23614 IEEE SENSORS JOURNAL, VOL. 21, NO.

20, OCTOBER 15, 2021

An Energy-Efficient Dynamic Clustering Protocol

for Event Monitoring in Large-Scale WSN
Zhiyi Qu , Huihui Xu, Xue Zhao, Hongying Tang , Jiang Wang , and Baoqing Li

Abstract —As a key technology, clustering has been an

effective way in large-scale wireless sensor networks (WSNs)
to extend the lifetime. However, the static cluster struc-
ture in most of the traditional method is formed without
considering the development of the event. In this paper,
we propose an Energy-Efficient Dynamic Clustering (EEDC)
protocol for event monitoring applications in large scale WSN.
In EEDC, a dynamic clustering method using Rough Fuzzy
C-Means and Genetic algorithm (RFCM-GA) is designed.
Firstly, the idea of fuzzy set and rough set in RFCM are used to
form the overlapping cluster, which can guarantee the quality
of coverage of the developing event. Secondly, we use GA to
perform a parallel search in each cluster to find the optimal
set of candidate cluster heads (CCHs). RFCM-GA can use its powerful global search capabilities and fast convergence
speed to obtain the best clustering results. Simulation results demonstrate that EEDC has higher energy efficiency and
prolongs the network lifetime compared to the existing approaches.
Index Terms — Dynamic clustering, event monitoring, rough fuzzy C-means, genetic algorithm, wireless sensor
networks, energy efficiency.

I. I NTRODUCTION perform sensing task due to the locality and randomness of

the event. The event may refer to a moving individual (such
T HE recent booming of the Internet of Things (IoT), 5G
communication technology, artificial intelligence (AI),
smart electronic devices, and so forth, imposes increasing
as a moving military target) or an area of varying scope (such
as a fire disaster area). The nodes in WSN are responsible for
requirements on communication network [1]–[4]. Due to the monitoring the location or scope of the event and reporting
promising application prospect and significant research value, it to the external network users [5]–[8]. A particular node
wireless sensor network (WSN) has attracted much attention called ‘sink’ acts as a gateway between the WSN and the
in many industries. Its application range includes military external network. The relevant data of the event is sent to
equipment, environmental monitoring, smart home, health users through sink. In a small-scale network, sink can obtain
care, intelligent transportation, etc. data by flooding throughout the whole network. However,
As one of the essential applications of WSN, event mon- this approach does not work well in large-scale networks
itoring only requires a part of the nodes to be activated to because information exchange will consume too much energy.
Therefore, how to choose the activated nodes in a larch-scale
Manuscript received May 21, 2021; revised June 28, 2021 and July 30, network and organize them into a dynamic structure to col-
2021; accepted July 30, 2021. Date of publication August 9, 2021; date
of current version October 18, 2021. This work was supported by the lect event data efficiently is a challenging task. In general,
National Natural Science Foundation of China under Grant 61901457. the existing methods can be classified into two categories:
The associate editor coordinating the review of this article and approv- tree-based and cluster-based methods.
ing it for publication was Prof. Jaime Lloret. (Corresponding authors:
Jiang Wang; Baoqing Li.)
Zhiyi Qu, Huihui Xu, and Xue Zhao are with the Science and A. Tree-Based Method
Technology on Microsystem Laboratory, Shanghai Institute of Microsys- The tree-based methods monitor the event by construct-
tem and Information Technology, Chinese Academy of Sciences,
Shanghai 200050, China, and also with the School of Electronic, Elec- ing the underlying tree structure between nodes [9], [10].
trical and Communication Engineering, University of Chinese Acad- As shown in Fig. 1, when an event occurs in the network,
emy of Sciences, Beijing 100049, China (e-mail: [email protected]; its nearby nodes will form a tree. A root node is first selected
[email protected]; [email protected]).
Hongying Tang, Jiang Wang, and Baoqing Li are with the Sci- based on the location of the event, and the leaf nodes are
ence and Technology on Microsystem Laboratory, Shanghai Institute added according to the shortest path algorithm. Each leaf node
of Microsystem and Information Technology, Chinese Academy of sends the event data to its parent node, and the root node will
Sciences, Shanghai 200050, China (e-mail: [email protected];
[email protected]; [email protected]). eventually collect all the data. After processing, the root node
Digital Object Identifier 10.1109/JSEN.2021.3103384 obtains the result and reports it to sink. In addition, the tree

1558-1748 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Aditya Engineering College. Downloaded on November 24,2021 at 05:14:04 UTC from IEEE Xplore. Restrictions apply.
QU et al.: EEDC PROTOCOL FOR EVENT MONITORING 23615

Fig. 2. Target at the junction of three clusters.

and does not change according to the development of events.

As shown in Fig. 2, when an event occurs at the junction of
several clusters, the nodes can sense it belong to three different
clusters. All nodes in the three clusters will be forced to remain
active, which will result in a waste of energy.
To improve the traditional methods, a novel energy-efficient
dynamic clustering protocol is proposed in this paper. The
main contributions are presented as follows:
(1) A dual back-off mechanism is designed for selecting the
initial CH. The back-off value includes two parts. The fixed
Fig. 1. Tree-based method: (a) Form a tree (b) Tree reconfiguration.
part is related to residual energy and distance to the event.
The random part is to prevent multiple nodes from declaring
structure will adjust as the event develops, and the roles of as CH at the same time.
root and leaf nodes will also change accordingly. (2) A dynamic clustering scheme based on RFCM-GA is
The tree-based methods can reduce the redundant overhead proposed to organize the nodes around the event into clusters.
and network load. Nevertheless, to establish and maintain Firstly, the number of CCHs is derived based on the event
a dynamic tree structure, many interactive messages are status and network topology. Then an overlapping cluster
required. In large-scale networks, this energy consumption of structure is generated by using the Rough Fuzzy C-means
interaction between nodes may be massive. algorithm (RFCM). After that, the Genetic Algorithm (GA)
will be used to search for the most suitable node in each cluster
to form the optimal CCH set.
B. Cluster-Based Method (3) A strategy of cluster structure adjustment is designed to
Clustering is one of the most commonly used energy-saving deal with the dynamic development of the event. We discuss
technologies, widely used to reduce energy consumption and this strategy under two cases of event development, the single
prolong network life in WSN [11]–[13]. It has been proved moving target and the continuous changing area.
that the CH selection problem in clustering is a well-known (4) In the simulation experiment, we validate the energy
NP-Hard problem [14]. The traditional clustering approach is efficiency of EEDC from the perspective of the number of
to consider several factors (such as residual energy, distance alive nodes, total energy consumption and the average residual
to sink, node degree, etc.) to select cluster head (CH) and energy in each round. In addition, the comparison of HND
form clusters in the network [15]–[19]. CH is responsible for and LND under different numbers of nodes is also provided.
collecting data from its cluster members (CMs). After data The experiment also compares the convergence speed of
fusion, CH reports the result to sink. If the event moves or RFCM-GA and GA.
expands outside of a cluster, other nearby clusters will also The rest of this paper is organized as follows. In Section II,
participate in the sensing task. This process is repeated until related works are introduced. The network and energy models
the event disappears or the network is dead. are given in Section III. The details of our algorithm are
Since the additional tasks of CH will consume extra energy, discussed in Section IV. In Section V, we analyze the per-
it is necessary to reselect them after several rounds of transmis- formance of our approach and simulate it in comparison with
sion. This periodic adjustment will bring additional commu- other existing algorithms. Finally, Section VI concludes the
nication costs. Besides, the cluster structure is generally static paper.

Authorized licensed use limited to: Aditya Engineering College. Downloaded on November 24,2021 at 05:14:04 UTC from IEEE Xplore. Restrictions apply.
23616 IEEE SENSORS JOURNAL, VOL. 21, NO. 20, OCTOBER 15, 2021

II. R ELATED W ORKS structure of the entire network does not have to change,
In this section, we will list several existing network proto- which can save much energy. The selection of CH considers
cols and analyze their performance in event monitoring in a the residual energy and node degree. The simulation result
large-scale network. shows that ADEC has good scalability and can balance energy
consumption.
Although ADEC avoids the vast energy consumption caused
A. LEACH
by global cluster structure adjustment, its clustering process
LEACH [20] is a classic low-power adaptive clustering pro- does not consider monitoring dynamic events. When the event
tocol. In LEACH, each node decides whether it can be a CH spreads across the annulus, ADEC will not be able to organize
by comparing a random number between 0 and 1 with a given an effective topology.
threshold. The following equation estimates the threshold T(n):
⎧ p
⎨ , if n ∈ G, D. Other Related Algorithms
T (n) = 1 − p × (r mod p )
1
(1)
⎩ In [13], an adaptive dynamic clustering scheme for target
0, otherwise, tracking is presented. A model for the prediction of the
where ‘r ’ is the current round, ‘ p’ is the desired probability target location and the formation of an adaptive cluster is
of the nodes to become CH, and ‘G’ is the set of nodes that designed to optimize the power consumption in the network.
have not become CH in the last 1/p rounds. The equation In [23], a time-division election transmission model was
(1) ensures that within 1/p rounds, each of the sensor nodes first proposed to reduce energy consumption caused by data
becomes the CH only once. LEACH balances network energy collision. Based on energy information and tracking quality,
consumption by rotating CHs. The experimental results show the energy-balanced farthest scheduling strategy was proposed,
that LEACH’s network survival time is 15% longer than the effectively reducing energy consumption and improving track-
general plane routing protocol network. ing accuracy. In CAICS [24], a collaborative node selection
However, LEACH assumes that all nodes can directly com- method based on the spatial correlation of sensors is proposed.
municate with sink nodes, which is impractical in large-scale Each CH selects the tracking nodes with the most prominent
networks. In addition, the selection of CHs is completely joint information utility to reduce energy consumption. A
random without considering the state of the event. novel protocol based on the k-nearest neighbours (KNN)
algorithm is proposed in [25]. KNN is utilised to achieve
efficient continuous tracking with coverage holes in the net-
B. EEAOC
work. An energy-efficient incremental clustering algorithm
EEAOC [21] is a clustering algorithm for dynamic event followed by Gaussian adaptive resonance theory is proposed
monitoring. In EEAOC, a 2-logical overlapping clustering in [26]. The cluster can learn, create, update, and retain
scheme is designed so that adjacent sensors in the event area incrementally through online learning to adapt to the dynamic
can be grouped into the same cluster for data fusion. All sensor event. In HCTT [27], a hybrid clustering strategy is proposed,
nodes will exchange information in the initial phase. Each in which a dynamic cluster structure is combined with an
node will save a slave-CH for backup. When the event spreads existing static cluster. Whenever the event spreads to the
out of the current cluster, the boundary node will send data current static cluster boundary, an on-demand dynamic cluster
to its slave-CH. In this way, an overlapping cluster structure is formed and temporarily responsible for monitoring.
can be generated based on the development of the event.
Experimental results show that EEAOC can achieve higher
efficiency in the application of event monitoring.
III. N ETWORK AND E NERGY M ODELS
However, in EEAOC, the cluster structure will not change
once it is formed. When a CH is about to die, the closest A. Network Model
CM will take its role. In this way, some CMs are no longer
In order to present our algorithms, the assumptions about
within the new CH’s communication range and become iso-
the network model are as follows:
lated nodes. The same situation holds true for the slave-CH.
(1) There are ‘N’ sensor nodes randomly distributed in an
After several rounds of transmission, more and more isolated
‘L × L’ monitoring area with high density. The sink is located
nodes will appear in the network. The isolated nodes will
in the centre of the network.
remain active till death, and the data they sensed will be
(2) All the sensor nodes and the sink are stationary after
lost. Therefore, the energy efficiency of the network cannot
deployment.
be guaranteed to the end.
(3) The sensor nodes are all homogeneous, and the sink
has enough energy to support the continuous operation of the
C. ADEC WSN.
In ADEC [22], the monitoring area is divided into annu- (4) The sensor nodes are aware of their location coordinates.
luses, and nodes in the same annulus are at the same level. The clocks of all nodes are synchronized.
Clustering is performed only between nodes of the same level. (5) The wireless channel is completely symmetric, and the
In this way, when any CH dies, the process of re-clustering transmission power can be adjusted according to the distance
will only be carried out in the corresponding annulus. The between two nodes.

Authorized licensed use limited to: Aditya Engineering College. Downloaded on November 24,2021 at 05:14:04 UTC from IEEE Xplore. Restrictions apply.
QU et al.: EEDC PROTOCOL FOR EVENT MONITORING 23617

(6) Each node gets a unique ID through the assignment

algorithm proposed in [28].

B. Energy Model
The energy dissipation of a sensor node for transmitting and
receiving ‘k’ bits over a distance ‘d’ are (the same as the radio
model in [21], [24]),

k(E elec + ε f s × d 2 ), d < d0
E T x (k, d) = (2)
k(E elec + εmp × d 4 ), d ≥ d0
E Rx = k × E elec (3)
where ‘E elec ’ is the energy spent per bit of the transmitting
circuit, ‘ε f s ’ and ‘εmp
’ are the energy coefficient of power
amplification, ‘d0 = εmp /ε f s ’ is the threshold distance of
the two energy models.

IV. P ROPOSED P ROTOCOL Fig. 3. The occurrence of the event activates the surrounding nodes.
In this section, the details of the proposed dynamic cluster-
ing protocol, namely EEDC, is discussed. Firstly, a broadcast
delay related to the sensing distance and residual energy of
nodes is introduced for initial CH competing. Next, a new
hybrid algorithm RFCM-GA is designed to solve the CCH
selection problem. In the proposed algorithm, the soft clus-
tering method supported by fussy set theory is combined
with robust global search ability to provide a better clustering
solution. After this, a strategy of cluster structure adjustment
is also presented to deal with the dynamic development of the
event.
The proposed EEDC protocol operates in several epochs,
and each epoch includes four steps: 1) network initialization,
2) initial cluster formation, 3) dynamic clustering scheme
based on RFCM-GA and 4) cluster structure adjustment.

A. Network Initialization
In network initialization, sink will flood a “Hello” message
in the network. By receiving and forwarding this message,
Fig. 4. Formation of the initial cluster.
each node can create a neighbour information table containing
nodes’ ID, residual energy and location. Each node takes the
neighbour with the most residual energy as its preferred relay factors, and satisfy ‘a +b = 1’. The first term on the right side
node to sink. Besides, each node can know its node degree by of equation (4) guarantees that the node with more residual
calculating the number of neighbours. energy and closer to the event becomes CH. The second term
Network initialization is run at the beginning of each epoch ‘Trand ’ is a relatively small random value, which is set to
before the event appears. Its function is to gather the necessary prevent nodes from simultaneously declaring as CH.
network information in preparation for the following steps. After ‘TD ’, the first node that broadcasts the “CH competing
message” will announce itself as the initial CH. The broadcast
B. Initial Cluster Formation range is the communication radius ‘RC ’. If a node receives a
“CH competing message” before ‘TD ’ expires, it will give
When an event occurs in the monitoring area, nodes within
up the CH competition and become a cluster member (CM).
‘R S ’ from the event centre will be activated, as shown in
CMs will send node ID, residual energy, node degree and
Fig. 3. To compete to become the initial CH, each node sends
position coordinates to CH. The initial cluster structure is
a delayed broadcast message, which is introduced to avoid
shown in Fig. 4. The small circle of radius ‘R S ’ centred on
collision and reduce message overhead. The delay value of
the event centre represents the range of nodes participating in
node ‘i ’ is given as [21]:
the sensing task. The large circle of radius ‘RC ’ centred on the
a Erein(i) +b dRe (i)
E
TD (i ) = e S + Trand , (4) CH indicates the cluster range.
Moreover, the transmission path from CH to sink is obtained
where ‘Ere (i )’ is the residual energy of node ‘i ’, ‘E in ’ is the after CH is determined. CH will send a “relay awaken mes-
initial energy, ‘de (i )’ is the distance from node ‘i ’ to the event sage” to the preferred relay node saved in network initializa-
centre, ‘R S ’ is the sensing radius, ‘a’ and ‘b’ are the weight tion. The relay node will continue to forward the message to

Authorized licensed use limited to: Aditya Engineering College. Downloaded on November 24,2021 at 05:14:04 UTC from IEEE Xplore. Restrictions apply.
23618 IEEE SENSORS JOURNAL, VOL. 21, NO. 20, OCTOBER 15, 2021

TABLE I
N OTATIONS AND D EFINITIONS IN EEDC

Fig. 5. Approximate distribution of CH and CCH.

its preferred relay node. In this way, a transmission path from
CH to sink is created. Some notations and definitions in EEDC be set as large as possible. Moreover, the distance between
are listed in Table I. CCHs should also be set appropriately to ensure uniform
distribution of CCHs. To ensure a certain degree of overlap,
we set the distance ‘d’ between CCHs as:
C. Dynamic Clustering Scheme Based on RFCM-GA
The problem studied in this section is to realize an efficient R S < d < RC . (8)
clustering while monitoring a dynamic event. Considering
the dynamic characteristics of the event, a novel clustering Substituting (5) and (7) into (8), we can obtain:
method by integrating RFCM and GA is proposed. To prevent π
σ RC < 2RC sin( ) < RC , (9)
event loss in monitoring, an overlapping area is required k
between clusters to improve the coverage capability. RFCM is and the range of ‘k’ is derived by the following formula:
a soft clustering algorithm supported by fuzzy set and rough
σ π 1
set theory, derived from the k-means algorithm. In RFCM, < sin( ) < . (10)
the overlap of clusters is acceptable. We combined the idea of 2 k 2
RFCM and GA to obtain the optimal selection result of CCHs. Thus, equation (10) can be:
CCHs are evenly distributed around the initial CH, and their σ π 1
roles are adjusted according to the development of the event. arcsin( ) < < arcsin . (11)
2 k 2
Step C can be divided into three phases:
1) Determine the Number of CCHs: We start with deriv- At last, the number of CCHs is derived:
ing the number of CCHs. To ensure network connectivity, π
6<k< . (12)
we express the relationship between ‘R S ’ and ‘RC ’ as: arcsin(σ/2)
R S = σ RC , σ ∈ (0.5, 1), (5) The upper limit of the number of CCHs is determined by the
relationship between ‘R S ’ and ‘RC ’ in a practical application.
where ‘σ ’ is a constant, representing the ratio of ‘R S ’ to ‘RC ’. According to equation (12), if higher efficiency is required,
The specific value of ‘σ ’ is related to the type of the event. ‘k’ should be set as 7; if better coverage of event is required,

Assuming that the number of CCHs is determined to be k, π
‘k’ should be set as ‘ arcsin(σ/2) ’.
a k-sided shape can be obtained by connecting the positions
of the CCHs. To facilitate the derivation of ‘k’, we assume 2) RFCM Based Cluster Formation: After determining the
that it is a regular k-sided inscribed the cluster. As shown number of CCHs, the cluster formation is carried out using
in Fig. 5, according to the properties of a regular k-sided RFCM. RFCM is obtained by introducing the idea of the
polygon, the angle ‘α’ in the triangle AO B can be expressed fuzzy sets and the rough sets into the k-means algorithm. The
as fuzzy membership in fuzzy sets can deal with overlapping
2π clustering; the upper and lower approximation in rough sets
α= . (6) enable handling uncertainty, vagueness, and incompleteness in
k
the class definition.
The distance between A and B can be calculated by In RFCM, nodes are randomly assigned to the nearest
equation (7): cluster, and the distance between the node and the cluster
α π centre is calculated. If the distance is less than the threshold,
d = 2RC sin( ) = 2RC sin( ). (7)
2 k the node belongs to the upper approximate set of the cluster.
To avoid energy consumption caused by frequent cluster Otherwise, the node is classified as the lower approximate set.
structure changes, the distance between CH and CCH should The algorithm will iterate until the objective function is less

Authorized licensed use limited to: Aditya Engineering College. Downloaded on November 24,2021 at 05:14:04 UTC from IEEE Xplore. Restrictions apply.
QU et al.: EEDC PROTOCOL FOR EVENT MONITORING 23619

Fig. 6. βi is represented by crisp lower bound and fuzzy boundary.

than a given value ‘δ’. The process of cluster formation is as

follows:
Let X = {x 1, . . . , x j , . . . , x n } be the set of ‘n’ nodes and
V = {v 1 , . . . , v i , . . . , v k } be the set of ‘k’ cluster centres. Let
‘ Ā(βi )’ and ‘ A(βi )’ be the upper and lower approximations
of cluster ‘βi ’, and ‘B(βi ) = { Ā(βi ) − A(βi )}’ denote the
boundary region of cluster ‘βi ’. Fig. 7. Pseudo-code for RFCM based cluster formation.
Fig. 6 shows the crisp lower bound and fuzzy boundary of
cluster ‘βi ’. The properties of RFCM are as follows: nodes in the boundary region because they definitely belong
⎧ to a particular cluster. Thus, the value is given by ‘0 < w̄ <
⎪
⎪ Ā(βi ) = X (B.1)
⎪
⎪ w < 1’. The membership value ‘μi j ’ indicates the degree of
⎪
⎪ ∀i = k, A(βi ) ∩ A(βk ) = ∅ (B.2)
⎪
⎪ association between nodes and clusters:
⎨
∀i, A(βi ) ∩ B(βi ) = ∅ (B.3) k 2
di j M−1 −1
⎪
⎪ ∃i, k, B(βi ) ∩ B(βk ) = ∅ (B.4) μi j = ( ( ) ) . (16)
⎪
⎪ dl j
⎪
⎪
⎪
⎪ ∀x j ∈ A(βi ), μi j = 1 (B.5) l=1
⎩ The greater ‘μi j ’, the higher the degree of association of the
∀x j ∈ B(βi ), μi j ∈ [0, 1] (B.6)
node to the cluster. The modified cluster centre calculation for
Property B.2 says that if a node ‘x j ’ belongs to the crisp RFCM is obtained by solving equation (13) concerning ‘v i ’:
lower approximation of cluster ‘βi ’, then ‘x j ’ is contained ⎧
in ‘βi ’ definitely. Property B.3 establishes that a node may ⎪
⎨ w × C + w̄ × D i f A(βi ) = ∅, B(βi ) = ∅
not be in both lower and boundary region of a cluster ‘βi ’. vi = C
RF
i f A(βi ) = ∅, B(βi ) = ∅ (17)
Property B.4 says that a node ‘x j ’ possibly belongs to ‘βi ’ and ⎪
⎩
D i f A(βi ) = ∅, B(βi ) = ∅,
potentially belongs to other clusters. Properties B.5 and B.6 are
1
of great importance in computing the objective function ‘ J R F ’. C = x j, (18)
The membership value of the nodes in the lower approximation |A − (βi )|
x j ∈A(βi )
is 1, while those in the boundary region are between 0 and 1. 1
RFCM partitions ‘n’ nodes into ‘k’ clusters by minimizing D= (μi j )Mx j , (19)
ni
the objective function: x j ∈B(βi )
⎧
⎪
⎨ w × A + w̄ × B i f A(βi ) = ∅, B(βi ) = ∅ where
JR F = A i f A(βi ) = ∅, B(βi ) = ∅ , (13) ni = (μi j )M, (20)
⎪
⎩
B i f A(βi ) = ∅, B(βi ) = ∅ x j ∈B(βi )
k
A= (μi j )M x j − v i
2
, (14) and ‘ A(βi ) ’ represents the cardinality of ‘A(βi ) ‘. The cluster
formation starts by randomly selecting ‘k’ nodes as the cluster
i=1 x j ∈A(βi )
centres. The membership value ‘μi j ’ is calculated by equation
k
(16) based on the relative distance of the node ‘x j ’ to the
(μi j )M x j − v i
2
B= , (15)
cluster centre ‘v i ’. Then the position of new cluster centres is
i=1 x j ∈B(βi )
calculated according to equation (17). The process will repeat
where ‘w’ and ‘w̄’ (w + w̄ = 1) indicate the relative until the position of cluster centres are stable. Fig. 7 shows
importance of the lower and boundary region, respectively. the pseudo-code for RFCM based cluster formation.
The fuzzifier ‘M’ rules the relative influence. The nodes in Since the overlapping area between clusters is necessary for
lower approximation are assigned a higher weight than the our model, it is a reasonable way to partition the nodes into

Authorized licensed use limited to: Aditya Engineering College. Downloaded on November 24,2021 at 05:14:04 UTC from IEEE Xplore. Restrictions apply.
23620 IEEE SENSORS JOURNAL, VOL. 21, NO. 20, OCTOBER 15, 2021

elimination process can be expressed as

E q,min
Ē q = , (26)
EC H
Nq,min
N̄q = , (27)
NC H
Fig. 8. Composition of a chromosome. Dq,min
D̄q = , (28)
RC
fuzzy clusters by RFCM. Moreover, the way of calculating and
updating the cluster centre position can meet our requirement where ‘E C H ’ and ‘NC H ’ are the remaining energy and node
for the distance between CCHs, so that CCHs can be evenly degree of CH, and ‘RC ’ is the communication radius, rep-
distributed to cover the event. After forming ‘k’ candidate resenting the farthest distance to the CH among CMs. The
clusters, the CCH selection based on GA is performed next. fitness function ‘ f (cq )’ is obtained by
3) GA Based CCH Selection: In this step, ‘k’ CCHs will f (cq ) = W1 Ē q + W2 N̄q + W3 D̄q
be selected from ‘k’ candidate clusters. Details are discussed E q,min Nq,min Dq,min
below. = W1 + W2 + W3 , (29)
EC H NC H RC
(1) Generate the initial population
We randomly select a node from each candidate cluster to where ‘W1 ’, ‘W2 ’ and ‘W3 ’ are the weights of the three
encode them into a chromosome. Nodes in the same cluster factors, and ‘W1 + W2 + W3 = 1’. The optimal CCH
only occupy one locus on the chromosome. The length of the set means that these nodes have more residual energy,
chromosome is ‘k’, same as the number of CCHs. The initial a smaller number of neighbour nodes, and a longer distance
population ‘P’ can be expressed as: to CH, which is obtained by maximizing the fitness function
‘ f (cq )’.
P = {c1 , c2 , . . . , c M }, (21) (3) Selection, crossover and mutation
The selection allows chromosomes with higher fitness to
where ‘M’ represents the population size, and ‘ci ’ is the i-th have a greater chance of survival. According to the roulette
chromosome in ‘P’, which can be further expressed as: wheel selection mechanism, the survival probability ‘ϕ’ is
given as:
ci = {ni,1 , ni,2 , . . . , ni,k } (22)
f (cq )
ϕ(cq ) =  , (30)
where ‘ni, j ’ represents the j-th gene locus (node) in the i-th f (cq )
M
chromosome. Assuming that ‘k = 8’, one of the chromosomes
can be expressed as shown in Fig. 8. where ‘ f (cq )’ is the fitness value of chromosome ‘cq ’.
(2) Construct fitness function We select ‘M’ chromosomes from ‘P’ according to equation
For the selection of CCH, we consider three factors: remain- (30) to form a selection population ‘P’:
ing energy, node degree, and distance to CH. The first two
P = ζ(P| f ) = {c1 , c2 , . . . , c M }. (31)
factors consider the energy and location requirements. The
function of third factor is to control the distance between CH Each chromosome in ‘P’ is randomly selected once accord-
and CCH to reduce the extra energy consumption caused by ing to probability ‘PC ’. The crossover process is performed by
cluster migration. exchanging the nodes on the same locus of two chromosomes.
We evaluate the overall performance of the node set based This also ensures that nodes in the same cluster always occupy
on three factors, so the selection of CCHs is transformed into the same gene locus. The selected ‘ci ’ and ‘c j ’ are the parent
a combinatorial optimization problem. We incorporate the idea chromosomes, and the cross operation produces two offspring
of the cask theory into the construction of the fitness function. chromosomes as follows:
The shortest plank determines the capacity of a barrel. The
same rule can also apply to the selection of CCHs. The ξ(ci , c j |PC ) ⇒ {ci , c j }. (32)
worst performance value of the three factors will determine Since the chromosome is composed of the node ID, node
the overall performance of the node set. According to this replacement is used to mutate chromosomes. We randomly
principle, we construct the fitness function as follows select two gene loci in the chromosome according to proba-
bility ‘PM ’, and replace them with other nodes in the same
E q,min = min E q, j , (23) cluster. The chromosomes generated after chromosomes ‘ci ’
j
Nq,min = min Nq, j , (24) and ‘c j ’ are mutated as follows:
j
Dq,min = min Dq, j , (25) M(ci |PM ) ⇒ ci , M(c j |PM ) ⇒ c j . (33)
j
Selection, crossover, and mutation operations will repeat
where ‘E q, j ’, ‘Nq, j ’, and ‘Dq, j ’ represent the remaining until one of the following conditions is met:
energy, node degree, and distance to CH of the j-th node
in chromosome ‘cq ’. ‘ E q,min ’, ‘ Nq,min ’ and ‘ Dq,min ’ are I ≥ Imax , (34)
the lowest value of these three factors. The data dimensional fbest ≥ T f , (35)

Authorized licensed use limited to: Aditya Engineering College. Downloaded on November 24,2021 at 05:14:04 UTC from IEEE Xplore. Restrictions apply.
QU et al.: EEDC PROTOCOL FOR EVENT MONITORING 23621

Fig. 10. The movement and expansion of the event triggers some CCHs
to become new CHs.

Fig. 9. Pseudo-code for GA based CCH selection.

where ‘I ’ is the number of iterations, ‘Imax ’ is the maximum

number of iterations, ‘ f best ’ is the best fitness value for each
iteration, and ‘T f ’ is the fitness threshold.
In step C, we divided the CCH clustering process into two
parts, cluster formation based on RFCM and CCH selection
based on GA. Firstly, RFCM uses the location information
of nodes to form overlapping clusters, which can ensure the
coverage quality of dynamically developing events. Secondly,
GA is used to find the optimal CCH set by minimizing the
fitness function (related to residual energy, node degree, and
distance to CH). By applying RFCM-GA, we can get the
optimal clustering result with a faster convergence rate. Fig. 9
shows the pseudo-code for GA based CCH selection.

D. Cluster Structure Adjustment

After the clustering process is completed, CH will send a
wake-up message to CCHs to notify them on standby. The
trigger condition for a CCH to become CH is that its distance
from the event centre is less than ‘R S ’, as shown in Fig. 10.
All the CCHs that have not been activated by the event will
remain active for a while, and then return to sleep.
In the case when the event keeps its shape while moving,
only the nearest CCH becomes the new CH and activates its
members. Then the previous CH and its associating members Fig. 11. Flow chart of EEDC.
will return to the dormant state. Therefore, in this process, only
one CH and its CMs are working in the network. Therefore, and the specific number of CHs will depend on the current
during the movement of the event, only one CH and its CMs shape and scope of the event area.
are working in the network. The complete flow chart of the proposed clustering protocol
In the case when the event is expanding, all the CCHs is shown in Fig. 11. The process of cluster formation, CCH
that sense the event will be activated as CHs. The remaining selection and cluster structure adjustment will be repeated
CCHs that are not affected by the event will stay dormant. around the new CH until the event disappears or the network
Consequently, there will be more than one CH in the network, dies.

Authorized licensed use limited to: Aditya Engineering College. Downloaded on November 24,2021 at 05:14:04 UTC from IEEE Xplore. Restrictions apply.
23622 IEEE SENSORS JOURNAL, VOL. 21, NO. 20, OCTOBER 15, 2021

TABLE II TABLE III

PARAMETER S ETTINGS OF E VENT D EVELOPMENT N ETWORK PARAMETERS

V. S IMULATION R ESULTS AND A NALYSIS

In this section, we evaluate the performance of EEDC via
MATLAB (version R2016b). A comparison among LEACH,
EEAOC, and ADEC is carried out under two different event
development ways. The basic parameter configuration of
The performance metrics include the network lifetime, average
EEDC, EEAOC and ADEC is the same as in LEACH. EEAOC
remaining energy and total energy consumption. In small-scale
is an event-driven clustering algorithm with a 2-logical cluster
WSNs, the network life is usually measured by FND (first
structure. The affiliation between nodes and different CHs
node dead). However, in large-scale networks, the death of a
will change according to the movement of the event. ADEC
single node cannot affect network connectivity and coverage.
is a clustering method with an annulus structure, which can
Therefore, we consider HND (half node dead) and LND (last
significantly reduce the energy consumption caused by the
node die) to measure network life. The specific parameters of
global network structure adjustment. Its data sensing process
the network are shown in Table III.
is independent of the event area.
Fig. 12 and Fig. 13 show the energy-efficiency performance
We assume that the event occurs in a circular region
of each protocol in case#1 and case#2, respectively. It can
and the event development is modelled by vector
be seen that EEDC outperforms all the other protocols in
‘[x e , ye , v e , e , re , vr ]’, where ‘(x e , ye )’ is the coordinates for
both cases based on HND and LND. Under case#1, EEDC
the event centre, ‘v e ’ is the movement speed of the event,
is improved by 78%, 31% and 19% against LEACH, ADEC
‘ e ’ is the movement direction angle of the event, ‘re ’ is the
and EEAOC for HND. For LND, EEDC is improved by 67%,
radius of the event area, ‘vr ’ is the expansion speed of the
9% and 25%, respectively. The result in (b) and (c) indicate
event. The analytical expression of the event area is:
that EEDC consumes less energy and retains the most residual
(x − x e )2 + (y − ye )2 ≤ re2 , 0 ≤ x e , ye ≤ M (36) energy than other protocols under two cases. This is because
the dynamic clustering algorithm RFCM-GA we proposed in
To demonstrate the performances of the protocols in differ- EEDC can avoid the enormous energy consumption caused
ent environments, two event development cases (case#1 and by frequent information exchange and cluster reconstruction
case#2) are considered under 800 nodes deployed in the in most static clustering algorithms.
network. In case#1, the event occurs in a local area and changes
Case#1: the event follows the movement of an individual with its location while maintaining its shape. Taking advantage
target. In this situation, the event moves randomly while of the event-driven data transmission scheme, the number of
maintaining its shape. ‘(x e , ye )’ is randomly generated, ‘v e ’ nodes is almost unchanged in EEAOC and EEDC. There-
and ‘re ’ are constant values, ‘ e ’ is a variable that varies fore, EEAOC’s performance for HND is even better than
randomly over time, and ‘vr = 0’. ADEC. However, in the later stage of the algorithm execution,
Case#2: the development of the event follows a continuous the maintenance of the static overlapping cluster structure
changing area. In this situation, the event moves while extend- brings heavy communication overhead in EEAOC. ADEC
ing its area. ‘(x e , ye )’ is randomly generated, ‘v e ’ is a constant adopts a lightweight intra-annulus cluster method, and the
value, ‘re ’ is a variable set as ‘re = r0 + vr × t’, where ‘r0 ’ cost of maintaining the cluster structure is small. So the
is the initial value of ‘re ’, ‘vr ’ is a constant value so that the performance of EEAOC for LND is surpassed by ADEC even-
event expands with a uniform speed. The specific parameters tually. However, no event-driven data transmission scheme is
of event development are shown in Table II. designed in ADEC so that more redundant nodes are activated
The simulations were conducted with nodes deployed over during the movement of the event, causing additional energy
a region of 500 × 500 m 2 . The initial energy of nodes is consumption. In comparison, the event-driven dynamic clus-
fixed at 0.1J. Sink is located at the centre of the network, tering scheme in EEDC can achieve higher energy efficiency
and its coordinate is (250,250). The deployment of nodes is than LEACH, EEAOC and ADEC.
different for each simulation run, and each result is averaged In case#2, the event changes with its location and size.
over ten runs. The time interval of a single round is set to 10s. Due to the continuous expansion of the event area, there will
We set ‘a’ and ‘b’ in equation (4) to 0.4 and 0.6, respectively. be many more activated nodes in the network than case#1.

Authorized licensed use limited to: Aditya Engineering College. Downloaded on November 24,2021 at 05:14:04 UTC from IEEE Xplore. Restrictions apply.
QU et al.: EEDC PROTOCOL FOR EVENT MONITORING 23623

Fig. 12. Performance of each protocols under case#1.

Since no event-driven data transmission scheme is designed in

LEACH and ADEC, their performance dropped significantly Fig. 13. Performance of each protocols under case#2.
under case#2. By comparing Fig. 12 and Fig. 13, LEACH’s
performance under case#2 drops significantly, the HND and surpass EEAOC. In comparison, due to the event-driven cluster
LND drop by 51% and 64%. The HND of ADEC also drops by formation scheme, EEAOC and EEDC are slightly affected.
17%. Furthermore, compared to case#1, the energy-efficiency However, the clustering process in EEAOC is executed only
performance of ADEC takes about 300 more rounds to once, and the CHs are rotated multiple times in the later stage

Authorized licensed use limited to: Aditya Engineering College. Downloaded on November 24,2021 at 05:14:04 UTC from IEEE Xplore. Restrictions apply.
23624 IEEE SENSORS JOURNAL, VOL. 21, NO. 20, OCTOBER 15, 2021

the event, avoiding unstable structure and the appearance of

isolated nodes.
The comparison of HND and LND of different node num-
bers for each protocol in case#1 is shown in Fig. 14. Same as
the previous analysis, EEAOC and ADEC perform better than
each other in HND and LND, respectively. It can be seen that
the proposed EEDC has superior performance in both HND
and LND with different numbers of nodes.
Fig. 15 shows the comparison of the fitness value in
the process of clustering using RFCM-GA and classical
GA. According to equation (27), the fitness value is the
performance evaluation of the CCH set. It can be seen from
Fig. 15 that RFCM-GA can find the optimal solution at a
faster convergence rate than GA. This is because RFCM-GA
divides clustering into two parts: cluster formation and CCH
selection, and these two parts take advantage of RFCM and
GA, respectively. RFCM can quickly determine the allocation
of nodes according to their locations, forming a cluster
structure. After that, GA is used to select CCHs from each
cluster to form chromosomes and searches for the best CCH
set in a parallel way. In comparison, the classic GA is slow to
obtain clustering results, and it cannot generate the overlapping
cluster structure required by EEDC. This is because the classic
GA just performs a random global search in all nodes.

VI. C ONCLUSION
This paper proposes a new dynamic clustering algorithm—
EEDC, used for event monitoring in large-scale WSN.
In EEDC, an RFCM-GA based CCH selection algorithm is
proposed. In RFCM-GA, the clustering process is divided
into two parts, cluster formation based on RFCM and CCH
selection based on improved GA. We firstly use RFCM to
form the overlapping clusters, and then the improved GA is
used to find the most suitable CCH in each cluster. A group of
Fig. 14. Comparison of HND and LND in the network under different
number of nodes. CCHs are obtained by minimizing the fitness function. Finally,
as the event develops, the cluster structure will be adjusted,
and some CCHs will become new CHs. This mechanism can
make the cluster structure of the network change dynamically
with the state of the event. The simulation results show that
EEDC achieves better energy efficiency than other clustering
protocols under two different event development cases. It is
concluded that EEDC is more suitable for event monitoring
applications in large-scale WSNs compared with existing
protocols.

R EFERENCES
[1] P. Prasad, “Recent trend in wireless sensor network and its applications:
A survey,” Sensor Rev., vol. 35, no. 2, pp. 229–236, Mar. 2015.
[2] L. K. Ketshabetswe, A. M. Zungeru, M. Mangwala, J. M. Chuma,
and B. Sigweni, “Communication protocols for wireless sensor net-
works: A survey and comparison,” Heliyon, vol. 5, no. 5, May 2019,
Art. no. e01591.
[3] R. Elhabyan, W. Shi, and M. St-Hilaire, “Coverage protocols for wireless
Fig. 15. Comparison of fitness values in the clustering process using sensor networks: Review and future directions,” J. Commun. Netw.,
RFCM-GA and classic GA. vol. 21, no. 1, pp. 45–60, Feb. 2019.
[4] H. Yang, Z. Feng, X. Teng, L. Guan, H. Hu, and M. Wu, “Three-
of the algorithm execution, resulting in an unstable cluster dimensional printing of high-mass loading electrodes for energy storage
structure and many isolated nodes. Compared with EEAOC, applications,” InfoMat, vol. 3, no. 6, pp. 631–647, Jun. 2021.
[5] O. Demigha, W. K. Hidouci, and T. Ahmed, “On energy efficiency in
EEDC has a more flexible and straightforward dynamic cluster collaborative target tracking in wireless sensor network: A review,” IEEE
structure, which selects CH according to the current state of Commun. Surveys Tuts., vol. 15, no. 3, pp. 1210–1222, 3rd Quart., 2013.

Authorized licensed use limited to: Aditya Engineering College. Downloaded on November 24,2021 at 05:14:04 UTC from IEEE Xplore. Restrictions apply.
QU et al.: EEDC PROTOCOL FOR EVENT MONITORING 23625

[6] M. Naderan, M. Dehghan, H. Pedram, and V. Hakami, “Survey of Zhiyi Qu received the B.S. degree from Jilin Uni-
mobile object tracking protocols in wireless sensor networks: A network- versity in 2017. He is currently pursuing the Ph.D.
centric perspective,” Int. J. Ad Hoc Ubiquitous Comput., vol. 11, no. 1, degree with Shanghai Institute of Microsystem
pp. 34–63, Sep. 2012. and Information Technology, Chinese Academy
[7] M. Nighot, A. Ghatol, and V. Thakare, “Distributed energy efficient of Sciences, and also with the School of Elec-
tracking in hybrid wireless sensor network (DEETH),” Int. J. Commun. tronic, Electrical and Communication Engineer-
Syst., vol. 31, no. 5, p. e3498, Mar. 2018. ing, University of Chinese Academy of Sciences.
[8] C. Lersteau, A. Rossi, and M. Sevaux, “Minimum energy target tracking His research interests include topology control
with coverage guarantee in wireless sensor networks,” Eur. J. Oper. Res., and protocol design of wireless sensor networks.
vol. 265, no. 3, pp. 882–894, Mar. 2018.
[9] S. Bhowmik and C. Giri, “Convoy tree based fuzzy target tracking in
wireless sensor network,” Int. J. Sens. Netw., vol. 24, no. 4, pp. 1–9,
Dec. 2017. Huihui Xu received the B.S. degree from
[10] X. Yu and J. Liang, “Genetic fuzzy tree based node moving strategy of Wuhan University in 2019. She is currently pur-
target tracking in multimodal wireless sensor network,” IEEE Access, suing the Ph.D. degree with Shanghai Insti-
vol. 6, pp. 25764–25772, May 2018. tute of Microsystem and Information Technology,
[11] X. Liu, “A survey on clustering routing protocols in wireless sensor Chinese Academy of Sciences, and also with
networks,” Sensors, vol. 12, no. 8, pp. 11113–11153, Aug. 2012. the School of Electronic, Electrical and Com-
[12] P. Nayak and A. Devulapalli, “A fuzzy logic-based clustering algorithm munication Engineering, University of Chinese
for WSN to extend the network lifetime,” IEEE Sensors J., vol. 16, Academy of Sciences. Her research interests
no. 1, pp. 137–144, Jan. 2016. include resource allocation, and algorithm design
[13] T. Ahmad and A. M. Abbas, “EEAC: An energy efficient adaptive cluster in event-driven wireless sensor networks.
based target tracking in wireless sensor networks,” J. Interdiscipl. Math.,
vol. 23, no. 2, pp. 379–392, May 2020.
[14] P. Lalwani, S. Das, H. Banka, and C. Kumar, “CRHS: Clustering and
routing in wireless sensor networks using harmony search algorithm,” Xue Zhao received the B.S. degree from
Neural Comput. Appl., vol. 30, no. 2, pp. 639–659, Jul. 2018. Shandong University in 2019. She is currently
[15] X. Zhao, S. Ren, H. Quan, and Q. Gao, “Routing protocol for het- pursuing the Ph.D. degree with Shanghai Insti-
erogeneous wireless sensor networks based on a modified grey wolf tute of Microsystem and Information Technology,
optimizer,” Sensors, vol. 20, no. 3, p. 820, Feb. 2020. Chinese Academy of Sciences, and also with
[16] S. M. Bozorgi, M. R. Hajiabadi, A. A. R. Hosseinabadi, and the School of Electronic, Electrical and Com-
A. K. Sangaiah, “Clustering based on whale optimization algorithm for munication Engineering, University of Chinese
IoT over wireless nodes,” Soft Comput., vol. 25, no. 7, pp. 5663–5682, Academy of Sciences. Her research interests
Jan. 2021. include wireless sensor networks and network
[17] G. P. Gupta and S. Jha, “Integrated clustering and routing protocol protocols.
for wireless sensor networks using cuckoo and harmony search based
metaheuristic techniques,” Eng. Appl. Artif. Intell., vol. 68, pp. 101–109,
Feb. 2018.
[18] V. K. Chawra and G. P. Gupta, “Hybrid meta-heuristic techniques Hongying Tang received the Ph.D. degree
based efficient charging scheduling scheme for multiple mobile wireless from Shanghai Jiao Tong University (SJTU),
chargers based wireless rechargeable sensor networks,” Peer-Peer Netw. China, in 2015. She is currently an Engineer
Appl., vol. 14, no. 3, pp. 1303–1315, Jan. 2021. with the Science and Technology on Microsystem
[19] G. P. Gupta and B. Saha, “Load balanced clustering scheme using Laboratory, Shanghai Institute of Microsystem
hybrid metaheuristic technique for mobile sink based wireless sensor and Information Technology, Chinese Academy
networks,” J. Ambient Intell. Humanized Comput., early access, of Sciences. Her research interests include
nos. 14–15, Apr. 2020, doi: 10.1007/s12652-020-01909-z. unmanned aerial vehicle (UAV) communications,
[20] M. A. Rahmadhani, L. V. Yovita, and R. Mayasari, “Energy consumption linear transceiver design, and wireless sensor
and packet loss analysis of LEACH routing protocol on WSN over networks.
DTN,” in Proc. 4th Int. Conf. Wireless Telematics (ICWT), Jul. 2018,
pp. 1–5.
[21] Y. Hu, Y. Niu, J. Lam, and Z. Shu, “An energy-efficient adaptive
Jiang Wang received the B.S. and M.S.
overlapping clustering method for dynamic continuous monitoring in
degrees from the University of Electronic Sci-
WSNs,” IEEE Sensors J., vol. 17, no. 3, pp. 834–847, Feb. 2017.
ence and Technology of China, Chengdu, China,
[22] Z. Zhou and Y. Niu, “An energy efficient clustering algorithm based
in 1998 and 2001, respectively, and the Ph.D.
on annulus division applied in wireless sensor networks,” Wireless Pers.
degree from Southeast University, Nanjing,
Commun., vol. 115, no. 3, pp. 2229–2241, Aug. 2020.
China, in 2006. He is currently a Professor
[23] M. Wei, X. He, and L. Xu, “Dynamic clustering target tracking based
and a Master’s Supervisor with Shanghai Insti-
on energy optimization in wireless sensor networks,” J. Comput. Appl.,
tute of Microsystem and Information Technology,
vol. 37, no. 6, pp. 1539–1544, Oct. 2017.
Chinese Academy of Sciences. His research
[24] J. Feng, B. Lian, and H. Zhao, “Coordinated and adaptive information
interests include single frequency networking for
collecting in target tracking wireless sensor networks,” IEEE Sensors J.,
future wireless communication systems, radio
vol. 15, no. 6, pp. 3436–3445, Jun. 2015.
resource management/cross-layer optimization for mobile self-organized
[25] Y. Liu, J.-S. Fu, and Z. Zhang, “K-nearest neighbors tracking in wire-
networks, and wireless sensor networks.
less sensor networks with coverage holes,” Pers. Ubiquitous Comput.,
vol. 20, no. 3, pp. 431–446, Jun. 2016.
[26] M. Akter, M. O. Rahman, M. N. Islam, M. M. Hassan, A. Alsanad, and Baoqing Li received the Ph.D. degree from
A. K. Sangaiah, “Energy-efficient tracking and localization of objects the State Key Laboratory of Transducer Tech-
in wireless sensor networks,” IEEE Access, vol. 6, pp. 17165–17177, nology, Shanghai Institute of Metallurgy, Chi-
Feb. 2018. nese Academy of Sciences, Shanghai, China,
[27] Z. Wang, W. Lou, Z. Wang, J. Ma, and H. Chen, “A hybrid cluster-based in 2000. He is currently a Professor and a Ph.D.
target tracking protocol for wireless sensor networks,” Int. J. Distrib. Supervisor with Shanghai Institute of Microsys-
Sens. Netw., vol. 2013, pp. 245–253, Mar. 2013. tem and Information Technology, Chinese Acad-
[28] Z. Liu, Q. Zheng, L. Xue, and X. Guan, “A distributed energy- emy of Sciences. His research interest includes
efficient clustering algorithm with improved coverage in wireless sensor application of wireless sensor networks.
networks,” Future Generat. Comput. Syst., vol. 28, no. 5, pp. 780–790,
May 2012.

Authorized licensed use limited to: Aditya Engineering College. Downloaded on November 24,2021 at 05:14:04 UTC from IEEE Xplore. Restrictions apply.