Вы находитесь на странице: 1из 14

714

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 2, FEBRUARY 2015

Maximum Lifetime Scheduling for Target Coverage


and Data Collection in Wireless Sensor Networks
Zaixin Lu, Wei Wayne Li, Senior Member, IEEE, and Miao Pan, Member, IEEE

AbstractTarget coverage and data collection are two fundamental problems for wireless sensor networks (WSNs). Target
coverage is needed to select sensors in a given area that can
monitor a set of interesting points. Data collection is needed to
transmit the sensed data from sensors to a sink. Since, in many
applications, sensors are battery powered, it is expected that a
WSN can work untended for a long period. This paper addresses
the scheduling problems for both target coverage and data collection in WSNs with the objective of maximizing network lifetime.
First, a polynomial-time approximation scheme is developed for
the case where the density of target points is bounded, and then, a
polynomial-time constant-factor approximation algorithm is developed for the general case. It is also proved that it is N P-hard
to find a maximum lifetime scheduling of target cover and data
collection for a WSN, even if all the sensors have the same sensing
radius and the same transmission radius. Further, the practical
efficiency of our algorithms is analyzed through simulation. These
extensive simulation results show better performances of our algorithms compared with other research findings.
Index TermsApproximation algorithm, data collection, lifetime maximization, N P-hard, target coverage, wireless sensor
network (WSN).

I. I NTRODUCTION
A. Background and Motivation

N many wireless sensor network (WSN) applications, sensors are spatially distributed in a finite area to monitor
physical or environmental conditions, such as pressure, humidity, temperature, etc., and to transmit the sensed data to
a base station cooperatively [6], [27]. In addition, at times,
a set of target points has to be monitored in a given area.
On the one hand, to provide deterministic quality-of-service
guarantees, every point of interest should be monitored by at
least one sensor at all times. On the other hand, the energy
consumption of sensors should be minimized since, in most
cases, sensors are battery powered. Therefore, sensors should
have their power supplies turned off when they are not in use to
Manuscript received October 11, 2013; revised February 24, 2014; accepted
April 22, 2014. Date of publication May 7, 2014; date of current version
February 9, 2015. This work was supported in part by the National Science
Foundation under Grant NSF-1137732, Grant NSF-1241626, and Grant NSF1343361 and in part by the Air Force Office of Scientific Research under Grant
FA8650-13-C-5800. The review of this paper was coordinated by Prof. J. Tang.
Z. Lu is with the National Science Foundation Center for Research on
Complex Networks, Texas Southern University, Houston, TX 77004 USA
(e-mail: luz@tsu.edu).
W. W. Li and M. Pan are with the Department of Computer Science and
the National Science Foundation Center for Research on Complex Networks,
Texas Southern University, Houston, TX 77004 USA (e-mail: liww@tsu.edu;
panm@tsu.edu).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TVT.2014.2322356

preserve energy [5], [36], [37]. Due to this limitation, a critical


issue becomes how to prolong the lifetime of WSNs while also
assuring the service quality of coverage. Thus, research on the
energy-efficient sensor coverage problem has been extensively
investigated in the literature (see, e.g., [3], [4], [10], [13][19],
[22][26], [28], [29], [31], [39], and [42]).
For a typical target coverage problem in WSNs, the network
lifetime is defined as the time duration that all the target points
are monitored. As pointed out in [14] and [36], network lifetime
can be prolonged by alternating the working modes of sensors
between settings of on and off. In other words, this means
to schedule the entire time duration into a number of rounds
and to only turn on the power supplies of a subset of sensors
to monitor the target points in each round. Fig. 1(a) shows
an example. There are three target points and three sensors,
in which sensors s1 , s2 , and s3 can monitor target points
(p1 , p2 ), (p2 , p3 ), and (p1 , p3 ), respectively. Supposing that all
the sensors can work two time units, then by alternating the
on and off modes, we can monitor all the target points for
three time units. A possible schedule would be the following: In
the first time unit, turn on (s1 , s2 ) and turn off s3 ; in the second
time unit, turn on (s2 , s3 ) and turn off s1 ; in the third time unit,
turn on (s1 , s3 ) and turn off s2 . According to this schedule,
all the sensors only work two time units, and all the target
points are monitored for three time units. If the sensor modes
are not switched, then any target point can only be monitored
for two time units. Therefore, to prolong the network lifetime,
it is necessary to develop efficient algorithms to schedule the
sensors and perform the monitoring tasks.
In addition to coverage, sensors are also required to transmit
the sensed data to a base station, which consumes a significant
amount of energy. To reduce that energy consumption, multihop
communications are always applied to transmit data from sensors to the base station [6]. Therefore, to let M denote the set
of sensors that monitor the target points in a time interval , we
need to select a path from each sensor in M to the base station,
such that all the sensed data in can be transmitted to the base
station, and the total energy consumed for both monitoring and
data transmission for every sensor involved does not exceed its
residual energy. If these paths are unwisely selected, the network lifetime will decrease. Fig. 1(b) shows such an example.
There are five sensors, two target points, and a base station. The
target points p1 and p2 are in the sensing ranges of sensors s1
and s2 , respectively. According to the network structure shown
in Fig. 1(b), s1 can directly transmit its sensed data to base
station B, but s2 cannot. There are two possible paths, i.e.,
(s2 s1 B) and (s2 s3 s4 s5 B), from s2 to
B. Intuitively, the first path is shorter than the second path, and

0018-9545 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

LU et al.: MAXIMUM LIFETIME SCHEDULING FOR TARGET COVERAGE AND DATA COLLECTION IN WSNs

715

3) To evaluate the practical efficiency of our algorithms, we


compared them with the state-of-the-art algorithm [29]
based on real-world scenarios. These extensive simulation results show that longer network lifetime can be
obtained by using our algorithms rather than using the
algorithm proposed in [29].
C. Paper Organization
The rest of this paper is organized as follows: Section II
introduces the previous work for coverage and lifetime maximization in WSNs. Section III introduces the system model
used in this study and the problem statement for MLCS. In
Section IV, we present our theoretical results. We first prove
the N P-hardness of MLCS in Section IV-A and then present a
primal and dual method for MLCS in Section IV-B, as well as
two approximation algorithms that can be applied in the primal
and dual method in Section IV-C and D, respectively. Finally,
we provide simulation results in Section V and conclude this
paper with a discussion about future research in Section VI.
II. R ELATED W ORK

Fig. 1.

Illustration of (a) target coverage and (b) data collection.

thus, the first path consumes less energy. However, using the
first path to transmit the sensed data from s2 to B will decrease
the network lifetime because it consumes the additional energy
of s1 . Therefore, we need to transmit data wisely when there
are multiple paths from a sensor to the base station.

B. Our Contribution
In this paper, we study the Maximum Lifetime Coverage
Scheduling (MLCS) problem for WSNs, considering both target coverage and data collection. Our contributions follow.
1) Generally, a sensor si can monitor a target point pj if and
only if pj is in the sensing range of si . In addition, the
energy consumed for transmitting a bit of data between
two sensors will depend on their Euclidean distance. In
this research, we find that MLCS in a Euclidean plane is
N P-hard, even if all the sensors have the same sensing
radius and the same transmission radius.
2) To overcome the computational intractability of MLCS,
we develop a polynomial-time approximation algorithm
having a constant-factor ratio to general MLCS. In
addition, we develop a polynomial-time approximation
scheme for MLCS, assuming the number of target points
is bounded by a constant in a unit area. This assumption
is of practical interest because in many real applications,
the target points are uniformly distributed. In addition, the
distance between any two target points should not be too
close; otherwise, we can merge them into a single target
point.

Coverage is a hot topic in WSNs. There are two major types


of coverage problems. One is how to deploy sensors to cover
a set of target points; the other is how to schedule sensors that
have been already deployed, so they will cover a certain area or
a set of target points. In [8], Kar and Banerjee studied the sensor
deployment problem and proposed a constant-factor approximation algorithm to minimize the number of sensors required to
cover a given set of target points. In [30], Nieberg et al. further
proposed an approximation scheme for this minimum sensor
deployment problem. In [9], Dasgupta et al. studied the sensor
deployment problem with the objective of prolonging network
lifetime. They developed a deterministic sensor deployment
strategy that outperformed the random sensor deployment strategy by nearly 40%.
Although using deterministic sensor deployment can prolong
network lifetime, in many applications, the sensors are randomly deployed in a large area. In the past decade, how to
select sensors with predetermined locations to cover an area
or a set of target points has become an important problem
for WSNs. To maximize network lifetime, Cardei et al. [14]
first studied how to schedule the working modes of sensors
to preserve energy and maintain the service quality of the
coverage. They defined two types of sensor modes i.e., on
and off, and developed a linear programming algorithm and
a greedy algorithm to switch the sensor modes, such that the
coverage lifetime could be prolonged. In [15], Berman et al.
showed that convex area coverage can be converted into target
coverage by adding O(M 2 ) target points, where M denotes
the number of sensors in the given area. They then proposed a
primal and dual method with the performance ratio 1/O(log N )
for the coverage lifetime maximization problem [14], where N
is the number of target points. Inspired by the works in [2] and
[15], Ding et al. [38] developed a constant-factor approximation
algorithm, assuming that all sensors have a uniform sensing
range. In addition to sensor mode scheduling, in [18], Lu et al.

716

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 2, FEBRUARY 2015

studied how to prolong coverage lifetime by configuring the


sensing ranges of sensors.
It is worthy to mention also that none of the aforementioned
research efforts considered the energy consumption for data
collection. All of them assumed that the sensors with the same
working mode have the same energy consumption rate, which
restricts their applications, since in the real world, the energy
consumption of a sensor depends on how much data is sensed,
received, and transmitted. To efficiently transmit the sensed
data to the base station, Chang and Tassiulas [12] studied the
routing problem in WSNs and developed a linear programming
algorithm and a shortest path algorithm to transmit data in
WSNs with the target of prolonging network lifetime. In [7],
Bhardwaj and Chandrakasan developed an optimal assignment
algorithm to analyze the lifetime bounds for data collection
in WSNs. In [11], the maximum data collection problem was
investigated further, and an approximation algorithm was developed by Sadagopan and Krishnamachari. In [29], Zhao and
Gurusamy considered the energy consumption of both target
coverage and data collection so as to maximize the network
))-factor approximation
lifetime. They developed a 1/O(H(N


) =
algorithm, where H(N
i=1 (1/i), and N is the maximum
number of targets in the sensing area of a sensor. In [20], Du and
Lin developed a differentiated coverage algorithm that provided
different coverage degrees for different areas in terms of the
importance of the area (i.e., some key areas are covered by
multiple sensors). In [26], Liu et al. further investigated the kcoverage and data collection problem in WSNs, where network
lifetime is defined as the duration up to the time when there
exists one target point that cannot be monitored by k sensors, or
the sensed data cannot be transmitted to the base station. After
that work, Lin et al. [40] considered the k-coverage problem
in hybrid sensor networks where each node is equipped with
various types of sensors. They found that the optimal solution
for the k-coverage was a pure Nash equilibrium, and thus, an algorithm was designed based on game theory. In addition to full
coverage, many research efforts have been done on constructing
energy-efficient WSNs to partially cover a given area or a subset
of target points (see, e.g., [15], [19], [33], [34], and [41]).
III. P ROBLEM F ORMULATION
In the study, we assume that the base station, wireless sensors, and target points are in a finite Euclidean plane. We also
assume that all the sensors have the same transmission radius
Rt and sensing radius Rs . Let B be the base station, S = {s1 ,
s2 , . . . , sM } be the set of sensors, and P = {p1 , p2 , . . . , pN }
be the set of target points. A sensor si covers a point pj if the
Euclidean distance between si and pj is no greater than Rs , and
it can send data to B if the Euclidean distance between si and
B is no greater than Rt . Similarly, two sensors si and sj can
communicate with each other if the Euclidean distance between
si and sj is no greater than Rt . Hence, we can formulate the
sensors and the base station as an undirected graph G(V, E),
where V = S {B} and (vi , vj ) E if and only if vi and
vj can communicate with each other. In addition, each sensor
si V is associated with a set P(si ) P of target points that
are covered by si . Without loss of generality, we assume that

Fig. 2. Coverage and data collection tree. (a) CDCT. (b) Sensor repeated
CDCT.

pi P, sj S, such that pi P(sj ), and si S, there is


at least one path available from si to B in G.
A. Coverage and Data Collection
Let M S denote a sensor set that covers all the target
points (i.e., for any target point pi P, there exists at least one
sensor sj M, such that pi P(sj )). A coverage and data
collection tree (CDCT) for M in an operation time interval
(OTI) , as defined in [29], is a tree rooted at B and used for
transmitting the sensed data from M to B during . It is clear
that every leaf node in a CDCT performs the monitoring task. In
addition, some nonleaf nodes may also perform the monitoring
task. Meanwhile, every node except for B sends data to its
parent node to perform the data collection task. As shown in
Fig. 2(a), nodes s1 , s2 , and s3 monitor target points p1 , p2 ,
and p3 , respectively. For data collection, s2 and s3 send their
sensed data to s1 , and s1 sends all the sensed data to B. In this
example, s1 is not a leaf node but still performs the monitoring
task to cover p1 .
To turn all the monitoring sensors into leaf nodes, we repeat
the sensors in a CDCT. Here, we further define the sensor
repeated version of CDCT as a tree rooted at B that satisfies
1) only the leaf nodes perform the monitoring task; 2) a sensor
can appear, at most, one time at the leaf nodes but may appear
multiple times at intermediate nodes; and 3) each node in S has
exactly one parent. Fig. 2(b) shows the sensor repeated version
of the CDCT in Fig. 2(a), where s1 appears three times in total
(once at a leaf node for monitoring p1 and twice at intermediate
nodes for transmitting the sensed data from s2 and s3 to B).
The CDCT in Fig. 2(b) is essentially the same as that shown
in Fig. 2(a), but it is described in a way that is better suited

LU et al.: MAXIMUM LIFETIME SCHEDULING FOR TARGET COVERAGE AND DATA COLLECTION IN WSNs

to formulate the energy consumption. In the remainder of this


paper, we use the repeated version of CDCT for the problem
formulation, algorithm design, and theoretical analysis.

the energy consumption model described in Section III-B, we


can formulate MLCS into a linear programming, i.e.,
Maximize

B. Energy Model
We consider the following energy consumption scenarios.
Initially, each sensor si S has a battery with a specific
residual energy e0 (si ). When sensors perform the monitoring
task, they not only consume energy but also generate data. We
assume that all the sensors are equipped with the same hardware
and software (i.e., have the same sampling frequency, coding
scheme, etc.) and, thus, the same data generation rate (bits
per unit time) and energy consumption rate (joules per unit
time). Therefore, the amount of data generated by a monitoring
sensor during an OTI is bits, and the amount of energy
consumed is joules. For data transmission, as pointed out
in [29], the amount of energy consumed by a sensor in receiving
a bit of data is fixed, but the amount of energy consumed
when transmitting a bit of data depends on the transmission
distance. Therefore, we let er (si , sj ) = er be the amount of
joules consumed by sensor sj for receiving a bit of data from
any sensor si , where er is a constant, and let et (si , sj ) = et +
d(si , sj ) be the amount of joules consumed by sensor si
when transmitting a bit of data to sensor sj , where et , , and
are constants, and d(si , sj ) is the Euclidean distance between
si and sj .
Let T = (L, N L, E) be a CDCT for an OTI , where L and
N L are the sets of leaf and nonleaf nodes, respectively, and E
is the set of transmission links. For a node v T , we denote
by p(v) the parent node of v (i.e., (v, p(v)) E) and s(v) the
sensor at node v. Then, the energy consumption rate of each
node v except for the root B can be formulated as

+ et (s(v), s (p(v))) ,
vL
e(v, T ) =
(er (s(v), s (p(v)))+et (s(v), s (p(v)))) , v N L.
(1)
Since a sensor may appear at multiple nodes in T , the total
energy consumption
rate of a sensor si according to T is

e(si , T ) = vT ,s(v)=si e(v, T ), and the amount of energy
consumed by si during is e(si , T ) .
C. Lifetime Maximization
Assume that all the sensors have nonzero residual energy
initially. We call a sensor alive at some point if its residual
energy is greater than zero and a target point uncovered if there
exists no path of alive sensors from the target point to the
base station. In this paper, the network lifetime of a WSN is
defined as the duration of network operation time until the first
uncovered target point appears. We next present here the formal
definition of MLCS.
Definition 1MLCS: Given a set P of target points and a set
S of sensors in a finite Euclidean plane, MLCS is the problem
of finding a schedule of CDCTs to cover all the target points,
such that the network lifetime is maximized.
Let T1 , T2 , . . . , TQ be the possible CDCTs that can cover all
the target points and 1 , 2 , . . . , Q be their OTIs. According to

717

subject to

Q

i
L(P, S)
i=1

Q


e(si , Tj ) j e0 (si ) si S.

(2)

j=1

The number of CDCTs may be exponential in the input size.


Therefore, it is computationally expensive to solve the given
linear programming directly when there is a large number of
sensors. However, this formulation is useful for developing
approximation algorithms for MLCS. We will discuss this point
later.
IV. T HEORETICAL R ESULTS
A. NP-Hardness
Here, we investigate the hardness of MLCS. It has been
shown in [29] that the MLCS problem is N P-hard when we
formulate the WSN as a general graph where two sensors can
communicate with each other (or a sensor covers a target point)
if and only if there is an edge between them. We next show that
the MLCS problem in a Euclidean plane is still N P-hard.
Theorem 1: MLCS is N P-hard even if all the sensors have
the same transmission radius and the same sensing radius.
Proof: We sketch a polynomial-time reduction from the
dominating set problem in unit disk graphs. A unit disk graph
is an undirected graph of uniform-size disks located in a Euclidean plane, in which every disk is a node, and two nodes
are connected by an edge if and only if they have a nonempty
intersection. A dominating set for a graph G(V, E) is a node
set D V , such that every node not in D is adjacent to one
or multiple nodes in D. Therefore, a dominating set of a unit
disk graph is a subset D of disks, such that every disk not in
D intersects with at least one disk in D. It is well known that
given a general graph G = (V, E) and an integer K > 0, decide
whether there exists a dominating set D of size K in G is N Phard, and this hardness holds, even if G(V, E) is a unit disk
graph [1].
Given a unit disk graph G and an integer K > 0 as the input
of a dominating set problem, we construct an MLCS problem as
follows. Let r denote the radius of disks in the unit disk graph.
For each disk di in V , we create a sensor si of sensing radius 2 r and transmission radius 4 r at the center of di
and create a target point pi at the same location. Let sl
be the leftmost sensor and (xl , yl ) the coordinates of sl .
We create an additional sensor sg at (xl + 4r, yl ) and a
base station B at (xl + 8r, yl ) (see Fig. 3). Let = K(er +
et + (2r) ) and
= K(er + et + (4r) ) ,
where is an arbitrary positive number, and other variables are the energy parameters defined in Section III-B.
We complete the reduction by setting the residual energy
e0 (si ) = + for sensors si = sg and setting e0 (sg ) =
,
where is the energy consumption rate for performing the
monitoring task. We next show that the unit disk graph G has a
dominating set of size K if and only if the MLCS problem has
a solution with network lifetime .

718

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 2, FEBRUARY 2015

MLCS, we study the dual problem of MLCS and solve MLCS


by using a primal and dual method. This section is devoted to
briefly introducing the primal and dual method proposed in [2]
for solving packing linear programming problems.
In linear programming (1), the inequality guarantees that the
energy consumed by every sensor si does not exceed its residual
energy e0 (si ). It can be formulated as Ax b and x 0,
where A is an M Q matrix, x = [1 , 2 , . . . , Q ] is the vector
denoting the OTIs, and b = [e0 (s1 ), e0 (s2 ), . . . , e0 (sM )]T is
the vector denoting the residual energy. This is a typical packing
linear programming, and its dual linear programming can be
formulated as follows:
Minimize

bT y

subject to AT y I, y 0

Fig. 3. Convert unit disk dominating set into MLCS. (a) Unit disk graph.
(b) MLCS.

Consider that G has a dominating set D of size K. Then,


for every disk di G, there exists a disk dj D, such that the
distance between the centers of di and dj is no greater than 2r.
Therefore, in the MLCS problem, the set of K sensors at the
centers of those K disks cover all the target points. To monitor
target points, each sensor generates bits of data for an OTI
. We assume that G is connected; otherwise, we can find the
minimum dominating set for each component independently.
Then, the K sensors can transmit their sensed data to sensor
sl via multihop communication such that any transmission
distance is no greater than 2r. The energy consumption is, at
most, + for any sensor other than sg . Therefore, all the
sensed data can be transmitted to sg , and sg can transmit exact
K bits of data to the base station B. Hence, the network
lifetime is . On the contrary, consider that all dominating
sets of G are larger than K. The additional sensor sg cannot
cover any target points. Therefore, we cannot use K sensors to
cover all the target points. Since sg is the only sensor that can
transmit data to B and sl is the nearest sensor to sg among all
the sensors, sg can receive and transmit, at most, K bits
of data to the base station. It directly follows that the network
lifetime is less than .
Therefore, G has a dominating set of size K if and only if
the network lifetime of MLCS is . The proof of Theorem 1 is
complete.

B. Primal and Dual Method
As mentioned in Section III-C, the MLCS problem can be
formulated as a linear programming, but the number of variables may be exponential in the input size. To efficiently solve

(3)

in which I is a vector, all elements in which are one. The


primal and dual method solves the primal problem by finding
the row of matrix AT with the minimum length according to the
T
vector y in the dual problem. Let ATi denote the ith
 rowTof A .
T
T
The length of Ai is defined as length(Ai , y) = j (A (i, j)
y(j)). Assume that we have an algorithm f that finds a row
in AT with length no greater than r mini (length(ATi , y)) for
any vector y, we can get a 1/r-factor approximation solution
to the primal problem by using algorithm f to solve the dual
problem. The detailed steps are presented in Algorithm 1.
Algorithm 1 Primal and Dual Method
Input: A parameter  > 0, a primal linear programming
max{I T x|Ax b, x 0} and its dual min{bT y|AT y
I, y 0}.
Output: A solution (A , x ) to the primal problem, where A is
a subset of columns of A, and x is the set of nonzero
values for x.
1: Let A and x ;
2: Assign a small value (/b(i)) > 0 to each element y(i) in
y such that bT y < 1;
3: while bT y < 1do
4: Apply algorithm f to find a row ATi in AT according to y;
5: Let c arg minj (b(j)/ATi (j)) and let xi (b(c)/
ATi (c));
6: Put (ATi )T into A and xi into x ;
7: For each nonzero element ATi (j) ATi , let y(j)
y(j)(1 +  (b(c) ATi (j)/b(j) ATi (c)));
8: end while
9: Let x (x / log1+ (1 + /)) and return (A , x );
The primal and dual method (Algorithm 1) works as follows:
It uses two sets A and x to store the selected columns in A
and the corresponding values in x, respectively. Initially, both
A and x are empty, i.e., all values in x are zero, and the values
in y are assigned according to a parameter (see Steps 1 and 2).
It then iteratively adds columns of A into A and updates x
and y according to algorithm f until the objective value of
the dual problem is greater than or equal to 1 (see Steps 38).

LU et al.: MAXIMUM LIFETIME SCHEDULING FOR TARGET COVERAGE AND DATA COLLECTION IN WSNs

At the end, it normalizes the values in x to guarantee a feasible


solution to the primal problem (see Step 9). Now, consider the
meaning of Algorithm 1 for our MLCS problem. Each row ATi
of AT denotes a CDCT Ti , whereas an element ATi (j) ATi
denotes the energy consumption rate e(sj , Ti ) of sensor sj
when Ti is used to perform the monitoring and data collection
tasks. We can consider y as a coefficient vector for the sensors,
i.e., y(j) is the coefficient of sj . Algorithm 1 first assigns an
initial coefficient /b(j) to each sensor sj , where b(j) = e0 (sj )
denotes the residual energy of sj . The two sets of A and x
hold the selected CDCTs and their OTIs, respectively. Initially,
both A and x are empty. After initialization, Algorithm 1
runs round by round. In each round, it selects a row ATi
corresponding to a CDCT Ti and places it into A . The OTI
i for Ti is equal to xi . In addition, the coefficient y(j) for
each sensor sj Ti is increased by  (b(c) ATi (j)/b(j)
ATi (c)) =  (e0 (sc ) e(sj , Ti )/e0 (sj )e(sc , Ti )), where c satisfies minj (e0 (sj )/e(sj , Ti )) = (e0 (sc )/e(sc , Ti )). When the
value of the objective function in the dual linear programming
is more than 1, the algorithm stops. Still, the total energy
consumption of some sensors may exceed their residual energy
according to A and x . Therefore, we need to normalize the
OTI of each selected CDCT by a factor of log1+ (1 + /).
Let w(sj , Ti ) = e(sj , Ti ) y(j) denote the weight of sj in Ti .
The length of a row ATi of AT is equal to the sum of weights
of all the sensors in Ti . Therefore, if we have an approximation
algorithm to the problem of finding minimum weight CDCT
(MWCDCT; Definition 2) for any coefficient vector y > 0, we
can get an approximation solution to MLCS.
Definition 2MWCDCT: Given a set P of target points and
a set S of sensors in a finite Euclidean plane, and given energy
parameters: , , er , et , , , and an M -dimensional vector
y > 0, where M is the number of sensors, MWCDCT is the
problem of finding
a CDCT T to cover all the target points in

P, such that si S (e(si , T ) y(i)) is minimized.
Theorem 2: There exists an (1 )/r-approximation algorithm to MLCS if there exists an r-approximation algorithm f
to the MWCDCT problem for any coefficient vector y > 0.
Proof: Assume that we have an r-approximation algorithm f to the MWCDCT for any coefficient vector y > 0.
Consider applying f in Algorithm 1. Let yk denote the coefficient vector in round k. Initially, we have bT y0 = M ,
where M is the number of sensors. Suppose in round k, f finds
a CDCT Ti according to yk1
 and let c = arg minj (e0 (sj )/
e(sj , Ti ), we have bT yk = sj S (e0 (sj ) yk1 (j)) + 

Let (y) denote
sj Ti (e0 (sc )e(sj , Ti ) yk1 (j)/e(sc , Ti )).

the weight of MWCDCT for vector y. Then, sj Ti (e(sj , Ti )
yk1 (j)) r (yk1 ). Let Tk denote the sum of OTIs in x
for the selected CDCTs until round k, we have
bT yk bT yk1 + 
r (yk1 )(Tk Tk1 )


bT y 0 +  r

(yj1 )(Tj Tj1 )

j=1...k

M +  r

j=1...k

(yj1 )(Tj Tj1 ) . (4)

719

Let = miny (bT y/(y)). Then, is the optimal objective value for the dual problem, and thus, (bT yj1 /(yj1 )) for
any j k. Therefore, by inequality (4), we have

 bT yj1
(Tj Tj1 )
bT y k M +  r

j=1...k


r
M +
bT yj1 (Tj Tj1 ) . (5)

j=1...k

If we let W (y0 )=M and W (yk ) = W (yk1 )+(r/)


W (yk1 )(Tk Tk1 ), we have W (y0 ) bT y0 . By induction,
it can be shown that W (yk ) bT yk for any k. In addition, we
have
W (yk ) = W (yk1 ) +
e
e
=e

r
W (yk1 )(Tk Tk1 )

r(Tk Tk1 )

rTk

rTk

W (yk1 )

W (y0 )
M.

(6)

Therefore, when Algorithm 1 stops at round k, we have 1


bT yk erTk / M , which implies that Tk ( ln(1/
M )/ r). After scaling, we have Tk (/r (1 + )2 ) when
(1 + ) ((1 + ) M )1/ . To complete the proof, it is
sufficient to show that (A , x ) is a feasible solution after the
scaling. If we suppose that Algorithm 1 selects a CDCT Ti in
a round l(l k), then the energy consumption of any sensor
sj is increased by e0 (sc ) e(sj , Ti )/e(sc , Ti ), and the coefficient yl (j) is increased by yl1 (j) e0 (sc ) e(sj , Ti )/eo (sj )
e(sc , Ti ). Let ek (sj ) be the total energy consumed by sj afk
ter round k. By induction, we have (1 + )e (sj )/e0 (sj )
yk (j) 1 + . Therefore, ek (sj ) log1+ (1 + /). In sum,
we derive an 1/(r (1 + )2 )-approximation to MLCS. It is
clear that we can choose  to make 1 (1/(1 + )2 ) for
any constant .

By Theorem 2, we can use an MWCDCT algorithm to solve
the MLCS problem. An approximation algorithm to MWCDCT
will bring us an approximation algorithm to MLCS with nearly
the same performance ratio.
C. PTAS to Target-Bounded MLCS
Here, we show that there is a polynomial-time approximation
scheme to target-bounded MLCS by presenting a polynomialtime (1 )-approximation algorithm to MWCDCT. Here,
target bounded means that the number of target points per unit
area is bounded by a constant. There are two phases in our approximation algorithm. We first transfer the weight of each path
into its leaf node. Note that in our CDCTs, only the leaf nodes
perform both the monitoring and data collection tasks, and the
other nodes only perform the data collection task. Therefore,
for each sensor si S, we create a path P (si ) = {si1 , si2 ,
= B, and the path weight
. . . , sia }, such that si1 = si , sia 
w(P(si )) = y(i1 )( +et (si1 , si2 ))+ j=2...a1 y(ij )(er(sij1 ,
sij ) + et (sij , sij+1 )) is minimized, where y is the coefficient
vector, er (sij1 , sij ) is the energy consumption rate when

720

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 2, FEBRUARY 2015

sij receives data from sij1 , and et (sij , sij+1 ) is the energy
consumption rate when sij transmits data to sij+1 . This way,
we can focus on the coverage problem and solve MWCDCT
by first finding a minimum weight sensor cover with weights
w(si ) = w(P (si )) and then converting it into an MWCDCT.
Before we present our approximation algorithm formally, some
notations must be introduced first.
Regardless of the communication constraints, we call two
sensors si and sj neighbors if and only if P(si ) P(sj ) = ,
where P() denotes the set of target points covered by
and define the 1-hop neighborhood of a sensor si by N (si ) =
of si }. For a set S
S of sensors, we
{sj S|sj is a neighbor

define N (S ) = si S  N (si ), and the r-hop neighborhood of


si (r 2) is recursively defined by N r (si ) = N (N r1 (si )).
For a set P
P of target points, we denote the minimum

weight sensor set that covers


 P by MWSC(P ) and its weight

by w(MWSC(P )) = si SC(P  ) w(si ). Based on the aforementioned notations, the pseudocode of our algorithm is given
in Algorithm 2.
Algorithm 2 Polynomial-Time Approximation Scheme
Input: A positive constant > 0, a set P of target points, a set
S of sensors and a vector y.
Output: A minimum weight CDCT.
1: Let SC (SC holds the selected sensors);
2: For each sensor si , create a minimum weight path P (si )
and let w(si ) = w(P (si ));
3: while P = do
4: For each target point pi (P \ PSC ) where PSC denotes
the target points covered by SC, let smin (pi ) denote the
minimum weight sensor that covers pi ;
5: Let pj arg maxpi (P\PSC ) w(smin (pi ));
6: Find the minimum positive integer r satisfying that
w(MWSC (P (N r+1 (smin (pj ))))) (1 + )w(MW
SC(P(N r (smin (pj )))));
7: Let SC SC MWSC(P(N r+1 (smin (pj )))) and let
P P \ P(N r+1 (smin (pj )));
8: end while
9: Let T (SC) be the CDCT where SC performs the monitoring task, and each sensor si SC transmits the sensed
data to the base station by path P (si ), return T (SC);
The polynomial-time approximation scheme (see
Algorithm 2) to construct a CDCT works as follows. It
uses a set SC to store the sensors that perform the monitoring
tasks, and initially, SC is empty (see Step 1). For each sensor
in S, it computes the weight by finding a minimum weight
path from that sensor to the base station (see Step 2). During
Steps 38, it continually adds sensors to SC until all the target
points are covered. Each time it considers a target point pj
that has to be covered by a sensor with the maximum weight
and finds an optimal solution to cover all the target points in
P(N r+1 (smin (pj ))). When all the target points are covered by
SC, it constructs the CDCT by adding paths from each sensor
in SC to the base station (Step 9). To show the correctness of
Algorithm 2, we first show the following lemmas.

Lemma 1: If SC S is a minimum weight sensor


set that

covers all the target points with weight w(SC) = si SC w(si )
and w(si ) = w(P (si )), then T (SC) is an MWCDCT.
Proof: Consider an arbitrary CDCT T (L, N L, E). Note
T is defined as w(si , T ) =
that the weight of sensor si in
e(si , T ) y(i) and e(si , T ) = vT ,s(v)=si e(v, T ), where
s(v) denotes the sensor at node v. Let yv = y(i) for nodes
{v
 T |s(v) = si }, then the sum of weights of T is equal to
vT e(v, T ) yv . For any leaf node v L, if we denote the
path from v 
to the root B in T by {v1 , v2 , . . . , vb }, where vb =
B, we have j=1b1 e(vj , T ) y(vj ) w(P (s(v))) because
P (s(v)) is a minimum weight path from s(v) to B according
to
 y. Therefore, the sum of weights of T is no less than
vL w(P (s(v))). Since in T , only the nodes in L perform
the monitoring task, L is a sensor set that
 covers all the target
points.
Therefore,
we
have
w(SC)

vL w(P (s(v)))

e(v,
T
)

y
.

v
vT
Lemma 2: Let the sensing radius be 1. If the density of target
points per unit area (1 1) is bounded by a constant c, for any
constant > 0, there exists a constant integer r() that only
depends on such that w(MWSC(P(N r()+1 (smin (pj )))))
is no greater than (1 + ) w(MWSC(P(N r() (smin (pj )))))
for any sensor smin (pj ) selected in Step 5 of Algorithm 2.
Proof: For an arbitrary positive constant and a sensor
smin (pj ) selected by Algorithm 2, assume that w(MWSC

(P(N r +1 (smin (pj ))))) is strictly greater than (1 + ) w(M

WSC(P(N r (smin (pj ))))) for integers r
= 1, 2, . . .. Then, on
the one hand, we have




w MWSC P N r +1 (smin (pj ))



> (1 + ) w MWSC P N r (smin (pj ))


> (1 + )r w (MWSC (P (N (smin (pj )))))




(1 + )r w (smin (pj ))

(7)

in which the last inequality is because the target point pj in


P(N (smin (pj ))) cannot be covered by any sensor with weight
less than w(smin (pj )) according to Step 5 in Algorithm 2. On

the other hand, the area covered by sensor set N r +1 (smin (pj ))

2
is, at most, (2r + 1) because two sensors are neighbors
if and only if the Euclidean distance between them is no
more than 2. Hence, there are, at most, c (2r
+ 1)2 target

points in P(N r +1 (smin (pj ))). According to Steps 4 and 5 in
Algorithm 2, any target point in P \ PSC can be covered by a
sensor with weight no more than w(smin (pj )). It follows that




w MWSC P N r +1 (smin (pj ))
c (2r
+ 1) w (smin (pj ))
2

O(r
2 ) w (smin (pj )) .


(8)

In sum (7) and (8), we have (1 + )r O(r


2 ). The left-hand
side is exponential in r
, and the right-hand side is polynomial in r
. The inequality is only true when r
= O((1/)
ln(1/)). Therefore, there exists a constant integer r() = O
((1/) ln(1/)), such that w(MWSC(P(N r()+1(smin (pj )))))
is no greater than (1+)w(MWSC(P(N r()(smin (pj ))))). 

LU et al.: MAXIMUM LIFETIME SCHEDULING FOR TARGET COVERAGE AND DATA COLLECTION IN WSNs

Theorem 3: Algorithm 2 is a polynomial-time (1 )approximation algorithm to the MWCDCT problem when the
number of target points is bounded by a constant in a unit area.
Proof: It is clear that SC covers all the target points
when Algorithm 2 stops. During Steps 38, it iteratively selects
a sensor smin (pj ) and find a sensor set MWSC(P(N r+1
(smin (pj )))) to cover target point set P(N r+1 (smin (pj ))),
such that the weight of MWSC(P(N r+1 (smin (pj )))) is, at
most, (1 + ) w(MWSC(P(N r (smin (pj ))))). In addition,
any sensor covers target points in P(N r (smin (pj ))) cannot
cover any target point out of P(N r+1 (smin (pj ))). Therefore,
the sum of weights of SC is greater than the weight of the
minimum weight sensor cover, at most, by a factor of (1 + ).
By Lemma 1, in Step 9 of Algorithm 2, T (SC) is a (1 + )approximation solution to MWCDCT. By Lemma 2, there are
only a constant number of target points in P(N r+1 (smin (pj )))
when the number of target points is bounded by a constant c in a
unit area. Therefore, each time MWSC(P(N r+1 (smin (pj ))))
can be efficiently computed in polynomial time. In sum,
Algorithm 2 is a polynomial-time (1 )-approximation algorithm to MWCDCT.

Theorem 4: There exists a polynomial-time approximation
scheme to MLSC when the density of target points is bounded.
Proof: It follows directly from Theorems 2 and 3.

D. Constant Approximation to General MLCS
Here, we present a polynomial-time constant-factor approximation algorithm to general MLCS, regardless of the density of
the target points. As with the previous approximation scheme,
our constant-factor approximation algorithm solves MLCS by
creating a minimum weight path from each sensor to the base
station and converting MWCDCT into the minimum weight
sensor coverage problem. In [21], Ambuhl et al. first developed
a 72-factor approximation algorithm to the minimum weight
dominating set problem in a unit disk graph, which could be
applied to solve the minimum weight sensor cover problem.
The algorithm is based on an optimal solution under the
condition that all the target points are in a strip, and all the
sensors are outside the strip. The state-of-the-art approximation
algorithm to the minimum weight sensor cover was developed
by Zou et al. [35]. However, it is a theoretical result and
not suitable for practical uses. To solve MLCS efficiently, we
develop a constant-factor approximation algorithm for the minimum weight sensor cover. Its performance is worse than the
algorithm proposed in [35], but it is easy to implement. Before
we introduce the details of our algorithm and its theoretical
analysis, we first want to introduce some of the properties when
using sensors outside a rectangle to cover the target points in
that rectangle.
Fig. 4 shows a rectangle in a Euclidean plane where the
outside region is divided into eight areas. First, if there is a
target point in the rectangle covered by a sensor in the South
area (or respectively, the East, West, North areas), we can draw
two lines across the target point and the border line of the
rectangle to form an isosceles right triangle, such that all the
target points in the isosceles right triangle are covered by that
sensor. As an example, p1 is a target point covered by a sensor

721

Fig. 4. Target coverage in a rectangle.

Fig. 5. Rectangle partition.

in the South area; hence, it is clear that all the target points in
(p1 , b, c) are covered by that sensor. Second, if there are two
target points in the rectangle covered by sensors in the South
area (or, respectively, the North area) and not covered by any
sensor in the East and West areas, we can draw two lines across
the two target points, respectively, to form an isosceles right
triangle, such that all the target points in the isosceles right
triangle are not covered by any sensor in the East or West areas.
As an example, assume that p1 and p2 are two target points
covered by two sensors in the South area and not covered by
any sensor in the East or West areas. Then, all the target points
in (a, b, e) are not covered by any sensor in the East or West
areas. We can prove it by contradiction. Assume that p is such
a target point in (a, b, c) covered by a sensor in the East area.
Then, we can draw two lines across p and the East borderline
to form an isosceles right triangle. Since p2 is in this triangle, it
is covered by that sensor according to the first property, which
is a contradiction. Based on these two properties, Huang et al.
[32] proved the following lemma.
Lemma 3: Given a set of target points in a rectangle and a
set of sensors outside the rectangle, we can select four target
points p1 , p2 , p3 , and p4 to draw two isosceles right triangles as
shown in Fig. 5 such that every target point lying inside the two
triangles is not covered by any sensor in the East and West areas
and that every target point lying outside the two triangles is not
covered by any sensor in the North and South areas.
We next present our constant-factor approximation algorithm
in Algorithm 3.
Algorithm 3 Constant-Factor Approximation
Input: A set P of target points, a set S of sensors and a vector y.
Output: A minimum weight CDCT.
1: Let SC (SC holds the selected sensors);
2: For each sensor si , create a minimum weight path P (si )
and let w(si ) = w(P (si ));
3: Randomly select a point (x0 , y0 ) on the southwest of all
the target points
partition the whole plane into rectan and
gles of size ( 2 ( 2/2)) starting from (x0 , y0 ).
4: For each rectangle containing target points in P, find a set
S
S of sensors with the minimum weight to cover all
the target points in the rectangle;

722

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 2, FEBRUARY 2015

5: Let SC SC S
, S S \ S
, and P P \ P(S
);
6: Let T (SC) be the CDCT where SC performs the monitoring task, and each sensor si SC transmits the sensed data
to the base station by the path P (si ), return T (SC);

The constant-factor approximation (see Algorithm 3) still


uses a set SC to store the sensors that perform the monitoring
tasks and computes the weights of sensors by finding minimum
weight paths (see Steps 1 and 2), which are the same as that
in Algorithm 2. After that, it randomly partitions the whole
Euclidean plane into small rectangles (see Step 3), finds sensors
to cover each rectangle sequentially, and adds them into SC (see
Steps 4 and 5). Finally, it converts SC into a CDCT T (SC)
and outputs T (SC) (Step 6). It is worth mentioning that in
Step 4, an optimal or approximation algorithm to the minimum
weight
setcover problem, where all the target points are in
a ( 2 ( 2/2)) rectangle, is needed for our approximation
algorithm design. When the number of target points is bounded
by a constant, it is clear that an optimal solution can be obtained
by enumerating all the possible solutions. However,
general
for
cases, the number of target points, even in a ( 2 ( 2/2))
rectangle, may be so large that enumerating all the possible
solutions is computationally expensive. We next show the
existence of a polynomial-time 2-approximation algorithm to
the minimum weight
cover problem when the target
sensor
points are in a ( 2 ( 2/2)) rectangle. For simplicity of
description, we assume that the sensing radius is 1.
Theorem 5: There exists a polynomial-time 2-factor approximation algorithm to
minimum
weight sensor cover if all the
target points are in a ( 2 ( 2/2))rectangle.
as two
Proof: Consider the ( 2 ( 2/2)) rectangle

consecutive squares sq1 and sq2 , each of size (( 2/2) ( 2/


2)). Then, a sensor si located in sq1 (respectively sq2 ) covers all
the target points in sq1 (respectively sq2 ). Assume that SC opt is
an optimal solution. There are three possible cases: 1) SC opt
consists of two sensors lying in sq1 and sq2 , respectively;
2) SC opt consists of a set of sensors lying outside the rectangle;
and 3) SC opt consists of a set of sensors in which only one
sensor is in the rectangle.
In the first case, the two sensors lying in sq1 and sq2 with the
minimum weight is the optimal solution.
In the second case, we assume that there are more than four
target points since, otherwise, an optimal solution can be found
in polynomial time. By Lemma 3, we can partition the target
points into two disjoint sets P1 and P2 such that every target
point in P1 is not covered by any sensor in SC opt lying inside
the East and West areas and that every target point in P2 is
not covered by any sensor in SC opt lying inside the North
and South areas. In [21], Ambuhl et al. developed a dynamic
programming (DP) algorithm to find the minimum weight set
of sensors lying outside a strip to cover all the target points
in the strip. Consider P1 and P2 as two sets of target points
in a horizontal strip and a vertical strip, respectively. We can
apply Ambuhls DP algorithm to find a minimum weight set
MWSC(P1 ) of sensors to cover P1 and a minimum weight
set MWSC(P2 ) of sensors to cover P2 . It is clear that both
w(MWSC(P1 )) and w(MWSC(P2 )) are no greater than

Fig. 6.

Number of rectangle a sensor covers.

w(MWSC opt ). Therefore, MWSC(P1 ) MWSC(P2 ) is a


2-factor approximation solution.
In the last case, we assume that SC opt has a sensor lying
in sq1 (we can prove it in the same way when SC opt has
a sensor lying in sq2 ), then there are only N (sq1 ) choices,
where N (sq1 ) denotes the number of target points in sq1 . We
can try each sensor lying in sq1 and use the same method
as in the second case to find a 2-approximation solution for
the remainder of target points in sq2 . The combination is a
2-approximation for the whole rectangle.
In all the three cases, we can find a 2-approximation or a
better solution. Therefore, Theorem 5 is proved.

We next analyze the performance ratio of Algorithm 3.
Theorem 6: Algorithm 3 is a polynomial-time constantfactor approximation algorithm to MWCDCT.
Proof: To prove Thoerem 6, we analyze how many rectangles is covered by a sensor in an optimal solution SC opt . Let
rec(x, y) be an arbitrary rectangle in the randomized partition,
as shown in Fig. 6. We calculate the expected number of
rectangles covered by a sensor in it. If we divide rec(x, y) into
four small rectangles with the same size, we just have to analyze
one small rectangle, and others will have the same result.
Consider the upper-left rectangle in rec(x, y). If a sensor is in
the shaded area, it covers, at most, eight rectangles. If a sensor
is in the unshaded area, it covers,at most,ten rectangles. The
proportion of shaded area is (2 2 2/ 2) 0.58. Therefore, with randomized partition, each sensor in SC opt covers
less than nine rectangles with probability greater than half
(see Fig. 6).
Let SC opt (rec(x,
 y)) be an optimal solution for rec(x, y).
Then, we have x,y w(SC opt (rec(x, y))) 9 w(SC opt ) for
a randomized partition. By Theorem 5, we can find a 2approximation
solution SC app
(rec(x, y)) for rec(x, y). Since

w(SC
(rec(x,
y)))
app
x,y
x,y (2w(SC app (rec(x, y))))
18 w(SC opt ), it follows from Lemma 2 that Algorithm 3
is a polynomial-time 18-factor approximation algorithm to
MWCDCT.

In the proof of Theorem 5, we roughly analyze the average
number of rectangles covered by a sensor in a random partition.
A more careful analysis can indeed reduce the approximation
ratio of Algorithm 3. For example, in the dark shaded area of
rec(x, y) in Fig. 6, a sensor only covers six rectangles.
Theorem 7: There exists a polynomial-time constant-factor
approximation algorithm to general MLCS.

LU et al.: MAXIMUM LIFETIME SCHEDULING FOR TARGET COVERAGE AND DATA COLLECTION IN WSNs

v L, s(v) = si

curacy, we set = 2 in Algorithm 2. In the constant-factor


approximation, we apply Algorithm 1 and use Algorithm 3
to solve the MWCDCT problem. We still set the parameter
= (1 + )((1 + )M )(1/ in Algorithm 1.
In addition to the approximation scheme and the constantfactor approximation, we implement the greedy-based algorithm proposed in [29] for comparison purposes. It also uses
the primal and dual method. However, instead of finding a
MWCDCT, it iteratively finds the paths from the base station
to sensors to cover the target points. For the sake of reducing
the total weight, it greedily selects a path each time, such that
the ratio of the path weight over the number of newly covered
target points is minimized. It has been shown in [29] that
)-approximation to
the greedy algorithm is a 
1/(1 + )2 H(N

is the maxi
MLCS, where H(N ) = 1iN (1/N ), and N
mum number of target points in the sensing area of a sensor.
The algorithms are written in JAVA, and the experiments are
conducted on a DELL Inspiron with Intel Core i5 2.5 GHz and
4-GB RAM. To the best of our knowledge, we do not find any
better approximation algorithm in the literature for scheduling
both target coverage and data collection with the objective of
maximizing network lifetime.

v N L, s(l(v)) = sj .
(9)

B. Energy Settings

Proof: The proof follows directly from Theorems 2


and 6.

By Theorems 4 and 7, we can develop a polynomial-time approximation scheme to MLCS when the number of target points
is bounded in a unit area, and we can develop a polynomial-time
constant-factor approximation algorithm to general MLCS. In
the theoretical analysis, we assume that all sensors have the
same data generation rate (i.e., sensors have the same sampling
frequency, coding scheme, etc.). A similar assumption was also
applied to design scheduling algorithms for various lifetime
maximization problems in WSNs [14], [29], [38]. However,
in some applications, different sensors may have different data
generation rates. To make the work more general for practical
applications, we can modify the problem energy model, and the
theoretical analysis still holds. In the problem formulation, we
can let i and i denote the data generation rate and the energy
consumption in sensing for each sensor si , then the energy
consumption rate for each node in a CDCT T = (L, N L, E)
can be formulated as
e(v, T
)
i + i et (s(v), s (p(v)))
= j (er (s(v), s (p(v)))

+ et (s(v), s (p(v))))

723

Here, s(v) denotes the sensor at node v, and l(v) denotes


the leaf node in the subtree rooted at v. The total energy
consumption rate of a sensor si according to T is equal to

vT ,s(v)=si e(v, T ). By a similar argument, it is easy to
verify that Algorithms 13 still can be applied.
V. S IMULATION
Here, we evaluate the practical efficiency of the proposed
algorithms. We are interested in the impact of both the network parameters and the algorithm parameters on the network
lifetime. The ranges of parameters are as follows: The number
of sensors, i.e., M , varies between 80 and 100; the number of
target points, i.e., N , varies between 10 and 100; the sensing
radius Rs varies between 15 and 25 m; the transmission radius
Rt varies between 30 and 50 m; and the parameter  of the
primal and dual method varies between 0.1 and 0.5. In addition
to our approximation scheme and constant-factor approximation, we also implement the state-of-the-art algorithm [29] for
MLCS. The algorithms and energy settings are described in
Section V-A and B, respectively, and the experiential results are
presented in Section V-C.
A. Algorithms
Based on the primal and dual method (see Algorithm 1)
and the MWCDCT algorithms (see Algorithms 2), we implemented an approximation scheme and a constant-factor approximation to MLCS. In the approximation scheme, we apply
Algorithm 1 with parameter = (1 + )((1 + )M )(1/) and
use the (1 + )-approximation scheme (Algorithm 2) to solve
the MWCDCT problem. Thus, the approximation ratio is
1/(1 + ) (1 + )2 . Considering both running time and ac-

The energy consumption parameters are chosen according


to the energy settings in [29]. As defined in Section III, the
energy consumption for transmitting a bit of data from sensor
si to sensor sj is et + d(si , sj ) . In the experiments, et is
set to 50 nJ, is set to 4, and is set to 100 pJ. The energy
consumption for receiving a bit of data for any sensor is set
to er = 150 nJ, and the energy consumption for sensing a bit
of data is set to es = 150 nJ. For any sensor performing the
monitoring task, it produces data at a rate of 10 k bits per unit
time, and initially, all the sensors have the same residual energy
of 20 J.
C. Experimental Results
We performed three experiments to evaluate the network
lifetime of the three algorithms. In the first experiment, we
compared the algorithms in random WSNs, where 80 sensors
and 50 target points were located in a 100 m 100 m Euclidean
plane uniformly at random, and the base station B was placed
in the middle of the area. For each algorithm, we simulated
100 times to get the Average Network Lifetime (ANL). In each
simulation, we guaranteed that every target point was covered
by at least one sensor and that every sensor could transmit data
to B directly or via multihop communication.
Figs. 79 show the simulation results of the first experiment.
Fig. 7 shows the ANL curves with the primal and dual parameter  varying from 0.1 to 0.5, sensing radius Rs = 20 m, and
transmission radius Rt = 40 m. First, in Fig. 7, we note that the
ANL for each algorithm gradually decreases as  increases. This
finding agrees with the intuition that increasing  will decrease
the performance guarantee. Second, the approximation scheme,
which has the best theoretical performance guarantee, always
outperforms the other two algorithms. The constant-factor

724

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 2, FEBRUARY 2015

Fig. 7.  = 0.1, , 0.5, Rs = 20 m, and Rt = 40 m.

Fig. 8.  = 0.1, Rs = 20 m, and Rt = 30 m, , 50 m.

Fig. 9.  = 0.1, Rs = 15 m, , 25 m, and Rt = 40 m.

approximation ranks second. When  is large, the ANL of the


constant-factor approximation is close to the approximation
scheme, and when  is small, the ANL of the constant-factor
approximation is close to the greedy algorithm (e.g., when  =

Fig. 10.  = 0.1, , 0.5, Rs = 20 m, and 40 m.

0.5, the constant-factor approximation outperforms the greedy


algorithm by about 5%; when  = 0.1, the constant-factor approximation outperforms the greedy algorithm by about 3%).
Fig. 8 shows the simulation results with Rt varying from 30
to 50 m ( = 0.1 and Rs = 20 m). We can see that the ANL
curves for all algorithms increase when Rt varies from 30
to 35 m, whereas the increments slow down when Rt varies
from 35 to 50 m. This circumstance is probably because there
are few data transmissions between distant sensors in all the
algorithms. The approximation scheme and the constant-factor
approximation still outperform the greedy algorithm. Fig. 9
shows the simulation results with Rs varying from 15 to 25 m
( = 0.1 and Rt = 40 m). First of all, the ANL curves of all
the algorithms greatly increase as Rs increases, which indicates
that the sensing radius is indeed important for the network
lifetime. When Rs is small, all the algorithms perform similarly,
e.g., when Rs = 15, the gap between the approximation scheme
and the greedy algorithm is less than 1%. However, when
Rs 18, the gap between them is about 4%6%.
From the first experiment, we can deduce that longer network
lifetime can be obtained by using our approximation algorithms
rather than using the greedy algorithm. Moreover, the parameter
 and the sensing radius Rs have a great impact on the network
lifetime. Last but not least, when the transmission radius is
greater than some threshold (35 m in this experiment), then the
network lifetime cannot be prolonged by further increasing the
transmission radius.
In the second experiment, we compared the three algorithms
in grid WSNs, i.e., the sensors and the target points were located
on grid points. This scenario is of practical interest because in
many applications, WSNs have grid topologies. We simulated
a 9 9 square grid in a 90 m 90 m Euclidean plane and
located 100 sensors and 100 target points, respectively, on the
grid points. The base station was placed in the middle of the
area. Since the impact of transmission radius Rt on the network
lifetime is little, its experiential results are omitted henceforth.
Figs. 10 and 11 demonstrate the simulation results from the
second experiment. Similar to the results of the first experiment,
the approximation scheme performs better than the constantfactor approximation and the greedy algorithm. As shown in

LU et al.: MAXIMUM LIFETIME SCHEDULING FOR TARGET COVERAGE AND DATA COLLECTION IN WSNs

Fig. 11.  = 0.1, Rs = 15 m, , 25 m, and Rt = 40 m.

Fig. 10, when  increases from 0.1 to 0.5 (Rs = 20 m and


Rt = 40 m), the approximation scheme outperforms the greedy
algorithm by 5%15%, and it outperforms the constant-factor
approximation by 3%9%. In Fig. 11, unlike the first experiment, when  and Rt remain unchanged and Rs increases, the
ANL curves of all algorithms are stepped. When Rs increases
from 15 to 19, the ANLs of all algorithms remain unchanged.
This is reasonable because every sensor covers exactly the same
target points when 15 Rs 19 in our grid WSN. When Rs
increases from 19 to 23, the ANLs of all the algorithms start
to increase. Specifically, when Rs increases from 21 to 23, the
ANLs are nearly doubled. This is also reasonable because when
Rs = 21 a sensor can cover 15 grid points and when Rs = 23,
a sensor can cover 23 grid points. From the second experiment,
we determine that our approximation algorithms are still better
than the greedy algorithm for prolonging the network lifetime
of grid WSNs. The sensing radius is indeed important for
network lifetime. A slight increment in the sensing radius may
actually and significantly prolong the network lifetime.
In a typical WSN, a significant amount of energy is consumed in transmitting the sensed data to the base station. Since
the sensors near the base station will transmit more data than
the sensors far away from the base station, we can prolong the
network lifetime by employing the sensors nonuniformly. It has
been shown in [28] that the network lifetime can be prolonged
in WSNs by using Gaussian distribution. Therefore, in the
third experiment, we compared the algorithms in Gaussian
distribution WSNs, in which the base station was placed in
the middle of a 100 m 100 m Euclidean plane, and 50
target points were uniformly located at random. For sensor
deployment, we first located 40 sensors uniformly at random
to guarantee the coverage and then located 40 sensors with
Gaussian distribution. Hence, the total number of sensors is
80 as with the first experiment. For each algorithm, we still
simulated 100 times to get the ANL.
The ANL improvements are exhibited in Figs. 12 and 13
in which symbol (U) denotes the simulation results of the
first experiment (i.e., all the 80 sensors are uniformly located),
and symbol (G) denotes the simulation results of the third
experiment (i.e., 40 sensors are uniformly located, and 40

725

Fig. 12.  = 0.1, , 0.5, Rs = 20 m, and Rt = 40 m.

Fig. 13.  = 0.1, Rs = 15 m, , 25 m, and Rt = 40 m.

sensors are located with Gaussian distribution). In Figs. 12


and 13, we can see clearly that using the Gaussian distribution
can prolong the network lifetime greatly, e.g., when  = 0.1,
Rt = 40 m, and Rs = 20 m, the ANL of approximation scheme
is prolonged about 13% by using the Gaussian distribution
instead of uniform distribution, and the ANL for the constantfactor approximation and greedy algorithm are prolonged even
more by applying the Gaussian distribution. In addition, our
approximation algorithms are still better than the greedy algorithm in terms of ANL in the Gaussian distribution WSNs. The
ANL of approximation scheme is greater than that of the greedy
algorithm by 1%7%, and the constant-factor approximation is
better than the greedy algorithm by 1%5%.
Finally, we wanted to evaluate the running time of the
proposed algorithms so that we could validate the scalability.
The evaluation was conducted by varying one parameter, respectively, and fixing others. Table I shows the running time
results. Greedy is faster than our approximation algorithms
in most cases because of its low computational complexity.
However, our approximation algorithms also scale well. The
average running time for all the algorithms highly depends on
, M , and Rs and relatively independently on N and Rt . This

726

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 2, FEBRUARY 2015

TABLE I
RUNNING T IME P ERFORMANCE ( S )

VI. C ONCLUSION
In this paper, we have addressed the sensor scheduling
problem, considering both target monitoring and data collection
to maximize network lifetime. We studied the problem both
theoretically and practically. On the one hand, we prove that
it is N P-hard to maximize the network lifetime even if all the
sensors and target points are in a Euclidean plane. On the other
side of the issue, we develop a polynomial-time approximation
scheme, assuming that the density of target points is bounded,
and a constant-factor approximation algorithm for the general
case. To evaluate the practical efficiency of our algorithms,
we compared them with the greedy algorithm proposed in
[29]. Extensive simulation results show that our approximation
algorithms are better than the greedy algorithm for maximizing
network lifetime. Our future work will extend this research to
the directional sensor coverage problem. In addition, we investigated the MLCS by assuming that the sensors have uniform
sensing radius in this work. As a direction for further research,
the MLCS can be studied without applying this assumption.

ACKNOWLEDGMENT
result makes sense because when  is small, and M and Rs
are large, the primal and dual method will construct a large
number of CDCTs. When Rt increases from 40 to 50 m, the
running time changes for all the algorithms are minor; however,
when Rs increases from 15 to 25, the average running time for
all the algorithms gradually increases. Intuitively, when N is
increasing, the average running time should be getting longer.
However, in our experiment, the average running time does not
monotonically increase when N increases, e.g., when N = 20,
the average running time for the three algorithms are 3.48,
3.89, and 3.62, respectively, and when N = 40, the average
running time for the three algorithms are 2.90, 3.48, and 3.17,
respectively. This is probably because when N = 20, the primal
and dual method constructs much less CDCTs than it constructs
when N = 10. In such a case, although a CDCT needs more
time to construct, the entire running time becomes shorter.
According to the experimental results, we can conclude that
the three algorithms each have their own advantages. The
approximation scheme has a higher running time complexity
than the greedy algorithm (65% in the worst case and 32%
on average), and it outperforms the greedy algorithm by about
1%18% in terms of network life. The running time complexity of the constant-factor approximation algorithm is slightly
higher than that of the greedy algorithm, whereas its lifetime
performance is better than the greedy algorithm but worse than
the approximation scheme in most cases. Therefore, the choice
between the three algorithms depends on real situations. When
the WSN is large, and running time is the first priority, the
greedy algorithm is the best option. When network lifetime is
the first priority, the approximation scheme is the best option.
However, the approximation scheme is not scalable for the case
that the number of target points is extremely large. In such a
case, the constant-factor approximation algorithm can find a
near-optimal solution in a short time.

The authors would like to thank the anonymous referees


for their valuable comments and suggestions, which helped
improve the quality of this paper.
R EFERENCES
[1] B. N. Clark and C. J. Colbourn, Unit disk graphs, Discr. Math., vol. 86,
no. 13, pp. 165177, Dec. 1990.
[2] N. Garg and J. Knemann, Faster and simpler algorithms for multicommodity flow and other fractional packing problems, in Proc. IEEE FOCS,
1998, pp. 300309.
[3] S. Meguerdichian, F. Koushanfar, M. Potkonjak, and M. B. Srivastava,
Coverage problems in wireless ad-hoc sensor networks, in Proc. IEEE
INFOCOM, 2001, pp. 13801387.
[4] S. Slijepcevic and M. Potkonjak, Power efficient organization of wireless
sensor networks, in Proc. IEEE ICC, 2001, pp. 472476.
[5] V. Raghunathan, C. Schurgers, S. Park, M. Srivastava, and B. Shaw,
Energy-aware wireless microsensor networks, IEEE Signal Process.
Mag., vol. 19, no. 2, pp. 4050, Mar. 2002.
[6] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, Wireless
sensor networks: A survey, Comput. Netw., vol. 38, no. 4, pp. 393422,
Mar. 2002.
[7] M. Bhardwaj and A. P. Chandrakasan, Bounding the lifetime of sensor
networks via optimal role assignments, in Proc. IEEE INFOCOM, 2002,
pp. 15871596.
[8] K. Kar and S. Banerjee, Node placement for connected coverage in
sensor networks, in Proc. Model. Optim. Mobile, Ad Hoc Wireless Netw.,
2003, pp. 556563.
[9] K. Dasgupta, M. Kukreja, and K. Kalpakis, Topology-aware placement
and role assignment for energy-efficient information gathering in sensor
networks, in Proc. IEEE ISCC, 2003, pp. 341348.
[10] X. Wang et al., Integrated coverage and connectivity configuration in
wireless sensor networks, in Proc. ACM SenSys, 2003, pp. 2839.
[11] N. Sadagopan and B. Krishnamachari, Maximizing data extraction
in energy-limited sensor networks, in Proc. IEEE INFOCOM, 2004,
pp. 17171727.
[12] J. H. Chang and L. Tassiulas, Maximum lifetime routing in wireless
sensor networks, IEEE/ACM Trans. Netw., vol. 12, no. 4, pp. 609619,
Aug. 2004.
[13] P. Berman, G. Calinescu, C. Shah, and A. Zelikovsky, Power efficient
monitoring management in sensor networks, in Proc. IEEE WCNC,
2004, vol. 4, pp. 23292334.
[14] M. Cardei, M. T. Thai, Y. Li, and W. Wu, Energy-efficient target
coverage in wireless sensor networks, in Proc. IEEE INFOCOM, 2005,
pp. 19761984.

LU et al.: MAXIMUM LIFETIME SCHEDULING FOR TARGET COVERAGE AND DATA COLLECTION IN WSNs

[15] P. Berman, G. Calinescu, C. Shah, and A. Zelikovsky, Efficient energy management in sensor networks, in Ad Hoc and Sensor Networks.
Commack, NY, USA: Nova, 2005.
[16] C. F. Huang and Y. C. Tseng, A survey of solutions to the coverage
problems in wireless sensor networks, J. Internet Technol., vol. 6, no. 1,
pp. 18, 2005.
[17] M. Cardei and D. Du, Improving wireless sensor network lifetime
through power aware organization, Wireless Netw., vol. 11, no. 3,
pp. 333340, May 2005.
[18] M. Lu, J. Wu, M. Cardei, and M. Li, Energy-efficient connected coverage
of discrete targets in wireless sensor netorks, Netw. Mobile Comput.,
vol. 3619, pp. 4352, 2005.
[19] H. Zhang and J. Hou, On the upper bound of -lifetime for large sensor
networks, ACM Trans. Sens. Netw., vol. 1, no. 2, pp. 272300, Nov. 2005.
[20] X. Du and F. Lin, Maintaining differentiated coverage in heterogeneous
sensor networks, EURASIP J. Wireless Commun. Netw., vol. 2005, no. 4,
pp. 565572, 2005.
[21] C. Ambuhl, T. Erlebach, M. Mihalak, and M. Nunkesser, Constantfactor approximation for minimum-weight (connected) dominating sets
in unit disk graphs, in Approximation, Randomization, Combinatorial
Optimization. Algorithms and Techniques, vol. 4110. Berlin, Germany:
Springer-Verlag, 2006, ser. Lecture Notes in Computer Science, pp. 314.
[22] H. Gupta, Z. Zhou, S. R. Das, and Q. Gu, Connected sensor cover: Selforganization of sensor networks for efficient query execution, IEEE/ACM
Trans. Netw., vol. 14, no. 1, pp. 5567, Feb. 2006.
[23] M. Cardei and J. Wu, Energy-efficient coverage problems in wireless ad
hoc sensor networks, Comput. Commmun., vol. 29, no. 4, pp. 413420,
2006.
[24] H. Liu, P. J. Wan, and X. Jia, Maximal lifetime scheduling for sensor
surveillance systems with k sensors to 1 target, IEEE Trans. Parallel
Distrib. Syst., vol. 17, no. 12, pp. 15261536, Dec. 2006.
[25] S. Tang and W. Li, QoS supporting and optimal energy allocation for
a cluster based wireless sensor network, Comput. Commun., vol. 29,
no. 13/14, pp. 25692577, Feb. 2006.
[26] H. Liu et al., Maximizing lifetime of sensor surveillance systems,
IEEE/ACM Trans. Netw., vol. 15, no. 2, pp. 334345, Apr. 2007.
[27] I. F. Akyildiz, T. Melodia, and K. Chowdhury, A survey on wireless
mltimedia senosr networks, Comput. Netw., vol. 51, no. 4, pp. 921960,
Apr. 2007.
[28] D. Wang, B. Xie, and D. P. Agrawal, Coverage and lifetime optimization
of wireless sensor networks with Gaussian distribution, IEEE Trans.
Mobile Comput., vol. 7, no. 12, pp. 14441458, Dec. 2008.
[29] Q. Zhao and M. Gurusamy, Lifetime maximization for connected target
coverage in wireless sensor networks, IEEE/ACM Trans. Netw., vol. 16,
no. 6, pp. 13781391, Dec. 2008.
[30] T. Nieberg, J. Hurink, and W. Kern, Approximation schemes for wireless networks, ACM Trans. Algorithms, vol. 4, no. 4, pp. 49:149:17,
Aug. 2008.
[31] Y. Cai, W. Lou, M. Li, and X. Li, Energy efficient target-oriented
scheduling in directional sensor networks, IEEE Trans. Comput., vol. 58,
no. 9, pp. 12591274, Sep. 2009.
[32] Y. Huang, X. Gao, Z. Zhang, and W. Wu, A better constant-factor approximation for weighted dominating set in unit disk graph, J. Comb.
Optim., vol. 18, no. 2, pp. 179194, Aug. 2009.
[33] H. Zhang, P. Nixon, and S. Dobson, Partial coverage in homological
sensor networks, in Proc. IEEE WIMOB, 2009, pp. 4247.
[34] Y. Li, C. Vu, C. Ai, G. Chen, and Y. Zhao, Transforming complete
coverage algorithms to partial coverage algorithms for wireless sensor
networks, Parallel Distrib. Syst., vol. 22, no. 4, pp. 695703, Apr. 2011.
[35] F. Zou et al., New approximations for minimum-weighted dominating sets and minimum-weighted connected dominating sets on unit disk
graphs, Theoret. Comput. Sci., vol. 412, no. 3, pp. 198208, Jan. 2011.
[36] W. Li, Several characteristics of active/sleep model in wireless sensor
networks, IFIP NTMS, pp. 15, 2011.
[37] Y. Zhang and W. Li, Modeling and energy consumption evaluation of
a stochastic wireless sensor network, EURASIP J. Wireless Commun.
Netw., vol. 2012, p. 282, 2012.
[38] L. Ding et al., Constant-approximation for target coverage problem in wireless sensor networks, in Proc. IEEE INFOCOM, 2012,
pp. 15841592.

727

[39] Y. Gu, M. Pan, and W. Li, Prolonging the lifetime of large scale wireless
sensor networks via base station placement, in Proc. IEEE VTC, 2013,
pp. 15.
[40] K. Lin, X. Wang, L. Peng, and X. Zhu, Energy-efficient K-cover problem
in hybrid sensor networks, Comput. J., vol. 56, no. 8, pp. 957967, 2013.
[41] Z. Lu et al., Routing-efficient CDS construction in disk-containment
graphs, Optim. Lett., vol. 8, no. 2, pp. 425434, Feb. 2014.
[42] Y. Gu, M. Pan, and W. Li, Maximizing the lifetime of delay-sensitive
sensor networks via joint routing and sleep scheduling, in Proc. IEEE
ICNC, 2014, pp. 540544.

Zaixin Lu received the B.Sc. degree in computer science from Yanshan University, Qinhuangdao, China,
in 2006 and the Ph.D. degree in computer science
from the University of Texas at Dallas, Richardson,
TX, USA, in 2013.
He is currently a Postdoctoral Fellow with the
National Science Foundation Center for Research
on Complex Networks, Texas Southern University,
Houston, TX. His research interests include optimization, approximation algorithm, and networking.

Wei Wayne Li (M99SM06) received the B.Sc.


degree from Shaanxi Normal University, Xian,
China, in 1982; the M.Sc. degree from Hebei University of Technology, Tianjin, China, in 1987; and
the Ph.D. degree from the Chinese Academy of
Sciences, Beijing, China, in 1994.
He is a Professor and the current Interim Chair
of the Computer Science Department and the
Director/PI of the National Science Foundation
Center for Research on Complex Networks, Texas
Southern University, Houston, TX, USA. He was an
Associate Professor with tenure with the Department of Electrical Engineering
and Computer Science, University of Toledo, Toledo, OH, USA, and a tenure
track Assistant Professor with the Department of Electrical and Computer
Engineering, University of Louisiana at Lafayette, Lafayette, LA, USA. He
is the author of more than 100 peer-reviewed papers in professional journals
and the proceedings of international conferences. His research interests include
dynamic control and optimization of energy-efficient wireless sensor networks;
evaluation, complexity, power connectivity, and coverage for wireless sensor networks; adaptation, design, and implementation of dynamic models in
wireless multimedia systems; theoretic foundations and advanced analysis
in real-time, hybrid, and embedded systems; and performance evaluation of
biomolecular networks, queueing networks, and reliability systems.
Dr. Li is currently serving as an Editor for several professional journals
and is serving or has served as a Steering Committee Member/General
Cochair/Technical Program Committee Cochair/Session Chair/Technical Program Committee Member for a number of professional conferences.

Miao Pan (S07M12) received the B.Sc. degree


in electrical engineering from Dalian University of
Technology, Dalian, China, in 2004; the M.Sc. degree in electrical and computer engineering from
Beijing University of Posts and Telecommunications, Beijing, China, in 2007; and the Ph.D. degree
in electrical and computer engineering from the University of Florida, Gainesville, FL, USA, in 2012.
He is currently an Assistant Professor with the Department of Computer Science, Texas Southern University, Houston, TX, USA. His research interests
include cognitive radio networks, cyberphysical systems, and cybersecurity.

Вам также может понравиться