Distributed unequal clustering algorithm in large-scale wireless sensor networks using fuzzy logic

Abstract

Clustering is a promising and popular approach to organize sensor nodes into a hierarchical structure, reduce transmitting data to the base station by aggregation methods, and prolong the network lifetime. However, a heavy traffic load may cause the sudden death of nodes due to energy resource depletion in some network regions, i.e., hot spots that lead to network service disruption. This problem is very critical, especially for data-gathering scenarios in which Cluster Heads (CHs) are responsible for collecting and forwarding sensed data to the base station. To avoid hot spot problem, the network workload must be uniformly distributed among nodes. This is achieved by rotating the CH role among all network nodes and tuning cluster size according to CH conditions. In this paper, a clustering algorithm is proposed that selects nodes with the highest remaining energy in each region as candidate CHs, among which the best nodes shall be picked as the final CHs. In addition, to mitigate the hot spot problem, this clustering algorithm employs fuzzy logic to adjust the cluster radius of CH nodes; this is based on some local information, including distance to the base station and local density. Simulation results demonstrate that, by mitigating the hot spot problem, the proposed approach achieves an improvement in terms of both network lifetime and energy conservation.

Full text

uncorrected proof
J Supercomput
https://doi.org/10.1007/s11227-018-2261-5
Distributed unequal clustering algorithm in large-scale
wireless sensor networks using fuzzy logic
Peyman Neamatollahi1·Mahmoud Naghibzadeh2
© Springer Science+Business Media, LLC, part of Springer Nature 2018
Abstract Clustering is a promising and popular approach to organize sensor nodes1
into a hierarchical structure, reduce transmitting data to the base station by aggregation2
methods, and prolong the network lifetime. However, a heavy traffic load may cause3
the sudden death of nodes due to energy resource depletion in some network regions,4
i.e., hot spots that lead to network service disruption. This problem is very critical,5
especially for data-gathering scenarios in which Cluster Heads (CHs) are responsible6
for collecting and forwarding sensed data to the base station. To avoid hot spot problem,7
the network workload must be uniformly distributed among nodes. This is achieved8
by rotating the CH role among all network nodes and tuning cluster size according to9
CH conditions. In this paper, a clustering algorithm is proposed that selects nodes with10
the highest remaining energy in each region as candidate CHs, among which the best11
nodes shall be picked as the final CHs. In addition, to mitigate the hot spot problem,12
this clustering algorithm employs fuzzy logic to adjust the cluster radius of CH nodes;13
this is based on some local information, including distance to the base station and local14
density. Simulation results demonstrate that, by mitigating the hot spot problem, the15
This work was supported by the Islamic Azad University, Mashhad Branch [Grant Number 91368.400.7].
The first version of this work was published in 7th International Symposium on Telecommunications
(IST) [38]. In comparison to the preliminary version, in the current version, we improved the presentation
throughout the manuscript, stated the contributions in details, provided a new section for the reference
protocols (Sect. 2), modified the clustering algorithm and membership functions of the fuzzy inference
system, compared the proposed approach with recent protocols (HEED, M-LEACH, and DUCF), and
evaluated the performance of UCF in different conditions and scenarios.
BMahmoud Naghibzadeh
naghibzadeh@um.ac.ir; naghibzadeh@mshdiau.ac.ir
1Department of Computer Engineering, Faculty of Engineering, Ferdowsi University of Mashhad,
Mashhad, Iran
2Department of Computer Engineering, Mashhad Branch, Islamic Azad University, Mashhad, Iran
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
P. Neamatollahi, M. Naghibzadeh
proposed approach achieves an improvement in terms of both network lifetime and16
energy conservation.17
Keywords Energy efficiency ·Fuzzy logic ·Hierarchical routing algorithm ·Hot18
spot problem ·Network lifetime ·Wireless sensor networks19
1 Introduction20
In the past years, rapid technological developments in low-power wireless commu-21
nication and microelectromechanical systems (MEMS) have led to the production of22
small-sized, inexpensive, and low-power sensor nodes with embedded communica-23
tion and computing and sensing capabilities. The hundreds or even thousands of such24
power-constrained sensor devices are required to organize a practical wireless sensor25
network (WSN), which is often designed for detecting and monitoring the physical26
parameters of a sensing area by the use of radio frequencies [1–3]. Much attention27
has been given to WSNs because of their varied potential applications, such as traffic28
control, intrusion detection, object tracking, environmental monitoring, healthcare,29
and industrial inventory management [4–7].
130
As these sensor nodes are often scattered in non-rechargeable environments and31
have been equipped with limited batteries, it is necessary that the nodes’ energy32
resource be wisely managed so as to prolong the nodes’ lifetime. Although this major33
issue has been highlighted in signal processing techniques and low-power hardware34
design, energy-efficient algorithms must also be applied at different network layers.35
In addition to generally decreasing energy dissipation, balancing energy consump-36
tion among all sensor nodes is very significant so that the network lifetime may be37
extended.38
1.1 Clustering39
Within a sensor node, the energy spent in the radio unit is much higher than that in40
other units. The amount of this energy dissipation is often proportional to the square41
distance between the transmitter and receiver nodes. To collect information efficiently42
in periodical data-gathering applications, network nodes are grouped into clusters. In43
cluster-based WSNs, the privilege for communicating with the distant base station (BS)44
is only granted to some sensor nodes called Cluster Heads (CHs) [8,9]. In addition,45
as each cluster is often formed from nodes in one geographical region, every CH may46
remove the existing high redundancy in the data reported by its member nodes. To47
accomplish this, data collection operations may be organized into two steps: intra- and48
inter-cluster communications. At first, the data transmitted by the regular (non-CH)49
nodes in each cluster are received in the respective CH and then the aggregated data50
are sent to the BS by every CH [10].51
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
Distributed unequal clustering algorithm in large-scale…
1.2 Problem statement52
Most of the previous research in the literature [11–16] has employed multi-hop for53
inter-cluster and single-hop for intra-cluster communication. The reason is that, due to54
wireless channel characteristics in large-scale WSNs, it is an energy-efficient operation55
for CHs to cooperate with each other to deliver data to the BS. However, the hot spot56
problem is created due to this many-to-one traffic pattern, in which the energy of some57
nodes in particular areas is depleted sooner than that of other nodes in the network, thus58
causing network partitions and diminished sensing coverage. For instance, hot spots59
are created by a heavier relay traffic imposed on nodes closer to the BS in multi-hop60
networks.61
To improve energy efficiency in WSN, many clustering algorithms [12,16–34]have62
been proposed. However, due to the unbalanced energy consumption, the appearance63
of hot spots in practice reduces the effectiveness of these algorithms. Some unequal64
clustering algorithms [11,13,14,35–37] have been proposed to mitigate this prob-65
lem. In these algorithms, the network is split into clusters of different sizes having an66
approximately equal workload. In other words, to save energy in intra-cluster com-67
munication for bearing the inter-cluster heavy workload, the size of clusters in the68
vicinity of the BS is smaller than that of clusters far from the BS (see Fig. 1).69
1.3 The contribution70
This paper proposes an Unequal Clustering approach using Fuzzy logic (UCF)71
[38]. The UCF is a self-organized competition-based distributed clustering algorithm72
designed for large-scale WSNs that periodically report environmental information73
to the BS. It requires no special node capabilities, such as heterogeneity or location74
awareness. By evenly distributing the workload over the nodes, the UCF mitigates the75
hot spot problem and thus prolongs the network lifetime. The main contributions of76
UCF may be summarized as follows:77
•Some of the current approaches [11,14] employ the nodes’ residual energy in78
combination with the location properties so as to determine the nodes’ cluster79
radius. Using these approaches, by diminishing the residual energy of nodes, the80
size of clusters also decreases which leads to increasing the number of clusters81
and multi-hop routes on the CHs. Therefore, the hot spot issue may be intensified82
in some cases. In addition to directly utilizing the residual energy of nodes in the83
CH election process, one of the current paper’s contributions during the clustering84
process is to use some of the node’s local information as fuzzy attributes so as to85
accurately compute the cluster radius. In other words, the residual energy of the86
nodes is taken into account for selecting the most eligible node as a CH in each87
neighborhood. Besides, the cluster radius of each node is calculated based on the88
location properties of nodes by using the fuzzy inference system (FIS).89
•Most of the existing unequal clustering approaches [11,14,36,39,40] choose the90
final CHs from among randomly selected candidate CHs. Sometimes, the net-91
work nodes’ lack of full coverage by these tentative CHs results in the generation92
of orphan nodes which do not belong to any cluster. In contrast, the UCF is a93
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
P. Neamatollahi, M. Naghibzadeh
Fig. 1 Unequal clustering
distributed clustering algorithm in which CHs are chosen primarily based on the94
residual energy of nodes, without the usage of a random function. On the other95
hand, due to the fair distribution of CHs in the network area, the proposed clustering96
algorithm does not permit orphan node generation.97
•The current unequal clustering algorithms [11,36,40] have often assumed the uni-98
form deployment of nodes in the area. However, in most of the WSNs’ practical99
applications, nodes are randomly dispersed [1]. Consequently, the unequal cluster-100
ing algorithms, which are only based on the nodes’ distance from the BS, may still101
suffer from unbalanced energy consumption due to the variation of the local den-102
sity in different regions. With respect to the random distribution of sensor nodes,103
this paper, in addition to the distance to the BS, considers the local density of nodes104
for determining the nodes’ cluster radius.105
•In contrast to many other algorithms [13,22,41], the clustering energy overhead of106
UCF is low due to a decrease in the number of messages exchanged for constructing107
the clusters.108
Simulation results show that the proposed approach balances energy consumption109
across the network. Therefore, in terms of network lifecycle and energy conservation,110
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
Distributed unequal clustering algorithm in large-scale…
the UCF overcomes the well-known clustering algorithms of M-LEACH [41], HEED111
[12], and DUCF [13] to which it is compared.112
1.4 Paper outline113
The rest of the paper is organized as follows. Section 2reviews the reference protocols114
which will be compared to the proposed protocol. Preliminaries of the scheme con-115
sisting of the radio energy model and system assumptions are introduced in Sect. 3.116
Section 4describes the fuzzy inference system and proposes a fuzzy logic-based117
unequal clustering algorithm. By employing some experiments in the simulation envi-118
ronment, Sect. 5evaluates the performance of the UCF. The paper is concluded in119
Sect. 6.120
2 Reference protocols121
This section explains the well-known and closely related distributed clustering pro-122
tocols with which this study is compared: M-LEACH, HEED, and DUCF. These123
protocols have been proposed with the aim of increasing network lifetime.124
2.1 Multi-hop low energy adaptive clustering hierarchy (M-LEACH)125
M-LEACH [41] is a multi-hop version of LEACH protocol [22] which is one of the126
primary and most famous protocols presented in the literature. It minimizes energy127
consumption by using dynamic clustering. The protocol’s operation is divided into two128
main parts: the setup phase and steady-state phase. In the first phase, the process of129
clustering is performed. To achieve this, each node selects a random number between 0130
and 1. If this value becomes less than a specified threshold (T(n)), that node introduces131
itself as a CH. T(n)is calculated as follows:132
T(n)=p
1−p×rmod1
pif n∈G
0 otherwise
(1)
133
ris the current round number, pis the desired percentage of CHs, and Gis the set of134
nodes that were not selected as the CH in the last 1/prounds. When a node becomes a135
CH, it informs all nodes in the network of its new status by the means of broadcasting136
an announcement message. When each regular node receives this message, it joins a137
cluster corresponding to the CH with the maximum signal strength. Afterward, the138
CH assigns a time slot to each member during which the member nodes can send their139
sensed data. Up to this step, the setup phase is completed and clusters are formed for140
the current round.141
In the steady-state phase, the sensor nodes sense the environment and send obtained142
information to the related CH. CHs then aggregate the received data and send them143
to the BS via multi-hop communication. After a certain period of time, in which the144
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
P. Neamatollahi, M. Naghibzadeh
steady state is completed, another round begins and hence the setup phase restarts. In145
contrast to the prior protocols, there are several advantages in using M-LEACH:146
•M-LEACH is a fully distributed protocol that does not require global nor even147
local network knowledge to select CHs.148
•Ease of employment is another advantage of M-LEACH.149
Listed below are the disadvantages of M-LEACH:150
•M-LEACH assumes uniform energy consumption for CHs; thus, it does not con-151
sider node energy while electing CHs.152
•Although M-LEACH clustering process terminates in a constant time, it does153
not guarantee good CH distribution. In other words, the probabilistic method of154
selecting CHs may result in two or more CHs being in close range of each other.155
This means that CHs may not be evenly and properly dispersed.156
2.2 Hybrid energy-efficient distributed (HEED)157
HEED [12] is a distributed clustering protocol that differs from M-LEACH in the158
way CHs are selected. The combination of the following factors plays an important159
role in the process of HEED’s cluster formation: the residual energy of nodes and the160
cost of intra-cluster communication. The communication cost is defined according161
to whether the variable power level is allowed or not. If this power level is constant,162
the communication cost is proportional to the node degree (in the case when load163
distribution is more important) or to 1/node degree (when dense clusters are more164
preferable). In the status of a variable power level, AMRP is defined as the appropriate165
cost. AMRP is the averaged minimum power level needed by all nodes of a cluster166
for sending data to the corresponding CH (Eq. 2). For every CH node, u, with M167
members, AMRP(u)is computed as:168
AMRP (u)=M
i=1MinPwr (i)
M(2)
169
In this protocol, every regular node selects the least cost CH to join. The data is sent170
to the BS by a multi-hop approach, in contrast to that of direct communication. The171
benefits of HEED are as follows:172
•It is a distributed protocol that utilizes two parameters to form efficient clusters.173
•The probability of locating two CHs in each other’s proximity is negligible. This174
means that CHs are fairly distributed in the network.175
The drawback of HEED is early death of some nodes from hot spot problem. This176
issue considerably decreases the network lifetime.177
2.3 Distributed unequal clustering using fuzzy approach (DUCF)178
To overcome the drawback of the other protocols, the authors in [13] presented an179
unequal clustering algorithm, termed DUCF. In this method, both CH selection and180
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
Distributed unequal clustering algorithm in large-scale…
cluster size computation are performed by using fuzzy logic. It uses residual energy,181
node degree, and distance to BS as input variables to calculate the chance of being182
a CH (an output parameter). In addition, it considers node degree and distance to183
BS to compute the size of clusters (another output parameter). In contrast to [11,14]184
which tune the cluster radius to mitigate the hot spot issue, DUCF limits the number185
of member nodes assigned to a given CH by the size of that cluster.186
At the beginning of clustering process, every sensor node introduces itself as a187
tentative CH via broadcasting an advertisement message to the neighbor nodes. This188
message bears the chance and ID of the transmitter node. A node with the higher189
chance wins the competition and announces itself as a CH. Each non-CH node selects190
the nearest CH to join it. In this stage, each CH accepts the joining requests with191
respect to its cluster size. If the request of a non-CH node is rejected by a CH, then it192
attempts to associate with another CH. In the worst case, an orphan node introduces193
itself as a CH.194
As a notable advantage of DUCF, it uses three input variables for determining the195
chance and cluster size to prolong the network lifetime. Therefore, it chooses the good196
CHs and computes the cluster size more carefully than the other existing approaches.197
However, the disadvantages of DUCF are as follows:198
•As all sensor nodes are selected as tentative CH at the beginning of clustering pro-199
cess, the clustering energy overhead of DUCF is high. In addition, the converging200
time of the clustering process may be high because, in the worst case, a non-CH201
node must exchange many messages to join a non-full cluster.202
•Since the orphan nodes will be finally turned into CHs, there is a possibility that203
some CHs to be located within the cluster range of the others. This issue wastes204
the network energy resources.205
•Although the metric of distance to BS is more important than the metric of node206
degree, DUCF prefers the node degree over the distance to BS in its fuzzy rules.207
This leads to the generation of large clusters in the BS neighborhood which208
decreases the network longevity.209
3 Preliminaries210
In this section, the characteristics of the radio energy model are introduced and then211
system assumptions used in the proposed approach are described.212
3.1 Radio energy model213
In this paper, the simplified radio model proposed in [16,22] is employed to model
214
communication energy dissipation. If the distance between the sender and receiver is215
more than the threshold value d0(i.e., d>d0), then the d4power loss (the multipath216
fading channel model) is utilized. Otherwise, the d2power loss (the free space model)217
is used. Therefore, the amount of energy consumed for transmitting a k-bit message218
over distance dis represented as:219
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
P. Neamatollahi, M. Naghibzadeh
ETx (k,d)=kEelec +kfs d2d<d0
kEelec +kmd d4d≥d0
(3)220
Eelec represents the transmitter and the receiver circuit energy consumption per bit.221
Also, md and fs are the amplification energy dissipation factor for the multipath radio222
and free space models, respectively. On the other hand, to receive a k-bit data message,223
the amount of energy spent is given by:224
ERx (k)=kEelec (4)225
3.2 System assumptions226
This paper essentially addresses WSN applications with energy-constrained sensor227
nodes that are randomly distributed over a geographic region for continuously mon-228
itoring the area. In these multi-hop networks, the basic operation is to periodically229
gather data by the nodes and to transmit the collected information to the BS for further230
processing.231
Generally, in these homogenous networks, all sensor nodes have the same capabili-232
ties; therefore, each sensor node must always play one of two different roles. The first233
is the role of a regular node, in which the node monitors the environmental parameters234
and sends the data to the related CH. The second is the role of a CH, in which the node235
operates as a cluster coordinator.236
As a cluster is often formed from the nodes in one geographical region, locating237
cluster members in the close vicinity of each other causes redundancy in the data sent238
to the related CH. This issue motivates the employment of data aggregation techniques239
to reduce the communication load by the elimination of the redundancies. It is assumed240
that the sensed data are highly correlated; thus, the CH can always aggregate the data241
gathered from its members before sending them to the BS. Some other assumptions242
are:243
•After the deployment phase, the BS and all sensor nodes are semi-stationary.244
•To employ the UCF, nodes do not need to be aware of their location and, hence,245
are not equipped with a GPS antenna. Instead, one of the approximate distance246
calculation approaches, such as that of used in [22], may be applied to compute247
the approximate distance between the nodes.248
4 The approach249
In a sensor network, clustering partitions sensor nodes into subnetworks, including250
a node selected as a CH and some regular nodes as cluster members. Similar to251
many previous works in the literature [13,22,42,43], the operation of the UCF is split252
into fixed time periods, i.e., rounds. At the beginning of every round, the clustering253
algorithm is triggered. In other words, to balance the energy consumption of the254
network, in each round, the responsibility of being a CH is given to the most qualified255
node in each cluster. Furthermore, to mitigate the hot spot problem, the UCF generates256
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
Distributed unequal clustering algorithm in large-scale…
clusters with unequal sizes. Clusters in the BS vicinity are smaller in size because they257
should consume less energy for intra-cluster communications so as to conserve energy258
required for relaying inter-cluster traffic. On the other hand, the random deployment of259
sensor nodes in the sensing environment creates clusters with different node densities.260
The higher the density of nodes in a cluster, the higher the CH energy consumption261
of that cluster. Therefore, the local density of nodes may be used to determine cluster262
size. In the UCF, by employing this local information (distance to the BS and local263
density), each node computes its cluster radius using the FIS. In this section, the FIS is264
first described and then the proposed unequal clustering algorithm is illustrated with265
a pseudo code.266
4.1 Fuzzy inference system267
The CHs’ distribution over the network should be controlled such that less local density268
and a farther distance to the BS leads to a larger cluster radius and, on the contrary,269
greater local density with a closer distance to the BS results in smaller cluster range.270
Therefore, a suitable scope of cluster radius must be chosen for the network nodes.271
In this paper, an FIS handles the uncertainties of calculating the cluster radius for272
each node. For a particular WSN, the value of message transmission radius during273
clustering process, denoted as R0, is considered as a static parameter. Because the BS274
broadcasts R0to the entire network in advance, all the sensor nodes are informed of275
this value. Thus, any sensor node Sican compute its cluster radius (i.e., its competition276
range) by using the output of the FIS (which is a positive value smaller than one) as277
follows:278
Si·R←R0×(ω×FIS(Si·RD,Si·LD)+ρ)(5)279
in which ωand ρare constantly determined with respect to the application. Besides,280
for any sensor node Si,theRD and LD are input parameters of FIS which will be281
described in the following. The Eq. (5) limits the amount of the nodes’ cluster radius
282
between R0×ρand R0×(ω+ρ).283
Since it was assumed that the sensor nodes are stationary, the nodes’ maximum284
and minimum distance to the BS is also considered as a fixed parameter. A signal is285
broadcast to all sensor nodes by the BS after the network deployment phase. Therefore,286
the approximate distance to the BS can be calculated by each sensor node. In addition287
to facilitating the computation of a suitable power level for communicating with the288
BS, the approximate distance to the BS is applied for the computation of cluster radius289
of each node.290
Consequently, the Relative Distance (RD) to the BS can be determined by any291
sensor node Si(a sensor node with the ID number i) with respect to Si’s distance292
to the BS and also the minimum and maximum network nodes’ distance to the BS.293
Therefore, sensor node Sican calculate its RD as:294
RD =d(Si,BS)−dmin
dmax −dmin
(6)295
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
P. Neamatollahi, M. Naghibzadeh
Fig. 2 Fuzzy inference system
d(Si,BS)denotes the distance between Siand the BS, whereas dmax and dmi n indicate296
the maximum and minimum distance to BS, respectively. On the other hand, by sending297
a packet within the radius of R0and receiving the replies, each node can simply298
compute the number of its adjacent nodes. Therefore, node Sican calculate its Local299
Density (LD)as:300
LD =|Neighbor (Si)|
N×γ(7)
301
|Neighbor (Si)|denotes the number of node Si’s neighbors, whereas the network302
nodes’ set is represented as S, in which S={S1,S2,...,SN}and |S|=N. Besides,
303
the desired percentage of the number of clusters is shown by γ. As depicted in Fig. 2,304
the two variables, the RD and LD of a node, are applied as the inputs of the FIS and the305
cluster radius factor of a node, i.e., CR, is the only output parameter. A high CR value306
indicates that, if the node can be elected as a CH, it will handle a large size cluster, a307
cluster with wide range.308
The illustrated fuzzy set in Fig. 3a describes the RD input variable. Close, ade-309
quate, and far are linguistic variables for this fuzzy set. The close and far linguistic310
variables employ a trapezoidal membership function, whereas for the adequate lin-311
guistic variable, a triangular membership function is applied. LD is the other fuzzy312
input variable. In Fig. 3b, a fuzzy set is depicted that describes the LD input variable.313
Low,medium, and high are defined as the linguistic variables for this fuzzy set. For314
the CR output variable, the fuzzy set is shown in Fig. 3c. Very high,high,rather high,315
medium high,medium,medium low,rather low,low, and very low are the nine cluster316
radius linguistic variables. A trapezoidal membership function is used for very high317
and very low and triangular membership functions are employed for the remaining318
linguistic variables. In order to simplify the implementation and reduce computation319
costs, this work often takes up the triangular membership functions.320
Generally, fuzzy rules can be produced either from experimental or heuristics data.321
In the current paper, the heuristic data are taken as the generation method for the322
predefined fuzzy if-then mapping rules according to the following principle: A node323
with a longer distance to the BS and less local density has a larger cluster radius when324
it is chosen as a CH. Nine fuzzy mapping rules, illustrated in Table 1, are determined325
according to two fuzzy input variables and one output variable. A single crisp number326
is then generated through the defuzzification process so as to take advantage of the327
output fuzzy variable in practice. Similar to some proposed algorithms in the literature328
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
Distributed unequal clustering algorithm in large-scale…
Fig. 3 Fuzzy set for input and output variables. aRelative distance. bLocal density. cCluster radius
[25,44], the defuzzification in the FIS is performed by the Center of Area (CoA)329
method.330
Figure 4shows an example of FIS operation by given data. In this figure, the RD and331
LD input variables have been set by 0.75 and 0.25, respectively. In this situation, the332
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
P. Neamatollahi, M. Naghibzadeh
Tab l e 1 Fuzzy rules RD LD CR
1 Far Low Very high
2 Far Medium High
3 Far High Rather high
4 Adequate Low Medium high
5 Adequate Medium Medium
6 Adequate High Medium low
7 Close Low Rather low
8 Close Medium Low
9 Close High Very low
Fig. 4 An example for calculation of cluster radius in the proposed FIS
CR output variable is achieved by 0.711 which is an approximately large value. Some333
other examples of FIS operation are summarized in Table 2. Therefore, the shortening334
of the distance to the BS and the increase in local density lead to the reduction in the335
cluster radius. As a result, if a CH is close to the BS, it then consumes less energy336
for intra-cluster communications as compared to a CH which is far from the BS. To337
explain, the target CH benefits from a low number of members and a closer distance338
to cluster members. In this way, some more energy can be saved for the inter-cluster339
relay traffic.340
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
Distributed unequal clustering algorithm in large-scale…
Tab l e 2 The examples of FIS
operation No. RD LD CR
1 0.95 0.20 0.906
2 0.90 0.50 0.776
3 0.85 0.90 0.635
4 0.70 0.15 0.710
5 0.55 0.50 0.551
6 0.45 0.85 0.362
7 0.30 0.10 0.445
8 0.15 0.45 0.306
9 0.05 0.95 0.051
4.2 Unequal clustering algorithm341
This section describes a distributed self-organization and a balanced clustering algo-342
rithm for WSNs which can handle random distribution of sensor nodes in the sensing343
area. In contrast to [11,12,22], this competitive algorithm chooses CHs primarily344
based on the nodes’ residual energy and without using any random function. Using345
the pseudocode given in Algorithm 1, this section describes the clustering algorithm346
in detail.347
Because the algorithm is a distributed, the relation between the nodes is established348
by message exchanges. In addition, all variables are defined locally for every node Si,349
i=1,2,...,N. First, the variables applied in the algorithm are denoted as follows.350
The time required for selecting CHs by Algorithm 1 is represented by TCH-election.351
As shown in Lines 6 and 7, this value assures that Algorithm 1 is converged at O(1).352
Each sensor node Simaintains the two sets of SCH and Scandidate_CH which hold all353
deterministic and candidate CHs in the neighborhood of node Si, i.e., those whose354
broadcast signals node Sican hear, respectively. Also, α(0<α<1)and β(β>355
1)are the sensor parameters which are constantly determined by a specific WSN356
application.357
Before the beginning of a clustering operation (Lines 6–26), each node Sicomputes358
its chance of becoming a CH, denoted as Si.itCHprob, depending on its residual359
energy:360
Si.CHprob ←α×MAX(Si.RE/ME,pmin)(8)361
in which ME is the maximum energy of a fully charged battery and Si.RE represents362
node Si’s current residual energy. Note that, as pmin is a fixed number (0<pmin <1),363
the value of itCHprob is not permitted to fall below α×pmin. This threshold improves364
the time required for the clustering process convergence.365
As explained, for any application, R0is predefined as the message transmission366
radius during the clustering process. Nevertheless, the length of a node’s competition367
range, i.e., cluster radius, is computed according to Line 5. Therefore, different com-368
petition ranges are employed to generate unequal sizes of clusters. During the CH369
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
P. Neamatollahi, M. Naghibzadeh
selection process, every message is scattered in the radius of R0, which enables each370
node Sito hear all messages from adjacent nodes. On the other hand, node Sjis defined371
as a neighbor node of Si,ifSjis in Si’s cluster radius or Siis in Sj’s cluster radius.372
In other words, Siand Sjare neighbor nodes if the distance between them, denoted373
as dSi,Sj, is less than the maximum competition range of Siand Sj, denoted as374
MAX(Si.R,Sj.R).375
Three types of messages (i.e., Candidate_CH_msg,CH_msg, and Quit_Election376
_msg) are exchanged among the neighbor sensor nodes in the clustering algo-377
rithm. Upon receiving either a Candidate_CH_msg or CH_msg from a neighbor378
node, each receiver node Siinserts the information related to the transmitter neigh-379
bor node into its Scandidate_CH or SCH , as appropriate. In contrast, upon receiving380
Quit_Election_msg from a neighbor node, the receiver eliminates the sender informa-381
tion from its Scandidate_CH (refer to Lines 18–26).382
During the clustering process, a node can become a candidate CH, a CH, or a383
regular node joining a CH. When the status of a node turns into regular or CH, it384
gives up the competition, as its role in relation to other nodes has been determined. At385
first, due to a greater CHprob, a node with a higher level of energy has more chances386
of becoming a candidate_CH when competing with other nodes in its vicinity (see387
Lines 8 and 9). If a node becomes a candidate_CH, it broadcasts its new status,388
augmented with its identification number and cluster radius, to the nodes in its range389
via a Candidate_CH_msg. In the next iterations, if this node receives any CH_msg from390
a CH in its neighborhood, it becomes a regular node. Before quitting the competition,391
the node broadcasts a Quit_Election_msg in its range (see Lines 10–13). By receiving392
this message, each receiver node removes the sender node’s information from its393
Scandidate_CH. On the contrary, if a node has not heard from any CH in its proximity394
and its CHprob exceeds one, its status turns to a CH, thus winning the competition395
and broadcasting a CH_msg to neighbors in its cluster range (see Lines 14–16). Each396
receiver candidate_CH node, if any, immediately drops out of the competition and397
informs all nodes in its vicinity by broadcasting a Quit_Election_msg. In other words,398
at the end of a region’s node competition, if a node is selected as a CH, no other CH399
will exist in that node’s cluster radius. This assures a suitable CH distribution in the400
sensing area, which is required to achieve the scalability property. Note that, in contrast401
to previous works [11,12,36,39,45] that employ a random function for selecting nodes402
as candidate CHs, this algorithm considers the residual energy of nodes to elect the403
appropriate candidate CHs.404
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
Distributed unequal clustering algorithm in large-scale…
405
406
After selecting CHs, each regular node picks the closest CH from its SCH .The407
node informs this CH to join it by sending a Join_CH_msg. The CH provides a time-408
division multiple access (TDMA) schedule and sends this to its cluster nodes. After all409
cluster members receive the TDMA schedule, the setup phase is accomplished and the410
steady-state operation (data transmission) begins. During this operation, the sensed411
data are periodically transmitted to the CHs. The CHs send the aggregated data to the412
BS over multi-hop routes. 2
413
5 Performance evaluation414
To evaluate the performance of the proposed approach, UCF is compared via thorough415
simulation in MATLAB with three well-cited distributed clustering algorithms, M-416
LEACH [41], HEED [12], and DUCF [13]. M-LEACH employs a probability model417
for CH selection and so its energy efficiency is not maximized. On the other hand,418
HEED uses an iterative CH selection algorithm. It selects the tentative CHs based on an419
initial probability which depends on the nodes’ residual energy. The communication420
cost is also considered to elect final CHs from among the tentative CHs. During421
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
P. Neamatollahi, M. Naghibzadeh
Fig. 5 The deployment of
network nodes in the field
simulation, the HEED’s node degree communication cost is applied. DUCF utilizes a422
fuzzy system with three fuzzy descriptors to perform the CH election and to determine423
the size of clusters; those are the node residual energy, the local density, and distance424
to the BS. In the M-LEACH, HEED and DUCF algorithms, the CHs relay data to the425
BS via multi-hop routing.426
5.1 Setup427
In this section, the simulation setup of the experiments is described. In these experi-428
ments, a square network field of a 400 m×400 m area is considered. The BS is located429
at coordinate (200, 450). Also, 400 nodes are randomly distributed in the field, as illus-430
trated in Fig. 5. The effect of the BS location and the number of nodes will be studied431
in Sects. 5.5 and 5.6 in which the BS location and the number of nodes are varied.432
The initial energy of sensor nodes’ battery is randomly considered between 2 and 4J.433
The other values used in the simulation parameters are given in Table 3.Itisworth434
mentioning that the radio model parameters are the same as those in [11–13,22,25].435
5.2 Cluster formation436
Figure 6depicts clusters formed by the algorithms in the sensor deployment illustrated437
in Fig. 5. The dead nodes are shown by green hollow points, whereas each regular438
node, displayed by a blue dot, can be a member of a CH node represented by a red439
dot. A cluster is recognized as a set of regular nodes connected to a CH via a blue440
link. This figure shows the nodes’ situation with respect to the simulated algorithms441
for the round number 400 of a simulation run. As indicated in this figure, clusters in442
UCF have different sizes; the clusters closer to the BS with a higher local density are443
smaller in size, while the clusters distant from the BS with a lower local density have444
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
Distributed unequal clustering algorithm in large-scale…
Tab l e 3 Parameter settings Type Parameter Value
Network topology Number of nodes 400
Network coverage (0, 0)–(400, 400) m
Base station location At (200, 450) m
Radio model Idle power 13.5 mW
Sleep power 15 μW
Eelec 50 nJ/bit
fs 10 pJ/bit/m2
mp 0.0013 pJ/bit/m4
EDA 5 nJ/bit/signal
do87 m
Application Initial energy [2, 4] J
Round time 20 s
Round 5 frame
Control packet size 25 bytes
Data packet size 500 bytes
R075 m
pmin 10−4
α0.05
β2
ω1.6
ρ0.4
a larger size. This is because the CH selection algorithm employs fuzzy descriptors445
in the FIS, i.e., relative distance and local density. As illustrated, in some regions, the446
number of sensor nodes is great. Hence, the nodes’ local density in these regions is447
high which leads to the decrease in cluster radius when UCF is applied. For example,448
see the cluster created at coordinate (173, 262) of the UCF section in Fig. 6.The449
cluster illustrated has a smaller range than its right cluster [at coordinate (360, 250)]450
mainly due to its higher local density (observe Fig. 5), nonetheless both clusters are451
adjacent to the BS. However, the down clusters are larger than the other clusters as452
their distance from the BS is farther. It is also worth mentioning that all nodes in UCF453
are alive up to round 400 of this run, while in other algorithms, some dead nodes are454
observed. The clusters created in HEED suffer from the hot spot problem because455
they are organized in the same size clusters. In other words, unequal clustering is not456
explored by this algorithm. As the CH selection in M-LEACH is randomly performed,457
in each round, the number of CHs is unpredictably variable. Besides, the CHs may458
be located in each other’s cluster range, such as the down clusters in the M-LEACH459
section of Fig. 6. On the other hand, as displayed, because forwarding the sensed data460
to the BS is performed in a multi-hop fashion, the nodes close to the BS die sooner.461
Although DUCF benefits from unequal clusters, its cluster formation method is such462
that the node degree is given a higher priority than the distance to the BS (in fuzzy463
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
P. Neamatollahi, M. Naghibzadeh
Fig. 6 Clusters formed in round number 400 by using different clustering algorithms
rules) for tuning the size of clusters. This results in the creation of large clusters in464
the BS vicinity (see the DUCF section of Fig. 6). Therefore, a high number of nodes465
die due to the heavy traffic imposed on nodes near the BS. As a result, DUCF can not466
gain the best performance in some scenarios.467
5.3 Energy conservation468
Figure 7a shows the ratio of the sensors’ energy spent in the clustering process to469
the whole energy dissipated during network lifetime, i.e., clustering overhead, of sim-470
ulated clustering algorithms up to round number 250. As illustrated in this figure,471
M-LEACH has the most clustering overhead when compared to the others. This is due472
to employing the network-wide messages during the clustering process. The cluster-473
ing overhead of DUCF is also high. This is mainly because of broadcasting the CH474
candidate message by every sensor node at the beginning of the clustering process.475
In contrast, the UCF benefits a low-cost clustering algorithm because it decreases the476
number of messages exchanged.477
The total energy dissipated by using the simulated protocols up to round number 250478
isshowninFig.7b. This amount is high for M-LEACH which mainly is the result of the479
unfair distribution of the CHs in the field. Besides, this figure depicts the overcome of480
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
Distributed unequal clustering algorithm in large-scale…
Fig. 7 Energy conservation. a
Clustering overhead. bTotal
energy consumption
UCF on the other simulated protocols in terms of energy conservation. This is caused481
by a low-cost clustering algorithm and the CHs’ fair distribution in the field.482
5.4 Network lifetime483
Figure 8illustrates the number of live nodes in different algorithms during the simula-484
tion time. M-LEACH has the lowest network lifetime due to non-fair CH distribution485
and high clustering overhead. DUCF performs better than LEACH and HEED because486
of unequal clustering. However, it still suffers from the high clustering overhead and487
the higher priority of node degree in comparison to the node distance to the BS for488
adjusting the cluster size. In contrast, there are some main features of UCF which489
result in a longer network lifetime: selection of CHs based only on the nodes’ resid-490
ual energy without using a random function, a lower clustering overhead, a better491
CH distribution in the area, and a balance of the transmission load by using unequal492
clustering.493
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
P. Neamatollahi, M. Naghibzadeh
Fig. 8 The number of live
nodes against the rounds
Fig. 9 The variations of BS
location
5.5 BS location494
Figure 9shows the network lifetime in BS location variation by using HNA metric495
(the round number in which half of the nodes remain alive). Although FND (the round496
number in which the first node dies) is an acceptable metric for evaluating the network497
lifetime in sparsely WSNs, in densely deployed networks, a single node death is not a498
significant issue. Therefore, HNA metric is often selected in these cases. The horizontal499
axis of Fig. 9depicts the distance from the center of the sensing region. The number of500
zero means that BS is at the center of sensing region [at coordinate (200, 200)], while501
the number of 300 is related to the BS location of (200, 500). As shown in this figure,502
with the variation in BS location, UCF can still balance the load on the sensor nodes503
and achieve the best network lifetime. However, by increasing the BS distance from504
the nodes’ deployment region, the energy consumed to deliver the sensed data to the505
BS grows which leads to the decrease in network lifetime of all simulated protocols.506
5.6 Scalability507
In this subsection, which focuses on scalability, Fig. 10 illustrates the superiority of508
UCF over the other simulated distributed clustering algorithms of M-LEACH, HEED,509
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
Distributed unequal clustering algorithm in large-scale…
Fig. 10 The scalability
and DUCF with respect to HNA. In the clustering phase, when the number of nodes510
grows, the number of regular sensor nodes belonging to each cluster also increases.511
In these situations, as DUCF limits the number of cluster members, the number of512
created orphan nodes increases. The status of these orphan nodes is usually turned into513
single CHs which imposes a higher traffic to the network. In general, UCF performs514
better than the other algorithms because it benefits from a low-cost unequal clustering515
algorithm which balances the load on the CHs and mitigates the hot spot problem.516
Therefore, UCF is scalable in terms of the number of nodes.517
6 Conclusion518
Because energy constraint is a significant challenge when designing WSNs, many stud-519
ies have been performed to obtain energy-efficient algorithms. Nevertheless, most of520
the previous clustering approaches have not considered load balancing in the network.521
Current approaches often select final CHs from among randomly selected candidate522
CHs. This random selection may result in inappropriate CH election with respect to523
energy conservation and CH distribution in the field. The proposed algorithm in the524
current paper is designed for WSNs with stationary sensor nodes randomly distributed525
in the field. The main objective of the proposed algorithm is to prolong the network526
lifetime by evenly distributing the workload and, hence, avoiding hot spot problem527
by the construction of unequal clusters. To attain this, the proposed algorithm mostly528
focuses on choosing proper CHs from available sensor nodes and adjusting the clus-529
ter radius to mitigate the hot spot problem with the help of fuzzy logic based on the530
node’s relative distance to the BS and local density. UCF selects the CHs from among531
the CH candidates by only considering the residual energy of sensor nodes without532
using the random function. To take into account the hot spot issue, CHs farther from533
the BS with less local density have a larger cluster radius than those closer to the BS534
with more local density. Therefore, small-size clusters can save some energy for inter-535
cluster communications. In the experiments, the UCF achieved 30% improvement in536
network lifetime, more than 56% reduction in clustering overhead, and 12% progress537
in total energy conservation when compared to other simulated protocols. Therefore,538
the results of the simulation show that UCF is more efficient than other well-known539
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
P. Neamatollahi, M. Naghibzadeh
distributed algorithms (DUCF, M-LEACH, and HEED) in terms of load balancing,540
network lifetime, and energy efficiency. In addition, based on the simulation results,541
UCF is a scalable protocol which its superiority does not depend on the BS location542
or the number of sensor nodes. The presentation of an unequal clustering algorithm543
for mobile sensor nodes is left as a future work.544
Acknowledgements The authors would like to thank Research and Technology Affairs, Mashhad Branch,545
Islamic Azad University, for supporting this research under Grant 91368.400.7.546
References547
1. Akyildiz IF, Vuran MC (2010) Wireless sensor networks. Wiley, Hoboken548
2. Mottola L, Pietro G (2011) Picco, programming wireless sensor networks. ACM Comput Surv 43:1–51.549
https://doi.org/10.1145/1922649.1922656550
3. Rostami AS, Badkoobe M, Mohanna F, Keshavarz H, Hosseinabadi AAR, Sangaiah AK (2018) Survey551
on clustering in heterogeneous and homogeneous wireless sensor networks. J Supercomput 74:277–552
323. https://doi.org/10.1007/s11227-017- 2128-1553
4. Anastasi G, Conti M, Di Francesco M, Passarella A (2009) Energy conservation in wireless sensor554
networks: a survey. Ad Hoc Netw 7:537–568. https://doi.org/10.1016/j.adhoc.2008.06.003555
5. Yick J, Mukherjee B, Ghosal D (2008) Wireless sensor network survey. Comput Netw 52:2292–2330.556
https://doi.org/10.1016/j.comnet.2008.04.002557
6. Wang F, Liu J (2011) Networked wireless sensor data collection: issues, challenges, and approaches.558
IEEE Commun Surv Tutor 13:673–687. https://doi.org/10.1109/SURV.2011.060710.00066559
7. Sudevalayam S, Kulkarni P (2011) Energy harvesting sensor nodes: survey and implications. IEEE560
Commun Surv Tutor 13:443–461. https://doi.org/10.1109/SURV.2011.060710.00094561
8. Afsar MM, Tayarani-N M-H (2014) Clustering in sensor networks: a literature survey. J Netw Comput562
Appl 46:198–226. https://doi.org/10.1016/j.jnca.2014.09.005563
9. Liu X (2015) Atypical hierarchical routing protocols for wireless sensor networks: a review. IEEE564
Sens J 15:5372–5383. https://doi.org/10.1109/JSEN.2015.2445796565
10. Masdari M, Bazarchi SM, Bidaki M (2013) Analysis of secure LEACH-based clustering protocols in566
wireless sensor networks. J Netw Comput Appl 36:1243–1260. https://doi.org/10.1016/j.jnca.2012.567
12.017568
11. Bagci H, Yazici A (2013) An energy aware fuzzy approach to unequal clustering in wireless sensor569
networks. Appl Soft Comput 13:1741–1749. https://doi.org/10.1016/j.asoc.2012.12.029570
12. Younis O, Fahmy S (2004) HEED: a hybrid, energy-efficient, distributed clustering approach for ad571
hoc sensor networks. IEEE Trans Mob Comput 3:366–379. https://doi.org/10.1109/TMC.2004.41572
13. Baranidharan B, Santhi B (2016) DUCF: distributed load balancing unequal clustering in wireless573
sensor networks using fuzzy approach. Appl Soft Comput 40:495–506. https://doi.org/10.1016/ j.asoc.574
2015.11.044575
14. Sert SA, Bagci H, Yazici A (2015) MOFCA: multi-objective fuzzy clustering algorithm for wireless576
sensor networks. Appl Soft Comput 30:151–165. https://doi.org/10.1016/j.asoc.2014.11.063577
15. Neamatollahi P, Taheri H, Naghibzadeh M, Yaghmaee M-H (2011) A hybrid clustering approach578
for prolonging lifetime in wireless sensor networks. In: 2011 International Symposium on Computer579
Networks and Distributed Systems (CNDS). https://doi.org/10.1109/CNDS.2011.5764566580
16. Neamatollahi P, Naghibzadeh M, Abrishami S, Yaghmaee MH (2017) Distributed clustering-task581
scheduling for wireless sensor networks using dynamic hyper round policy. IEEE Trans Mob Comput.582
https://doi.org/10.1109/TMC.2017.2710050583
17. Zhao M, Yang Y, Wang C (2015) Mobile data gathering with load balanced clustering and dual data584
uploading in wireless sensor networks. IEEE Trans Mob Comput 14:770–785. https://doi.org/10.1109/585
TMC.2014.2338315586
18. Xu Z, Chen L, Chen C, Guan X (2016) Joint clustering and routing design for reliable and efficient587
data collection in large-scale wireless sensor networks. IEEE Internet Things J 3:520–532. https://doi.588
org/10.1109/JIOT.2015.2482363589
19. Kang SH, Nguyen T (2012) Distance based thresholds for cluster head selection in wireless sensor590
networks. IEEE Commun Lett 16:1396–1399. https://doi.org/10.1109/LCOMM.2012.073112.120450591
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
Distributed unequal clustering algorithm in large-scale…
20. Pal V, Singh G, Yadav RP (2015) Balanced cluster size solution to extend lifetime of wireless sensor
592
networks. IEEE Internet Things J 2:399–401. https://doi.org/10.1109/JIOT.2015.2408115593
21. Hoang DC, Yadav P, Kumar R, Panda SK (2014) Real-time implementation of a harmony search594
algorithm-based clustering protocol for energy-efficient wireless sensor networks. IEEE Trans Ind595
Inform 10:774–783. https://doi.org/10.1109/TII.2013.2273739596
22. Heinzelman WB, Chandrakasan AP, Balakrishnan H (2002) An application-specific protocol archi-597
tecture for wireless microsensor networks. IEEE Trans Wirel Commun 1:660–670. https://doi.org/10.598
1109/TWC.2002.804190599
23. Du T, Qu S, Liu F, Wang Q (2015) An energy efficiency semi-static routing algorithm for WSNs based600
on HAC clustering method. Inf Fusion 21:18–29. https://doi.org/10.1016/j.inffus.2013.05.001601
24. Nayak P, Devulapalli A (2016) A fuzzy logic-based clustering algorithm for WSN to extendthe network602
lifetime. IEEE Sens J 16:137–144. https://doi.org/10.1109/JSEN.2015.2472970603
25. Taheri H, Neamatollahi P, Younis OM, Naghibzadeh S, Yaghmaee MH (2012) An energy-aware dis-604
tributed clustering protocol in wireless sensor networks using fuzzy logic. Ad Hoc Netw 10:1469–1481.605
https://doi.org/10.1016/j.adhoc.2012.04.004606
26. Senouci MR, Mellouk A, Senouci H, Aissani A (2012) Performance evaluation of network lifetime607
spatial-temporal distribution for WSN routing protocols. J Netw Comput Appl 35:1317–1328. https://608
doi.org/10.1016/j.jnca.2012.01.016609
27. Chanak P, Banerjee I, Sherratt RS (2017) Energy-aware distributed routing algorithm to tolerate net-610
work failure in wireless sensor networks. Ad Hoc Netw 56:158–172. https://doi.org/10.1016/j.adhoc.611
2016.12.006612
28. Cenedese A, Luvisotto M, Michieletto G (2017) Distributed clustering strategies in industrial wireless613
sensor networks. IEEE Trans Ind Inform 13:228–237. https://doi.org/10.1109/TII.2016.2628409614
29. Neamatollahi P, Abrishami S, Naghibzadeh M, Yaghmaee MH, Younis O (2017) Hierarchical615
clustering-task scheduling policy in cluster-based wireless sensor networks. IEEE Trans Ind Inform.616
https://doi.org/10.1109/TII.2017.2757606617
30. Neamatollahi P, Naghibzadeh M, Abrishami S (2017) Fuzzy-based clustering-task scheduling for618
lifetime enhancement in wireless sensor networks. IEEE Sens J 17:6837–6844. https://doi.org/10.619
1109/JSEN.2017.2749250620
31. Neamatollahi P, Taheri H, Naghibzadeh M, Abrishami S (2014) A distributed clustering scheme for621
wireless sensor networks. In: 2014 6th Conference on Information and Knowledge Technology IKT622
2014. https://doi.org/10.1109/IKT.2014.7030326623
32. Toloueiashtian M, Motameni H (2017) A new clustering approach in wireless sensor networks using624
fuzzy system. J Supercomput. https://doi.org/10.1007/s11227-017- 2153-0625
33. Randhawa S, Jain S (2018) Energy-efficient load balancing scheme for two-tier communication in626
wireless sensor networks. J Supercomput 74:386–416. https://doi.org/10.1007/s11227-017- 2134-3627
34. Wang J, Cao J, Ji S, Park JH (2017) Energy-efficient cluster-based dynamic routes adjustment approach628
for wireless sensor networks with mobile sinks. J Supercomput 73:3277–3290. https://doi.org/10.1007/629
s11227-016- 1947-9630
35. Zahedi ZM, Akbari R, Shokouhifar M, Safaei F, Jalali A (2016) Swarm intelligence based fuzzy routing631
protocol for clustered wireless sensor networks. Expert Syst Appl 55:313–328 3
632
36. Chen G, Li C, Ye M, Wu J (2009) An unequal cluster-based routing protocol in wireless sensor633
networks. Wirel Netw 15:193–207634
37. Liao Y, Qi H, Li W (2013) Load-balanced clustering algorithm with distributed self-organization for635
wireless sensor networks. IEEE Sens J 13:1498–1506. https://doi.org/10.1109/JSEN.2012.2227704636
38. Naghibzadeh M, Taheri H, Neamatollahi P (2014) Fuzzy-based clustering solution for hot spot problem637
in wireless sensor networks. In: 2014 7th International Symposium on Telecommunications. IEEE, pp638
729–734. https://doi.org/10.1109/ISTEL.2014.7000798639
39. Mao S, Zhao C, Zhou Z, Ye Y (2012) An improved fuzzy unequal clustering algorithm for wireless640
sensor network. Mob Netw Appl. https://doi.org/10.1007/s11036- 012-0356-4641
40. Li C, Ye M, Chen G, Wu J, Chengfa L, Mao Y, Guihai C, Jie W (2005) An energy-efficient unequal642
clustering mechanism for wireless sensor networks. In: IEEE International Conference on Mobile643
Ad Hoc and Sensor System Conference 2005. IEEE, p 604. https://doi.org/10.1109/MAHSS.2005.644
1542849645
41. Xiangning F, Yulin S (2007) Improvement on LEACH protocol of wireless sensor network. In: Inter-646
national Conference on Sensor Technologies and Applications (SENSORCOMM 2007). IEEE 2007,647
pp 260–264. https://doi.org/10.1109/SENSORCOMM.2007.4394931648
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof

uncorrected proof
P. Neamatollahi, M. Naghibzadeh
42. Elhabyan RSY, Yagoub MCE (2015) Two-tier particle swarm optimization protocol for clustering and
649
routing in wireless sensor network. J Netw Comput Appl 52:116–128. https://doi.org/10.1016/j.jnca.650
2015.02.004651
43. Taheri H, Neamatollahi P, Yaghmaee MH, Naghibzadeh M (2011) A local cluster head election algo-652
rithm in wireless sensor networks. In: 2011 CSI International Symposium on Computer Science and653
Software Engineering, pp 38–43. https://doi.org/10.1109/CSICSSE.2011.5963987654
44. Neamatollahi P, Taheri H, Toreini E, Naghibzadeh M, Yaghmaee MH (2011) A novel fuzzy metric to655
evaluate clusters for prolonging lifetime in wireless sensor networks. In: 2011 International Symposium656
on Artificial Intelligence and Signal Processing. IEEE, pp 118–123. https://doi.org/10.1109/AISP.657
2011.5960995658
45. Lee J-S, Cheng W-L (2012) Fuzzy-logic-based clustering approach for wireless sensor networks using659
energy predication. IEEE Sens J 12:2891–2897. https://doi.org/10.1109/JSEN.2012.2204737660
123
Journal: 11227 Article No.: 2261 TYPESET DISK LE CP Disp.:2018/1/25 Pages: 24 Layout: Small-X
Author Proof