On the Planning of Wireless Sensor Networks:

Energy-Efficient Clustering under the Joint

Routing and Coverage Constraint
Ali Chamam, Student Member, IEEE, and Samuel Pierre, Senior Member, IEEE
AbstractMinimizing energy dissipation and maximizing network lifetime are important issues in the design of applications and
protocols for sensor networks. Energy-efficient sensor state planning consists in finding an optimal assignment of states to sensors in
order to maximize network lifetime. For example, in area surveillance applications, only an optimal subset of sensors that fully covers
the monitored area can be switched on while the other sensors are turned off. In this paper, we address the optimal planning of
sensors states in cluster-based sensor networks. Typically, any sensor can be turned on, turned off, or promoted cluster head, and a
different power consumption level is associated with each of these states. We seek an energy-optimal topology that maximizes
network lifetime while ensuring simultaneously full area coverage and sensor connectivity to cluster heads, which are constrained to
form a spanning tree used as a routing topology. First, we formulate this problem as an Integer Linear Programming model that we
prove NP-Complete. Then, we implement a Tabu search heuristic to tackle the exponentially increasing computation time of the exact
resolution. Experimental results show that the proposed heuristic provides near-optimal network lifetime values within low computation
times, which is, in practice, suitable for large-sized sensor networks.
Index TermsWireless sensor networks (WSNs), coverage, clustering, routing, network lifetime, energy efficiency, optimization,
mathematical programming, Tabu search heuristic.

1 INTRODUCTION
W
IRELESS sensor networks (WSNs) consist of a large
number of limited capability (power and processing)
MicroElectroMechanical Systems (MEMS) capable of mea-
suring and reporting physical variables related to their
environment. In surveillance applications, sensors are
deployed in a certain field to detect and report events like
presence, movement, or intrusion in the monitored area [1].
As depicted in Fig. 1, data collected by sensors are
transmitted to a special node equipped with higher energy
and processing capabilities called processing node (PN)
or sink [1]. The PN collects, filters, and compiles data sent
by sensors in order to extract useful information. Due to
their energy constraints, wireless sensors usually have a
limited transmission range, making multihop data routing
toward the PN more energy efficient than direct transmis-
sion (one hop). Energy conservation in WSN is critical and
has been addressed by substantial research [2], [3].
Generally, energy conservation is dealt with on five
different levels [1], [2]:
1. efficient scheduling of sensor states to alternate
between sleep and active modes;
2. energy-efficient routing, clustering, and data aggre-
gation;
3. efficient control of transmission power to ensure an
optimal trade-off between energy consumption and
connectivity;
4. data compression (source coding) to reduce the
amount of uselessly transmitted data;
5. efficient channel access and packet retransmission
protocols on the Data Link Layer.
The scope of this paper includes both the first and the
second levels. We address the global problem of maximiz-
ing network lifetime under the joint clustering, routing, and
coverage constraint. We consider a sensor network that is
deployed in a certain area to monitor some given events.
When the network is dense, sensing ranges of neighbor
sensors usually overlap. This means that when an event
occurs at a point 1 of , it will be detected and reported by
all the sensors whose sensing range encompasses 1. This
redundant transmission results in useless energy consump-
tion. To save network energy and increase its lifetime, we
propose to switch on only a subset of sensors that covers
while all other sensors are turned off. Fig. 2 depicts an
example of full-covering sensor set. On the other hand,
clustering has been proven energy efficient in WSN [4], [5],
[6]. In cluster-based WSN, sensors are organized in clusters
each having one sensor promoted as CH. All non-CH nodes
transmit their data to their CH, which routes it to the remote
PN. Clustering can provide for substantial energy saving
[4], [5], [6], [7] since only CH sensors are involved in routing
and relaying data. Moreover, clustering alleviates band-
width, enables its reuse, and can, thus, increase system
capacity [8]. Besides, the fact that only the CH is transmit-
ting information out of the cluster helps avoid collisions
between the sensors inside the cluster and helps avoid the
IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 8, NO. 8, AUGUST 2009 1077
. The authors are with the Department of Computer Engineering, Ecole
Polytechnique of Montreal, C.P. 6079, Succursale Centre-Ville, Montreal,
QC H3C 3A7, Canada. E-mail: {ali.chamam, samuel.pierre}@polymtl.ca.
Manuscript received 26 June 2008; revised 20 Nov. 2008; accepted 9 Dec.
2008; published online 6 Jan. 2009.
For information on obtaining reprints of this article, please send e-mail to:
[email protected], and reference IEEECS Log Number TMC-2008-06-0251.
Digital Object Identifier no. 10.1109/TMC.2009.16.
1536-1233/09/$25.00 2009 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS
uncovered hole problem [9]. However, since CHs consume
more energy in aggregating and routing data, it is
important to have an energy-efficient mechanism for CHs
election and rotation [8], [10]. In flat networks, sensors route
data to the PN through their peer sensors using one of the
many routing protocols proposed in the literature [4]. In
contrast, in clustered networks, CHs transmit aggregated
data to the PN, either directly (one hop) [10] or in multihop
[11]. For the sake of minimizing energy consumption,
both optimal number and optimal placement of CHs have
to be sought.
In a cluster-based monitoring sensor network, any
energy-efficient sensor scheduling mechanism has to
guarantee a certain area coverage rate. Besides, the
connectivity of every sensor to a CH has to be ensured at
any time. Furthermore, for data to be routed from any CH
to the PN, all CHs have to belong to a single connected
graph. Hence, for sensors states allocation to be optimal,
coverage, connectivity of sensors to CHs, and routing have
to be taken into account within the same global planning
process. When coverage and connectivity are dealt with
separately, the obtained configuration may not be optimal.
For example, an optimal covering subset of sensors can fail
to guarantee network connectivity because some nodes are
switched off or the optimally designated CHs may belong
to the set of switched-off sensors.
Many papers addressed separately energy-efficient rout-
ing [4], clustering [5], [6], and area coverage [2], [12], [13],
[14]. Many other works [15], [16] addressed the integrated
problem of maintaining area coverage and network con-
nectivity but only on flat networks and did not take
advantage of the potential energy saving and ease of
manageability of cluster-based networks [1], [6], [7]. To the
best of our knowledge, the problem of maximizing sensor
network lifetime under the integrated constraint of cluster-
ing, coverage, and routing has not been addressed within
the same global optimization process. In this paper, we
address the optimal planning of cluster-based WSN under
the joint routing and coverage constraint. In our architec-
ture, any sensor can be active, switched off, or upraised as
CH, and only CHs can route data. We seek an optimal
allocation of states to sensors, which maximizes network
lifetime, while ensuring simultaneously full area coverage,
connectivity of every sensor to a CH, and connectivity of the
overlay network composed of CHs.
This paper is organized as follows: In the next section,
we present some related work. In Section 3, we outline our
problem and enumerate our assumptions. In Section 4, we
mathematically formulate our problem as an Integer Linear
Programming (ILP) model. In Section 5, we present our
resolution method based on a Tabu search algorithm. In
Section 6, we discuss our simulation results. Finally,
Section 7 concludes this paper and points out some
future directions.
2 RELATED WORK
The problemof maintaining both area coverage and network
connectivity under energy constraint in WSN has been
extensively addressed in the literature and many protocols
were proposed to alternate sensor states between active and
sleep in order to maximize network lifetime. For example,
Xing et al. [15] provide a geometric analysis of the relation-
ship between coverage and connectivity, and propose the
Coverage Configuration Protocol (CCP) that dynamically
configures the network to guarantee different degrees of
coverage depending on the application requirements. In
CCP, every node decides its state (Active or Sleep) based on
the coverage degree of the intersection points of its sensing
circle with those of its neighbors. When coupled with any
connectivity maintenance protocol, CCP offers connectivity
and 1-coverage. Lu et al. [17] present Scalable Coverage
Maintenance (SCOM), a localized coverage maintenance
algorithmwhere sensors use the same redundancy eligibility
rule as in [15] to decide whether to turn on or turn off. SCOM
implements, for each sensor, a back-off timer proportional to
its residual energy. The back-off timer allows sensors with
lower residual energies to decide about their states before
sensors with more energy, making them more likely to turn
off than the other sensors, if they find themselves redundant.
Chamam and Pierre [18] propose a centralized heuristic
which dynamically calculates a near-optimal subset of
sensors that guarantees a predefined coverage rate while
ensuring network connectivity when the transmission range
is greater than or equal to twice the sensing range. Yan et al.
[16] propose to schedule sensors activities (Active/Sleep) so
that every point in a grid-monitored area is covered at any
time. Neighbor sensors exchange a random reference time
T
ic)
within cyclic rounds of constant duration T and decide
to be active for a certain time duration within T. The round
period T is equally shared among all the neighbor sensors
that cover a common grid. Even though the proposed
schedule balances consumed energy over neighbor nodes, it
does not take into account the residual energy of sensors
1078 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 8, NO. 8, AUGUST 2009
Fig. 1. Multihop routing of collected data.
Fig. 2. A full-covering sensor set.
when calculating the activation time period of every node,
which makes nodes with less residual energy more prone to
expiration. However, all the works cited above do not
address cluster-based architectures. Cluster formation is
typically based on the energy reserve of sensors and sensors
proximity to the cluster head [5], [6]. Energy-efficient
cluster-based routing algorithms for WSN have been widely
addressed in literature [4]. For instance, Low-Energy
Adaptive Clustering Hierarchy (LEACH) [10], one of the
most popular hierarchical routing algorithms for WSN,
proposes to form clusters of sensor nodes based on the
received signal strength and use local cluster heads as
routers to the sink. This saves energy since the transmissions
will only be operated by cluster heads rather than all sensor
nodes. Even though LEACH is completely distributed, it
uses single-hop communication between cluster heads and
the sink, which is energy consuming and not applicable to
networks deployed in large regions. Power-Efficient GAth-
ering in Sensor Information Systems (PEGASIS) [19] and its
variant Hierarchical-PEGASIS are two improvements of
LEACH. Rather than forming multiple clusters, PEGASIS
forms chains of sensor nodes so that each node transmits to
and receives from a neighbor and only one node is selected
from that chain to transmit to the PN. But still, communica-
tion between the elected CH and the PN is made in one hop,
which is not suitable for large networks. Energy-efficient
sensor state scheduling mechanisms in cluster-based WSN
also raised much interest in the research community. For
example, Yao and Giannakis [8] proposed a scheduling
algorithm for the one-level-clustered WSN, where sensors
have different data sequence lengths to transmit within a
period of time T. The authors propose an Inverse-Log
algorithm that finds, for every sensor, a set of optimal
time allocations that minimizes the dissipated energy of the
whole network over the period T. However, all sensors
are activated during the time period T and no coverage
constraint is considered. Besides temporal scheduling, other
publications propose a spatial scheduling scheme based on
the selective activation of sensors to maximize network
lifetime [2], [20], [21]. Tian and Georganas [21] propose a
localized algorithm that finds an optimal subset of sensors
ensuring full area coverage or, if not possible, the least
uncovered points. If the whole range covered by a sensor is
covered by a subset of its neighbors, then the sensor decides
to turn off. A random back-off time ensures that two nodes
do not make the decision to turn off at the same time. When
implemented over LEACH [10], the protocol proposed in
[21] shows some energy saving. However, this protocol is
not optimal because of the uncontrolled coverage redun-
dancy due to the random aspect of switch on/off of sensors.
In [20], sensors sensing ranges follow a certain distribution
derived from the channel characteristics and the log-normal
path loss. An event occurring outside a certain range is still
detected with a corresponding probability. The cumulative
detection probabilities are shown to increase the mean area
coverage which obviously decreases the number of sensors
activated within the covering subset, thus reducing the
consumed energy. However, the proposed algorithm is not
optimal because it only takes into account each sensors
range and not its residual energy making nodes with very
little residual energies prone to expiration. Hwang et al. [9]
propose a cluster-based coverage-preserved node schedul-
ing scheme. This mechanism assumes a dense network and
assigns states (Active, Sleep, Cluster head) to sensors in a
distributed and self-organized manner. The algorithm starts
by dividing sensors into clusters and defines, for each
cluster, a number of sponsor sensor sets that may be turned
on simultaneously. Only one among the computed sponsor
sets is elected to be turned on until it completely runs out of
energy. Even though Hwang et al. [9] provide an efficient
coverage and clustering mechanism, they assume that
sensed data are routed to the sink in one hop, which may
be energy intensive for the relay nodes.
3 PROBLEM STATEMENT AND ASSUMPTIONS
In this paper, we consider a WSN deployed in an area to
monitor certain critical activities or events. As shown in [3],
for the case of Rockwells WINS seismic sensors, a sensors
radio can be in one of the following four activity modes,
characterized by their respective power consumptions:
Transmit (0.38-0.7 W), Receive (0.36 W), Idle (0.34 W), and
Sleep (0.03 W). We note that when sensors are transmitting,
receiving, or idle, they have roughly the same energy
consumption and can then be associated to a same state,
Active, in which the sensors radio is switched on.
Moreover, in this paper, we consider a cluster-based
topology in which CHs route the data they receive from the
non-CH sensors of their cluster to the PN through an
overlay network solely composed of CHs. CH election and
cluster formation are very important issues that deeply
affect network lifetime of WSN and different approaches
exist to implement these stages. For example, it is possible
to use a fixed distribution of the sensor nodes and CHs, or
use a dynamic algorithm for CH election. If CHs were
chosen a priori and fixed throughout the system lifetime,
they would quickly exhaust all their energy making them
no longer operational. In this work, we propose to
dynamically designate the set of CHs according to their
residual energies, their distance to their neighboring non-
CH active nodes, and their position within the graph
formed by CHs. As depicted in Fig. 3, we will consider,
without loss of generality, that each sensor can be in one of
the three states: Sleep, Active, and Cluster Head (CH) having,
respectively, power consumptions 1
o|ccj
, 1
cti.c
, and 1
CH
per time unit, where 1
o|ccj
( 1
cti.c
< 1
CH
.
To control energy dissipation of the sensors that perform
data relaying, we restrict the routing task to CHs. For this,
there must exist a route from any CH to the PN. The most
straightforward solution for that is to have a connected
graph linking all the CHs. In our problem modeling, we
propose that any admissible configuration must exhibit a
spanning tree connecting all CHs, as shown in Fig. 4. This
CHAMAM AND PIERRE: ON THE PLANNING OF WIRELESS SENSOR NETWORKS: ENERGY-EFFICIENT CLUSTERING UNDER THE JOINT... 1079
Fig. 3. Sensor states.
makes the overlay network formed by the CHs sufficient to
route data from any sensor toward the PN. The spanning
tree construction is explained in the next section.
On the other hand, we assume that all sensors have the
same sensing range 1
:
and that their detection model
follows a binary probability function, also called Disk model
[2], [15], in which an cti.c sensor i detects any event that
occurs within its sensing range 1
:
with the same probability
1
d
1, whatever the distance from the event is. And any
event occurring outside 1
:
is not detected by i. Let us note
that this binary probability sensing model is a simplified
representation of a sensors detection probability and it
does not necessarily reflect the received signal loss model
on a real wireless channel. In fact, on a real wireless
channel, because of signal fading and interference, events
occurring far away from a sensor are less likely to be
detected than events occurring close to it. In other words, an
event occurring outside a sensors range is still detected
with a certain probability which decreases when the
distance to the sensor increases. Many papers in the
literature addressed probabilistic coverage in wireless
sensor networks [20], [22] and proposed coverage protocols
based on this model. However, to simplify the representa-
tion of the sensing model, substantial work in the literature
[2], [15], [23] make the assumption of disk model. In this
paper, we also make this simplifying assumption but we
will show later that our optimization model and our
proposed heuristic can be easily adapted to handle
probabilistic coverage.
Definition 1. A set of sensors o
c
is a covering set of area if and
only if 8 point 1 2 . 9 i 2 o
c
such as i covers 1.
In critical surveillance applications, it is important to
guarantee that the monitored area is fully covered by
sensors at every instant of the network lifetime. Hence, in
our problem, the optimal network configuration must
. contain a full-covering set of active sensors;
. contain a set of CHs so that every sensor is
connected to a CH;
. ensure that all CHs belong to a spanning tree over
which data will be routed toward the PN.
Our objective is to find the network-lifetime-optimal
allocation of sensors states (Active, Sleep, CH) that meets
these three conditions. Before modeling our problem, we
make the following assumptions:
1. Each sensor has a unique ID, known to the PN and
to the sensor itself.
2. The position of each sensor is fixed and known to the
PN. The location information can be obtained either
through a Global Positioning System (GPS), as
assumed in [2] (but this technique is still expensive
due to the high cost of placing a GPS on each sensor),
or using one of the many GPS-free localization
techniques proposed in the literature [24], [25], [26].
However, in this paper, we do not specifically
address any localization technique and assume that,
whatever the localization technique used, sensors
location information is available at the PN.
3. Active sensors capture events occurring in their
sensing range and transmit data associated with
these events straightaway, without any buffering,
because sensors are usually not equipped with large
(and costly) buffers.
4. All sensors have the same sensing range 1
:
and the
same transmission range 1
t
. All CHs have the
same transmission range 1
CH
t
1
t
.
5. Only the CHs can perform data routing. Routing
over the overlay network composed of CHs can be
performed using one of the energy-efficient routing
protocols for WSN proposed in the literature [4].
However, we do not address any specific routing
protocol, we only guarantee the existence of a
routing topology.
6. Each sensor has an initial energy 1
0
. The PN has no
energy limitation. Besides, we assume that, when a
sensor is cti.c, it has a constant energy dissipation
during a unit of time, no matter how the events
distribution is.
7. The network is dense enough so that when all the
sensors are Active, the monitored area is fully
covered. Besides, we assume that the graph repre-
senting the sensor network is connected (two sensors
being connected when they are within the transmis-
sion range of each other).
8. Network lifetime is defined as the time separating
the instant the network starts operating and the
instant at which the network cannot be covered
anymore because of the expiration of some nodes.
9. We assume ideal MAC layer conditions, i.e., perfect
transmission of data on a node-to-node wireless link.
10. We assume that sensors have ideal sensing capabil-
ities, i.e., inside the sensing range, the quality of
sensing does not depend on the distance from
the sensor.
4 PROBLEM MODELING
Our problem consists in finding the optimal allocation of
states to sensors, which maximizes network lifetime under
the integrated constraint of coverage, clustering, and
routing. We call this problem OPT-ALL-RCC. To maximize
network lifetime, we need a trade-off between total energy
consumption and energy balancing among sensors. For
example, to ensure area coverage, we would prefer to
activate more sensors having higher residual energy (and
consuming a higher total energy) than few sensors having
little residual energy (consuming less total energy but more
prone to expiration). On the other hand, any admissible
1080 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 8, NO. 8, AUGUST 2009
Fig. 4. Overlay network composed of the selected cluster heads.
solution of our model has to ensure full coverage of the
monitored area and the existence of a spanning tree
connecting all CHs. To model this problem, we first define
the following sets and constants:
. Let o f1..`g be the set of sensors.
. Let C f1..`g be the set of cells composing the
monitored area.
. Let 8 i 1..joj. 8 , 1..joj. 8 c 1..jCj,
d
i,

1 if sensor i can reach sensor , in one hop
0 else.

1
i,

1 if CH i can reach CH , in one hop
0 else.

,
ic

1 if sensor i covers cell c
0 else.

Then, we define our binary decision variables 8 i 1..joj,

8 , 1..joj:
A
i

1 if sensor i is cti.c
0 else.

Y
i

1 if sensor i is a CH
0 else.

7
i,

1 if sensor i is connected to CH ,
0 else.

\
i,

1 if CH i is connected to CH , within
a spanning tree
0 else.
8
>
<
>
:
To balance energy consumption among nodes, we
choose to minimize an objective function that is a linear
combination of sensors scores. The score of a sensor i is
defined by
ocoici 1oq 1
1d
i
1i
i

.
where
1d
i

1
cti.c
if sensor i is cti.c but not CH
1
CH
if sensor i is CH
1
o|ccj
% 0 else.
8
<
:
As shown in Fig. 5, the logarithmic nature of this
score function will tend to accentuate the importance of the
residual energy when the latter is small and will give more
importance to the dissipated energy when the residual
energy is high enough. In other words, when the residual
energy is low, sensors will be selected essentially according
to their residual energies, favoring the activation of sensors
having relatively high residual energy and when the
residual energy is relatively high, the optimal solution will
tend to activate as less sensors as possible. Thus, we can
model our problem by the following optimization system:
Minimize:
X
joj
i1
Y
i
.1oq 1
1
CH
1i
i

A
i
Y
i
.1oq 1
1
cti.c
1i
i

. 1a
Subject to:
8 c 1..jCj.
X
joj
i1
A
i
.,
ic
! 1. 1b
8 i 1..joj. Y
i
A
i
. 1c
X
joj
i1
Y
i
.d
i0
! 1. 1d
8 i 1..joj. 8 , 1..joj. , 6 i. 7
i,
A
i
Y
i
. 1e
8 i 1..joj. 8 , 1..joj. , 6 i. 7
i,
Y
,
. 1f
8 i 1..joj. 8 , 1..joj. , 6 i. 7
i,
d
i,
. 1g
8 i 1..joj.
X
joj
,1
,6i
7
i,
Y
i
A
i
. 1h
8 , 1..joj.
X
joj
i1
,6i
7
i,
`
ior
. 1i
8 i 1..joj. 8 , 1..joj. , 6 i. \
i,
Y
i
. 1j
8 i 1..joj. 8 , 1..joj. , 6 i. \
i,
Y
,
. 1k
8 i 1..joj. 8 , 1..joj. , 6 i. \
i,
1
i,
. 1l
8 H o. 8 : 2 H.
X
,2H
X
/2H
/,
\
,/

X
i2H
Y
i
!
Y
:
. 1m
X
,2joj
X
/2o
/,
\
,/

X
,2o
Y
,
1. 1n
X. Y 2 f0. 1g
joj
. Z. W2 f0. 1g
joj
2
. 1o
The expression of the objective function (1a) aims at
balancing the energy consumption over the network. Equa-
tions (1b)-(1o) are the model constraints. Constraint (1b)
CHAMAM AND PIERRE: ON THE PLANNING OF WIRELESS SENSOR NETWORKS: ENERGY-EFFICIENT CLUSTERING UNDER THE JOINT... 1081
Fig. 5. Shape of the score function.
guarantees a full coverage of the monitored area such that
every elementary cell is covered by at least one Active sensor.
Let us note here that this coverage constraint is valid under
the assumption of binary probability (disk) sensing model,
where every cell is either covered by a sensor (if it is located
inside its sensing range) or not covered at all (if it is outside
that sensing range). With probabilistic coverage, the cover-
age constraint must ensure that every cell is covered up to a
certain predefined coverage rate. Hence, the constraint (1b)
should be replaced with
8 c 1..jCj. 1io/ c i: co.cicd ! co. iotc. 2
where co. iotc is a predefined threshold coverage rate. Let
1 f1
ic
g be a coverage probability matrix known a priori,
where 1
ic
probability that sensor i covers cell c. Then, (2)
is equivalent to
8 c 1..jCj. 1
Y
i1..joj
1 A
i
.1
ic
! co. iot. 3
Once linearized, the new coverage constraint (3) will
replace constraint (1b) in the above model to have it handle
probabilistic coverage.
The rest of our constraints are as follows: constraint (1b)
ensures that there exists at least a CH located one hop away
from the PN. Constraints (1e)-(1h) ensure that every Active
and non-CH sensor is connected to at least one CH within
its range. Constraint (1i) gives an upper bound on clusters
sizes. Equations (1j)-(1n) describe the routing constraint
ensuring that the overlay network composed of CHs is
connected, and hence, there exists a tree-like partial
subgraph. Equations (1o) are the integrality constraints.
To ensure that a spanning tree connecting all the CHs exists
in any solution, constraints (1m) and (1n) require the
enumeration of all the subsets of o. Even though these
constraints represent the theoretical conditions to have a
spanning tree in any graph (no cycles and a connected
graph), they quickly result in a combinatorial explosion of
the number of constraints due to the exponentially increas-
ing number of subsets of o. To circumvent this problem, we
will proceed differently: we represent the routing constraint
of our problem as a multiflow routing problem. We
consider that a virtual flow has to be routed from any CH
to, at least, one CH which is one hop from the sink. Indeed,
the optimal graph configuration that allows a flow to be
routed between any pair of nodes of a connected graph
where links have infinite capacity is the minimal-cost
spanning tree. To model this virtual flow routing problem,
we define a binary variable representing the use of the
wireless link |/ to convey a flow i. ,, where i. ,. /, and | are
CHs and i. , are, respectively, the source and destination of
the flow. Let
8 i. ,. /. | 2 f1..jojg. i 6 ,. / 6 |.
\
/|
i,

1 if flow i. , passes through the link /|
0 else.

4
The following constraints ensure that the network
contains a spanning tree connecting all CHs:
8 i 1..joj. 8 , 1..joj. , 6 i. 8 / 1..joj.
P
|2o
|6/
\
/|
i,

P
|2o
|6/
\
|/
i,

0. if / 6 i and / 6 ,.
Y
i
.Y
,
. if / i.
Y
i
.Y
,
. if / ,.
8
<
:
5a
8 i 1..joj. 8 , 1..joj. , 6 i. 8 / 1..joj.
8 | 1..joj. \
/|
i,
Y
i
.
5b
8 i 1..joj. 8 , 1..joj. , 6 i. 8 / 1..joj.
8 | 1..joj. / 6 |. \
/|
i,
Y
,
.
5c
8 i 1..joj. 8 , 1..joj. , 6 i. 8 / 1..joj.
8 | 1..joj. | 6 /. \
/|
i,
Y
/
.
5d
8 i 1..joj. 8 , 1..joj. , 6 i. 8 / 1..joj.
8 | 1..joj. | 6 /. \
/|
i,
Y
|
.
5e
8 i 1..joj. 8 , 1..joj. , 6 i. 8 / 1..joj.
8 | 1..joj. | 6 /. \
/|
i,
1
/|
.
5f
V 2 f0. 1g
joj
4
. Y 2 f0. 1g
joj
. 5g
Equation (5a) is the flow constraint ensuring that a feasible
path exists between any pair of CHs to convey an
elementary unit of flow. Remaining constraints (5b)-(5e)
limit the relevance of this virtual flow problem to the
overlay network. Constraint (5f) ensures that CHs / and |
are neighbors for a flow to pass on the link /|. Finally, (5g)
are the integrality constraints. In (5a), we have a nonlinear
term that we need to linearize. For this, we define
8 i 1..joj. 8 , 1..joj. l
i,
Y
i
.Y
,
. 6a
To have a logical equivalence between l
i,
and A
i
.Y
,
, we
add the following constraints:
8 i 1..joj. 8 , 1..joj. l
i,
Y
i
. 6b
8 i 1..joj. 8 , 1..joj. l
i,
Y
,
. 6c
8 i 1..joj. 8 , 1..joj. l
i,
! Y
i
Y
,
1. 6d
Taking into account the virtual-flow-related constraints
(5a)-(5f) and the linearized constraints (6a)-(6d), we end up
with a linearized model of OPT-ALL-RCC.
Proposition 1. OPT-ALL-RCC is NP-Complete.
Proof. To prove the NP-Completeness of OPT-ALL-RCC, we
will derive a polynomial reduction to the set covering
problem which is known to be NP-Complete [27]. For
this, we propose to define an instance 1 of the set
covering problem, build an instance 1 of OPT-ALL-RCC,
and then show that any algorithm that resolves 1 is able
to resolve 1.
Any instance 1 of the set covering problem is defined
by a set of nodes, a set of node subsets, and a cost for
each node subset. Let o f:
1
..:
`
g be a set of ` sensors
(nodes) and J fo
1
..o
`
g be a set of ` node subsets
1082 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 8, NO. 8, AUGUST 2009
(o
,
o. , 1..`) such that 8 , 1..`. o
,
is built from
the empty set as follows:
. The node :
,
is inserted into o
,
.
. All nodes :
i
2 o f:
,
g that are one-hop neigh-
bors of :
,
(i.e., d
:,:i
1) are added to o
,
.
For each , 1..`, we define the cost of subset o
,
by
c, 1oq1 1
CH
,1i
,
. The instance 1 being defined,
let us consider the instance 1 of OPT-ALL-RCC, with the
following input:
. `
ior
= Infinity.
. Each cell of the monitored area is covered by one
and only one sensor.
. The residual energy of nodes :
i
2 o is equal to
1
0
1
CH
1
cti.c
.
. The transmission power of a CH is such that any
potential CHs can reach each other, that is,
8 ,. i 2 f1..`g. , 6 i, if :
,
and :
i
are elected
CH, then 1
:,:i
1.
Since each sensor covers a single cell, any admissible
solution will have all its sensors turned on to satisfy the
full coverage constraint. We clearly see that any algo-
rithm that is able to resolve the above instance 1 of OPT-
ALL-RCC can resolve any instance 1 of the set covering
problem. Indeed, for every sensor :
/
1 / `, desig-
nated as cluster head in the solution of 1, o
/
is a member
of the minimum-cost covering set. Such a reduction
proves that OPT-ALL-RCC is NP-Complete. tu
5 PROPOSED HEURISTIC
As the considered problem is NP-Complete, we propose
a Tabu search heuristic, called TABU-RCC, to tackle
to exponentially increasing processing time of the
exact solution. TABU-RCC will be run by the PN to find a
near-optimal sensor state configuration. As shown in
Algorithm 1, TABU-RCC starts with an admissible solution
and iteratively performs movements that consist in chan-
ging the state of one sensor at a time. The best solution
found after the predefined number of iterations is trans-
posed on sensors to form the new network configuration.
The network will operate with this configuration for a
predefined period T during which residual energies of
active nodes and CHs will decrease, then TABU-RCC is run
again to find a new configuration based on the new values
of residual energies. This new configuration will be kept for
another period T and so forth. The periodic execution of
TABU-RCC by the PN requires sensors-related information
(e.g., residual energies) to be transmitted periodically to the
PN (upstream communication) and the newly computed
sensor states to be transmitted to the sensors (downstream
communication). In our architecture, sensor-related data
will be collected exactly in the same manner as the sensed
data, i.e., using the cluster-based hierarchical structure of
the network. Active sensors that have data to report will
send it to the PN via their respective CHs. They will append
the value of their respective residual energies to the data
packets they are sending. When they have no data to send,
they will synchronize their energy information with the PN
periodically. As for the turned-off sensors, their energy
consumption is constant and very low, and can therefore be
estimated by the PN whenever needed, as long as they are
idle. As far as state assignment to sensors is concerned, the
PN will broadcast a notification message holding associa-
tions between sensors and their newly computed states
(Sensor ID, Sensor State). Only those sensors whose state has
changed will have an entry in the notification message
while all other sensors will keep their ongoing state. The
notification message will be routed to the sensors via their
respective CHs.
Algorithm 1: TABU-RCC: Tabu search algorithm of CH
election under routing and coverage constraints
. Initial solution: the Tabu algorithm starts with a
configuration where all sensors are activated as cluster
heads. This configuration is obviously admissible;
. Admissible configuration: a configuration o is defined
by the states of its sensors (o|ccj, cti.c, or CH). Only
feasible configurations (i.e., satisfying model
constraints (1b) to (1i)) are considered;
. Score function: a configuration is evaluated using the
score function given by (1a);
. Neighborhood investigation: a search movement
` < i. n. . consists in changing the state (Sleep,
Active, or CH) of a single sensor i from state n to
state . such that the model constraints (1b) to (1i) are
satisfied;
. Aspiration criterion: Tabu movements are allowed
when the score of the resulting configuration is lower
than the score of the best solution :

found so far over

the whole search process;
. Stop criterion: The search algorithm stops after a
predefined number of iterations.
6 SIMULATION RESULTS
In this section, we will evaluate the quality of TABU-RCC
with respect to the exact (optimal) solution which provides
a lower bound of the objective function. To find the optimal
solution, we implement our linear integer-variable model
using ILOG CPLEX [28]. CPLEX is a mathematical
programming and optimization tool that solves linear
problems with continuous, integer, or mixed variables,
using the branch-and-bound method.
Then, we draw the variation of TABU-RCCs network
lifetime when, respectively, the sensing range and the
cluster size vary. These measures will show how worth it is
to invest in sensors with higher sensing range or higher
processing capacity.
Finally, we compare our Tabu search algorithm to EESH
[29], a clustering algorithm recently proposed in the
literature. EESH elects the set of cluster heads and ensures
network connectivity. However, EESH does not handle
optimal area coverage. To tackle this issue, we propose an
enhanced version of EESH that computes an optimal area-
covering subset of nodes (among the noncluster head
nodes) by resolving the linear system proposed in [18].
The remaining nodes are then switched off. Our choice of
EESH as a comparison reference for our algorithm was
CHAMAM AND PIERRE: ON THE PLANNING OF WIRELESS SENSOR NETWORKS: ENERGY-EFFICIENT CLUSTERING UNDER THE JOINT... 1083
motivated by the fact that EESH is one of the most recent
clustering algorithms proposed in the literature and that
EESH has been shown to outperform, in terms of network
lifetime (which is our main objective to maximize), two of
the most popular and well-recognized clustering algorithms
for WSN: LEACH [10] and HEED [30]. For all the
simulations made, TABU-RCC was run with a tabu list
size of 15 and a number of iterations of 10,000.
6.1 Comparative Performance Evaluation: TABU-
RCC With Respect To Its Lower Bound (CPLEX)
To evaluate the quality of our heuristic, we considered
different network sizes of 9, 25, 49, 64, 100, and 225 sensors
and compared the objective function provided by TABU-
RCC to the lower bound provided by CPLEX. The measure
of the objective functions has no practical interest but is
only useful to assess the quality of the heuristic with respect
to the exact solution (lower bound). It was difficult to run
simulations on bigger network sizes because of the
exponentially increasing processing time of CPLEX, as
shown in Table 1. In fact, we can clearly see that, while
TABU-RCC always converges in few seconds, CPLEXs
computation time grows exponentially, which deeply
justifies the use of a heuristic. Let us note here that for the
network sizes of 9, 25, and 49 sensors, the lower bound used
is the value of the objective function given by the solution of
the integer-variable system 1, while for network sizes of 64,
100, and 225 sensors, the resolution of the integer-variable
system could not be made in a reasonable time, so we made
a continuous relaxation of the system 1 and we used its
solution to evaluate the solution provided by TABU-RCC.
In fact, the relaxed system provides an obvious lower
bound and it takes much less time to be resolved than the
original integer-variable system.
This explains why the lower bound processing time for
64 sensors reported in Table 1 seems lower than the
processing time of the lower bound for 49 sensors.
As shown in Fig. 6, the objective function of TABU-
RCCs solutions is very close to that of the lower bound.
Also, Fig. 7 shows that the average energy consumed by
TABU-RCCs solutions is very close to that of the lower
bound. Thus, we can infer that TABU-RCC provides good
solutions, close to their lower bound. Even though it was
predictable that an exact ILP resolution would not be
suitable for large-sized networks, we computed the objec-
tive function of CPLEXs solutions for small and medium-
sized networks only to assess the quality of the solutions
provided by TABU-RCC with respect to optimality.
6.2 Performance Evaluation of TABU-RCC:
Impact of the Sensing Range
In practice, the value of the sensing range could vary
depending on the physical properties of the sensors and on
the type of signal they are sensing. To measure the impact of
the sensing range, we considered three network sizes of,
respectively, 225, 625, and 900 sensors. To measure the
network lifetime, we run TABU-RCCiteratively and at every
iteration, a sensors energy is decremented by 1
CH
if it is
designated as CH and by 1
cti.c
if it is assigned the state
cti.c. Iterations stop when the monitored area cannot be
covered anymore because of the energy exhaustion of
some sensors. Fig. 8 shows the variation of the network
lifetime with the sensing range, for the considered networks.
We see that, for all network sizes, when 1
:
increases, the
network lifetime increases as well. This is due to the fact that,
when1
:
increases, less sensors are activatedtoensure the full
area coverage and the connectivity. Hence, less energy is
dissipated, which, in average, increases network lifetime.
6.3 Performance Evaluation of TABU-RCC:
Impact of the Maximum Cluster Size
In practice, a CH does not have an infinite capacity due to
many physical factors like interference, limited CHs
processing capabilities, collisions on the MAC layer, etc.
To evaluate the impact of the maximum cluster size on the
performance of our heuristic, we measured the network
lifetime for different network sizes of, respectively, 225, 625,
1084 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 8, NO. 8, AUGUST 2009
TABLE 1
Computation Time of the
Lower Bound (CPLEX) and TABU-RCC
Fig. 6. Value of the objective function generated by TABU-RCC versus
CPLEX.
Fig. 7. Average consumed energy using the solution generated by
TABU-RCC versus CPLEX.
and 900 sensors. In Fig. 9, we observe that, for all
configurations, when `
ior
increases, the network lifetime
increases as well. This is due to the fact that, when `
ior
increases, less CHs are needed to ensure connectivity since
a CH can serve more sensors. The leap up in network
lifetime becomes salient when the cluster size tends to
infinity. In fact, when `
ior
! 1, one single CH can ensure
the connectivity of all the sensors that can reach it.
6.4 Comparative Performance Evaluation:
TABU-RCC versus EESH
To evaluate our TABU-RCC in terms of network lifetime,
we compared it to EESH [29]. EESH functions as follows:
nodes are promoted cluster heads according to their
respective residual energies, their respective degrees, the
distance to their neighbors, and the residual energies of
these neighbors. For that, EESH evaluates a cost function for
every sensor in the network and iteratively elects the node
having the greatest cost as CH. This process terminates
when all non-CH sensors in the network are connected to at
least one cluster head. As EESH does not consider optimal
area coverage, we made a small modification of EESH that
consists of computing, in each cluster provided by EESH, an
optimal subset of non-CH sensors that will be activated
while the other sensors are turned off. Fig. 10 depicts the
network lifetime provided, respectively, by TABU-RCC,
EESH, and the modified version of EESH, for different
network sizes. We clearly see that TABU-RCC outperforms
the modified version of EESH by providing three times
longer network lifetime, for all network sizes. This
considerable gain is one of the main benefits of our
centralized heuristic.
7 CONCLUSION
In this paper, we proposed a novel centralized mechanism
for near-optimal state assignment to sensors in large-scale
cluster-based monitoring wireless sensor networks. Our
mechanism is based on a tabu algorithm that computes a
near-optimal network configuration in which each sensor
can be activated, put in sleep mode or promoted as cluster
head. Our mechanism maximizes network lifetime while
ensuring the full coverage of the monitored area and the
connectivity of the obtained configuration. Connectivity is
fulfilled through an optimally computed spanning tree
connecting all the cluster heads. Simulations show that our
mechanism provides for acceptable results with respect to
the exact solutions of the derived ILP model, within low
computation times. Despite its centralized aspect, our
mechanism exhibits low complexity and low computation
times making its practical implementation adaptable for
large-scale networks. As future research directions, we
intend to develop a more sophisticated heuristic to improve
the network lifetime. Furthermore, we intend to consider
distance-dependent probabilistic event detection, where the
probability that a sensor detects an event is function of the
distance of that sensor from the event. Furthermore, we
intend to work on distributed algorithms that address
energy-efficient clustering under the joint coverage and
routing constraint.
REFERENCES
[1] C.Y. Chong and S.P. Kumar, Sensor Networks: Evolution,
Opportunities, and Challenges, Proc. IEEE, vol. 91, no. 8,
pp. 1247-1256, Aug. 2003.
[2] M. Cardei, M.T. Thai, Y. Li, and W. Wu, Energy-Efficient Target
Coverage in Wireless Sensor Networks, Proc. IEEE INFOCOM,
vol. 3, pp. 1976-1984, 2005.
[3] V. Raghunathan, C. Schurgers, S. Park, and M.B. Srivastava,
Energy-Aware Wireless Microsensor Networks, IEEE Signal
Processing Magazine, vol. 19, pp. 40-50, 2002.
CHAMAM AND PIERRE: ON THE PLANNING OF WIRELESS SENSOR NETWORKS: ENERGY-EFFICIENT CLUSTERING UNDER THE JOINT... 1085
Fig. 8. Variation of the network lifetime with the sensing range, using
TABU-RCC.
Fig. 9. Variation of the network lifetime with the maximum cluster size,
using TABU-RCC.
Fig. 10. Comparison of TABU-RCC to EESH and the modified EESH in
terms of network lifetime.
[4] K. Akkaya and M. Younis, A Survey of Routing Protocols in
Wireless Sensor Networks, Ad Hoc Network J., vol. 3, no. 3,
pp. 325-349, 2005.
[5] L.M.C. Arboleda and N. Nasser, Comparison of Clustering
Algorithms and Protocols for Wireless Sensor Networks,
Proc. Canadian Conf. Electrical and Computer Eng., pp. 1787-1792,
May 2006.
[6] M. Younis, P. Munshi, G. Gupta, and S.M. Elsharkawy, On
Efficient Clustering of Wireless Sensor Networks, Proc. Second
IEEE Workshop Dependability and Security in Sensor Networks and
Systems, pp. 78-91, 2006.
[7] H. Karl and A. Wiling, Protocols and Architectures for Wireless Sensor
Networks. Wiley, 2005.
[8] Y. Yao and G.B. Giannakis, Energy-Efficient Scheduling for
Wireless Sensor Networks, IEEE Trans. Comm., vol. 53, no. 8,
pp. 1333-1342, 2005.
[9] S.F. Hwang, Y.Y. Su, Y.Y. Lin, and C.R. Dow, A Cluster-Based
Coverage-Preserved Node Scheduling Scheme in Wireless Sensor
Networks, Proc. Third Ann. Intl Conf. Mobile and Ubiquitous
Systems: Networking and Services, pp. 1-7, July 2006.
[10] W.R. Heizelman, A. Chandrakasan, and H. Balakrishnan,
Energy-Efficient Communication Protocol for Wireless Micro
Sensor Networks, Proc. IEEE Hawaii Intl Conf. System Sciences,
Jan. 2000.
[11] V. Mhatre and C. Rosenberg, Homogeneous vs. Heterogeneous
Clustered Sensor Networks: A Comparative Study, Proc. IEEE
Intl Conf. Comm., vol. 6, pp. 3646-3651, June 2004.
[12] M. Inanc, M.I. Magdon, and B. Yener, Power Optimal Con-
nectivity and Coverage in Wireless Sensor Networks, Technical
Report TR-03-06, Dept. of Computer Science, Rensselaer Poly-
technic Inst., 2003.
[13] J. Carle and D. Simplot, Energy Efficient Area Monitoring
by Sensor Networks, Computer, vol. 37, no. 2, pp. 40-46,
Feb. 2004.
[14] X.Y. Li, P.J. Wan, and O. Frieder, Coverage in Wireless Ad-Hoc
Sensor Networks, IEEE Trans. Computers, vol. 52, no. 6, pp. 753-
763, June 2002.
[15] G. Xing, X. Wang, Y. Zhang, C. Lu, R. Pless, and C. Gill,
Integrated Coverage and Connectivity Configuration for Energy
Conservation in Sensor Networks, ACM Trans. Sensor Networks,
vol. 1, no. 1, pp. 36-72, Aug. 2005.
[16] T. Yan, T. He, and J.A. Stankovic, Differentiated Surveillance for
Sensor Networks, Proc. First Intl Conf. Embedded Networked Sensor
Systems (SenSys 03), pp. 51-62, 2003.
[17] J. Lu, J. Wang, and T. Suda, Scalable Coverage Maintenance for
Dense Wireless Sensor Networks, Proc. Third Ann. IEEE Comm.
Soc. Conf. Sensor and Ad Hoc Comm. and Networks (SECON 06),
vol. 2, pp. 651-660, Sept. 2006.
[18] A. Chamam and S. Pierre, Dynamic Sensor Activation for
Maximizing Network Lifetime under Coverage Constraint, Proc.
Intl Symp. Comm. and Information Technologies (ISCIT 07), pp. 971-
976, Oct. 2007.
[19] S. Lindsey and C.S. Raghavendra, PEGASIS: Power Efficient
Gathering in Sensor Information Systems, Proc. IEEE Aerospace
Conf., Mar. 2002.
[20] N. Ahmed, S.S. Kanhere, and S. Jha, Probabilistic Coverage in
Wireless Sensor Networks, Proc. IEEE Conf. Local Computer
Networks (LCN 05), vol. 1, pp. 672-681, 2005.
[21] D. Tian and N.D. Georganas, A Coverage-Preserving Node
Scheduling Scheme for Large Wireless Sensor Networks, Proc.
ACM Intl Workshop Wireless Sensor Networks and Applications
(WSNA 02), pp. 32-41, 2002.
[22] M. Hefeeda and H. Ahmadi, A Probabilistic Coverage Protocol
for Wireless Sensor Networks, Proc. IEEE Intl Conf. Network
Protocols (ICNP 07), pp. 41-50, Oct. 2007.
[23] Y.-S. Yen, S. Hong, R.-S. Chang, and H.-C. Chao, An Energy
Efficient and Coverage Guaranteed Wireless Sensor Network,
Proc. IEEE Wireless Comm. and Networking Conf. (WCNC 07),
pp. 2923-2928, Mar. 2007.
[24] N. Patwari, J.N. Ash, S. Kyperountas, A.O. Hero, III, R.L. Moses,
and N.S. Correal, Locating the Nodes: Cooperative Localization
in Wireless Sensor Networks, IEEE Signal Processing Magazine,
vol. 22, no. 4, pp. 54-69, 2005.
[25] J. Bachrach and C. Taylor, Localization in Sensor Networks,
Handbook of Sensor Networks: Algorithms and Architectures,
I. Stojmenovic, ed., Wiley Interscience, 2005.
[26] H. Chu and R. Jan, A GPS-Less, Outdoor, Self-Positioning
Method for Wireless Sensor Networks, J. Ad Hoc Networks, vol. 5,
no. 5, pp. 547-557, July 2007.
[27] T.H. Cormen, C.E. Leiserson, R.L. Rivest, and C. Stein, The Set-
Covering Problem, Introduction to Algorithms, second ed.,
Section 35.3, pp. 1033-1038, MIT Press and McGraw-Hill, 2001.
[28] http://www.ilog.com/products/cplex/, 2009.
[29] W. Zhou, H.-M. Chen, and X.-F. Zhang, An Energy Efficient
Strong Head Clustering Algorithm for Wireless Sensor Net-
works, Proc. Intl Conf. Wireless Comm. Networking and Mobile
Computing (WiCom 07), pp. 2584-2587, Sept. 2007.
[30] O. Younis and S. Fahmy, HEED: A Hybrid, Energy-Efficient,
Distributed Clustering Approach for Ad Hoc Sensor Networks,
IEEE Trans. Mobile Computing, vol. 3, no. 4, pp. 366-379, Oct.-Dec.
2004.
Ali Chamam received the BEng degree in 1999
from the Ecole Nationale dIngenieurs de Tunis,
Tunisia, and the MSc degree in 2003 from the
Ecole Polytechnique de Montreal, Canada,
where he is currently working toward the PhD
degree. His research interests include mobile
computing, ad hoc and sensor networks, and
network planning. He is a student member of
the IEEE.
Samuel Pierre received the BEng degree in civil
engineering in 1981 from the Ecole Polytechni-
que de Montreal, Quebec, Canada, the BSc and
MSc degrees in mathematics and computer
science in 1984 and 1985, respectively, from
LUniversite du Que bec a` Montre al, Canada, the
MSc degree in economics in 1987 from the
University of Montreal, and the PhD degree in
electrical engineering in 1991 from the Ecole
Polytechnique de Montreal. He is currently a
professor of computer engineering at the Ecole Polytechnique de
Montreal, where he is also the director of the Mobile Computing and
Networking Research Laboratory (LARIM) and the holder of the
NSERC/Ericsson chair in next generations and mobile networking
systems. His research interests include wireline and wireless networks,
mobile computing, artificial intelligence, and telelearning. He is a fellow
of the Engineering Institute of Canada (EIC). He is a regional editor of
the Journal of Computer Science, an associate editor of IEEE
Communications Letters, IEEE Canadian Journal of Electrical and
Computer Engineering, and IEEE Canadian Review, and serves on the
editorial board of Telematics and Informatics (Elsevier Science). He is a
senior member of the IEEE.
> For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/publications/dlib.
1086 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 8, NO. 8, AUGUST 2009