0% found this document useful (0 votes)

30 views

A Survey On Clustering Techniques

Uploaded by

jyothi

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

30 views

A Survey On Clustering Techniques

Uploaded by

jyothi

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 22

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/JIOT.2017.2726014, IEEE Internet of
Things Journal

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 1

A Survey of Clustering Techniques in WSNs and

Consideration of the Challenges of Applying Such to
5G IoT Scenarios
Lina Xu, Rem Collier, and Gregory M. P. O’Hare, Member, IEEE

Abstract—Wireless Sensor Network (WSN) systems are typ- batteries for extended periods of time. In order to prolong a
ically composed of thousands of sensors that are powered by networks longevity, clustering techniques have been introduced
limited energy resources. To extend the networks longevity, to achieve energy efficient communication between sensors.
clustering techniques have been introduced to enhance energy
efficiency. This paper presents a survey on clustering over the
last two decades. Existing protocols are analysed from a Quality
of Service (QoS) perspective including three common objectives,
those of energy efficiency, reliable communication and latency
awareness. This review reveals that QoS aware clustering de-
mands more attention. Furthermore, there is a need to clarify how
to improve Quality of user Experience (QoE) through clustering.
Understanding the users’ requirements is critical in intelligent
systems for the purpose of enabling the ability of supporting
diverse scenarios. User awareness or user oriented design is one
remaining challenging problem in clustering.
In additional, this paper discusses the potential challenges
of implementing clustering schemes to Internet of Things (IoT)
systems in 5G networks. We indicate that clustering techniques
enhanced with smart network selection solutions could highly
benefit the QoS and QoE in IoT. As the current studies for WSNs
are conducted either in homogeneous or low level heterogeneous
networks, they are not ideal or even not able to function in
highly dynamic IoT systems with a large range of user scenarios.
Moreover, when 5G is finally realised, the problem will become
more complex than that in traditional simplified WSNs. Several
challenges related to applying clustering techniques to IoT in 5G Fig. 1: An Example of Cluster-Based WSN
environment are presented and discussed.
Keywords—WSNs, Clustering, QoS, QoE, IoT, HetNets, 5G. For large WSN systems, topology control ought to be
applied to balance the network load, increase the network
I. I NTRODUCTION scalability and prolong the network lifetime. Clustering tech-
niques are also one of the approaches to topology control,
W IRELESS Sensor Networks (WSNs) are networks com-
posed of distributed micro-devices embedded with var-
ious sensing abilities (called sensors), which are used to mon-
which can organise a WSN into a cluster-based network. Task
scheduling, data gathering and Transmission Power Control
itor the environment and send the information back to the end (TCP) algorithms can be implemented in this structure in
users. WSN technologies were introduced more than 20 years order to achieve specific objectives. A clustering algorithm
ago and many projects have been proposed and undertaken can partition sensors into different clusters/groups, as shown in
that embrace this technology [1] [2]. Green computing [3] was Figure 1. In each cluster, a Cluster Header (CH) is elected to
introduced in 2008 with the purpose of reducing the use of be in charge of generating a transmission schedule, gathering
limited resources and maximising energy efficiency during the data from all the sensors in the cluster and transmitting the
lifetime of a system. WSNs typically include a large number assembled data back to the Based Station (BS). Based on the
of sensors that are equipped with limited energy resources, clustered structure, the system can maintain a longer life by
but are required to operate without recharging or replacing scheduling the duty cycle between the sensors within a cluster,
without harming the functionality of the network. In addition
Lina Xu, Rem Collier and Gregory M. P. O’Hare are with the Department to saving energy from scheduling, a sensor can also reduce
of Computer Science, University College Dublin, Ireland, e-mail: {lina.xu, energy consumption from communication since it only needs
rem.collier, gregory.ohare}@ucd.ie.
Copyright (c) 2012 IEEE. Personal use of this material is permitted. to communicate with a local CH rather than a far located BS.
However, permission to use this material for any other purposes must be Clustering algorithms are proposed and have become essential
obtained from the IEEE by sending a request to [email protected]. in WSNs primarily for two reasons:

2327-4662 (c) 2017 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/JIOT.2017.2726014, IEEE Internet of
Things Journal

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 2

1) Sleep scheduling: Scheduling the duty cycle between a specific combination of these parameters. An
the sensors in a cluster can not only prolong the adaptive CH election approach is designed to
network lifetime, but can also harmonise the power adapt to the variations of the network and the
usage throughout the network. environment.
2) Density control: Activating partial sensors in the net- A CH election algorithm can be either centralised or
work can lower the density of the network and in turn distributed. For centralised algorithms, the BS normally
avoid the conflict at the Medium Access Control (MAC) has access to the information pertaining to the entire
layer. Achieving this can reduce the communication la- network. A centralised algorithm can provide a global
tency and energy consumption caused by retransmission optimal solution, but may consume more energy and
[4]. time. For distributed algorithms, without the global in-
Engery saving is the most original temptation motivating formation of the network, the sensors can only achieve
clustering. This type of heuristic—reducing long-distance com- a local optimal performance based on the local infor-
munication ratio, inspires many other related studies. After- mation. However, they can be more reliable, energy and
wards, a wide body of clustering algorithms has become time efficient.
available. Different heuristics can be adopted depending on the 2) Cluster formation/CH selection: After being elected,
clustering purposes. However, the ultimate goal for clustering the CHs will advertise themselves by broadcasting their
techniques is to divide sensors in a WSN to different clusters. information to other sensors. Each sensor will gather the
After joining a cluster, a sensor normally only needs to information from all the CHs within its communication
communicate with its own CH. A CH can communicate with range and then decide which CH to join based on some
the BS directly or through other CHs, as shown in Figure 1. communication properties. Several metrics can be used
The routing between the sensors in the same cluster is called to determine the communication properties between a
intra-cluster routing. The routing between the CHs and the sensor and a CH, such as communication cost, hop
BS is called inter-cluster routing. The routing scheme can be count or even physical distance. In some cases, the
either single hop or multihop, and is dependent upon several size of clusters is also considered when sensors join
factors, such as the objectives of a clustering algorithm or the a cluster.
communication capability of the sensors. Algorithms with different CH election or cluster formation
In general, a clustering algorithm has two main phases: schemes will exhibit diverse performances. For example, com-
1) CH election: The CH is the leader of a cluster and it pared to random CH election scheme, energy aware CH
is in charge of gathering data and transmitting the data election, which can elect high residual energy sensors as CHs,
to the BS. The CHs will consume more energy than can balance the power usage throughout a network better.
normal sensors and therefore, run out of power sooner. However, the election process may need more energy and
CHs are normally rotated between different sensors to time. All clustering algorithms have their own specific CH
balance the power usage on each sensor. Which sensors election schemes and cluster formation schemes. Depending
ought to be elected as CHs requires careful investiga- on the requirements of a system, different strategies may be
tion. Generally, one of three CH election schemes is applied in order to achieve specific objectives.
adopted: The inter-cluster communication and intra-cluster commu-
a) Deterministic: In this scheme, the CHs are preset nication can be single hop or multihop. To realise multihop
and placed at fixed locations in a network. When communication in clustering, several algorithms, such as Max-
super nodes (with powerful processing ability and Min [5] and Khopca [6] are proposed. Max-Min algorithm
high energy storage) exist, the deterministic CH has introduced a heuristic to achieve k-hop cluster network
election can maintain the stability of the network structure with a time complexity of O(k) rounds, reducing from
and avoid the energy and time consumption as- O(n) rounds in earlier work [7] (where n is the number of
sociated with the frequent CH election. However, nodes). In this algorithm, a sensor can join a CH at most k
this case is not common since the sensors in fields wireless hops away.
are normally homogeneous and the super nodes Clustering is one of the major approach to green computing
can die for unexpected reasons. in WSNs which can be harnessed in many systems. Those
b) Random: CHs can be selected among sensors techniques can extend the longevity of a WSN through partially
based on randomly generated values. If a network activating the sensors in the network. Existing clustering
is homogeneous, random CH election scheme is algorithms, including review papers, focus on solving energy
a simple and beneficial strategy. efficiency problems, ignoring other Quality of Services (QoS)
c) Adaptive: Adaptive CH electing scheme provides requirements, like transmission reliability or network latency.
an alternative approach from that of the random Reviewing existing studies from the QoS perspective is re-
scheme. Instead of electing CHs based on random quired. Besides, being able to detect the user’s preference and
values, the adaptive CH election is based on some being aware of the system context also become more attractive
particular parameters, such as remaining residual features in intelligent systems.
energy or distance to the BS. With different Since the concept of Internet of Things (IoT) has been
system objectives, a clustering algorithm can use proposed, the corresponding systems are widely deployed to

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 3

assistant people’s everyday life. IoT systems differ from WSN to explore in this area. For the selected survey papers in Ta-
systems for its high diversity and usability. Instances for ble II, as you can see, they include a large number of clustering
IoT include smart home/buildings, connected cars, etc. With algorithms. There are also some new studies in recently two
the development of future networks/5G networks, migrating years (2015 and 2016) on clustering that are not covered by the
current IoT systems to those advanced communication plat- existing reviewing papers, such as RINtraR [36], SenCar [37],
form is the trend towards wireless. In this paper, we have FL-LEACH [38], BEEM [39] and PathQuality [40]. Survey
also discussed the differences between WSNs and IoT along paper [35] has stated that the existing heuristics in the state of
with the challenges of applying clustering techniques into IoT the art are diverse. However, most of them are following the
systems based on 5G networks. same idea and many clustering algorithms have similar heuris-
The remainder of this paper is organised as follows. Related tic. The essential parts for the most well accepted heuristics are
work is delivered in Section II. The clustering techniques control information exchanging and CH detection and selection
are analysed in details by classifying the existing approaches in 1-hop range. Information exchanging allows decentralised
into two categories: Voronoi structure based and Non-Voronoi clustering, which is the most adopted scheme in existing work.
structure based in Section III. Section IV presents QoS criteria In order to exchange information between nodes, different
in clustering algorithms from a network perspective, includ- systems may have their own requirements on the devices, such
ing energy efficiency, transmission reliability and network as synchronisation (strong, loose or none), location or energy
latency. Section V indicates that one QoS criterion from level awareness.
a user’s perspective—Quality of user Experience (QoE), is LEACH and similar algorithms are sharing the same heuris-
missing from current work. Since the clustering techniques tic, which requires time synchronisation and homogeneous de-
are normally utilised in a large network, in order to examine velopment. The heuristics are mainly determined by the main
the usability and scalability of an algorithm, a simulator is objectives of the clustering algorithms, such as load balancing,
normally be used. The advantages and disadvantages of the energy efficiency or mobility awareness. The common analysis
different simulators are compared in Section VI. Finally, a topics in existing survey papers include: convergence time,
comparison of 27 existing clustering algorithms is summarised node mobility, cluster overlapping, location awareness, energy
and several design guidelines are drawn from the comparison efficient, failure recovery, balanced cluster/cluster size, cluster
in Section VII. stability, cluster count, load balancing, deliver delay, intra/inter
Section VIII has revealed the major differences between routing schemes, objectives and complexity, etc. However,
WSN systems and IoT systems. Section X discusses the besides energy efficiency, other discussed metrics are mostly
possibility and challenges to apply clustering algorithms to IoT about topology.
systems and the potential challenges, specially when deployed In this paper, we are mainly focusing on QoS oriented
in 5G enabled environment. Conclusions of this paper is finally metrics to analysis clustering algorithm, including energy ef-
drawn in Section X. ficiency, transmission reliability and network latency. Besides,
we indicate that QoE awareness also requires more studies
when undergoing the transformation from WSN to IoT. We
TABLE I: List of Previous Survey Papers
aim to provide guidelines for researchers who are new in this
Year Survey papers area to make their design decisions.
2006 Arboleda 2006 [8], Younis 2006 [9]
2007 Ameer 2007 [10] III. C LUSTERING T ECHNIQUES
2008 Kumarawadu 2008 [11], Deosarkar 2008 [12]
Clustering algorithms normally structure networks into
2009 Jiang 2009 [13]
2010 Lotf 2010 [14], Boyinbode 2010 [15] Voronoi diagrams. However, some algorithms apply non-
2012 Naeimi 2012 [16], Aslam 2012 [17], Mundada 2012 [18] Voronoi structures, like chain or spectrum structures. In this
Ramesh 2012 [19], Liu 2012 [20] section, clustering techniques are classified into two categories
2013 Sudhanshu 2013 [21], Jan 2013 [22], Jain 2013 [23] based on the post network structure: Voronoi based approaches
Kumari 2013 [24], Jindal 2013 [25], Subha 2013 [26] and Non-Voronoi based approaches. The detailed implemen-
2014 Dhawan 2014 [27], Nayyar 2014 [28], Afsar 2014 [29], tations of several highly referenced clustering algorithms are
2015 Liu 2015 [30], Santar 2015 [31] also presented.
Ouafaa 2015 [32], Zanjireh 2015 [33]
2016 Pradhan 2016 [34] [35]
A. Beyond The Structure
Before we discuss clustering algorithms based on the fi-
nalised network structure, a brief introduction to clustering
II. R ELATED WORK heuristics is presented here. A heuristic is composed with the
We have reviewed most of the existing survey papers for information and rules used to form clusters. Those information
clustering in WSNs for the last decade, as shown in Table I. can be from the sensors themselves or data collected from
There are many overlapping studies and investigations. Many the network. Many clustering algorithms are sharing similar
of them lack of deep analysis and comprehensive introduction. heuristic with certain variations on how the information ex-
These four selected survey papers shown in Table II ([10], [20], change, or the metrics applied. The most popular heuristic is
[29] and [30]) are good to start with when researchers are about single hop inter-cluster and intra-cluster communication, like

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 4

TABLE II: Summaries for Selected Existing Survey Papers on Clustering for WSNs
Survey Paper Major contributions Covered Papers
Ameer [10] 1. Earlier clustering work before LEACH mainly 1997: Adaptive clustering [41]
2007 addressed node failure problem. Clustering 1998: CLUBS [42]
algorithms after LEACH start to focus 2000: LEACH [43]
more on energy efficiency. 2001: HCC [44], TEEN [45], MOBIC [46]
2. Multihop intra-cluster topology is rare. 2002: APTEEN[47], Deterministic [48], GS3 [49],
Multihop inter-cluster communication is well LEACH-C [50], PEGASIS [51]
applied. 2003: EEHC [52]
Afsar [29] 1. Energy efficiency/maxing lifetime is the 2004: ACE [53],FLOC [54], HEED [55], SEP [56]
2014 the dominate objective for clustering. 2005: BCDCP [57], CAWT, DWEHC [58], EDACH [59],
2. No reviewed clustering algorithm can operate EECS [60], EEUC [61], Ex-HEED [62], Gupta [63]
in a heterogeneous network. TL-LEACH [64], TASC, TTDD [65], UCS [66]
3. Multihop intra-cluster communication is rarely 2006: MOCA [67]
supported. While multihop inter-cluster 2007: CCS [68], CMEER [69], EACLE [70]
communication is well applied. EDC [71], PEACH [72], TCCA
4. The CH is only in charge of data aggregation 2008: EEMC [73], EEDUC [74], MRPUC [75]
and transmission. PRODUCE [76], S-WEB [77]
5. Device mobility is hardly concerned in solutions. 2009: EEDCF [78], PEBECS [79]
6. Distributed implementation is the mainstream 2010: CCN [80], EAUCF [81], PANEL [82], Unequal-LEACH [83]
in clustering. 2011: AWARE [84], EADUC [85], EC [86], EECABN [87]
Liu [20] [30] 1. Algorithms have small delivery delays LUCA [88], MBC [89], Spatial-clustering [90]
2012 tend to have low energy efficiency. 2012: EBCAG [91], EEBCDA [92], LEACH-DT [93]
2015 2. Distributed implementation is the mainstream TCAC [94], UHEED [95]
in clustering. 2013: ACDA [96], DSBCA [97], LCM [98]
3. Tree and train based implementations 2014: EEDC [99]
have lower scalability than grid based ones.

T
LEACH. The heuristic of LEACH as one of the very first reg (p) = dom (p, q)
energy efficiency studies in clustering, has motivated many q∈S−{p}
variations and extensions. Other heuristics may improve the From the definition, for any point x that x ∈ R3 ∩ x ∈ / S
existing one from several perspectives, such as balance of one can have δ (x, p) ≤ δ (x, q) where p ∈ S ∩ q ∈ S − {p} if
cluster size, maximal hop numbers, or catering for specific x ∈ reg (p). Here S is called the seed-point set, which is the
scenarios. Regardless of the heuristic that an approach adopts, CH set in a WSN.
the network structure constructed at the end, can be classified
into Voronoi based ones or Non-Voronoi based ones.

B. Voronoi Based Approaches

27 clustering algorithms are reviewed in this paper and 88%
of them are Voronoi based. Chain based and spectrum based
algorithms account for 8% and 4% respectively, as show in
Table III.
Voronoi diagrams (2-Dimensional or 3-Dimensional), as a
very important data structure in computational geometry, are
mainly used for solving clustering and scheduling problems
in computer science. The formal definition for 3D Voronoi
diagrams amended from [101] the definition for 2D is:
Let S denote a set of n points (called sites) in the Euclidean
space R3 . For two distinct sites p, q ∈ S, the dominance of p Fig. 2: Voronoi Diagrams
over q is defined as the subset of the space being at least as
close to p as to q. Formally
dom (p, q) = x ∈ R3 | δ (x, p) ≤ δ (x, q)
Figure 2 shows an example of a Voronoi diagram. The points
on the surface will join the closest point from set S (the black
for δ denoting the Euclidean distance function. The region dots). For example, M is the closest black dot for all the red
of a site p ∈ S is the portion of the space lying in all the dots in the cluster of M , compared to other black dots. Clus-
dominances of p over the remaining sites in S. Formally tering algorithms can totally adopt this idea, using Euclidean

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 5

TABLE III: Structure Review on Existing Clustering algo-

rithms
Algorithm Distributed Structure Intra Inter
cluster cluster
LEACH [43] Yes Voronoi 1-hop 1-hop
LEACH-C [50] No Voronoi 1-hop 1-hop
LEACH-TL [64] Yes Voronoi 1-hop 2-hop
HEED [55] Yes Voronoi 1-hop multihop
EECS [60] Yes Voronoi 1-hop 1-hop
EEHC [52] Yes Voronoi multihop 1-hop
DWEHC [58] Yes Voronoi 1-hop multihop
PANEL [82] Yes Voronoi multihop 1-hop
UCS [66] Yes Voronoi 1-hop multihop
EEUC [61] Yes Voronoi 1-hop multihop
ACE [53] Yes Voronoi 1-hop 1-hop
BCDCP [57] No Voronoi 1-hop multihop
PEGASIS [51] Yes Chain multihop 1-hop
one layer
TEEN [45] Yes Voronoi 1-hop multihop Fig. 3: Voronoi-Based Cluster Network
APTEEN [47] Yes Voronoi 1-hop multihop
TTDD [65] Yes Voronoi 1-hop multihop
CCS [68] Yes Chain multihop multihop
multi-layers C. Non-Voronoi Based Approaches
HGMR [100] Yes Voronoi 1-hop multihop
S-WEB [77] Hybrid Spectrum 1-hop multihop Chain and spectrum are two non-Voronoi structures that are
PEACH [72] Yes Voronoi 1-hop multihop used in clustering algorithms. In a chain structure based net-
MRPUC [75] Yes Voronoi 1-hop multihop work, no CHs are elected before the clusters are constructed.
MOCA [67] Yes Voronoi multihop 1-hop
FLOC [54] Yes Voronoi 1-hop 1-hop All the sensors are organised into chains as shown in Figure 4
Adaptive [41] Yes Voronoi 1-hop multihop from PEGASIS and Figure 5 from CCS. A sensor only needs
AWARE [84] Yes Voronoi multihop 1-hop to communicate with its left and right neighbours. The data
BEEM [39] Yes Voronoi 1-hop multihop flows in one direction. If a sensor receives data from its two
RINtraR [36] Yes Voronoi multihop 1-hop adjacent neighbours on the chain, it will transmit the data to
the BS, performing like a CH.
In PEGASIS [51], every sensor in the network transmits
distance as a partition metric. However, other metrics, such its data to one of its neighbours. Through this way, the
as energy consumption for communication and hop count, can gathered data is transferred from one node to another through
also be applied to replace the use of Euclidean distance. This a chain. A designated node (a node that receives data from
type of clustering algorithm is defined as Voronoi structure both neighbours) will send the assembled data back to the BS.
based approaches. This structure is convenient for relevant data PEGASIS is claimed as a distributed algorithm. However, it is
fusion and energy saving for two reasons: assumed that every single node has the global information of
1) The sensors close to each other have a better chance of the network.
having relevant data. CCS [68] is a centralised clustering algorithm based on
2) Generally the power consumption on communication PEGASIS. Instead of using a single chain structure, CCS
are low and the link quality is high. utilises a multihp chain structure, as shown in Figure 5.
Therefore, Voronoi diagrams are the dominating structure Regarding the BS as the centre of the network, each sensor
adopted in clustering algorithms. assigns itself a level number according to the signal strength
LEACH [43] and HEED [55] are two classic Voronoi received from the BS. Through this way, the sensors in the
structure based clustering algorithms. Several common phases network are organised into a hierarchy structure. For each
are included in Voronoi based clustering algorithms: level, the sensors perform transmission and fusion in the same
way as that in PAGASIS. The sensor that is elected as the
1) The first step is to construct the CH set. Then the CHs CH will gather data from all the other sensors on the same
will broadcast their information. After that, each sensor level and then transmit the assembled data to the CH in the
in the network will join one of the CHs based on the 1-lower level. Once being assigned a level, a sensor will not
received information. change its level unless the location of the BS changes. This
2) The sensors and their CH in the same cluster form structure suffers from a problem that the sensors near the BS
a star shape network as shown in Figure 3. In some can die soon from forwarding packets for the sensors in the
scenarios, the sensors can commutation with the CHs higher levels. Only total power consumption of the network
through multihop connections. is measured to evaluate this algorithm. There is no evidence
3) A sensor normally can only belong to one cluster at one showing a balanced power usage throughout the network.
time. There is no overlapping between different clusters. In a spectrum structure based network, the sensors are

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 6

Fig. 4: PEGASIS Single Layer Chain Structure From [20] Fig. 5: CCS Multi-layer Chain Structure From [20]

partitioned based on both the distance and the angle to the D. Influential Clustering Algorithms
BS. The angle is captured from a scanning sweep from the
BS at a specific time. Figure 6 shows the method used in A wide body of research on clustering algorithms for WSNs
the S-WEB clustering algorithm. S-WEB [77] is a spectrum is available; in particular, four noticeable survey papers have
structure based clustering algorithm. The first step in S-WEB been published, shown in Table II. In Table IV, a list of existing
is similar to that in CCS. All sensors are organised into clustering algorithms is analysed from the perspectives that are
layers based on their distance (measured by signal strength) not covered by the existing survey papers, including simula-
to the BS. Then the BS does a 360-degree scanning sweep tion environment, basic structure and QoS objectives (energy
by sending out a signal at one angel at a specific time. The efficiency, transmission reliability and network latency). For
sensors are clustered into cells based on their layer and the each algorithm, in order to evaluate their new approach, one or
scanning angle. In each cluster cell, the sensor with the highest more existing algorithms are normally used as a benchmark to
residual energy will be elected as the CH. All the sensors compare with. Table III shows the control manner (distributed,
have the responsibility of forwarding packets. The CHs may centralised or hybrid), structure, intra-cluster and inter-cluster
rotated between different sensors. However, the clusters are routing scheme of each algorithm. The algorithms that are
fixed during the entire operation of the network. The structure used as benchmarks are elaborated in the Table IV. As it
of the network is fixed after performing S-WEB, which cannot can be seen from Table IV, both TTDD and HGMR support
adapt to the dynamic changes in the network, like node failure energy efficient, reliable and latency aware communication.
or node death. The evaluation of S-WEB is only based on the However, TTDD is a solution based on a mobile BS that has
comparison with a non-cluster based routing—Direct Routing. a specific defined travel route. In such a case, the location
of the sensors, the user scenario and the network condition
have to be constant in order for TTDD to operate correctly.
HGMR can only cluster the sensors that are involved in the
current transmission. Furthermore, it uses multicast routing,
which could significantly increase the traffic load in large
networks. Therefore, LEACH and HEED are often selected
as two benchmarks in existing studies for several reasons:
1) LEACH and HEED are referred to as benchmarks in
44% and 20% of the 27 algorithms reviewed in this
paper respectively.
2) They are two classic clustering algorithms that have
inspired many other algorithms.
3) They are distributed algorithms and they have low
requirements on the applied WSNs.
4) They are convenient to deploy in real networks and easy
to extend in oder to cater to different scenarios.
5) They are Voronoi structure based, in which Transmis-
sion Power Control (TPC) can have the best perfor-
mance [102].
These two algorithms will be discussed in details. A detailed
Fig. 6: S-WEB Spectrum Cluster Structure [77] comparison between different algorithms will be drawn in
Section VII.
1) LEACH: Low-Energy Adaptive Clustering Hierarchy
(LEACH) provides an elegant approach to clustering routing

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 7

TABLE IV: QoS Review on Existing Clustering Algorithms

Algorithm Simulator Energy Reliability Latency QoE Awareness Benchmarks
LEACH MATLAB Yes No No No Direct, MTE, Static
LEACH-C NS-2 Yes No Yes No LEACH, Static
LEACH-TL NS-2 Yes No No No LEACH
HEED MATLAB∗ Yes No No No LEACH
EECS MATLAB Yes No No No LEACH
EEHC MATLAB∗ Yes No No No MAX-MIN-D
DWEHC NS-2 Yes No No No HEED
PANEL TOSSIM Yes No No No HEED
UCS MATLAB Yes No No No Equal cluster size
EEUC MATLAB∗ Yes No No No LEACH, HEED
ACE MATLAB∗ Yes No No No HCP
BCDCP MATLAB Yes No No No LEACH, LEACH-C, PEGASIS
PEGASIS MATLAB∗ Yes No Yes No Direct, LEACH
TEEN NS-2 Yes No No No LEACH, LEACH-C
APTEEN NS-2 Yes No No No LEACH, LEACH-C, TEEN
TTDD NS-2 Yes Yes Yes No no comparison
CCS MATLAB∗ Yes No No No PEGASIS
HGMR NS-2∗ Yes Yes Yes No HRPM, GMR
S-WEB MATLAB Yes No No No Short, Direct
PEACH MATLAB∗ Yes No No No EEUC, HEED, LEACH, PEGASIS
MRPUC MATLAB Yes No No No HEED
MOCA MATLAB Yes No No No No comparison
FLOC MATLAB No No No No No comparison
Adaptive MATLAB Yes Yes No No PRNET, Cluster TDMA
AWARE TOSSIM Yes Yes No No Unaware LEACH
RINtraR MATLAB Yes Yes No NO LEACH
BEEM MATLAB Yes No Yes NO LEACH, HEED
* indicates that the paper did not specify a simulator, but the simulation can be performed in the given one

that has inspired many adaptations, like LEACH-C, LEACH- no sensing data transmission in the set-up phase, the related
TL and HEED. It is a decentralised Voronoi structure based energy and time cost is considered as the overhead of LEACH.
clustering algorithm. Instead using Euclidean distance as a To minimise the impact of the overhead from CH election
metric, LEACH uses received signal strength. Most current and cluster forming, the steady-state phase is set long enough
radio chips provide a specific register to store the received compared to the set-up phase.
signal strength value, called Received Signal Strength Indicator In an ideal environment without any obstacles, the value of
(RSSI). This value will be updated once a new packet is RSSI between two sensors is determined by the distance and
received. LEACH necessitates 5 steps to construct the cluster the transmission power of the transmitter. The relation between
structure: RSSI and distance can be represented by the Log Distance Path
1) Each sensor elects itself to be a CH with a specific Loss Model [103]:
probability, which is set to be 5% in the experiments.
d
2) After the CHs have been elected, they will broadcast Pr (d) [dBm] = Pr (d0 ) [dBm] − 10η log − Xσ (1)
their information to the reset of the sensors in the d0
network. where Pr (d) is the received transmission power measured in
3) Based on the information received from all the CHs in dBm at distance d. It is also the value for RSSI. Pr (d0 ) is
its communication range, a sensor will decide which RSSI at the reference distance d0 . η, also referred to as Path
CH to join. Loss exponent, is a constant parameter value determined by
4) The CHs will create a transmission schedule for the the environment. For example, in a free space its value is 2;
sensors in their respective clusters. All the sensors in while in a well constructed building, its value can be from 4
the same cluster communicate with their CH in a single- to 6. Xσ is a zero-mean Gaussian distributed random variable
hop TDMA manner. (in dBm) with a standard deviation value of σ. This variable
5) The CHs collect and fuse the data from the sensors and is used only when there is a shadowing effect. Otherwise, this
then send the assembled data to the BS using long- variable will be zero.
distance CDMA communication. Figure 7 depicts one such example whereby five CHs are
The first four steps together are called the set-up phase and elected in the network. The sensors are clustered into different
the last one is called the steady-state phase. Since there is groups based on their RSSI values to all the five CHs. As

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 8

mentioned before, RSSI is determined by the distance in an propose their own solutions for clustering. It has become the
ideal environment without any obstacles or noise, if transmit- most popular clustering algorithm because of its extensibility.
ting data at a fixed transmission power. Therefore, in an ideal However, LEACH, is not an environment adaptive or user-
environment, the network structure after performing LEACH friendly protocol. In order to achieve specific expected perfor-
is a regular Voronoi diagram. In the real world, since RSSI mance, more work needs to be done to enhance the algorithm
is not totally determined by the distance due to the influence in order to be implemented in real WSNs.
of the environment, the network structure will be a analogous 2) HEED: Hybrid Energy-Efficient Distributed (HEED) is
Voronoi diagram instead of a strict one. a clustering algorithm based on LEACH that can support
multihop inter-cluster communication. Instead of using ran-
dom CH election scheme, HEED can select the sensors with
high residual energy to be the CHs through an iteration CH
election scheme. In this iteration CH election scheme, CHs are
classified into two types: tentative or f inal CHs. Some initial
CHs are selected randomly as tentative CHs. The probability
to become a CH is:
Eresidual
CHprob = Cprob (3)
Emax
where Eresidual is the current battery level and Emax is the
initial battery level. Cprob is the optimal probability that a
sensor will elect itself as a CH. However, Cprob is only used
to limit the number of the initial CHs and it has no direct
impact on the number of the f inal CHs. Cprob is set to be
5%, which is the same as that in LEACH.
The pseudocode of the CH iteration election process is
shown in Figure 8. After each iteration, every sensor doubles
Fig. 7: LEACH Clustering their probability of becoming a CH, which is CHprob . The
minimal value of CHprob is set to a specific threshold (Pmin =
10−4 ) to bound the number of iterations. If a sensor is not
A node becomes a CH for the current round if the randomly covered by any CHs and its CHprob is higher than a randomly
generated number is less than the following threshold: generated number, it will elect itself to be a tentative CH.
( If its Cprob has reached 1 and it is still not covered by any
P
1 if n ∈ G tentative or f inal CHs, it will claim itself as a f inal CH
T (n) = 1−P (r mod P ) (2) and then broadcast its information. A tentative CH becomes a
0 otherwise
f inal CH once its Cprob reaches 1. After its Cprob is increased
where G indicates the set of the nodes that have not been to 1, a sensor will terminate the iteration process. A tentative
elected as CHs in the last 1/P rounds. P is the desired percent- CH will give up to be a CH if it discovers a f inal CH in
age of the CHs. Through experimental analysis, LEACH has its communication range. At the end, the sensors that are not
proved that when P = 5%, the system has the best normalised covered by any f inal CHs will elect themselves as CHs.
energy dissipation. The iterative CH election can elect those sensors with
Compared to 1) direct transmission to the BS with full high residual energy to be CHs and therefore, balance the
transmission power 2) Minimum Transmission Energy (MTE) power usage throughout the network. To further reduce the
that uses minimum transmission power to complete the data transmission cost, a second parameter, that of, the energy
transmission and 3) static clustering with fixed CHs and consumption for intra-cluster communication, is considered
network structures, LEACH can provide a better load balancing when a sensor chooses its CH. This iteration CH election
throughout the network and an improved networks longevity. scheme guarantees that the probability of the phenomena that
LEACH assumes that every sensor has two communica- two sensors, within each other’s communication range, both
tion modes: short distance and long distance. If the RSSI become CHs is rare. Hence it can be deduced that the CHs are
between two sensors is higher than a threshold, the sensors distributed evenly in the network. In that case, all the clusters
can switch to short distance communication mode by lowering will have similar size.
the transmission power level. Otherwise long-distance CDMA A single-hop TDMA intra-cluster communication scheme is
communication mode is required in order to guarantee the applied, which is the same as that in LEACH. Different from
network connectivity. By long-distance communication, the LEACH, HEED adopts multihop inter-cluster communication
CHs can connect to the BS with direct single hop. In a cluster, scheme between the CHs and the BS. Significant overhead is
each sensor sends data to its CH in a TDMA manner to involved in HEED due to the heavy broadcast in each iteration.
avoid the collisions with others. Outside the clusters, the CHs Along with the energy decreasing while operating, the number
communicate with the BS by CDMA MAC protocol. of the iterations required in the initial phase consequently
The basic idea of LEACH has inspired many researchers to increases. Therefore, the clustering overhead and the network

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 9

Fig. 8: HEED Pseudocode [55]

delay are also increased. when designing related algorithms or protocols. How to extend
network longevity is one of the most challenging topics in
IV. Q O S IN C LUSTERING the WSN field. In general, it can be achieved from two
In general, Over The Top (OTT) applications have some perspectives:
specific Quality of Service (QoS) requirements for a given 1) Power usage reduction: If the power consumption
WSN. These requirements motivate the WSN algorithms and on each individual sensor is reduced, consequently the
protocols to achieve particular QoS objectives. The common lifetime of the entire network will be extended.
objectives in a system include network lifetime, transmission 2) Load balancing: The QoS of a system may be sig-
reliability and network latency. Depending on the user require- nificantly undermined if some sensors that have vital
ments, a system may be designed to function in different responsibilities for the network connectivity or coverage
ways. Nowadays, QoS awareness is essential in intelligent die. Therefore, balancing the energy consumption on
systems. [20] emphasises that QoS supported clustering should each sensor to balance the power usage throughout
be addressed in the future work. Table IV shows it is hard the network can maintain the QoS, which can improve
to consider all these three objectives and limited work has longevity from another perspective.
been done on that. The major reason for this is that clustering As shown in Table IV, with the exception of FLOC, all other
techniques are normally used to extend network lifetime. Other clustering algorithms have addressed the energy efficiency
issues are often ignored. Transmission reliability and network problem. FLOC primarily focuses on fast forming clusters
latency problems have as yet failed to drawn sufficient research locally instead of providing QoS clustering. Several techniques
attention. More work need to be done to support QoS in are normally utilised in existing clustering algorithms to pro-
clustering algorithms. long network lifetime.
• BS location: The BS is the final destination of all the
A. Network Lifetime data, so its location is critical to the overall performance.
Clustering techniques are used as a means to prolong Considering the location of the BS can optimise the
network lifetime. From when a system starts until the first/last clustering structure and therefore, further reduce the
sensor dies is normally referred to as the lifetime of a network. power consumption from transmission.
Network lifetime is crucial in some WSNs, such as wild • CH rotation: Since the CHs are in charge of data
life animal detection systems, as replacing sensors in these gathering, assembling and transmission, they normally
network is difficult and expensive. Such systems are required consume more energy than other normal sensors. To
to operate without user maintenance for a long time. Therefore, balance the energy consumption throughout the network,
longevity is always one of the most important considerations CH rotation between sensors is necessary.

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 10

• Energy aware CH election: The sensors with high on sensing, processing and communication. However, this
residual energy should be elected as CHs to further even estimation, also called power management, is complicated
the power usage. to be implemented on real sensors. By using the location
• Cluster size: The CH of a big cluster consumes more information, in DWEHC, the sensors are evenly distributed
energy than the one of a small cluster. To balance the into the clusters. Therefore, the power consumption on the
cluster size, the CHs should be well distributed in the CHs is better balanced, compared to that in HEED.
network. The chance that two sensors in each other’s From the above studies, it can be seen that sufficient work
communication range both become CHs should be low. has been done to improve longevity from the traditional
• Data fusion: To reduce the energy consumption on perspective of network lifetime. The coverage of the network
transmission, the CHs should be able to fuse the data is always neglected. BEEM has introduced a new concept,
and send the assembled data to the BS. coverage sensitive longevity, which not only evaluate the
• TPC: Through TPC, a sensor can communicate with network longevity using the number of alive sensors but also
its CH with the minimal transmission power level while using the network coverage. We argue this concept should
maintaining the transmission quality. replace the transitional definition for network lifetime.
• Node density: Some work assumes that the sensors are
deployed uniformly. The node density is a constant value
in each single unit area. However, in most real world
WSN systems, the sensors are randomly distributed.
Node density in some area may be higher than that in
some other area. The sensors in a low density area will
have a bigger impact on the network coverage than the
ones in a high density area. Taking node density into
account in clustering algorithms can balance the sensor
distribution and maintain the network coverage [39].
It is difficult to cover all of the techniques in one single
algorithm. Since the CHs normally consume more energy
than the normal sensors, the CH election scheme has a big
impact on the power consumption in a network. This explains
why most existing work put a lot of effort into the CH
election schemes. Several typical algorithms are summarised in
Table V from the perspective of energy efficiency. LEACH and Fig. 9: Current Available LQEs [104].
HEED are two classic Voronoi based clustering approaches.
DWEHC improved HEED by supporting multihop inter-cluster
and intra-cluster communication. BEEM have proposed a new
perspective of network longevity—coverage sensitive cluster- B. Network Transmission Reliability
ing. Regardless of the number of sensors, if losing network In general, clustering algorithms are designed with specific
coverage, a system will lose part of the sensing data. That objectives in mind, for example, lifetime extending, load
data may be critical in terms of the overall performance of the balancing or scalability increasing [10]; however, transmission
system. Therefore, coverage sensitive longevity is essential in reliability is often overlooked. Little existing work considers
WSN systems. In addition to energy efficiency, some systems transmission reliability as a criterion when evaluating the
also have high requirements on transmission reliability. For performances of the approaches. For many WSN deployments,
example, in a disaster monitoring system, once an event is such as disaster or military scenarios, communication reliabil-
detected, the information has to be transmitted successfully. ity is essential in terms of QoS. Besides, for some civil ap-
Otherwise, serious damage may happen. In real-time systems, plications, like critical infrastructure monitoring, transmission
such as multimedia or road lighting, latency awareness is reliability is a crucial metric. In such systems, transmission
essential for user experience. Coverage sensitive longevity, quality has a high priority from the user’s perspective. How-
reliable transmission and latency awareness are important ever, enabling reliable transmission in cluster-based networks
issues in terms of QoS. BEEM was introduced with an aim to has received limited attention. Individual clustering algorithms
increase WSN coverage sensitive longevity. use contrasting metrics to decompose the networks into inter-
PEGASIS and CCS are chain based approaches. S-WEB connected clusters. Examples of such metrics include distance,
has a spectrum based structure. PEGASIS has analysed the hop count and cluster size [15]; however, link quality metrics
impact of the BS’s location on the performance of an al- are not considered.
gorithm, but it fails to account the location of the BS as a [104] has literally reviewed most of the available Link
factor in clustering. Since CCS and S-WEB both have layer Quality Estimators (LQEs) that can be used as link quality
based network structure, the location of the BS can affect the metrics. The current available LQEs are shown in Figure 9.
clustering results. HEED indicates that the sensors do not need Packet Receive Rate (PRR), Radio Signal Strength Indicator
any special component to measure the residual power. The (RSSI) and Link Quality Indicator (LQI) are three common
battery level can be estimated from the energy consumption metrics used to estimate link quality. PRR is computed as

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 11

TABLE V: Energy Efficiency Techniques Used in Clustering Algorithms

Algorithm LEACH HEED DWEHC BEEM PEGASIS CCS S-WEB
Structure Voronoi Voronoi Voronoi Voronoi Chain Chain Spectrum
Distributed Yes Yes Yes Yes Yes No No
Node Density No No No Yes No No No
CH Rotation Yes Yes Yes Yes Yes Yes No
Data Fusion Yes Yes Yes Yes Yes Yes No
TPC Yes Yes Yes Yes No No No
Energy Aware No Yes Yes Yes Yes Yes Yes
Cluster size No Yes Yes Yes No No No
Sensing None Energy Energy Energy Location Engery Energy
Abilities Location Location Location
Simulation MATLAB MATLAB NS-2 MATLAB MATLAB MATLAB MATLAB
Benchmarks Direct LEACH HEED LEACH Direct PEGASIS Direct
MTE, Static HEED LEACH

the ratio of the number of successfully received packets to C. Network Latency

the number of transmitted packets. RSSI indicates the signal Low network latency is an essential criterion in some WSN
strength of the received packets. Its value is stored in the RSSI systems, such as road lighting systems. Once detecting a
register, which is available in most popular radio chips. LQI car on the road, the system needs to turn on the lights
meanwhile presents the correctness of the received packets. For immediately. In such a latency sensitive system, maintaining
the CC2420, LQI is measured based on the first eight symbols network delay under an acceptable threshold is one of the
of a received packet. LQI is a good LQE when a large quality major objectives. In a cluster-based network, the sensors need
of data is available. In contrast, a good RSSI estimation can to transmit data to the CHs and then the CHs can transmit
be obtained over a small number of measurements and can the data to the BS. Therefore, the network latency of the data
converge quicker than LQI [105]. Compared to RSSI and LQI, transmission in a cluster-based network is longer than that in a
PRR has a higher correlation with transmission quality. direct transmission based network. The cluster-based structure
can extend network lifetime, but conversely increase network
latency. For these reasons, for latency sensitive systems based
LEACH uses Radio Signal Strength Indicator (RSSI) as on clustering structure, the objective normally is to support
a metric to cluster sensors. However, LEACH adopts TCP latency awareness, rather than necessarily to minimise it. QoE
scheme, using RSSI to reduce energy consumption as an awareness in such a circumstance is to be able to customise
energy metric rather than a LQE metric. Moreover, LEACH the network latency to meet users’ requirements.
assumes that all the sensors have long-distance single hop Table IV shows that most existing work fails to consider
communication ability to avoid the complexity of multihop latency when evaluating the performance of their approaches.
communication. TTDD presents an approach that can provide Centralised clustering algorithm LEACH-C can balance the
a reliable transmission with the help of the mobile BS. As cluster size by distributing the CHs evenly over the network.
mentioned before, HGMR is a routing protocol rather than a From the experimental results of the throughput per unit
clustering algorithm. Even though HGMR claims that it can energy and the throughput per unit time, it can be seen that
provide reliable transmission, it cannot be applied to the entire LEACH-C is more energy and latency efficient, compared to
network. Although AWARE concludes that their solution can LEACH. PEGASIS is a power efficient and delay sensitive
achieve a better Packet Receive Rate (PRR), it is not clear algorithm. It states that minimising energy or delay in isolation
why this is the case. Additionally, the design of AWARE has drawbacks on the performance of the system. Therefore,
does not explicitly address the issue of reliable transmission. PEGASIS uses energy×delay as a metric to select two out of
Adaptive clustering provides a code division access scheme for its neighbours to form a chain structure. In the chain structure,
multimedia systems based on mobile radio networks, rather each node can assemble the data that comes in from one
than based on WSNs. Using the standby routing, Adaptive neighbour on the chain with its own data and then transmit the
clustering shows a low average packets loss rate but high assembled data to the other neighbour. PEGASIS indicates that
end-to-end network delay. RINtraR has proposed a solution the network delay for a packet is dominated by the number of
for intra-clustering routing to improve transmission quality by transmission times (or called hop count) since 1) there are no
transmitting data through high quality multihop route rather queuing delays and 2) the processing and propagation delays
than poor quality single hop route. Even though clustering are negligible compared to the multihop transmission delay.
techniques have been a hot research topic for more than twenty TTDD also provides for short network latency clustering by
years, reliable transmission supported clustering has not drawn utilising the mobility of the BS. It cannot be implemented in a
enough attention. Providing transmission reliability for intra- network with a fixed BS. HGMR is a single-hop inter-cluster
cluster or inter-cluster routing requires more work in the future. and intra-cluster based routing algorithm. By using multicast, it

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 12

can support short latency transmission. HGMR is not specified VI. S IMULATION
as a clustering algorithm. With the exception of PEGASIS, no Since clustering techniques are usually utilised in a large-
other algorithms account latency as a metric when performing scale network and it is not convenient to deploy a real one for
clustering. Although the experimental performances may show testing, the evaluation of an algorithm is normally undertaken
good results on network latency, the reason and some further in a simulation. The performance of a proposed solution in the
analysis are still missing. real network can be predicted from the simulation results. The
Several approaches in clustering that are not covered by the selected simulator should not only be applicable for testing,
existing surveys can also support short latency communication. but also convenient to use and modified.
[106] adopts the same idea as LEACH-C, through balancing
cluster size to reduce network latency. [107] uses several
multiple mobile BSs to collect data from the network. By A. Comparison of Simulators Utilised in Clustering
utilising the border nodes, [108] has proposed a solution that Within this paper, 27 different clustering algorithms are
can provide alternative routes for some sensors in the network under discussed. All of the 27 clustering algorithms tested
to reduce latency. As a result, the border nodes will consume their approaches in simulations. 60% of the tests are based on
more energy than other normal sensors. [109] is focusing on MATLAB, followed by 28% based on NS-2. The remaining
fast cluster formation instead of fast data transmission. [110] experiments were implemented in TOSSIM or other simulation
can provide low latency communication for the heterogeneous tools constructed by the researchers themselves. MATLAB and
networks where supper nodes are deployed. The BS is assumed NS-2 are commonly used simulation tools in clustering [116]
to locate in the centre of the sensing field and the super [117]. [117] reviewed all the papers published in SENSOR-
nodes are deployed around the BS in a uniform manner. In COMM 2011 that were not limited in clustering area. Its shows
a heterogeneous network with supper nodes deployed, no CH that MATLAB and NS-2 are still the most used simulators.
election is necessary since the supper nodes are the default Since MATLAB is convenient to use and easy to extend to
CHs. [111] and [112] have proposed solutions to provide short add a power measurement component, the evaluations for most
latency scheduling rather than short latency communication. of the existing work were performed in MATLAB or can be
[113] reveals that the geographical locations of the sensors and accomplished in MATLAB (if not specified) when only energy
the node density both have an impact on the network latency consumption is being evaluated.
in a cluster-based network. As a system grows, more QoS services will be required from
the clustering algorithms. To evaluate the QoS performance
of an algorithm, the applied simulator should either support
the measurements for communication properties or can be ex-
V. C ONTEXT AWARE /U SER C ONFIGURABLE F EATURES IN tended to add the corresponding functionalities. Some metrics
C LUSTERING that are used in clustering algorithms can only be captured
from the communication connection between sensors, such as
Besides the three QoS objectives discussed in Section IV, Radio Signal Strength Indicator (RSSI). RSSI is often used
Context awareness and user configurability are also important as a Link Quality Estimator (LQE). Ideally, it is determined
criteria from the user side. Those are often used to evaluate by transmission power, distance and other parameters, such as
QoE. As indicated in Table IV, currently, none of the existing antenna height and environment impact factor. However, in the
clustering algorithms can provide an interface for the users to real world, it is highly influenced by the outside surroundings,
interact with the systems. [114] presents a solution to achieve like obstacles and noises.
user configurability by integrating the improvements on net- The comparison of three popular simulators, which are
work lifetime, transmission reliability and network latency. MATLAB, NS-2 and TOSSIM, is presented in Table VI.
Since clustering is a major approach to energy efficiency in Neither TOSSIM nor NS-2 have a power model that can
WSNs [115], most of the existing work has emphasis on simulate power consumption in WSNs. Since there is no
energy saving, ignoring the diverse requirements from the radio communication is supported in MATLAB, the RSSI
users. The users normally expect that the WSN performs in value between two sensors is computed from the distance by
a particular way, through which to benefit their applications adopting Log Distance Path Loss Model [103]. In this model,
furthest. For example, in a multimedia system, the users RSSI is mainly determined by the distance and transmission
want immediate response in a cluster-based network rather power. This model can only be applied to the networks that are
than simply reducing power consumption. On the other hand, deployed in an ideal environment, where no obstacles or noise
the users may highlight transmission reliability in a disaster exist. NS-2 is a general network simulator, which has a radio
detection system. Scenarios such as these involve a tradeoff communication model. However, it is not specially designed
between competing demands. Therefore, power consumption is for WSNs. The simulation processing becomes slow when the
no longer the only consideration. An advanced WSN should be network scales over to 100 nodes [116]. In NS-2, the RSSI
sufficiently intelligent to understand the users’ preferences and value is calculated from the Friis transmission equation, which
adapt to these changes. Currently, there is no work showing any is only accurate for long-distance communication scenarios.
interest in adapting to user preference in clustering algorithms. Now NS-3 is becoming more popular and it is should be used
User scenarios and system context are overlooked, without to replace NS-2 . TOSSIM is specially designed to simulate
which QoE awareness is hard to be guaranteed. WSNs that are composed with only TinyOS nodes. Currently,

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 13

TABLE VI: Comparison of Existing Simulators

Simulator RSSI Specified Power Limitations
calculation for WSN Model
MATLAB Log Distance Path No Numeric No radio
Loss Model, calculated Model communication
from distance
NS-2 Friis equation, calculated No No Slow for large
from distance, network
antenna height
TOSSIM Not supported Yes No Constrained to
Berkeley nodes

TABLE VII: Energy Dissipated in MATLAB

Protocols Transmission/Receiving (Eelect ) Transmit Amplifier (amp )
LEACH 50nJ/bit 100pJ/bit/m2
HEED 50nJ/bit 10pJ/bit/m2 (d < d0 )
0.0013pJ/bit/m4 (d > d0 )

only sensors like Micaz, TelosB that are based on TinyOS. The power consumption [50]. The calculation is specified as:
simulator allows the users to test and verify the code that will
run on real sensors. However, it is more focusing on examining ET x = Eelec ∗ k + amp ∗ k ∗ dn (5)
the performance of an individual node rather than the whole When d < d0 , amp = 10pJ/bit/m2 and n = 2. When d >
network, and the sensor platform is constrained to a specific d0 , amp = 0.0013pJ/bit/m4 and n = 4. The value of d0 is
hardware type. TOSSIM is based on TinyOS, which has a deep a constant distance, which is determined by the surrounding
learning curve. This is also the primary reason why it is not environment. HEED has also adopted this power consumption
widely used. Besides, it can neither model RF cancellation nor model.
support RSSI value generation.
As a result of the above analysis, MATLAB is the most VII. C OMPARISON OF E XISTING C LUSTERING
popular one for simulating WSN clustering systems for the A LGORITHMS
following reasons: Many survey papers have been published during the last
1) MATLAB can adopt Log Distance Path Loss Model to decade. The comparison between different algorithms is gen-
calculate RSSI values in order to be used in algorithms. erally focusing on clustering objectives, CH election, intra-
2) It provides a numeric power consumption model, which cluster and inter-cluster routing. Table IV shows the compar-
can be easily modified to fit into different scenarios. ison of cluster structure, QoS supported situation, simulation
3) The processing time for large networks is quick. environment and benchmark algorithms in parallel experiments
4) It is not constrained to a specific type of sensors or for the 27 clustering algorithms reviewed in this paper. These
networks. A more general evaluation can be performed. criteria are not included in the existing survey papers. The
5) MATLAB suits for agile programming. New models or statistics analysis of the 27 clustering algorithms for network
components can be implemented and plugged in freely. structure, control manner, simulation environment, intra-cluster
and inter-cluster routing is shown in Figure 10.
From the comparison of existing work, several conclusions,
B. Power Consumption Model in MATLAB
which will contribute to construct the solutions in later studies,
LEACH indicates that different assumptions about the radio are drawn:
characteristics will affect the performance of an algorithm. Two 1) Voronoi diagrams are easy to implement and convenient
most referenced clustering algorithms, LEACH and HEED to maintain as a structure for clustering, comparing with
adopted different models to compute energy consumption other structures. In a Voronoi structure, the sensors in
for long-distance communication. A shown in Table VII, in the same cluster are normally physically close to each
LEACH, the calculation of transmission power consumption other. Those sensors will have a high chance to have
is simplified as: relevant data. The communication cost between close
ET x = Eelec ∗ k + amp ∗ k ∗ d2 (4) sensors can be low.
2) Distributed clustering manner is more tolerant and
where k is the message length measured in bits and d is the robust.
distance from the transmitter to the receiver. In the future work, 3) The most often used simulator is MATLAB for its
LEACH research team provided an advanced study on radio shallow learning curve and high extensibility.

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 14

(a) Structure (b) Control Manner (c) Simulator (d) Intra-Cluster Routing (e) Inter-Cluster Routing

Fig. 10: Statistics Analysis of 27 Studied Clustering Algorithms

4) Multihop inter-cluster communication is essential to VIII. F ROM WSN S TO I OT

guarantee the connectivity of a network without en- As it is already presented in Section I, a WSN is a
abling long-distance communication ability on the sen- network composed of autonomous wireless micro-devices,
sors. which can monitor the surrounding environment, record the
5) LEACH and HEED are two popular algorithms that are data and transmit the information throughout the network
referred to as benchmarks in evaluations. to a central Base Station (BS), which is also sometimes
referred to as a sink node [1] [2]. WSNs were originally
Besides, some problems existing in current work are also proposed for military surveillance purposes [118]. Due to
unveiled: its first success, this technology was then envisioned for
other purposes, such as habitat monitoring [119], weather
1) Multihop inter-clustering routing is implemented in monitoring [120], agriculture monitoring [121] and wildlife
many existing work. However, multihop intra-cluster monitoring [122]. WSN technologies were introduced more
communication is not well supported. than 20 years ago and many projects have been proposed and
2) As is evident, in QoS supported clustering algorithms, undertaken that embrace this technology. However, more effort
energy efficiency is well accomplished, while transmis- was still required before WSNs became a truly mainstream
sion reliability and network delay are less studied. technology. WSN technologies envisaged as enabling another
3) None of the 27 algorithms can provide interface for digital revolution in order to facilitate people’s daily life. Only
the users to interact with the algorithms in order to a small number of users were showing interest in WSN area.
customise the performance on energy efficiency, net- Because of the complexities associated with it, application
work delay and transmission reliability. QoE awareness developers were required to be domain experts to build quality
is missing. services based on WSNs. For this reason, only limited usages
and few applications were available for normal users. Hence,
Through the above analysis and comparison, it is convincing researchers were facing the dilemma that WSN technologies
that a Voronoi based distributed clustering algorithm, which could benefit and advance society, but bringing them from the
can support QoS and QoE awareness, enabling multihop intra- laboratory to the marketplace.
cluster and inter-cluster routing, is demanded. Specifically,
A similar and successful paradigm to compare against is
we conclude that user oriented/QoE aware design and QoS
the case of the Internet. The Internet was invented in the late
supported services should attract more attention in the future
1960s. However, it did not become universally popular until
study of clustering for WSNs.
1995 when the Internet access was more free and convenient.
In the following sections, the audience will amazingly find The number of Internet users has increased impressively for
that those problems in clustering techniques for WSNs are the last decade owing to numerous OTT applications developed
exactly aligned with what we need to address applying such for it, such as World Wide Web, electronic mail and social
techniques to IoT systems in 5G environment. As WSNs networking1 . The Internet has brought great convenience to
evolve into IoT and communication technologies march to society and its importance is self-evident.
5G, clustering techniques also need to grow new charms in The revolution that is necessary to ensure WSN technologies
order to adapt to those changes. In the next sections, firstly flourish in a similar manner requires more effort in simplifying
the transformation between WSNs and IoT is presented. The design, implementation, deployment and usability. The societal
possibility and advantages to migrate IoT systems to 5G impact of WSNs is defined as the number of the users, which
platform is also discussed. We further present the challenges itself is determined by the quality and quantity of the avail-
that researchers are facing if applying clustering techniques in
IoT in 5G. 1 http://www.internetworldstats.com/stats.htm

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 15

able applications. Currently, it is extremely urgent to inspire migrating IoT systems to 5G platform have been discussed a
developers to build more useful OTT applications to improve lot from a general perspective [125]. In this section, we will
people’s life quality or experience. Therefore, the concepts show the necessity of clustering techniques in IoT systems that
of Internet of Things (IoT) [123] and Internet of Everything are based on 5G networks.
(IoE) [124] were proposed and now they are extremely popular.
The instances include smart buildings/home, smart cities, smart A. An Overview on 5G
roads and connected cars. WSNs were the former existence and
the technology foundation for IoT. The fundamental difference For the first time in history LTE has brought the entire
between WSNs and IoT is that the dynamics and diversities in mobile industry to a single technology footprint resulting in
IoT are much higher than that in WSNs, which includes the unprecedented economies of scale. The converged footprint
following aspects: of LTE has made it an attractive technology baseline for
several segments that had traditionally operated outside the
1) Applications: Traditional WSN applications are mainly
commercial cellular domain. There is a growing demand for
focusing on monitoring the environment and then col-
a more versatile Machine-to-Machine (M2M) platform. The
lecting the data. The sensors are deployed and fixed
challenge for industrial is the lack of convergence across the
in the field. The volume of the data transmitted in the
M2M architecture design that has not materialised yet. It is
network is small. However, IoT applications can range
expected that LTE will remain as the baseline technology for
from smart kitchen to smart city monitoring. All the
wide area broadband coverage also in the 5G area.
applications in these systems have specific requirements
on the network performance. For instance, smart road
lighting applications in smart road system require short
network delays. Smart monitoring applications in smart
city system require high network bandwidth.
2) IoT devices: WSN systems normally have homoge-
neous sensors, in terms of sensing ability, processing
capability and power supply. This feature has simplified
the design of the relevant protocols, such as routing
protocols. However, for IoT systems, this feature is no
longer applied. For example, in a smart home system,
the fridge, the vacuuming robot and the light sensing
devices all have their own characteristics in this system.
Furthermore, the devices in IoT systems are no longer
limited to those basic sensors. Fig. 11: An Overview on 5G Network
3) Communication ability: Since the features of the
connected devices in IoT systems vary greatly, their
communication ability, including communication range, Figure 11 presents an overview of 5G network infrastructure.
communication band, communication power consump- Mobile operators now aim to create a blend of pre-existing
tion can be different. technologies covering 2G, 3G, 4G, WiFi and others to allow
4) Number of connected devices: By the end of 2020, higher coverage and availability, and higher network density
according to Cisco, there will be over 50 million in terms of cells and devices with the key differentiator
connected devices. Managing such a large number of being greater connectivity as an enabler for M2M services
devices along with the generated data is challenging in [126]. New machine type communication technologies are also
IoT systems. invited, such as LTE-M and NB-IoT. An array of antennae
5) Objectives: As mentioned, WSN systems are mainly supporting High-order MIMO (Multi-Input, Multi-Output) is
used for monitoring environment and collecting data. installed in a device and multiple radio connections are estab-
User profile and system context are rarely under dis- lished between the device and the cellular base station allowing
cussion. One of the most important objectives for IoT paralleled data transmission. Meanwhile, operators, vendors
systems is to improve people’s life and their personal and academia are combining efforts to explore technical so-
experience. User oriented and context aware design is lutions for 5G that could use frequencies above 6GHz and
demanded in IoT systems. reportedly as high as 300GHz. This platform will need to
provide a network management and control layer to coordinate
IX. C LUSTERING FOR I OT S YSTEMS IN 5G the activities from the application layer and the services from
Among the current clustering strategies for WSNs, many of underlying infrastructure. This layer should be implemented
them assume the sensors in the networks are homogeneous. between network transport layer and application layer. It
If a heterogeneous network is considered, the degree of the provides functionalities/components as such network selection,
diversity is still rather low (For example, some heterogeneous traffic monitoring and user analysis, etc. For example, smart
network simply assumes that part of the sensors have double network selection can be implement as the following way:
power supplies than others). This assumption has become unre- The network selection component can match OTT usage to
alistic recently, especially in IoT systems. The challenges when a suitable network interface based on the characteristics of

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 16

the OTT application itself, the network conditions and the on the transmission links. Reliability can also be improved
user profile. The traffic monitoring component can provide through cognitive radio [151].
the network condition information. As you can see, those As we can see from the above evidence, 5G can provide
components will work together to achieve intelligent 5G. a feasible and reliable backhaul infrastructure for many IoT
On top of the 5G network, many systems can be supported, systems. It is sensible to design our future protocols for
for example, smart homes and enterprise networks. In those IoT systems based on that infrastructure, keeping QoS/QoE
systems, running multiple applications at the same time is also awareness as one design principle.
common. The network selection should know the applications
and user scenarios in order to allocate suitable underlining
network services for communication. This information can B. The Need For Clustering for IoT in 5G
be accomplished by the user analysis component. Besides,
the network context itself, such as congestion level, is also As cellular involved backhaul can provide full connectivity,
important when selecting the network interface. Therefore, it is becoming the main trend. As show in Figure 12, there
traffic monitoring is necessary. 5G aims to enhance the degree are normally three layers in the communication system. Layer
of automatic adaptation and configuration. User oriented and 1 is composed with the sensors/devices in the field. Layer 2
QoE aware design is one of the main tasks. The management is deployed by the mobile operators beside the field, normally
layer should decouple software functions from the hardware in the form of small or micro cells, supporting 3GPP standard
resource layer. Meanwhile, it also needs to provide network communication. Layer 3 is the Evolved Packet Core (EPC) net-
performance analysis and optimisation. The 5G networks aim work (core network), where all the information and data would
to advance IoT technologies further in the market [127]. be collected. The backhaul infrastructure is considered as the
3GPP standard/5G based backhaul has become popular as combination of layer 2 and layer 3. Layer 1 is referred to as the
a great solution for connectivity problem in IoT systems. last hop in wireless communication. 3GPP communication can
[128] indicates that the next generation of mobile networks be applied to all these three layers. The M2M communication
(5G), will need not only to develop new radio interfaces only exists in Layer 1 and Layer 2. Figure 12 shows a standard
or waveforms to cope with the expected traffic growth but infrastructure for 3GPP involved IoT backhaul. Nowadays, the
also to integrate heterogeneous networks from End to End 3GPP organisation, mobile operators and academics are trying
(E2E) with distributed cloud resources to deliver E2E IoT to realise 5G by the end of 2020. 5G with an enabler for
and mobile services. [129] has provided a backhaul solution M2M communication, has features such as 1 to 10Gbps speed,
through mobile network for smart building applications. The 1 millisecond latency, 100% coverage and reliability, which
proposed network architecture will improve services for users aims to support and provide good QoE for a large range of
and also will offer new opportunities for both service providers applications and usages. Hence, this infrastructure shown in
and network operators. [130] utilises LTE uplink resource as Figure 12 has been well accepted and studied as mentioned
the communication backhaul for IoT systems. [131] uses LTE- in Section IX-A. [152] has introduced a novel two-stage re-
Femtocell composed networks to collect data from health- clustering algorithm to reduce high load on cells in hotspot
caring IoT systems. They consider a heterogeneous network areas and improve user satisfaction. Smart coordinating mul-
in which a macro-cell tier is overlaid with a very dense tier tiple access points can improve the overall spectral efficiency.
of small cells. [132] presents three scheduling algorithms for In such a 3GPP involved IoT backhaul, we indicate that
a two-tier HetNet to maximise network throughput of IoT clustering techniques are still necessary and beneficial for the
devices. [133] highlights new applications for optical wireless following reasons:
communication (OWC) as a mobile backhaul for WiFi, LTE 1) Energy efficiency: In IoT systems, many sensors/de-
and 5G and as a new access technology in IoT systems where vices are still deployed remotely, requiring lifetime
it enables secure and reliable communications at low latency. for years. 5G aims to provide M2M communication,
The next generation networks/5G is definitely the ulti- allowing devices to have up to 10 years’ battery life.
mate means to connect everything together, specially for IoT Therefore, energy efficiency still is a challenging prob-
systems [134] [135]. Small cells take a huge role in such lem in IoT systems regarding of QoS. In addition, from
infrastructure [136] [137]. Besides, Software Defined Network green computing perspective, with 50 billion connected
(SDN) and Network Function Virtualisation (NFV) techniques devices, if one device could reduce 1% energy con-
are also widely used in such scenarios [138] [139] to improve sumption, it can save the world 1 billion dollars’ worth
the scalability. Smart backhaul solutions have been proposed electricity. [153] has introduced a clustering algorithm
to improve users’ utility/QoE [140] [141]. Combining cloud for IoT systems in MIMO scenarios to extend network
technology with the backhaul is also becoming dominant [142] longevity and maintain network coverage.
[143] [144]. [145] indicates that network selection/resource 2) Distributed processing: In recently 10 years, big data
allocation in 5G network is a challenging research topic has been a really hot topic and people generally expect
that needs to be addressed. In addition to the infrastructure, treasures in the data. However, not all the data is useful.
resource management in the backhaul is also critical [146]. Treasure hunting in massive amounts of meaningless
[147] [148] and [149] have proposed corresponding solutions data can be costing. In addition, data transmission
for that. [150] aims to balance the workload of different remote through the network and data maintaining in the server
radio heads (RRHs) in the backhaul to alleviate the pressure are also expensive. It is essential to filter out worthless

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 17

or redundant data from the source, rather than transmit- systems. With 3GPP standard communication, the deployment
ting it back to the core network. becomes more flexible and the connectivity problem is solved
3) Management hierarchy: As we have discussed, the straightway. When migrating IoT systems to 5G networks, if
major difference between WSNs and IoT is the diver- applied, several challenging problems in clustering are needed
sities from the devices themselves to the OTT applica- to be addressed.
tions. Besides, with such a large number of devices, an The first challenge comes from the fundamental nature of
extensible and dynamic hierarchic structure can achieve IoT systems—the vastly diversity. The things in the field
effective and efficient management. are highly heterogeneous. Some of the nodes can have low
To relieve the problems caused by the above issues in IoT capabilities and some of them can be extremely advanced.
systems, clustering techniques can still be applied. As shown in In order to connect everything in smart cities, cheap and en-
Figure 13, the three layers are still organised as before. How- ergy efficient transmit-only-devices will be massively deployed
ever, at layer 1, the clusters are formed and one CH is elected [154] alone with other highly advanced sensors. Due to the
for each cluster. The CH in each cluster will still be in charge hardware limitation, those devices cannot receive information
of data gathering and confusion. With clustering in the field, from others, which makes decentralised clustering challenging.
firstly, the M2M communication is only happening in layer 1. One opposite example to a transmit-only-device is a super
The number of cross layer communication is largely reduced, sensors. It is a more powerful node with the task of collecting
which in turn saves a lot of energy on the devices. Secondly, sensing data and, in some cases, with the physical capability
clustering structure is beneficial for data gathering and data to act as a relay node [155]. Because of their high strength
processing locally. The CH can dismiss redundant data and in memory and computing abilities, they normally are treated
avoid overloading the 3GPP backhaul network. Thirdly, the as CHs in networks. Therefore, comparing with traditional
cluster structure in layer 1 can be further stratified, based on WSN, IoT systems are more complex and comprehensive. The
the size of the network, the device types, application types, OTT applications are no longer just about data collection and
communication abilities, etc. The hierarchy structure can be transmission. We need to consider more complicated scenarios
extended deeper depending on the requirements. and user cases. When clustering, from the perspective of
reducing redundant data, sensors/devices with similar usage
should be grouped together.
The second challenge is that cost for transmission. The
energy cost is still a critical concern in IoT systems when
deployed in 5G networks. Besides, since mobile network is
also involved, the financial cost should be well controlled.
For example, the use of LTE will be more expensive than
that of WiFi. In practice, LTE can be used as the default
communication means. If possible (the CH is in range and
the user requirements can be satisfied), some of the devices
can switch to Bluetooth or ZigBee, which are much power
saving solutions. In such a condition, in order to utilise
Fig. 12: 3GPP Assisted IoT Backhaul WITHOUT Clustering. MIMO techniques, the devices in the same cluster should well
distributed to use different network interfaces. This approach
cannot only avoid interfering but also balance the load on those
networks.
The third challenge is how to improve user utility. QoE has
been emphasised greatly when formulating 5G related policies
and regulations. Users are requiring specialised services ac-
cording to their behaviours. In order to cater for characterised
usage, user profile should be in consideration. The user profile
can include information such as, user priority, user behaviours,
user scenarios, etc. The first issue needs to be addressed is
how to formularise user utility as a measurable metric. Then
the clustering scheme should consider the user requirements
for the networks. For example, if TDMA MAC protocol is
applied in each cluster, it is not a good approach to group all
Fig. 13: 3GPP Assisted IoT Backhaul WITH Clustering. the users that have high requirements on latency into the same
cluster. Those users should be in a relevantly smaller cluster.
In extreme case, they should be able to transmit with the layer
2 access points directly.
C. Challenges in Clustering Towards 5G The fourth challenge is how to utilise the intelligent com-
As we have discussed in the last section, clustering tech- ponents in the core network. In order to make the 5G network
niques are required and beneficial even for highly dynamic IoT smart and be aware of the context, extra components in

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 18

the core network have been proposed to provide additional especially when WSNs are evolving to IoT, with an aim to
functionalities. The information from those components will support a larger number of users and more complex user
be available for devices to query. For example, a network scenarios. It is extremely challenging if considering a higher
monitoring component can detect the congestion levels for all degree of heterogeneous for the devices and the networks.
the network access interfaces. Be aware of this information QoE supported features, such as user configurability, should
should also help the designed clustering algorithms to be be implemented in order to advance clustering techniques
smart. Supposedly a device needs high network throughput for to again adaptivity to various conditions and user scenarios.
the application and uses LTE as the default communication Clustering algorithms now need to cater for multiple and
interfaces. It should not switch to WiFi once it receives differing scenarios with varying user preferences, which is
notifications from the core network indicating that currently essential when implementing a smart clustering strategy to
WiFi network is congested [156]. understand the environment and the users. To the end, we
The fifth challenge is how to manage and utilise mobility in discussed the need for clustering techniques in IoT systems
the networks. Some work has been done specifically addressing and the challenges when migrating to 5G platforms. Future
mobility problem in WSNs [155] [157] [158]. With the devel- research directions are also outlined indicating where this
opment of IoT systems, supporting the connectivity of the fast technique should progress in order to profit IoT systems in
moving things (such as vehicles in the connected cars system) the future network.
and utilising the mobility to improve the communication
efficiency are challenging. R EFERENCES
One of the general objectives for 5G is to improve QoS and [1] C.-Y. Chong and S. Kumar, “Sensor networks: evolution, opportunities,
QoE. Clustering techniques, in order to adapt to 5G and more and challenges,” Proceedings of the IEEE, vol. 91, no. 8, pp. 1247–
complicated scenarios in IoT systems, the above challenges 1256, Aug 2003.
need to be addressed. The corresponding research in those [2] I. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “Wireless
directions should be further investigated. Cross layer design sensor networks: a survey,” Computer Networks, vol. 38, no. 4, pp.
393 – 422, 2002.
is also highly recommended to address multiple issues col-
[3] S. Murugesan, “Harnessing green it: Principles and practices,” IT
laboratively. We sincerely indicate that more advanced studies Professional, vol. 10, no. 1, pp. 24–33, Jan. 2008.
should be undertaken in the current scenario with 5G rather
[4] L. Xu, D. Delaney, G. O’Hare, and R. Collier, “The impact of
than the traditional WSN usages. transmission power control in wireless sensor networks,” in Network
Computing and Applications (NCA), 2013 12th IEEE International
X. C ONCLUSION Symposium on, 2013, pp. 255–258.
[5] A. D. Amis, R. Prakash, T. H. Vuong, and D. T. Huynh, “Max-
In this paper, a general-to-specific review of clustering min d-cluster formation in wireless ad hoc networks,” in INFOCOM
algorithms in WSNs was conducted. Firstly, we presented a 2000. Nineteenth Annual Joint Conference of the IEEE Computer and
summarised survey for clustering techniques in WSNs. The Communications Societies. Proceedings. IEEE, vol. 1. IEEE, 2000,
services provided by existing algorithms were analysed from pp. 32–41.
four QoS angles: network lifetime, transmission reliability, [6] M. R. Brust, H. Frey, and S. Rothkugel, “Adaptive multi-hop clus-
tering in mobile networks,” in Proceedings of the 4th international
network latency and the QoE awareness perspective. Upon conference on mobile technology, applications, and systems and the
this analysis and the comparison, several findings have been 1st international symposium on Computer human interaction in mobile
revealed: technology. ACM, 2007, pp. 132–138.
1) Limited work concerns network coverage when evalu- [7] D. Baker and A. Ephremides, “The architectural organization of a
ating network lifetime. mobile radio network via a distributed algorithm,” IEEE Transactions
on communications, vol. 29, no. 11, pp. 1694–1701, 1981.
2) Latency awareness and transmission reliability are not
[8] L. M. C. Arboleda and N. Nasser, “Comparison of clustering algo-
well supported in clustering. rithms and protocols for wireless sensor networks,” in 2006 Canadian
3) User scenario/profile awareness has drawn little atten- Conference on Electrical and Computer Engineering, May 2006, pp.
tion in current clustering research. 1787–1792.
4) Limited work shows interest for clustering in heteroge- [9] O. Younis, M. Krunz, and S. Ramasubramanian, “Node clustering
neous networks with high degree of diversities. in wireless sensor networks: recent developments and deployment
challenges,” IEEE Network, vol. 20, no. 3, pp. 20–25, May 2006.
These findings should motivate future work in clustering.
[10] A. A. Abbasi and M. Younis, “A survey on clustering algorithms for
As QoS has become a crucial evaluator in intelligent WSN wireless sensor networks,” Computer Communications, vol. 30, 2007.
systems. Unfortunately, existing work only proposes solutions [11] P. Kumarawadu, D. J. Dechene, M. Luccini, and A. Sauer, “Algorithms
to extend network lifetime from a narrow perspective. More- for node clustering in wireless sensor networks: A survey,” in 2008
over, limited work has been undertaken in clustering with a 4th International Conference on Information and Automation for
main objective to improve transmission reliability or reduce Sustainability, Dec 2008, pp. 295–300.
network latency. However, being able to optimise network per- [12] B. P. Deosarkar, N. S. Yadav, and R. P. Yadav, “Clusterhead selection
formance from one single dimension among energy efficiency, in clustering algorithms for wireless sensor networks: A survey,” in
2008 International Conference on Computing, Communication and
transmission quality and network latency is not sufficient to Networking, Dec 2008, pp. 1–8.
enable intelligence in clustering. It has been stated that an [13] C. Jiang, D. Yuan, and Y. Zhao, “Towards clustering algorithms in
intelligent system should be able to understand the user’s wireless sensor networks-a survey,” in 2009 IEEE Wireless Communi-
requirements and adapt to the changes within the system, cations and Networking Conference, April 2009, pp. 1–6.

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 19

[14] J. J. Lotf, M. Hosseinzadeh, and R. M. Alguliev, “Hierarchical routing [33] M. M. Zanjireh and H. Larijani, “A survey on centralised and dis-
in wireless sensor networks: a survey,” in 2010 2nd International tributed clustering routing algorithms for wsns,” in 2015 IEEE 81st
Conference on Computer Engineering and Technology, vol. 3, April Vehicular Technology Conference (VTC Spring), May 2015, pp. 1–6.
2010, pp. V3–650–V3–654. [34] N. Pradhan, K. Sharma, and V. K. Singh, “A survey on hierarchical
[15] O. Boyinbode, H. Le, A. Mbogho, M. Takizawa, and R. Poliah, clustering algorithm for wireless sensor networks,” ENERGY, vol. 134,
“A survey on clustering algorithms for wireless sensor networks,” in no. 4, 2016.
Network-Based Information Systems (NBiS), 2010. [35] V. Sucasas, A. Radwan, H. Marques, J. Rodriguez, S. Vahid, and
[16] S. Naeimi, H. Ghafghazi, C.-O. Chow, and H. Ishii, “A survey on the R. Tafazolli, “A survey on clustering techniques for cooperative
taxonomy of cluster-based routing protocols for homogeneous wireless wireless networks,” Ad Hoc Networks, vol. 47, pp. 53–81, 2016.
sensor networks,” Sensors, vol. 12, no. 6, pp. 7350–7409, 2012. [36] L. Xu, M. O’Grady, G. O’Hare, and R. Collier, “Reliable multihop
[17] M. Aslam, N. Javaid, A. Rahim, U. Nazir, A. Bibi, and Z. A. Khan, intra-cluster communication for wireless sensor networks,” in Com-
“Survey of extended leach-based clustering routing protocols for wire- puting, Networking and Communications (ICNC), 2014 International
less sensor networks,” in 2012 IEEE 14th International Conference Conference on, Feb 2014, pp. 858–863.
on High Performance Computing and Communication 2012 IEEE 9th [37] M. Zhao, Y. Yang, and C. Wang, “Mobile data gathering with
International Conference on Embedded Software and Systems, June load balanced clustering and dual data uploading in wireless sensor
2012, pp. 1232–1238. networks,” IEEE Transactions on Mobile Computing, vol. 14, no. 4,
[18] M. R. Mundada, S. Kiran, S. Khobanna, R. N. Varsha, and S. A. pp. 770–785, 2015.
George, “A study on energy efficient routing protocols in wireless [38] P. Nayak and A. Devulapalli, “A fuzzy logic-based clustering algorithm
sensor networks,” International Journal of Distributed and Parallel for wsn to extend the network lifetime,” IEEE Sensors Journal, vol. 16,
Systems (IJDPS) Vol, vol. 3, pp. 311–330, 2012. no. 1, pp. 137–144, Jan 2016.
[19] K. Ramesh and D. K. Somasundaram, “A comparative study of [39] L. Xu, G. O’Hare, and R. Collier, “A balanced energy-efficient
clusterhead selection algorithms in wireless sensor networks,” arXiv multihop clustering scheme for wireless sensor network,” 7th IFIP
preprint arXiv:1205.1673, 2012. Wireless and Mobile Networking Conference (WMNC), 2014.
[20] X. Liu, “A survey on clustering routing protocols in wireless sensor [40] G. Sirisha, R. B. Babu, and K. R. Rao, “Establishing path quality
networks,” Sensors, vol. 12, no. 8, 2012. management in wireless sensor networks through cluster head deter-
[21] S. Tyagi and N. Kumar, “A systematic review on clustering and routing mination,” Indian Journal of Science and Technology, vol. 9, no. 5,
techniques based upon leach protocol for wireless sensor networks,” 2016.
Journal of Network and Computer Applications, vol. 36, no. 2, pp. [41] C. R. Lin and M. Gerla, “Adaptive clustering for mobile wireless net-
623–645, 2013. works,” IEEE Journal on Selected Areas in Communications, vol. 15,
[22] M. A. Jan and M. Khan, “A survey of cluster-based hierarchical routing pp. 1265–1275, 1997.
protocols,” IRACST–International Journal of Computer Networks and [42] R. Nagpal and D. Coore, “An algorithm for group formation and
Wireless Communications (IJCNWC), vol. 3, no. 2, pp. 138–143, 2013. maximal independent set in an amorphous computer,” Cambridge, MA,
[23] N. Jain, P. Sinha, and S. Gupta, “Clustering protocols in wireless sen- USA, Tech. Rep., 1998.
sor networks: A survey,” International Journal of Applied Information [43] W. Heinzelman, A. Chandrakasan, and H. Balakrishnan, “Energy-
System (IJAIS), vol. 5, no. 2, 2013. efficient communication protocol for wireless microsensor networks,”
[24] P. Kumari, M. P. Singh, and P. Kumar, “Survey of clustering algorithms in System Sciences, 2000. Proceedings of the 33rd Annual Hawaii
using fuzzy logic in wireless sensor network,” in 2013 International International Conference on, jan. 2000, p. 10 pp. vol.2.
Conference on Energy Efficient Technologies for Sustainability, April [44] S. Banerjee and S. Khuller, “A clustering scheme for hierarchical
2013, pp. 924–928. control in multi-hop wireless networks,” in INFOCOM 2001. Twentieth
[25] P. Jindal and V. Gupta, “Study of energy efficient routing protocols annual joint conference of the IEEE computer and communications
of wireless sensor networks and their further researches: a survey,” J societies. Proceedings. IEEE, vol. 2. IEEE, 2001, pp. 1028–1037.
Comput Sci Commun Eng, vol. 2, 2013. [45] A. Manjeshwar and D. Agrawal, “Teen: a routing protocol for
[26] C. Subha, S. Malarkan, and K. Vaithinathan, “A survey on energy enhanced efficiency in wireless sensor networks,” in Parallel and
efficient neural network based clustering models in wireless sensor Distributed Processing Symposium., Proceedings 15th International,
networks,” in Emerging Trends in VLSI, Embedded System, Nano Elec- 2001, pp. 2009–2015.
tronics and Telecommunication System (ICEVENT), 2013 International [46] P. Basu, N. Khan, and T. D. Little, “A mobility based metric for
Conference on. IEEE, 2013, pp. 1–6. clustering in mobile ad hoc networks,” in Distributed computing
[27] H. Dhawan and S. Waraich, “A comparative study on leach routing systems workshop, 2001 international conference on. IEEE, 2001,
protocol and its variants in wireless sensor networks: a survey,” pp. 413–418.
International Journal of Computer Applications, vol. 95, no. 8, 2014. [47] A. Manjeshwar and D. Agrawal, “Apteen: a hybrid protocol for
[28] A. Nayyar and A. Gupta, “A comprehensive review of cluster-based efficient routing and comprehensive information retrieval in wireless,”
energy efficient routing protocols in wireless sensor networks,” IJR- in Parallel and Distributed Processing Symposium., Proceedings In-
CCT, vol. 3, no. 1, pp. 104–110, 2014. ternational, IPDPS 2002, Abstracts and CD-ROM, 2002, pp. 8 pp–.
[29] M. M. Afsar and M.-H. Tayarani-N, “Clustering in sensor networks: [48] B. P. Deosarkar, N. S. Yadav, and R. Yadav, “Clusterhead selection
A literature survey,” Journal of Network and Computer Applications, in clustering algorithms for wireless sensor networks: A survey,”
vol. 46, pp. 198 – 226, 2014. in Computing, Communication and Networking, 2008. ICCCN 2008.
[30] X. Liu, “Atypical hierarchical routing protocols for wireless sensor International Conference on. IEEE, 2008, pp. 1–8.
networks: A review,” IEEE Sensors Journal, vol. 15, no. 10, pp. 5372– [49] Z. Hongwei and A. Arora, “Gs3: Scalable self-configuration and self-
5383, Oct 2015. healing in wireless networks,” in Proceedings of the 21 Annual Sym-
[31] S. P. Singh and S. Sharma, “A survey on cluster based routing protocols posium on Principles of Distributed Computing, Mobile and Wireless
in wireless sensor networks,” Procedia Computer Science, vol. 45, pp. Communications Network, 2002.
687 – 695, 2015. [50] W. Heinzelman, A. Chandrakasan, and H. Balakrishnan, “An
[32] I. Ouafaa, L. Jalal, K. Salah-ddine, and E. H. Said, “The comparison application-specific protocol architecture for wireless microsensor net-
study of hierarchical routing protocols for ad-hoc and wireless sensor works,” Wireless Communications, IEEE Transactions on, vol. 1, no. 4,
networks: A literature survey,” in Proceedings of the The International pp. 660–670, 2002.
Conference on Engineering & MIS 2015. ACM, 2015, p. 32.

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 20

[51] S. Lindsey, C. Raghavendra, and K. M. Sivalingam, “Data gathering Communication Technology, The 9th International Conference on,
algorithms in sensor networks using energy metrics,” IEEE Trans. vol. 1, 2007, pp. 260–265.
Parallel Distrib. Syst., vol. 13, no. 9, pp. 924–935, Sep. 2002. [69] T. Kang, J. Yun, H. Lee, I. Lee, H. Kim, B. Lee, B. Lee, and K. Han,
[52] S. Bandyopadhyay and E. Coyle, “An energy efficient hierarchical “A clustering method for energy efficient routing in wireless sensor
clustering algorithm for wireless sensor networks,” in INFOCOM networks,” in Proceedings of the 6th WSEAS International Conference
2003. Twenty-Second Annual Joint Conference of the IEEE Computer on Electronics, Hardware, Wireless and Optical Communications.
and Communications. IEEE Societies, vol. 3, 2003, pp. 1713–1723 World scientific and engineering academy and society (WSEAS),
vol.3. 2007, pp. 133–138.
[53] H. Chan and A. Perrig, “Ace: An emergent algorithm for highly [70] K. Yanagihara, J. Taketsugu, K. Fukui, S. Fukunaga, S. Hara, and K.-i.
uniform cluster formation,” in Wireless Sensor Networks, ser. Lecture Kitayama, “Eacle: Energy-aware clustering scheme with transmission
Notes in Computer Science, H. Karl, A. Wolisz, and A. Willig, Eds. power control for sensor networks,” Wireless Personal Communica-
Springer Berlin Heidelberg, 2004, vol. 2920, pp. 154–171. tions, vol. 40, no. 3, pp. 401–415, 2007.
[54] M. Demirbas, A. Arora, and V. Mittal, “Floc: A fast local clustering [71] Q. Chen, J. Ma, Y. Zhu, D. Zhang, and L. M. Ni, “An energy-efficient
service for wireless sensor networks,” in Workshop on Dependability k-hop clustering framework for wireless sensor networks,” in European
Issues in Wireless Ad Hoc Networks and Sensor Networks (DIWAN- Conference on Wireless Sensor Networks. Springer, 2007, pp. 17–33.
S/DSN, 2004. [72] S. Yi, J. Heo, Y. Cho, and J. Hong, “Peach: Power-efficient and
[55] O. Younis and S. Fahmy, “Heed: a hybrid, energy-efficient, distributed adaptive clustering hierarchy protocol for wireless sensor networks,”
clustering approach for ad hoc sensor networks,” Mobile Computing, Computer Communications, vol. 30, no. 1415, pp. 2842 – 2852, 2007.
IEEE Transactions on, vol. 3, no. 4, pp. 366–379, 2004. [73] Y. Jin, L. Wang, Y. Kim, and X. Yang, “Eemc: An energy-efficient
[56] G. Smaragdakis, I. Matta, and A. Bestavros, “Sep: A stable election multi-level clustering algorithm for large-scale wireless sensor net-
protocol for clustered heterogeneous wireless sensor networks,” Boston works,” Computer Networks, vol. 52, no. 3, pp. 542–562, 2008.
University Computer Science Department, Tech. Rep., 2004. [74] S. Lee, J. Lee, H. Sin, S. Yoo, S. Lee, J. Lee, Y. Lee, and S. Kim,
[57] S. Muruganathan, D. Ma, R. Bhasin, and A. Fapojuwo, “A central- “An energy-efficient distributed unequal clustering protocol for wire-
ized energy-efficient routing protocol for wireless sensor networks,” less sensor networks,” World Academy of Science, Engineering and
Communications Magazine, IEEE, vol. 43, no. 3, pp. S8–13, 2005. Technology, vol. 48, pp. 443–447, 2008.
[58] P. Ding, J. Holliday, and A. Celik, “Distributed energy-efficient hier- [75] B. Gong, L. Li, S. Wang, and X. Zhou, “Multihop routing protocol
archical clustering for wireless sensor networks,” in Proceedings of with unequal clustering for wireless sensor networks,” in Computing,
the First IEEE international conference on Distributed Computing in Communication, Control, and Management, 2008. CCCM ’08. ISECS
Sensor Systems, ser. DCOSS’05. Berlin, Heidelberg: Springer-Verlag, International Colloquium on, vol. 2, 2008, pp. 552–556.
2005, pp. 322–339. [76] J.-H. Kim, S. H. Chauhdary, W.-C. Yang, D.-S. Kim, and M.-S.
[59] K. T. Kim and H. Y. Youn, “Energy-driven adaptive clustering hierar- Park, “Produce: a probability-driven unequal clustering mechanism
chy (edach) for wireless sensor networks,” in International conference for wireless sensor networks,” in Advanced Information Networking
on embedded and ubiquitous computing. Springer, 2005, pp. 1098– and Applications-Workshops, 2008. AINAW 2008. 22nd International
1107. Conference on. IEEE, 2008, pp. 928–933.
[60] M. Ye, C. Li, G. Chen, and J. Wu, “Eecs: an energy efficient clustering [77] H. Le, D. Hoang, and R. Poliah, “S-web: An efficient and self-
scheme in wireless sensor networks,” in Performance, Computing, organizing wireless sensor network model,” in Network-Based Infor-
and Communications Conference, 2005. IPCCC 2005. 24th IEEE mation Systems, ser. Lecture Notes in Computer Science, M. Takizawa,
International, 2005, pp. 535–540. L. Barolli, and T. Enokido, Eds. Springer Berlin Heidelberg, 2008,
[61] C. Li, M. Ye, G. Chen, and J. Wu, “An energy-efficient unequal vol. 5186, pp. 179–188.
clustering mechanism for wireless sensor networks,” in Mobile Adhoc [78] T. Kaur and J. Baek, “A strategic deployment and cluster-header selec-
and Sensor Systems Conference, 2005. IEEE International Conference tion for wireless sensor networks,” IEEE Transactions on Consumer
on, 2005, pp. 8 pp.–604. Electronics, vol. 55, no. 4, pp. 1890–1897, Nov 2009.
[62] H. Huang and J. Wu, “A probabilistic clustering algorithm in wireless [79] Y. Wang, T. Yang, and D. Zhang, “An energy efficient and balance
sensor networks,” in IEEE Vehicular Technology Conference, vol. 62, hierarchical unequal clustering algorithm for large scale sensor net-
no. 3. IEEE, 2005, p. 1796. work,” Inform. Technol. J, vol. 8, no. 1, pp. 28–38, 2009.
[63] I. Gupta, D. Riordan, and S. Sampalli, “Cluster-head election using [80] N. Shigei, H. Morishita, and H. Miyajima, “Energy efficient clustering
fuzzy logic for wireless sensor networks,” in Communication Networks communication based on number of neighbors for wireless sensor net-
and Services Research Conference, 2005. Proceedings of the 3rd works,” in Proceedings of International MultiConference of Engineers
Annual. IEEE, 2005, pp. 255–260. and Computer Scientists. Berlin: Springer-Verlag, 2010.
[64] V. Loscri, G. Morabito, and S. Marano, “A two-levels hierarchy for [81] H. Bagci and A. Yazici, “An energy aware fuzzy unequal clustering
low-energy adaptive clustering hierarchy (tl-leach),” in Vehicular Tech- algorithm for wireless sensor networks,” in Fuzzy systems (FUZZ),
nology Conference, 2005. VTC-2005-Fall. 2005 IEEE 62nd, vol. 3, 2010 IEEE international conference on. IEEE, 2010, pp. 1–8.
2005, pp. 1809–1813. [82] L. Buttyn and P. Schaffer, “Panel: Position-based aggregator node
[65] H. Luo, F. Ye, J. Cheng, S. Lu, and L. Zhang, “Ttdd: two-tier data election in wireless sensor networks,” in in Proc. of the 4th IEEE
dissemination in large-scale wireless sensor networks,” Wirel. Netw., International Conference on Mobile Ad-Hoc and Sensor Systems,
vol. 11, no. 1-2, pp. 161–175, Jan. 2005. PisaItaly, 2007.
[66] S. Soro and W. Heinzelman, “Prolonging the lifetime of wireless [83] P. Ren, J. Qian, L. Li, Z. Zhao, and X. Li, “Unequal clustering scheme
sensor networks via unequal clustering,” in Parallel and Distributed based leach for wireless sensor networks,” in Genetic and Evolution-
Processing Symposium, 2005. Proceedings. 19th IEEE International, ary Computing (ICGEC), 2010 Fourth International Conference on.
2005, pp. 8 pp.–. IEEE, 2010, pp. 90–93.
[67] A. Youssef, M. Younis, M. Youssef, and A. Agrawala, “Wsn16-5: [84] I. Urteaga, N. Yu, N. Hubbell, and Q. Han, “Aware: Activity aware
Distributed formation of overlapping multi-hop clusters in wireless network clustering for wireless sensor networks,” in Local Computer
sensor networks,” in Global Telecommunications Conference, 2006. Networks (LCN), 2011 IEEE 36th Conference on, 2011, pp. 589–596.
GLOBECOM ’06. IEEE, 2006, pp. 1–6. [85] J. Yu, Y. Qi, G. Wang, Q. Guo, and X. Gu, “An energy-aware
[68] S.-M. Jung, Y.-J. Han, and T.-M. Chung, “The concentric clustering distributed unequal clustering protocol for wireless sensor networks,”
scheme for efficient energy consumption in the pegasis,” in Advanced International Journal of Distributed Sensor Networks, 2011.

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 21

[86] D. Wei, Y. Jin, S. Vural, K. Moessner, and R. Tafazolli, “An energy- [107] W. M. Aioffi, C. A. Valle, G. R. Mateus, and A. S. da Cunha,
efficient clustering solution for wireless sensor networks,” IEEE trans- “Balancing message delivery latency and network lifetime through
actions on wireless communications, vol. 10, no. 11, pp. 3973–3983, an integrated model for clustering and routing in wireless sensor
2011. networks,” Computer Networks, vol. 55, no. 13, pp. 2803 – 2820,
[87] W. Zhou, “Energy efficient clustering algorithm based on neighbors 2011.
for wireless sensor networks,” Journal of Shanghai University (English [108] A. Tufail, “Reliable latency-aware routing for clustered wsns,” Interna-
Edition), vol. 15, pp. 150–153, 2011. tional Journal of Distributed Sensor Networks, vol. 2012, no. 681273,
[88] S. Lee, H. Choe, B. Park, Y. Song, and C.-k. Kim, “Luca: An energy- 2012.
efficient unequal clustering algorithm using location information for [109] T. Zhong, S. Wang, S. Xu, H. Yu, and D. Xu, “Time delay based
wireless sensor networks,” Wireless Personal Communications, vol. 56, clustering in wireless sensor networks,” in Wireless Communications
no. 4, pp. 715–731, 2011. and Networking Conference, 2007.WCNC 2007. IEEE, March 2007,
[89] S. Deng, J. Li, and L. Shen, “Mobility-based clustering protocol for pp. 3956–3960.
wireless sensor networks with mobile nodes,” IET wireless sensor [110] M. Yao, C. Lin, Y. Tian, L. Wu, and Y. Chen, “Energy and delay
systems, vol. 1, no. 1, pp. 39–47, 2011. minimization in cluster-based wireless sensor networks,” in Green
[90] Y. Ma, Y. Guo, X. Tian, and M. Ghanem, “Distributed clustering- Computing and Communications (GreenCom), 2012 IEEE Interna-
based aggregation algorithm for spatial correlated sensor networks,” tional Conference on, Nov 2012, pp. 588–594.
IEEE Sensors Journal, vol. 11, no. 3, pp. 641–648, 2011. [111] O. Dousse, P. Mannersalo, and P. Thiran, “Latency of wireless sensor
[91] T. Liu, Q. Li, and P. Liang, “An energy-balancing clustering approach networks with uncoordinated power saving mechanisms,” in in Pro-
for gradient-based routing in wireless sensor networks,” Computer ceedings of Mobihoc, 2004, 2004, pp. 109–120.
Communications, vol. 35, no. 17, pp. 2150–2161, 2012. [112] Y. Li, L. Guo, and S. Prasad, “An energy-efficient distributed algo-
[92] J. Yuea, W. Zhang, W. Xiao, D. Tang, and J. Tang, “Energy efficient rithm for minimum-latency aggregation scheduling in wireless sensor
and balanced cluster-based data aggregation algorithm for wireless networks,” in Distributed Computing Systems (ICDCS), 2010 IEEE
sensor networks,” Procedia Engineering, vol. 29, pp. 2009–2015, 30th International Conference on, June 2010, pp. 827–836.
2012. [113] D. Padmanabhan and F. Labeau, “Effect of location on the latency in
[93] S. H. Kang and T. Nguyen, “Distance based thresholds for cluster head cluster-based wsns,” in Sensors, 2013 IEEE, Nov 2013, pp. 1–4.
selection in wireless sensor networks,” IEEE Communications Letters, [114] L. Xu, D. Lillis, G. M. O’Hare, and R. W. Collier, “A user configurable
vol. 16, no. 9, pp. 1396–1399, 2012. metric for clustering in wireless sensor networks.” in SENSORNETS,
[94] D. P. Dahnil, Y. P. Singh, and C. K. Ho, “Topology-controlled adaptive 2014, pp. 221–226.
clustering for uniformity and increased lifetime in wireless sensor [115] N. Heo and P. Varshney, “Energy-efficient deployment of intelligent
networks,” IET Wireless Sensor Systems, vol. 2, no. 4, pp. 318–327, mobile sensor networks,” Systems, Man and Cybernetics, vol. 35, no. 1,
2012. pp. 78–92, 2005.
[95] E. Ever, R. Luchmun, L. Mostarda, A. Navarra, and P. Shah, “Uheed [116] M. Korkalainen, M. Sallinen, N. Karkkainen, and P. Tukeva, “Survey
- an unequal clustering algorithm for wireless sensor networks,” in of wireless sensor networks simulation tools for demanding applica-
SENSORNETS, 2012. tions,” in Networking and Services, 2009. ICNS ’09. Fifth International
[96] Y.-C. Chen and C.-Y. Wen, “Distributed clustering with directional Conference on, 2009, pp. 102–106.
antennas for wireless sensor networks,” IEEE Sensors Journal, vol. 13, [117] A. Abuarqoub, F. Alfayez, M. Hammoudeh, T. Alsboui, and A. Nis-
no. 6, pp. 2166–2180, 2013. bet, “Simulation issues in wireless sensor networks: A survey,” in
[97] Y. Liao, H. Qi, and W. Li, “Load-balanced clustering algorithm SENSORCOMM 2012 , The Sixth International Conference on Sensor
with distributed self-organization for wireless sensor networks,” IEEE Technologies and Applications, 2012, pp. 222–228.
sensors journal, vol. 13, no. 5, pp. 1498–1506, 2013. [118] K. Romer and F. Mattern, “The design space of wireless sensor
[98] S.-S. Wang and Z.-P. Chen, “Lcm: a link-aware clustering mechanism networks,” Wireless Communications, IEEE, vol. 11, no. 6, pp. 54
for energy-efficient routing in wireless sensor networks,” IEEE Sensors – 61, dec. 2004.
Journal, vol. 13, no. 2, pp. 728–736, 2013. [119] J. Polastre, R. Szewczyk, A. Mainwaring, D. Culler, and J. Anderson,
[99] M. Yu, K. K. Leung, and A. Malvankar, “A dynamic clustering “Analysis of wireless sensor networks for habitat monitoring,” in
and energy efficient routing technique for sensor networks,” IEEE Wireless Sensor Networks, C. S. Raghavendra, K. M. Sivalingam, and
Transactions on wireless communications, vol. 6, no. 8, 2007. T. Znati, Eds. Springer US, 2004, pp. 399–423.
[100] D. Koutsonikolas, S. Das, H. Charlie, and I. Stojmenovic, “Hierar- [120] G. Booch, R. Maksimchuk, M. Engle, B. Young, J. Conallen, and
chical geographic multicast routing for wireless sensor networks,” K. Houston, Object-oriented analysis and design with applications,
in Sensor Technologies and Applications, 2007. SensorComm 2007. third edition, 3rd ed. Addison-Wesley Professional, 2007.
International Conference on, 2007, pp. 347–354. [121] L. Ruiz-Garcia, L. Lunadei, P. Barreiro, and I. Robla, “A review of
[101] F. Aurenhammer, “Voronoi diagrams: a survey of a fundamental wireless sensor technologies and applications in agriculture and food
geometric data structure,” ACM Comput. Surv., vol. 23, 1991. industry: State of the art and current trends,” Sensors, vol. 9, no. 6,
pp. 4728–4750, 2009.
[102] M. Meghji and D. Habibi, “Transmission power control in multihop
wireless sensor networks,” in Third International Conference on Ubiq- [122] K. H. Pollock, J. D. Nichols, T. R. Simons, G. L. Farnsworth, L. L.
uitous and Future Networks, 2011. Bailey, and J. R. Sauer, “Large scale wildlife monitoring studies:
statistical methods for design and analysis,” Environmetrics, vol. 13,
[103] T. S. Rappapor, Wireless Communications: Principles and Practice.
no. 2, pp. 105–119, 2002.
Englewood Cliffs: Prentice Hall, 1996.
[123] A. Zanella, N. Bui, A. Castellani, L. Vangelista, and M. Zorzi,
[104] N. Baccour, A. Koubâa, L. Mottola, M. A. Zúñiga, H. Youssef, C. A.
“Internet of things for smart cities,” IEEE Internet of Things Journal,
Boano, and M. Alves, “Radio link quality estimation in wireless sensor vol. 1, no. 1, pp. 22–32, Feb 2014.
networks: A survey,” Trans. Sen. Netw., 2012.
[124] D. Evans, “The internet of everything: How more relevant and valuable
[105] K. Srinivasan and P. Levis, “Rssi is under appreciated,” in EmNets,
connections will change the world,” Cisco IBSG, pp. 1–9, 2012.
2006.
[125] M. Chiang and T. Zhang, “Fog and iot: An overview of research
[106] S. A. Nikolidakis, D. Kandris, D. D. Vergados, and C. Douligeris,
opportunities,” IEEE Internet of Things Journal, vol. 3, no. 6, pp.
“Energy efficient routing in wireless sensor networks through balanced
854–864, Dec 2016.
clustering,” Algorithms, vol. 6, no. 1, pp. 29–42, 2013.

JOURNAL OF IEEE INTERNET OF THINGS, VOL. 6, NO. 1, JULY 2017 22

[126] D. Warren and C. Dewar, “Understanding 5g: Perspectives on future [143] M. Peng, Y. Li, J. Jiang, J. Li, and C. Wang, “Heterogeneous cloud
technological advancements in mobile,” GSMA, Dec 2014. radio access networks: a new perspective for enhancing spectral and
energy efficiencies,” IEEE Wireless Communications, vol. 21, no. 6,
[127] F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski,
pp. 126–135, December 2014.
“Five disruptive technology directions for 5g,” IEEE Communications
Magazine, vol. 52, no. 2, pp. 74–80, 2014. [144] M. Hassanalieragh, A. Page, T. Soyata, G. Sharma, M. Aktas,
G. Mateos, B. Kantarci, and S. Andreescu, “Health monitoring and
[128] R. Munoz, J. Mangues-Bafalluy, R. Vilalta, C. Verikoukis, J. Alonso-
management using internet-of-things (iot) sensing with cloud-based
Zarate, N. Bartzoudis, A. Georgiadis, M. Payaro, A. Perez-Neira,
processing: Opportunities and challenges,” in 2015 IEEE International
R. Casellas, R. Martinez, J. Nunez-Martinez, M. R. Esteso, D. Pubill,
Conference on Services Computing, June 2015, pp. 285–292.
O. Font-Bach, P. Henarejos, J. Serra, and F. Vazquez-Gallego, “The
cttc 5g end-to-end experimental platform: Integrating heterogeneous [145] M. N. Tehrani, M. Uysal, and H. Yanikomeroglu, “Device-to-device
wireless/optical networks, distributed cloud, and iot devices,” IEEE communication in 5g cellular networks: challenges, solutions, and
Vehicular Technology Magazine, vol. 11, no. 1, pp. 50–63, Mar 2016. future directions,” IEEE Communications Magazine, vol. 52, no. 5,
pp. 86–92, 2014.
[129] R. Fantacci, T. Pecorella, R. Viti, and C. Carlini, “A network ar-
chitecture solution for efficient iot wsn backhauling: challenges and [146] I. Loumiotis, E. Adamopoulou, K. Demestichas, and M. Theologou,
opportunities,” IEEE Wireless Communications, vol. 21, no. 4, pp. Optimal Backhaul Resource Management in Wireless-Optical Con-
113–119, Aug 2014. verged Networks. Cham: Springer International Publishing, 2015,
pp. 254–261.
[130] E. Piri and J. Pinola, “Performance of lte uplink for iot backhaul,”
in 2016 13th IEEE Annual Consumer Communications Networking [147] S. Andreev, M. Gerasimenko, O. Galinina, Y. Koucheryavy, N. Hi-
Conference (CCNC), Jan 2016, pp. 6–11. mayat, S. P. Yeh, and S. Talwar, “Intelligent access network selection
in converged multi-radio heterogeneous networks,” IEEE Wireless
[131] M. Hindia, T. Rahman, H. Ojukwu, E. Hanafi, and A. Fattouh,
Communications, vol. 21, no. 6, pp. 86–96, December 2014.
“Enabling remote health-caring utilizing iot concept over lte-femtocell
networks,” PloS one, vol. 11, no. 5, p. e0155077, 2016. [148] G. Yu, Y. Jiang, L. Xu, and G. Y. Li, “Multi-objective energy-
efficient resource allocation for multi-rat heterogeneous networks,”
[132] W. Kim, “Adaptive resource scheduling for dual connectivity in het-
IEEE Journal on Selected Areas in Communications, vol. 33, no. 10,
erogeneous iot cellular networks,” International Journal of Distributed
pp. 2118–2127, Oct 2015.
Sensor Networks, 2016.
[149] W. Xia, J. Zhang, S. Jin, C. K. Wen, F. Gao, and H. Zhu, “Bandwidth
[133] V. Jungnickel, D. Schulz, J. Hilt, C. Alexakis, M. Schlosser, L. Grobe,
allocation in heterogeneous networks with wireless backhaul,” in 2016
A. Paraskevopoulos, R. Freund, B. Siessegger, and G. Kleinpeter,
IEEE Global Communications Conference (GLOBECOM), Dec 2016,
“Optical wireless communication for backhaul and access,” in 2015
pp. 1–6.
European Conference on Optical Communication (ECOC), Sept 2015,
pp. 1–3. [150] C. Ran, S. Wang, and C. Wang, “Balancing backhaul load in heteroge-
neous cloud radio access networks,” IEEE Wireless Communications,
[134] M. Grob, “The road to 5g: Providing the connectivity fabric for
vol. 22, no. 3, pp. 42–48, June 2015.
everything,” in 2015 IEEE Hot Chips 27 Symposium (HCS), Aug 2015,
pp. 1–26. [151] R. Bonnefoi, C. Moy, and J. Palicot, “Advanced metering infrastructure
backhaul reliability improvement with cognitive radio,” in Smart
[135] M. R. Palattella, M. Dohler, A. Grieco, G. Rizzo, J. Torsner, T. Engel,
Grid Communications (SmartGridComm), 2016 IEEE International
and L. Ladid, “Internet of things in the 5g era: Enablers, architecture,
Conference on. IEEE, 2016, pp. 230–236.
and business models,” IEEE Journal on Selected Areas in Communi-
cations, vol. 34, no. 3, pp. 510–527, March 2016. [152] S. Bassoy, M. Jaber, M. A. Imran, and P. Xiao, “Load aware self-
organising user-centric dynamic comp clustering for 5g networks,”
[136] A. Cimmino, T. Pecorella, R. Fantacci, F. Granelli, T. F. Rahman,
IEEE Access, vol. 4, pp. 2895–2906, 2016.
C. Sacchi, C. Carlini, and P. Harsh, “The role of small cell tech-
nology in future smart city applications,” Transactions on Emerging [153] L. Xu, G. M. O’Hare, and R. W. Collier, “A smart and balanced
Telecommunications Technologies, vol. 25, no. 1, pp. 11–20, 2014. energy-efficient multihop clustering algorithm (smart-beem) for mimo
iot systems in future networks,” in MDPI Sensors, vol. 17, 2017.
[137] R. Fantacci, T. Pecorella, R. Viti, and C. Carlini, “Short paper:
Overcoming iot fragmentation through standard gateway architecture,” [154] J. Zhao, C. Qiao, R. S. Sudhaakar, and S. Yoon, “Improve efficiency
in 2014 IEEE World Forum on Internet of Things (WF-IoT), March and reliability in single-hop wsns with transmit-only nodes,” IEEE
2014, pp. 181–182. Transactions on Parallel and Distributed Systems, vol. 24, no. 3, pp.
520–534, 2013.
[138] A. Hakiri, P. Berthou, A. Gokhale, and S. Abdellatif,
“Publish/subscribe-enabled software defined networking for efficient [155] M. F. Munir and F. Filali, “Increasing connectivity in wireless sensor-
and scalable iot communications,” IEEE Communications Magazine, actuator networks using dynamic actuator cooperation,” in Vehicular
vol. 53, no. 9, pp. 48–54, September 2015. Technology Conference, 2008. VTC Spring 2008. IEEE. IEEE, 2008,
pp. 203–207.
[139] K. Wang, Y. Wang, D. Zeng, and S. Guo, “An sdn-based architecture
for next-generation wireless networks,” IEEE Wireless Communica- [156] L. Xu, J. Xie, X. Xu, and S. Wang, “Enterprise lte and wifi in-
tions, vol. 24, no. 1, pp. 25–31, February 2017. terworking system and a proposed network selection solution,” in
Proceedings of the 2016 Symposium on Architectures for Networking
[140] P. Huskov, T. Maksymyuk, I. Kahalo, and M. Klymash, “Smart back-
and Communications Systems, ser. ANCS ’16. New York, NY, USA:
hauling subsystem for 5g heterogeneous network,” in The Experience
ACM, 2016, pp. 137–138.
of Designing and Application of CAD Systems in Microelectronics,
Feb 2015, pp. 481–483. [157] R. Lent and E. C.-H. Ngai, “Lightweight clustering in wireless sensor-
actuator networks on obstructed environments,” in Wireless Pervasive
[141] A. Lea, K. Negus, H. Tapse, B. Varadarajan, and R. Vaughan,
Computing, 2008. ISWPC 2008. 3rd International Symposium on.
“Spectrum options for wireless backhaul of small cells,” in Antennas
IEEE, 2008, pp. 26–32.
and Propagation (EuCAP), 2014 8th European Conference on. IEEE,
2014, pp. 3310–3311. [158] O. Banimelhem, M. Mowafi, E. Taqieddin, F. Awad, and M. Al Rawab-
deh, “An efficient clustering approach using genetic algorithm and
[142] M. Peng, Y. Li, Z. Zhao, and C. Wang, “System architecture and key
node mobility in wireless sensor networks,” in Wireless Communica-
technologies for 5g heterogeneous cloud radio access networks,” IEEE
tions Systems (ISWCS), 2014 11th International Symposium on. IEEE,
network, vol. 29, no. 2, pp. 6–14, 2015.
2014, pp. 858–862.