critical drawbacks as discussed in the last part of
section III. As an alternative, DC performs locally. A BP
articial neural network is built and the gradient descending
algorithm is incorporated to develop the control strategy of
DC. Simulation results reveals that DC assures similar steady
E
SR
as CC while DC is less sensitive towards packet losses
than CC.
REFERENCES
[1] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. Wireless
sensor networks: A survey. Computer Networks, 38(4):393422, 2002.
0 5 10 15 20 25 30 35 40 45 50
0
20
40
60
80
100
k
E
S
R

(
%
)
=50%
=40%
=20%
=0%
(a) CC with 50%.
0 5 10 15 20 25 30 35 40 45 50
0
20
40
60
80
100
k
E
S
R

(
%
)
=60%
=50%
=40%
=20%
=0%
(b) DC
Fig. 6. Performances of the control systems with regard to various packet
loss ratios , when = 0.8, = 0.5, x [30, 30], and all the noises
power equal 0.
[2] E.M. Petriu, N.D. Georganas, D.C. Petriu, D. Makrakis, and V.Z.Groza.
Sensor-based information appliances. IEEE Instrumentation and Mea-
surement Magazine, 3(4):3135, 2000.
[3] J. S. Sandhu, A. M. Agogino, and A. K. Agogino. Wireless sensor
networks for commercial lighting control: Decision making with multi-
agent systems. In Proceedings of AAAI-04, July 2004.
[4] R. Kumar, A. Kansal, and M. Srivastava. Distributed control applications
over ad-hoc sensor actuator networks. Technical Report of NESL,
UCLA, Jan. 2004.
[5] T. Melodia, D. Pompili, and I. F. Akyildiz. A communication architec-
ture for mobile wireless sensor and actor networks. In Proceedings of
IEEE SECON, Reston, VA, USA, September 2006.
[6] X. Liu and A. Goldsmith. Wireless communication tradeoffs in dis-
tributed control. In 42nd IEEE CDC, volume 1, 2003.
[7] B. Sinopoli, C. Sharp, L. Schenato, S. Schaffert, and S. S. Sastry.
Distributed control applications within sensor networks. Proceedings
of the IEEE, 91(8):12351246, August 2003.
[8] Y. Lin and S. Megerian. Low cost distributed actuation in large-scale
ad hoc sensor-actuator networks. In Proceedings of IEEE WirelessCom,
volume 2, June 2005.
[9] M. Coates, F. Thouin, and R. Thommes. Optimal actuation strategies
for sensor/actuator networks. In Proceedings of IWASN, San Jose, CA,
USA, July 2006.
[10] C.J. Rozell and D.H. Johnson. Evaluating local contributions to global
performance in wireless sensor and actuator networks. In Proceedings
of the International Conference on Distributed Computing in Sensor
Systems (DCOSS), San Francisco, CA, June 2006.
[11] B. P. Gerkey and M. J Mataric. A market-based formulation of sensor-
actuator network coordination. In Proceedings of the AAAI-02, pages
2126, Palo Alto, CA, USA, 2002.
[12] G. Welch and G. Bishop. An introduction to the kalman lter. Technical
report, Technical report TR 95-041, Department of Computer Science,
University of North Carolina, 2002.
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the ICC 2008 proceedings.

Potrebbero piacerti anche