Content uploaded by Osama Moh'd Alia
Author content
All content in this area was uploaded by Osama Moh'd Alia on Apr 11, 2018
Content may be subject to copyright.
Ann. Telecommun.
https://doi.org/10.1007/s12243-017-0611-6
A dynamic harmony search-based fuzzy clustering protocol
for energy-efficient wireless sensor networks
Osama Moh’d Alia1
Received: 15 June 2017 / Accepted: 1 October 2017
© Institut Mines-T´
el´
ecom and Springer-Verlag France SAS 2017
Abstract In the development of cluster-based energy-
efficient protocols for wireless sensor networks (WSNs),
a particularly challenging problem is the dynamic organi-
zation of sensors into a wireless communication network
and the routing of sensed information from the field sen-
sors to a remote base station (BS) in a manner that prolongs
the lifetime of WSNs. This paper presents a new energy-
efficient clustering protocol for WSNs, which can minimize
total network energy dissipation while maximizing network
lifetime. The protocol is divided into two parts. The first
deals with constructing an infrastructure for the given WSN.
A newly developed algorithm, based on a harmony search
(HS), automatically determines the optimal number of clus-
ters and allocates sensors into these clusters. This algorithm
also eliminates the need to set the number of clusters a
priori. The second part is concerned with the process of
sending sensed data from nodes to their cluster head and
then to the BS. A decentralized fuzzy clustering algorithm
is proposed, where the selection of cluster heads in each
round is locally made in each cluster during the network
lifetime. Simulation results demonstrate that the proposed
protocol can achieve an optimal number of clusters, prolong
the network lifetime and increase the data delivery at the
BS, when compared to other well-known clustering-based
routing protocols.
Osama Moh’d Alia
sm alia@yahoo.com; oalia@ut.edu.sa
1Department of Computer Science, Faculty of Computers and
Information Technology, University of Tabuk, PO box 741,
Tabuk 71491, Saudi Arabia
Keywords Wireless sensor networks ·Energy-efficient
routing protocols ·Harmony search algorithm ·Fuzzy
clustering
1 Introduction
Wireless sensor networks (WSNs) have potential appli-
cations in civilian and military domains, including envi-
ronmental monitoring, surveillance, healthcare, intelligent
building control, traffic control and object tracking [1–3].
WSNs consist of a large number of autonomous sensor
nodes equipped with sensing capabilities, wireless com-
munication interfaces and limited processing and energy
resources.
One or more powerful base stations (BSs) act as the final
destination of the data. WSNs are used for distributed and
cooperative sensing of physical phenomena and events of
interest [4].
Energy consumption needed to pass sensory data to the
final destination for each node is a major concern when
designing WSN routing protocols [5–7]. Economical usage
of sensor nodes is paramount in order to prolong the oper-
ational lifetime of the network. Moreover, the sensor nodes
are equipped with limited energy sources (lightweight bat-
tery); therefore, replacing these batteries may be inappli-
cable or impossible [8–10]. Thus, optimal battery energy
usage can prolong use and improve performance efficiency
of WSNs. Designing routing algorithms, which can con-
sume less energy, while maintaining the efficiency and
robustness of WSNs, is hence highly desirable.
During the last decade, WSN researchers have focused on
solving the above-mentioned problem by proposing power
management protocols alongside the evolution of hardware-
based solutions [11–13]. Cluster-based techniques, which
Ann. Telecommun.
are some of the most innovative techniques in this aspect,
have been proven to be scalable and efficient [13,14].
In such techniques, all sensors are organized into clusters.
Each cluster center, called the cluster head (CH), performs
specific tasks to acquire data from sensors within its own
cluster, carry out data aggregation and transmit fused data
directly to the BS. This allows most nodes to transmit across
small distances and reduce the amount of data sent through
the network to save battery energy. Furthermore, the pro-
cess of alternating the role of being a CH between cluster
member nodes further reduces energy consumption, since
non-CH member nodes can go into sleep mode for a longer
period of time.
In the last few years, a relatively large number of clus-
tering routing protocols have been developed for WSNs,
such as [14–21]; for more references, see [13,22]. One pio-
neering approach was presented by [23,24] and involves
a low-energy adaptive clustering hierarchy (LEACH). A
LEACH essentially selects sensor nodes as CHs by rotation,
such that the high-energy dissipation in communicating with
the BS is spread to all sensor nodes in the network. This
distribution enables the LEACH to be superior to other pro-
tocols since: (1) each node equally shares the load imposed
upon CHs to some extent; and (2) non-CH member nodes
can manage their communication interfaces in compliance
with their allocated time slots, hence avoiding excessive
energy dissipation [13]. However, a LEACH also has a few
drawbacks, as follows: (1) to some extent, it is not applica-
ble to large region networks, since it performs the single-hop
inter-cluster routing method;(2) it cannot ensure real load
balancing when dealing with sensor nodes with different
amounts of initial energy;(3) it is hard for predetermined
CHs to be uniformly distributed throughout the network
since CH election is performed in terms of probabilities;
and (4) the idea of dynamic clustering brings with it extra
overheads [13].
These disadvantages have inspired other researchers to
improve subsequent clustering routing protocols. Among
the popular ones are hybrid energy-efficient distributed
clustering [25], distributed energy-efficient clustering
(DEEC) [26], power-efficient gathering in sensor informa-
tion systems [27], the threshold-sensitive energy-efficient
sensor network [28], the stable election protocol [29], the
neuro-fuzzy energy-aware clustering scheme [30], distance-
based and low-energy adaptive clustering [31], the energy-
efficient distributed clustering algorithm based on the fuzzy
approach with nonuniform distribution [32], the distributed
clustering protocol using voting and priority [33], priority-
based congestion control dynamic clustering [34] and the
distributed energy-efficient clustering protocol [35]. Other
protocols are discussed in the work of [13,22].
Despite improvements, many of these protocols ignore
the problem of determining the optimal number of clusters
for the given WSN. Naturally, clustering algorithms require
the number of clusters to be determined by the network
designer before the clustering process begins [36–38].
Ensuring the longevity of the WSN is a critical factor,
given that, if there are fewer clusters, non-CH nodes prob-
ably consume too much energy when transmitting sensory
data to their CHs because most clusters will have to be
large. Conversely, if there are too many clusters, the energy
consumed by non-CH nodes in order to transmit data to
their CHs decreases. At the same time, there is a possi-
bility that more energy will be consumed by CHs, where
CHs are required to fuse the sensory data and transmit them
over a large distance to the BS [39]. Therefore, in order to
minimize energy consumption and prolong the WSN’s life-
time, the clustering scheme must be able to find the optimal
number of clusters before the clustering process begins.
To create an efficient and energy-aware routing protocol
in WSNs, while avoiding the above-mentioned weaknesses,
we seek mechanisms from nature and adapt them to suit
the challenges of WSNs. Since clustering problems can be
classified as optimization problems[40–42], meta-heuristic
algorithms are widely believed to be among the best types of
clustering algorithms. Such algorithms are able to solve NP-
hard problems with satisfactory near-optimal solutions and
significantly less computational time than exact algorithms
[38]. The HS algorithm is a relatively new meta-heuristic
algorithm developed by Geem et al. [43] to solve opti-
mization problems. Ever since its inception, the HS has
encouraged many researchers to develop HS-based applica-
tions for many optimization problems [14,44,45]. In a HS,
a candidate solution to an optimization problem corresponds
to a musical harmony composed of notes played by a group
of musicians. Each decision variable in a candidate solution
is quite similar to when a musician plays a note. In this con-
text, the value range of each decision variable is equivalent
to the pitch range of each note.
The HS approach is a successful meta-heuristic algo-
rithm, which can explore the search space of the given
data set (i.e., WSN) in a parallel optimization environ-
ment, where each solution (harmony) vector is generated by
intelligently exploring and exploiting a search space.
This paper proposes a new energy-efficient clustering
protocol for WSNs,1can minimize total network energy
dissipation, while maximizing network lifetime. The pro-
tocol is divided into two parts. In the first part, a newly
Dynamic Clustering algorithm using the HS for WSNs,
termed DCHS-WSN, is proposed. This algorithm automat-
ically determines the number of clusters that can result in
minimum total network energy dissipation. This algorithm
1The initial version of the proposed protocol, which was presented in
[46]
Ann. Telecommun.
builds the appropriate infrastructure for the given WSN
by allocating sensor nodes to their most appropriate clus-
ters. Note that this part is only performed once at the
beginning of the protocol at a BS, with the resultant infras-
tructure remaining the same without any changes during the
network’s lifetime.
The second part of the proposed protocol deals with send-
ing sensed data from nodes to their CHs, and then to the
BS. It runs into rounds, with each round subdivided into
a CH selection phase and a steady-state phase. In the CH
selection phase, the role of the CH is rotated between the
cluster nodes to evenly distribute the energy load among the
sensor nodes in each cluster. The selection process for the
new CHs is performed locally in each cluster, where a new
multi-criteria objective function is proposed to enhance the
quality of the selected CHs. In the steady-state phase, each
sensor node senses the targeted environment, then transmits
the sensed data to the cluster head to which it belongs. Then,
the cluster heads, in turn, aggregate and send the sensed data
to the BS.
To summarize the contribution of this research, one can
say that the proposed protocol should minimize total net-
work energy dissipation while maximizing the network’s
lifetime. The following key factors can highlight the pref-
erences of the proposed protocol over other competitor
protocols, such as a LEACH [23,24]. First, the optimal
number of clusters can be determined automatically with-
out any intervention from a network designer; second, the
sensor nodes can be evenly distributed to most proper clus-
ters; and third, the CH selection is selected locally in each
cluster, based on the residual energy and the geographical
location of each candidate sensor with regard to the BS and
every other cluster’s member.
This paper is arranged as follows. Section 2sets out
the preliminaries of the network and the radio model being
explored. Section 3provides an overview of the HS algo-
rithm. A detailed description of the proposed protocol is
giveninSection4. A simulation study is presented in
Section 5. We conclude our findings in Section 7.
2 Preliminaries
This section presents the assumptions and radio dissipation
model of the network under consideration.
2.1 Assumptions
1. The BS is located far from the sensor nodes and immo-
bile.
2. All nodes in the network are homogeneous and energy-
constrained.
3. The channel propagation is symmetric.
4. Nodes have location information with respective energy
levels.
5. Nodes have no mobility.
2.2 Radio energy dissipation model
In the radio energy model used in this work, as presented
in [23,24], the transmitter dissipates energy to run the
radio electronics and the power amplifier, while the receiver
dissipates energy to run the radio electronics. The energy
consumption for transmitting a message of bbit over a
distance dis:
ETx =Eelec ×b+Efs ×b×d2,d≤d0(1)
ETx =Eelec ×b+Emp ×b×d4,d>d
0(2)
For receiving this message:
ERx =Eelec ×b(3)
where Eelec is the energy expended to operate the
transceiver circuit, and Efs and Emp are the energy expen-
ditures of transmitting 1 bit of data to achieve an acceptable
bit error rate, which depends on the transmission distance
in the case of the free space model and multipath fading
model, respectively [23,24]. If the transmission distance is
less than the threshold of d0, the free space model is applied;
otherwise, we use the multipath model. The threshold d0is
calculated by equating the two expressions in Eqs. 1and 2
at d=d0as follows:
d0=Efs/Emp (4)
Data aggregation, which is performed by the CH to
reduce the total amount of sent data is calculated as Eda =5
nJ/bit/message. This is based on the assumption that the
data from neighboring sensors will often be highly corre-
lated; therefore, the overall data collected by a cluster of n
nodes, where each node collects bbits of data, can be com-
pressed to bbits, regardless of the number of nodes in that
cluster [47].
A new parameter is proposed in this paper to represent
the energy consumption of CH when the selection of a new
CH for the next round takes place. We assume that the CH
selection energy expenditure is set as ECH−Elec =5nJ ×
No.Above.Ave,whereN o.Above.Ave represents the num-
ber of candidate CHs within a cluster, which are above the
average energy of alive nodes.
3 Harmony search algorithm
As mentioned, the HS algorithm mimics the improvisa-
tion process of musicians in an intelligent way. Just as
Ann. Telecommun.
musical harmony is improved over time, the solution vec-
tor is improved, iteration by iteration. In general, the HS
algorithm has five steps [48], which can be described as
follows:
Step 1: Initialize the HS and optimization problem
parameters
The optimization problem is treated as a minimization (or
maximization) problem. Minimize (or maximize) f(a)sub-
ject to ai∈Ai,i=1,2,...,N,wheref(a) is the
objective function, (a) is the set of decision variables (ai),
Ais the set of possible ranges of any decision variable,
Ai=ai(1), ai(2),...a
i(M),(M) is the number of possible
ranges for each decision variable, and (N) is the num-
ber of decision variables. Then, the parameters of the HS
are initialized. These parameters are the harmony mem-
ory size (HMS), the harmony memory consideration rate
(HMCR), the pitch adjustment rate (PAR) and the number
of improvisations (NI).
Step 2: Initialize the harmony memory
The harmony memory (HM) is a matrix of solutions with
aHMS,asshowninEq.5. In this step, the solutions are
randomly constructed and rearranged in reverse order to the
HM, based on the objective function values.
HM =⎛
⎜
⎜
⎜
⎝
a1
1a1
2··· a1
N
a2
1a2
2··· a2
N
.
.
..
.
.··· .
.
.
aHMS
1aHMS
2··· aHMS
N
fa1
fa2
.
.
.
faHMS
⎞
⎟
⎟
⎟
⎠
(5)
Step 3: Improvise a new harmony
In this step, the HS generates (improvises) a new harmony
vector, aNew =aNew
1,a
New
2,a
New
3,...,a
New
Nbased on
three operators, which are the memory consideration, ran-
dom consideration and pitch adjustment. In the memory
consideration, the values of the new harmony vector are
randomly inherited from the historical values stored in the
HM with the probability of HMCR. The value of the deci-
sion variable aNew
1is chosen from a1
1,a
2
1,a
3
1,...,a
HMS
1,
which is stored in the HM. The next decision variable,
aNew
2, is chosen from a1
2,a
2
2,a
3
2,...,a
HMS
2, while the other
variables, aNew
3,a
New
4,..., are chosen consecutively in the
same manner with the probability of HMCR ∈[0,1].The
HMCR parameter is a probability of selecting one value
from the decision variable, aNew
i, based on the historical
values stored in the HM. The usage of the HM is analo-
gous to the step where the musician uses his or her memory
to ‘generate’ an excellent tune. This important step ensures
that good harmonies are considered as the elements of the
new harmony vectors.
All of this is a cumulative process. Other values that are
not chosen according to memory considerations are chosen
according to their possible range by random consideration
with a probability of 1-HMCR. This step is called random-
ization, which increases the diversity of the solutions and
drives the system to further explore various diverse solu-
tions, thereby attaining global optimality. Furthermore, the
new vector components that are selected out of the memory
consideration operator are examined to be pitch-adjusted
with a probability of the PAR ∈[0,1],asinEq.6.
(aNew
i)=(aNew
i)±rand() ∗bw (6)
Here, bw is an arbitrary distance bandwidth used to
improve the performance of the HS. The PAR parameter
simulates the music by ‘changing the frequency’, which
means generating a slightly different value for the new har-
mony vector component. The PAR explores more solutions
in the search space for that purpose.
Step 4: Update the HM
The generated harmony vector, aNew =
aNew
1,a
New
2,a
New
3,...,a
New
N, replaces the worst har-
mony in the HM, but only if its fitness value (measured in
terms of the objective function) is better than that of the
worst harmony.
Step 5: Check the stopping criterion
Repeat Steps 3 and 4 until the maximum NI is reached. x
4 The proposed protocol
As mentioned earlier, the proposed protocol is divided into
two parts. The first part builds the appropriate network
infrastructure. The description of this part is given in the
following section. The second part of the proposed protocol
deals with sending sensed data from sensor nodes to their CHs
and then to the BS. This is given in the next section as well.
The flowchart of the proposed protocol is given in Fig. 1.
4.1 Building network infrastructure (DCHS-WSN)
The proposed DCHS-WSN algorithm is the first part of our
protocol to construct the infrastructure of the WSN, where
the optimal number of clusters, as well as the membership of
each node to these clusters, is determined. Suppose a WSN
consists of Nrandomly distributed nodes over an area of
M×Mmeters. These sensor nodes send a short message
called an advertisement message to the BS with information
Ann. Telecommun.
Fig. 1 An overview of the proposed protocol
about their geographical location. Based on the information
received from the sensor nodes, the DCHS-WSN algo-
rithm automatically determines the appropriate number of
clusters at the BS, then allocates sensor nodes into these
clusters.
This infrastructure becomes permanent during the life-
time of the network, where each node retains its location
from one cluster to another, while the role of the CH is
rotated between cluster nodes to evenly distribute the energy
load among sensor nodes in each cluster. The selection pro-
cess of the new CHs for the upcoming rounds is performed
locally in each cluster and independently from the BS, as
described in the CH selection phase. A full description of
the DCHS-WSN algorithm is given in the following section.
4.1.1 DCHS-WSN algorithm
This section describes how the DCHS-WSN algorithm is
designed and applied to optimally cluster the WSNs. As
with the HS, DCHS-WSN has five steps. These steps are
described as follows:
4.1.2 Step 1: Initialize the DCHS-WSN parameters
The DCHS-WSN has the same parameters as the HS: HMS,
HMCR, PAR and NI.
4.1.3 Step 2: Initialization of the HM
Each HM vector encodes the CHs of the given WSN. How-
ever, since the number of these CHs is unknown for the
given WSN, the range of the possible number of clusters that
the given WSN may possess is tested. Therefore, the length
of each HM vector can vary according to the randomly gen-
erated number of clusters for each vector. To initialize the
HM with feasible solutions, which is the index of a sensor
node that refers to the actual node location, each HM vector
initially encodes a number of CHs, denoted by CHno,such
that:
CHno =(rand ×(CH M ax N o −CH MinNo)) +CHMinNo (7)
where rand is a function that generates a random number ∈
[0,1],CHMaxNo is an estimate of the maximum number
of clusters (upper bound) and CHMinNo is the minimum
number of clusters (lower bound). In this simulation, the
value of the upper bound is set to √N,whereNis the num-
ber of nodes in the given WSN, as recommended by the
authors in [49], while the value of the lower bound is set to
2. Therefore, the number of clusters CHno will range from
CHMinNo to CHMaxNo. Even though the vector length
is allowed to vary, for a matrix representation, each vector
length in the HM must be made equal to the maximum num-
berofclusters(CHMaxNo). As a result, the remnants of
unused vector elements (referred to as “unused”) are repre-
sented with the $ sign. For example, if we have a WSN with
100 nodes, covering an area of 200m×200m, then the upper
bound will be √100 =10 and the HM vector may be as
follows: $,5,29,$,$,78,50,41,66,$.
This vector means that the given WSN can be clustered
into six clusters, where, for instance, the actual geograph-
ical location, x-axis and y-axis, for index 29 can be {120,
60},{10, 90}for index 66, and so on for the rest. These
indices represent the geographical location of the candidate
CHs encoded in this sample vector.
Ann. Telecommun.
After the HM is generated with a different number of
candidate CHs in each HM vector, the goodness of each
vector is measured as the quality of the clustering result,
which each HM vector represents, calculated by the pro-
posed objective function and saved in the HM, as explained
in the following subsection (i.e., Evaluation of Solutions).
4.1.4 Step 3: Improvise a new harmony
The new harmony vector, aNew =(aNew
1,a
New
2,a
New
3,
...,a
New
CHMaxNo), is a vector with the candidate CHs, while
the values of this vector are generated depending on the
improvisation rules of the HS, which are modified to fit the
nature of the WSN. Each new harmony vector aˆ{New}
inherits the values of its components from the vectors
stored in the HM, with the probability of HMCR. For
example, the value of aNew
1is chosen randomly from its
possible range, which is a1
1,a
2
1,a
3
1,...,a
HMS
1;foraNew
2,
it is chosen randomly from its possible range, which is
a1
2,a
2
2,a
3
2,...,a
HMS
2; it continues in the same manner for
the rest of the components. In the opposite case of the proba-
bility of 1-HMCR, the values of the new vector components
are selected from the possible range R, which represents the
indices of the sensor nodes. This process can be summarized
as follows:
aNew ←aNew ∈a1,a
2,a
3,...,a
HMS w.p HMCR
aNew ∈Rw.p(1−HMCR)
(8)
The selected new vector components (out of the memory
consideration operator) are examined to see if they are pitch-
adjusted with the probability of the PAR. If a generated
random number, rand ∈[0,1], is within the probability
of the PAR, then the new candidate CH will be adjusted to
the nearest node, based on the minimum Euclidean distance
between the new candidate CH and the other nodes in the
given WSN. The pitch adjustment process is summarized as
follows:
aNew
i←min N
i=1CH −niw.p HM CR
No change w.p (1−HMCR)
(9)
Another important issue worth mentioning is that, when
the inherited components of the new harmony vector have
unused values (‘$’), no pitch adjustment takes place. Once
the new harmony vector is generated, a count is performed
on the generated number of CHs in the new vector. If it is
less than the minimum number of CHs, i.e., ‘CHMinNo’,
the new vector will be rejected. Otherwise, the new vector
will be accepted and an objective function computed using a
cluster validity measurement, as described in the following
subsection (i.e., Evaluation of Solutions).
4.1.5 Step 4: Update the HM
In this step, the new vector is compared with the worst HM
solution in terms of the fitness function. If it is better, the
new vector is included in the HM, while the worst harmony
is excluded. Note that Steps 3 and 4 are repeated until the
maximum number of improvisations (NI) is reached.
4.1.6 Step 5: Check the stopping criteria
After the NI is reached, the best solution determined by the
maximum value of the fitness function of each HM solution
vector is finally selected to be the best solution vector.
4.2 Evaluation of solutions
The evaluation of each HM vector indicates the degree of
goodness of the solution it represents. In order to evaluate
the goodness of each HM vector, the unused elements (‘$’)
that may appear in the harmony vector are removed and
the remaining components representing the CHs are used to
cluster the given WSN. In this paper, we use the fuzzy tech-
nique in the clustering process, where the sensor node, xi,is
assigned to one or more clusters with a membership grade.
The fuzzy membership value for each node is calculated as
follows: [50]:
uij =1
C
k=1xi−CHj
xi−CHk2
m−1
(10)
where Crepresents the actual number of CHs used to
cluster the given WSN, and mis the fuzziness value of
fuzzy clustering, which is set to 1.5. Since each generated
vector holds a number of CHs, which are used to clus-
ter the given WSN, the quality of the vector is measured
using one of the clustering validity indices. This is because
the clustering validity indices have the ability to measure
the quality of the clustering results by calculating what is
called the inter-cluster distances, which refer to the sepa-
ration between clusters and the intra-cluster distances, i.e.,
the compactness within the cluster members. Therefore, the
validity index measurement is used as the fitness function in
this study.
In this case, the objective is to maximize the separation
measurement and minimize the compactness measurement.
This, in turn, helps to minimize the energy required by
sensor nodes to pass sensed data to their CH. It is worth
mentioning that, at this stage of building the infrastructure
of the given WSN, all nodes have the same energy level, as
Ann. Telecommun.
assumed in this study. Therefore, this factor is not involved
in the fitness calculation.
Several cluster validity indices for the fuzzy clustering
assessment have been proposed (e.g., see [51,52] and refer-
ences therein). In this paper, we use the PBMF index [49],
which is defined as:
PBMF(C)=1
C×E1
Ec×Dcp
(11)
where Cis the number of clusters. Here:
Ec=
c
j=1
N
i=1
um
ij
xi−CHj
(12)
and:
DC=maxi,l CHi−CHl(13)
where CHiis the center of the i-th cluster, while the power
‘p’, which is used to control the contrast between the dif-
ferent cluster configurations, is set at 2. E1is a constant
term for a particular WSN, which is used to prevent the
index value approaching 0; it is set at 10,000. The value
of m, which is the fuzziness weighting exponent, is exper-
imentally set at 1.5. Dcmeasures the maximum separation
between two clusters over all possible pairs of clusters. EC
measures the sum of Cwithin-cluster distances (compact-
ness). The maximum value of the PBMF index indicates the
correct clustering results, i.e., the correct number of CHs
that can be gained. Consequently, maximization of the fit-
ness function is desirable in order to achieve a near-optimal
solution.
4.3 Sensing and sending data
As mentioned earlier, the second part of the proposed proto-
col deals with sending sensed data from nodes to their CHs
and then to the BS. The proposed protocol involves running
into rounds, with each round subdivided into a CH selection
phase and a steady-state phase.
In the CH selection phase, the role of the CH is rotated
between the cluster nodes to evenly distribute the energy
load among the sensor nodes in each cluster. The selection
process of the new CHs is performed locally in each cluster,
where a new multi-criteria objective function is proposed to
enhance the quality of the selected CHs. Meanwhile in the
steady-state phase, the sensing and transmitting of data from
each sensor node to their cluster head are performed where
cluster heads in turn aggregate and send the sensed data to
the BS. More details of these two phases are given in the
following sections.
4.4 Phase 1: CH selection phase
After the infrastructure of the WSN is developed by the
DCHS-WSN algorithm in the BS, the CHs’ selection pro-
cess for the upcoming rounds is performed locally in each
cluster. Each CH (the one that is elected from the previous
round) will calculate the average energy level of all alive
nodes in that cluster. Only the nodes with residual energy
that is higher than the average level qualify as a candidate
CH, cdi∈CDc. The competition between candidate nodes
to become a CH is based on the following factors:
•Residual energy
•Location of each candidate node within a cluster
•Location of each candidate node with regards to the BS.
These factors are the main components of our proposed
objective function,2which is used in the selection process
of CHs. The proposed objective function is described as
follows:
CHobj =max
∀cdi∈CDcEcdi×q
α×f1+(1−α) ×f2(14)
where:
f1=
nc
j=1
xj−cdi
(15)
f2=cdi−BS(16)
In this objective function, Ecdiis the residual energy of
the candidate cluster head cdi∈cluster c. The higher the
value of this parameter, the greater the probability of a node
to become a CH. The term qis a constant term for a par-
ticular WSN and is used to avoid the objective function to
approach zero; q=1000. f1is the Euclidean distance of
all nodes in a particular cluster c,xi∀i∈cluster c,totheir
candidate cluster head cdi. This factor, f1, is responsible to
measure the intra-cluster distance (compactness) of a partic-
ular cluster when a candidate cluster head cdiis chosen to
be a CH. This in turn affects the energy expenditure required
to pass the sensed data from each sensor node to their CH.
Therefore, the minimization of f1is desired, while choos-
ing a cdithat minimizes this factor is also desirable. The
other factor, f2, is the Euclidean distance from the candidate
cluster head cdito the network BS. This in turn affects the
energy expenditure required by the CH to pass the aggre-
gated data to the network BS. Therefore, minimization of
this factor, f2, is desirable, as well as choosing the cdithat
minimizes this factor. The last factor of this objective func-
tion is the constant α, which represents the amount of the
influence of f
1sand f
2son the objective function.
2The initial version of the proposed objective function was presented
in [46,53]
Ann. Telecommun.
Overall, finding the maximum value of the objective
function CHobj in each round of the proposed protocol for
each cluster Cis desirable, as it indicates that the candidate
cluster head cdiis the best among other candidate competi-
tors. It is worth mentioning here that the cost of finding the
optimal candidate CH in each cluster is calculated as men-
tioned earlier in Section 2B, where the CH selection energy
expenditure is set to ECH−Elec =5nJ ×N o.Above.Ave.
After the optimal CHs are selected, a joint message is
sent by the current CHs to all alive sensor nodes in their
respective clusters, which contain information about the
new CHs, as well as the time schedule to transfer the data.
Once the joint message reaches a sensor node, the node
extracts the new CH identifier and transmission time sched-
ule, then stores this information in its memory to forward
data during the steady-state phase.
4.5 Phase 2: Steady-state phase
Once all nodes receive the joint message, and the transmis-
sion schedule is initialized, the sensor nodes activate their
radio component for a very short period of time to perform
data sensing and transmission to the CHs. At that time, the
CHs must be awake so as to receive the data from the nodes
in their clusters. Once the CHs receive all the data, they per-
form data aggregation, where all individual signals in each
cluster are combined into a single representative signal. This
process, as assumed in this study, is to enhance the common
signal and reduce the uncorrelated noise among the signals.
The resultant data are sent from the CHs to the BS. This
reduces the amount of information being transferred and, in
turn, reduces energy consumption.
Both the CH selection and steady-state phases are
repeated in each round of the proposed protocol throughout
the network’s lifetime.
5 Simulation results
The simulation results presented in this section are divided
into three parts as follows. The first demonstrates the effect
of the number of clusters on the performance of clustering-
based WSN protocols. The second part highlights the effec-
tiveness of the proposed DCHS-WSN algorithm in finding
the optimal number of clusters and choosing the appropriate
cluster members for various WSN settings. The final part of
this section demonstrates the efficiency of the proposed pro-
tocol compared to other state-of-the-art protocols in terms
of two main factors: (1) first and last node death; and (2)
packets sent to BS.
Before going any further, a description of the WSN settings is
given toenablea better understanding of the experimental setup.
6 Experimental setup
Two different simulations were run using MATLAB ver-
sion R2010a. The first simulation involved a WSN with 100
sensor nodes scattered randomly across a 300m×300mnet-
work (Fig. 4). The second simulation was performed with
a WSN with 200 sensor nodes scattered randomly across
a 500m×500mnetwork (Fig. 6). In these two simula-
tions, two nodes were not allowed to be in a same location,
meaning that the horizontal and vertical coordinates of each
sensor node were randomly selected between 0 and the
maximum value of the dimension (i.e., 300 for the first sim-
ulation and 500 for the second simulation). The allowed
minimum distance between each sensor node was set to
eight meters in the first simulation and 10 meters in the sec-
ond simulation. The BS location for the first simulation was
set at 175,400, while, in the second simulation, it was set at
400,650.
In both simulations, the coefficient αfrom Eq. (10) was
set at α=0.75 in order to give the compactness factor
more influence than the location of the candidate cluster
head cdiwith regard to the BS. Similar to [23,24], the
radio energy parameters used in both simulations were set
at: Eelec =50pJ/bit,Efs =10pJ /bit/m2,andEmp =
0.0013pJ/bit/m4. The size of the message that nodes sent
to their cluster heads, as well as the size of the (aggre-
gate) message that a cluster head sent to the BS, was set at
b=4,000bits/message. The packet header for each type
of packet was 25 bytes in length. The DCHS-WSN param-
eters were also experimentally set at: HMCR =0.85,
PAR =0.30, HMS =50 and NI =100000.
6.1 Impact of the number of clusters on WSN protocols
In this part, the impact of the number of clusters on the per-
formance of WSN protocols was investigated, based on the
two simulation network areas, 300m×300mand 500m×
500m. The numbers of clusters varied between one and 15,
while the clustering protocol was run for 10 different topolo-
gies (random scattering of sensor nodes on the simulated
area) for each simulation. Figures 2and 3, which show the
number of rounds when the first node died as a function of
the number of clusters, show that the simulation agrees with
the assumptions stated in the introduction section (i.e., the
number of clusters has a major impact on the performance
of the clustering-based WSN protocols).
Figure 2shows that the optimum number of clusters was
around three to five for the 300m×300mnetwork. By way
of illustration, in the simulation ‘run # 7’, the first node died
after approximately 700 rounds, when the number of clus-
ters was four. When there was only one cluster, the first
node died after 143 rounds. This is because the non-cluster
Ann. Telecommun.
Fig. 2 The effect of the number
of clusters on the lifetime of the
300m×300mWSN
0
100
200
300
400
500
600
700
800
123456789101112131415
Rounds
Number of Clusters
run #1 First Node Died run #4 First Node Died
run #5 First Node Died run #7 First Node Died
head nodes often had to transmit data over very long dis-
tances to reach the CH node. This drained energy; and, when
there were more than six clusters, not as much local data
aggregation was performed. In Fig. 3, the optimum num-
ber of clusters was shown to be around four to six for the
500m×500mnetwork. Here, for instance, the simulation
‘run# 1’ found the optimal number of clusters for the given
network to be five when the first node died after approxi-
mately 90 rounds. The first node, on the other hand, died
after approximately 20 rounds when the number of clusters
was between four and eight. It is noteworthy that the first
node died after 700 rounds in the optimal case of the first
simulation, while it took 90 rounds in the second simulation.
This is because, in both simulations, the initial energy for
each sensor node was equally set to 2, while the location of
the BS and the sensors’ coverage area were both different.
6.2 Simulations of the DCHS-WSN algorithm
This part evaluates the performance of the proposed DCHS-
WSN algorithm in finding the optimal number of clusters.
Since the HS is heuristic in nature, we performed 30 tri-
als with several random network topologies in order to
obtain the best solution. Figure 4shows how 100 sen-
sor nodes were randomly deployed in the 300m×300m
network. Figure 5shows how the DCHS-WSN algorithm
clustered the given network and how their infrastructure was
built. Figure 6shows how 200 sensor nodes were randomly
deployed in the 500m×500mnetwork. Figure 7shows how
the DCHS-WSN algorithm clustered the given network and
built their infrastructure. Figure 8shows the number of clus-
ters obtained by the DCHS-WSN algorithm in each trail for
both simulations. From the result of the 300m×300mWSN
Fig. 3 The effect of the number
of clusters on the lifetime of the
500m×500mWSN
0
10
20
30
40
50
60
70
80
90
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Rounds
Number of Clusters
run #1 First Node Died run #3 First Node Died
run #4 First Node Died run #7 First Node Died
Ann. Telecommun.
Fig. 4 Unclustered 300m×300mnetwork
simulation shown in a blue color in Fig. 8, it can be seen
that the number of clusters obtained by the DCHS-WSN
algorithm varied between three and five; note that this is the opti-
mal range as mentioned in the previous section. Furthermore,
from the result of the 500m×500mWSN simulation shown
in a red color in Fig. 8, it can be seen that the number of clusters
obtained by the DCHS-WSN algorithm varied between four
and six, which is also the optimal range as mentioned in the
previous section. The plot confirms that the proposed algo-
rithm can produce the optimal number of clusters, which in
turn results in greater energy savings in the network.
6.3 Comparative simulations with clustering-based
routing protocols
After the optimal number of clusters was obtained by
DCHS-WSN, a demonstration of the system lifetime,
Fig. 5 DCHS-WSN clustered 300m×300mnetwork
Fig. 6 Unclustered 500m×500mnetwork
defined by the first and last dead nodes, was considered.
When the nodes use up their limited energy during the
course of the simulation, they can no longer transmit or
receive data. For these simulations, energy is consumed
whenever a node transmits or receives data, or performs data
aggregation or CH selection. Furthermore, we tracked the
rate at which the data packets are transferred to the BS. The
more data the BS receives, the more accurate its view of the
remote environment will be.
Figure 9-a and -b show how the DEEC and LEACH pro-
tocols compare with our protocol DCHS-WSN in terms of
the number of rounds before the occurrence of a first dead
node, as well as the number of rounds until the last dead
node. It can be seen from the figures for both simulations
that the network lifetime for the DCHS-WSN protocol was
significantly better than for the DEEC and LEACH pro-
tocols. Figure 9-a, which represents the simulation in the
Fig. 7 DCHS-WSN clustered 500m×500mnetwork
Ann. Telecommun.
0
1
2
3
4
5
6
7
1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930
Obtained Number of Clusters by DCHS-WSN
Optimal Number of Clusters (300mX300m)
Optimal Number of Clusters(500mX500m)
Fig. 8 Obtained number of clusters using the DCHS-WSN algorithm
for the two simulations
300m×300marea network, shows that the first node died
after 45 rounds for the DEEC protocol and after 176 rounds
for the LEACH protocol, while the first node died after 440
rounds for the DCHS-WSN protocol. Figure 9-a also shows
176
45
440
1756
1101
2319
0
500
1000
1500
2000
2500
LEACH DEEC DCHS-WSN
Rounds
First and Last Node Died in 300mx300m Network
First Node Died Last Node Died
21 14
60
415
1101
1382
0
200
400
600
800
1000
1200
1400
1600
LEACH DEEC DCHS-WSN
Rounds
First and Last Node Died in 500mx500m Network
First Node Died Last Node Died
Fig. 9 Comparison of the first and last dead nodes for the two
simulations
0
1000
2000
3000
4000
5000
6000
LEACH DEEC DCHS-WSN
Total Packets Sent to BS
Total Packets sent to BS (300mX300m) Total Packets sent to BS (500mX500m)
Fig. 10 Total data delivered to the BS for the two simulations
that the last node died after 1,101 rounds for the DEEC
protocol and after 1,756 rounds for the LEACH protocol; for
the DCHS-WSN protocol, it occurred after 2,319 rounds.
The performance of the DCHS-WSN, in terms of its
capability to deliver data to the BS and energy efficiency,
was compared with the DEEC and LEACH protocols.
Figure 10 shows the total data received by the BS for the dif-
ferent simulated networks. It can be seen that, in both cases
of the different network area, the DCHS-WSN achieved bet-
ter data delivery than the DEEC and LEACH protocols. For
instance, Figure 10 shows the number of packets delivered
to the BS by the DCHS-WSN protocol, with 5,224 packets
for the 300m×300marea network simulation, compared to
2,605 packets for the 500m×500marea network simula-
tion. The difference is obvious, especially when the initial
energy for each node in both simulations is known to be the
same (e.g., 2J). However, it can also be seen that, in both the
300m×300mand 500m×500mcases, the DCHS-WSN
protocol exploited the network energy as much as possible,
compared to the other protocols, which in turn resulted in
better data delivery. Thus, the DCHS-WSN protocol rep-
resents a worthy approach for the efficient utilization of
network energy resources.
7Conclusion
In WSNs, it is very important to develop routing protocols
that can conserve the energy of the nodes as much as pos-
sible in order to improve the network lifetime. In response,
we designed a new dynamic clustering protocol for WSNs
based on the HS algorithm. The proposed protocol elimi-
nates the need to determine the number of clusters in the
simulation a priori. This number is considered a key fac-
tor, which may affect the performance of clustering-based
WSN protocols. Furthermore, a decentralized cluster-based
protocol is proposed, where the selection process of cluster
Ann. Telecommun.
heads in each simulation round is conducted within each
cluster, instead of the BS. This is predicated on a new
multi-objective function, in which the network energy con-
sumption, intra-cluster distance and cluster-to-BS distance
are its main factors. Simulation results have shown that
the proposed algorithm can generate an optimal number of
clusters in each round during the simulations. Moreover,
the proposed protocol offers an improvement in the net-
work lifetime compared to other popular algorithms, such
as LEACH and DEEC.
This improvement is based on different factors. Firstly,
significant energy saving is achieved by the DCHS-WSN
algorithm through the use of dynamic clustering, where the
optimal number of clusters is chosen and each node joins the
most appropriate cluster. This is based on the strength of the
HS algorithm in terms of exploring the search space (sen-
sor nodes) and discovering the most appropriate network
infrastructure. The objective function used in the DCHS-
WSN algorithm is also oriented towards minimizing the
distance from non-cluster head nodes to cluster heads, as
well as the distance from cluster heads to the BS. Sec-
ondly, the decentralized technique proposed in our protocol
leads to lower network energy expenditure. This is due to
the reduction in energy required to send control messages
from alive nodes to their cluster head, where the process
of selecting a new CH is performed, as opposed to the BS
giving information about their residual energy. Thirdly, the
multi-objective function adopted in our proposed protocol
attempts to produce a set of good compromises or trade-offs,
where the values of all the objective functions are accept-
able to the system requirements. As a result, this protocol
can obtain an optimal set of cluster heads that are evenly
distributed across the network with optimal network con-
figuration, which further reduces the total network energy
dissipation.
Acknowledgments The author gratefully acknowledges the support
received in the course of this study from the Sensor Networks and
Cellular Systems Research Center, the University of Tabuk, and the
Ministry of Education in Saudi Arabia.
References
1. Akyildiz IF, Su W, Sankarasubramaniam Y, Cayirci E (2002)
Wireless sensor networks: a survey. Comput Netw 38(4):393–422
2. Shi H, Liang G, Wang H (2014) A novel traffic identi-
fication approach based on multifractal analysis and com-
bined neural network. Annals Telecommunications-annales des
T`
e,l`
ecommunications 69(3-4):155–169
3. Kim JW, Barrado JRR, Jeon DK (2016) Time and energy effi-
cient relay transmission for multi-hop wireless sensor networks.
Sensors 16(7):985
4. Saleem M, Di Caro GA (2011) Farooq swarm intelligence based
routing protocol for wireless sensor networks: survey and future
directions. Inf Sci 181(20):4597–4624
5. Maskooki A, Soh CB, Gunawan E, Low KS (2015) Adaptive
routing for dynamic on-body wireless sensor networks. IEEE J
Biomed Health Inform 19(2):549–558
6. Loutfi A, Elkoutbi M, BenOthman J, Kobbane A (2014)
An energy aware algorithm for OLSR clustering. Annals
of Telecommunications-annales des T`
e,l`
ecommunications 69(3-
4):201–207
7. Razaque A, Elleithy KM (2014) Energy-efficient boarder node
medium access control protocol for wireless sensor networks.
Sensors 14(3):5074–5117
8. Kulkarni RV, Venayagamoorthy GK (2011) Particle swarm opti-
mization in wireless-sensor networks: a brief survey. IEEE Trans
Syst Man Cybern Part C Appl Rev 41(2):262–267
9. Ahmed MHU, Razzaque MA, Hong CS (2013) DEC-MAC: delay-
and energy-aware cooperative medium access control protocol for
wireless sensor networks. Annals of Telecommunications-annales
des T`
e,l`
ecommunications 68(9-10):485–501
10. Umar IA, Mohd Hanapi Z, Sali A, Zulkarnain ZA (2016) FuGeF:
a resource bound secure forwarding protocol for wireless sensor
networks. Sensors 16(6):943
11. Yick J, Mukherjee B, Ghosal D (2008) Wireless sensor network
survey. Comput Netw 52(12):2292–2330
12. Al-Karaki JN, Kamal AE (2004) Routing techniques in wirless
sensor networks: a survey. IEEE Wirel Commun 11(6):6–28
13. Liu X (2012) A survey on clustering routing protocols in wireless
sensor networks. Sensors 12(8):11113–11153
14. Alia OM (2017) Dynamic relocation of mobile BS in wireless sen-
sor networks using a cluster-based harmony search algorithm. Inf
Sci 385:76–95
15. Fu P, Cheng Y, Tang H, Li B, Pei J, Yuan X (2017) An effective
and robust decentralized target tracking scheme in wireless camera
sensor networks. Sensors 17(3):639
16. Ren P, Qian J (2016) A power-efficient clustering protocol for
coal mine face monitoring with wireless sensor networks under
channel fading conditions. Sensors 16(6):835
17. Xu L, O’Hare GM, Collier R (2017) A smart and balanced energy-
efficient multihop clustering algorithm (smart-BEEM) for MIMO
IoT systems in future networks. Sensors 17(7):1574
18. Srie Vidhya Janani E, Ganesh Kumar P (2015) Energy efficient
cluster based scheduling scheme for wireless sensor networks. Sci
World J 2015, Article ID 185198, p 9
19. Yu J, Feng L, Jia L, Gu X, Yu D (2014) A local energy con-
sumption prediction-based clustering protocol for wireless sensor
networks. Sensors 14(12):23017–23040
20. Mustapha I, Ali BM, Rasid MFA, Sali A, Mohamad H (2015)
An energy-efficient spectrum-aware reinforcement learning-based
clustering algorithm for cognitive radio sensor networks. Sensors
15(8):19783–19818
21. Gao Y, Wkram CH, Duan J, Chou J (2015) A novel energy-aware
distributed clustering algorithm for heterogeneous wireless sen-
sor networks in the mobile environment. Sensors 15(12):31108–
31124
22. Ogundile OO, Alfa AS (2017) A survey on an energy-efficient and
energy-balanced routing protocol for wireless sensor networks.
Sensors 17(5):1084
23. Heinzelman WR, Chandrakasan A, Balakrishnan H (2000)
Energy-efficient communication protocol for wireless microsen-
sor networks. In: Proceedings of the 33rd Annual Hawaii Interna-
tional Conference on System Sciences, vol 2, pp 1–10
24. Heinzelman WB, Chandrakasan AP, Balakrishnan H (2002) An
application-specific protocol architecture for wireless microsensor
networks. IEEE Trans Wirel Commun 1(4):660–670
25. Younis O, Fahmy S (2004) HEED: A hybrid, energy-efficient,
distributed clustering approach for ad-hoc sensor networks. IEEE
Trans Mob Comput 3(4):366–379
Ann. Telecommun.
26. Qing L, Zhu Q, Wang M (2006) Design of a distributed energy
efficient clustering algorithm for heterogeneous wireless sensor
networks. Comput Commun 29(12):2230–2237
27. Lindsey S, Raghavendra C, Sivalingam KM (2002) Data gathering
algorithms in sensor networks using energy metrics. IEEE Trans
Parallel Distributed Syst 13(9):924–935
28. Manjeshwar A, Agrawal DP (2001) TEEN: a routing protocol
for enhanced efficiency in wireless sensor networks. In: Proceed-
ings of the 15th International Parallel and Distributed Processing
Symposium (IPDPS), San Francisco, CA, USA, pp 2009–2015
29. Smaragdakis G, Matta I, Bestavros A (2004) SEP: a stable election
protocol for clustered heterogeneous wireless sensor networks. In:
Proceedings of the International Workshop on SANPA, pp 251–
261
30. Julie EG, Selvi S (2016) Development of energy efficient clus-
tering protocol in wireless sensor network using neuro-fuzzy
approach. Sci World J 2016, Article ID 5063261, p 8
31. Liaqat M, Gani A, Anisi MH, Ab Hamid SH, Akhunzada A,
Khan MK, Ali RL (2016) Distance-based and low energy adap-
tive clustering protocol for wireless sensor networks. PloS one
11(9):e0161340
32. Zhang Y, Wang J, Han D, Wu H, Zhou R (2017) Fuzzy-logic based
distributed energy-efficient clustering algorithm for wireless sen-
sor networks. Sensors 17(7):1554
33. Hematkhah H, Kavian YS (2015) DCPVP: distributed clustering
protocol using voting and priority for wireless sensor networks.
Sensors 15(3):5763–5782
34. Beulah Jayakumari R, Jawahar Senthilkumar V (2015) Priority
based congestion control dynamic clustering protocol in mobile
wireless sensor networks, Sci World J, 2015, Article ID 596138, p
10
35. Han R, Yang W, Wang Y, You K (2017) 2017 DCE: a distributed
energy-efficient clustering protocol for wireless sensor network
based on double-phase cluster-head election. Sensors 17(5):998
36. Alia OM, Mandava R, Ramachandram D, Aziz ME (2009)
Dynamic fuzzy clustering using harmony search with application
to image segmentation. In: Proceedings of the IEEE International
Symposium on Signal Processing and Information Technology,
ISSPIT09, pp 538–543
37. Alia OM, Mandava R, Aziz M (2011) A hybrid harmony search
algorithm for MRI brain segmentation. Evol Intell., Springer-
Verlag 4(1):31–49
38. Hruschka ER, Campello RJ, Freitas AA, Carvalho AC (2009) A
survey of evolutionary algorithms for clustering. IEEE Trans Syst
Man Cybern Part C Appl Rev 39(2):133–155
39. Latiff NA, Tsimenidis CC, Sharif BS, Ladha C (2008) Dynamic
clustering using binary multi-objective particle swarm optimiza-
tion for wireless sensor networks. In: IEEE International Sym-
posium on Personal, Indoor and Mobile Radio Communications
(PIMRC), pp 1–5
40. Falkenauer E (1998) Genetic algorithms and grouping problems.
Wiley, New York
41. Huang S, Tao M (2017) Competitive swarm optimizer based gate-
way deployment algorithm in cyber-physical systems. Sensors
17(1):209
42. Saleem K, Derhab A, Orgun MA, Al-Muhtadi J, Rodrigues JJ,
Khalil MS, Ali Ahmed A (2016) Cost-effective encryption-based
autonomous routing protocol for efficient and secure wireless
sensor networks. Sensors 16(4):460
43. Geem ZW, Kim JH, Loganathan GV (2001) A new heuristic
optimization algorithm: harmony search. Simulation 76(2):60–68
44. Alia OM, Al-Ajouri A (2017) Maximizing wireless sensor net-
work coverage with minimum cost using harmony search algo-
rithm. IEEE Sensors J 17(3):882–896
45. Alia OM, Mandava R (2011) The variants of the harmony search
algorithm: an overview. Artif Intell Rev 36(1):49–68
46. Alia OM, Shaaban Z, Basheer A, Al-Ajouri A, Alsswey A (2014)
Musicians’-inspired clustering protocol for efficient energy wire-
less sensor networks. In: Proceedings of the Fourth International
Conference on Communications and Networking (ComNet’14),
pp 1–6
47. Wang A, Heinzelman WB, Sinha A, Chandrakasan AP (2001)
Energy-scalable protocols for battery-operated microsensor net-
works. J VLSI Signal Process Syst Signal, Image Video Technol
29(3):223–237
48. Geem ZW, Tseng CL, Park Y (2005) Harmony search for general-
ized orienteering problem: best touring in China. In: International
Conference on Natural Computation. Springer, Berlin, pp 741–
750
49. Pakhira MK, Bandyopadhyay S, Maulik U (2004) Validity index
for crisp and fuzzy clusters. Pattern Recogn 37(3):487–501
50. Bezdek JC (1981) Pattern recognition with fuzzy objective func-
tion algorithms. Kluwer Academic Publishers, Norwell
51. Wang W, Zhang Y (2007) On fuzzy cluster validity indices. Fuzzy
Sets Syst 158(19):2095–2117
52. Halkidi M, Batistakis Y, Vazirgiannis M (2001) On clustering
validation techniques. J Intell Inf Syst 17(2):107–145
53. Alia OM (2014) A decentralized fuzzy C-means-based energy-
efficient routing protocol for wireless sensor networks. Sci World
J, 2014, Article ID 647281, p 9
