IET Communications

Research Article

WDAT-OMS: A two-level scheme for efficient ISSN 1751-8628

Received on 21st April 2019
Revised 16th February 2020
data gathering in mobile-sink wireless sensor Accepted on 30th March 2020
E-First on 3rd June 2020
networks using compressive sensing theory doi: 10.1049/iet-com.2019.0433
www.ietdl.org

Shima Pakdaman Tirani1, Avid Avokh1 , Sahebeh Azar1

1Department of Electrical Engineering, Najafabad Branch, Islamic Azad University, Najafabad, Iran
E-mail: [email protected]

Abstract: This study addresses the problems of energy and delay in wireless sensor networks equipped with mobile sinks. The
authors jointly consider the compressive sensing (CS) theory, cluster-based routing, and sink mobility to propose a data
collection method named ‘weighted data aggregation trees with optimal mobile sink(s) (WDAT-OMS)’. The proposed scheme
relies on a two-level architecture in which sensors are clustered at the first level. WDAT-OMS uses the CS theory along with
load-balanced data aggregation trees to route packets from sensors to the corresponding cluster heads (CHs). In this regard,
they present an efficient metric named ‘energy-and distance-aware CH selection’ to fairly distribute the energy consumption
among different sensors. At the second level, one or more sinks traverse the network to collect the aggregated data of CHs. As
an advantage, WDAT-OMS not only balances the energy consumption among different sensors but also increases the network
scalability. Numerical results demonstrate that the proposed algorithm reduces energy consumption in comparison with
‘centralised clustering algorithm’, ‘energy-aware CS-based data aggregation’, and ‘energy-balanced high-level data aggregation
tree ‘by 66%, 62%, and 63% for an average number of clusters, respectively. It also decreases the sink delay in comparison with
the ‘single-hop data-gathering problem’ by 10%.

 Nomenclature concentration in the air [1, 2]. Energy consumption is one of the
main challenges in WSNs, which affects the network's lifetime.
B number of polling points Owing to the redundancy in WSNs, data of different sensors should
b a constant parameter be compressed and aggregated [1, 3]. In this regard, the
C number of clusters compressive sensing (CS) theory is an efficient solution used to
CH set of cluster heads (CHs) reconstruct the sparse signals at a rate less than the Nyquist rate
CM( j) members of jth cluster [4]. It combines the sampling and compression processes in one
d(i, j) Euclidean distance between sensors i and j step to minimise the number of samples of signals. This
E0 initial energy of sensors significantly reduces the amount of data to send, and subsequently,
ERx‐elec electrical energy required to run the circuitry in decreases energy consumption.
receiver On the other hand, traffic distribution in WSNs is naturally
ETx‐elec electrical energy required to run the circuitry in unbalanced. The sensors near the sink act as relay nodes and
transmitter receive/transmit more data than the other sensors. As a result, these
Er(i) residual energy of sensor i sensors quickly lose their energy. This is identified as ‘energy hole
ERx(L) energy consumption for receiving an L-bit packet problem’, which reduces the performance of the network.
ETx(L, d) energy consumption for transmitting an L-bit packet Therefore, an efficient routing protocol should be considered to
over distance d improve the load-balancing in the network. In this regard, cluster-
L packet length based routing protocols divide the network into several clusters to
M number of compressive sensing measurements locally balance the traffics among different sensors [5]. Over the
N number of sensors in the network last decade, several methods have been proposed for cluster head
NB(CH j) set of neighbour polling points of the jth CH (CH) selection, which generally use the residual energy and
NP(i) nearest neighbour polling point located at the distance factors [6–10]. However, unlike the traditional methods,
transmission range of sensor i we propose a new CH selection strategy, which not only uses both
P set of polling points distance and residual energy metrics but also considers the sink
Pr(i, r) punishment factor of sensor i at round r mobility and a punishment factor to more efficiently control the
R transmission range of sensors energy consumption in the network.
RSP reduced set of polling points In traditional WSNs, a multi-hop routing is used to deliver the
S number of mobile sinks data of sensors to the sink node. This increases the number of
V speed of mobile sinks transmissions and consequently, intensifies the energy consumption
ρ compressive ratio in the network. On the other hand, two-hop routing protocols, e.g.
ϵamp energy consumption of the transmission amplifier low-energy adaptive clustering hierarchy (LEACH), have also been
proven to be inefficient in comparison with the modern schemes. In
1 Introduction recent years, a number of research studies have been conducted to
more efficiently collect data in WSNs, including the use of
Wireless sensor networks (WSNs) consist of small sensors with multiple constant sinks and mobile sinks [11, 12]. In particular, the
sensing, limited computing, and wireless communication abilities. use of a mobile sink for data collection introduces a trade-off
These nodes sense the environment and send the collected data to between energy consumption and delay. Although the mobile sink
the sink. WSNs have many applications in environmental is not suitable for some real-time applications in terms of delay, it
monitoring including the information collection of soil humidity, is an efficient technique to reduce the energy consumption in non-
monitoring the coastline changes, and checking the gas real time applications, which generally collect data of an

IET Commun., 2020, Vol. 14 Iss. 11, pp. 1826-1837 1826

© The Institution of Engineering and Technology 2020
environment for a long time. A mobile sink can efficiently traverse The rest of the paper is organised as follows. In Section 2, the
the network and stop at the pre-determined polling points to previous related works are surveyed. In Section 3, we first describe
receive the buffered sensed data through a single-hop transmission. the network architecture and energy consumption model. Then, the
In this way, using a shorter path for the sink tour reduces delay in CS theory is briefly introduced. In Section 4, the proposed
the network. algorithm is described in detail. Extensive simulations on the
In the last decade, several research studies have been presented performance of the proposed algorithm are presented in Section 5.
to increase the lifetime in WSNs [10, 13–19]. However, there are Finally, some concluding remarks are provided in Section 6.
still some research gaps that need to be addressed. The ability to
make appropriate use of the CS theory, mobile sink, and load- 2 Related work
balanced cluster-based routing can improve the performance of
WSNs. One drawback of the previous works is that the literature In recent years, several studies have been presented on the CS and
either address these issues separately or study two cases together. data aggregation in WSNs [17–19]. Ebrahimi and Assi [17]
However, keeping them all together can significantly increase the propose an efficient data gathering scheme using the CS theory
network lifetime. Thus, an efficient algorithm is required to jointly named ‘minimum spanning tree projection’. This algorithm creates
address the above-mentioned problems. Taking the described a number of minimum spanning trees, each of which rooted at a
challenges into account, this study deals with two basic problems randomly selected projection node. Each projection node collects
in WSNs, i.e. the energy consumption of sensors and delay of sink data of interest nodes and sends the packets to the sink through the
tour. By jointly employing the CS theory, cluster-based routing, shortest path tree (SPT). Zhang et al. [18] propose the ‘ring
and mobile sink(s), we propose an efficient algorithm, called topology-based CS’ data collection scheme to reduce the total
‘weighted data aggregation trees with optimal mobile sink(s) number of transmissions. The proposed algorithm uses dual
(WDAT-OMS)’, which uses a two-level hybrid architecture. At the compensation-based CS measurement to reconstruct the original
first level, the network is divided into several clusters. An efficient signals with high accuracy. A novel Bayesian CS data gathering
metric, namely ‘energy-and distance-aware CH selection (EDCS)’, strategy is introduced in [19]. It is an efficient algorithm for
is proposed to fairly distribute the energy consumption among minimising the tour path of the mobile sink. Sensors measure the
different sensors. In each cluster, the data of sensors is transmitted parameters of the field and forward data to the nearest gathering
to the corresponding CH using the CS theory and weighted data nodes by the SPT routing algorithm. A mobile sink collects data
aggregation tree. At the second level, one or more mobile sinks from the gathering nodes at each round. As an advantage, the
optimally traverse the network to collect the aggregated data of proposed scheme increases the network lifetime.
CHs. There have been some studies on jointly considering the CS
The main related works are [13–16]. Pakdaman Tirani and theory and cluster-based routing in WSNs [13, 14, 20, 21]. Two
Avokh [13] consider a simple mobility pattern for the sink node data collection methods named ‘energy-aware CS-based data
while Xie and Jia [14] adopt a stationary sink. In [15], considering aggregation (ECDA)’ and ‘energy-balanced high-level data
a constant sink and clustering the network, an energy-aware aggregation tree (EHDT)’ are proposed in [13]. These schemes not
emergency transmission scheme is proposed. In the current study, only balance the energy consumption among different sensors but
unlike [13–15], we present a hybrid architecture to use the also increase the network lifetime. Both algorithms are compatible
advantages of both data aggregation trees and mobile sinks. The with constant and mobile sinks. However, Pakdaman Tirani and
presented scheme not only makes a load-balanced tree for each Avokh [13] consider a simple mobility pattern for the mobile sink.
cluster but also provides an optimal tour for the mobile sink(s) by EHDT applies a weighted routing tree for transmitting data
enjoying an integer linear programming (ILP) model. Hop-count- projections from CHs to the sink. This distributes the energy
based routing results in hotspots in the network and reduces its consumption among different sensors fairly. In [14], two
lifetime [13, 14]. In this regard, we propose an efficient metric to centralised and distributed clustering algorithms are presented to
fairly construct the weighted data aggregation routing trees. reduce the number of transmissions in the network. Each sensor
Furthermore, unlike [16] in which the proposed objective function first transmits its data to the corresponding CH using a cluster-
only considers the tour delay of the mobile sink, this work presents based routing method and without the CS theory. Then, CHs use
a multi-objective function to jointly decrease both energy the CS theory to transmit data to the sink. In [13, 14], the hop-
consumption of sensors and tour delay of the mobile sink(s). In count metric is used for routing the packets between sensors and
[16], the sink interacts directly with all sensors. However, in our CHs. This metric leads to create hotspots in the network. In [20],
model thanks to the use of the proposed two-level architecture, it the authors first propose a data aggregation method named
has only an interaction with CHs, which have a large number of ‘weighted compressive data aggregation’. It benefits from the
buffered packets. This allows the sink to travel much less distance advantage of the sparse random measurement matrix to reduce
and its delay will be reduced. Joint use of the CS theory, mobile energy consumption. Then, they present another data aggregation
sink, and load-balanced cluster-based routing is another superiority method named ‘cluster-based weighted compressive data
of this work compared to [16]. aggregation’ to make a significant improvement in the performance
In summary, the main contributions of this work are as follows: of the network. Sun et al. [21] propose the ‘sparsest random
sampling scheme for cluster-based compressive data gathering
• We present the WDAT-OMS algorithm in the form of a two-level (SRS-CCDG)’ in WSNs. In this work, the network is divided into
hybrid architecture that jointly considers the CS theory, mobile several clusters. At each round of data gathering, a random subset
sink, and load-balanced cluster-based routing. of sensors sense the environment and transmit the measurements to
• An efficient metric for CH selection is introduced to distribute the the corresponding CHs. Then, each CH transmits the aggregated
energy consumption among different sensors fairly. data to the sink. According to the simulation results, SRS–CCDG
• We propose a load-balanced strategy to develop data aggregation significantly reduces energy consumption and improves the
routing trees. system's robustness to the node failures. Singanamalla et al. [15],
considering a constant sink node, divide the network into a series
• An optimal tour for sink mobility is determined by considering a
of circular layers. They propose an energy-aware emergency
multi-objective function, which minimises both energy
transmission scheme, which decreases the energy consumption and
consumption of sensors and delay of the mobile sink.
increases transmission accuracy.
• We extend the problem for WSNs equipped with multiple mobile Over the past few years, several methods have been presented
sinks to increase the network scalability. to select CHs in WSNs [6–10, 22–26]. They generally use the
• We present extensive simulations to evaluate the proposed residual energy and distance factors. For example, in [8], a
algorithm compared with the other state-of-the-art algorithms in distributed CH selection algorithm is presented, which uses the
terms of energy consumption, energy consumption variance, delay distances between sensors and sink to improve the energy
of the sink, and time complexity. consumption in the network. The authors of [9, 22] extend the
method of CH selection in LEACH by considering the residual

IET Commun., 2020, Vol. 14 Iss. 11, pp. 1826-1837 1827

© The Institution of Engineering and Technology 2020
energy of sensors. In [10], a CH selection algorithm called ‘hybrid counterclockwise direction along the border of the sensing field to
grouped grey wolf search optimisation’ is proposed to improve the collect data from the closest CHs via single-hop transmissions. In
lifetime of WSNs. This algorithm focuses on the factors of [37], an energy-efficient cluster-based routing mechanism is
distance, delay, energy, and security to select CHs in such a way proposed for WSNs with obstacles. The authors use a scheduling
that improves the performance of the network. In [23], a clustering mechanism based on the spanning graphs. They present a heuristic
mechanism is proposed to partition the sensors into unequal-size tour-planning algorithm for the mobile sink to find the obstacle-
clusters so that the closer CHs to the sink can preserve more energy avoiding shortest route. However, none of these works use the CS
for inter-cluster data forwarding. The clustering scheme presented theory. On the other hand, the authors of [33, 34, 36] consider a
in [24] selects CHs with more residual energy through local radio simple mobility pattern for the mobile sink. Furthermore, the
communication while achieving well CH distribution. The load- authors of [34, 37] only use a single mobile sink to collect the
aware rotation of CH (LAR-CH) method was presented in [25]. It sensed data from CHs. Certainly, using multiple mobile sinks not
sets a dynamic threshold for CH-rotation to reduce the premature only increases the scalability of the network but also reduces the
death of CHs. LAR-CH uses the current energy of CHs to network delay.
determine the dynamic threshold for CH-rotation. Behera et al. In addition to the works discussed above, there are also some
[26] propose an efficient CH selection scheme that rotates the CH references that study energy consumption from the other
position among the sensors with higher energy. The proposed perspectives. For example, Patan and Babu [38] propose an energy-
scheme considers the initial energy, residual energy, and optimum efficient traffic-aware resource scheduling and re-streaming stream
number of CHs to select the next group of CHs for the network. structure to replace a default scheduling strategy of a storm named
However, none of these methods jointly consider the problems of ‘re-storm’. This strategy is investigated in three parts: (i)
sink mobility and load-balancing. presenting a mathematical relation between energy consumption,
As mentioned before, another challenge of WSNs is the low response time, and high traffic streams. (ii) Providing various
unbalanced energy consumption in different sensors. Joint approaches for reducing energy consumption without affecting the
clustering and load balancing is the main purpose in [27]. Gherbi et response time and high performance in overall stream computing
al. [27] propose a hierarchical routing algorithm named for big data. (iii) Deploying energy-efficient traffic-aware
‘hierarchical energy-balancing multipath’. This algorithm scheduling on the storm platform. In [39], the authors propose a
decreases the energy consumption, balances the energy dissipation self-adaptive technique named ‘energy-efficient internet of things’,
among different sensors, and increases the network lifetime. which aims to minimise the energy harvesting in the network. In
Siavoshi et al. [28] present an energy-efficient clustering protocol [40], a new scheme is presented, namely ‘resistive to selective drop
that improves both load-balancing and network lifetime. It uses attack’, to provide effective security against the selective drop
several virtual circles to form a multi-layer circular structure across attack in wireless ad-hoc networks. As an advantage, the proposed
the network. In [29], a balanced multi-path routing algorithm is scheme is compatible with different routing schemes, e.g. ‘ad hoc
proposed by considering the residual energy and hop-count on-demand distance vector’ and ‘dynamic source routing’. It
metrics. The precision of this algorithm grows with increasing the reduces the overhead, jitter, and energy consumption in the
path length and network density. Network scalability is another network.
advantage of the proposed algorithm in [29].
In the field of using the mobile sink, some works aim to 3 Network model and assumptions
heuristically improve the network performance [30, 31], while
other works present the optimisation solutions [16, 32]. In [30], a In this section, we first present the preliminaries of our work
mobile sink moves on a pre-determined linear route; while in [31], including the network model, assumptions, and energy
multiple controlled data mobile ubiquitous local area network consumption model. Then, the CS theory is briefly described.
extension is used in parallel routes. In this scheme, sensors deliver
the measured data to the sinks through the single-hop or multi-hop 3.1 Network model
transmissions.
In [16], an efficient data gathering mechanism is proposed for We consider a WSN consisting of N sensors and S mobile sinks,
WSNs equipped with single or multiple mobile collectors. A where the sensors are randomly distributed in a a × a square area.
mobile data collector works like a mobile base station. It starts the Each node is equipped with an omni-directional antenna. The
data gathering process from a static sink and periodically traverses network is modelled as a bidirectional graph, i.e. G(V, E), where V
the network to poll sensors one by one. Finally, it returns and consists of sensors and sinks and E represents the wireless links
uploads data to the static sink. Ma et al. [16] mainly focus on between them. Two nodes are directly connected and form a
minimising the length of data gathering tour by an ILP model and communication link if they are within the transmission range of
refer to this as a single-hop data-gathering problem (SHDGP). For each other denoted by R. It is assumed that there is no interference
the large-scale networks, Ma et al. [16] use several mobile in the network by utilising the orthogonal codes in data
collectors instead of a single one. In this method, the mobile transmission. In addition, no packet will be lost during the
collectors transmit their data for a mobile collector located in the transmissions.
vicinity of the sink. In [32], taking a mobile base station into In this study, according to Fig. 1, a two-level hybrid architecture
account, the problem of load-balanced data collection is is used. At the first level, the sensors are organised into C clusters.
investigated. The main idea is using the existing multi-hop routing Each sensor transmits its data to the corresponding CH via a
schemes to achieve further improvements in terms of the network weighted data aggregation tree. At the second level, the network
lifetime. Luo and Hubaux [32] present some constraints on the sink has B ‘polling points’ deployed as a grid arrangement. Polling
sites so that each site covers a subset of sensors. The mobile sink points are defined as the locations where the mobile sink stops and
stops at each site to collect the corresponding data. receives the buffered data of CHs. In fact, each mobile sink starts a
There are some works which jointly consider the clustering and tour to periodically traverse a part of the network and collects the
mobile sink to improve the network performance [33–37]. In [33], aggregated data of CHs through a single-hop transmission. Since
the authors propose an efficient clustering method along with data there must be at least one polling point in the neighbourhood of
aggregation using multiple mobile sinks. In [34], a mobile sink- each sensor, the grid length, i.e. the distance between neighbour
based routing protocol is proposed to prolong the network lifetime polling points in the same row or column, is assumed to be equal to
in clustered WSNs. It uses a single mobile sink to collect the the transmission range of sensors.
sensed data from CHs. During the data gathering process, the This work relies on a real traffic model in which sensors send
mobile sink maintains the residual energy of CHs to move to the the measured data to the sink, round by round. The procedure of
high-energy CHs. The proposed algorithm in [35] aims to find the measuring the environmental parameters is periodically repeated at
location of the mobile sink and the optimum number of CHs for specific times. The time between two consecutive sensing of the
minimising the energy consumption and the overhead of control environment is defined as a round. Given the intended application
packets. In [36], the mobile sinks move in a clockwise or for WSNs, i.e. non-real-time monitoring of the environment, it is

1828 IET Commun., 2020, Vol. 14 Iss. 11, pp. 1826-1837

© The Institution of Engineering and Technology 2020
 sparse signal can be recovered by only M ≪ N random
measurements, i.e. Y = [y1, y2, …, yM ]T ∈ RM . This vector is
generated as Y = Φ X, where Φ = [φi j] ∈ RM × N is called as the
measurement matrix. Thus, the ith element of the measurement
N
vector Y can be stated as yi = ∑ j = 1 φi j x j. Measurement matrix Φ
satisfies the restricted isometry property (RIP) of order 2k if there
exists a constant δk ∈ (0, 1) such that [42]

(1 − δk) ∥ X ∥ 22 ≤ ∥ Φ X ∥ 22 ≤ (1 + δk) ∥ X ∥ 22, (3)

N 1/ p
where ∥ X ∥ p = ∑i = 1 xi p represents the lp-norm of vector
X . The existence of RIP for the random matrices, e.g. Gaussian
matrix with uniformly and independently distributed elements and
Bernoulli matrix with ±1 elements, has been proved [20]. A k-
sparse signal can be reconstructed from only M measurements by
utilising the following l1 optimisation expression [43]:
Fig. 1  Typical example for the applied two-level architecture consisted of ^
the first-level data aggregation trees and sink tour Θ = arg min ∥ Θ ∥ 1 subject to Y = Φ Ψ Θ, (4)

N ^
assumed that the duration of a round is such that the mobile sink where ∥ Θ ∥ 1 = ∑i = 1 ∣ θi ∣ and X = Ψ Θ.
has the opportunity to traverse the network and collects the It should be noted that the practical performance of the CS
buffered data of CHs. However, we aim to minimise this time. theory depends on the sparseness of signals and reconstruction
In general, the energy consumed by sensors is influenced by algorithms. Certainly, increasing the number of CS measurements
three factors, including data sensing, data processing, and wireless leads to a better quality of data recovery.
communication. However, the first two factors are negligible
compared to the energy consumption for wireless communication
[3, 20]. We use the first-order radio model, in which the transmitter
4 Weighted data aggregation trees with optimal
dissipates the energy to run the radio electronics and the power mobile sink(s) (WDAT-OMS)
amplifier, and the receiver consumes a different amount of energy In this section, we first present the problem description and then,
to run the radio electronics. According to this model, the energy introduce the WDAT-OMS algorithm in detail.
consumption for transmitting an L-bit packet over distance d is
modelled as follows [5, 20]: 4.1 Problem description
ETx(L, d) = ETx‐elec(L) + ETx‐amp(L, d) Given the explained model in Section 3, the goal is to reduce the
(1) energy consumption of sensors and tour delay of the mobile sink in
= L × ETx‐elec + L × ϵamp × d 2,
WSNs. For this purpose, we use a two-level hybrid architecture in
which the first level organises sensors into C clusters. Each sensor
where ETx‐elec is the electrical energy required to run the circuitry in
transmits its data to the corresponding CH via a weighted data
the transmitter and ϵamp shows the energy consumption of the aggregation tree. At the second level, one (or multiple) sink(s)
transmission amplifier. The electrical energy factor in the traverses the network to collect the aggregated data from CHs. The
transmitter, i.e. ETx‐elec, depends on different factors such as digital main issues to be discussed include
coding, modulation, and spreading of the signal, whereas the term
of amplifier energy, i.e. ϵampd 2, uses the square of the distance • Employing an efficient method for clustering and CH selection to
between transmitter and receiver to consider the acceptable bit fairly distribute the energy consumption among different sensors.
error rate. • Developing a load-balanced data aggregation routing tree in each
In addition, the energy for receiving an L-bit packet can be cluster that connects the sensors to the corresponding CH.
computed as follows [5, 20]: • Using the hybrid CS theory along with the first-level routing
process.
ERx(L) = L × ERx‐elec, (2) • Determining an appropriate path for the sink tour to jointly
reduce both energy consumption of sensors and sink delay in the
where ERx‐elec is the electrical energy required to run the circuitry in network.
the receiver. Following the other references [5, 20, 41], we ignore
the energy consumption of baseband signal processing blocks such 4.2 Algorithm description
as source coding and pulse shaping. In fact, these factors are quite
small compared to the energy consumption of the radio frequency In this subsection, the WDAT-OMS algorithm is proposed to
circuitry. jointly address the problems of CS theory, load-balanced cluster-
Nomenclature summarises the notations that will be used in this based routing, and sink mobility in WSNs. Unlike sensors that
paper. have a simple structure, the mobile sink can be a mobile robot or a
vehicle equipped with a rechargeable battery, powerful central
processing unit (CPU), large memory, and global positioning
3.2 CS theory
system (GPS) device. Thus, in this work, the decision-making
CS theory presents a novel technique to efficiently recover a sparse centre of the network is the sink node. In fact, the basic
signal from its CS measurements at a rate less than the Nyquist rate calculations, including the CH selection, optimal tour etc., are
[17]. Consider a data vector collected by the sensors, i.e. made in the sink node.
X = [x1, x2, …, xN ]T ∈ RN . [ ⋅ ]T indicates the transpose operator. Similar to the other protocols, some preliminary signalling is
This vector has a representation as X = Ψ Θ in the proper basis of required at the initial deployment of sensors and before monitoring
the environment. For example, in the beginning, the sink node gets
Ψ = [ψ i j] ∈ RN × N , in which Θ = [θ1, θ2, …, θN ]T is the coefficient the location of sensors by one of the conventional localisation
vector of X in the basis of Ψ. X is a k-sparse vector, if it has at most services [44]. Given the grid layout of polling points, it is also well
k non-zero elements (k ≪ N) in Θ. According to the CS theory, a k- aware of their position. Owing to the static nature of sensors and

IET Commun., 2020, Vol. 14 Iss. 11, pp. 1826-1837 1829

© The Institution of Engineering and Technology 2020
polling points, the distances between them remain constant during node distances, we adopt the distance between sensors and polling
the network activity and this information does not need to be points to consider the sink mobility in CH selection. This makes
updated frequently. sensors near the polling points more likely to be selected as CH. As
WDAT-OMS partitions the sensors into C clusters based on a result, according to the first-order radio model, the energy
their location. For this purpose, it first randomly selects C nodes as consumption at the second level of the network is reduced.
the initial set of selected nodes (SNs). Certainly, this initial set However, by increasing the number of clusters, the number of
cannot be the optimal choice and should be modified by the sensors in each cluster is reduced. Therefore, the sensors near the
following steps: polling points are repeatedly selected as CHs at different rounds. In
fact, as the number of clusters increases, the distance factor
(1) The Euclidean distance from each sensor to the current SNs is diminishes the role of the energy factor. This reduces load
measured. Each sensor joins to the nearest SN. In this way, each balancing ability. Thus, we have a trade-off between energy
SN forms a cluster that should be modified. consumption and load-balancing. To solve this problem and control
(2) For each cluster, a new SN is determined in such a way that the this trade-off, we add the punishment factor, which reduces the
sum of the Euclidean distance from all of its sensors to the new SN chance of selecting the current CH at the next rounds. Pr(i, r)
is minimised. In fact, this step aims to select a closer SN to the represents the punishment factor of sensor i at round r. It is initially
centre of the corresponding cluster. set to 1 for all sensors. If a sensor is selected as CH at the current
(3) Two above steps are repeated until there are no more changes in round, its Pr is divided by an empirical constant parameter (b)
SNs. Repeating the previous steps allows the algorithm to
m(i, r)
gradually converge to the best clusters. 1
Pr i, r + 1 = Pr(i, r), (6)
b
Given the fixed location of sensors, the clustering algorithm is
implemented only after their deployment in the network. However, where
as we will see later, the role of CH is rotated between sensors,
round by round. In general, various factors influence the optimal 1, if sensor i is selected as CH at round r,
number of clusters, including the number of sensors, density of m(i, r) = (7)
0, otherwise .
sensors, and number of CS measurements. As the number of
clusters increases, the size of each cluster will decrease. This Since the sensors are randomly distributed in the network, a
reduces the number of transmissions and energy consumption in sensor may be very close to the polling point. In this case,
each cluster. On the other hand, the number of transmissions and d(i, NP(i)) tends to zero and the fraction of Er(i)/d(i, NP(i))
consequently, the energy consumption at the second level of the becomes infinite. As a result, the effect of Pr and Er(i) in CH
network will be increased. However, thanks to the use of the hybrid selection is ignored. Hence, ‘1’ was added to the denominator.
CS method at the first level and the optimal tour for the mobile
sink at the second level of the network, the overall energy
consumption is reduced. From the delay point of view, increasing 4.2.2 Phase 2 – constructing the first-level data aggregation
the number of clusters leads to an increase in the number of CHs. trees: Unlike the previous works [13, 14], which use the hop-count
This increases the delay of the sink tour. Thus, there is a trade-off based routing trees, WDAT-OMS constructs the weighted data
between the energy consumption of sensors and the delay of the aggregation trees for routing packets from sensors to the
sink tour. corresponding CHs. This balances the energy consumption among
After the clustering, WDAT-OMS is run in three phases at each different sensors of the network.
round. A round is defined as the time between two consecutive Let consider link (x, y) weighted by Cost(x, y). WDAT-OMS
sensing of the environment. In the first phase, CHs are determined jointly uses the residual energy and inter-sensor distance to weight
by considering the residual energy of sensors, Euclidean distance this link as follows:
between sensors and polling points, and a punishment factor. In the
second phase, each sensor transmits its data to the corresponding d(x, y)
Cost(x, y) = . (8)
CH by using the first-level data aggregation tree and the hybrid CS Er(x) + Er(y)
method. In the third phase, the mobile sink starts an optimal tour to
traverse the network. It stops at the selected points to collect the This weighting method selects the sensors with more residual
buffered data in CHs, and finally, returns to its starting point. In the energy on the paths with shorter links. As a result, the energy will
following, different phases of WDAT-OMS are explained in detail. be consumed at the same rate in different sensors. On the other
hand, according to the first-order radio model, choosing the shorter
links reduces the energy consumption in the network.
4.2.1 Phase 1 – CH selection: In WSNs, CHs consume more
Based on the proposed metric, each cluster finds the best
energy than the other sensors. Thus, CH selection significantly
modified SPT by the Dijkstra algorithm. This data aggregation tree
affects the network's lifetime. WDAT-OMS uses the EDCS metric
connects all sensors of a cluster to the corresponding CH through
to select an efficient CH for each cluster by jointly considering the
the minimum-cost paths. Let P j(i, CH) be the set of all possible
residual energy, Euclidean distance, and punishment factor. This
makes the role of CH to rotate between different sensors and they paths connecting sensor i to CH in cluster j. The optimal path,
lose their energy synchronously. However, it should be noted that denoted by poptj (i, CH), is defined as the minimum-cost path among
during a round, CHs remain fixed and they may only change at the all paths in P j(i, CH), i.e.
beginning of each round. The proposed EDCS metric is as follows:

CH( j) = arg max Pr(i, r) ×

Er(i)
, (5)
popt
j (i, CH) = arg min
p j(i, CH) ∈ P j(i, CH)
∑ Cost(x, y) . (9)
i ∈ CM( j) 1 + d i, NP(i) (x, y) ∈ p j(i, CH)

where CH( j) represents the selected CH for cluster j, Er(i) denotes Certainly, by reducing the energy of sensors at different rounds,
the residual energy of sensor i, and NP(i) is the nearest polling different routing trees are created in each cluster. At each round,
point located at the transmission range of sensor i. Also, d i, NP(i) considering the new CHs and current residual energy of sensors,
shows the Euclidean distance between sensor i and NP(i). In fact, WDAT-OMS provides the routing trees in such a way that not only
among the members of cluster j, i.e. CM( j), a sensor with more balances the energy among different sensors but also decreases the
EDCS is selected as the next CH. energy consumption in the network.
In EDCS, the residual energy factor causes the high-energy In this way, each sensor transmits its data to the corresponding
sensors to have more chance to be selected as CH. This leads the parent sensor. Intermediate sensors aggregate the received data by
sensors to lose their energy fairly, and subsequently, the network their own sensed data and hierarchically send the result towards the
lifetime will be increased. Unlike the works, which use the inter- sink node. For this purpose, WDAT-OMS uses the hybrid CS

1830 IET Commun., 2020, Vol. 14 Iss. 11, pp. 1826-1837

© The Institution of Engineering and Technology 2020
 gathering applications. The mobile sink starts an optimal tour from
an appropriate point (a polling point with the closest CH) and stops
at the selected polling points to collect the buffered data from CHs.
Finally, it returns to the starting point. Once the sink receives all M
measurements from CHs, the original data of sensors can be
recovered. Certainly, at each round, the optimal tour of sink
changes depending on the location of the new CHs. WDAT-OMS
recalculates the sink trajectory after identifying CHs at the
beginning of each round, and it will remain unchanged until the
end of the round.
Towards this goal, we use the optimisation model presented in
[16] with some modifications. Let us consider the set of polling
points as P = {P0, P1, P2, …, PB − 1}. Actually, every polling point
cannot be a good candidate for stopping the mobile sink. For
example, consider a typical polling point which has no CH in its
neighbourhood. Adding this point to the optimisation problem not
only has no benefits but also increases its complexity. Therefore,
Fig. 2  First-level data aggregation tree by applying the hybrid CS method depending on the location of CHs at each round, the polling points
for a typical cluster should be considered that have at least one CH in their
neighbourhood. This set is called a ‘reduced set of polling points
method, in which both non-CS and plain-CS strategies are applied (RSP)’.
[17]. The first strategy is used for the intermediate sensors that Given the CHs obtained in phase 1, RSP, and set of neighbour
transmit less than M packets and the second one is adopted for the CHs for each candidate polling point, the goal is to find a set of
sensors with more than M packets. In the non-CS strategy, the polling points and determine the sequence to visit them, such that
intermediate sensors send the packets without any data the objective function is minimised. In fact, we aim to not only
compression. While, according to the plain-CS, the intermediate reduce the energy consumption at the second level of network, but
sensors use the CS theory to send M packets to the parent sensors. also minimise the length of the sink tour. Following [16], let us
Thus, the sink reconstructs the original data using M coded packets consider a complete directed graph, i.e. Gp = (RSP, A), where
instead of N packets (M ≪ N). Each sensor multiplies its data, i.e. ai j ∈ A represents the arc which connects polling point i to polling
x j, with the jth column vector of the measurement matrix Φ, i.e. point j with cost W(i, j). Suppose the binary variable xi, j which
[φ1 j, φ2 j, …, φM j]T, and produces a vector with size M. shows whether arc ai j belongs to the optimal tour. It equals to one,
In real networks, each sensor generates its measurement if ai j belongs to the optimal tour and zero otherwise. Also, consider
coefficient using a pseudorandom number generator seeded with its the binary variable I j that represents which polling points are
own address. Hence, there is no need for transmitting the located on the sink trajectory. I j equals to one, if polling point j is
measurement matrix to all sensors. In fact, during the data located on the sink trajectory and zero otherwise. Constraints (10)–
gathering process, each sensor requires to storage a column of the (15) model a closed route for the sink tour [16].
measurement matrix, which takes up a little memory space, and it Given the binary nature of I j, constraint (10) guarantees that
is compatible with its memory capacity. An intermediate sensor only one arc is imported to each selected polling point on the sink
waits to receive all vectors from its children and adds them to its tour. In fact, the sink tour is a closed-loop without any extra
own vector. It transmits the resulting vector to its parent sensor branch. In this context, constraint (11) guarantees that only one arc
using M packets. As explained in Section 3.2, matrix Φ has N is originated from each selected polling point on the sink tour [16].
columns and M rows so that each column is assigned to a sensor of
the network and each row corresponds to a CS measurement. Once ∑ xi, j = I j ∀ j ∈ RSP, (10)
sink receives all CS measurements, it can recover the original data i ∈ RSP, i ≠ j
of sensors by a conventional recovery algorithm. This CS-based
data aggregation method benefits from the fact that the encoding ∑ xi, j = Ii ∀i ∈ RSP . (11)
process in each sensor is carried out in a distributed manner by j ∈ RSP, i ≠ j
some simple and low-cost operations such as addition and
multiplication. In fact, the main calculations and recovery process Constraint (12) causes each CH to be located in neighbouring of
are pushed to the decoding phase on the sink node which does not at least one polling points selected on the sink tour, such that every
have the problem of limiting energy and memory. It should also be CH can directly communicate with the mobile sink [16]:
noted that the sensors do not require matrix Ψ for encoding. It is
only required at the sink for data recovery. ∑ Ii ⩾ 1 ∀CH j ∈ CH, (12)
Fig. 2 shows an example of applying the hybrid CS method for i ∈ NB(CH j)
a typical cluster. In this figure, the number written near each link
denotes the number of transmitted packets of that link. The solid where CH shows the set of CHs and NB(CH j) presents the set of
black points represent the aggregator sensors that use the plain-CS polling points located in the transmissions range of jth CH.
strategy and the white points represent the sensors that use the non- Constraints (13)–(15) prevent the creation of sub-tours in the
CS strategy. In the highlighted part of the figure, sensors 1 and 4 mobility path of the sink. In these constraints, yi, j shows the flow
transmit one packet to sensor 6 using the non-CS strategy. It value from polling point i to polling point j on arc ai j. In particular,
transmits the received packets and its own packet to sensor 8.
constraint (13) causes the flow to be passed only through the
Sensor 2 collects the packet from its children in a similar way.
polling points selected on the tour. In this constraint, RSP
Also, sensor 3 sends one packet using the non-CS strategy to
indicates the number of members in RSP. Constraint (14) specifies
sensor 8. By considering M = 6, sensor 8 transmits six packets
that the number of flow units entering the starting polling point (P0)
using the plain-CS strategy to sensor 9. Similarly, this sensor
transmits six packets to CH. is equal to the number of selected polling points on the sink tour.
Finally, constraint (15) leads to out one unit flow from each of the
other points on the tour [16].
4.2.3 Phase 3- second-level data transmission: WDAT-OMS
uses a mobile sink to collect the aggregated data from CHs at the yi, j ⩽∣ RSP ∣ xi, j ∀i, j ∈ RSP, (13)
second level of the network. Although the mobile sink is not
suitable for some real-time applications due to its delay, it is an
efficient technique to reduce the energy consumption in data
IET Commun., 2020, Vol. 14 Iss. 11, pp. 1826-1837 1831
© The Institution of Engineering and Technology 2020
 ∑ y j, P0 = ∑ I j, (14)
CHs, polling point 8 is the best stop point for the sink. In addition,
j ∈ RSP {P0} j ∈ RSP {P0} given the location of CHs in the network, selecting polling points
19, 20, and 21 instead of 14, 20, and 26 leads to a further reduction
∑ y j i − ∑ yk j = I j
, , ∀ j ∈ RSP {P0} . (15)
in delay and energy consumption.
In the cases where the network is equipped with multiple
i k
mobile sinks, it is equally partitioned into several parts. By
For the purposes described above, we propose the following considering the distribution of CHs, RSP is formed for each part to
objective function to jointly minimise both energy consumption of independently run the described optimisation model. Each mobile
sensors and delay of sink tour: sink traverses a part of the network to collect the buffered data of
the corresponding CHs via single-hop transmissions.
minimise ∑ W(i, j) subject to (10) − (15), (16)
The use of this developed scheme for large-scale networks not
only decreases the delay of the sink tour, but also reduces the
i, j ∈ RSP, i ≠ j
complexity of optimisation problem. As an example, consider Fig.
where 4 where 400 sensors are randomly distributed in a 500 m × 500 m
area. This figure compares the second level of the proposed scheme
(mi + m j)αd(i, j) β for two cases of single-sink WSN (Fig. 4a) and double-sink WSN
W(i, j) = xi, j . (17) (Fig. 4b), assuming 20 CHs, 36 polling points (6 × 6) with a grid
(ni + n j)γ length of 100 m, and R = 100 m. In Fig. 4a, a single sink starts its
tour from polling point 11 and collects the buffered data from CHs
d(i, j) represents the Euclidean distance between polling points i 1, 8, and 18. Then, it stops at polling point 16 and collects data of
and j, and mi shows the total Euclidean distances between polling CHs 2, 7, and 11. In a similar way, it stops at the other selected
point i and its neighbouring CHs. Also, ni denotes the number of polling points and collects data from the corresponding CHs via
CHs in the neighbourhood of polling point i. In fact, the proposed single-hop transmissions. Finally, it returns to the starting point. By
objective function prefers to select the polling point, which not assuming the speed of 1 m/s, the tour delay of this case is equal to
only has more number of neighbouring CHs, but also has less 1536.4 s. In contrast, Fig. 4b shows the network equipped with two
distance to its CHs. In this regard, the term of (mi + m j) gives more mobile sinks under the same conditions. In each part, a sink is
chance to the polling point, which has less distances from its responsible for collecting data from the corresponding CHs. Table
neighbouring CHs. According to the first-order radio model, this 1 shows the details of this figure. In this case, the tour delay of
reduces the energy consumption at the second level of the network. both sinks is equal to 1006.4 s. However, since both sinks
(ni + n j) selects the polling points that have more CHs in their simultaneously traverse the corresponding areas, the total delay is
neighbourhood. This reduces the number of polling points on the equal to the maximum delay of the sub-networks, i.e. 1006.4 s.
sink tour. As a result, its length will be reduced. d(i, j) also makes a Here, after describing all three phases of the WDAT-OMS
shorter tour and consequently, decreases the sink delay. It should be algorithm, it should be noted again that the decision-making centre
noted that α, β, and γ are the adjusting binary powers that control of the network is the sink node. Unlike sensors that have a simple
the weight of different factors to achieve different metrics. structure, the mobile sink can be a mobile robot or a vehicle
Consider Fig. 3 to get more insight into the problem. By equipped with a rechargeable battery, powerful CPU, large
assuming a network consisted of 20 random CHs and 36 polling memory, and GPS device. Thus, the basic calculations, including
points (6 × 6), Figs. 3a and b show the sink tour obtained by the cost of links, CH selection, optimal tour etc., are made in the
WDAT-OMS and SHDGP [16], respectively. The grid length and sink node. To clarify the problem, let us reconsider the proposed
transmission range are assumed to be equal. In Fig. 3a, the mobile EDCS metric in (5). In the following, we particularly describe
sink starts its tour from polling point 8. Compared to the tour in where and how CHs are determined. However, the cost of links,
Fig. 3b, it travels less distance to traverse the network and collects and subsequently, the first-level data aggregation trees are
data from CHs. In Fig. 3b, the selection of polling points 1, 14, and determined in a similar way.
20 is inefficient and increases both delay and energy consumption. According to (5), three factors are required to determine the CH
As an example, CH 5 is located in the neighbourhood of polling of each cluster: residual energy of sensors, Euclidean distance
points 1, 2, 7, and 8. However, with respect to the other adjacent between each sensor and its neighbour polling points, and
punishment factor of sensors. However, in the first round, only the

Fig. 3  A typical example for sink tour considering 20 CHs and 36 polling points (6 × 6) formed by
(a) WDAT-OMS algorithm, (b) SHDGP algorithm [16]

1832 IET Commun., 2020, Vol. 14 Iss. 11, pp. 1826-1837

© The Institution of Engineering and Technology 2020
Fig. 4  A typical example for data aggregation in both levels of the network by assuming 400 sensors (N = 400), 20 clusters (C = 20), and 36 polling points
(6 × 6)
(a) Single-sink WSN, (b) Double-sink WSN

Table 1 List of selected polling points and corresponding sensors. Once the mobile sink reaches a selected polling point, it
CHs in Fig. 4b assuming 20 CHs, 36 polling points, and two stops and sends a beep to wake up the transceiver of the
mobile sinks neighbouring CHs. It collects the buffered data and control
Selected polling Corresponding Corresponding information from CHs, and at the same point calculates the new
point sink CHs configuration of the corresponding clusters for the next round.
Subsequently, the mobile sink locally returns the result to the
11 1 8, 18
corresponding CHs. Each CH notifies its sensors of the new
5 1 1, 16 configuration through the existing first-level routing tree. Finally,
10 1 2, 11 the mobile sink moves to the next points to complete its trajectory.
8 1 4, 13 At the beginning of the next round, it determines its new optimal
16 1 7 trajectory based on the presented scheme and starts to aggregate
18 1 3, 19 data of new CHs.
23 2 15
21 2 12 5 Simulation results
19 2 10 In this section, several scenarios are presented to compare the
26 2 5 performance of the proposed algorithm with ‘centralised clustering
34 2 14 algorithm (CCA)’ [14], ‘ECDA with constant sink in centre’ [13],
29 2 6, 9, 17, 20 ‘EHDT with constant sink in (0,0)’ [13], and SHDGP [16]. With
the aim of a fair comparison, the CCA algorithm is modified to be
run round by round. Furthermore, to evaluate the described strategy
for providing the first-level data aggregation trees, we use the hop-
distance factor is required for determining CHs. This is because all
count metric once again for constructing these trees. In the
sensors have approximately the same initial energy and their
following, this is referred to as ‘data aggregation trees with optimal
punishment factor is also equal to one. At the next rounds, CHs are
mobile sink(s) (DAT-OMS)’ to be distinct from WDAT-OMS. In
selected based on the aggregated control information at the
fact, in DAT-OMS, each sensor transmits its data to the
previous round. In fact, at round r, CHs are selected based on the
corresponding CH through the pure SPT and by utilising the hybrid
aggregated control information at round r − 1. For this purpose,
CS method. It should be noted that the first and third phases of
during the data gathering process at round r − 1, the information of
DAT-OMS are completely similar to the WDAT-OMS algorithm.
all sensors is also collected locally. This information includes the
The performance of the algorithms is evaluated in terms of
residual energy of each sensor and the round numbers in which that
energy consumption, energy consumption variance, and tour delay.
sensor is selected as CH. Certainly, the mobile sink is well aware
Both energy consumption and energy consumption variance are
of the location of sensors and polling points by the described
calculated after the 10th round. From the perspective of statistics
preliminary signalling at the initial deployment of sensors. Each
theory, the variance is a suitable metric that shows how far a set of
sensor knows the length of packets it must send to its parent. It is
numbers are spread out from their average value. In fact, a
also well aware of its distance to its parent. Thus, each sensor can
easily calculate how much energy it will have at the end of the reduction in the energy consumption variance (σE2 ) demonstrates a
current round. The number of times a sensor is selected as CH is better load-balancing in the network. This metric is defined as
also used to calculate its punishment factor. Accordingly, each follows:
sensor transmits its information along with the sensed data to the
N
corresponding CH via the first-level routing tree at round r − 1. 1
N i∑
2
σE2 = E0 − Er(i) − Ec , (18)
This information is very small compared to the sensed data. =1
However, we consider a fractional of each packet for this control
information. The current CH of each cluster buffers this where E0 and Ec are the initial energy and average energy
information and sends it along with the aggregated data to the
consumption of sensors, respectively. Also, since the speed of sink
mobile sink.
(V ) is assumed to be 1 m/s, tour delay is calculated as the sum of
As mentioned before, the clusters are constant during the
network activity and only the role of CH is rotated between the

IET Commun., 2020, Vol. 14 Iss. 11, pp. 1826-1837 1833

© The Institution of Engineering and Technology 2020
Table 2 Simulation parameters
Parameters Value Parameters Value
L 1024 bits γ 1
ETx‐elec 50 nJ/bit R 100 m
ϵamp 100 pJ/bit/m2 N 400
E0 5 kJ V 1 m/s
ERx‐elec 50 nJ/bit b 2
α 1 ρ 10
β 1 a 500 m

Fig. 6  Energy consumption variance as a function of the number of

sensors for 15 clusters (C = 15) in scenario 1

Fig. 5  Energy consumption as a function of the number of sensors for 15

clusters (C = 15) in scenario 1

the Euclidean distances between the consecutive polling points on

the tour.
By assuming R = 100 m and B = 36 (6 × 6), N sensors are
randomly distributed over a 500 m × 500 m square area. We
consider a grid arrangement for polling points in which each
polling point is separated from its adjacent neighbours in horizontal
and vertical directions by the same distance named grid length.
Since there should be at least one polling point in the
neighbourhood of each sensor, the grid length is assumed to be
equal to the transmission range of sensors, i.e. 100 m.
In the following, considering four parameters N, ρ, C, and S, we
present several scenarios to study different aspects of the Fig. 7  Energy consumption variance as a function of the number of
algorithms. N is the number of sensors, ρ = N /M represents the sensors for four clusters (C = 4) in scenario 1
compressive ratio, C denotes the number of clusters, and S shows
the number of mobile sinks. In all scenarios, the simulation problems of CS theory, efficient clustering, CH selection, and sink
parameters are set according to Table 2 unless otherwise is mobility. For CH selection, both algorithms use the EDCS metric,
described. It is assumed that a part of the 1024-bit packets (100 bits which jointly considers the residual energy, Euclidean distance, and
on average) is overhead required for synchronisation, preambles, punishment factor. The mobile sink traverses the network optimally
addressing etc. Also, for simplicity and following the other works and stops at the selected polling points to collect the buffered data
[5, 20], the electrical energy required to run the circuitry in both of the closest CHs. This reduces the number of transmissions and
transmitter and receiver is assumed to be 50 nJ/bit. consequently, decreases energy consumption. However, WDAT-
Since the sensors are randomly distributed in the network, we OMS consumes less energy than DAT-OMS and it balances the
run the algorithms several times and finally, the average of results energy among the sensors more fairly. In WDAT-OMS, each sensor
is presented. Each data point in the graphs is obtained by averaging transmits its data to the corresponding CH via a weighted routing
the results of 50 individual runs. The error bar at each data point tree. This tree consists of sensors with more residual energy on the
represents the standard deviation of the results. In each run, the paths with shorter links. This improves the load-balancing in the
network topology is checked to be connected. If the generated network compared to the hop-count metric (Fig. 6). DAT-OMS
topology has at least one isolated sensor, it is regenerated until a uses the hop-count metric that leads to trees consisting of long-
connected network is obtained. distance hops. According to the first-order radio model, this
Scenario 1: In the first scenario, assuming C = 15, we increases the energy consumption in the network (Fig. 5).
investigate the effect of the number of sensors on the network It is also worth noting that as the number of clusters decreases,
performance. The results are shown in Figs. 5 and 6. By increasing the superiority of WDAT-OMS becomes more apparent in terms of
the number of sensors, more data should be sent over the network. load balancing. For example, Fig. 7 shows the energy consumption
This not only exacerbates the energy consumption (Fig. 5), but also variance for C = 4 in the same conditions. In fact, as the number of
reduces the load-balancing in the network (Fig. 6). clusters decreases, the number of sensors in each cluster increases.
The results show that both energy consumption and energy Thus, the first-level data aggregation trees can balance the energy
consumption variance in the proposed algorithms are smaller than among different sensors more efficiently.
CCA, ‘ECDA with constant sink in centre’, and ‘EHDT with
constant sink in (0,0)’. WDAT-OMS and DAT-OMS address the
1834 IET Commun., 2020, Vol. 14 Iss. 11, pp. 1826-1837
© The Institution of Engineering and Technology 2020
 Fig. 10  Energy consumption as a function of the network density for 25
Fig. 8  Energy consumption as a function of the number of clusters for 400
clusters (C = 25) in scenario 3
sensors (N = 400) in scenario 2

Fig. 11  Energy consumption variance as a function of the network density

Fig. 9  Energy consumption variance as a function of the number of for 25 clusters (C = 25) in scenario 3
clusters for 400 sensors (N = 400) in scenario 2
It is clear that ‘EDCS with punishment factor’ has the best
Scenario 2: In this scenario, assuming N = 400, the number of performance in terms of energy consumption and energy
clusters varies from 7 to 21. The results are shown in Figs. 8 and 9. consumption variance. Unlike the strategy presented in [13, 20],
As the number of clusters increases, each of them becomes EDCS jointly considers the Euclidean distances between sensors
smaller. Accordingly, the number of transmissions at the first level and polling points, residual energy of sensors, and punishment
of the network is reduced. In the proposed algorithms, thanks to the factor to select CHs. Given the use of the mobile sink, it prefers to
use the optimal mobile sink, CHs transmit the buffered data select the sensors with less distance to the polling points as CHs.
through the single-hop transmission. Thus, by increasing the Thus, according to the first-order radio model, the second-level
number of clusters, the energy consumption of these algorithms is energy consumption is reduced. Furthermore, both energy
decreased. However, other algorithms use the second-level routing consumption and energy consumption variance in ‘EDCS with
tree to transmit the CS measurements from CHs to the sink. By punishment factor’ are smaller than ‘EDCS without punishment
increasing the number of clusters, this tree grows. Thus, due to the factor’. In particular, by eliminating the punishment factor, energy
increasing number of transmissions, the energy consumption also consumption variance is significantly increased. This factor
increases. reduces the chance of re-selecting the sensors previously chosen as
Furthermore, as justified before, both energy consumption and CH. Thus, the role of CH rotates between different sensors and
energy consumption variance of the proposed algorithms, i.e. DAT- they lose their energy synchronously.
OMS and WDAT-OMS, are smaller than the other algorithms for a In confirmation of the results obtained in scenario 1, the energy
different number of clusters. consumption increases as the number of sensors increases. In fact,
Scenario 3: In this scenario, assuming C = 25, we study the by increasing the number of sensors, more data should be sent over
efficiency of the proposed CH selection metric, i.e. EDCS, in the network. This not only increases the energy consumption (Fig.
comparison with the other metrics for two cases of N = 200 and 10), but also reduces the load-balancing in the network; thus, the
N = 600. With the aim of a fair comparison, we use the CH energy consumption variance will be increased (Fig. 11).
selection metric of [13, 20] in our algorithm. We also consider the Scenario 4: In this scenario, assuming N = 400 and C = 10, we
EDCS metric for both cases of ‘with’ and ‘without’ the punishment compare the DAT-OMS and WDAT-OMS algorithms for different
factor to evaluate the effect of this factor on the performance of the values of α and different number of sinks. The results are shown in
network. Figs. 10 and 11 compare the energy consumption and Figs. 12 and 13. Let us consider the objective function presented in
energy consumption variance for both low density network (16) and (17). As α increases, the chance of selecting the closer
(N = 200) and high density network (N = 600), respectively. polling points to CHs will be increased. Thus, according to the

IET Commun., 2020, Vol. 14 Iss. 11, pp. 1826-1837 1835

© The Institution of Engineering and Technology 2020
Fig. 12  Energy consumption as a function of parameter α for 400 sensors
(N = 400) and ten clusters (C = 10) in scenario 4 Fig. 14  Delay of sink as a function of the number of clusters in scenario 5

Fig. 13  Delay of sink as a function of parameter α for 400 sensors Fig. 15  Time complexity as a function of the number of clusters for 100
(N = 400) and ten clusters (C = 10) in scenario 4 sensors (N = 100) and 300 sensors (N = 300) in scenario 6

first-order radio model, energy consumption is reduced (Fig. 12). Scenario 6: In the last scenario, we study the time complexity
This is associated with an increase in the sink delay (Fig. 13). For of DAT-OMS and WDAT-OMS for a different number of clusters.
α = 0.5, the most energy consumption and the least delay is The results are presented in Fig. 15 for two cases of N = 100 and
observed. N = 300. All measurements are made on a computer equipped with
On the other hand, by increasing the number of mobile sinks, an Intel core i7 2.40 GHz processor and 6 GB of memory. As the
the number of partitions in the network is increased. This divides a number of clusters increases, the time complexity of both
single tour into multiple parallel sub-tours. Since all sinks algorithms intensifies. Increasing the number of clusters leads to an
simultaneously traverse the corresponding area, the total delay is increase in the number of CHs. This aggrandises the dimensions of
equal to the maximum delay of sub-tours. Thus, the sink delay of the adopted ILP optimisation model in the third phase of the
‘DAT-OMS with S = 4’ and ‘WDAT-OMS with S = 4’ are less algorithms.
than the other cases. It is worth noting that both DAT-OMS and Unlike DAT-OMS, which uses the hop-count metric for
WDAT-OMS have the same method for CH selection and sink constructing the first-level data aggregation trees, WDAT-OMS
trajectory formation. This causes both algorithms to have the same adopts the metric introduced in (8) to efficiently weight different
sink delay (Fig. 13). links of the network. Although a weighted data aggregation tree
Scenario 5: In this scenario, we compare the tour delay in the provides better load-balancing in the network, it increases the time
proposed algorithms and SHDGP [16] for a different number of complexity. Thus, the WDAT-OMS algorithm has more time
clusters and a different number of sinks. With the aim of a fair complexity than the DAT-OMS algorithm. Furthermore, as the
comparison, SHDGP is used at the second level of the network to number of sensors increases, the time required for calculating the
directly collect the aggregated data of CHs. The results are shown weighted graph and constructing the first-level routing trees is also
in Fig. 14. In this figure, due to the similar delay of both DAT- increased. Accordingly, as shown in Fig. 15, the time complexity
OMS and WDAT-OMS algorithms, we show them as ‘DAT/ of the algorithms for N = 300 is more than N = 100.
WDAT-OMS’. The proposed algorithms use an improved objective
function, which not only aims to reduce the length of the sink tour,
but also minimises the number of selected polling points. Thus, in
6 Conclusion
the same conditions, i.e. using a single sink, DAT/WDAT-OMS has In this study, we proposed an efficient algorithm named WDAT-
less delay than the SHDGP algorithm. Furthermore, as the number OMS to jointly address the problems of CS theory, cluster-based
of clusters increases, the number of CHs also increases. This leads routing, and sink mobility in WSNs. At the first level, WDAT-
to a longer trajectory for the sink. As a result, its delay increases. OMS organises the sensors into several clusters. In each cluster, a
On the other hand, as justified before, increasing the number of weighted routing tree is used which results in fair energy
sinks significantly reduces the delay for a different number of consumption by different sensors. Thus, the energy consumption
clusters. variance is reduced. At the second level, one or more mobile sinks
traverse the network to collect the aggregated data of CHs through

1836 IET Commun., 2020, Vol. 14 Iss. 11, pp. 1826-1837

© The Institution of Engineering and Technology 2020
a single-hop transmission. The simulation results show that the [20] Abbasi, S., Abouei, J.: ‘Toward cluster-based weighted compressive data
aggregation in wireless sensor networks’, Ad Hoc Netw., 2016, 36, pp. 368–
energy consumption of WDAT-OMS is less than CCA [14], 386
‘ECDA with constant sink in centre’ [13], and ‘EHDT with [21] Sun, P., Wu, L., Wang, Z., et al.: ‘Sparsest random sampling for cluster-based
constant sink in (0,0)’ [13]. By increasing the number of clusters, compressive data gathering in wireless sensor networks’, IEEE Access, 2018,
each cluster covers a less number of sensors. This generally 6, pp. 36383–36394
[22] Handy, M.J., Hasse, M., Timmermann, D.: ‘Low energy adaptive clustering
reduces the energy consumption in the network. However, the hierarchy with deterministic cluster-head selection’. Proc. 4th Int. Workshop
delay of the sink tour will be increased. We also compared the tour on Mobile and Wireless Communications Network, Stockholm, Sweden,
delay of the proposed algorithm with SHDGP [16] for a different 2002, pp. 368–372
number of clusters and a different number of sinks. The proposed [23] Li, C., Ye, M., Chen, G., et al.: ‘An energy-efficient unequal clustering
mechanism for wireless sensor networks’. Proc. IEEE Int. Conf. on Mobile
algorithm reduces the sink delay thanks to the use of an improved Adhoc and Sensor Systems Conf., Washington, DC, USA, 2005, pp. 598–604
objective function. Furthermore, the use of multiple mobile sinks [24] Ye, M., Li, C., Chen, G., et al.: ‘EECS: an energy efficient clustering scheme
breaks down a single tour into several parallel sub-tours and in wirless sensor networks’. Proc. 24th IEEE Int. Performance, Computing,
significantly reduces the network delay. and Communications Conf. (PCCC 2005), Phoenix, AZ, USA, 2005, pp. 535–
540
[25] Pachlor, R., Shrimankar, D.: ‘LAR-CH: a cluster-head rotation approach for
7 References sensor networks’, IEEE Sens. J., 2018, 18, (23), pp. 9821–9828
[26] Behera, T.M., Mohapatra, S.K., Samal, U.C., et al.: ‘Residual energy based
[1] Lazarescu, M.T.: ‘Design of a WSN platform for long-term environmental cluster-head selection in WSNs for IoT applications’, IEEE Internet Things J.,
monitoring for IoT applications’, IEEE J. Emerging Sel. Top. Circuits Syst., 2019, 6, (3), pp. 5132–5139
2013, 3, (1), pp. 45–54 [27] Gherbi, C., Aliouat, Z., Benmohammed, M.: ‘An adaptive clustering approach
[2] Barrenetxea, G., Ingelrest, F., Schaefer, G., et al.: ‘Wireless sensor networks to dynamic load balancing and energy efficiency in wireless sensor networks’,
for environmental monitoring: the sensor scope experience’. 2008 IEEE Int. Energy, 2016, 114, pp. 647–662
Zurich Seminar on Communications (IZS), Zurich, Switzerland, 2008, pp. [28] Siavoshi, S., Kavian, Y., Sharif, H.: ‘Load-balanced energy efficient
98–101 clustering protocol for wireless sensor networks’. IET Wirel. Sens. Syst., 2016,
[3] Lu, Y., Kuonen, P., Hirsbrunner, B., et al.: ‘Benefits of data aggregation on 6, (3), pp. 67–73
energy consumption in wireless sensor networks’, IET Commun., 2017, 11, [29] Laouid, A., Dahmani, A., Bounceur, A., et al.: ‘A distributed multi-path
(8), pp. 1216–1223 routing algorithm to balance energy consumption in wireless sensor
[4] Haupt, J., Bajwa, W., Rabbat, M., et al.: ‘Compressed sensing for networked networks’, Ad Hoc Netw., 2017, 64, pp. 53–64
data’, IEEE Signal Process. Mag., 2008, 25, (2), pp. 92–101 [30] Wang, C., Shih, J., Pan, B., et al.: ‘A network lifetime enhancement method
[5] Wu, X., Xiong, Y., Huang, W., et al.: ‘An efficient compressive data gathering for sink relocation and its analysis in wireless sensor networks’, IEEE Sens.
routing scheme for large-scale wireless sensor networks’, Comput. Electr. J., 2014, 14, (6), pp. 1932–1943
Eng., 2013, 39, (6), pp. 1935–1946 [31] Jea, D., Somasundara, A.A., Srivastava, M.B.: ‘Multiple controlled mobile
[6] Ding, P., Holliday, J., Celik, A.: ‘Distributed energy-efficient hierarchical elements (data mule) for data collection in sensor networks’. Proc. IEEE/
clustering for wireless sensor networks’. Proc. Int. Conf. on Distributed ACMDCOSS, Berlin, Germany, 2005, pp. 244–257
Computing in Sensor Systems, Berlin, Germany, 2005, pp. 322–339 [32] Luo, J., Hubaux, J.P.: ‘Joint mobility and routing for lifetime elongation in
[7] Ramesh, K., Somasundaram, K.: ‘A comparative study of cluster head wireless sensor networks’. Proc. IEEE 24th Annual Joint Conf. of the IEEE
selection algorithms in wireless sensor networks’, Int. J. Comput. Sci. Eng. Computer and Communications Societies, Miami, FL, USA, 2005, pp. 1735–
Surv. (IJCSES), 2011, 2, (4), pp. 153–164 1746
[8] Kang, S.H., Nguyen, T.: ‘Distance based thresholds for cluster head selection [33] Muthu Krishnan, A., Ganesh Kumar, P.: ‘An effective clustering approach
in wireless sensor networks’, IEEE Commun. Lett., 2012, 16, (9), pp. 1396– with data aggregation using multiple mobile sinks for heterogeneous WSN’,
1399 Wirel. Pers. Commun., 2016, 90, (2), pp. 423–434
[9] Ying, L., Haibin, Y.: ‘Energy adaptive cluster-head selection for wireless [34] Nazir, B., Hasbullah, H.: ‘Mobile sink based routing protocol (MSRP) for
sensor networks’. Proc. Sixth Int. Conf. on Parallel and Distributed prolonging network lifetime in clustered wireless sensor network’. Proc. 2010
Computing, Applications and Technologies (PDCAT’05), Dalian, China, Int. Conf. on Computer Applications and Industrial Electronics, Kuala
2005, pp. 634–638 Lumpur, Malaysia, 2010, pp. 624–629
[10] Shankar, A., Jaisankar, N., Khan, M.S., et al.: ‘Hybrid model for security- [35] Abo-zahhad, M., Ahmed, S.M., Sabor, N., et al.: ‘Mobile sink-based adaptive
aware cluster head selection in wireless sensor networks’, IET Wirel. Sensor immune energy-efficient clustering protocol for improving the lifetime and
Syst., 2019, 9, (2), pp. 68–76 stability period of wireless sensor networks’, IEEE Sens. J., 2015, 15, (8), pp.
[11] Gu, Y., Ren, F., Ji, Y., et al.: ‘The evolution of sink mobility management in 4576–4586
wireless sensor networks: a survey’, IEEE Commun. Surv. Tutor., 2016, 18, [36] Wang, J., Cao, J., Ji, S., et al.: ‘Energy-efficient cluster-based dynamic routes
(1), pp. 507–524 adjustment approach for wireless sensor networks with mobile sinks’, J.
[12] Barani, H., Jaradat, Y., Huang, H., et al.: ‘Effect of sink location and Supercomput., 2017, 73, (7), pp. 3277–3290
redundancy on multi-sink wireless sensor networks: a capacity and delay [37] Xie, G., Pan, F.: ‘Cluster-based routing for the mobile sink in wireless sensor
analysis’, IET Commun., 2018, 12, (8), pp. 941–947 networks with obstacles’, IEEE Access, 2016, 4, pp. 2019–2028
[13] Pakdaman Tirani, S., Avokh, A.: ‘On the performance of sink placement in [38] Patan, R., Babu, M.R.: ‘Re-storm: real-time energy efficient data analysis
WSNs considering energy-balanced compressive sensing-based data adapting storm platform’, J. Teknol., 2016, 78, (10), pp. 139–146
aggregation’, J. Netw. Comput. Appl., 2018, 107, pp. 38–55 [39] Suresh, K., Babu, M.R., Patan, R.: ‘EEIoT: energy efficient mechanism to
[14] Xie, R., Jia, X.: ‘Transmission-efficient clustering method for wireless sensor leverage the internet of things (IoT)’, Proc. 2016 Int. Conf. on Emerging
networks using compressive sensing’, IEEE Trans. Parallel Distrib. Syst., Technological Trends (ICETT), Kollam, India, 2016
2014, 25, (3), pp. 806–815 [40] Poongodi, T., Khan, M.S., Patan, R., et al.: ‘Robust defense scheme against
[15] Singanamalla, V., Patan, R., Khan, M.S., et al.: ‘Reliable and energy-efficient selective drop attack in wireless ad hoc networks’, IEEE Access, 2019, 7, pp.
emergency transmission in wireless sensor networks’, Internet Technol. Lett., 18409–18419
2019, 2, (2), pp. 1–6 [41] Avokh, A., Mirjalily, G.: ‘Dynamic balanced spanning tree (DBST) for data
[16] Ma, M., Yang, Y., Zhao, M.: ‘Tour planning for mobile data-gathering aggregation in wireless sensor networks’. Proc. 2010 5th Int. Symp. on
mechanisms in wireless sensor networks’, IEEE Trans. Veh. Technol., 2013, Telecommunications (IST’2010), Tehran, India, 2010, pp. 391–396
62, (4), pp. 1472–1483 [42] Candes, E., Wakin, M.: ‘An introduction to compressive sampling’, Signal
[17] Ebrahimi, D., Assi, C.: ‘Compressive data gathering using random projection Process. Mag., 2008, 25, (2), pp. 21–30
for energy efficient wireless sensor networks’, Ad Hoc Netw., 2014, 16, pp. [43] Qaisar, S., Bilal, R., Iqbal, W., et al.: ‘Compressive sensing: from theory to
105–119 applications, a survey’, Commun. Netw., 2013, 15, (5), pp. 443–456
[18] Zhang, P., Wang, J., Guo, K.: ‘Compressive sensing and random walk based [44] Han, G., Jiang, J., Zhang, C., et al.: ‘A survey on mobile anchor node assisted
data collection in wireless sensor networks’, Comput. Commun., 2018, 129, localization in wireless sensor networks’, IEEE Commun. Surv. Tutor., 2016,
pp. 43–53 18, (3), pp. 2220–2243
[19] Gu, X., Zhou, X., Yuan, B., et al.: ‘A bayesian compressive data gathering
scheme in wireless sensor networks with one mobile sink’, IEEE Access,
2018, 6, pp. 47897–47910

IET Commun., 2020, Vol. 14 Iss. 11, pp. 1826-1837 1837

© The Institution of Engineering and Technology 2020