IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO.

5, JUNE 2014

2407

An Energy-Efficient Mobile-Sink Path Selection

Strategy for Wireless Sensor Networks
Hamidreza Salarian, Kwan-Wu Chin, and Fazel Naghdy

AbstractSeveral studies have demonstrated the benefits of using a mobile sink to reduce the energy consumption of nodes and to
prevent the formation of energy holes in wireless sensor networks
(WSNs). However, these benefits are dependent on the path taken
by the mobile sink, particularly in delay-sensitive applications, as
all sensed data must be collected within a given time constraint.
An approach proposed to address this challenge is to form a hybrid
moving pattern in which a mobile-sink node only visits rendezvous
points (RPs), as opposed to all nodes. Sensor nodes that are not
RPs forward their sensed data via multihopping to the nearest
RP. The fundamental problem then becomes computing a tour
that visits all these RPs within a given delay bound. Identifying
the optimal tour, however, is an NP-hard problem. To address
this problem, a heuristic called weighted rendezvous planning
(WRP) is proposed, whereby each sensor node is assigned a weight
corresponding to its hop distance from the tour and the number of
data packets that it forwards to the closest RP. WRP is validated
via extensive computer simulation, and our results demonstrate
that WRP enables a mobile sink to retrieve all sensed data within
a given deadline while conserving the energy expenditure of sensor
nodes. More specifically, WRP reduces energy consumption by
22% and increases network lifetime by 44%, as compared with
existing algorithms.
Index TermsData collection, mobile sink, scheduling, wireless
sensor networks (WSNs).

I. I NTRODUCTION

IRELESS sensor networks (WSNs) are composed of a

large number of sensor nodes deployed in a field. They
have wide-ranging applications, some of which include military
[1][3], environment monitoring [4], [5], agriculture [6], [7],
home automation [8], smart transportation [9], [10], and health
[11]. Each sensor node has the capability to collect and process
data, and to forward any sensed data back to one or more sink
nodes via their wireless transceiver in a multihop manner. In
addition, it is equipped with a battery, which may be difficult
or impractical to replace, given the number of sensor nodes and
deployed environment. These constraints have led to intensive
research efforts on designing energy-efficient protocols [12].
In multihop communications, nodes that are near a sink tend
to become congested as they are responsible for forwarding
data from nodes that are farther away. Thus, the closer a sensor
node is to a sink, the faster its battery runs out, whereas those
Manuscript received March 11, 2012; revised March 9, 2013, June 20, 2013,
September 15, 2013, and November 9, 2013; accepted November 14, 2013.
Date of publication November 21, 2013; date of current version June 12, 2014.
The review of this paper was coordinated by Dr. L. Li.
The authors are with the School of Electrical, Computer, and Telecommunications Engineering, University of Wollongong, Wollongong, NSW 2522,
Australia (e-mail: fazel@uow.edu.au).
Digital Object Identifier 10.1109/TVT.2013.2291811

Fig. 1. Example showing a mobile sink performing data collection in a WSN.

A source node determines and sends all data to a suitable site.

farther away may maintain more than 90% of their initial energy
[13]. This leads to nonuniform depletion of energy, which
results in network partition due to the formation of energy holes
[14], [15]. As a result, the sink becomes disconnected from
other nodes, thereby impairing the WSN. Hence, balancing the
energy consumption of sensor nodes to prevent energy holes is
a critical issue in WSNs.
To this end, previous works such as [15] and [16] employ
one or more mobile sinks. These mobile sinks survey and
collect sensed data directly from sensor nodes and thereby help
sensor nodes save energy that otherwise would be consumed
by multihop communications. Fig. 1 shows the feasible sites
of a mobile sink in an example WSN. Specifically, the squares
denote the feasible sites that the mobile sink will visit and
stop for data collection. The data forwarding path from sensor
nodes to the sink is dependent on the sinks current position.
This requires sensor nodes to dynamically plan one or more
data forwarding paths to each feasible site whenever the sink
node changes its position over time. As demonstrated by [15],
a mobile sink that moves at the periphery of a sensor field maximizes the lifetime of sensor nodes. Intuitively, by changing the
position of the sink over time, the forwarding tree will involve
a different set of sensor nodes and, hence, will help to balance
energy consumption [17][21].
The traveling path of a mobile sink depends on the realtime requirement of data produced by nodes. For example,
in hard real-time applications such as a fire-detection system
[22], environmental data need to be collected by a mobile sink

0018-9545 2013 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.

2408

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 5, JUNE 2014

Fig. 2. Hybrid movement pattern for a mobile sink node. Source nodes
generate and send sensed data to the nearest RP. Each RP buffers received data
and waits until the mobile sink arrives.

quickly. Moreover, a mobile-sink node may change its position

after a certain period of time and select another data collection/
feasible site. The feasible sites and corresponding sojourn time
are dependent on the residual energy of sensor nodes [23]
[27]. In general, limitations such as the maximum number of
feasible sites [28], maximum distance between feasible sites,
and minimum sojourn time [24] govern the movement of a
mobile sink.
In WSNs with a mobile sink, one fundamental problem is
to determine how the mobile sink goes about collecting sensed
data. One approach is to visit each sensor node to receive sensed
data directly [29]. This is essentially the well-known traveling
salesman problem (TSP) [30], where the goal is to find the
shortest tour that visits all sensor nodes. However, with an
increasing number of nodes, this problem becomes intractable
and impractical as the resulting tour length is likely to violate
the delay bound of applications. To this end, researchers have
proposed the use of rendezvous points (RPs) to bound the tour
length [31], [32]. This means a subset of sensor nodes are
designated as RPs, and non-RP nodes simply forward their data
to RPs. A tour is then computed for the set of RPs, as shown
in Fig. 2. As a result, the problem, which is called rendezvous
design, becomes selecting the most suitable RPs that minimize
energy consumption in multihop communications while meeting a given packet delivery bound. A secondary problem here is
to select the set of RPs that result in uniform energy expenditure
among sensor nodes to maximize network lifetime.
In this paper, we call this problem the delay-aware energyefficient path (DEETP). We show that the DEETP is an NPhard problem and propose a heuristic method, which is called
weighted rendezvous planning (WRP), to determine the tour of
a mobile-sink node. In WRP, the sensor nodes with more connections to other nodes and placed farther from the computed
tour in terms of hop count are given a higher priority. Thus, this
paper is summarized as follows.
We define the problem of finding a set of RPs to be visited
by a mobile sink. The objective is to minimize energy
consumption by reducing multihop transmissions from
sensor nodes to RPs. This also limits the number of RPs

such that the resulting tour does not exceed the required
deadline of data packets.
We propose WRP, which is a heuristic method that finds
a near-optimal traveling tour that minimizes the energy
consumption of sensor nodes. WRP assigns a weight to
sensor nodes based on the number of data packets that
they forward and hop distance from the tour, and selects
the sensor nodes with the highest weight.
We mathematically prove that selecting the sensor node
that forwards the highest number of data packets and have
the longest hop distance from the tour reduces the network
energy consumption, as compared with other nodes. Moreover, we show that, in contrast to cluster-based (CB) [32],
rendezvous design for variable tracks (RD-VT) [33], and
rendezvous planning utility-based greedy (RP-UG) [31]
algorithms, WRP is guaranteed to find a tour if the latter
exists.
We demonstrate via computer simulation the properties
and effectiveness of WRP against the CB [32], RD-VT
[33], and RP-UG algorithms [31]. Our results show that
WRP achieves 14% more energy savings and 22% better
distribution of energy consumption between sensor nodes
than the said algorithms.
The remainder of this paper is structured as follows.
Section II reviews methods that employ a hybrid model to collect
sensed data. Section III presents DEETP. In Section IV, we
illustrate WRP and present a detailed analysis of WRP and key
properties. Finally, in Section V, we compare the performance
of WRP with previous works before concluding in Section VI.
II. L ITERATURE R EVIEW
Existing methods on using a mobile sink in WSNs can be
grouped into two categories: 1) direct, where a mobile sink
visits each sensor node and collects data via a single hop;
and 2) rendezvous, where a mobile sink only visits nodes
designated as RPs. The main goal of protocols in category 1 is
to minimize data collection delays, whereas those in category 2
aim to find a subset of RPs that minimize energy consumption
while adhering to the delay bound provided by an application
[17]. In the following, we review the challenges faced by these
protocols.
A. Direct
Initial studies used a mobile sink that visits sensor nodes
randomly and transport collected data back to a fixed sink
node. An example is the use of animals as mobile-sink nodes
to assist in data collection from sensor nodes scattered on a
large farm [16]. To reduce the latency of visiting each sensor
node randomly, researchers have proposed TSP-based data collection methods. In essence, the problem is reduced to finding
the shortest traveling path that visits each sensor node [29].
For example, TSP with neighborhood [34] involves finding
the shortest traveling tour for a mobile-sink node that passes
through the communication range of all sensor nodes. Another
TSP-based algorithm [35] called label-covering considers a
WSN as a complete graph. For each edge, it calculates a cost

SALARIAN et al.: ENERGY-EFFICIENT MOBILE-SINK PATH SELECTION STRATEGY FOR WSNs

and associates a label set. The cost of an edge is the Euclidean

distance between nodes, whereas the label set contains sensor
nodes whose transmission range intersects with the given edge.
The label-covering algorithm selects the minimum number of
edges where their associated label set covers all sensor nodes.
B. Rendezvous
The problem with collecting data directly from sensor nodes
is that it becomes impractical when there are a large number
of sensor nodes. Visiting each sensor node increases the mobile
sinks traveling path length and results in sensor nodes experiencing buffer overflow due to data collection delays. To address
this problem, researchers have proposed a rendezvous-based
model, in which a mobile sink only visits a subset of sensor
nodes called RPs. The sensor nodes outside the mobile sink
path send their data via multihop communications to these RPs.
Studies [31][33], [36][38] deploying this approach can be
classified according to the mobile sinks trajectory, i.e., whether
it moves along a fixed path or its path is unconstrained by any
external factors.
1) Fixed: In the studies conducted in [36][38], the path
of the mobile sink is fixed, and sensor nodes are randomly
deployed near the sinks traveling path. Sensor nodes that are
inside a mobile sinks communication range play the role of
RPs and collect data from other sensor nodes. An example
application is a traffic management system where mobile sinks
are public buses that roam a city to collect data from sensor
nodes placed on buildings [36]. In these approaches, the length
of the traveling path is not dependent on the buffer size of sensor
nodes or application deadline. Hence, the buffer of RPs may
overflow or packets may expire before they are collected by the
sink.
Xing et al. [33] propose RD-FT, where the movement of
a mobile sink is governed by application deadline. They also
consider obstacles that restrict the movement of a mobile sink
along a predefined path. The objective is to find a set of RPs on
the fixed path such that the length of data forwarding paths from
sensor nodes to RPs is minimized and that the traveling time
between RPs is limited to the required packet delivery time.
2) Unconstrained: In [33], a WSN with a static sink node
and a mobile element (ME) is assumed to collect data from
RPs. Moreover, RPs perform data aggregation. An algorithm
called RD-VT is proposed with the objective of identifying a
traveling path that is shorter in duration than the packet delivery
time. The algorithm first constructs a Steiner minimum tree
(SMT) rooted at the sink node. RD-VT then starts from the
sinks position and traverses the SMT in preorder until the
shortest distance between visited nodes is equal to the required
packet delivery time. Since, in an SMT, a Steiner point may
be a physical position and does not correspond to the position
of a sensor node, RD-VT replaces these virtual RPs with the
closest sensor nodes. A major limitation of RD-VT is that
traversing the SMT in preorder leads to the selection of RPs
that in turn results in long data forwarding paths to sensor nodes
located in different parts of the SMT. As a result, RD-VT causes
nodes to have an unbalanced data forwarding load and energy
consumption.

2409

Xing et al. [31] propose rendezvous planning with a constrained ME path (RP-CP). Similar to RD-VT, the authors
consider a WSN with a fixed sink node and a ME. The RP-CP
first constructs a routing tree that is rooted at the sink node and
connects all sensor nodes. Then, each edge of the routing tree is
assigned a weight that corresponds to the number of nodes that
use that edge to forward their data to the sink node. The ME is
restricted to moving only on the edges of the tree. To construct
the MEs traveling path, RP-CP first sorts all edges according
to their weight. It then selects the edges with the highest weight
until the length of the selected edges becomes equal or less than
the required packet delivery time. The problem with RP-CP is
that the traveling path of the ME is restricted to routing tree
edges. This also means that the ME will visit the sensor nodes
on the selected edges twice.
In [31], the authors propose an improvement to RP-CP, which
is called the RP-UG algorithm. Initially, a geometric tree, which
is rooted at the fixed sink node, is constructed, and all edges on
the tree are split into multiple short intervals, which are denoted
as Lo . All points that join two edges with length Lo are considered RP candidates. RP-UG starts from the sinks position and,
in each step, adds the RP that minimizes the distance of sensor
nodes from selected RPs and also results in the shortest traveling tour between RPs. RP-UG uses a TSP solver to calculate
the tour length. To finalize the tour, RP-UG replaces virtual RPs
with the closest sensor nodes and marks them as RPs. RP-UG
does not balance the energy consumption rate of sensor nodes,
which has a significant impact on network lifetime. Specifically,
the network lifetime is determined by the sensor node with the
highest energy consumption rate, e.g., n, assuming all nodes
have the same initial energy level. In this regard, RP-UG does
not aim to minimize the energy consumption rate of node n. In
addition, when using RP-UG with a small Lo value, the number
of RPs increases significantly, and the complexity of RP-UG
grows exponentially. The algorithm uses a TSP solver N times
in each iteration, where N is the number of RPs. Hence, RP-UG
has a running time complexity of O(N 2 O(TSP)).
A CB algorithm is proposed in [32], in which a binary search
procedure is used to select RPs. Fig. 3 shows how CB works
in a network with ten nodes where the maximum allowed tour
length is 90 m. In the first iteration, using binary search on a
range from 0 to 10, CB selects five random cluster heads. In this
case, we have nodes 2, 3, 4, 5, and 7 [see Fig. 3(a)]. Other nodes
then establish a path to their closest cluster head in terms of hop
count. After the clusters are determined, CB starts from the sink
nodes position and selects one node from each cluster as an RP,
which is the closest node to the set of selected RPs. Fig. 3(b)
shows that node 8 from cluster {7, 8}, which is the closest node to
the sink, is selected as an RP and also node 6 from cluster {5, 6}
and so forth. Fig. 3(b) shows the final tour that have a length
of 127 m, which is larger than 90 m. Therefore, CB reduces the
number of clusters to two. According to Fig. 3(c), nodes 7 and 1
are selected randomly as new cluster heads. Fig. 3(d) shows that
the shortest possible tour between clusters is a tour including
nodes 2 and 9 that has a length of 55 m. CB therefore increases
the number of clusters to three because the tour length is less
than 90 m. Fig. 3(e) shows that nodes 4, 3, and 10 are selected
as new cluster heads, and the final tour length is 128 m, which is

2410

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 5, JUNE 2014

nodes. Replacing a virtual RP with a real sensor node in

RD-VT [33] and RP-UG [31] effectively approximates
an SMT with a shortest-path tree (SPT). We will show
through simulation that, when SMT is used and virtual
nodes are in the final tour, the energy consumption of
nodes can be reduced considerably, as compared with
using an SPT.
III. P ROBLEM F ORMULATION
Let us consider a WSN in which sensor nodes generate data
packets periodically. Each data packet must be delivered to the
sink node within a given deadline. There is a mobile sink that
roams around a WSN to collect data from a set of RPs. The
objective is to determine the set of RPs and associated tour that
visits these RPs within the maximum allowed packet delay.
A. Assumptions
Before describing DEETP, we first outline some assumptions.

Fig. 3. Tour finding steps in CB.

larger than 90 m. This causes CB to reduce the number of

clusters to two again. However, given that, CB has already
found a tour for two clusters, it stops and outputs {2, 9} as
the final tour. Note that this tour does not pass through dense
parts of the network consisting of nodes with a larger number of
neighbors such as nodes 10, 6, and 7. This problem causes long
data forwarding paths from sensor nodes to the RPs and nonuniform energy depletion, which reduces the lifetime of the WSN.
C. Contributions of This Paper
In this paper, we propose a hybrid unconstrained movement
pattern for a mobile sink with the aim of balancing the energy
consumption of sensor nodes. Our approach makes the following contributions, as compared with the work reported in the
literature.
We prefer nodes that have a high degree. This is critical as
sensor nodes in dense parts of a WSN generate the highest
number of packets. Hence, giving priority to sensor nodes
in these parts during tour computation will help to reduce
congestion points, and in turn, reduces energy consumption and improves WSN lifetime. In addition, it helps to
mitigate the energy-hole problem.
We consider hop distance between sensor nodes and RPs
as fewer hop counts reduce multihop transmissions. RPUG [31] minimizes network energy consumption by reducing the physical distance between sensor nodes and
RPs. However, due to the existence of obstacles, physical
distance is not a reliable indicator of energy consumption.
Apart from node density and hop count, when using an
SMT, we use virtual RPs in the final tour to increase
performance, and we do not replace them with real sensor

1) The communication time between the sink and sensor

nodes is negligible, as compared with the sink nodes
traveling time. Similarly, the delay due to multihop communications including transmission, propagation, and
queuing delays is negligible with respect to the traveling
time of the mobile sink in a given round.
2) Each RP node has sufficient storage to buffer all sensed
data.
3) The sojourn time of the mobile sink at each RP is sufficient to drain all stored data.
4) The mobile sink is aware of the location of each RP.
5) All nodes are connected, and there are no isolated sensor
nodes.
6) Sensor nodes have a fixed data transmission range.
7) Each sensor node produces one data packet with the
length of b bits in time interval D.
B. Notation
We model a WSN as G(V, E), where V is the set of homogeneous sensor nodes, and E is the set of edges between nodes in
V (see Table I for a summary of each notation). If sensor node
i sends b bits to node j, its energy consumption is [23]


ETX (i, j) = b 1 + 2 di,j

(1)

where di,j is the physical distance between sensor node i and j,

and 1 is the energy consumption factor indicating the power
per bit incurred by the transmitting circuit. The expression
2 di,j indicates the energy consumption of the amplifier per
bit, where 2 is the energy consumption factor of the amplifier
circuit. Here, is the path-loss exponent, which usually ranges
between 2 and 4, depending on the environment. Moreover, the
power consumption incurred by node i to receive b bits from
node j is
ERX (i, j) = b

(2)

SALARIAN et al.: ENERGY-EFFICIENT MOBILE-SINK PATH SELECTION STRATEGY FOR WSNs

TABLE I
S UMMARY OF N OTATION

2411

where C(i, Tmj ) is a function that returns the number of

children that node i has in the data forwarding tree rooted at
its corresponding RP mj .
C. Delay-Aware Energy-Efficient Traveling Path
The objective is to find a tour M = m0 , m1 , m2 , . . . ,
mn , m0 , where mi V , such that 1) the tour M is not longer
than lmax , and 2) the energy consumption for sending sensed
data from
 sensor nodes to the tour M , as defined by (ETX +
ERX ) iV H(i, M ), is minimized within time interval D.
DEETP is NP-hard by a reduction from TSP. Note that the
minimum energy consumption occurs when all sensor nodes
are designated as an RP. This is because they do not incur any
energy expenditure related to the forwarding of packets from
other nodes. In this case, the goal is then to determine whether
there is a tour that is not longer than lmax . Henceforth, in the
following, we propose a novel heuristic method to approximate
the optimal solution.
IV. W EIGHTED R ENDEZVOUS P LANNING
WRP preferentially designates sensor nodes with the highest
weight as a RP. The weight of a sensor node is calculated by
multiplying the number of packets that it forwards by its hop
distance to the closest RP on the tour. Thus, the weight of sensor
node i is calculated as
Wi = NFD(i) H(i, M ).

where is a factor that represents the energy consumption per

bit of the receiving circuit.
The mobile-sink node moves with a constant speed v. Hence,
the maximum length of the traveled path l is
lmax = D v.

(3)

A mobile-sink node starts its movement from a node m0

V and before time D returns to its starting point. Each sensor
node sends its generated data packets to the closest RP through
multihop transmissions. We define a function called H(i, M )
that returns the closest RP in terms of hop count to the sensor
node i, where M is the set of RPs. Specifically
H(i, M ) = {hi,mj | mk M, hi,mj hi,mk }

(4)

where hi,j is the hop distance between nodes i and j.

For each RP mi , our algorithm constructs a data forwarding
tree Tmi comprising of the closest sensor nodes to said RP. The
number of data packets NFD(i) that sensor node i forwards to
the closest RP mi in each time interval D is equal to its own
generated data packet plus the number of its children in the data
forwarding tree Tmi . Specifically
NFD(i) = C(i, Tmj ) + 1

(5)

(6)

Based on (6), sensor nodes that are one hop away from an RP
and have one data packet buffered get the minimum weight.
Hence, sensor nodes that are farther away from the selected
RPs or have more than one packet in their buffer have a higher
priority of being recruited as an RP.
From (1) and (2), the energy consumption is proportional
to the hop count between source and destination nodes, and
the number of forwarded data packets. Hence, visiting the
highest weighted node will reduce the number of multihop
transmissions and thereby minimizes the energy consumption.
In addition, as dense areas give rise to congestion points due to
the higher number of nodes, energy holes are more likely to occur in these areas. Hence, a mobile sink that preferentially visits
these areas will prevent energy holes from forming in a WSN.
Algorithm 1 shows how WRP works. It takes as input
G(V, E), and it outputs a set of RPs. WRP first adds the fixed
sink node as the first RP (see line 6). Then, in lines 917, it
adds the highest weighted sensor node. After that, WRP calls
TSP() (see line 21) to calculate the cost of the tour. If the tour
length is less than the required length lmax , the selected node
from lines 917 remains as an RP (see lines 2232). Otherwise,
it is removed from the tour (see lines 3337).
After a sensor node is added as an RP, WRP removes those
RPs from the tour that no longer receives any data packets
from sensor nodes(see lines 2730). This is because adding a
sensor node to the tour may reduce the number of data packets
directed to these RPs. Consequently, this step affords WRP
more opportunities to add other nodes into the tour. Note that
the variable removed is used to guarantee that an RP will be

2412

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 5, JUNE 2014

deleted from the tour only once. If a removed RP is added to

the tour for the second time, because its corresponding variable
removed is true, it will not be removed from the tour again.
In this way, all sensor nodes will be added to the tour when the
required tour length for a mobile sink is bigger than the time to
visit all sensor nodes.

Fig. 4.

Example of WRP operating in a WSN with ten nodes.

As Fig. 4(b) shows, the tour length of M is smaller than

the required tour length (56 < 90), meaning node 10 stays
in the final tour (lines 2232). In the second iteration, WRP
recalculates the weight of sensor nodes because node 10 is now
part of the tour. In this iteration, WRP selects node 6 as the
next RP, which has the highest weight. As Fig. 4(c) shows, the
tour length of M = [Sink, 10, 6] is larger than the required tour
length (119 > 90). Consequently, WRP removes node 6 from
the tour M = [Sink, 10] (lines 3337). In the third iteration, the
weight of sensor nodes will not change because node 6 is not
selected as an RP but it stays marked and will not be selected.
WRP selects node 8 because it has the highest weight and is
not marked [see Fig. 4(d)]. The TSP function returns 76 m for
M = [Sink, 10, 8], which is less than 90 m. Therefore, node 8
is added to the tour. The process continues, yielding a final tour
of M = [Sink, 8, 7, 10, 9] with a tour length of 81 m, which is
less than the required tour length [see Fig. 4(e)].
As shown in Fig. 4, the final tour computed by WRP always
includes sensor nodes that have more data packets to forward
than other nodes as RPs. This ensures uniform energy consumption and mitigates the energy-hole problem. This is the key
advantage of WRP over CB, RD-VT, and RP-UG. In Section V,
we will show that this feature of WRP allows it to save 30%
more energy than CB.
A. Analysis
Fig. 4 shows an example of how WRP finds a traveling tour
for a mobile sink. The maximum tour length is lmax = 90 m.
WRP starts from the sink node and adds it to the tour, i.e., M =
[Sink]. Then, an SPT rooted at the sink node is constructed
[see Fig. 4(a)]. In the first iteration, WRP adds node 10 to the
tour because it has the highest weight, yielding M = [Sink, 10].

The time complexity of our algorithm is dependent on how

many times WRP calls the TSP solver to calculate a tour that
visits all RPs. The worst case is when all sensor nodes are
marked but not selected as an RP, which means WRP will iterate
for |V | times to check the possibility of adding nodes into a tour.
After a node is selected as an RP, WRP again unmarks other

SALARIAN et al.: ENERGY-EFFICIENT MOBILE-SINK PATH SELECTION STRATEGY FOR WSNs

sensor nodes and restarts the search process (see lines 1237).
This means our algorithm uses the TSP solver for a maximum
of n2 times, where n = |V |. Therefore, the time complexity
of WRP is O(n2 O(TSP)). Hence, if we use Christofidess
heuristic [39], which has time complexity of O(n3 ), the resulting time complexity is O(n5 ). In our experiments, we used a
local-search-based heuristic TSP solver [40].
We like to point out that WRP always finds a tour when there
is at least one possible tour in the network. This is because
WRP checks the possibility of adding all sensor nodes to the
tour. This is significant when compared with CB and RD-VT
because the latter two algorithms fail in the following scenario.
In CB, if the only possible tour consists of only the sink and a
neighbor in the same cluster, CB will not be able to find this tour
because two sensor nodes from the same cluster cannot be in the
final tour. As for RD-VT, it will return no tour if the distance
of the first sensor node in the SMT, as it starts its depth-first
traversal, exceeds lmax .
We now prove that visiting the most weighted nodes by a mobile sink results in the least energy consumption, as compared
with visiting any other nodes.
Theorem 1: Visiting sensor node P with weight wp reduces
energy consumption more than visiting sensor node Q with
weight wq , where wp > wq .
Proof: Recall that sensor node P forwards NFD(P ) data
packets to its closest RP. Therefore, the energy consumption of
sensor nodes on the path from node P to the closest RP is
Ep = (ETX + ERX ) (NFD(P ) H(P, M )) .

(7)

However, if sensor node P becomes an RP, the energy consumption incurred by data packets at P is zero. Similarly, for
sensor node Q that forwards NFD(Q) data packets to its closest
RP, we have
Eq = (ETX + ERX ) (NFD(Q) H(Q, M )) .

(8)

From (6), the weight of sensor node P is wp = NFD(P )

H(P, M ), and the weight of sensor node Q is wq = NFD(Q)
H(Q, M ). We know that wp > wq ; therefore, it can be concluded that Ep > Eq , which means selecting sensor node P as
an RP, which has a higher weight than Q, leads to less network
energy consumption.

We now show the difference between WRP and the optimal
solution. We first prove the following lemma.
Lemma 1: Let WRPop be a version of WRP that uses an
optimal TSP solver. If there is an optimal tour named C with
length Lc lmax comprising of all sensor nodes as RP, then
WRPop is guaranteed to find tour C.
Proof: Assume there are n sensor nodes in a WSN and
tour C = m0 , m1 , m2 , . . . , mn , m0 , where m0 is the sink node.
Then
Lc =

n1


dmi ,mi+1 + dmn ,m0 .

(9)

WRPop , after picking the sink, will select node mi=1 to

include in the tour as it has the highest weight before running
TSP() (see line 21 of Algorithm 1). The returned cost will be

2413

less than Lc as the tour connecting the set of nodes cannot

be longer than the tour containing all nodes by the triangle
inequality. Hence, WRPop will add mi=2 and so forth until

i = n. WRPop then terminates as Tn = |V |.
Note that, in Lemma 1, the requirement on there being an
|E|
optimal TSP solver can be relaxed if we assume that i=0 i
lmax . In other words, the sum of all distances between nodes is
less than the required tour length.
Note that, intuitively, it would seem that the maximum difference in energy consumption occurs when the final tour returned
by WRPop is not composed of any sensor nodes while the
optimal tour visits all sensor nodes. However, as per Lemma 1,
this will not happen.
Theorem 2: Assume a sensor node P that has the longest hop
distance from the sink, and the average hop distance between
sensor nodes and the sink is k; then, the maximum difference
between the network energy consumption of WRPop and the
optimal is within ((2 K (|V | 1) + 1)/(|V | + 2)).
Proof: The network energy consumption when the mobile
sink visits only sensor node P is
ENetwork(p) = (ETX + ETX ((|V | 1) k))
+ (ERX ((|V | 1) k)) .

(10)

On the other hand, the minimum amount of energy consumed

by visiting all sensor nodes except node P is
ENetwork(|V |1) = ETX (|V | + 1) + ERX .

(11)

This means sensor node P has to send all its data packets
to the closest RP, whereas other sensor nodes send their data
packets directly to the mobile sink. From (10) and (11), the ratio
of energy consumption in WRP in comparison to the optimal
model is
Ratio
=

ETX (1 + (|V | 1) k) + ERX ((|V | 1) k)

.
ETX (|V | + 1) + ERX
(12)

If we consider ETX
= ERX , (12) is equal to
Ratio
=

2 K (|V | 1) + 1
.
|V | + 2

(13)


V. E VALUATION
We compare WRP against three existing methods that have
the same objective as ours, namely CB, RD-VT, and RP-UG,
using a custom simulator written in C++.1 We consider a
connected WSN where nodes are placed uniformly on a sensor field of size 200 200 m2 . We note that interconnecting
disconnected sensor nodes using a mobile node is a wellknown and separate problem. Having said that, we remark that
WRP can be also made to interconnect disconnected nodes if
the required delivery time for data packets is greater than the
1 Our

simulator is available upon request.

2414

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 5, JUNE 2014

TABLE II
S IMULATION PARAMETERS

shortest traveling tour to visit all sensor nodes. The reason that
we have assumed uniform sensor-node distribution is because
energy holes are more likely to form when nodes are distributed
uniformly [41]. Experimental results in [13] demonstrated that,
if sensor nodes are distributed uniformly, up to 90% of residual
energy is unused when the first sensor node dies. In addition,
we adopt uniform distribution to ensure fair comparison with
RD-VT, CB, and RP-UG.
Similar to [42], we have set the radio parameters as per the
CC1000 radio, which is used by Mica2 Motes [43]. Each sensor
node generates one data packet every T time, which is then
forwarded to an RP via an SPT. We assume nodes are aware of
the mobile sinks movement and, hence, arrival time. We record
and compare the network energy consumption every T time.
Moreover, there are a maximum of 200 sensor nodes, which
is reasonable for most applications (see [44, Tab. I]).
To measure network lifetime, we assume that all sensor
nodes have a fully charged battery with 100 J of energy. Other
parameters are summarized in Table II. We set the mobile sinks
speed to 1 m/s. We further assume that it visits each RP. Given
a transmission range of 20 m, which is feasible for Mica2 [43]
or TelosB [45] nodes, the mobile sink will be in a sensor nodes
transmission range for 20 s. Assuming a data transmission rate
of 40 Kb/s, each sensor node will be able to send 3413 data
packets with a length of 30 b to the mobile sink in 20 s. This
means that the mobile sink has sufficient time to drain the buffer
of all sensor nodes even when there are 200 sensor nodes. To
reduce the run time of RP-UG, we set L0 to 20 m, which
corresponds to the transmission range of sensor nodes.
We use standard deviation (SD) to measure the imbalance
between the energy consumption of sensor nodes, i.e., a wide
variation means some parts of a WSN is likely to exhaust its
energy sooner. The metric SD is calculated as follows:

2
iV (EN[i] )
(14)
SD =
|V |
where EN[i] is the energy consumption of node i, V is the set
of sensor nodes, and is the average energy consumption of
sensor nodes.
In our evaluation, we consider two scenarios involving an
SPT and an SMT for the RD-VT model: RP-UG and WRP.
In WRP, we find the Steiner points and treat them as real
nodes. This means that Steiner points have a weight and are
not replaced with real sensor nodes in the final tour.

Fig. 5.

Network energy consumption between brute force and WRP.

Two sets of experiments are carried out. Initially, the number

of nodes is limited to 20, and WRP is compared against a bruteforce approach that yields the optimal tour with 2.5 min as
the required tour length. In the second experiment, the number
of nodes is increased to 200, and WRP is compared against
RD-VT and CB with 5 min as the required tour length. In all
experiments, we designate the node with the highest ID as the
sink node, and the results are an average of ten simulation runs
over different topologies.
A. Performance Under SPT
Fig. 5 shows the energy consumption of sensor nodes for
WRP versus brute force. Both algorithms yield higher energy
consumption when the number of sensor nodes increases as
the length of data forwarding paths from sensor nodes to RPs
increases. The energy consumption of WRP is very close to
the brute-force approach. In particular, brute force only outperforms WRP by 5%.
Fig. 6 shows the difference, as per SD, between sensor nodes
energy consumption. A small SD value means uniform energy
consumption and longer network lifetime. The performance of
WRP is only 16% less than the optimal or brute-force approach.
This is because, in WRP, sensor nodes that forward more
data packets and cause more multihop transmissions than other
sensor nodes are likely to be designated as an RP.
Fig. 7 shows the energy consumption for WRP, CB, RDVT, and RP-UG with a large number of sensor nodes. RDVT leads to the highest energy consumption because of its
preorder traversal of the SPT and long data forwarding paths
from sensor nodes to the RPs. WRP recorded 47% reduction in
energy consumption, as compared with RD-VT. CB has better
performance than RD-VT because, in its finalization process, if
the required delivery time is not violated, it replaces the selected
RP in each cluster with a node closer to the cluster head to reduce the number of multihop transmissions. CBs performance
is 28% better than RD-VT in terms of energy consumption.
Recall that in Section II-B2, CB does not consider node density
or hop counts when selecting RPs. As a result, WRP achieves

SALARIAN et al.: ENERGY-EFFICIENT MOBILE-SINK PATH SELECTION STRATEGY FOR WSNs

Fig. 6.

Standard deviation of sensor nodes energy consumption.

Fig. 7.

Network energy consumption for WRP, CB, RD-VT, and RP-UG.

a 10% reduction in energy consumption, as compared with CB.

RP-UG and WRP have nearly the same performance; WRP
reduces energy consumption by an additional 6%, as compared
with RP-UG.
The ability of the four algorithms to uniformly distribute
energy consumption is shown in Figs. 8 and 9. WRP distributes
energy more uniformly than the other approaches, specifically,
12% more than RP-UG, 28% more than CB, and 53% better
than RD-VT. Similarly, RP-UG does not aim to balance the
energy consumption rate of sensor nodes. RP-UG adds sensor
nodes that are close to the sink as RPs, which may not necessarily have the highest energy consumption rate. Moreover,
although RP-UG considers nodes that are on many routing
paths, WRP preferentially selects nodes with high energy consumption rate and hop count from the sink. As shown in Fig. 9,
WRP improves network lifetime by 13%, as compared with RPUG. With regard to CB, as shown in Fig. 3, the random clusterhead selection process causes nonuniform energy consumption.
Moreover, two sensor nodes from the same cluster cannot be in
the final tour. Hence, when there are a large number of sensor

2415

Fig. 8. Standard deviation of sensor nodes energy consumption for WRP,

CB, RD-VT, and RP-UG models.

Fig. 9. Network lifetime for WRP, CB, RD-VT, and RP-UG.

nodes, energy holes are likely to occur around the RP for a

given cluster. In contrast, WRP avoids this scenario by having
the mobile sink visits dense parts of a WSN, which helps to
reduce the number of multihop transmissions. In RD-VT, long
forwarding paths are observed from sensor nodes to RPs, which
results in nonuniform energy consumption and 25% reduction
in network lifetime, as compared with CB.
In this experiment, we simulated the algorithms in a network
with 110 sensor nodes and data packets with a packet delivery
time ranging from 100 to 300 s. Figs. 10 and 11 show network
energy consumption and network lifetime for WRP, CB, RDVT, and RP-UG. Consistent with the result shown in Fig. 7,
WRP yields the best performance among all algorithms. The
energy consumption for RD-VT reduces by 21% when the
required packet delivery time is changed from 100 to 300 s,
whereas WRP, RP-UG, and CB observed a reduction of 41%,
33%, and 37%, respectively. WRP observed a superior performance, as compared with other algorithms, even with small
packet delivery times. This is because WRP always checks the
possibility of adding the node with the highest weight first.

2416

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 5, JUNE 2014

Fig. 10. Network energy consumption for WRP, CB, RD-VT, and RP-UG
under different required delivery times for data packets.

Fig. 11. Network lifetime for WRP, CB, RD-VT, and RP-UG under different
required delivery times for data packets.

Finally, we recorded the simulation time of each algorithm

(see Fig. 12). For RP-UG, we set L0 = 20 m, i.e., the transmission range of sensor nodes. The value of 20 m is the maximum
possible for L0 because otherwise, edges bigger than L0 are
split into edges with length L0 . Virtual nodes are then added
as necessary to connect these new edges. Consequently, this
process, depending on the value of L0 , increases run time
significantly. In fact, even with L0 set to 20 m, the running
time of RP-UG is six times bigger than WRP, 36 times more
than CB, and 72 times longer than RD-VT. This is because,
in each iteration, RP-UG calculates the utility of each sensor
node by calling a TSP solver. RD-VT has the lowest running time because it only calls the TSP solver once in each
iteration.
B. Performance Under SMT
We have also considered using the SMT for RP-UG, RDVT, and WRP; this tree is constructed using the function in

Fig. 12.

Simulation time for WRP, CB, RD-VT, and RP-UG.

[46]. Specifically, the Steiner tree function uses the principal

of an equilateral triangle, a circle, and a line to construct a
Steiner point for a set containing three points on the minimum
spanning tree (MST). When an SMT is formed, there are two
types of Steiner points. The first type corresponds to real sensor
nodes, which are called real Steiner points, and the other type
is simply a physical position with no sensor nodes, which are
called virtual Steiner points.
For each Steiner point in WRP, a neighbor set is assigned
after the formation of the SMT. In RP-UG, RD-VT, and WRP,
virtual Steiner points that are not in the final tour are deleted
from the SMT. On the other hand, we handle virtual Steiner
points that are part of a tour in the following manner. In RD-VT
and RP-UG, virtual RPs are replaced with the closest physical
sensor nodes. In WRP, when a sensor node notices that the
succeeding hop destination for its data packets is a virtual RP,
the sensor node stores its data until the mobile sink arrives at the
virtual RPs position. Upon arrival, the sensor node forwards its
data to the mobile sink.
Fig. 13 shows a comparison between WRP, RP-UG, RD-VT,
and CB when the SMT is used for data forwarding. The results
show that the SMT has a better performance than the SPT. This
is because a Steiner point is the shortest interconnection point
between neighboring nodes. As a result, the total length of the
SMT in comparison to the SPT is shorter. This causes RD-VTSMT and RP-UG-SMT to visit more RPs, and causes nodes
to have 29% and 8% less energy consumption than RD-VTSPT and RP-UG-SPT, respectively. Moreover, RD-VT-SMT
outperforms CB. For the same reason, WRP-SMT conserves
13% more energy than WRP-SPT and 22% more than CB.
Moreover, WRP has 24% better performance than RD-VT and
14% better performance than RP-UG when they all use the
SMT. This is because WRP does not replace virtual RPs with
the closest physical sensor nodes. Instead, as stated previously,
a mobile sink is programmed to visit virtual RP positions and
to collect data from nearby sensor nodes.
Figs. 14 and 15 show the difference between the energy
consumption of sensor nodes and the network lifetime for SMT
experiments. In particular, the SD recorded is less than those

SALARIAN et al.: ENERGY-EFFICIENT MOBILE-SINK PATH SELECTION STRATEGY FOR WSNs

Fig. 13. Network energy consumption for WRP, CB, RD-VT, and RP-UG in
SMT scenarios.

2417

when using the SPT. This is because virtual Steiner points,

which are in the final tour, do not receive data from other sensor
nodes. Instead, these points are visited by the mobile sink. In
comparison, when we use the SPT, RPs have a higher energy
consumption, as compared with other sensor nodes. However,
for the SMT case, when virtual points are in the final tour, we
observe fewer RPs; thus, actual nodes that act as RPs reduces
and thereby reduces consumed energy. Moreover, the energy
consumption between sensor nodes is distributed uniformly.
For these reasons, the SD for WRP-SMT is 16% less than
WRP-SPT and 39% less than CB (see Fig. 14). The SD for RDVT-SMT is 28% less than RD-VT-SPT, and for RP-UG-SMT,
it is 5% less than RP-UG-SPT. This is because visiting more
RPs leads to shorter data forwarding paths and, thereby, better
network lifetime. The difference between the SD of WRP-SMT
and RD-VT-SMT is 44%, and between WRP-SMT and RP-UGSMT, it is 22%, which is less than the results recorded in SPT
experiments. This thus confirms that the better performance
gained by WRP in the SMT scenario is due to the use of
virtual RPs.

VI. C ONCLUSION

Fig. 14. Standard deviation of sensor nodes energy consumption in SMT

scenarios.

In this paper, we have presented WRP, which is a novel algorithm for controlling the movement of a mobile sink in a WSN.
WRP selects the set of RPs such that the energy expenditure of
sensor nodes is minimized and uniform to prevent the formation
of energy holes while ensuring sensed data are collected on
time. In addition, we have also extended WRP to use an SPT
and an SMT. Apart from that, we have also considered visiting
virtual nodes to take advantage of wireless coverage. Our results, which are obtained via computer simulation, indicate that
WRP-SMT reduces the energy consumption of tested WSNs by
22% in comparison to CB. We also benchmarked WRP against
existing schemes in terms of the difference between sensornode energy consumption. Our simulation results show that
WRP uniformly distributes energy consumption by 39% and
44% better than CB and RD-VT, respectively.
As a future work, we plan to enhance our approach to include
data with different delay requirements. This means a mobile
sink is required to visit some sensor nodes or parts of a WSN
more frequently than others while ensuring that energy usage is
minimized, and all data are collected within a given deadline.
Moreover, we plan to extend WRP to the multiple mobile
sinks/rovers case. This case, however, is nontrivial as it involves
subproblems such as interference and coordination between
rovers. Having said that, we note that WRP remains applicable
if a large WSN is partitioned into smaller areas where each
area is assigned a mobile sink. WRP can be thus run in each
area. We defer the evaluation of such an approach to a future
paper.

ACKNOWLEDGMENT

Fig. 15. Network lifetime for WRP, CB, RD-VT, and RP-UG in SMT
scenarios.

The authors would like to thank the reviewers for providing constructive comments that have improved this paper
significantly.

2418

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 5, JUNE 2014

R EFERENCES
[1] 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.
[2] S. Diamond and M. Ceruti, Application of wireless sensor network to
military information integration, in Proc. 5th IEEE Int. Conf. Ind. Inform., Vienna, Austria, Jun. 2007, vol. 1, pp. 317322.
[3] I. Bekmezci and F. Alagz, Energy efficient, delay sensitive, fault tolerant
wireless sensor network for military monitoring, Int. J. Distrib. Sens.
Netw., vol. 5, no. 6, pp. 729747, 2009.
[4] A. Mainwaring, D. Culler, J. Polastre, R. Szewczyk, and J. Anderson,
Wireless sensor networks for habitat monitoring, in Proc. 1st ACM Int.
Workshop Wireless Sens. Netw. Appl., New York, NY, USA, Sep. 2002,
pp. 8897.
[5] J. Zhang, W. Li, Z. Yin, S. Liu, and X. Guo, Forest fire detection system
based on wireless sensor network, in Proc. 4th IEEE Conf. Ind. Electron.
Appl., Xian, China, May 2009, pp. 520523.
[6] L. Ruiz-Garcia, L. Lunadei, P. Barreiro, and I. Robla, A review of wireless sensor technologies and applications in agriculture and food industry:
State of the art and current trends, Sensors, vol. 9, no. 6, pp. 47284750,
Jun. 2009.
[7] N. Wang, N. Zhang, and M. Wang, Wireless sensors in agriculture and
food industryrecent development and future perspective, Comput. Electron. Agriculture, vol. 50, no. 1, pp. 114, Jan. 2006.
[8] A. Wheeler, Commercial applications of wireless sensor networks using
zigbee, IEEE Commun. Mag., vol. 45, no. 4, pp. 7077, Apr. 2007.
[9] W. Chen, L. Chen, Z. Chen, and S. Tu, Wits: A wireless sensor network
for intelligent transportation system, in Proc. 1st Int. Multi-Symp. Comput. Comput. Sci., Hangzhou, China, Jun. 2006, vol. 2, pp. 635641.
[10] B. Hull, V. Bychkovsky, Y. Zhang, K. Chen, M. Goraczko, A. Miu,
E. Shih, H. Balakrishnan, and S. Madden, Cartel: A distributed mobile
sensor computing system, in Proc. 4th Int. Conf. Embedded Netw. Sens.
Syst., New York, NY, USA, Oct. 2006, pp. 125138.
[11] L. Yu, N. Wang, and X. Meng, Real-time forest fire detection with
wireless sensor networks, in Proc. Int. Conf. Wireless Commun., Netw.
Mobile Comput., Wuhan, China, Sep. 2005, vol. 2, pp. 12141217.
[12] I. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, A survey
on sensor networks, IEEE Commun. Mag., vol. 40, no. 8, pp. 102114,
Aug. 2002.
[13] J. Lian, K. Naik, and G. B. Agnew, Data capacity improvement of
wireless sensor networks using non-uniform sensor distribution, Int. J.
Distrib. Sens. Netw., vol. 2, no. 2, pp. 121145, Apr.Jun. 2006.
[14] J. Li and P. Mohapatra, An analytical model for the energy hole problem
in many-to-one sensor networks, in Proc. 62nd IEEE Veh. Technol. Conf.,
Dallas, TX, USA, Sep. 2005, vol. 4, pp. 27212725.
[15] J. Luo and J.-P. Hubaux, Joint mobility and routing for lifetime elongation in wireless sensor networks, in Proc. IEEE INFOCOM, Miami, FL,
USA, Mar. 2005, vol. 3, pp. 17351746.
[16] R. C. Shah, S. Roy, S. Jain, and W. Brunette, Data mules: Modeling and
analysis of a three-tier architecture for sparse sensor networks, Ad Hoc
Netw., vol. 1, no. 2/3, pp. 215233, Sep. 2003.
[17] X. Li, A. Nayak, and I. Stojmenovic, Sink Mobility in Wireless Sensor
Networks. Hoboken, NJ, USA: Wiley, 2010, pp. 153184.
[18] E. Ekici, Y. Gu, and D. Bozdag, Mobility-based communication in wireless sensor networks, IEEE Commun. Mag., vol. 44, no. 7, pp. 5662,
Jul. 2006.
[19] S. Jain, R. Shah, W. Brunette, G. Borriello, and S. Roy, Exploiting
mobility for energy efficient data collection in wireless sensor networks,
Mobile Netw. Appl., vol. 11, no. 3, pp. 327339, Jun. 2006.
[20] I. Chatzigiannakis, A. Kinalis, and S. Nikoletseas, Sink mobility protocols for data collection in wireless sensor networks, in Proc. 4th ACM
Int. Workshop Mobility Manage. Wireless Access, Terromolinos, Spain,
Oct. 2006, pp. 5259.
[21] I. Chatzigiannakis, A. Kinalis, and S. Nikoletseas, Efficient data propagation strategies in wireless sensor networks using a single mobile sink,
Comput. Commun., vol. 31, no. 5, pp. 896914, Mar. 2008.
[22] J. Capella, A. Bonastre, J. Serrano, and R. Ors, A new robust, energyefficient and scalable wireless sensor networks architecture applied to a
wireless fire detection system, in Proc. Int. Conf. Wireless Netw. Inform.
Syst., Shanghai, China, Dec. 2009, pp. 395398.
[23] Z. Wang, S. Basagni, E. Melachrinoudis, and C. Petrioli, Exploiting sink
mobility for maximizing sensor networks lifetime, in Proc. 38th Annu.
Hawaii Int. Conf. Syst. Sci., Hawaii, USA, Jan. 2005, p. 287a.
[24] S. Basagni, A. Carosi, E. Melachrinoudis, C. Petrioli, and Z. M. Wang,
Controlled sink mobility for prolonging wireless sensor networks lifetime, Wireless Netw., vol. 14, no. 6, pp. 831858, Dec. 2008.

[25] Y. Bi, L. Sun, J. Ma, N. Li, I. A. Khan, and C. Chen, Hums: An

autonomous moving strategy for mobile sinks in data-gathering sensor
networks, EURASIP J. Wireless Commun. Netw., vol. 2007, pp. 64574-1
64574-15, 2007.
[26] Y. Yun and Y. Xia, Maximizing the lifetime of wireless sensor networks
with mobile sink in delay-tolerant applications, IEEE Trans. Mobile
Comput., vol. 9, no. 9, pp. 13081318, Sep. 2010.
[27] M. Gatzianas and L. Georgiadis, A distributed algorithm for maximum
lifetime routing in sensor networks with mobile sink, IEEE Trans. Wireless Commun., vol. 7, no. 3, pp. 984994, Mar. 2008.
[28] J. Luo, J. Panchard, M. Pirkowski, M. Grossglauser, and J.-P. Hubaux,
Mobiroute: Routing towards a mobile sink for improving lifetime in sensor networks, in Distributed Computing in Sensor Systems, P. Gibbons,
T. Abdelzaher, J. Aspnes, and R. Rao, Eds. Berlin, Germany: SpringerVerlag, 2006, pp. 480497.
[29] M. Ma and Y. Yang, Data gathering in wireless sensor networks with
mobile collectors, in Proc. IEEE Int. Symp. Parallel Distrib. Process.,
Miami, FL, USA, Apr. 2008, pp. 19.
[30] S. Arora, Polynomial time approximation schemes for euclidean traveling salesman and other geometric problems, J. ACM, vol. 45, no. 5,
pp. 753782, Sep. 1998.
[31] G. Xing, T. Wang, Z. Xie, and W. Jia, Rendezvous planning in wireless
sensor networks with mobile elements, IEEE Trans. Mobile Comput.,
vol. 7, no. 12, pp. 14301443, Dec. 2008.
[32] K. Almiani, A. Viglas, and L. Libman, Energy-efficient data gathering with tour length-constrained mobile elements in wireless sensor networks, in Proc. 35th IEEE Conf. LCN, Denver, CO, USA,
Oct. 2010, pp. 582589.
[33] G. Xing, T. Wang, W. Jia, and M. Li, Rendezvous design algorithms for
wireless sensor networks with a mobile base station, in Proc. 9th ACM
Int. Symp. Mobile ad hoc Netw. Comput., Hong Kong, China, May 2008,
pp. 231240.
[34] S. Nesamony, M. Vairamuthu, and M. Orlowska, On optimal route of a
calibrating mobile sink in a wireless sensor network, in Proc. 4th Int.
Conf. Netw. Sens. Syst., Braunschweig, Germany, Jun. 2007, pp. 6164.
[35] R. Sugihara and R. Gupta, Improving the data delivery latency in
sensor networks with controlled mobility, in Distributed Computing in Sensor Systems, S. Nikoletseas, B. Chlebus, D. Johnson, and
B. Krishnamachari, Eds. Berlin, Germany: Springer-Verlag, 2008,
pp. 386399.
[36] A. Chakrabarti, A. Sabharwal, and B. Aazhang, Using predictable
observer mobility for power efficient design of sensor networks, in
Proc. 2nd Int. Conf. Inform. Process. Sens. Netw., Palo Alto, CA, USA,
Apr. 2003, pp. 129145.
[37] A. Chakrabarti, A. Sabharwal, and B. Aazhang, Communication
power optimization in a sensor network with a path-constrained mobile observer, ACM Trans. Sens. Netw., vol. 2, no. 3, pp. 297324,
Aug. 2006.
[38] S. Gao, H. Zhang, and S. Das, Efficient data collection in wireless
sensor networks with path-constrained mobile sinks, IEEE Trans. Mobile
Comput., vol. 10, no. 4, pp. 592608, Apr. 2011.
[39] D. Johnson, G. Gutin, L. McGeoch, A. Yeo, W. Zhang, and A. Zverovitch,
Experimental analysis of heuristics for the atsp, in The Traveling Salesman Problem and Its Variations, G. Gutin, A. Punnen, D.-Z. Du, and
P. M. Pardalos, Eds. New York, NY, USA: Springer-Verlag, 2004,
pp. 445487.
[40] I. Cholissodin, To solving traveling salesman problem (TSP) for 10 till
100 city with localsearch., Dec. 2007. [Online]. Available: http://www.
planet-source-code.com/vb/scripts/ShowZip.asp?lngWId=3&
lngCodeId=11881&strZipAccessCode=tp%2FT118817122
[41] X. Wu, G. Chen, and S. Das, Avoiding energy holes in wireless sensor
networks with nonuniform node distribution, IEEE Trans. Parallel Distrib. Syst., vol. 19, no. 5, pp. 710720, May 2008.
[42] W. Liang, J. Luo, and X. Xu, Prolonging network lifetime via a controlled mobile sink in wireless sensor networks, in Proc. IEEE Global
Telecommun. Conf., Miami, FL, USA, Dec. 2010, pp. 16.
[43] Mica2 868, 916 mhz, CROSSBOW, San Jose, CA, USA, Tech.
Rep., 2007. [Online]. Available: http://www.investigacion.frc.utn.edu.ar/
sensores/Equipamiento/Wireless/MICA2_Datasheet.pdf
[44] H. Salarian, K.-W. Chin, and F. Naghdy, Coordination in wireless sensoractuator networks: A survey, J. Parallel Distrib. Comput., vol. 72, no. 7,
pp. 856867, Jul. 2012.
[45] Telosb. [Online]. Available: http://www.memsic.com/products/wirelesssensor-networks/wireless-modules.html
[46] B. Bell, Steiner minimal tree problem. Scientific Programming Track,
CSS Senior Seminar Project., Jan. 16, 1999. [Online]. Available: http://
cse.Taylor.edu/bbell/steiner/

SALARIAN et al.: ENERGY-EFFICIENT MOBILE-SINK PATH SELECTION STRATEGY FOR WSNs

Hamidreza Salarian received the B.S. degree

in computer engineering (hardware major) from
Ferdowsi University of Mashhad, Mashhad, Iran, in
2005 and the M.S. degree in computer architecture
from Isfahan University of Technology, Isfahan, Iran,
in 2008. He is currently working toward the Ph.D.
degree with the School of Electrical, Computer,
and Telecommunications Engineering, University of
Wollongong, Wollongong, Australia.
His research interests include energy consumption
in wireless sensor networks and wireless sensor actuator networks, particularly the effect of mobility in increasing system lifetime.

Kwan-Wu Chin received the B.S. degree (with first

class honors) and the Ph.D. degree (with distinction
and the vice chancellors commendation) from the
Curtin University of Technology, Perth, Australia.
He then joined Motorola Research Laboratory as
a Senior Research Engineer, where he developed
zero-configuration home networking protocols and
designed new medium-access-control protocols for
wireless sensor networks. Since 2004, he has been
with the University of Wollongong, Wollongong,
Australia, first as a Senior Lecturer and then promoted as an Associate Professor in 2011. He is a holder of four U.S. patents and
the author of more than 65 articles in numerous conferences and journals. His
current research interests include medium-access-control protocols for wireless
networks, routing protocols for delay-tolerant networks, radio-frequency identification anticollision protocols, and discrete optimization problems related to
computer communications.

2419

Fazel Naghdy was born in Tehran, Iran. He received

the B.Sc. degree from Tehran University in 1976;
the M.Sc. and Ph.D. degrees from the University of
Bradford, Bradford, U.K., in 1972 and 1982, respectively; and the M.Edu. degree from the University of
Wollongong, Wollongong, Australia.
He has a demonstrated track record and leadership in research, teaching, and management. He
is currently a Professor of robotics and intelligent
systems and the Director of the Center for Intelligent
Mechatronics Research with the University of Wollongong. He is the author of more than 250 papers in international journals
and conferences and several book chapters. His research has had its focus
on machine intelligence and control, particularly in embedded mechatronics
and robotics systems. His current research interests include haptically rendered
virtual manipulation of clinical and mechanical systems and intelligent control
and learning in nonlinear and nonstructured systems.
Dr. Naghdy has served on a large number of international scientific committees of various international conferences. He is a Contributing Editor
to the IEEE T RANSACTIONS ON M ECHATRONICS E NGINEERING and the
International Journal of Intelligent Automation and Soft Computing.