Energy efficient protocol for deterministic

HEFEEDA ET AL.: ENERGY-EFFICIENT PROTOCOL FOR DETERMINISTIC AND PROBABILISTIC COVERAGE IN SENSOR NETWORKS 579

Energy-Efficient Protocol for Deterministic
and Probabilistic Coverage in Sensor Networks
Mohamed Hefeeda, Senior Member, IEEE, and Hossein Ahmadi

Abstract—Various sensor types, e.g., temperature, humidity, and acoustic, sense physical phenomena in different ways, and thus,
are expected to have different sensing models. Even for the same sensor type, the sensing model may need to be changed in different
environments. Designing and testing a different coverage protocol for each sensing model is indeed a costly task. To address this
challenging task, we propose a new probabilistic coverage protocol (denoted by PCP) that could employ different sensing models. We
show that PCP works with the common disk sensing model as well as probabilistic sensing models, with minimal changes. We analyze
the complexity of PCP and prove its correctness. In addition, we conduct an extensive simulation study of large-scale sensor networks
to rigorously evaluate PCP and compare it against other deterministic and probabilistic protocols in the literature. Our simulation
demonstrates that PCP is robust, and it can function correctly in presence of random node failures, inaccuracies in node locations, and
imperfect time synchronization of nodes. Our comparisons with other protocols indicate that PCP outperforms them in several aspects,
including number of activated sensors, total energy consumed, and network lifetime.

Index Terms—Sensor networks, coverage in sensor networks, probabilistic coverage, coverage protocols.

Ç

1 INTRODUCTION

M ANY real-life applications have been proposed for
wireless sensor networks, including forest fire detec-
tion, area surveillance, and natural habitat monitoring [1].
for the sensing ranges of sensors, with minimal changes in
its operation. The generality is important in real-life
applications, since different sensor types require different
Two of the important research problems in such sensor sensing models. Even for the same sensor type, the sensing
network applications are ensuring area coverage and model may need to be changed in different environments or
maintaining the connectivity of the network. Area coverage when the technology changes. Thus, the generality of the
means that nodes use their sensing modules to detect events proposed protocol enables the designers of sensor networks
occurring in the monitored area. Each sensor is assumed to to avoid the costly and complex tasks of designing,
have a sensing range, which depends on the phenomenon implementing, and testing a different coverage protocol
being sensed and the environment conditions. Maintaining for each sensing model.
the sensor network connected is also important because The proposed protocol, called Probabilistic Coverage
information about the detected events may need to be Protocol (PCP), works for the disk sensing model used in
communicated to processing centers for possible actions. many of the previous works in the literature, e.g., [2], [3],
Connectivity is achieved by the radio communication [4], [5], [6], [7], [8]. This model, depicted in Fig. 1a, assumes
modules installed in the sensors. Each sensor is assumed that the sensing range is a uniform disk of radius rs . The
to have a communication range, which is totally different simple disk sensing model is appealing because it makes
from the sensing range in general. coverage maintenance protocols, e.g., [2], [3], [4], less
This paper presents a new efficient and general coverage complicated to design and analyze. It also makes analytical
protocol, which also considers the network connectivity. The and asymptotic analysis, e.g., [5], [6], tractable. However, it
proposed protocol is efficient because it reduces the energy may not be realistic in some environments or it could
consumed by sensor nodes and prolongs the network become too conservative in modeling the sensing range of
some sensors. Therefore, better models for sensing ranges
lifetime. Energy efficiency is critical for successful deploy-
may be needed in some sensor network applications.
ment and operation of large-scale sensor networks that are
Several studies [9], [10], [11], [12], [13] have argued that
typically battery-powered. The protocol is general because it
probabilistic sensing models capture the behavior of sensors
can employ different deterministic and probabilistic models
more realistically than the deterministic disk model. For
example, through experimental study of passive infrared
. M. Hefeeda is with the School of Computing Science, Simon Fraser (PIR) sensors, the authors of [13] show that the sensing
University, 250-13450 102nd Ave., Surrey, BC V3T 0A3, Canada. range is better modeled by a continuous probability
E-mail: mhefeeda@cs.sfu.ca.
distribution, which is a normal distribution in the case of
. H. Ahmadi is with the Department of Computer Science, University of
Illinois at Urbana-Champaign, 201 N. Goodwin Ave., Urbana, IL 61801. PIR sensors. The authors of [9], [10] use an exponential
E-mail: hahmadi2@uiuc.edu. sensing model, where the sensing capacity degrades
Manuscript received 16 Jan. 2009; revised 21 June 2009; accepted 26 June according to an exponential distribution after a certain
2009; published online 1 July 2009. threshold, as shown in Fig. 1b. Whereas the authors of [12]
Recommended for acceptance by Y.-C. Tseng. propose a polynomial function to model the probabilistic
For information on obtaining reprints of this article, please send e-mail to:
tpds@computer.org, and reference IEEECS Log Number TPDS-2009-01-0027. nature of the sensing range, as shown in Fig. 1d.
Digital Object Identifier no. 10.1109/TPDS.2009.112. Furthermore, the authors of [11] assume that the sensing
1045-9219/10/$26.00 ß 2010 IEEE Published by the IEEE Computer Society
Authorized licensed use limited to: SIMON FRASER UNIVERSITY. Downloaded on May 12,2010 at 04:12:29 UTC from IEEE Xplore. Restrictions apply.

580 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 21, NO. 5, MAY 2010

Fig. 1. Some of the sensing models used in the literature. (a) Disk model. (b) Exponential model in [9], [10]. (c) Staircase model in [11]. (d) Model in [12].

range can be modeled as layers of concentric disks with define the probabilistic coverage problem and show how our
increasing diameters, and each layer has a fixed probability protocol can solve it. In Section 5, we prove the correctness of
of sensing, as shown in Fig. 1c. the protocol and provide bounds on its convergence time and
To support probabilistic sensing models, we introduce message complexity. We also prove the condition on the
the notion of probabilistic coverage of a target area with a communication range needed for our coverage protocol to
given threshold , which means that an area is considered provide network connectivity as well. In Section 6, we
covered if the probability of sensing an event occurring at evaluate our protocol and compare it against others using
any point in the area is at least . This notion not only simulation. We conclude the paper in Section 7.
enables probabilistic sensing models but also provides a
coverage-network lifetime control knob, which is . This knob
allows the sensor network operators to trade off the 2 RELATED WORK
coverage level for longer network lifetimes, if the applica- Coverage in sensor networks has received significant
tions using the sensor network could tolerate this trade-off. research attention, see [14] for a survey. We summarize
For example, reducing makes the coverage protocol the most relevant works to ours in the following sections.
activate fewer sensors to monitor the area, and hence,
saves energy and extends the network lifetime. We show 2.1 1-Coverage Using the Disk Sensing Model
that the proposed protocol can work with various prob- The studies in [5], [6] conduct asymptotic and analytical
abilistic sensing models. In particular, our protocol requires analyses to provide necessary and sufficient conditions for
the computation of a single parameter from the adopted
coverage in various environments. While these studies
sensing model, while everything else remains the same. We
provide useful insights and guidelines, which we indeed
show how this parameter can be derived in general, and we
benefited from, they do not propose specific coverage
actually do the calculation for the exponential sensing
model [9], [10]. This model is chosen because it is protocols. In [15], optimal deployment patterns for different
conservative in terms of estimating the sensing capacity, ratios of the communication and sensing ranges are
and it has been used before in another probabilistic proposed. Exact sensor placement is difficult, if at all
coverage protocol (CCANS [10]). This enables us to possible, in many realistic environments such as battlefields
compare our protocol against CCANS, which is the only and forests. In addition, Bai et al. [15] do not present
fully specified probabilistic coverage protocol that we are dynamic protocols to maintain coverage as the time passes
aware of. Also because it is conservative, the exponential and some sensors fail.
sensing model can be used as a first approximation for Several distributed coverage protocols have been pro-
many other sensing models. posed for the disk model, including [2], [4], [7], [8], [16], [17],
We analyze the complexity of the proposed protocol and [18]. For example, OGDC [4] tries to minimize the overlap
prove its correctness. We also derive the condition under between the sensing circles of activated sensors, while CCP
which the protocol maintains the connectivity of the [2] deactivates redundant sensors by checking that all
network in addition to covering the monitored area. intersection points of sensing circles are covered. CCP can
Furthermore, we conduct an extensive simulation study of provide coverage with degrees higher than 1 as well. Sensors
large-scale sensor networks to rigorously evaluate our in PEAS [7] probe their neighbors to decide whether to be in
protocol and compare it against other deterministic and active or sleep mode. The coverage algorithms in [16] solve a
probabilistic protocols in the literature. Our simulation variation of the set k-cover problem, where sensors are
demonstrates that the proposed protocol is robust, and it partitioned into k covers and individual covers are itera-
can function correctly in presence of random node failures, tively activated to achieve 1-coverage of the monitored area.
inaccuracies in node locations, and imperfect time synchro- The authors of [17] propose three node scheduling schemes
nization of nodes. Our comparisons with other protocols that estimate the distance to the nearest neighbor, number of
indicate that our protocol outperforms them in several neighbors, or the probability of a node being off duty and use
aspects, including number of activated sensors, total energy one of these metrics to put some sensors in sleep mode. The
consumed, and network lifetime. coverage algorithm in [18] tries to find uncovered spots and
The rest of the paper is organized as follows: We activate sensors in these areas using information from
summarize the related work in Section 2. In Section 3, we nearby active sensors.
present the proposed protocol in the context of the disk All of the above protocols assume the disk sensing
sensing model because it is easier. In Section 4, we formally model. Our protocol, in contrast, is general and can adopt


HEFEEDA AND AHMADI: ENERGY-EFFICIENT PROTOCOL FOR DETERMINISTIC AND PROBABILISTIC COVERAGE IN SENSOR NETWORKS 581

the disk model as well as probabilistic models. To show this hood. The work in [3] presents two distributed k-coverage
generality, we compare our protocol against the more recent algorithms. The first one is a distributed greedy algorithm,
OGDC and CCP protocols, because, according to the which requires carrying around a central state. The second
performance evaluations in [2], [4], they outperform the algorithm, called distributed priority algorithm (DPA), is
earlier ones. localized and more robust. DPA, which is also used in [30]
The early work [19] considers a slightly different with some modifications to activate a minimal subset of
definition of coverage: finding maximal support and sensors to answer a query, employs multihop neighborhood
maximal breach paths for which the observability is information to turn off nodes that are not needed to k-cover
maximum and minimum, respectively. The authors of [21] the area. A more recent k-coverage algorithm was presented
improve the work in [19] and present a more efficient in our previous work [31].
algorithm. Furthermore, Li et al. [21] present a distributed Because of the hardness of the problem, most of these
algorithm for the maximal support path. We consider a works assume the disk sensing model; the concept of k-
different problem: area coverage. Thus, these algorithms are coverage under probabilistic sensing models is not yet well
not comparable to ours. defined. In this paper, we focus on 1-coverage with
2.2 Coverage Using Probabilistic Sensing Models probabilistic sensing models and leave the extension to
the k-coverage case for future work.
Probabilistic coverage with various sensing models has also
Finally, a closely related problem to coverage is con-
been studied in [10], [11], [12], [22]. The work in [12]
nectivity. k-connectivity (k ! 1Þ means that there are at least
analytically studies the implications of adopting probabilistic
k disjoint paths between any pair of nodes in the network.
and disk sensing models on coverage. The study in [22]
For the disk sensing and communication models, it has been
presents closed-form equations for computing the probabil-
ity of any point in the area to be covered given N deployed proved that if the communication range of sensors is at least
sensors, which could be heterogeneous and may not twice the sensing range and the monitored area is convex,
necessarily follow the disk sensing model. These studies do then k-coverage implies k-connectivity [2], [4], [10]. In this
not propose specific coverage protocols. In [11], the sensing paper, we prove the conditions under which probabilistic
range is modeled as layers of concentric disks with increasing coverage ensures 1-connectivity.
diameters, where the probability of sensing is fixed in each
layer. A coverage evaluation protocol is also proposed. 3 PCP WITH DISK SENSING MODEL
Although the authors mention that their coverage evaluation
In this section, we present our new PCP, in the context of
protocol can be extended to a dynamic coverage protocol,
the disk sensing model because it is simpler.
they do not specify the details of that protocol. Therefore, we
could not compare our protocol against theirs. The closest 3.1 Overview of PCP
work to ours is [10], where the authors design a probabilistic It has been shown before, e.g., in [15], that covering an area
coverage protocol called CCANS. A brief description of with disks of same radius (rs ) can optimally be done by
CCANS is presented in Section 6.4. We compare our protocol placing disks on vertices of a triangular lattice, where the side
against CCANS. pffiffiffi
of the triangle is 3rs . Optimality here means the minimum
2.3 k-Coverage and Network Connectivity number of disks required. The idea of PCP is to activate a
Coverage with various degrees (k-coverage), where every subset of deployed sensors to construct an approximate
point is sensed by at least k sensors, has also been studied, triangular lattice on top of the area to be covered. PCP starts
see the survey in [14]. The problem of verifying k-coverage by activating any sensor in the area, which we refer to as an
is studied in [23]. Each sensor is modeled as a disk and it is activator. This sensor activates six other sensors located at
proved that the area is k-covered if the perimeter of all vertices of the hexagon centered at that sensor. Each
disks is k-covered. An improved modeling is presented in activated sensor, in turn, activates other sensors at vertices
[24], where the authors use the concept of order-k Voronoi of its own hexagon. As illustrated in Fig. 2, this process
diagrams [25] to build a verifier algorithm. In [26], the continues till the activated sensors form a virtual triangular
authors first propose a k-coverage determination algorithm lattice over the whole area.
and then present a distributed sleep control protocol to We refer to the distance between the vertices of the
achieve k-coverage by exchanging several types of mes- triangular lattice as the maximum separation between
sages. In [27], the authors formulate the k-coverage active nodes, and it is denoted by s. The value of s is
problem of a set of n grid points as an integer linear determined from the sensing range rs of sensors. Under the
programming to determine the minimum cost of sensors to diskpsensing model, the maximum separation is set to
cover all grid points. ffiffiffi
s ¼ 3rs . The lattice is approximate because it is con-
In [28], the authors address the problem of selecting the
structed in a distributed manner and is controlled by sensor
minimum number of sensors to activate to achieve k-
coverage, which is shown to be NP-hard. The authors deployment. The initial sensor deployment is not assumed
present a distributed algorithm, which uses a pruning to be on a lattice; it could be any distribution. In our
method similar to the algorithms used for constructing simulations, we deploy sensors uniformly at random.
connected dominating sets, e.g., [29]: nodes are assigned The above description is idealistic and makes several
unique priorities and they broadcast their neighbor set assumptions. We list these assumptions below and describe
information. Then each node can go to a sleep mode by how we address them. We also note that a similar idea of
checking whether the coverage and connectivity can be activating sensors on a triangular lattice was used by the
maintained by other higher priority nodes in its neighbor- OGDC protocol [4]. However, the optimization proposed



node sets its state to be START and selects a random start-
up timer Ts inversely proportional to its remaining energy
level. The node with the smallest Ts will become active and
broadcast an activation message to all nodes in its
communication range. The sender of activation message is
called the activator. The activation message contains the
coordinates of the activator, and it tries to activate nodes at
vertices of the hexagon centered at the activator, while
putting all other nodes within that hexagon to sleep. A node
receiving the activation message can determine whether it is
a vertex of the hexagon by measuring the distance and
angle between itself and the activator. The angle is
measured starting from the positive x-axis and going
anticlockwise. If the angle is multiple of =3 and the
distance is s, then the node sets its state to ACTIVE and it
becomes a new activator. Otherwise, it goes to SLEEP state.
In real deployment, nodes may not always be found at
vertices of the triangular lattice because of randomness in
Fig. 2. A simplification of the node activation process in PCP. Activated node deployment or because of node failure. PCP tries to
nodes try to form a triangular lattice over the area. activate the closest nodes to hexagon vertices in a dis-
tributed manner as follows: Every node receiving an
later in Section 3.2 is new and yields significant perfor- activation message calculates an activation timer Ta as a
mance improvement as shown in Section 6. In addition, the function of its closeness to the nearest vertex of the hexagon
authors of [4] do not address probabilistic sensing models, using the following equation:
as we do in Section 4. À Á
Ta ¼ a d2 þ da 2
2 ;
v ð1Þ
.Single starting node. In the beginning of the
where dv and da are the euclidean distances between the
protocol, only one node starts as an activator. In
Section 3.3, we extend our protocol to handle node and the vertex, and the node and the activator,
multiple starting nodes, which is important for respectively;
is the angle between the line connecting the
large-scale sensor networks. node with the activator and the line connecting the vertex
. Nodes are time synchronized at a coarse-grained with the activator; and a is a constant.1 Note that the closer
level. In Section 6, we verify that only coarse-grained the node gets to the vertex, the smaller the Ta will be. After
synchronization is needed and we study the robust- computing Ta , a node moves to WAIT state and stays in this
ness of our protocol to clock drifts. state till its Ta timer either expires or is canceled. When the
. Nodes know their locations. This is not hard to smallest Ta timer expires, its corresponding node changes
achieve in practice given efficient localization its state to ACTIVE. This node then becomes a new activator
schemes such as those in [32], [33], any of them can and broadcasts an activation message to its neighbors.
be used with our protocol. The protocols that we When receiving the new activation message, nodes in WAIT
compare ours against [2], [4], [10] also assume that state cancel their Ta timers and move to SLEEP state.
nodes know their locations. Note that our protocol
does not require accurate knowledge of global posi- 3.2 Optimization Using -Circles
tions because the position information is used only in We propose a new optimization of the PCP protocol. This
local decisions to activate nodes, as will become clear optimization puts more sensors in sleep mode faster,
later. In Section 6, we analyze the robustness of our
shortens the protocol convergence time, and thus, saves
protocol to inaccuracies in node location information.
more energy. We first introduce the concept of -circle in
We now present some details of the proposed protocol.
the following definition:
PCP works in rounds of R seconds each. R is chosen to be
much smaller than the average lifetime of sensors. In the Definition 1 (-circle). The smallest circle drawn anywhere in
beginning of each round, all nodes start running PCP the monitored area such that there is at least one node inside it
independent of each other. A number of messages will be is called the -circle, where is the diameter of the circle.
exchanged between nodes to determine which of them
should be on duty (i.e., active) during the current round,
1. The intuition behind this formula is as follows: We need the activation
and which should sleep till the beginning of the next round. timer Ta to rank points in terms of their deviation from the lattice vertex. For
The time it takes the protocol to determine active/sleep each point, the timer has to be related to the number of points with better
nodes is called the convergence time, and it is desired to be as positions. Since the number of points around the lattice vertex having
distance less than dv is proportional to d2 , the waiting should be
small as possible. After convergence, no node changes its v
proportional to d2 . In addition, the angle
is between 0 and 2 while the
v
state and no protocol messages are exchanged till the scale of dv can change in different applications. Therefore,
is multiplied by
beginning of the next round. the distance between sensor and the activator da to make it on the same
scale as dv . The number of points with better
inside a -circle is
In PCP, a node can be in one of the four states: ACTIVE, proportional to
2 . Thus, the activation timer is formed by the summation of
SLEEP, WAIT, or START. In the beginning of a round, each d2 and the scaled angle ðda
Þ2 .
v



Fig. 3. Choosing the closest node to a triangle vertex.

The diameter is computed from the deployment
distribution of nodes. We compute for two common
deployment schemes: grid and uniform distribution. for
Fig. 4. The state diagram of the PCP protocol. In each state, we mark
other schemes can be derived in a similar way. We assume
which of the sensing, sending, and receiving modules is on.
that there are n nodes to be deployed on the monitored area,
which is an l Â pffiffiffi pffiffiffiFor the grid distribution, nodes are
l square.
decreasing node density with the time. For example, can
deployed on a n Â n virtual grid.ffiffiffi p The spacing between be doubled after certain number of rounds of the protocol.
any two adjacent grid points is l=ð n À 1Þ. To compute ,
consider any grid cell that is composed of four nodes This only requires each node to keep a counter on the
pffiffiffi pffiffiffi number of elapsed rounds. Also note that during transition
forming a small square of size l=ð n À 1Þ Â l=ð n À 1Þ.
Clearly, setting larger than the diagonal of this small between rounds, active nodes in the finished round stay
square ensures that a -circle drawn anywhere ffiffiffi the grid active for a short period in the new round while they are
p on pffiffiffi
will contain at least one node. Therefore, ¼ 2l=ð n À 1Þ participating in the protocol. This period is approximately
for grid deployment. Next, we consider the case when equal to the expected convergence time. After this short
nodes are deployed uniformly at random in the range period, these nodes will move to states determined by the
½0; 2Š, i.e., the mean distance between adjacent nodes is , protocol in the new round. This is done to prevent any
whereas the maximum distance does not exceed 2. Using a outages in coverage during transition.
similar argument as in the grid distribution, should be
pffiffiffi
2 2. To uniformly distribute n nodes over an ffiffiffi Â l square, 3.3 Multiple Starting Nodes
pffiffiffi p l pffiffiffi
should be l=ð n À 1Þ, which results in ¼ 2 2l=ð n À 1Þ. For large-scale sensor networks, it may be desired to have
Note that randomness in the deployment distribution multiple starting nodes such that the coverage protocol
results in larger values. converges faster in each round. Faster convergence means
We employ the -circle concept to minimize the number that nodes move quicker from START or WAIT state to
of nodes in WAIT state. That is, nodes decide quickly to be either SLEEP or ACTIVE state, which reduces the total
either in ACTIVE or SLEEP state. This saves energy because energy consumed in the network. Multiple starting nodes,
nodes in WAIT state must have their wireless receiving however, could increase the number of activated sensors
modules turned on, while all modules are turned off in because of the potential overlap between subareas that are
SLEEP state. The savings in energy are significant as shown covered due to different starting nodes.
in Section 6. PCP achieves this optimization by making only The number of starting nodes in a round can be
nodes inside -circles near to the six vertices of the hexagon controlled by setting the range of the start-up timer Ts . Ts
stay in WAIT state, all others move to SLEEP state once they
is chosen randomly between 0 and a constant s . Suppose
determine that they are outside of all -circles. Nodes inside
that we want to compute the value of s such that each
-circles compute activation timers, as described above, to
round of PCP starts with k nodes, on average. Let us assume
choose the closest node the vertex to be active. Fig. 3 shows
that the average convergence time of PCP is Tc . Note that if
one of the six -circles of a given activator. Note that the
centers of the -circles are located at a distance of s À =2 the start-up timer Ts of a node is less than Tc , this node will
from the activator and at angles that are multiple of =3. become a starting node before the protocol converges. The
The state diagram of the PCP protocol is illustrated in Fig. 4. expected number of nodes with Ts smaller than Tc is
The figure shows the status of the sensing, sending, and k ¼ ðTc =s Þn, which yields s ¼ nTc =k. In Section 6, we
receiving modules in each state of the node. verify that our protocol consumes the energy of nodes in a
Note that the PCP protocol does not require that to be uniform manner, therefore, it keeps more nodes alive for
static throughout the lifetime of the sensor network. longer periods and prolongs the network lifetime. We also
Rather, can be changed to account for node failures or study the impact of multiple starting nodes on the number



nodes on the triangular lattice. Once we compute s, PCP
functions in the same manner as described in Section 3.
Note that computing s depends only on the sensing
model used. In the following section, we derive s for the
exponential sensing model [9], [10] as an example.
Computing s for other sensing models can be done in a
similar way. We should emphasize that the operation of the
PCP protocol does not change by changing the sensing
model. The only parameter that needs to be determined and
given to PCP is the maximum separation between any two
active sensors s.

4.2 Computing Maximum Separation
This section presents the details of computing the max-
Fig. 5. The sensing capacity of three sensors that use the exponential imum separation s between any two active nodes for the
sensing model and deployed at vertices of an equilateral triangle. The exponential sensing model, which is defined as
least-covered point by these three sensors is at the center of the
triangle. 1; for d rs ;
pðdÞ ¼ ð2Þ
eÀðdÀrs Þ ; for d rs ;
of activated nodes, convergence time, and total energy
where pðdÞ is the probability of detecting an event
consumed in the network.
happening at a distance d from the sensor, rs is a threshold
below which the sensing capacity is strong enough such
4 PCP WITH PROBABILISTIC SENSING MODELS that any event will be detected with probability 1, and is a
In this section, we define the notion of probabilistic factor that describes how fast the sensing capacity decays
coverage and extend the proposed PCP protocol to support with distance. We call the sensing capacity decay factor.
The exponential model is shown in Fig. 1b. We consider this
probabilistic sensing models.
sensing model for two reasons. First, it has been adopted
4.1 Definitions and Overview before in [9], [10], which allows us to conduct a fair
We start by stating the following two definitions: comparison between our protocol and the protocol in [10].
Second, it is conservative as it assumes that the sensing
Definition 2 (probabilistic coverage). An area A is
capacity decreases exponentially fast beyond rs , which
probabilistically covered by n sensors with threshold parameter
Q means that the achieved actual coverage will be higher than
ð0 1Þ if P ðxÞ ¼ 1 À n ð1 À pi ðxÞÞ ! for every
i¼1 the estimated by the theoretical analysis. In addition, since
point x in A, where pi ðxÞ is the probability that sensor i the exponential sensing model is conservative, it can be
detects an event occurring at x. used as a first approximation for other sensing models such
as those in [11], [12], [13]. Therefore, sensor network
P ðxÞ in the above definition measures the collective designers may not need to compute the exact value of the
probability from all sensors to cover point x, pi ðxÞ is maximum separation parameter for mathematically com-
specified by the adopted sensing model, and the coverage plex sensing models, and instead use the exponential
threshold parameter depends on the requirements of the sensing model.
target application. If we set ¼ 1 and pi ðxÞ as a binary The following theorem provides the maximum separa-
function that takes on either 0 or 1 in the above definition, tion between any two active nodes s for the exponential
we get the commonly used deterministic coverage defini- sensing model:
tion with the disk sensing model. Theorem 1 (maximum separation). Under the exponential
Definition 3 (least-covered point). A point x within an area A sensing model defined in (2), the maximum separation between
is called the least-covered point of A if P ðxÞ P ðyÞ for all any two active sensors on the triangular lattice to ensure that
y 6¼ x in A. the probability of ffiffiffiffiffiffi
pffiffiffi psensing at the least-covered point is at least
lnð1À 3 1ÀÞ
is 3ðrs À Þ.
Fig. 5 demonstrates the concept of the least-covered
Proof. To prove this theorem, we need to find the location
point by showing the sensing capacity of three nodes
of the least-covered point. We prove in the Appendix
deployed on an equilateral triangle. The three sensors use pffiffiffi
the exponential sensing model in Fig. 1b. that this location is at a distance of s= 3 from each vertex
The proposed PCP protocol works again by building an of the equilateral triangle. The probability of sensing at
Àðpffi Àrs Þ 3
s

approximate triangular lattice in a distributed manner, as the least-covered point is then 1 À ð1 À e 3 Þ , which
explained in Section 3. However, in the probabilistic case, should be greater than or equal to . By manipulating
PCP ensures that the least-covered point in the monitored this inequality, ffiffiffiffiffiffi get the maximum separation
p we
pffiffiffi 3
area has a probability of being sensed that is at least . In s ¼ 3ðrs À lnð1À 1ÀÞÞ. u
t
order for PCP to provide this assurance, we need to Note that the exponential sensing model reduces to the
compute the maximum separation s between any two active disk model when we set ¼ 1. From Theorem 1, it is easy



pffiffiffi
to see that s ¼ 3rs under the disk sensing model, which is node creates a connected subset of activated nodes. Thus,
the same value used by our protocol in Section 3 and the we need to prove that the union of subsets activated by
same as the optimality condition proved in [4], [15]. different starting nodes is also connected. We prove this
Remark. Note that as the coverage threshold decreases, the by contradiction. Consider any two connected subsets A
maximum separation between activated sensors in- and A0 that are activated by two different activators. Let
creases. Therefore, a fewer number of active sensors will u 2 A and v 2 A0 be the nearest nodes in the two subsets.
be needed to maintain the coverage of the monitored Assume that the PCP protocol terminates and the network
area. This implies that the lifetime of the sensor network is not connected, i.e., A is disconnected from A0 . Thus, the
could be extended by reducing if the application can distance between u and v is more than their communica-
tolerate this reduction. Therefore, the coverage threshold tion range: distðu; vÞ rc . Since the protocol has termi-
nated, there is no node in the WAIT state. Therefore, there
can be used as a control knob, which can trade off the
are six activated neighbors of u with a distance at most s;
coverage quality for longer network lifetimes. In Section 6,
otherwise, some nodes around u are still in WAIT state.
we assess the savings in the number of active sensors for
Let u0 be the neighbor with the least distance to v. We
different values of .
identify two cases as follows:

5 ANALYSIS OF THE PCP PROTOCOL 1. u0 2 A0 . Since distðu; u0 Þ s and distðu; vÞ rc ! s,
we have distðu; vÞ distðu; u0 Þ. Thus, u0 2 A0 is
In this section, we first state and prove the condition under closer to u 2 A than v 2 A0 . This is a contradiction
which the activated nodes by the PCP protocol form a because u and v are assumed to be the closest nodes
connected network. Then, we prove the correctness of the in A and A0 .
PCP protocol and provide bounds on its convergence time, 2. u0 2 A. Consider the triangle uu0 v, and recall that
message complexity, and number of nodes activated in each any triangle has the following property:
round. All theorems in this section are validated using
simulation in Section 6. distðu0 ; vÞ2 ¼ distðu; vÞ2 þ distðu; u0 Þ2
5.1 Network Connectivity Analysis À 2 cosðu0 uvÞdistðu; vÞdistðu; u0 Þ:
Under the disk sensing model, previous studies [2], [4], [10] Since distðu; u0 Þ s, we have
have shown that if the communication range of sensors is at
least twice the sensing range and the surveillance area is distðu0 ; vÞ2 distðu; vÞ2 þ s2
convex, then coverage implies that the network is con- À 2 cosðu0 uvÞdistðu; vÞs:
nected. These results may not hold in the case of PCP
because it uses probabilistic sensing models. The following The angle between lines uv and uu0 , called u0 uv,
theorem provides the condition on the communication is less than or equal to 60 degrees. Otherwise,
range to ensure that PCP results in a connected network of there is another neighbor of u nearer than u0 to
activated sensors. The theorem assumes that the commu- v. Therefore, cosðu0 uvÞ ! 0:5 and distðu0 ; vÞ2
nication range of nodes is a circle with radius rc : distðu; vÞ2 þ sðs À distðu; vÞÞ. Since
Theorem 2 (network connectivity). The subset of nodes
activated by PCP will result in a connected network if the distðu; vÞ rc ! s;
communication range of nodes rc is greater than or equal to the we have sðs À distðu; vÞÞ 0. Therefore,
maximum separation between any two active nodes s, where s
is computed from the sensing model. distðu0 ; vÞ distðu; vÞ:
Proof. First, we prove that the subset of nodes activated In other words, u0 is closer to v than u, which is a
by PCP is connected when there is a single starting contradiction. u
t
node in each round. We use induction in the proof.
Initially, we have one node activated, which is 5.2 Correctness and Complexity Analysis
connected. Suppose at step k, we have a connected We carry out our analysis in terms of the input parameters
subset Ak of active nodes formed after k steps of ; ; s, and l, and the protocol parameter a , which is the
sending activation messages. We show by contradiction maximum value of the activation timer. is determined
that the subset Akþ1 constructed in step k þ 1 is also from the deployment distribution of sensors as explained in
connected. Suppose that Akþ1 is not connected. Since Section 3.2. The maximum separation between any two
Ak is connected, there are some nodes (denoted by the
active nodes s is computed from the adopted probabilistic
set V ) that are activated in step k þ 1 and not
sensing model as explained in Section 4.2. is the
connected to Ak . Consider any v 2 V . v must have
been activated by an activator (say u) in Ak because v is probabilistic coverage threshold, which is application
activated in step k þ 1. Since v is at a distance of at dependent. l is the length of the area to be covered, which
most s from u, v is reachable from u because rc ! s. is assumed to be a square for simplicity of the analysis. We
Since v is chosen arbitrarily from V , all nodes in V are assume that the area is large compared to the sensing
reachable from Ak . That is, Akþ1 is connected, which radius, and therefore, we ignore the boundary effects. We
contradicts the assumption. further assume that a message transferred between two
Second, we consider the case for multiple starting neighboring nodes takes at most m time units, which
nodes. From the previous case, we know that each starting includes propagation and transmission delays.



The following theorem proves the correctness of PCP
and provides an upper bound on its convergence time. PCP
is considered correct if terminates with every point in the
area has a probability of being sensed at least . After
convergence, nodes do not change their states and no
protocol messages are exchanged till the beginning of the
next round:
Theorem 3 (correctness and convergence time). The PCP
protocol converges in at most lða 2 þ m Þ=ðs À Þ time units
with every point in the area has a probability of being sensed at
least .
Proof. First, we prove the correctness part. PCP incremen-
tally constructs a triangular lattice of active nodes. This
triangular lattice will eventually cover the whole area
because each node begins a round with setting a start-up
timer Ts , and if Ts expires, the node becomes active (i.e.,
it will be a vertex of a triangle). The Ts timer of a node n1
can be canceled only if another node n2 has become Fig. 6. Connectivity among nodes activated by PCP.
active and n2 is at a vertex of the triangle that contains n1 .
pffiffiffi pffiffiffi
Now we need to show that each triangle of the lattice is 3=6 Â 2l2 = 3s2 ¼ l2 = 3s2 . The number of nodes activated
covered. Consider any triangle. Since nodes activated by by PCP is computed in a similar way, but with a triangular
an activator are at a distance of at most s from the lattice with spacing at most s À . Thus, the number of
pffiffiffi
activator, the triangle formed by activated nodes will activated nodes by PCP is at most l2 = 3ðs À Þ2 .
have side lengths of at most s. Recall that s is computed For message complexity, we notice that there is only
from the sensing model to ensure that the coverage one message sent by each activated node. Thus, the total
probability at the least-covered point in a triangle is at number of messages sent in a round is equal to the
least . Therefore, the coverage probability in the whole number of activated sensors. u
t
triangle is at least .
Second, we bound the convergence time. Within each
round, PCP runs in steps. In each step, an activation 6 EVALUATION
message is sent, and at least one node is activated in each In this section, we evaluate our protocol and compare it
of the six directions. Consider one direction. In the worst against others in the literature using extensive simula-
case, the newly activated node is at a distance of s À tions. We first describe our experimental setup. Then, we
from the old node. Thus, in the worst case, PCP needs verify the correctness of our protocol and validate the
l=ðs À Þ steps if the first activated node is at the border. theoretical bounds derived in Section 5. Next, we study
The maximum time to complete one step occurs when the robustness of our protocol against node failures,
the node chosen to be active happens to have the largest inaccuracy in nodes location information, and clock drifts.
value for the activation timer Ta , which is a 2 (computed Then, we compare our protocol against a probabilistic
from (1)). Adding the message transmission time m to coverage protocol called CCANS [10]. Finally, we com-
the maximum value of the activation timer yields a pare our protocol versus two recent deterministic cover-
worst-case time for any step as ða 2 þ m Þ. Multiplying age protocols: OGDC [4] and CCP [2].
this value by the number of steps l=ðs À Þ yields the
6.1 Experimental Setup
worst-case convergence time of PCP. u
t
We have implemented our PCP protocol in NS-2 [34] and in
The following theorem provides upper bounds on the our own packet-level simulator in C++. The source code for
number of activated sensors and number of messages both implementations is available at [35]. Some results from
exchanged by PCP in a round: the NS-2 implementation (Fig. 6) with reasonable network
sizes (up to 1,000 nodes) are presented. Most results,
Theorem 4 (activated nodes and message complexity). The however, are based on our own simulator because it
number of nodes activated by the PCP protocol is at most
pffiffiffi supports much larger networks, which we need to
l2 = 3ðs À Þ2 , which is the same as the number of exchanged rigorously evaluate our protocol.
messages in a round. We use the following parameters in the experiments,
Proof. The number of nodes to cover an l Â l area is equal to unless otherwise specified. We uniformly at random
the number of vertices of a triangular lattice with spacing s. deploy 20,000 sensors over a 1 km Â 1 km area. We use
pffiffiffi
This number is l2 = 3s2 and computed as follows: Since the
pffiffiffi
two sensing models: The disk sensing model with a sensing
area of an equilateral triangle with side s is s2 3=2 and the range of rs ¼ 15 m and the exponential sensing model with
triangles completely tile the area, the total number of
pffiffiffi sensing capacity decay factor ¼ 0:05, and we set rs ¼
triangles required is 2l2 = 3s2 . Since there are three nodes 15 m as the threshold value below which sensing is
used in each triangle and each node is also used in six achieved with probability 1. We employ the energy model
different triangles, the total number of nodes is in [7] and [4], which is based on the Mote hardware



specifications. In this model, the node power consumption
in transmission, reception, idle, and sleep modes is 60, 12,
12, and 0.03 mW, respectively. The initial energy of a node
is assumed to be 60 Joules, which allows a node to operate
for about 5,000 seconds in reception/idle modes.
When we compare various coverage protocols, we
assume that the wireless communication channel has a
bandwidth of 40 Kbps. Since the message sizes in all
protocols are almost the same, we assume that the average
message size is 34 bytes, which is the same size used in [4].
We ignore the propagation delay because it is negligible for
the 1 km Â 1 km area considered in the simulation. This
results in a message transmission time m ¼ 6:8 ms.
We repeat each experiment 10 times with different seeds
and report the averages in all of our results. We also report
the minimum and maximum values if they do not clutter
the figures. Note that the simulated sensor network in each Fig. 7. Savings in number of active nodes because of using the
experiment replica has 20,000 nodes, and the measured exponential sensing model for different values of and .
statistics are collected from all of them. Therefore, we
believe that combining the data from 10 different replicas figure indicates that even for a conservative value of ¼
and each with 20,000 nodes yields statistically significant 0:05 and for ¼ 0:99, a saving of up to 30 percent in number
results (we did not see large variances in our results). of active nodes can be achieved, which means less energy
Finally, we mention that in most experiments, each single consumed and ultimately longer lifetimes for the sensor
replica took several hours of running time on a decent network. It is expected that the savings will be higher for
multicore Linux server. Furthermore, processing the huge other probabilistic sensing models in which the sensing
traces created in these large-scale experiments consumed capacity decays slower than exponential.
many CPU hours. In addition, PCP provides a controllable knob: the
coverage level , which can be employed to trade off
6.2 Validation and Savings Achieved by PCP reliability of sensing the area with the sensor network
We validate that PCP indeed achieves the requested lifetime. For example, Fig. 7 shows that the number of
coverage level for all points in a monitored area for activated sensors can be decreased if the coverage threshold
deterministic as well as probabilistic sensing models. We is reduced. Reducing is feasible in applications that can
also study the potential gain of adopting probabilistic tolerate a small probability of not detecting an event
sensing models. happening at a point, as it can be sensed at other points.
Coverage and Connectivity. In the first experiment, we fix Theory versus Simulation. We compare the number of
the coverage threshold at a specific value, run our protocol activated nodes and the convergence time resulted from
till it converges, and measure the resulting coverage in the simulation versus our theoretical analysis in Section 5. Some
whole area. To approximate area coverage, we measure the of the results are shown in Fig. 8, other results are similar.
coverage of all points of a very dense grid deployed on top The results show that the upper bounds proved in
of the area. A point is considered covered if the probability Theorems 3 and 4 are only worst-case values, and our
of sensing at this point is at least . The dense grid points protocol performs better on the average case.
have spacing of 0:03rs ¼ 0:5 m. We conduct this experiment
for several values of : 0.9, 0.99, 0.999, and 1.0. Note that 6.3 Robustness of PCP
¼ 1 denotes the deterministic (disk) sensing model. In all We show that our protocol is robust against many practical
cases (the figure is not shown), PCP ensured that the whole aspects. We also show that the protocol consumes the
area is covered with the requested coverage threshold . energy of nodes in a uniform manner, and functions
In addition, we check the connectivity of the nodes correctly when multiple nodes start as activators, which is
activated by PCP when the communication range varies important for large-scale sensor networks.
from 15 to 40 m. The maximum separation s in this Location Inaccuracy. We use the same setup described in
experiment is set to 30 m. We measure connectivity as the Section 6.1 except that we add random errors to the ðx; yÞ
fraction of active nodes that are connected. We plot the coordinates of each of the 20,000 deployed nodes. The error
results in Fig. 6. We show the minimum, average, and can be positive or negative, and it is chosen randomly in the
maximum values obtained from the 10 iterations. Confirm- interval ½0; ermax Š. We vary ermax between 0 and 20 m, that
ing our analysis in Theorem 2, our protocol achieves full is, a node could have as much as 20 m of error on any (or
connectivity when rc ! s. both) of its coordinates. For every value of ermax , we run our
Savings and Flexibility Achieved by PCP. We conduct an protocol till it converges, and compute the fraction of the
experiment to assess the potential savings in number of area covered. As shown in Fig. 9a, PCP maintains the whole
active nodes because of using the exponential sensing area covered even in the presence of large location errors.
model instead of the disk sensing model. Fig. 7 shows the This shows the robustness of PCP against location inaccu-
results for different values of the coverage threshold , and racy. There is a slight cost, though, for location inaccuracy.
for a range of values for the sensing decay factor . The We compute the average number of sensors activated by the



Fig. 8. Validation of the theoretical analysis: (a) and (b) fraction of sensors activated by PCP, and (c) convergence time of PCP.

Fig. 9. Robustness of PCP against: (a) inaccurate node locations and (b) imperfect time synchronization. Note that there are two y-axes in each plot.

protocol to maintain coverage. We normalize this number are either in SLEEP or ACTIVE states. Therefore, some of
by the number of sensors needed when there are no location the late nodes may become unnecessarily active.
errors. The results are also shown in Fig. 9a (note that some Random Node Failures. Nodes deployed in real fields
figures have two y-axes). As shown in the figure, location might get damaged, burned, or just fail at any time. We
inaccuracy could increase the number of active sensors. simulate failures at arbitrary times during the lifetime of the
This increase is not large in most practical cases: There is network. In particular, we randomly choose a fraction f of
less than 9 percent increase in the number of active sensors the nodes to be failed during the first 100 rounds of the
for location errors up to 10 m. protocol. We randomly schedule a failure time for each
Imperfect Time Synchronization. Exact, or fine-grained,
node. We change f between 0 and 60 percent. For each
time synchronization of nodes in large-scale sensor net-
value of f, we run our protocol and periodically check the
works is costly to achieve in practice. In this experiment, we
assess the impact of the granularity of time synchronization coverage of the whole area. The results, shown in Fig. 10,
on our protocol. In our protocol, nodes need to know the
start of the round so that they begin executing the protocol.
Nodes will start at exactly the same time if their clocks are
perfectly synchronized. We let clocks of nodes drift with
different random values in the interval ½0; dmax Š, where dmax
is the maximum clock drift. We vary dmax between 0 and
500 ms. For every value of dmax , we run our protocol till it
converges, and compute the fraction of the area covered. As
indicated in Fig. 9b, PCP is fairly robust against clock drifts
because it can tolerate these drifts and maintain coverage. In
addition, for practical clock drifts (up to 300 ms), there is
virtually no increase in the number of activated sensors. For
larger clock drifts, the cost is not significant as shown in
Fig. 9b. Note that PCP converges in about 300 ms, on
average. This explains why the number of active sensors
starts to increase for clock drifts beyond 300 ms: Some
nodes with high clock drifts may start executing the
protocol after others have already terminated it, i.e., they Fig. 10. Robustness of PCP against random node failures.



nodes k from 1 to 9 and plot the number of sensors
activated by PCP to ensure coverage normalized by the
number of active sensors when k ¼ 1. In the same figure,
we plot the normalized convergence time. As expected,
increasing the number of starting points increases the
number of active sensors but makes the protocol converges
faster. In Fig. 12b, we repeat the same experiment but
measure the normalized 80 percent lifetime, which is the
time it takes for the coverage in the network to drop below
80 percent. The figure shows that reducing the conver-
gence time is more beneficial for the network lifetime than
reducing the number of active sensors. This is because
before convergence, many nodes are either in WAIT or
ACTIVE state before the protocol converges, which
consume more energy.
Fig. 11. Uniform energy consumption and network lifetime using PCP.
6.4 Comparing PCP versus Another Probabilistic
indicate that even with high failure rates, PCP maintains Coverage Protocol (CCANS)
coverage in almost all rounds. We compare our PCP protocol against the probabilistic
Uniform Energy Consumption. In this experiment, we coverage protocol (CCANS), proposed in [10], in terms of
show that our protocol distributes the load uniformly across the number of activated sensors, network lifetime, and
all deployed nodes. This is critical in order to keep nodes energy consumption. The idea of CCANS is to start all
alive for the longest possible period, and thus, to prolong nodes in active mode, then iteratively deactivate nodes that
the network lifetime and achieve more reliable coverage. are not needed for coverage. A token is circulated among
We measure the load on a node by the energy consumed by nodes in the network in a certain manner. The node holding
that node. Once a node runs out of energy, it is assumed to the token calculates the coverage on the grid points around
be dead. We run our protocol till all nodes are dead. After it. If coverage is achieved at these points, it broadcasts a
each round of the protocol, we count the number of alive notification to its neighbors, passes the token to another
nodes. We plot the average of the number of alive nodes node, and deactivates itself. All redundant nodes are
versus round number in Fig. 11. As the figure shows, most deactivated when the token visits each node in the network.
of the nodes stay alive till round number 60. Then, they
We implemented CCANS in C++ and validated our
gradually die. This means that the protocol did not
implementation of CCANS by obtaining the same results
overutilize some nodes in early rounds, otherwise, they
in [10]. To conduct fair comparisons, we make CCANS
would have died earlier. Note that the energy of a node is
check only for coverage and not for connectivity.
enough for it to be active in about only five rounds. In
In our comparison, we use the same exponential sensing
addition, Fig. 11 shows that the coverage is maintained in
most of the area throughout the network lifetime. model for CCANS and PCP with the same parameters. The
Multiple Starting Nodes. Finally, we analyze the impact of parameters used for CCANS are taken from [10], and they
multiple starting nodes on the performance of the PCP are: th ¼ 1 and tmax ¼ m . The parameters used for PCP are:
protocol. Multiple starting nodes are desired for large-scale a ¼ m =2 and s ¼ n=Er , where Er is the fraction of the
networks. In Fig. 12a, we change the number of starting remaining energy in the node. is computed as explained in

Fig. 12. Impact of multiple starting nodes on the performance of PCP: (a) convergence time and fraction of activate nodes and (b) network lifetime.



Fig. 13. Comparison between PCP and CCANS: (a) number of activated nodes, (b) total remaining energy, and (c) network lifetime.

Fig. 14. Comparison among PCP, OGDC, and CCP: (a) total remaining energy in all nodes and (b) energy consumption per millisecond on a smaller
timescale.

Section 3.2: For a uniform distribution of 20,000 nodes in a
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi C++. We validated our implementation of OGDC and CCP
1 km Â 1 km area, we have ¼ 2 Â 1;000 2=20; 000 ¼ 20 m. by obtaining the same results in their respective papers. We
We plot in Fig. 13a the average number of nodes use the disk sensing model for all protocols. To conduct a
activated by PCP and CCANS for different values of the fair comparison and remove the overhead imposed by CCP
sensing decay factor . As the figure shows, PCP activates a and OGDC to maintain connectivity, we assume that the
much smaller number of nodes than CCANS, while communication range is twice as the sensing range in all
ensuring the same level of probabilistic coverage. This is experiments for all protocols. The round length is
significant because it indicates that the sensor network 100 seconds for both PCP and OGDC. We set the
could last much longer using our protocol. To validate this parameters p0 in OGDC and s in PCP such that both
claim, we study the fraction of the remaining energy in protocols have a single starting node.
nodes as the time progresses from 0 to 1,000 seconds in We focus our comparison on the energy consumption of
Fig. 13b. The figure shows that as CCANS activates more deployed nodes under different coverage protocols. In
nodes and exchanges more messages than PCP, the node Fig. 14a, we plot the fraction of remaining energy in nodes
energy is depleted at a much faster rate. For example, after as the time progresses. The figure shows that our PCP
1,000 seconds, the average energy of a node is 60 percent of protocol is much more energy conserving than CCP and
its original energy if the sensor network uses CCANS to
OGDC. To better understand the dynamics of energy
maintain coverage, while this average is 90 percent if our
consumption in the network, we plot in Fig. 14b the rate
PCP protocol is used. Finally, the lifetime of the sensor
of energy consumption in terms of energy units per
network is shown in Fig. 13c, where we plot the fraction of
millisecond. The peak in the energy consumption for CCP
the area covered with time. The lifetime of the network
represents the sending of HELLO messages. We notice that
under our protocol is much larger than under CCANS.
the main reason that makes OGDC consumes more energy
6.5 Comparing PCP versus Other Deterministic than CCP and PCP is that it takes longer time to converge,
Coverage Protocols (OGDC and CCP) which is shown by the high energy consumption over a
We have implemented two recent coverage protocols: longer period. To study this issue further, we analyze the
OGDC [4] and CCP [2] that were shown to outperform dynamic change in node states with time. Recall that the
others in the literature. Both protocols are implemented in energy model that we use in the comparison assigns



Fig. 15. State transitions with time for three deterministic coverage protocols: (a) PCP, (b) OGDC, and (c) CCP.

different energy consumption levels for transmitting, coverage case is expected to yield even higher savings in the
receiving, idling, or sleeping. For instance, the amount of number of activated sensors than in the 1-coverage case.
energy consumed in receiving mode is 400 times more than Another extension is to consider probabilistic communication
the energy used in the sleep mode. models, in addition to the probabilistic sensing models, in
We plot in Fig. 15 the number of nodes in each state of the design and operation of the protocol.
the three protocols versus the time. At the beginning of the
execution of these protocols, a large amount of energy is
consumed because all sensors in the field are active [4], [2]. APPENDIX
Fig. 15 explains why PCP achieves the energy saving in Here, we show that under the exponential sensing model
Fig. 14a. For instance, comparing OGDC (Fig. 15b) versus defined in (2), the least-covered point by three sensors
PCP (Fig. 15a), we see that nodes decide to go to sleep much located at vertices of an equilateral triangle is at the center
faster in PCP than in OGDC. Since the total number of of the triangle.
deployed nodes is much larger than the activated subset of From geometric properties of triangles, the following
them, even a small difference in convergence time will relationship holds between any interior point c at distances
make a significant difference in the energy consumption. x; y; z from the vertices of the triangle and the triangle side s
Finally, the convergence time of PCP, OGDC, and CCP (see Fig. 16):
can be inferred from Fig. 15, by finding the time at which
the states of all nodes are decided. The figure shows that the s2 ðx2 y2 þ s2 z2 Þ þ s2 ðs2 y2 þ x2 z2 Þ þ s2 ðs2 x2 þ y2 z2 Þ
ð3Þ
convergence time of PCP and CCP is less than 500 ms, while À s2 ðx4 þ y4 þ z4 Þ À s6 ¼ 0:
it is more than 1,500 ms for OGDC.
Using Definition 2, the probability of sensing at point c is
À ÁÀ ÁÀ Á
7 CONCLUSION AND FUTURE WORK P ðcÞ ¼ 1 À 1 À eÀðxÀrs Þ 1 À eÀðyÀrs Þ 1 À eÀðzÀrs Þ : ð4Þ
In this paper, we proposed and evaluated a fully distributed, Note that using (3), we can replace z in (4) as a function
probabilistic coverage protocol. A key feature of our protocol
of x and y, making P ðcÞ a function of only two variables.
is that it can be used with different sensing models, with
minimal changes. We analyzed our protocol and showed that Now we show that P ðcÞ is minimum at the center of the
it converges fast and has a small message complexity. We triangle by showing that the partial derivatives of P ðcÞ with
verified our analytical results using simulations. We also respect to x and y are zeros, and the second derivatives are
implemented our protocol and three other coverage proto- positive, when x ¼ y ¼ z. Due to symmetry, we show that
cols: one of them is probabilistic (CCANS) and the other two only for x:
are deterministic (OGDC and CCP). Our extensive experi-
mental study shows that our protocol activates less sensors
than the others while maintaining the same level of coverage,
and consumes much less energy. In addition, we showed
through simulation that a probabilistic coverage model may
result in significant savings in the number of activated
sensors, which reduces energy consumption and extends the
network lifetime. It also provides a flexible way to control the
number of activated sensors versus the level of coverage
achieved by the protocol.
The work in this paper can be extended in several
directions. For example, the analysis and design of our
coverage protocol can be extended to the probabilistic k-
coverage case. k-coverage is needed in several sensor
network applications to enhance reliability and accuracy Fig. 16. Location of the least-covered point in an equilateral triangle
of the network. Using probabilistic sensing models in the k- formed by three sensors.



@P ðcÞ À ÁÀ Á
¼ ÀeÀðxÀRÞ 1 À eÀðyÀRÞ 1 À eÀðzÀRÞ We know that for any ; R 0 and x ! R:
@x ð5Þ
@z ÀðzÀRÞ À ÁÀ Á
À e 1 À eÀðxÀRÞ 1 À eÀðyÀRÞ : 2 0;
@x
@z eÀðxÀRÞ 0;
In order to find @x , we differentiate (3) with respect to x: À Á
1 À eÀðxÀRÞ 0:
@z 2 @z @z
2xy2 þ 2 s z þ 2xz2 þ 2 x2 z þ 2s2 x þ 2 y2 z
@x @x @x Moreover, it is known that for any w 2 R, eÀw ! 1 À w.
@z @z Àxðy2 þ z2 þ s2 À 2x2 Þ Therefore,
À 4x3 À 4 z3 ¼ 0 ) ¼ :
@x @x zðs2 þ x2 þ y2 À 2z2 Þ
ð6Þ 1ÀeÀðxÀRÞ ðx À RÞ x
At the center point, we have x ¼ y ¼ z, which yields 2 À Á
)2À 1 À eÀðxÀRÞ 0:
x
@z Àxðx2 þ x2 þ s2 À 2x2 Þ
¼ ¼ À1: This means that all four terms in (9) are positive. Therefore,
@x xðs2 þ x2 þ x2 À 2x2 Þ
the second derivative of coverage probability is positive.
Substituting @x ¼ À1 into (5) results in @P ðcÞ ¼ 0. In a
@z
@x
similar manner, it is easy to show that the second derivative ACKNOWLEDGMENTS
of P ðcÞ with respect to x is positive:
This work is partially supported by the Natural Sciences
@ 2 P ðcÞ À ÁÀ Á and Engineering Research Council (NSERC) of Canada.
¼ 2 eÀðxÀRÞ 1 À eÀðyÀRÞ 1 À eÀðzÀRÞ
@x2
@z 2 ÀðxÀRÞ À Á REFERENCES
À e 1 À eÀðyÀRÞ eÀðzÀRÞ
@x [1] I. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci,
@2z À ÁÀ Á “Wireless Sensor Networks: A Survey,” Computer Networks,
À 2 eÀðzÀRÞ 1 À eÀðxÀRÞ 1 À eÀðyÀRÞ vol. 38, no. 4, pp. 393-422, Mar. 2002.
@x
2 [2] G. Xing, X. Wang, Y. Zhang, C. Lu, R. Pless, and C. Gill,
@z À ÁÀ Á “Integrated Coverage and Connectivity Configuration for Energy
þ 2 eÀðzÀRÞ 1 À eÀðxÀRÞ 1 À eÀðyÀRÞ Conservation in Sensor Networks,” ACM Trans. Sensor Networks,
@x
vol. 1, no. 1, pp. 36-72, Aug. 2005.
@z 2 ÀðzÀRÞ ÀðxÀRÞ À Á [3] Z. Zhou, S. Das, and H. Gupta, “Connected K-Coverage Problem
À e e 1 À eÀðyÀRÞ : in Sensor Networks,” Proc. Int’l Conf. Computer Comm. and
@x
Networks (ICCCN ’04), pp. 373-378, Oct. 2004.
ð7Þ [4] H. Zhang and J. Hou, “Maintaining Sensing Coverage and
@2 z Connectivity in Large Sensor Networks,” Ad Hoc and Sensor
We differentiate (3) twice to find @x2 : Wireless Networks: An Int’l J., vol. 1, nos. 1/2, pp. 89-123, Jan. 2005.
2 [5] S. Shakkottai, R. Srikant, and N. Shroff, “Unreliable Sensor Grids:
@z @2z @z @2z Coverage, Connectivity, and Diameter,” Ad Hoc Networks, vol. 3,
2y2 þ 2 s2 þ 2 2 s2 z þ 2z2 þ 4 xz þ 2 2 x2 z no. 6, pp. 702-716, Nov. 2005.
@x @x @x @x [6] S. Kumar, T.H. Lai, and J. Balogh, “On K-Coverage in a Mostly
2 2
2
@z @z @ z @z Sleeping Sensor Network,” Proc. ACM MobiCom, pp. 144-158, Sept.
þ 4 xz þ 2 x2 þ 2s2 þ 2 2 y2 z þ 2 y2 2004.
@x @x @x @x [7] F. Ye, G. Zhong, J. Cheng, S. Lu, and L. Zhang, “PEAS: A Robust
2 Energy Conserving Protocol for Long-Lived Sensor Networks,”
@2z @z
À 12x2 À 4 2 z3 À 12 z2 ¼ 0: Proc. Int’l Conf. Distributed Computing Systems (ICDCS ’03), pp. 28-
@x @x 37, May 2003.
@z [8] D. Tian and N. Georganas, “A Coverage-Preserving Node
Again we have x ¼ y ¼ z, and @x ¼ À1 at the center point: Scheduling Scheme for Large Wireless Sensor Networks,” Proc.
First ACM Int’l Workshop Wireless Sensor Networks and Applications,
@2z 2 2@ 2 z 3 pp. 32-41, Sept. 2002.
2x2 þ 2ðÀ1Þ2 s2 þ 2 s x þ 2x2 À 4x2 þ x [9] Y. Zou and K. Chakrabarty, “Sensor Deployment and Target
@x2 @x2
2 Localization in Distributed Sensor Networks,” ACM Trans.
@ z Embedded Computing Systems, vol. 3, no. 1, pp. 61-91, Feb. 2004.
À 4x2 þ 2x2 þ 2s2 þ 2 2 x3 þ 2x2 À 12x2
@x [10] Y. Zou and K. Chakrabarty, “A Distributed Coverage- and
Connectivity-Centric Technique for Selecting Active Nodes in
@2z 3 2 @ 2 z 24x2 À 4s2 Wireless Sensor Networks,” IEEE Trans. Computers, vol. 54, no. 8,
À 4 2 x À 12x ¼ 0 ) 2 ¼ :
@x @x 2s2 x pp. 978-991, Aug. 2005.
pffiffiffi [11] N. Ahmed, S. Kanhere, and S. Jha, “Probabilistic Coverage in
Since the triangle is equilateral, s ¼ 3x. Thus, Wireless Sensor Networks,” Proc. IEEE Conf. Local Computer
Networks (LCN ’05), pp. 672-681, Nov. 2005.
@ 2 z 24x2 À 12x2 2 [12] B. Liu and D. Towsley, “A Study on the Coverage of Large-Scale
¼ ¼ : ð8Þ Sensor Networks,” Proc. IEEE Int’l Conf. Mobile Ad-Hoc and Sensor
@x2 6x3 x
Systems (MASS ’04), pp. 475-483, Oct. 2004.
Finally, we substitute (8) into (7) and use x ¼ y ¼ z: [13] Q. Cao, T. Yan, T. Abdelzaher, and J. Stankovic, “Analysis of
Target Detection Performance for Wireless Sensor Networks,”
!
@ 2 P ðcÞ À Á 2 À Á Proc. Int’l Conf. Distributed Computing in Sensor Networks, pp. 276-
2
¼ 2 eÀðxÀRÞ 1 À eÀðxÀRÞ 2 À 1 À eÀðxÀRÞ : 292, June 2005.
@x x [14] M. Cardei and J. Wu, “Energy-Efficient Coverage Problems in
ð9Þ Wireless Ad Hoc Sensor Networks,” Elsevier Computer Comm.,
vol. 29, no. 4, pp. 413-420, 2006.



[15] X. Bai, S. Kumar, D. Xuan, Z. Yun, and T. Lai, “Deploying Mohamed Hefeeda (S’01, M’04, SM’09) re-
Wireless Sensors to Achieve Both Coverage and Connectivity,” ceived the BSc and MSc degrees from Mansoura
Proc. ACM MobiHoc, pp. 131-142, May 2006. University, Egypt, in 1994 and 1997, respec-
[16] Z. Abrams, A. Goel, and S. Plotkin, “Set K-Cover Algorithms for tively, and the PhD degree from Purdue Uni-
Energy Efficient Monitoring in Wireless Sensor Networks,” Proc. versity, West Lafayette, Indiana, in 2004. He is an
Int’l Symp. Information Processing in Sensor Networks (IPSN ’04), assistant professor in the School of Computing
pp. 424-432, Apr. 2004. Science, Simon Fraser University, Surrey, British
[17] D. Tian and N. Georganas, “Location and Calculation-Free Node Columbia, Canada, where he leads the Network
Scheduling Schemes in Large Wireless Sensor Networks,” Elsevier Systems Lab. His research interests include
Ad Hoc Networks, vol. 2, pp. 65-85, 2004. multimedia networking over wired and wireless
[18] T. Wu and K. Ssu, “Determining Active Sensor Nodes for networks, peer-to-peer systems, network security, and wireless sensor
Complete Coverage without Location Information,” Int’l J. Ad networks. He has served on several technical program committees of
Hoc and Ubiquitous Computing, vol. 1, nos. 1/2, pp. 38-46, 2005. major conferences in his research areas, including ACM Multimedia,
[19] S. Meguerdichian, F. Koushanfar, M. Potkonjak, and M. Srivastava, ACM Multimedia Systems, ACM/SPIE Multimedia Computing and
“Coverage Problems in Wireless Ad-Hoc Sensor Networks,” Proc. Networking (MMCN), the IEEE Conference on Network Protocols (ICNP),
IEEE INFOCOM, pp. 1380-1387, Apr. 2001. and the IEEE Conference on Communications (ICC). He is an associate
[20] D. Mehta, M. Lopez, and L. Lin, “Optimal Coverage Paths in Ad- editor of the International Journal of Advanced Media and Communica-
Hoc Sensor Networks,” Proc. IEEE Int’l Conf. Comm. (ICC ’03), May tion and the guest editor of that journal’s special issue on High-Quality
2003. Multimedia Streaming in P2P Environments. His paper on the hardness
[21] X. Li, P. Wan, and O. Frieder, “Coverage in Wireless Ad Hoc of optimally broadcasting multiple video streams with different bit rates
Sensor Networks,” IEEE Trans. Computers, vol. 52, no. 6, pp. 753- won the Best Paper Award in the IEEE Innovations 2008 Conference. In
763, June 2003. addition to publications, he and his students develop actual systems,
[22] L. Lazos and R. Poovendran, “Stochastic Coverage in Hetero- such as PROMISE, pCache, svcAuth, pCDN, and mobile TV testbed, and
geneous Sensor Networks,” ACM Trans. Sensor Networks, vol. 2, contribute the source code to the research community. The mobile TV
no. 3, pp. 325-358, Aug. 2006. testbed software developed by his group won the Best Technical Demo
[23] C. Huang and Y. Tseng, “The Coverage Problem in a Wireless Award in the ACM Multimedia 2008 Conference. He is a senior member
Sensor Network,” ACM Mobile Networks and Applications of the IEEE and a member of the ACM SIGCOMM and SIGMM.
(MONET), special issue on wireless sensor networks, vol. 10,
no. 4, pp. 519-528, Aug. 2005. Hossein Ahmadi received the BSc degree from
[24] A. So and Y. Ye, “On Solving Coverage Problems in a Wireless Sharif University, Iran, in 2005, and the MSc
Sensor Network Using Voronoi Diagrams,” Proc. Workshop Internet degree from Simon Fraser University, Canada,
and Network Economics (WINE ’05), pp. 584-593, Dec. 2005. in 2007. He is currently working toward the PhD
[25] T. Okabe, B. Boots, K. Sugihara, and S.N. Chiu, Spatial Tessellations: degree at the University of Illinois, Urbana-
Concepts and Applications of Voronoi Diagrams, second ed. John Champaign. His research interests are in the
Wiley, 2000. area of wireless sensor networks.
[26] C. Huang, Y. Tseng, and H. Wu, “Distributed Protocols for
Ensuring Both Coverage and Connectivity of a Wireless Sensor
Network,” ACM Trans. Sensor Networks, vol. 3, no. 1, Mar. 2007.
[27] K. Chakrabarty, S. Iyengar, H. Qi, and E. Cho, “Grid Coverage for
Surveillance and Target Location in Distributed Sensor Net-
works,” IEEE Trans. Computers, vol. 51, no. 12, pp. 1448-1453, Dec. . For more information on this or any other computing topic,
2002. please visit our Digital Library at www.computer.org/publications/dlib.
[28] S. Yang, F. Dai, M. Cardei, J. Wu, and F. Patterson, “On Connected
Multiple Point Coverage in Wireless Sensor Networks,” Int’l J.
Wireless Information Networks, vol. 13, no. 4, pp. 289-301, May 2006.
[29] F. Dai and J. Wu, “An Extended Localized Algorithm for
Connected Dominating Sets Formation in Ad Hoc Wireless
Networks,” IEEE Trans. Parallel and Distributed Systems, vol. 15,
no. 10, pp. 908-920, Oct. 2004.
[30] H. Gupta, Z. Zhou, S. Das, and Q. Gu, “Connected Sensor Cover:
Self-Organization of Sensor Networks for Efficient Query Execu-
tion,” IEEE/ACM Trans. Networking, vol. 14, no. 1, pp. 55-67, Feb.
2006.
[31] M. Hefeeda and M. Bagheri, “Randomized K-Coverage Algo-
rithms for Dense Sensor Networks,” Proc. IEEE INFOCOM,
pp. 2376-2380, May 2007.
[32] A. Savvides, C. Han, and M. Strivastava, “Dynamic Fine-Grained
Localization in Ad-Hoc Networks of Sensors,” Proc. ACM
MobiCom, pp. 166-179, July 2001.
[33] L. Doherty, L.E. Ghaoui, and K. Pister, “Convex Position
Estimation in Wireless Sensor Networks,” Proc. IEEE INFOCOM,
pp. 1655-1663, Apr. 2001.
[34] NS-2, http://nsnam.isi.edu/nsnam/, 2009.
[35] Network Systems Lab, http://nsl.cs.sfu.ca/wiki, 2009.


Energy efficient protocol for deterministic

Recommended

More Related Content

What's hot (18)

Viewers also liked (8)

Similar to Energy efficient protocol for deterministic (20)

More from ambitlick (20)

Recently uploaded (20)

Energy efficient protocol for deterministic