EECS: An Energy Efficient Clustering Scheme in Wireless Sensor

Networks
Mao Ye', Chengfa Lil, Guihai Chenl and Jie Wu2
'National Laboratory of Novel Software Technology, Nanjing University, China
'Department of Computer Science and Engineering, Florida Atlantic University, USA

Abstract the corresponding cluster head and the cluster head

Data gathering is a common but critical operation transmits the aggregated data t o the base station
in many applications of wireless SenSOT networks. In- (BS). HEED [5] selects cluster heads through O(1)
novative techniques that improve energy eficiency to time iteration according t o some metric and adopts
prolong the network lifetime are highly required. Clus- the multi-hop communication t o further reduce the
tering is an eflective topology control approach in wire- energy consumption. PEGASIS [6] improves the per-
less sensor networks, which can increase network scal- formance of LEACH and prolongs the network lifetime
ability and lifetime. In this paper, we propose a novel greatly with a chain topology. But the delay is signif-
clustering schema EECS for wireless sensor networks, icant although the energy is saved. There are some
which better suits the periodical data gathering ap- other related work [7-91 which efficiently use energy
plications. Our approach elects cluster heads ujth. through clustering.
more residual energy through local radio communica- In this paper, we propose and evaluate an energy
tion while achieving well cluster head distribution; fur- efficient clustering scheme (EECS) for periodical data
ther more, it introduces a novel method to balance the gathering applications in WSNs. In the cluster head
load among the cluster heads. Simulation results show election phase, a constant number of candidate nodes
th,at EECS outperfoms LEACH significantly with pro- are elected and compete for cluster heads according
longing the network lifetime over 35%. to the node residual energy, The competition process
is localized and without iteration, thus it has much
1 Introduction lower message overhead. The method also produces
Continued advances of MEMS and wireless com- a near uniform distribution of cluster heads. Further
munication technologies have enabled the deployment in the cluster formation phase, a novel approach is
of large scale wireless sensor networks (WSNs) [l]. introduced to balance the load among cluster heads.
The potential applications of WSNs are highly varied, EECS is fully distributed and more energy efficient
such as environmental monitoring, target tracking and and the simulation results show that i t prolongs the
military 121. Sensors in such a network are equipped network lifetime as much as 135% of LEACH.
with sensing, data processing and radio transmission The remainder of this paper is organized as foHows.
units while the power is highly limited. Due to the Section 2 outlines the data gathering issues in WSNs.
sensors' limited power: innovative techniques that im- Section 3 exhibits the details of EECS and Section 4
prove energy efficiency to prolong the network lifetime analyzes the properties of EECS. Section 5 evaluates
are highly required. the performance of EECS. Finally, Section 6 gives the
Data gathering is a common but critical operation conclusion and future work.
in many applications of WSNs, where data aggrega-
tion and hierarchical mechanism are commonly used
techniques. Data aggregation can eliminate the data 2 Problem Outline
redundancy and reduce the communication load [3]. Data gathering is a typical application in WSNs.
Hierarchical (clustering) mechanisms are especially ef- Sensors periodical sense the environment and transmit
fective in increasing network scalability and reducing the data to the base station (BS), and t h e BS analyzes
data latency, which have been extensively exploited. the data to draw some conclusions about the activity
LEACH [4] which is the first clustering protocol, pro- in the area. ?Ve make a few assumptions about the
poses a twephase mechanism based on single-hop network model and introduce t h e radio model before
communication. The plain node transmits the data to the problem statements.

0-7803-899f-3/05/$20.000 2005 IEEE 535

2.1 Network Model
Table 1: Meanings of the Notations
To simplify the network model, we adopt a few
reasonable assumptions as follows: T)N sensors are Notation Meaning
uniformIy dispersed within a square field A ; 2)AH T a threshold between 0 and 1
sensors and BS are stationary after deployment; ETesidual the residual energy of node
3) The communication is based on the single-hop; CH the set of cluster heads
4)Communication is symmetric and a sensor can com- the sum of members in cluster j
mJ
pute the approximate distance based on the received P the set of plain nodes
signal strength if the transmission power is given; the bh node in CH
CHi
5)AIl sensors are location-unaware; 6)All sensors are the f h node in P
of equal significance.
Pj
d(x:y) the distance between node x and y
We use a simplified model shown in 141 for the ra- EX(x) the expectation of x
dio hardware energy dissipation as follows. We refer Etcompete broadcast radius of candidate nodes
readers to {4]for more details. To transmit an l-bit
data to a distance d, the radio expands:

power level. By this way each node can compute

the approximate distance to the BS based on the re-
ceived signal strength. It helps nodes to select the
(1)
The first item presents the energy consumption of proper power level to communicate with the BS. As
radio dissipation, while the second presents the en- will shown in Section 3.2, we will use this distance
ergy consumption for amplifying radio. Depending to balance the load among cluster heads. In cluster
on the transmission distance both the free space cfs head election phase, well distributed cluster heads are
and the multi-path fading channel models are elected with a little control overhead. And In clus-
used 1111. When receiving this data, the radio ex- ter formation phase, a novel weighted function is in-
pends: ER, ( E ) = 1 x Eelec.Additionally, the operation troduced to form load balanced clusters. Detailed de-
of data aggregation consumes the energy as EDA. scriptions of these two phases are in the following sub-
sections.
2.2 Problem Statement
Once a sensor node runs out its energy, we con- 3.1 Cluster head election
sider the network is dead because some area cannot In this phase, several cluster heads are elected.
be monitored any more. Periodical data gathering ap- Nodes become CANDIDATE nodes with a probability
plications in large scale sensor networks appeal the de- T and then broadcast the CDMPETEHEADMSGs within
sign of scalable, energy efficient clustering algorithms. radio range Rcompetet o advertise their wills. Each
Thus our primal goals in EECS are as follows: 1)fully CANDIDATE node checks whether there is a CANDIDATE
distributed manner. Sensors interact with each other node with mare residual energy within the radius
through4ocalized communication: 2) low control over- Rcompete. Once the CANDIDATE node finds a more pow-
head It is desirable t o reduce control overhead to erful CANDIDATE node, it will give up the competition
extend the time of data gathering: 3) load balanced without receiving subsequential CDMPETEHEADMSGs.
clustering mechanism. Balance the load among the Otherwise, it will be elected as HEAD in the end.
sensors, especially among the cluster heads. In the
next section, we will describe the EECS algorithm in 3.2 Cluster formation
details. In this phase, each HEAD node broadcasts the
HEADAD-MSG across the network, while the PLAIN
3 EECS Details nodes receive all the HEADADJSGs and decide which
EECS is a LEACH-like clustering scheme, where cluster to join. Most of existed metric for P L A I N nodes
the network is partitioned into a set of clusters with to make decisions is the distance metric. For exam-
one cluster head in each cluster. Communication be- ple in [4]or [7], the PLAIN nodes choose the cluster
tween cluster head and BS is direct (single-hop). For head that require minimum communication according
easy reference, we summarize the notations in Table to the received signal strength. However, pursuing ef-
1. ficient energy consumption of the PLAIN nodes only
In the network deplofment phase, the 3s broad- may lead HEAD nodes exhausted quickly during the
casts a “hello” message to all the nodes at a certain datu transmission phase.

536
 In the data transmission phase, the consumed en- 3.3 Synchronization issues
ergy of cluster head i : E ( C H i ) is as follows, assuming Synchronization between each phase should be
d(CHi,BSI > ~ C T O S S o U e T . guaranteed that each node has enough time t o com-
plete the procedure; while within each phase, synchro-
E ( C H i )= ~ l E e i e +(nzi
c + ~ ) ~ E +E(Eeiec
DA +~,,d) nization among the nodes is not necessary and idle
(2) nodes will turn to sleep till the phase ends. In EECS,
Observing formula 2: energy consumption of it is achieved by having the BS periodicaIly broadcast
E(CHi) is composed of three parts: data receiving, synchronization signals t o all nodes.
data aggregation and data transmission. In the field,
several cluster heads may be near the BS, while some 4 EECS Analysis
are far away. The energy expended during data trans- In this section, we anaIyze the performance of
mission for far away cluster heads is significant, espe- EECS in details and explain how to set the param-
cially in large scale networks. Since d ( C H i , BS) has eters T and Rcompete.
been fixed after cluster head election , we should jus- Lemma 1. The control overhead complexity across
tify the cluster size for each cluster head t o balance the network 2s O ( N ) ,where N is the number of nodes.
their load across the network. The larger d(CHi,BS)
is, the smaller member size mi the cluster head C H , Proof. Observing EECS, every node sends out
should accommodate. small constant-length control messages each round
Energy consumption of the PLAIN node Pj during without iteration. Each HEAD node sends three
transmitting the data to CHi obey the formula 1. Let messages which are COMPETEHEADMSG, HEADADMSG
E(P,) be the energy consumed by Pj. If Pj always and SCHEDULEJSC; each CANDIDATE node sends
chooses the cluster head CHbest with min {B(p’)], two messages which are COMPETEHEADMSG
CHbest may be exhausted due to long distance data and JOINXLWTERHSG; while the others send
transmission to the BS and immoderate cluster size, JOIN-CLUSTERJSGs only. Clearly, the total control
although the energy of Pj is saved. Thus, PLAIN node overhead is NT + N , whose asymptotic order is
Pj in EECS chooses the cluster head by considering O(W- U
not only saving its own energy but also balancing the
Good quality HEAD nodes should be guaranteed by
workload of cluster heads,i.e. two distance factors:
enough competition of the CANDIDATE nodes. Since
d ( P j ,C H i ) and d(CHi,B S ) .
T is the onIy crucial factor which affects the sum of
We introduce a weighted function cost(j, i) for the
CANDIDATE nodes, it must be large enough to guaran-
PLAIN node Pj t o make a decision, which is
tee enough CANDIDATE nodes. On the other hand, the
larger T is, the more overhead is produced in the clus-
+
cost (j , 2) = w x j(d( Pj , CHi ) ) ( I -w ) x g (d( cHi,B S)) , ter head election phase. So, we must properly set T
(3) to reduce the overhead with guaranteed HEAD quality.
and Pj chooses CHi with min {cost} t o join.
In LEACH, there is ’no interaction during the clus-
In formula 3, f and g are two normalized functions
ter head election. So the control overhead is near opti-
for the distance d ( P j , C H i ) and d(CH;,BS) respec-
mal, which is 2NP + N ( 1 - P ) = N P -t- N , where P is
tively:
similar to T in [4]. Thus the overhead of EECS is only
(l+T)/(l+P)times of LEACH. In HEED, HEAD nodes
are elected with iteration. Although t h e communica-
tion is localized and the algorithm terminates in O(1)
iteration, HEED stiIl produces much more overhead
with the upper bound Nzter xN. Clearly, our approach
f subfunction in cost guarantees that members is better than HEED. The above property shows that
choose the closest cluster head in order to minimize the control overhead of EECS is low significantly.
energy consumption of the cluster members, While g Lemma 2. There is at most one cluster heud in every
subfunction makes the nodes join the cluster head with Rcmpete rudzo covered range.
small d(CHi, SS) to aiIeviate the workload of the clus-
ter heads farther from the 8 s . w is the weighted factor Proof. Let S be the set of all sensor nodes. And for
for the tradeoff between f and g. The experiments in ‘dx E C H , let Cz = {yld(y,z) 5 Rcompete,yE SI.
Section 6 will show that the optimal value of w de- For contradiction, we assume that there is a node
pends on the specific network scale. y E C, which is also a cluster head. According

537
to the competition metric in duster head election, am, I
XCEreszdual > z * E r e s t d u a l , v z E CLC.Since ?I f cz,
then z.ETeszdual> y.Eresrdual.The communication is
symmetric in the network model of EECS. If y is the
cluster head, g.Er#zsrdual> x.E,.,,,dual as x is within
the distance RcomZrete, which is a contradiction.
So, for Vx f C H , Vy E C,, there is y # CH. 0 D oz ~4 06
I
a8
1

In [4],the author proves that there is an optimal (b) large scene

number of cluster heads kopf in a given scene. Since
EECS is a LEACH-like protocol, we want t o elect kept
cluster heads every round. According to Lemma 2, Figure 1: The impaction of T on the network life-
Rcompete affects the cluster heads directly. So we com- time:(a) normal scene, (b) large scene
pute the optimized value of Rcompete, denoted by Ropt
in the following lemma.
LEACH. In order to explain the relations between the

Rcompetet whach is
runge of ICHI.
a,
Lemma 3. Thew is an optimal mnge Rapt for
where kept is the optimal
network scale and the parameters in EECS, we run
each kind of simulation in two different scenes, which
are normal scale scene (scene 1) and large scale scene
(scene 2) respectively. The parameters of simulations
Proof. Let P(CAND1DATE) be the probability of are listed in TABEL.2, and the parameters of the radio
one node being CANDIDATE hode, so the sum of model are the same as LEACH [4]. Unless otherwise
CANDIDATE nodes n is P ( C A N D I D A T E ) x N . In specified, every simulation result shown below is the
the Rcompeteradius range, there are m nodes in average of 100 independent experinients where each
CANDIDATE state(boundary cases are ignored), where experiment uses a different randomly-generated uni-
= rR:"mPP*P form topology of sensor nodes.
A
Since all nodes have the same capxity, these m
nodes have equal probability to be HEAD, then the Table 2: Parameters of Simulations
probability of one node being HEAD node P ( H E A D ) =
P(HEAD1CANDIDATE)= n R ; o mA p e t c N - So the ex- Parameter I scene 1 I scene 2
pectation of the sum of cluster heads EX(ICH1) = Area 1 100 x 100 I 200 . . x 200
. .

NxP(HEAD)= A Location of BS (50,200) (100,350)

nR':"<te. N 400(600) lOOO(1500)
In order to optimize energy consumption, we want
to let E X ( I C H ( )equal to kop* in [4].Combining the
Initial enerm i 0 . 5 , ~' i 1.b .3 '
induction in [4]and'the formula of EX(ICH1) , we can Eelec 50 nJ/bit
find that the optimal radius Rapt is E. U
10 p J / b i t / m
0.0013 p J / b i t / m
87 m
In LEACH, cluster heads are elected simply at ran-
dom. As a result, the distribution of the cluster heads EDA 1 5 n Jlbitlsignal
are not ensured and may be non-uniform. Some mem- Packet size I 4000 bits
bers have to expend much more energy t o communi-
cate with t h e corresponding cluster heads far away.
The last two lemmas show that there is one and only Lifetime is the criterion for evaluating the perfor-
cluster head within any Rcompetewith high probabil- mance of sensor networks. In the simulation, we mea-
sure the lifetime in terms of round when the first node
ity. Thus the cluster heads in EECS are distributed
evenly. dies. We use the energy utilization rate q to evaluate
the efficiency of energy consumption which is defined
5 Simulation as the ratio of the total energy consumed when the
In this section, we evaluate the performance of first node dies to the initial total energy. A high q
EECS protocol implemented with MATLAB. For sim- implies that energy consumption is distributed well
plicity, we assume the probability of signal collision across the network.
and interference in the !tireless channel is ignorable. We first examine the impact of T on the network
And we adapt the same MAC protocols in EECS as in lifetime, as the scales are different. We have done two

538
 (b) large scene

Figure 4: The number of clusters in each round in

Figure 2: The impaction of Rcompeteon the network both EECS and LEACH (scene 1)
lifetime:(a) normal scene, (h) large scene

In Figure 3, the experiment shows the efficiency of

cost introduced to balance the load among the clus-
ter heads, where the dash line denoted as the method
without considering the cluster heads' load balance is-
sue. We set w at 0.8 in scene 1 and 0.6 in scene 2
respectively. Comparing the without - cost method
(w = 1) with the with - cost method, we find that the
cost indeed extends the network lifetime. The value of
(a) normal scene (b) large scene w is determined by the specific scene. While the net-
work grows larger, the difference among d(CHi, BS)s
impacts the load balance among the cluster heads
Figure 3: The impaction of cost on the network life- more and more distinctly. So w should be decreased
time:(a) normal scene, (b) large scene and the PLANE node will consider more about the load
of cluster head when joining the cluster. That's why
the value of w is bigger in scene 1 than in scene 2.
independent experiments in different scales. In nor- In this paper, the cost function is simple, and we will
mal scale, N = 400,600, Rcompeie= 26,22, w = 0.8; optimize the cost function in the next work.
in large scale, N = 1000,1500, Rcompete= 40,35, Finally, we compare the performance of EECS with
w = 0.6. As T varies from 0.05 to 0.75, Figure 1 the original-LEACH [4]based on the same a s s u m p
shows the relation between T and the network lifetime. tions in [4]. In scenel, kept = 6, T = 0.2, Rcompete-

-
There is an optimal range for the value of T , which is
about 0.1 0.3 in the given scene. According to the
explanation about T in Sectionl, T must be properly
26 and w = 0.8; in scene2, kept = 9, T = 0.15,
Rcompete= 40, w = 0.6. In Figure 4, it exhibits the
distribution of the number of clusters in random s e
set with guaranteed HEAD quality and low overhead. lected 100 rounds in both EECS and LEACH. Shown
Another point needed to be mentioned that the o p as the figure, the number of clusters varies widely in
timal value Toptdecreases when the network density each simulation run in LEACH; on t h e other hand,
increases. It can be explained that there is an optimal the cluster number varies narrowly a t the kept range
sum of CANDIDATE nodes in a given network coverage in EECS. In LEACH, the clusters in each round is not
size. controlled although the expectation is aware; while in
In the experiment shown in Figure 2, we demon- EECS, we use the Rapt radio radius to set up kept clus-
strate Lemma 3 by observing the relation between ters in all probability in each round. Figure 5 shows

- -
Rcompeieand the network lifetime. In scene 1, N =
400 and kept = 4 7, so the optimal value Rapt
the variation of total number of sensors still alive when
the simulation time lapses. In scenel, EECS prolongs

-
is between 2 1
-
28; In scene 2, N = 1000 and
k,t = 6 10, so Ropt is between 36 46. Observing
the impact on network lifetime when Rcolnpetevaries,
the lifetime over 35% against LEACH. T h e energy uti-
lization rate is about 93% in EECS, while only 53%
in LEACH. T h e reason is that EECS always achieves
Figure 2 suggests that the optimal value of Rcompete the well distributed cluster heads with considering the
is about 25 in scene 1 and about 40 in scene 2. Both residual energy; further, we consider t o balance the
results fall into the optimal range computed prior. load among the cluster heads with weighted function.

539
 4Y)r . . . 1 I ” ’ I Ministry of Education.
UXI -. 7 I
References
D. Estrin, et. al., “Instrumenting the world with
wireless sensor networks,” in the International
Conference on Acoustics, Speech and Signal Pro-
cessing (ICASSP) 2001.
Akyildiz I.F., et. al.,“A Survey on Sensor Net-
(a) normal scene (b) large scene works,” IEEE Communications hdagazine, Au-
gust, pp. 102-114, 2002.

Figure 5 : Performance comparison of EECS and B. Krishnamachari, et. al., “The impact of data
LEACH:(a) normal scene, (b) Iarge scene aggregation in wireless sensor networks,” in the
22nd International Conference o n Distributed
Computzng Systems Workshops (ICDCS W’OZ).
In Figure 5-b, the efficiency of EECS is more dis- W. Heinzelman, et. al., LLAnapplication-specific
tinct when the network scale grows. In [5], the au-
protocol architecture for wireless microsensor net-
thor mentions that the original LEACH outperforms works,” I E E E Transactions on IVzreless Commu-
HEED When based on the same assumptions in [4] nications, 1(4):660-669, 2002.
which is identical with EECS. In order to save energy
further, HEED adopts the mu,lti-hop communication 0. Younis, et. al., “HEED: A Hybrid, Energy-
among the cluster heads during the inter-cluster com- Efficient, Distributed Clustering Approach for Ad
munications in the data transmission phase. Notice Hoc Sensor Networks,” IEEE Transactions on
that we focus on the cluster set-up algorithm but not Mobile Computing, 3(4):660-669, 2004.
the data transmission approach in our current work.
S. Lindsey, et. al. ,“PEGASIS: Power-Efficient
Future work will consider the multi-hop technique in Gathering in Sensor Information Systems,” IEEE
the inter-cluster communication. Readers should r+ Aerospace Conference Proceedings, Vol. 3, 9-16
fer t o [12] for details about the multi-hop routing in pp. 1125-1130, 2002.
clustered networks.
S. Bandyopadhyay, et. al., “An Energy-Efficient
6 Conclusion and Future Work Hierarchical Clustering Algorithm for Wireless
In this paper, we present a novel distributed, energy Sensor Networks,” in IEEE INFOCOM’03.
efficient and load balanced clustering scheme applied
€or periodical data gathering. EECS produces a uni- H. Chan, et. al., “ACE: An Emergent Algorithm
for Highly Uniform Cluster Formation,“ in the
form distribution of cluster heads across the network
First European Workshop o n Sensor Networks
through localized cbmmunication with littie overhead.
(E WSN), January 2004.
What’s more, a novel approach has been introduced
t o distribute the energy consumption among the sen- J. Kamimura, et. al., “Energy-Efficient Cluster-
sors in the cZuster formation phase. Simulation re- ing Method for Data Gathering in Sensor Net-
sults show that EECS prolongs the network lifetime works,” in the Annual International Conference
as much as 135% of LEACH and the total energy is o n Broadband Networks, October 2004.
efficiently consumed.
J. Hill, et. al., “System Architecture Directions
A11 of our contributions here are focused on the for Networked Sensors”, Architectural support for
cluster set-up stage. There are still much space to
Pmgmmming Languages and Operating Systems,
improve the performance of data transmission. In the
pp. 93-104,2000.
large scale sensor networks, multi-hop communication
is a mainstream technique for energy saving. We will [Ill T. Rappaport, Wireless Communication: Princi-
remove the assumption of single-hop and design an en- pIes & Practice. Prentice-Hall, 1996.
ergy efficient protocol for both intra-cluster and inter-
cluster data transmission in the future work. [12] V. Mhatre, et. al., ”Design guidelines for wireless
sensor networks: communication, clustering and
Acknowledgments aggregation,” Ad Hoc Networks Joumul, Elsevier
The work is partly support by China 973 project Science, 2(1):45-63,2004.
(No. 2002CB312002) and T M P O Y T award of China