This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2632198, IEEE
Sensors Journal

An Energy-Efficient Adaptive Overlapping

Clustering Method for Dynamic Continuous
Monitoring in WSNs
Yuan Hu, Yugang Niu, James Lam, Fellow, IEEE, and Zhan Shu, Member, IEEE

 even 90% of total initial energy of the network is left [9]. Hence,
Abstract—Clustering is a key technique to improve energy it is a major concern of how to reduce and balance energy
efficiency in wireless sensor networks (WSNs). In continuous consumption in WSNs [6]–[11].
monitoring applications, the clusters should be formed
dynamically according to the event development for A. Clustering in WSNs
energy-efficient data gathering. In this paper, an energy-efficient Clustering is an efficient approach to prolong the network
adaptive overlapping clustering method (EEAOC) is proposed in
lifetime for large-scale WSNs [12]. In a cluster-based WSN, the
WSNs for continuous monitoring applications. In EEAOC, a
2-logical-coverage overlapping clustering topology is established cluster head (CH) is responsible for collecting data from the
such that the adjacent sensors in the event area can be grouped cluster members (CMs) in each cluster. Then the CH
into the same cluster for data fusion and the cluster migration aggregates the collected data and sent it to the sink directly or
operation can be processed without changing the overlapping via relay nodes. The clustering methods can prolong network
structure among clusters. Moreover, to further reduce energy lifetime, balance energy consumption and provide scalability of
consumption, a hybrid data reporting strategy that switches
the network [12].
between time-driven and event-driven schemes is introduced
based on the QoS requirements in continuous monitoring In general, the operations in clustering protocols are divided
applications. Simulation results show that EEAOC achieves a into four steps: information collection, CH selection, cluster
longer network lifetime cycle. formation and data transmission [12] [13]. CHs consume much
more energy than ordinary nodes due to extra data collection
Index Terms—Clustering algorithm, event-driven clustering, and fusion tasks. Hence, many criteria related to, for example,
event tracking, wireless sensor networks residual energy, node density, distance and location have been
proposed for selecting proper CHs to balance the energy
consumption [14].
I. INTRODUCTION
B. Types of Application in WSNs
W sensor networks (WSNs) consist of a large
IRELESS
amount of sensors to perform distributed sensing tasks,
which are widely used in health-care monitoring, wildlife
WSNs are application-oriented networks, and their typical
applications can be derived according to the monitored physical
tracking, environmental monitoring and industrial detection, process including event-based applications, continuous
for instance [1]–[4]. As sensors in WSNs are powered by monitoring applications, query-driven applications, and hybrid
battery and are usually deployed over a large unpopulated or applications [15]. In event-based applications, the reading in a
sometimes hostile region, it is impossible to recharge or replace particular area will be used only when an event is detected. In
them in most cases [1]–[6]. In addition, the characteristics of query-driven applications, each node transmits its gathering
multi-hop and many-to-one communication mode often result data to the sink when a query message is received from the sink.
in unbalanced energy depletion [6] [7]. Nodes at hotspots (near In event-based and query-driven applications, the network load
the sink) tend to exhaust the battery energy quickly, leading to is light most of the time, and will become heavy only when the
the 'Funnel Effect' [8]. No data can be transmitted to the sink data reporting condition is triggered [16] [17]. On the contrary,
in continuous monitoring applications, sensors send data to the
sink periodically [18]. In hybrid applications, burst data may
This work is supported by National Natural Science Foundation (NNSF) be generated in addition to periodically sensed data.
from China (61273073, 61374107), HKU CRCG 201411159139, and Cheung
Kong Scholars Programme. C. Motivations
Yugang Niu and Yuan Hu are with Key Laboratory of Advanced Control
and Optimization for Chemical Process of Ministry of Education, East China In some continuous monitoring applications, the detected
University of Science and Technology, Shanghai, China (e-mail: event may spread to a larger region or/and may change its
[email protected]; [email protected]).
James Lam is with Department of Mechanical Engineering, University of
physical properties [19]. For instance, in the forest fire alarm
Hong Kong, Hong Kong (e-mail: [email protected]). system, fire may spread from one region to another, or extend to
Zhan Shu is with Electro-Mechanical Engineering Group, Faculty of a larger region. In a safety supervision system of a chemical
Engineering and the Environment, University of Southampton，UK (e-mail: factory, the leaking gas cloud may dispense to the atmosphere.
[email protected]).

1530-437X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2632198, IEEE
Sensors Journal

In a rainstorm and flood warning system, the flooded area may are sent to the same CHs for fusion. Once the event spreads to
change according to seasonal and climate variations. In those neighboring regions, a collaborative CH re-adjustment and
cases, the detected events change gradually with their size and cluster migration technique is used to ensure effective cluster
location. formation associated with a 2-logical overlapping clustering.
To balance the intra-cluster energy consumption, CH rotation is
operated without changing the overlapping structure among
clusters when the CHs in the event area have depleted their
energy.
(2) Aiming at prolonging the network lifetime of WSNs for
continuous monitoring applications under some guaranteed
monitoring accuracy, a hybrid data reporting strategy that
switches between time-driven and event-driven schemes is
(a) Event migration (b) Event extension introduced.
Fig. 1: Event development in a cluster-based WSN The rest of this paper is organized as follows. In Section 2,
In traditional clustering schemes, clusters are formed without related works are introduced, and critical problems in these
considering event development. Consequently, adjacent nodes schemes are described. In Section 3, we present the network
in the same event area may belong to different clusters, while model and some definitions. The details of our protocol are
nodes outside the event area may be forced to send data to CHs. proposed in Section 4. In Section 5, we analyze the
As shown in Fig. 1(a)–(b), in a cluster-based WSN, an event performance of EEAOC and simulate it in comparison with
occurs in a small region and the event area moves or extends existing protocols. Finally, the conclusion is given in Section 6.
gradually, the event area will be covered by different group of
sensors as time progresses. In some data gathering rounds, II. RELATED WORKS
improper cluster formation may appear due to event In this section, we classify the main commonly used
development. As an example shown in Fig. 2, the relevant data clustering schemes based on intra-cluster topology (single-hop,
sensed from the same event area is transmitted to three different multi-hop) and data reporting schemes (time-driven,
CHs for fusion, and nodes that do not detect the event in these event-driven).
three clusters also transmit their sensed data to CHs.
Furthermore, most of the existing clustering schemes dismiss A. Single-hop and Multi-hop Clustering Schemes
all clusters to re-format new ones after certain transmission According to the intra-cluster topology, clustering
rounds. This periodic global cluster re-adjusting scheme leads techniques can be classified into single-hop and multi-hop
to huge communication overhead. Moreover, most approaches [13]. Single-hop clustering schemes as in [20]–[22]
event-driven clustering schemes take the whole event area as assign CMs to communicate with their corresponding CHs
one cluster for data gathering. The assumption is not practical directly. LEACH [20] is a classic single-hop clustering protocol
when the event is detected in a large area. Therefore, how to that assigns CHs according to a predefined probability ensuring
design more energy-efficient data reporting schemes in that all nodes in the network become CHs with the same chance.
cluster-based WSNs for dynamic continuous monitoring Each CM sends data to its CH periodically after cluster
applications still needs to be further investigated. formation. In DSBCA [21], the network is divided into unequal
single-hop clusters for load-balancing based on the node
density and the distance from the sink. A dynamic single-hop
clustering DECS [22] was proposed for heterogeneous WSNs,
wherein CHs are selected through the forecasted network
residual energy.
In large-scale WSNs, multi-hop intra-cluster communication
model is adopted to avoid long distance transmission. In
multi-hop clustering schemes as in [23]–[25], there are relay
nodes in each cluster to maintain multi-hop sensors-to-CHs
connectivity. The hierarchical cluster-based routing algorithm
in [23] constructs a multi-hop routing tree for inter-cluster
Fig. 2: Improper cluster formation
communication, in which multiple criteria are used to select
Motivated by above discussion, an energy-efficient adaptive CHs and an adjustment degree is set to modify intra-cluster and
overlapping clustering method (EEAOC) is proposed in this inter-cluster energy consumption. HEHC [24] is a
paper, which focuses on continuous monitoring with possibly heterogeneous-aware multi-level clustering scheme, which
event fluctuation. The key features of EEAOC algorithm are adds extra cluster layers and super CHs for obtaining better
summarized as follows: energy utilization. The objective of DEECIC [25] is to preserve
(1) An energy-efficient 2-logical-coverage overlapping the coverage and achieve efficient energy consumption by
clustering is designed to adapt to event fluctuation, which adopting 2-hop intra-cluster communication mode. Generally,
ensures that data collected by adjacent sensors in the event area

1530-437X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2632198, IEEE
Sensors Journal

multi-hop clustering schemes are more complicated for tracking algorithms, only one cluster is active for tracking the
building and maintaining the network topology. moving event at each instant, leading to poor performance
However, both single-hop and multi-hop clusters are formed when the moving event lies in several clusters [33]. Moreover,
independent of event developments, which are not suitable for most target tracking algorithms consider how to track the
dynamic continuous monitoring applications. As shown in Fig. moving target without considering the development of the
3, an event is detected in a small region in the beginning shown event area, and the target motion is not considered in the
as Area 1. All the sensors in the event area can be grouped into event-based state estimation algorithms.
one single-hop cluster. Gradually, the event may spread to a In most continuous monitoring applications, QoS
larger area shown as Area 2. The sensors in Area 2 can be requirement varies according to the changes of monitoring
grouped into one multi-hop cluster. Finally, the event may process. For example, in a temperature monitoring application
spread to an even larger region shown as Area 3. It is for vegetable shed automatic control system, when temperature
inappropriate to group the sensors in Area 3 into just one cluster. is within the normal range (10 ℃ -30 ℃ , for instance), the
As a result, the previous one multi-hop cluster formation should vegetables are growing under an appropriate growth condition.
be dismissed and the sensors in Area 3 should be grouped to Farmers care more about the tread of the temperature variation
multiple clusters. Hence, a dynamic clustering scheme should rather than the exact value of temperature. When the
be designed to adapt to the event development. temperature exceeds the normal range, the values at each
sampling instant should be sent to users. To further reduce the
energy consumption, some trade-off can be made between
improving the energy efficiency and guaranteeing the detection
performance [18]. Therefore, in the beginning, only a few data
should be sent to CH to judge the event development tendency,
the event-driven data reporting scheme can be used to save
energy. Once an abnormal phenomenon is detected, data
reporting switches to the time-driven scheme for accurate
Fig. 3 Event area development measurement. After the abnormal phenomenon disappears, data
B. Time-driven and Event-driven Clustering Schemes transmission will switch back to the event-driven scheme.
In EEAOC, a novel 2-logical-coverage cluster formation
The aforementioned clustering schemes in [20]–[25] are
scheme and cluster migration scheme are designed to form and
based on time-driven data reporting schemes. The nodes sense
activate proper clusters according to event development. The
and send data to the sink periodically. However, in some
overlapping cluster formation operation is illustrated in Fig. 4.
applications, the event may not be detected in the whole area.
Firstly, non-overlap clusters are formed similar to traditional
Event-driven clustering protocols are designed to further
cluster formation schemes, as in Fig. 4(a). Secondly, each node
improve energy efficiency, in which data collection depends on
chooses the next-closest CH to form the 2-logical-coverage
the occurrence of events [13]. The TEEN [26] and APTEEN
overlapping clusters, as in Fig. 4(b).
[27] protocols utilize two thresholds (Hard Threshold and Soft
Threshold) to reduce the number of transmission messages.
Hard Threshold is a sensed attribute for the node to send data to
its CH while Soft Threshold is a small change of the sensed
attribute that triggers the node to switch on the transmission
mode. By introducing the inverse path for event notifications
and the shortest path for the delivery of events, the HPEQ
protocol [28] speeds up new subscriptions in the sensor region
under low transmission latency. However, to the best of the
authors’ knowledge, none of the event-driven clustering
(a)Non-overlap cluster formation (b) Overlap cluster formation
schemes consider how to group sensors in the event area into Fig.4: Cluster formation with 2-logical coverage overlaps
clusters. Most of them assumed that the sensors in event area
can be organized into one cluster or the detected phenomenon III. PRELIMINARIES
occurs intermittently in the fully monitored area. Both of the
assumptions are not suitable for continuous monitoring A. Network Model
applications when the detected event changes gradually with We consider an network with N sensor nodes uniformly
their size and location. distributed in a M×M square monitoring area with high density,
Many target tracking or estimation algorithms as reported in which satisfies the following assumptions:
[29] [30] were proposed for monitoring or tracking moving (1) There is a sink located in the middle of the monitoring
events. Event-driven clustering and event-based sampling are area. The sensors and the sink are stationary after deployment.
effective approaches used to balance the energy efficiency and (2) The sensors are initially charged with identical energy,
desired monitoring accuracy in the target tracking or estimation and not equipped with any GPS-like hardware. In addition, the
algorithms [31]–[34]. However, in most of cluster-based target sink does not have energy constraints.

1530-437X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2632198, IEEE
Sensors Journal

(3) The links are symmetric (two nodes can communicate BN (i, j )  { x | x  S (i ) & x  S ( j )}, (3)
with each other using the same power). Nodes are time where S(i) and S(j) are the sets of sensors in cluster i and cluster
synchronous with the help of MAC protocols. j, respectively.
(4) Power control is allowed to vary the transmission power To detect event development, the whole network is divided
based on the distance between two nodes. The conflict control into 2-logical-coverage subnets. That is, each node belongs to
and channel selection would be dealt by using MAC protocols. two adjacent clusters simultaneously. The role of boundary
(5) Each node is assigned with a unique ID via ID nodes includes two aspects: as 1-hop forwarding nodes when
assignment algorithms for cluster-based networks as in [25]. the single hop clusters are combined into 2-hop clusters; as
By using the radio model in [20], both the free space and beacon nodes to indicate the moving direction of the event.
multi-path channel energy consumption models are used Definition 2 (Overlapping degree). The overlap degree for
according to the distance between transmitters and receivers, as cluster i is defined as
shown in Fig. 5. OD (i )  ni / ki , (4)
ET ER where ni is the total number of nodes in cluster i, ki is the
l l number of overlapping areas in cluster i.
l  Eelec l   fs  d n d l  Eelec
In the EEAOC protocol, clusters are divided into the same
Fig.5: Energy Consumption Model number of overlapping regions, which means that ki is a
constant (denoted as k0). The value of k0 is dependent on the
The energy dissipation for transmitting l packets over moving speed of the detected event. When an event moves at a
distance d is high speed, the cluster migration should be operated at high
 l ( Eelec   fs  d 2 ), d  d 0 frequency to avoid improper cluster formation. The value of k0
ET   . (1)
4 should be greater such that the cluster migration is more
l ( Eelec   mp  d ), d  d 0
energy-efficient and flexible to adapt to the high event
The energy dissipation for receiving l packets is development rate.
E R  l  E elec , (2) Definition 3 (Cluster size). The cluster size is defined as the
where Eelec is the energy spent per bit to run the circuitry of total number of sensors in the cluster.
each node,  fs and  mp are the energy required by amplifier in
IV. THE EEAOC PROTOCOL
two transmission energy modes, d 0   mp /  is the
fs EEAOC operates in several epochs, and each epoch contains
threshold distance for two energy models. E DA denotes the the cluster set-up phase and the steady data transmission phase.
energy consumption for aggregating one packet, and EDA  ET . An epoch of EEAOC includes five steps (information
collection, CH selection, cluster formation, data transmission,
The notations and definition in EEAOC are listed in Table 1. CH re-adjustment and cluster migration).
TABLE 1
NOTATION AND DEFINITION IN THE EEAOC PROTOCOL A. Information Collection Step
Notation Definition
In this step, local and global information of the network will
N Total number of sensors in the network
be obtained. First, the sink floods a ‘Hello message’ to the
S(i) Set of sensors in cluster i
network. Based on the Received Signal Strength Indication
ni Number of sensors in cluster i (RSSI), the Euclidean distance between the ordinary nodes and
CHi CH in cluster i the sink can be estimated [21]. Second, each node broadcasts a
ki Number of overlapping regions in cluster i ‘Hand-shaking message’ to its neighbors within Rs which
RS Communication radius of sensors contains the information of node ID and residual energy.
RC CH competing radius
B. Cluster Head Selection Step
Ere(i) Residual energy of sensor i
Total number of CHs in the network
For load balancing, CHs should be distributed uniformly in
KCH
Number of sensors in the CH competing area
the monitoring area. A competing radius RC is set to limit the
m0
broadcasting range of ‘CH competing message.’ In this work,
n0 Number of sensors in each cluster
we set 2 RC  RS for ensuring network connectivity, and set
f Number of data frames sent from a CH to the sink in a round
u
2 RC  d 0 for avoiding long distance intra-cluster
m i Number of sensors in the uth overlapping region for cluster i
communication. The exact value of RC is determined by the
rp (t ) Reading sensed by the pth sensor in the network at instant t QoS requirements of users and the features of sensors in the
network.
B. Definitions
To save the competing overhead and avoid collision, a
Definition 1 (Boundary nodes). A node x is a boundary node, if broadcast delay related to nodes’ residual energy is introduced.
x is located in the overlapping area between cluster i and cluster The broadcast delay of node i is given as
j. The set of boundary nodes in the overlapping area between
the two clusters i and j is Ti  e1/ Ere (i) , (5)

1530-437X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2632198, IEEE
Sensors Journal

where Ere(i) is the residual energy of node i. If a node receives As the number of nodes is m0 when the cluster radius is RC,
‘competing messages’ before its waiting time expires, it will the number of sensors in each cluster is less than m0 in Case I.
give up the competition operation. Otherwise, it will broadcast Case II: the Euclidean distance between any pair of closest
the ‘competing message’ to announce itself as a CH in RC. Due CHs is farther than 2RC and there exists an ordinary node with
to the limited competing radius and broadcast delay, nodes with distance from both CHs greater than RC.
higher residual energy have higher chance to serve as CHs, and Since the node belongs to neither of the two clusters, it will
the selected CHs are uniformly distributed in the network. announce itself as a CH. Therefore, the two CHs are not the
Suppose that there are KCH CHs selected in the network. In closest pair, which contradicts the above assumption.
the ideal situation of CH distribution, the sensors in the network Case III: the Euclidean distance between any pair of two
are covered by KCH non-overlapping circles with radius RC. Let closest CHs is farther than 2RC. No nodes are with distance
u denotes the smallest integer greater than u. The number of from both CHs greater than RC.
sensors in the circles (denoted as m0) can be estimated as In this case, the maximum Euclidean distance from a node to
its closest CH is less than RC, the number of cluster members
  R2  (6)
m0   N  2C  . for each cluster is less than m0.
 M 
Therefore, the cluster size for each cluster is less than m0.
As a result, the optimal number of CHs can be estimated as In the second time interval [T1, T2], overlapping regions are
opt  M2  (7) formed among clusters. All nodes send ‘overlapping
K CH  N / m0   2  .
  RC 
cluster-join messages’ to their next-closest CHs (slave CHs)
To format the 2-logical-coverage clusters, each node belongs to which contain less than n0 CMs. The slave-CHs will broadcast
‘overlapping reply messages’ to confirm the overlapping
two CHs simultaneously. Since each cluster has the same cluster
cluster formation. In the end, each cluster contains n0 CMs.
size, the cluster size under ideal CH distribution is a constant
(denoted as n0), which can be estimated as The number of boundary nodes in every overlapping area is
obtained by the corresponding CHs, and the adjacent CHs are
 2N    R2  (8) connected by the boundary nodes.
n0   opt   2  N  2C   2m0 .
K
 CH   M  Remark 2: In some cases, the total number of sensors N may
Substituting (8) into (4), we obtain not be a multiple of n0, which means that the network cannot be
n n   NRC2  divided into clusters with equal cluster size. Anyhow, KCH -1
OD (i )  i  0  2  2 
. (9)
ki k 0  k0 M  overlapping clusters will be grouped with cluster size n0 and the
It is seen from (9) that, for a given k0, the overlapping degree is last cluster will be grouped with cluster size N  ( KCH  1)n0 .
a constant (denoted as OD0) and approximately proportional to In the last time interval [T2, T3], equal-sized overlapping
the square of the competing radius. regions are divided. If an overlapping region has more than
Once more than K CH opt opt
( K CH  K CH ) nodes are selected as CHs, OD0 boundary nodes, its slave-CH will send an ‘overlapping
opt dismissal message’ to the redundant boundary nodes (with the
K CH  K CH redundant CHs should be eliminated. An ordinary
farthest distance between themselves and the slave-CH). After
node will broadcast a ‘cluster member message’ in its receiving the dismissal messages, all the redundant boundary
communication radius RS. Each CH hears the message and
opt
nodes broadcast ‘overlapping tuning messages’ in RC. A CH
counts the total number of received messages, K CH  K CH CHs will reply ‘overlapping join messages’ to the redundant
receiving the least number of ‘cluster member messages’ will boundary nodes if the following conditions are satisfied: (1) its
opt
be eliminated. Finally, K CH CHs will broadcast ‘cluster head corresponding cluster is overlapped with the cluster which the
messages’ in their communication radius RS to declare as the boundary nodes belongs to; (2) the boundary nodes in the
final CHs. overlapping region is less than OD0.
Each CH contains an overlapping cluster table with the
C. Cluster Formation Step information of which overlapping regions its members belong
The 2-logical overlapping cluster formation is operated in to, while each CM establishes a boundary table containing all
three time intervals. IDs of its neighbors in the same overlapping region.
In the first time interval [0, T1], non-overlapping clusters are Remark 3: Although the cluster formation schemes is more
formed. Each node will connect to its closest CH (master CH) complicated than non-overlapping cluster formation schemes,
by sending a ‘cluster-join message’. If a node has more than its overhead still can be reduced because it only needs to
one nearest CHs, it will choose the one with the smallest ID. execute once during the whole working time.
Remark 1：Under uniform distribution, the cluster size for
D. Data Transmission Step
each cluster is no more than m0 after the first cluster formation
round. We will prove the result by analyzing three cases: In continuous monitoring applications, the detection
Case I: the Euclidean distance between any pair of closest accuracy and energy efficiency can be traded according to the
CHs is less than 2RC, and the cluster radius for each event development (the event is defined as the value of detected
non-overlapping cluster is less than RC. variable exceeds a pre-set threshold r0). Two data reporting
schemes are used to balance the requirements of detection

1530-437X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2632198, IEEE
Sensors Journal

accuracy and energy efficiency. As shown in Fig. 6, if the Boundary nodes in the overlapping areas are used to indicate
measured variable is in the normal range, the sensors transmit the direction of event development. Every boundary node
to CHs only when they sense a change (the difference between maintains a boundary node table to store the information about
two adjacent sampling instant is greater than the reference the IDs of its master-CH, slave-CH and neighbor boundary
value rth). Nodes will communicate at a low frequency for nodes (BNNs) in the same overlapping region. According to the
saving energy while guaranteeing network connectivity. Once average reading of the boundary nodes in the same overlapping
the event is detected, sensors in the event area send data to CHs region, the direction of the event development can be estimated
periodically at a higher frequency, all the freshly sensed values and the proper adjacent CHs are activated for event monitoring.
at each sampling instant will be sent to users for improving the For cluster i, suppose the number of boundary nodes in its
monitoring accuracy. uth ( u  1, 2,..., k 0 ) overlapping area is miu , the reading of the
pth sensor at the sampling instant t is rp(t). Hence, the difference
of average reading at the adjacent sampling instant t and (t+1)
in the uth overlapping area (denoted as r（ u t  1）
) is
u
1 mi (11)
r（
u t  1） u  (rp (t  1)  rp (t )).
mi p 1
Fig. 6: Data transmission in EEAOC
The difference of average reading at sampling instant t
As the monitoring area is overlapped by 2-logical-coverage between the uth and vth overlapping areas in cluster i (denoted
clusters and each CM only needs to send the data to one of its as quv (t ) ) is
belonging CHs,  K CH / 2  CHs will be activated to play the data
miu v

fusion role (active CHs) and the rest work as ordinary nodes 1 1 mi v (12)
(inactive CHs) in the data transmission step. The active CHs
quv (t ) 
miu
 rpu (t ) 
p 1
 r (t ).
miv p 1 p
should be distributed uniformly in the network. First, the CH The parameterΔru indicates whether the detected event will
with the highest ID will send a ‘cluster initialization message’ expand. When the reading of a sensor is above r0, an event is
to its neighbor CH with the least overlapping regions to be an detected in the network. If Δ ru decreases continuously, the
active CH. And the neighbor CHs will repeat this initialization event may disappear sooner. Most of the sensors in these areas
operation until  K CH / 2  CHs are selected. Then, the active will go into the sleep mode while a few of them transmit the
CHs will send ‘connectivity-request message’ to their members sensed data at a low frequency for network connectivity. The
to inform them which CHs the sensed data should be sent to. parameterΔquv indicates the motion direction of the event. IfΔ
Once a node detects the event, it will send a ‘query message’ to quv decreases continuously, it means that the event will move to
its CH at first, and the CH will broadcast a ‘reply message’ to the adjacent cluster sharing the overlapping region v . As shown
inform its members to switch the data reporting frequency to a in Fig. 7, an event is detected in a small region covered by
high level. Cluster B. The CH of Cluster B judges the direction of event
Each CH creates a TDMA schedule for time slots allocation. development based on the boundary node tables. When the
To reduce the interference among clusters, each cluster average readings of node-a and node-b decrease while the
communicates through different channels. In the inter-cluster average readings of node-c and node-d increase, the event will
communication slot, each CH sends the aggregated data to the spread to the area covered by Cluster C. Therefore, the CH of
sink directly. Cluster B sends an ‘event moving’ message to the CH of
E. CHs Re-adjustment and Cluster Migration Step Cluster C to activate all CMs in Cluster C for data gathering in
the next round.
As CHs consume more energy than ordinary nodes, CHs
rotation is essential to balance the energy consumption.
However, frequent updating of CHs results in additional energy
consumption [22]. We use the total energy depletion of CHs in
the last round as the threshold for re-adjusting CHs. The energy
depletion of CHi during the last round is
2
E CH (i )  ( n 0 lE elec  n 0 (l  1) E DA  l  mp  d CH )  f , (10)
where dCH is the distance from CHi to the sink, and there are f
frames sent from CHi to the sink in the last round.
When the residual energy for CHi is less than ECH(i), the Fig. 7: Cluster migration with event development
nearest CM with residual energy more than ECH(i) in the active
cluster is selected to be CH in the next round. The current CH To prevent misjudgment of event development, a judgment
will send a ‘CH-switching message’ to its members. window is introduced to control the aggressiveness level of
When the event spreads to another region, a demand-driven cluster migration. The window contains a sequence of F
cluster migration strategy is utilized to ensure that the adjacent number of Δ rj and L number of Δ qjf in consecutive time
active nodes are divided into the same cluster as far as possible. intervals. If the sequence of Δ rj strictly monotonically

1530-437X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2632198, IEEE
Sensors Journal

increases and the sequence of Δqjf are strictly non-increasing, the event development. From Lines 17 to the end of the code,
the cluster migration can be manipulated. each CM will decide its role in the coming round after receiving
When the event spreads to a larger area, the adjacent 1-hop the corresponding control messages.
clusters are combined into a 2-hop cluster via the boundary
nodes. As shown in Fig. 8, at the beginning, the event area is
covered by the blue chain line cluster. Once the event spreads to
ni , OD (i), Ti
a larger area, the three green dotted line clusters will combine
with the blue one to form a 2-hop cluster. The nodes can
communicate with CH directly or through the nearest boundary
nodes. The boundary nodes aggregate the forwarding data and
send it to CHs. If the event spreads to an even larger region,
more than one 2-hop clusters are combined to ensure all the
event development is detected.

ki  k0

r  rth r rth
Fig. 8: Cluster migration with event expansion
Algorithm 1: Cluster migration and CH re-adjustment algorithm
1: for each CH do
Ere  Eth
2: if Ere<Eth then
3: Send ‘CH-switching message’ & ‘Cluster-join messages’
4: end
5: if Δr>rth &Δq>qth then
6: send ‘event-moving message’ to the next active CH;
7: If k0 ‘event-moving message’ should be sent then
8: send ‘cluster extension messages’ to neighbor CHs; Fig. 9: Flow chart of EEAOC
9: end
10: end The flow chart of EEAOC protocol is shown in Fig. 9. It can
11: if receive ‘cluster extension messages’ then be seen that the information collection step, cluster head
12: broadcast ‘cluster extension messages’ to its CMs;
13: be 1-hop or 2-hop CMs; selection step and cluster formation step only need to be
14: end operated once after the sensors are deployed over the
15: end monitoring area, which can reduce the complexity of the
16: for each CMs do EEAOC protocol. The data transmission step will be rotated
17: if receive ‘CH-switching messages’ then
18: if CM_ID=CH_swicth_ID then over time while the CH re-adjustment and cluster migration
19: be a CH in the coming round; steps will be operated only when the CH re-adjustment and
20: else cluster migration conditions are satisfied.
21: send ‘cluster-join message’ to the new CH;
22: end
23: end
V. PROTOCOL ANALYSIS AND SIMULATIONS
24: if receive ‘event-moving message’ then A. Protocol Analysis
25: send data to the specified CH;
26: end (1) There is only one CH in each CH competing radius.
27: If receive ‘cluster extension message’ then In the CH selection step, the node with the highest residual
28: send data to the 1-hop node; energy will broadcast competing message within the competing
29: end
30: end radius. All the nodes in its competing radius will give up the
competing operation. As a result, the node with the highest
Algorithm 1 provides the pseudo-code for cluster migration residual energy is the only CH in its competing radius.
and CH re-adjustment of EEAOC protocol. From Lines 1 to 5, (2) All nodes in the network are mapped to the cluster
the CH re-adjustment operation starts by current CHs with topology by EEAOC.
residual energy lower than the threshold. From Lines 6 to 11, In EEAOC, on one hand, all the nodes receiving the CH
the cluster migration operation starts by active CHs when the competing messages can be connected by at least one CH. On
event development condition is met. From Lines 12 to 16, the the other hand, all the nodes which do not receive the CH
cluster extension operation starts when all directions detected competing messages will announce themselves as CH, and can

1530-437X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2632198, IEEE
Sensors Journal

be connected by neighboring CHs. The isolate points are same as the process in LEACH protocol. Moreover, we set hard
avoided in the network. threshold HT=1 and soft threshold ST=0.01. We assume that
(3) The message complexity of EEAOC is O(N). the reading of the nodes in the event area is 2sin(0.1r) and the
The details of control messages during the cluster-setup steps reading of the nodes outside the event area is 0. The DEECIC
are summarized in Table 2. protocol is a multi-hop time-driven clustering protocol. Like
TABLE 2 the LEACH protocol, all the nodes in the monitoring area send
CONTROL MESSAGES DURING THE CLUSTER SET-UP STEPS data to the sink node periodically. As a result, the data reporting
Cluster Number of Message process is independent of the event area.
Message types
setup steps messages Producers The process of event development is modeled as a circular
hello 1 sink node region moving with uniform rectilinear speed. Ten simulation
Information
Collection hand-shaking N
ordinary runs are performed and statistics are averaged over these 10
nodes runs. The specific parameters are set as in Table 3.
ordinary
CH competing KCH(KCH<N) TABLE 3
nodes
NETWORK PARAMETERS
CH ordinary
cluster member N-KCH Parameter Value
Selection nodes
cluster head K opt
CH
opt
( K CH  K CH ) CH (100×100) m2 with 100 nodes
Network size (two cases)
opt (200×200) m2 with 200 nodes
cluster-join N -K CH CM
BS position center of the monitoring area
overlapping
N all nodes Initial energy 0.5J
cluster-join
opt
overlapping K CH
reply
CH l 4000 bits
Cluster
formation overlapping *
K CH（KCH
* opt
 KCH ） RC 20 m
slave-CH
-dismissal
overlapping boundary We assume that the detected event occurs in a circular region
N*( N*<N)
-tuning nodes and the event development is modeled by an array [xe, ye, re, vc,
overlapping K（
* opt
K *  K CH ） slave-CH vr], where (xe, ye) is the coordinates for center point of the event
-join
opt
0~ K CH
area, re is the radius of event area, vc is the movement speed of
CH switching current CH
CH
opt
the event area, and vr is the expansion speed of the event size. vc
readjustment event moving 0~ K CH current CH and vr are constant such that the detected event fluctuates with
& cluster
cluster opt uniform rectilinear speed. As a result, the analytical expression
migration 0~ K CH current CH
extension of the event area is
Remark 4: In the CH re-adjustment and cluster migration ( x  xe ) 2  ( y  ye ) 2  re2 , 0  xe , ye  M (16)
step, the minimum number of control messages is 0 when the Four event development cases are utilized to evaluate the
current CHs continue to be the CHs in the next round. At most performance of the EEAOC protocol. The specific parameters
opt
K CH opt
‘CH-switching messages’, K CH ‘event moving messages’ of the event development are set as in Table 4.
opt
and K CH ‘Cluster extension messages’ should be sent when all TABLE 4
PARAMETER SETTINGS OF EVENT DEVELOPMENT
CHs need to be adjusted. Value
Based on Table 2, the message complexity of EEAOC OTotal Parameter
Case 1 Case 2 Case 3
satisfies: (x0, y0) (25,25) (50,50) (25,25)
OTotal  4 N  K *  N *  K CH
* opt
 K CH  1  4 N  1 (13) r0 20m 10m 10m
ve 0.03m/round 0 0.03m/round
OTotal  4 N  K *  N *  K CH
* opt
 4 K CH  1  10 N  1 (14)
vc 0 0.02m/round 0.03m/round
Combining inequalities (13) and (14), we obtain
4 N  1  OTotal  10 N  1. (15)  Case 1: the event moves while maintaining its shape. In this
situation, ye and re are constant and xe is a variable set as
Thus, the overall communication overhead complexity for
the cluster set-up steps in the network is O(N). xe=x0+vc×t, where x0 is the initial value of xe. The schematic

B. Simulations diagram is shown in Fig. 10(a).

 Case 2: the event expands to a larger area while keeping its
In this section, we evaluate the performance of EEAOC via centre. In this situation, xe and ye are constant values while re is a
MATLAB. A comparison among LEACH (time-driven, single
hop) and DEECIC (time-driven, multi-hop) and TEEN variable set as re=r0+vr×t , where r0 is the initial value of re.
(event-driven) is performed in two simulation scenarios under The schematic diagram is shown in Fig. 10(b).
different event scalabilities and changing rates. In this  Case 3: the event moves while extending its area, and it will
simulation, we will take the same configuration parameters in spread to the whole network eventually. The schematic diagram
LEACH. The pseudo-codes of LEACH are publicly available is shown in Fig. 10(c).
in http://csr.bu.edu/sep/LEACH.m by Gergios Smaragdakis. In  Case 4: All sensors in the monitoring area detect the event
the simulations for TEEN, the cluster formation process is the and send the gathered data to the sink periodically.

1530-437X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2632198, IEEE
Sensors Journal

0.5
EEAOC
0.45 TEEN

Average residual energy of living nodes

DEECIC
0.4
LEACH

0.35

0.3

0.25

0.2

0.15
(a) Event moving while keeping its shape
0.1

0.05

0
0 500 1000 1500 2000 2500 3000
Rounds

(a) Network with 100 nodes

0.5

0.45

Average residual energy of living nodes

0.4

0.35

0.3

0.25

0.2
(b) Event spreads to larger area while keeping its central point
0.15
EEAOC
0.1
TEEN
0.05 DEECIC
LEACH
0
0 500 1000 1500 2000 2500 3000
Rounds

(b) Network with 200 nodes

Fig. 11: Average residual energy of living nodes over rounds
(c) Hybrid of scenarios (a) and (b) 100

Fig. 10: Event area development 90

80
When a node runs out of its energy, it is taken as dead. In this
70
Number of nodes alive

work, we take the time when 10% nodes die (TND) to evaluate
60
the network lifetime because the death of 10% nodes has a great
50
influence on the connectivity of network [9]. Moreover, we
40
assume that all packets received by a CH are well aggregated
30
into a constant length. EEAOC
TEEN
20
First, we evaluate the EEAOC protocol with two network DEECIC
10 LEACH
sizes under Case 3 to investigate its adaptability to the network
0
size. The results in Fig. 11 show that EEAOC retains most 0 500 1000 1500
Rounds
2000 2500 3000

residual energy compared with other protocols in both network (a) Network with 100 nodes
sizes. The results in Fig. 12 indicate that most of the nodes are
200
still alive under the network size with 200 nodes and the second
180
most of the nodes are still alive under the network size with 100
160
nodes. Comparing the results in Fig. 11 and Fig. 12, we can see
140
that LEACH performs better in small-scale networks due to the
Number of nodes alive

120
single hop intra-cluster communication feature and DEECIC
performs better in large-scale networks because of the 2-hop 100

intra-cluster communication feature. Taking the advantage of 80

event-driven data reporting schemes, the TEEN protocol has 60 EEAOC

TEEN
more living nodes than LEACH. Since EEAOC combines the 40 DEECIC
LEACH
virtues of event-driven data reporting scheme and event-based 20

cluster formation mechanism, it can decide whether or not to 0

0 500 1000 1500 2000 2500 3000
transmit to the CHs continuously based on the data attributes, Rounds

and form clusters to single hop or 2-hop according to the event (b) Network with 200 nodes
Fig. 12: Number of nodes alive over rounds
development. Hence, EEAOC can prolong the network lifetime
in both network sizes.

1530-437X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2632198, IEEE
Sensors Journal

Second, we simulate EEAOC under the four event energy efficiency are well balanced, making it suitable for
development cases mentioned above. Figs. 13–16 show the continuous monitoring applications.
number of living nodes and average residual energy of the 0.5
living nodes for EEAOC, TEEN, DEECIC and LEACH under 0.45
four event development cases. The results in Figs. 13 (a), 14 (a)

Average residual energy of living nodes

0.4
and 15 (a) indicate that the EEAOC protocol retains most 0.35
average energy under Case 1–Case 3. In Case 1–Case 3, the 0.3
event occurs in a local area and changes dynamically with its
0.25
location and size. The sensors in the event area send data to the
0.2
sink periodically while a few sensors outside the event area will
0.15
communicate with the sink at a low frequency. By considering EEAOC
0.1 TEEN
the residual energy during the CH selection phase, DEECIC DEECIC
0.05 LEACH
retains more energy compared with LEACH. As event-driven
0
reporting scheme is used, no data will be transmitted most of 0 500 1000 1500
Rounds
2000 2500 3000

the time for TEEN. As a result, TEEN achieves better energy (a) Average residual energy of living nodes
efficiency compared with LEACH and DEECIC. Compared
100
with TEEN, more communication traffic is required by the
90
EEAOC protocol due to the QoS requirement of continuous
80
monitoring. However, since EEAOC can organize sensors into
70
more reasonable clusters, it achieves even higher energy

Number of nodes alive

60
efficiency compared with the TEEN protocol. Since the vast
50
majority of nodes exhaust their energy after 2500 rounds while
40
a few nodes maintain their initial energy as they are always
30
located outside the event area, there is a significant increase of EEAOC
20 TEEN
the average residual energy for living nodes after the nodes DEECIC
10 LEACH
with low residual energy die. In some cases, a few nodes are
0
located outside the event area all the time. After all other nodes 0 500 1000 1500 2000 2500 3000
Rounds
die, the average residual energy of living nodes is equal to the
(b) Number of nodes alive
initial energy (for example, in Case 1). Fig.13: Living nodes and average residual energy under Case 1
The results in Figs. 13 (b), 14(b) and 15 (b) show that
0.5
although the first node die time of EEAOC (1040th round) is EEAOC
0.45
almost the same to LEACH (1039th round) in some specific TEEN
Average residual energy of living nodes

DEECIC
0.4
cases (Case 1), EEAOC achieves a prolonged network lifetime LEACH

0.35
because TND is much longer in all cases, compared with
0.3
LEACH and DEECIC. We also discover that, in a small
0.25
monitoring area with sensors randomly distributed, the 2-hop
0.2
intra-cluster clustering protocol (DEECIC) sometimes cannot
0.15
achieve better result than the single-hop inter-cluster clustering
0.1
protocol. Due to the flexible cluster formation scheme, EEAOC
0.05
achieves higher energy efficiency than LEACH and DEECIC.
0
The results in Fig. 16 show that our method achieves better 0 500 1000 1500 2000 2500 3000
Rounds
performance even in worst case (Case 4, all the sensors should
(a) Average residual energy of living nodes
be actively sending data at every round). As TEEN transmits
100
data to the sink only when the nodes sense a change in the
90
network, the TND of TEEN is a little longer than EEAOC.
80
However, since no information is obtained within a certain
70
period of time, data distortion will be inevitable, as shown in
Number of nodes alive

60
Fig.17, when an event is detected in the network. In LEACH
and DEECIC, all readings are sent to the sink to achieve high 50

accuracy regardless of the requirements of users. In TEEN, no 40

information will be obtained within a certain period of time 30

EEAOC
TEEN
which is not suitable for continuous monitoring applications. In 20
DEECIC
LEACH
EEAOC, the sensed data at every sampling instant are sent to 10

the sink to achieve high monitoring accuracy when the event is 0

0 500 1000 1500 2000 2500 3000
Rounds
detected while fewer data are sent to the sink to detect the
(b) Number of nodes alive
tendency of event development. The monitoring accuracy and Fig.14 Living nodes and average residual energy under Case 2

1530-437X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2632198, IEEE
Sensors Journal

0.5
EEAOC
0.45 TEEN
Average residual energy of living nodes

DEECIC
0.4
LEACH

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
0 500 1000 1500 2000 2500 3000
Rounds

(a) Number of nodes alive

100

90

80

70
Number of nodes alive

60

50

40
EEAOC
30 TEEN
DEECIC Fig. 17: Information obtained with different protocols
20 LEACH
The comparison of TND among EEAOC and other protocols
10
is depicted in Fig. 18 (mean, standard deviation, 95%
0
0 500 1000 1500 2000 2500 3000 confidence interval). It is shown that EEAOC can obtain better
Rounds
performance when the detected event occurs in a large event
(b) Average residual energy of living nodes
Fig.15: Living nodes and average residual energy under Case 3 area (the TND of our EEAOC protocol can be prolonged nearly
by 1401% and 210% under Case 1, 200% and 290% under Case
0.5
EEAOC 2, 140% and 220% under Case 3, 50% and 100% when
0.45 TEEN
compared with LEACH and DEECIC respectively). To meet
Average residual energy of living nodes

DEECIC
0.4
LEACH the continuous monitoring requirement, certain amount of data
0.35
should be sent to sink even if no event is detected in the
0.3
network. As a result, EEAOC obtains a little shorter TND than
0.25
the event-driven TEEN protocol (the TND of our EEAOC
0.2
protocol is nearly 20%, 5%, 20% and 20% shorter compared
0.15
with TEEN protocol under Case 1–Case 4, respectively). The
0.1 comparison results have illustrated the high energy efficiency
0.05 of the proposed EEAOC protocol in continuous monitoring
0
0 500 1000 1500 2000 2500 3000
applications.
Rounds

(a) Average residual energy of living nodes

100

90

80

70
Number of nodes alive

60

50

40

30 EEAOC
TEEN
20
DEECIC
LEACH
10

0
0 500 1000 1500 2000 2500 3000
Rounds
Fig.18: Network lifetime under the four cases
(b) Number of nodes alive
Fig.16 Living nodes and average residual energy under Case 4

1530-437X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2632198, IEEE
Sensors Journal

Finally, to evaluate the influence of the cluster migration dynamic continuous monitoring applications compared with
scheme in EEAOC protocol, we simulate EEAOC without existing clustering protocols.
considering the cluster migration scheme under Case 3, and
compare with the full version of EEAOC protocol. The results REFERENCES
in Fig. 19 indicate that the cluster migration scheme makes a [1] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, E. Cayirci, Wireless
great contribution in prolonging the network lifetime. The sensor networks: A survey, Computer Networks, 38(4): 393–422, 2002.
cluster migration scheme can improve the energy efficiency [2] J. Yick, B. Mukherjee, D. Ghosal, Wireless sensor network survey,
since it is adaptive to the event development. Associated with Computer Networks, 52(12): 2292–2330, 2008.
our cluster migration scheme, the nodes in the event area are [3] Q. Liu, Z. Wang, X. He and D. H. Zhou, Event-based recursive
grouped into proper clusters for data gathering, which can distributed filtering over wireless sensor networks, IEEE Transactions on
reduce unnecessary data transmission from irrelevant nodes. Automatic Control, 60(9): 2470–2475, 2015.
0.5 [4] Z. Wang and Y. Niu (editors), Special Issue on “Distributed Estimation
0.45
and Filtering for Sensor Networks”, International Journal of Systems
Science, 42(9): 1421–1425, 2011.
Average residual energy of living nodes

0.4
[5] V. Potdar, A. Sharif, E. Chang, Wireless sensor networks: a survey, In
0.35
Proceedings of the International Conference on Advanced Information
0.3
Networking and Applications Workshops, pp.636–641, 2009.
0.25
[6] C. F. Wang, J. D. Shih, B. H. Pan, and T. Y. Wu, A network lifetime
0.2
enhancement method for sink relocation and its analysis in wireless
0.15
sensor networks, IEEE Sensors Journal, 14(6): 1932–1943, 2014.
0.1
EEAOC with cluster migration scheme
[7] Y. Hu, Y. Niu, F. H. Gan, Congestion control for wireless sensor
0.05 EEAOC without cluster migration scheme networks: a survey, In Proceedings of the 26th Chinese Control and
0 Decision Conference, pp. 4853– 4858, 2014.
0 500 1000 1500 2000 2500 3000
Rounds
[8] A. F. Liu, P. H. Zhang, Z. G. Chen, Theoretical analysis of the lifetime
(a) Number of nodes alive and energy hole in cluster based wireless sensor networks, Journal of
100
Parallel and Distributed Computing, 71(10): 1327–1355, 2011.
90
[9] A. F. Liu, X. Y. Wu, Z. G. Chen, W. H. Gui, Research on the energy hole
80
problem based on unequal cluster-radius for wireless sensor networks,
Computer Communications,33(3): 302–321, 2010.
70
Number of nodes alive

[10] M. E. Keskin, İ. K. Altınel, N. Aras, C. Ersoy. Wireless sensor network

60
lifetime maximization by optimal sensor deployment, activity scheduling,
50
data routing and sink mobility, Ad Hoc Networks, 17(6): 18–36, 2014.
40 [11] K. Malav, D. Gupta, V. Murray, Energy efficient routing in zigbee
30 wireless sensor network- a review. International Journal of Advanced
20 Research in Computer and Communication Engineering, 4(4): 336–343,
EEAOC with cluster migration scheme
10 EEAOC without cluster migration scheme
2015.
0 [12] M. M. Afsar, M. Tayarani-N, Clustering in sensor networks: a literature
0 500 1000 1500 2000 2500 3000
Rounds survey. Journal of Network and Computer Applications, 46:198–226,
(b) Average residual energy of living nodes 2014.
Fig.19: Living nodes and average residual energy under Case 3 [13] A. Abbasi, M. Younis, A survey on clustering algorithms for wireless
sensor networks. Computer Communications, 30(14–15): 2826–2841,
VI. CONCLUSION 2007.
[14] N. C. Rathore, P. Gour, B. S. Meena, Improved clustering protocol for
In this work, we have presented a dynamic clustering method
delay minimization, International Journal of Computer Networking,
EEAOC for dynamic continuous monitoring applications. In
Wireless and Mobile Communications, 3(5): 41–48, 2013.
the proposed EEAOC, the collaborative CH re-adjustment and
[15] M. A. Kafi, D. Djenouri, J. B. Othman, N. Badache, Congestion control
cluster migration technique is used to activate appropriate
protocols in Wireless Sensor Networks: a survey, IEEE Communications
clusters for continuous monitoring via the
Surveys & Tutorials, 16(3): 1369–1390, 2014.
2-logical-overlapping clustering scheme. As a result, the cluster
topology can be switched according to the event development [16] B. D. Lee, K. H. Lim, An energy-efficient hybrid data-gathering protocol
with low communication overhead. The hybrid data reporting based on the dynamic switching of reporting schemes in Wireless Sensor
scheme is operated subsequently to balance the energy Networks, IEEE Systems Journal, 6(3): 378–387, 2012.
efficiency and detection accuracy. By doing so, an [17] A. R. Wilson, Event triggered analog data acquisition using the
energy-efficient operation can be achieved by EEAOC. The exponential moving average, IEEE Sensors Journal, 14(6): 2048–2055,
simulation results have confirmed that EEAOC obtains better 2014.
energy efficiency than time-driven clustering protocols and [18] N. Bouabdallah, M. E. Rivero-Angeles, B. Sericola, Continuous
better detection accuracy than event-driven protocols. It is thus monitoring using event-driven reporting for cluster-based wireless sensor
concluded that the EEAOC protocol is more suitable for the

1530-437X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2016.2632198, IEEE
Sensors Journal

networks, IEEE Trans. on Vehicular Technology, 58(7): 3460–3479, [27] A. Manjeshwar, D. P. Agrawal, APTEEN: A hybrid protocol for efficient
2009. routing and comprehensive information retrieval in wireless sensor
[19] G. Amato, S. Chessa, C. Gennaro, C. Vairo, Querying moving events in networks, In Proceedings on Parallel and Distributed Processing
Wireless Sensor Networks, Pervasive and Mobile Computing, 16: 51–75, Symposium, 2002.
2015. [28] B. A. Pazzi, R. W. N. Araujo, R. B. HPEQ: A hierarchical periodic,
[20] A. P. Chandrakasan, A. C. Smith, W. B. Heinzelman, An application event-driven and query-based wireless sensor network Protocol, In
specific protocol architecture for wireless micro sensor networks, IEEE Proceedings of IEEE Conference on Local Computer Networks, pp.
Transactions on Wireless Communications, 1(4): 660–669, 2004. 560–567, 2005.
[21] Y. Liao, H. Qi, W. Q. Li, Load-balanced clustering algorithm with [29] C. Huang, D. W. C. Ho, J. Q. Lu, Partial-Information-Based distributed
distributed self-organization for wireless sensor networks, IEEE Sensors filtering in two-targets tracking sensor networks, IEEE Transactions on
Journal, 13(5): 1498–1506, 2013. Circuits and Systems I, 59(4): 820–832, 2012.
[22] H. Zhen, Y. Li, G. J. Zhang, Efficient and dynamic clustering scheme for [30] S. Chen, D. W. C. Ho, C. Huang, Fault reconstruction and state estimator
heterogeneous multi-level Wireless Sensor Networks, Acta Automatica design for distributed sensor networks in multi-targets tracking, IEEE
Sinica, 39(4): 454–460, 2013. Transactions on Industrial Electronics, 62(11): 7091–7102, 2015.
[23] M. Sabet, H. R. Naji, A decentralized energy efficient hierarchical [31] J. Yan, C. Chen, X. Luo, H. Liang, X. Guan, X. Yang, Topology
cluster-based routing algorithm for wireless sensor networks, optimisation-based distributed estimation in relay assisted wireless
International Journal of Electronics and Communications, 69(5): sensor networks, IET Control Theory & Applications, 8(18): 2219–2228,
790–799, 2015. 2014.
[24] M. A. Mamun, Y. Koi, N. Nakaya, Y. Hagihara, G. Chakraborty, HEHC: [32] W. Zhou, J. Mou, T. Wang, T. Ji, J. Fang, Target-synchronization of the
Heterogeneous-aware enhanced hierarchical clustered scheme for distributed wireless sensor networks under the same sleeping-awaking
wireless sensor networks, In Proceedings of SICE Annual Conference, method, Journal of the Franklin Institute-Engineering & Applied
pp.1517–1522, 2011. Mathematics, 349(6): 2004–2018, 2012.
[25] Z. X. Liu, Q. C. Zhang, L. X., X. P. Guan, A distributed energy-efficient [33] T. Jing, H. Snoussi, C. Richard, Prediction-based cluster management for
clustering algorithm with improved coverage in Wireless Sensor target tracking in wireless sensor networks, Wireless Communications
Networks, Future Generation Computer Systems, 28(5): 780–790, 2012. and Mobile Computing, 12: 797–812, 2012.
[26] A. Manjeshwar, D. P Agrawal, TEEN: A routing protocol for enhanced [34] S. Lee, I. Hwang, Event-based state estimation for stochastic hybrid
efficiency in wireless sensor networks, In Proceedings on Parallel and systems, IET Control Theory & Applications, 9(13): 1973–1981, 2015.
Distributed Processing Symposium, pp.2009–2015, 2001.

1530-437X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.