A Clustering Topology for Wireless Sensor Networks

New Semantics over Network Topology

Ionel Tudor Calistru, Paul Cotofrei and Kilian Stoffel

Information Management Institute, University of Neuchatel, Neuchatel, Switzerland

Keywords:

Sensor Networks, Semantic Sensor Web, Network Topology, Energy-aware Routing, Routing protocols, Data

Mining, Clustering, DBSCAN.

Abstract:

Sensor networks are a primary source of massive amounts of data about the real world that surrounds us,

measuring a wide range of physical parameters in real time. Given the hardware limitations and physical envi-

ronment in which the sensors must operate, along with frequent changes of network topology, algorithms and

protocols must be designed to provide a robust and energy efﬁcient communications mechanism. With a view

to addressing these constraints, this paper proposes a routing technique that is based on density based spatial

clustering of applications with noise (DBSCAN) algorithm. This technique reveals several network topology

semantics, enables the splitting of sensors responsibilities (communication/routing and sensing/monitoring),

reduces the level of energy wasted on sending messages through the network by data aggregation only in

cluster-head nodes and last but not the least, brings along very good results prolonging the network lifetime.

1 INTRODUCTION

Wireless sensor networking is an emerging technol-

ogy that has a wide range of potential applications

(Sheth et al., 2008; Calbimonte et al., 2011) in-

cluding environment monitoring (e.g. meteorology,

civic planning, trafﬁc management, calamities detec-

tion), medical systems (e.g. health monitoring), smart

spaces, home automation or homeland security. Such

networks will consist of a large number of heteroge-

neous sensor nodes, sensors that organize themselves

into a multihop wireless network. Each node has one

or more embedded processors, low-power radios and

is normally battery operated. Typically, these nodes

are very densely deployed and have sensing, commu-

nicating and data processing capabilities. They can

gather different type of information such as pressure,

humidity, temperature, speed, location and they must

coordinate to perform a common task.

Although many protocols and algorithms have

been proposed for traditional wireless ad-hoc net-

works, they are not well suited to the unique fea-

tures and application requirements of sensor net-

works. Given the hardware and computational limi-

tations (small size, low power, limited computational

and memory capacities), the constraints of the phys-

ical environment in which the sensors must operate,

along with frequent changes of network topology,

optimal routing algorithms and protocols (Akyildiz

et al., 2002; Stojmenovic, 2005) must be designed to

provide a robust and energy efﬁcient communications

mechanism.

2 MOTIVATION

Prolonged network lifetime, topology awareness,

scalability and load balancing are important require-

ments for many sensor network applications. We

have also to take into consideration that the most

applications usually deploy a larger number of sen-

sors than the optimum necessary (Akkaya and You-

nis, 2005). Also many of the sensors that are used

to discover alternative networks paths to the Sink

(central node) waste their energy without contribut-

ing neither to routing or sensing phase (Cardei et al.,

2002; Cardei and Wu, 2006). To satisfy these require-

ments, several solutions have been proposed (Akkaya

and Younis, 2005; Abbasi and Younis, 2007) (You-

nis and Fahmy, 2003; Shin et al., 2006) (Cardei and

Wu, 2006) that exploit the trade of among energy, ac-

curacy, and latency. Different topologies were pro-

posed with hierarchical topologies as the most popu-

lar one. In this paper we address a cluster based topol-

ogy with a clear semantics which enables us to bet-

ter achieve the previously mentioned goals. Accord-

153

Cotofrei P., Calistru I. and Stoffel K..

A Clustering Topology for Wireless Sensor Networks - New Semantics over Network Topology.

DOI: 10.5220/0004423101530160

In Proceedings of the 2nd International Conference on Data Technologies and Applications (DATA-2013), pages 153-160

ISBN: 978-989-8565-67-9

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

ing to the general classiﬁcation of clustering scheme

done in (Abbasi and Younis, 2007), different cluster-

ing techniques have been addressed in function of the

network model, clustering objective or the taxonomy

of the clustering attributes.

Among the large number of clustering algorithms

for sensor networks proposed in the literature, such as

LCA (Baker and Ephremides, 1981), CLUBS (Nag-

pal and Coore, 1998), EEHC (Bandyopadhyay and

Coyle, 2003) - all having a linear convergence rate -

or LEACH (Heinzelman et al., 2002), HEED (Younis

and Fahmy, 2004), EECPL(Bajaber and Awan, 2010),

N-LEACH (Tripathi et al., 2012) - with a constant

convergence time - our proposed cluster topology is

based on the DBSCAN (Density Based Spatial Clus-

tering of Applications with Noise) algorithm (Sander

et al., 1998), a simple and widely used density-based

clustering algorithm.

In the context of sensor networks, DBSCAN has

already been applied with different purposes; for ex-

ample, Apiletti(Apiletti et al., 2011) used it to detect

sensor correlations whereas Almuzaini(Almuzaini

and Gulliver, 2011) applies it to a range-based sensor

nodes localization algorithm. However, to the best of

our knowledge, DBSCAN has not been used before as

a in-network clustering strategy of sensor nodes. Its

deﬁnition of a cluster is based on the notion of density

reachability and it can ﬁnd arbitrarily shaped clusters,

which makes it suitable in the context of randomly

deployed wireless sensor networks. Moreover, the al-

gorithm does not require to know the number of clus-

ters a priori and it has a good efﬁciency on very large

datasets.

Taking all this facts into account, we propose a

DBSCAN based communication schema, where sen-

sor nodes are organized into clusters. The cluster-

head nodes are dedicated solely to transmit (route)

messages between them in order to reach the Sink,

and the border-nodes are responsible with the sensing

(monitoring) activity. Our objective is to prove that

this communication strategy ﬁts well sensor networks

requirements and improves considerably the network

lifetime as well as balances the energy consumption.

Moreover, several network topology semantics can be

revealed by analyzing the results of a such technique.

The rest of the paper is structured as follows. The

next section will present a detailed description of DB-

SCAN algorithm and its application to our routing

technique for wireless sensor networks. Several simu-

lation results will be presented in Section 4, followed

by different discussions and improvements. Finally,

the last section summarizes our work and proposes

some promising future research directions.

3 DBSCAN

One of the major data mining methods is the clus-

tering, deﬁned as the unsupervised learning task of

grouping the objects from a dataset into meaningful

sub-classes. There has been a lot of research on clus-

tering algorithms for decades but their application to

sensor networks rise the following new requirements:

i. Discovery of clusters with arbitrary shape, as the

clusters’ shape in sensor networks may be non-

convex, spherical, linear, elongated etc.

ii. Good efﬁciency on very large networks, with sig-

niﬁcantly more than just a few thousand objects.

Introduced in (Ester et al., 1996), the clustering al-

gorithm DBSCAN relies on a density-based notion

of clusters. For each point p of a cluster, the den-

sity in its ε-neighborhood (the number of points sit-

uated at a distance less than ε from p) has to exceed

some threshold MinP. DBSCAN requires two input

parameters (ε and MinP) and supports the user in de-

termining an appropriate value for them. Designed

to discover clusters of arbitrary shape as well as to

distinguish noise, DBSCAN is also efﬁcient for large

spatial datasets.

In (Sander et al., 1998) it is shown that we can

use any binary predicate which is symmetric and re-

ﬂexive in the deﬁnition of a neighborhood relation.

For example, when clustering polygons, the neigh-

borhood may be deﬁned by the intersect predicate.

Furthermore, instead of simply counting the objects

in the neighborhood of an object, we can use other

measures to deﬁne the ”cardinality” of that neighbor-

hood. Thus, the generalized GDBSCAN algorithm

(Sander et al., 1998) can cluster point objects as well

as spatially extended objects according to both spatial

and non-spatial attributes. In the following subsection

we present the notion of density-connected sets and

in second subsection we give a detailed description of

GDBSCAN algorithm.

3.1 Density-connected Sets

The concept of density-connected sets is a general-

ization of the concept of density-based clusters (Ester

et al., 1996; Sander et al., 1998). The generalization

concerns the neighborhood relation N

(p, p

)- the dis-

tance is replaced by any binary predicate, symmetric

and reﬂexive - and the density measure - the ordinal

cardinality is replaced by any function C

: 2

→ R

In the following, we assume D to be a ﬁnite set of

objects characterized by spatial and non-spatial at-

tributes.

The deﬁnition of a cluster in (Ester et al., 1996)

is restricted to the special case of a distance based

DATA2013-2ndInternationalConferenceonDataManagementTechnologiesandApplications

154

Figure 1: Core Objects and Border Objects (Source:

(Sander et al., 1998)).

neighborhood, N

(p) = {p

∈ D : |p − p

| ≤ ε}. A

distance based neighborhood is a natural notion of a

neighborhood for point objects, but if clustering spa-

tially extended objects such as a set of polygons of

largely differing sizes it may be more appropriate to

use neighborhood predicates like intersects or meets

for ﬁnding clusters of polygons.

A natural, but naive approach to deﬁne a density-

connected set S ⊆ D as a generalization of a density-

based cluster is to require, for each object p ∈

S, that the cardinality measure C

of the N

(p)-

neighborhood to be less than a given threshold MinC.

However, this approach fails because a density-

connected set contains two types of objects: ”inside”

the set (core object) and ”on the border” of the set

(border objects). In general, an N

-neighborhood of a

border object has a signiﬁcantly lower C

value than

an N

-neighborhood of a core object (see Fig. 1).

Therefore, the value MinC must be set to a relatively

low value in order to include all objects belonging to

the same density-connected set. This value, however,

will not be characteristic for the respective density-

connected set - particularly in the presence of noise

objects.

The deﬁnition of a density-connected set must

precise how to decide if a given object p belongs

to a given set S by deﬁning the form of the re-

lation ”connecting” two objects from the same set.

The binary relations density-reachable and density-

connected introduced in (Sander et al., 1998) are used

to deﬁne a density-connected set with respect to N

neighborhood, C

cardinality function and threshold

MinC.

Deﬁnition 1 (directly density-reachable). An object p

is directly density-reachable from an object q if p ∈

(q) and C

(q)) ≥ MinC.

Obviously, directly density-reachable is symmetric

for pairs of core objects. In general, however, it is

not symmetric if one core object and one border ob-

ject are involved (see Fig. 2).

Figure 2: Directly density-reachable relation (Source:

(Sander et al., 1998)).

Figure 3: Density-reachability relation (Source: (Sander

et al., 1998)).

Deﬁnition 2 (density-reachable). An object p is

density-reachable from an object q if there is a chain

of objects p

= q, p

, . . . , p

= p, such that ∀i =

1, . . . , n : p

i+1

is directly density-reachable from p

Density-reachability is a canonical extension of direct

density-reachability. This relation is transitive, but it

is not symmetric in general (see Fig 3), except for core

objects (a chain from q to p can be reversed if p is also

a core object).

Two border objects p and q of the same density-

connected set S may be not density reachable from

each other. However, two objects belonging to the

same density-connected set must be ”connected” by

a symmetric binary relation. This relation, denoted

density-connectivity, requires a core object in S from

which both objects of S are density-reachable.

Deﬁnition 3 (density-connected). An object p is

density-connected to an object q if there is an object o

such that both p and q are density-reachable from o.

Density-connectivity is a symmetric relation. For

density reachable objects, the relation of density-

connectivity is also reﬂexive (see Fig 4).

Figure 4: Density-connectivity (Source: (Sander et al.,

1998)).

Deﬁnition 4 (density-connected set). The set S ⊆ D

is density-connected if satisﬁes the following condi-

tions:

i. Maximality: ∀p, q ∈ D: if p ∈ S and q is density-

reachable from p then q ∈ S.

ii. Connectivity: ∀p, q ∈ S. p is density-connected to

Note that a density-connected set S contains at least

one core object: since S contains at least one object p,

this one must be density-connected to itself via some

object o (which may be equal to p). Thus, at least o

has to satisfy the core condition, C

(o)) ≥ MinC.

3.2 GDBSCAN

The clustering algorithm GDBSCAN (Generalized

AClusteringTopologyforWirelessSensorNetworks-NewSemanticsoverNetworkTopology

155

Density Based Spatial Clustering of Applications

with Noise), introduced by Sander (Sander et al.,

1998), was designed to discover density-connected

sets in a spatial, possible noise, databaset.

To ﬁnd a density-connected set, GDBSCAN starts

with an arbitrary object p and retrieves all objects

density-reachable from p with respect to N

, C

and

MinC. If p is a core object, this procedure yields a

density-connected set. If p is not a core object, no ob-

jects are density-reachable from p and p is assigned to

NOISE. This procedure is iteratively applied to each

object which has not yet been classiﬁed.

GDBSCAN (SetOfObjects, Nh, Cw, MinC)

//SetOfObjects is UNCLASSIFIED

ClusterId := nextId(NOISE)

FOR i FROM 1 TO SetOfObjects.size DO

Object := SetOfObjects.get(i);

IF Object.ClId = UNCLASSIFIED THEN

IF ExpandCluster

(SetOfObjects,Object,ClusterId, Nh, Cw, MinC)

THEN ClusterId := nextId(ClusterId)

END IF

END FOR

END; // GDBSCAN

SetOfObjects is either the whole database or a dis-

covered cluster from a previous run. Nh and MinC are

the global density parameters and Cw is a pointer to

a function Cw(Objects) that returns the cardinality

of the set Objects. ClusterIds are values from an

ordered and countable datatype (e.g. implemented by

Integers) satisfying

UNCLASSIFIED < NOISE < otherIds.

Each object is marked with a clusterId Object.ClId.

The function nextId(ClusterId) returns the suc-

cessor of ClusterId in the ordering of the datatype

(implemented as Id := Id+1), while the func-

tion SetOfObjects.get(i) returns the i

ele-

ment of SetOfObjects. Also the call to function

ExpandCluster is constructing a density-connected

set for a core object Object.

Obviously the efﬁciency of the above algorithm

depends on the efﬁciency of the neighborhood query

because such a query is performed exactly once for

each object in SetOfObjects satisfying the selection

condition. If an indexing structure is used allowing

the execution of a neighborhood query in O(logn), an

overall runtime complexity of O(n log n) is obtained.

Therefore, the basic algorithm DBSCAN (Ester et al.,

1996), seen as a specialization of the algorithm GDB-

SCAN for the parameters Nh = N

(·), Cw = ordinal

cardinality and MinC = MinP, has an overall runtime

complexity of O(nlog n).

3.3 DBSCAN based Routing Technique

The main common task of a routing protocol for WSN

is to provide a robust and energy efﬁcient communi-

cation mechanism that enables the collaboration be-

tween a large number of wireless sensors, randomly

distributed in a region called Sensor Field, in order to

send the collected information to a central processing

node, called Sink. To achieve this, one of the well

known strategy is to group sensor nodes into clusters.

We are introducing a new routing technique for wire-

less sensor networks (WSN), technique that, to the

best of our knowledge, is the ﬁrst approach based on

the DBSCAN clustering algorithm.

The idea of our approach is to form density-based

clusters of sensor nodes so that we can enable the

splitting of sensors responsibilities: make the border-

nodes responsible of sensing/monitoring the events

and the cluster-head nodes (core nodes) aware of ag-

gregating/fusion and communication/routing the mes-

sages to the Sink node.

As we have already mentioned, our approach is

following very closely DBSCAN’s deﬁnition of a

cluster that is based on the notion of density reach-

ability. The notion of directly density-reachable is

achieved by the design of WSN, where each sensor

node is aware only of its neighbors with whom it can

interact (by sending or collecting information) using

only local communication strategies. As soon as a

sensor node detects a sufﬁcient number (MinP) of

neighbors, it becomes a cluster-head node, that is ded-

icated - in our approach - solely to transmit (route)

or to aggregate (fusion) information messages. The

border-nodes ( sensor nodes that make part of the

cluster but are not cluster-heads ) are only responsible

with sensing (monitoring) tasks. When a new event is

detected, they just have to transmit the message (de-

scribing the event) to the nearest cluster-head node.

The density-reachable notion is assured by a com-

munication between cluster-head nodes (from cluster-

head to cluster-head). To directly communicate to

each other, two cluster-head nodes must be neigh-

bors, so part of the same cluster. In order to reach

the Sink, every cluster-head node must have estab-

lished its routing paths. This procedure is executed

after the clustering phase, when cluster-head nodes

are elected. In a ﬁrst phase, the Sink is broadcast-

ing a special (recognition) message to its cluster head

neighbors. In a second phase, as soon as a cluster-

head node receives this kind of message (signed by

the Sink), it updates its routing table and then for-

wards it to all its cluster-head neighbors. After a lim-

ited number of steps all the cluster-head nodes, that

are density-reachable from Sink, have their routing

DATA2013-2ndInternationalConferenceonDataManagementTechnologiesandApplications

156

tables established and are ready to route the events

detected by the border-nodes.

One of the key advantages of DBSCAN, in the

context of randomly deployed sensor networks, is that

it can ﬁnd arbitrarily shaped clusters. It can even ﬁnd

clusters completely surrounded by (but not connected

to) a different cluster. Due to the MinP parameter, the

so-called single-link effect (different clusters being

connected by a thin line of points) is reduced. Also,

DBSCAN has the notion of noise, which is very help-

ful for identifying the sensors that should not waste

their energy trying to monitor different events, and

could may be helpful for other tasks.

In the next section we will provide some sets of re-

sults obtained after simulating the activity of a sensor

network under different scenarios. We will see how

our approach reveals several network topology se-

mantics, reduces the level of energy wasted on send-

ing messages through the network by data aggrega-

tion only in cluster-head nodes and last, but not the

least, brings along very good results prolonging the

network lifetime.

4 EXPERIMENTS

Our experiments consist of different simulations that

are based on the following general scenario: A large

number of wireless sensors distributed in a region

(Sensor Field) will collaborate for a common applica-

tion such as environmental monitoring. They should

collect and send information about different events

detected, to a central processing node (Sink).

To probe our routing algorithm we have used

Repast Suite

that is a family of advanced, free and

open source agent-based modeling and simulation

platforms. It can be found in two main editions:

Repast Simphony (North et al., 2013) and Repast

for High Performance Computing (HPC) (Collier and

North, 2012). The output of our simulation was col-

lected in a HDF5

ﬁle, that is a format for ﬂexible

and efﬁcient I/O when dealing with high volume and

complex data.

The battery consumption process have been im-

plemented as a credit point system, where each ac-

tivity of the sensor node has assigned an amount of

points. Each and every sensors have an initial maxi-

mum battery capacity. Activities such as sleep mode,

send/receive messages and sensing events are deﬁned.

During simulation, the battery charge is decreased

gradually according to the sensor activities. Also sev-

eral global counters like number of messages sent

http://repast.sourceforge.net/

http://hdfgroup.org/HDF5/

Figure 5: Sensors’ energy evolution (Directed Diffusion).

through the network or number of unique messages

sent through the network are continuously updated.

To underline the improvements of the technique

that we propose, we consider as a benchmark an-

other routing algorithm, an adaptation of Directed

Diffusion approach (Akkaya and Younis, 2005; In-

tanagonwiwat et al., 2003; Estrin et al., 1999). The

basic idea consists of sending messages of recogni-

tion (interest) between the sensor nodes. Based on

these messages (known also as gradient ﬁelds), each

sensor node builds up its own routing list to reach

the central node, Sink. The criterion taken into ac-

count is Minimum Energy Path (MEP), deﬁning the

path that consumes less energy for sending packets

between source and sink node. For simplicity, energy

required to transmit data is the same between any two

nodes that can communicate.

Figure 5 presents the evolution of sensors energy

when the routing list is constructed based on directed

diffusion approach. As all nodes are tacking part to all

network activities, the behavior of all sensors is pretty

much the same. After only 50 ticks (simulation ticks)

all the sensors ﬁnished already their power. We can

remark also that ﬁrst sensor node that dies ﬁnishes its

energy after around 28 ticks. The total amount of en-

ergy spent on routing messages through the network

by the alive sensors riches 235 (see Fig. 6).

Figure 6: Energy spent by alive sensors to route messages.

AClusteringTopologyforWirelessSensorNetworks-NewSemanticsoverNetworkTopology

157

4.1 DBSCAN over Sensor Networks

In this ﬁrst scenario, all nodes were taking part

to the network activity, even though many of them

were useless. Keeping in mind that our main goal

is to maximize network lifetime and to reduce the

level of energy consumption, much better results

can be achieved by splitting the sensor responsibili-

ties in two tasks: communication/routing and sens-

ing/monitoring. And this is what our DBSCAN ap-

proach does.

Figure 7: DBSCAN Clusters.

Figure 7 shows a sample of DBSCAN output over a

sensor network. Sensor nodes are organized into two

main clusters (colored nodes). The black nodes rep-

resent the ”noise”, i.e. the sensors that should not par-

ticipate to any action and that will be kept in a sleepy

state. The so-called ”core nodes” or cluster-heads are

market in the ﬁgure with ”C” and their only respon-

sibility is to route the received data to the Sink. The

”border nodes” (clustered nodes that are not cluster-

heads) are aware of monitoring different events that

may occur. They also transmit a message directly to

their cluster-head, informing about the events already

detected. In the clustering phase, their state is ”off”.

The performance criterion used to compare DB-

SCAN clustering schema against the Directed Diffu-

sion Routing Algorithm is related to the total energy

spent by alive sensors for sending messages through

the network (Fig. 8). The latest results show us that

Figure 8: Energy spent by alive sensors to route messages

(DBSCAN).

Figure 9: Sensors’ energy evolution (DBSCAN).

the network lifetime is considerably prolonged; the

ﬁrst node dies only after 80 simulation ticks (Fig. 9).

5 DISCUSSION

As mentioned in the section before, our ﬁrst results

are related to the Directed Diffusion Routing Algo-

rithm. We remind here that the number of alive sen-

sors starts to decrease around tick 28 (Fig. 5) and that

the energy spent with sending messages is about 235

(Fig. 6). In the following we will see how current

results can be improved by several approaches.

Data Fusion (Akyildiz et al., 2002) is a technique

that can be used to aggregate and to avoid sending

duplicates messages. One sensor node will not broad-

cast a message about the same event more than once.

Consequently, the number of messages sent in the net-

work is reduced, saving also energy consumption, es-

pecially of those sensor nodes that are frequently used

in routing data. By implementing Data Fusion into

our ﬁrst scenario, the simulation results show that we

can have now all sensors alive for a longer period

(about 37 ticks), the energy consumption being more

uniform. Of course, the major advantage is the reduc-

ing of the energy used for sending messages between

nodes, that is about 100 (57% less).

Another important aspect is that, in our previous

examples, all sensor nodes were active but not all of

them were needed to contribute. As we have moti-

vated, the technique that we propose, based on DB-

SCAN approach, has the notion of noise. This is

very helpful for identifying the sensors that should not

waste their energy trying to monitor different events,

and that may be helpful for further tasks. In our sce-

nario, about 20 nodes out of 50 may remain in a

sleepy state conserving their energy by not participat-

ing to routing or monitoring activities.

We must underline that implementing Data Fu-

sion on each sensor node might be too expensive.

A remarkable improvement of our DBSCAN based

schema is that the messages are routed to Sink ”from

DATA2013-2ndInternationalConferenceonDataManagementTechnologiesandApplications

158

Table 1: Repast HPC Simulation Results.

# Dirrected Diffusion(DD) DD with Data Fusion DBSCAN DBSCAN Data Fusion

Msg Evts(Unique) Power Msg Evts(Unique) Power Msg Evts(Unique) Power Msg Evts(Unique) Power

1 0 0(0) 90% 0 0(0) 90% 0 0(0) 92.1 % 0 0(0) 92.1 %

2 121 8(3) 74.1% 113 3(3) 74.5% 77 3(2) 80.5 % 70 2(2) 80.6 %

3 153 11(4) 62.6% 134 4(4) 63.5% 90 4(2) 72.1 % 84 2(2) 72.5 %

4 214 13(5) 49.6% 160 5(5) 52.2% 113 6(4) 63.2 % 102 4(4) 63.8 %

5 363 16(8) 33% 304 9(9) 35.2% 216 12(6) 50.9 % 189 6(6) 52.2 %

6 390 16(8) 23% 333 9(9) 25.8% 231 13(7) 43.1 % 197 7(7) 44.2 %

7 400 16(8) 17% 343 9(9) 28.2% 240 15(8) 36.1 % 203 8(8) 37.7 %

cluster-head to cluster-head”, implying a much less

number of sensors than for Directed Diffusion ap-

proach. By limiting the number of steps in the routing

paths, the information arrive much more quickly to

the central node. This brings us a great advantage, and

we can see that the energy consumed by broadcasting

messages between sensors, in this last scenario, is less

than 35 (Fig. 8), which represents 35% of the best re-

sult for Directed Diffusion with Data Fusion (around

100). Figure 9 conﬁrms our DBSCAN results: net-

work lifetime is considerably prolonged by having all

sensors alive for about 80 ticks, also better than the

best result for Directed Diffusion with Data Fusion

(around 37 ticks).

However, the hardware limitations did not permit

us to evaluate these techniques with a large scale net-

work. In order to have a deeper analysis of our rout-

ing strategy, we have considered to develop a Repast

HPC Simulation, by enabling our previous solution

to work in a parallel distributed environment. Taking

the advantage of a computer cluster network, we were

able to work with an increased number of sensors ran-

domly deployed. The summary of our results, includ-

ing the ﬁrst seven simulation steps for a network with

eight hundred nodes (simulated sensors), is presented

in Table 1.

Directed Diffusion and DBSCAN strategies were

both simulated with and without Data Fusion. All the

four cases, presented in Table 1, were running in par-

allel. The columns Msg indicate the number of mes-

sages that have been sent so far between sensor nodes.

The table summarizes also the number of events de-

tected by the network and sent to central Sink node,

how many of them are unique (have not been detected

before) as well as the remaining network power.

As we may remark, a bigger number of events are

detected in earlier steps in the case of Directed Diffu-

sion strategy. The explanation for this fact is that the

events are generated by a uniform distributed random

process. In the earlier steps, when events occur very

close to Sink, it is better to send the information di-

rectly to the Sink and not to a cluster-head. On the

other hand, when events occur far away from Sink,

the clustering can speed up the events routing, fewer

hops being needed to rich the central node.

The results listed in Table 1 show us also that

the communication level between sensor nodes (Msg)

is considerably reduced when using DBSCAN, and

this is reﬂected also in the surplus of network energy.

However, several interesting results may be obtained

by adjusting DBSCAN parameters like MinP (num-

ber of neighbors a sensor node must have in its neigh-

borhood to act as a cluster-head) and ε (the sensor ra-

dius within sensors may communicate). At this point

we can say that a DBSCAN based communication

strategy improves considerably the network lifetime

as well as balances the energy consumption.

6 CONCLUSIONS AND FUTURE

WORK

To provide a robust and energy efﬁcient commu-

nications mechanism for wireless sensor network

we proposed a routing technique derived from the

density based spatial clustering of applications with

noise (DBSCAN) algorithm. Based on a series of

simulation-based experiments, we could conclude

that the implementation of DBSCAN clustering tech-

nique to WSN reveals several network topology se-

mantics, by enabling the routing from cluster-head to

cluster-head and letting the sensing responsibilities to

the border-nodes. Moreover, the DBSCAN strategy

reduces the level of energy wasted on sending mes-

sages through the network by data aggregation only

in cluster-head nodes and brings along very good re-

sults prolonging the network lifetime.

Based on the successful results of this preliminary

research, several possibilities for future work may be

identiﬁed:

• The implementation of a technique that is con-

sidering, periodically or dynamically (in speciﬁc

conditions), re-clustering of sensor nodes (since

the network topology changes in time, due to loss

of several sensors).

• The extension of the routing technique to a

dynamic self-reorganization technique that may

learn how to adjust the DBSCAN algorithm pa-

rameters (MinP , ε) in order to optimize a speciﬁc

AClusteringTopologyforWirelessSensorNetworks-NewSemanticsoverNetworkTopology

159

streaming data mining task executed on a speciﬁc

network topology.

ACKNOWLEDGEMENTS

We are thankful to Prof. Dr. Cornelius Croitoru

(”Alexandru Ioan Cuza” University, Iasi, Romania)

for the contribution and supervision of Ionel Tudor

Calistru previous work, bachelor and master thesis,

related to wireless sensor networks domain.

REFERENCES

Abbasi, A. and Younis, M. (2007). A survey on cluster-

ing algorithms for wireless sensor networks. Comput.

Commun, 30(14):2826–2841.

Akkaya, K. and Younis, M. (2005). A survey on routing

protocols for wireless sensor networks. Ad Hoc Net-

works, 3:325–349.

Akyildiz, I. F., Su, W., Sankarasubramaniam, Y., and

Cayirci, E. (2002). A survey on sensor networks.

IEEE Communications Magazine, 40:102–114.

Almuzaini, K. and Gulliver, T. (2011). Range-based local-

ization in wireless networks using the dbscan cluster-

ing algorithm. In Vehicular Technology Conference

(VTC Spring), 2011 IEEE 73rd, pages 1–7.

Apiletti, D., Baralis, E., and Cerquitelli, T. (2011). Energy-

saving models for wireless sensor networks. Knowl.

Inf. Syst., 28:615–644.

Bajaber, F. and Awan, I. (2010). Energy efﬁcient clustering

protocol to enhance lifetime of wireless sensor net-

work. Journal of Ambient Intelligence and Humanized

Computing, 1:239–248.

Baker, D. and Ephremides, A. (1981). The architectural

organization of a mobile radio network via a dis-

tributed algorithm. IEEE Trans. on Communications,

29(11):1694–1701.

Bandyopadhyay, S. and Coyle, E. (2003). An energy ef-

ﬁcient hierarchical clustering algorithm for wireless

sensor networks. In Proc. of the 22nd Annual Joint

Conf. of the IEEE Computer and Communications So-

cieties (INFOCOM), volume 3, pages 1713–1723.

Calbimonte, J.-P., Jeung, H., Corcho, O., , and Aberer, K.

(2011). Semantic sensor data search in a large-scale

federated sensor network. 4th International Workshop

on Semantic Sensor Networks 2011, 11:23–38.

Cardei, M., MacCallum, D., Cheng, M. X., Min, M., Jia,

X., Li, D., and Du, D.-Z. (2002). Wireless sensor net-

works with energy efﬁcient organization. Journal of

Interconnection Networks, 3:213–229.

Cardei, M. and Wu, J. (2006). Energy-efﬁcient coverage

problems in wireless ad-hoc sensor networks. Com-

puter Communications, 29:413–420.

Collier, N. and North, M. (2012). Parallel agent-based sim-

ulation with repast for high performance computing.

Simulation: Transactions of the Society for Modeling

and Simulation International.

Ester, M., Kriegel, H. P., Sander, J., and Xu, X. (1996).

A density-based algorithm for discovering clusters in

large spatial databases with noise. In Second Interna-

tional Conference on Knowledge Discovery and Data

Mining, pages 226–231.

Estrin, D., Govindan, R., Heidemann, J., and Kumar., S.

(1999). Next century challenges: Scalable coordina-

tion in sensor networks. In Proceedings of the ﬁfth

annual ACM/IEEE international conference on Mo-

bile computing and networking, pages 263–270.

Heinzelman, W., Chandrakasan, A., and Balakrishnan, H.

(2002). An application-speciﬁc protocol architecture

for wireless microsensor networks. IEEE Trans. on

Wireless Communications, 1(4):660–670.

Intanagonwiwat, C., Govindan, R., Estrin, D., Heidemann,

J., and Silva, F. (2003). Directed diffusion for wire-

less sensor networking. IEEE/ACM Transactions on

Networking (TON), 11:2–16.

Nagpal, R. and Coore, D. (1998). An algorithm for group

formation in an amorphous computer. In In Proc. of

the 10th Int. Conf. on Parallel and Distributed Com-

puting Systems (PDCS).

North, M., Collier, N., Ozik, J., Tatara, E., Altaweel, M.,

Macal, C., Bragen, M., and Sydelko, P. (2013). Com-

plex adaptive systems modeling with repast simphony.

Complex Adaptive Systems Modeling, 1(3):1–26.

Sander, J., Ester, M., Kriegel, H.-P., and Xu, X. (1998).

Density-based clustering in spatial databases:the algo-

rithm gdbscan and its applications. Data Mining and

Knowledge Discovery, 2:169–194.

Sheth, A., Henson, C., and Sahoo, S. (2008). Semantic sen-

sor web. IEEE Internet Computing, 12:78–83.

Shin, K., Abraham, A., and Han, S. Y. (2006). Self orga-

nizing sensors by minimization of cluster heads using

intelligent clustering. Journal of Digital Information

Managenment, 4:87–923.

Stojmenovic, I. (2005). Handbook of Sensor Networks: Al-

gorithms and Architectures. John Wiley & Sons.

Tripathi, R., Singh, Y., and Verma, N. (2012). N-leach,

a balanced cost cluster-heads selection algorithm for

wireless sensor network. In In National Conference

on Communications (NCC).

Younis, O. and Fahmy, S. (2003). Distributed clustering

for scalable, long-lived sensor networks. Technical

report, Purdue University.

Younis, O. and Fahmy, S. (2004). Heed: a hybrid, energy-

efﬁcient, distributed clustering approach for ad hoc

sensor networks. IEEE Trans. on Mobile Computing,

3(4):366–379.

DATA2013-2ndInternationalConferenceonDataManagementTechnologiesandApplications

160