Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Sustainable Computing: Informatics and Systems 22 (2019) 62–74

Contents lists available at ScienceDirect

Sustainable Computing: Informatics and Systems


journal homepage: www.elsevier.com/locate/suscom

Scalable and sustainable wireless sensor networks for agricultural


application of Internet of things using fuzzy c-means algorithm
Anagha Rajput, Vinoth Babu Kumaravelu ∗
School of Electronics Engineering, VIT University, Vellore, Tamil Nadu, India

a r t i c l e i n f o a b s t r a c t

Article history: The integration of sensing, communication and Internet is innovatively merging into a new technology
Received 23 July 2017 called Internet of things (IoT). Wireless sensor networks (WSNs) are the main physical monitoring infras-
Received in revised form 6 August 2018 tructure of IoT. Resource constrained sensor nodes have to be utilized in energy efficient manner so as to
Accepted 24 February 2019
maximize the monitoring network’s lifetime. Thus, for large scale monitoring applications of agriculture,
Available online 3 March 2019
forest and environment, it is required to have sustainable WSNs, where maximum number of sensors is
alive over a large period of time. In the new era of IoT, WSN is popularly preferred and used in precision
Keywords:
agriculture for farmland monitoring. In this proposed work, an attempt is made to design a cost effec-
Fuzzy c-means (FCM)
Half of the nodes dead (HND)
tive clustering algorithm to obtain energy efficient sustainable WSN while maximizing node density and
Internet of things (IoT) coverage area. The first objective of the proposed algorithm is to optimize energy efficiency by reducing
Perceived probability data transmission distance of sensor nodes using fuzzy c-means (FCM) clustering algorithm. The second
Scalability objective is to select a suitable cluster head node (CHN) based on perceived probability to attain network
Sustainable wireless sensor networks scalability. The results obtained shows that proposed algorithm is more energy efficient than other simi-
(WSNs) lar approaches. The comparative result statistics prove that proposed algorithm outperforms in terms of
half of the nodes dead (HND) and last node dead (LND) for scalable scenarios. Thus, it can be effectually
used in farm monitoring IoT systems.
© 2019 Elsevier Inc. All rights reserved.

1. Introduction such IoT application—precision agriculture, is focused. Precision


agriculture is been developed for resource optimization like water,
An IoT platform is expected to be one of the disruptive tech- fertilizers, and soil qualities. WSNs are particularly well suited to
nology in coming few years. In the new era of IoT, the daily lives precision agriculture for large open space monitoring since wired
of human beings will be fundamentally made comfortable with deployment would be expensive and inefficient. Here, agricultural
various IoT applications. IoT has been developing as large scale fields, grazing lands and monitored sites can reach several tens
and ubiquitous network, which will be widely used in agricul- of hectares, so the number of sensor nodes deployed varies from
ture, environmental surveillance, home and industrial automation, dozens to thousands. This is why scalability is an important issue
transportation systems, etc. [1]. In the emerging IoT systems, the when developing algorithms to support a high quantity of sensor
access points should have omnipresent connectivity to huge num- nodes and ensure full coverage of the monitoring area [5–7]. WSN
ber of wireless sensor nodes that are collecting data. Every physical is the network with resource constrained sensor nodes. They are
quantity will be measured and pushed on Internet cloud. Such data equipped by a low powered controller, sensors and short ranged
acquisition from any remote place will become as easy as handling radio transceiver. These on board circuits operate on batteries. The
cellular data on Internet [2–4]. Thus, the main monitoring infras- sensed data is collected at final destined entity called sink. Sink
tructure of IoT systems is WSN. The recent research concentrates on can be a base station, gateway or an access point that is linked to
the sustainability of WSN because of its limited resources. Unlike Internet. For a large scale WSN, it is impractical to replace batteries
other networks, WSNs are designed for specific applications and of sensor nodes after they are drained off completely. Thus, reduc-
therefore, WSN deployment has to satisfy a set of requirements that ing energy consumption of these nodes become crucial to enhance
differ from one application to another. In this research work, one network lifetime [8–11].
Many researches on energy consumption issues have been con-
ducted in WSN [12–29]. Sensor nodes utilize their energy in two
∗ Corresponding author. main operations. One is computation of data in its processor and
E-mail address: vinothbab@gmail.com (V.B. Kumaravelu). another is communicating its data to other network devices. The

https://doi.org/10.1016/j.suscom.2019.02.003
2210-5379/© 2019 Elsevier Inc. All rights reserved.
A. Rajput, V.B. Kumaravelu / Sustainable Computing: Informatics and Systems 22 (2019) 62–74 63

Fig. 1. Clustering concept in WSN.

remaining energy, etc. It uses this information to form clusters of


the sensor nodes and selects a CHN for every cluster. Thus, sink is
responsible for CHN selection and scheduling of the overall data
transmission from sensor nodes toward sink [12–14]. Chang and
Pei [15] proposed a distance based centralized clustering scheme
to maximize network lifetime. The main drawback of centralized
approach is that it is not scalable. The protocols hold good only for
small number of sensor nodes in the network. As the node density
increases, the bottleneck congestion at sink is unavoidable. This
degrades the performance of the protocol. For large scale IoT sys-
tems, node density is high and also varying in some applications.
Thus, centralized clustering techniques are unsuitable for scalable
WSN used for IoT applications.
In distributed clustering approaches, sensor nodes individually
make decisions based on local information. The formation of clus-
ters and the election of CHN affect the energy consumption of
the sensor network. The CHN selection can be performed in two
methodologies. The first one is to elect a CHN first and then form
cluster using query driven models. The other method is to form
cluster first and then select efficient node among those as CHN.
Heinzelman et al. [16] proposed Low Energy Adaptive Clustering
Hierarchy (LEACH) based on former method. LEACH introduced
concept of clustering into WSN. CHNs and clusters are formed
based on the random probability. Other nodes in the network elect
their CHN based on received signal strength of the advertisement
message. The network lifetime is improved significantly compared
to direct communication protocol. Handy et al. [17] introduced
energy consideration in the CHN selection approach. Here, the CHN
is selected based on energy ratio that improves network lifetime
with respect to conventional LEACH protocol. The critical issues
with these protocols are that the cluster is not formed uniformly.
Some clusters consist large number of sensor nodes while some has
very few sensor nodes. This is called as unbalanced network load,
Fig. 2. (a) Clustering of nodes using center point estimation techniques: Cluster- which causes sensor nodes to transmit data through longer dis-
ing using Voronoi diagram. (b) Clustering of nodes using center point estimation tances. Due to this, the energy saving is not achieved reasonably.
techniques: Clustering using FCM algorithm. Most of the random probabilistic approaches result in unbalanced
load distribution due to which sensor nodes die at early rounds
clustering technique is one of the widely used approaches for [18–23].
energy saving of sensor nodes. Fig. 1 shows cluster based routing Chang [24] proposed a Distributed Cluster Computing Energy
of data toward the sink. Every cluster has a nominated single node Efficient Routing Scheme (DCCEERS) to balance the network load.
called as CHN. CHN is responsible for collecting data from the nodes The cluster formation process starts with a random counter time-
in its cluster. Then, it aggregates all collected information into single out. The clusters are formed using query messages and computing
data packet and transmit to the sink. center of gravity among the nodes. Center of gravity is the virtual
Clustering means grouping the sensor nodes which are near center point of the cluster. This center of gravity and energy level
most to each other. A sensor node is elected from among the group of node is used to compute selection factor of each node. The node
members and declared as CHN of that cluster. The study of cluster- with maximum selection factor is then selected as CHN. The pro-
ing in WSN leads to the classification of clustering as centralized tocol do not have control over number of CHNs elected per round
approaches and distributed approaches. In centralized approaches, as many nodes’ counter may get zero at the same time. Few of the
sink collects information of all the sensor nodes such as location, nodes might remain isolated as they are out of transmission range
64 A. Rajput, V.B. Kumaravelu / Sustainable Computing: Informatics and Systems 22 (2019) 62–74

of CHNs in the network. Though energy saving is achieved, issue of


isolated nodes is not considered.
Jia et al. [25] proposed Dynamic Cluster Head Selection Method
(DCHSM) for static cluster formation using Voronoi diagram. This
follows the latter methodology of CHN selection. The monitoring
region is divided into optimal number of Voronoi cells. The CHN is
then selected in two stages. First stage of CHNs is selected based
on perceptual probability. The perceptual probability depends on
distance of nodes from the center point of the cluster and the sens-
ing range of the sensor. Second stage of CHNs is elected based on
survival time estimation. Second stage of CHNs are selected when
first stage CHNs completely drain their energy. This results in few
nodes failing at early rounds. Dividing geographical region exactly
into Voronoi cells is not a practical approach for implementation.
There is another method of clustering used for optimization.
These are fuzzy based and categorized as hierarchical, partition
based and grid based clustering algorithms [26–28]. These are
hard clustering methods as member functions of these methods
take on value of either 0 or 1. This classification does not work Fig. 4. The perceived probability model of the proposed algorithm.
in real situations. To overcome this, fuzzy clustering methods
have emerged, which incorporates fuzzy set theory. Fuzzy clus-
(b) The CHN of a cluster is the sensor node that spends more energy
tering technique determines affinities of different objects with
than other nodes. Thus, CHNs die early compared to other nodes
mathematical methodologies described by the member functions
in the network. The probabilistic model used in the literatures
[29–33].
are not efficient because the conditions used to select a CHN are
Motivated from the above literature, it is studied that forma-
either sensor node’s energy or distance from cluster members.
tion of clusters around virtual center points will help to reduce
Due to such selection procedure, nodes are frequently utilized
transmission distance of sensor nodes as well as maximize energy
for CHN operation until they die out. This results falling rate
savings. This is particularly beneficial for wide monitoring regions.
of alive nodes in the network. For designing sustainable WSN,
But computation of virtual center points also face many limitations.
there is still scope to develop models to select appropriate CHNs
This work addresses following issues:
to enhance network lifetime.

As mentioned in literature, there are several fuzzy clustering


(a) Center point computation for cluster formation is one of the algorithms. To resolve above stated issues, FCM algorithm is imple-
popular approaches used by researchers [15,24–28]. Center mented. FCM algorithm is widely used because it is robust for any
point is a virtual geographical point location. Few related work ambiguity [34]. This algorithm is an extension of crisp k means clus-
assumes a node situated at center points [25]. Nodes near most tering method used in fuzzy set domain. The applications where
to the center point together form the cluster. Computing appro- FCM is popularly used are image clustering [35], pattern recogni-
priate location for center point of a cluster is essential as it tion [36], etc. The use of FCM in clustering of sensor nodes has also
affects the transmission distances of the nodes associated with proved beneficial [37,38].
it. Most related work focus on center point estimation based on In our proposed algorithm, firstly static clusters are formed in
geometry of the geographical area. The node locations are not the network. From each cluster, a CHN is then elected to forward
considered while calculating center points in the given region. the cluster data toward the sink. The one time data transmission
This does not optimize reduction in transmission distance of from sensor nodes toward sink via CHNs is termed as one round of
nodes placed at boundaries of geometrical partition. An illus- data collection. This is implemented in three phases—initial setup
tration of the case where nodes are crowded at boundary of the phase, setup phase and data transmission phase. In initial setup
partition is given in Fig. 2. The nodes marked by circle in Fig. 2(a) phase, FCM algorithm is used to cluster the sensor nodes and esti-
are the crowded sensor nodes at boundaries of two polygo- mate center point of each cluster. Thus, at the end of initial setup
nal clusters formed using Voronoi diagram. They are at large phase, a clustered network is obtained. This phase is executed only
distance from their respective static center point of the cell. once at the start of the algorithm and thus is named as initial setup
Instead, those boundary nodes from two polygons together can phase. One time computation cost to form static clusters is afford-
form a new cluster with suitable virtual center point as shown able than communication cost required to form dynamic clusters
in Fig. 2(b). Comparing the average distances of nodes toward in each round because energy utilized for transmission is directly
their respective center points in both cases, former has large proportional to the distance. As illustrated in Fig. 2(b), the clusters
average distance than in the latter case. Thus, it is sensible to are formed by the sensor nodes that are closely located to each
determine center points according to the locations of the sensor other. The advantage of cluster formation based on FCM algorithm
nodes using FCM algorithm to obtain better clustering. is reduction in the intra-cluster communication distances.

Fig. 3. Radio communication model.


A. Rajput, V.B. Kumaravelu / Sustainable Computing: Informatics and Systems 22 (2019) 62–74 65

Table 1 posed algorithm in detail. Section 4 presents simulation parameters


Computational complexity calculation of the proposed algorithm.
and comparative results. The conclusions and future scopes are
Operations Clustering FCM Perceived discussed in Section 5.
algorithm probability
model for CHN
selection
2. Energy model

Number of (N − 1)(T − 1) + 4T − 1 + 2(N − 1) + 1 3N + 2(C − 1)


The proposed algorithm is designed for the farm field monitor-
additions
Number of NT(m + 1) + 2T + 1 + N(2m − 1) + 1 4 ing IoT system based on WSN. To make the network sustainable
multiplications for long period of time, efficient utilization of sensor nodes’ bat-
Number of 1 4 tery energy is the prime priority. The sensor nodes are assumed
comparisons
to be deployed randomly and outdoor spreading over a hectares of
farming land. In WSN, data transmission consumes large amount of
The setup and data transmission phases are executed con- energy than data processing. Sensor nodes transmit their data wire-
secutively. In our work, each round consists of setup phase and lessly upon short range of 10–100 m. Thus, popularly used radio
data transmission phase. In setup phase, perceived probability communication models in WSN are free space propagation model
model is implemented to elect an appropriate CHN in the network. or multipath fading channel model [24]. A radio communication
The model calculates three parameters—perceived probability (P), model used in the proposed work is shown in Fig. 3.
perceived value (PV) and probable energy (PE). The perceived prob- The energy utilization of the sensor node is calculated for the
ability indicates the level of centrality of the node in the cluster. In following three operations:
this work, centrality of the node is categorized in terms of maxi-
mum, moderate or minimum reachability of the node. A node with (a) Data transmission from cluster members to a CHN
high perceived probability is more centrally placed in the cluster. The energy spent by a sensor node to transmit its data to its
Thus, it has maximum reachability over the cluster members. Sim- CHN is calculated as,
ilarly, a node with low perceived probability is reasonably away
EtxCHN = Z(Eelec + Eampl ) (1)
from the cluster center. Thus, such nodes have minimum reacha-
bility over the cluster members. A node with moderate reachability where EtxCHN is the total energy required to transmit Z bits of
is neither too central nor far away from the cluster center. Unlike data. Eelec is the energy required by electronic hardware circuit
other probabilistic models in literature, the perceived probability of the sensor device. Eampl is the energy required by ampli-
depends on sensor node’s distance from the center point, sens- fier at the transmitter front-end circuit. This value is calculated
ing range and hardware uncertainty factor as defined in (14). The depending on the distance between transmitter and receiver. It
sensing range is important factor as it decides the category of is given as,
the centrality of the node. For real time agricultural application 
consideration, sensing range is extended to tolerable range using εfs D2 , D ≤ D0
Eampl = (2)
uncertainty factor of sensor hardware unit. εmp D4 , D > D0
The perceived probability is further used to calculate PV and
PE of the nodes. It is explained in detail in Section 3.2. The sensor where D is the transmitter and receiver separation distance.
nodes are homogeneous in nature as these are used to monitor εfs and εmp are the energy parameters for free space propaga-
agricultural entities. It is also assumed that the sensor nodes have tion and multipath fading channel models respectively. D0 is
same initial energy during deployment. This may raise stalemate the minimum reference distance and is calculated as,
situations among the sensor nodes, which are competing for CHN εfs
selection. Thus, to resolve the tie situations and minimize random D0 = (3)
εmp
CHN selection among the nodes with same energy or perceived
probability values, PV and PE are suitably used to elect a CHN. The (b) Data reception at receiver of CHN
data transmission phase is similar as in [16]. The results in Section 4 The data transmitted by all the cluster members is received at
shows that the proposed algorithm proves better clustering, which the CHN receiver circuit. This operation utilizes some amount
ultimately improves network lifetime. of energy which is calculated as,
The rest of the sections illustrate the proposed work as follows:
Section 2 presents the energy model. Section 3 illustrates the pro- Erx = ZEelec (4)

where Erx is the total energy required to receive Z bits of data


Table 2 from a sensor node.
Parameters used for simulation. (c) Data transmission from a CHN to the sink
Parameters Values After receiving data from all its cluster members, CHN aggre-
gates all information into single data packet and transmit it to
Monitoring region size 100 m × 100 m
Number of sensor nodes 200, 300, 400, 500 sink. The amount of energy required by a CHN to aggregate and
Sink location (50,150), (0,0) send data is calculated as,
Initial energy of sensor node 0.5 J
Electronic circuit’s energy 50 nJ EtxSink = Z(Eelec + Eaggr + Eampl ) (5)
Data aggregation energy 5 nJ/bit/message
Free space communication energy 10 pJ/bit/m2 where EtxSink is the energy required by a CHN to transmit Z bits
Multipath communication energy 0.0013 pJ/bit/m4 of data to the sink. Eaggr is the data aggregation energy utilized
Control packet size 200 bits by a CHN.
Data packet size 2000 bits
Fuzzy factor (m) 2
Consider a WSN deployed in the monitoring region. The total
Uncertainty factor of sensor node (˛) 0.2 number of sensor nodes in the network is denoted by N. The
Exponent factor () 0.1 total number of clusters formed in the network is denoted by
Improvement threshold of objective function 1 × 10−5 T. A cluster consists of cluster members and one CHN. The total
Termination threshold for FCM (q) 100
number of sensor nodes in a cluster is denoted by C. For every
66 A. Rajput, V.B. Kumaravelu / Sustainable Computing: Informatics and Systems 22 (2019) 62–74

processors of pre-deployment sensor nodes. The sensor nodes are


then randomly deployed in the monitoring region. It is assumed
that a sensor node can estimate its location after deployment. A
sink is placed outside the monitoring area to collect the sensed
data from the nodes. After network deployment, sink broadcasts
‘HELLO’ message for the activation of nodes in the network. The
WSN is made functional by the proposed algorithm in three phases.
These are initial setup phase, setup phase and transmission phase.
The initial setup phase is executed once at the start of the algo-
rithm. In this phase, every deployed node runs FCM algorithm after
power activation. It is assumed that each node knows the location
information of all the other nodes in the network. The inputs given
to FCM algorithm are number of clusters to be formed and node
locations. The FCM algorithm partitions sensor nodes into given
number of clusters in such a way that the intra-cluster distances
are optimally reduced. The output of FCM algorithm is the clus-
tered network. Each cluster is identified by its cluster center point.
A node assigns itself the cluster ID based on the center point to
which it is associated. This cluster ID is used in control packet data
for cluster member identification. The distributed static clusters are
thus formed in the initial setup phase as illustrated in following Sec-
tion 3.1. It is affordable to spend one time computation energy cost
of sensor nodes than spending communication energy cost using
query system for cluster formation in every round.
After cluster formation, setup phase and transmission phase is
executed consecutively for every round. In setup phase, a CHN is
elected for every cluster based on P and PV or P and PE parameters.
Each node calculate P as defined in (14) and categorizes itself as
maximum or moderate or minimum reachable node with respect
to associated center point of its cluster. A node located nearest to
cluster center point is also more centrally located with respect to
its cluster. Such a node having high centrality is maximally reach-
able to all the cluster members. Thus, the perceived probability
of such nodes is high. Similarly, nodes located away from cluster
Fig. 5. (a) Clustering of 200 nodes in a region of size 100 m × 100 m: Clustering center points are assigned lower probability values. This value is
using FCM algorithm. (b) Clustering of 200 nodes in a region of size 100 m × 100 m: further used to calculate PV and PE of the node. These parameters
Clustering using Voronoi diagram. are suitably involved in CHN election to break the tie cases occur-
ring due to same energy levels or perceived probabilities in the
round, the total energy consumption of the cluster is calculated initial round. The sensor nodes are assumed to be homogeneous
as, and have same initial energy after deployment. This condition may
raise critical situations of tie between CHN redundant nodes having
Ecluster = EtxSink + (C − 1)EtxCHN + (C − 1)Erx (6) same energy levels or perceived probabilities. The CHN for initial
round is elected based on PV while for remaining rounds, CHN is
where Ecluster is the total energy consumed to transmit sensed
elected based on PE. It is illustrated in following Section 3.2. The
data from sensor nodes in a cluster toward sink via CHN. In a
PV or PE values are communicated to other redundant nodes of the
cluster with C number of sensor nodes, (C − 1) nodes are clus-
cluster. A node with highest PV or highest PE value declares itself as
ter members and one is CHN. Thus, energy is utilized for (C − 1)
CHN for the current round. It then prepares a transmission sched-
transmissions from cluster members to CHN and (C − 1) recep-
ule for cluster members and broadcast. The transmission phase is
tions by the CHN. Energy is also utilized for one transmission
similar to that of LEACH protocol [16].
from CHN to sink. The total energy utilization of the WSN for
Following are the assumptions considered for algorithm.
one round is given as,


T • Sensor nodes are randomly deployed in the monitoring area.
Eround = Eclusterj (7) • The network is homogeneous and all sensor nodes have same
j=1 initial energy.
• For farmland monitoring, sensor nodes and sink deployed are
where Eround is the total energy consumption of all the clus- static in nature.
ters in the network to transmit sensed data toward sink for one • All the nodes are battery limited electronic devices while sink is
round. an access point server situated at remote location.

3. Proposed algorithm 3.1. Cluster formation using FCM

In this section, the detailed construction of the proposed algo- FCM is a clustering algorithm developed by Dunn [30]. Later
rithm is elaborated. A WSN is to be used to monitor farmlands for it was redeveloped by Bezdek. It is used when number of clus-
precision agriculture. Hence, network must be scalable as well as ters is pre-determined. Agricultural land is systematically divided
sustainable for uninterrupted monitoring operation. The constant for different field operations and growing food grains. Thus, FCM
parameters used in the algorithm are updated in the memory of the algorithm is well suitable to divide sensor nodes spread over such
A. Rajput, V.B. Kumaravelu / Sustainable Computing: Informatics and Systems 22 (2019) 62–74 67

Table 3
Network lifetime in terms of FND, HND and LND (nodes = 200).

WSN LEACH [16] DCCEERS [24] DCHSM [25] Proposed

FND HND LND FND HND LND FND HND LND FND HND LND

1 254 1203 2122 380 1250 3279 56 1653 2702 90 2690 2799
2 155 1235 1775 379 1241 2802 62 1662 2563 78 2669 2795
3 209 1276 2002 382 1223 2495 55 1665 2556 108 2672 2799
4 348 1260 2382 406 1305 2767 75 1695 2656 102 2681 2799
5 190 1313 1750 232 1274 2676 63 1708 2636 108 2655 2799
6 226 1316 1832 348 1229 3040 70 1655 2548 85 2630 2799
7 245 1256 1871 350 1299 2664 70 1697 2545 91 2693 2799
8 237 1271 2058 382 1290 2689 71 1699 2673 97 2692 2798
9 223 1277 1799 266 1265 3307 72 1614 2604 93 2674 2796
10 208 1220 1864 298 1225 2764 61 1716 2677 103 2685 2798
Average 229.5 1262.7 1945.5 342.3 1260.1 2848.3 65.5 1676.4 2616 95.5 2674.1 2798.1

Table 4
Network lifetime in terms of FND, HND and LND (nodes = 300).

WSN LEACH [16] DCCEERS [24] DCHSM [25] Proposed

FND HND LND FND HND LND FND HND LND FND HND LND

1 186 1295 1853 297 1257 2556 47 1705 2290 72 2689 2799
2 223 1235 2234 330 1263 2827 43 1755 2650 69 2684 2799
3 250 1234 2136 253 1298 2509 48 1762 2773 67 2669 2799
4 183 1237 1804 181 1268 2943 43 1840 2662 70 2659 2799
5 188 1325 1868 134 1300 2583 55 1726 2627 76 2636 2798
6 229 1323 1956 208 1282 2935 49 1718 2621 78 2689 2799
7 149 1213 1780 249 1224 2546 51 1698 2621 66 2641 2799
8 222 1266 1905 149 1224 2831 45 1782 2538 70 2658 2799
9 152 1264 1779 271 1304 2823 52 1716 2701 71 2685 2799
10 258 1258 2092 224 1274 2463 53 1727 2660 74 2693 2799
Average 204 1265 1940.7 229.6 1269.4 2701.6 48.6 1742.9 2614.3 71.3 2670.3 2798.9

Table 5
Network lifetime in terms of FND, HND and LND (nodes = 400).

WSN LEACH [16] DCCEERS [24] DCHSM [25] Proposed

FND HND LND FND HND LND FND HND LND FND HND LND

1 227 1301 2029 221 1654 2758 36 1760 2716 65 2694 2798
2 165 1279 1819 204 1256 2795 35 1786 2774 60 2672 2798
3 287 1240 2338 181 1264 2831 40 1732 2677 62 2700 2798
4 216 1259 1878 241 1261 3192 37 1796 2716 53 2694 2799
5 245 1284 1894 185 1285 2274 42 1734 2565 60 2693 2798
6 219 1300 1837 158 1263 2270 40 1734 2772 61 2687 2799
7 248 1262 1836 201 1251 2782 35 1737 2592 55 2679 2798
8 154 1254 1843 141 1240 2184 34 1746 2695 58 2681 2798
9 266 1253 1960 185 1216 2079 31 1721 2732 52 2671 2798
10 204 1282 1920 235 1250 2918 35 1793 2781 52 2699 2798
Average 223.1 1271.4 1935.4 195.2 1294 2608.3 36.5 1735.9 2702 57.8 2687 2798.2

Table 6
Network lifetime in terms of FND, HND and LND (nodes = 500).

WSN LEACH [16] DCCEERS [24] DCHSM [25] Proposed

FND HND LND FND HND LND FND HND LND FND HND LND

1 145 1200 1847 133 1211 2662 24 1796 2768 48 2691 2799
2 185 1277 1850 287 1274 2655 30 1750 2796 47 2696 2799
3 249 1294 1922 166 1282 3003 33 1791 2779 46 2661 2799
4 141 1284 1769 189 1273 2536 30 1737 2670 48 2661 2799
5 243 1285 1993 190 1295 2635 30 1818 2731 49 2668 2799
6 225 1313 2177 204 1270 2866 34 1766 2711 52 2684 2799
7 252 1309 1900 185 1290 3094 28 1877 2723 46 2658 2799
8 282 1279 2110 211 1214 2736 31 1742 2782 42 2654 2799
9 241 1283 1893 107 1239 2848 33 1787 2709 46 2653 2799
10 236 1288 1926 192 1258 2560 29 1841 2787 52 2680 2799
Average 219.9 1281.2 1938.7 186.4 1260.6 2759.5 30.2 1790.5 2745.6 47.6 2670.6 2799

farmlands into known number of clusters. This algorithm associates Each node location has two-dimensional coordinates. Number of
each node with one of the cluster center point. Here, FCM is exe- clusters is a mere integer. The FCM algorithm initially sets T ran-
cuted to classify N number of sensor nodes into T number of clusters. dom cluster center points. The node locations are then considered
Inputs to FCM algorithm are node locations and number of clusters. iteratively to estimate its membership value toward each cluster
68 A. Rajput, V.B. Kumaravelu / Sustainable Computing: Informatics and Systems 22 (2019) 62–74

Table 7
Network lifetime in terms of FND, HND and LND (region size = 100 m × 100 m).

WSN LEACH [16] DCCEERS [24] DCHSM [25] Proposed

FND HND LND FND HND LND FND HND LND FND HND LND

1 326 1420 2731 310 1519 2755 73 1782 2404 104 2681 2796
2 215 1465 1938 329 1583 2882 84 1779 2261 101 2687 2798
3 226 1481 1931 319 1571 2765 72 1760 2427 124 2677 2792
4 231 1453 2005 437 1539 2854 80 1765 2334 117 2685 2798
5 265 1442 2117 360 1511 2982 80 1795 2190 120 2680 2799
Average 252.6 1452.2 2144.4 351 1544.6 2847.6 77.8 1776.2 2323.2 113.2 2682 2796.6

Table 8
Network lifetime in terms of FND, HND and LND (region size = 100 m × 150 m).

WSN LEACH [16] DCCEERS [24] DCHSM [25] Proposed

FND HND LND FND HND LND FND HND LND FND HND LND

1 230 1235 1797 380 1291 2812 72 1482 2256 93 2603 2795
2 277 1303 1871 358 1307 2521 72 1480 1977 78 2589 2795
3 115 1318 1727 202 1220 2501 64 1431 2079 86 2606 2799
4 153 1301 1746 364 1208 2473 68 1442 2018 103 2636 2792
5 210 1325 1777 394 1146 2735 77 1462 2230 109 2645 2798
Average 197 1296.4 1783.6 339.6 1234.4 2608.4 70.6 1459.4 2112 93.8 2615.8 2795.8

Table 9
Network lifetime in terms of FND, HND and LND (region size = 200 m × 200 m).

WSN LEACH [16] DCCEERS [24] DCHSM [25] Proposed

FND HND LND FND HND LND FND HND LND FND HND LND

1 98 584 1520 39 530 2925 27 877 1944 41 2320 2796


2 67 528 1385 50 488 2538 25 873 1910 42 2286 2796
3 57 590 1318 64 532 2730 25 914 1880 38 2408 2796
4 116 579 1902 99 561 2536 28 808 1984 35 2343 2791
5 85 751 1877 58 670 2528 26 904 1991 37 2378 2796
Average 84.6 606.4 1600.4 62 556.2 2651.4 26.2 875.2 1941.8 38.6 2347 2795

Table 10
Network lifetime in terms of FND, HND and LND (region size = 200 m × 250 m).

WSN LEACH [16] DCCEERS [24] DCHSM [25] Proposed

FND HND LND FND HND LND FND HND LND FND HND LND

1 45 460 1314 51 443 2357 17 692 1957 15 2223 2792


2 32 443 1229 37 416 2265 13 708 2021 20 2306 2793
3 32 443 1229 37 416 2265 13 708 2021 20 2306 2793
4 56 520 1798 44 494 2542 17 671 2076 21 2279 2783
5 71 467 1415 47 414 2513 16 709 2037 22 2427 2789
Average 47.2 466.6 1397 43.2 436.6 2388.4 15.2 697.6 2022.4 19.6 2308.2 2790

center point. Unlike conventional clustering algorithms, FCM com- will be changing for every iteration in order to minimize objective
putes membership values from 0 to 1. A value of 0 indicates no function. Distances of every node from T cluster center points are
membership and 1 indicates complete membership. In between calculated using matrices X and Y. These distances are then used to
values indicate proportionate membership. The set of center points formulate objective function as,
are calculated for every iteration so as to minimize the objective
function of the algorithm. The FCM algorithm is mathematically
illustrated as below:

N

T
The first input data to the FCM algorithm is node locations. It is OF = (Uij )m d(xi , yj )2 (10)
a matrix of dimension N × 2 and represented as,
i=1 j=1

X = {x1 , x2 , . . ., xi , . . ., xN } (8)

where N is the total number of sensor nodes in the network. X is


where OF is the objective function for a current iteration. Uij is the
a matrix in which all the node locations are stored. The variable xi
degree of membership that the ith sensor node pertains with jth
indicates two-dimensional coordinates of ith node.
center point. The degree of membership is expressed in the order
The algorithm starts by initializing T random center points. It is
of m, which is a fuzzy factor ranging as m ∈ [1, ∞] [31]. Many stud-
represented by matrix of dimension T × 2 and represented as,
ies show that m value is considered to be [2,2.5]. The fuzzy factor
Y = {y1 , y2 , . . ., yj , . . ., yT } (9) considered in our algorithm is 2, which is optimal value for short
distance computations as studied in [31]. The term d(xi , yj ) is the
The matrix Y stores iterative T cluster center points. The value yj distance between ith sensor node and jth center point. The mem-
indicates present co-ordinates of the jth center point. These values bership matrix element Uij and the iterative center point yj are
A. Rajput, V.B. Kumaravelu / Sustainable Computing: Informatics and Systems 22 (2019) 62–74 69

formulated as,
1
Uij = , i = 1, 2, . . ., N and j = 1, 2, . . ., T
T  ||xi −yj || 2/m−1
o=1 ||xi −yo ||
(11)

N
i=1
Uijm xi
yj = N , j = 1, 2, . . ., T (12)
i=1
Uijm

To converge the algorithm output at the most optimal value,


the partial derivative of objective function OF with respect to Uij
and yj is executed for every iteration using (11) and (12). yo is the
cluster center point of the jth cluster estimated in the past itera-
tion. The denominator of (11) is computed using the set of cluster
center points estimated in the previous iteration. Following are the
conditions need to be satisfied by algorithm in every iteration.
 N
i=1
Uij = 1, j = 1, 2, . . ., T
(13)
0 ≤ Uij ≤ 1, i = 1, 2, . . ., T and j = 1, 2, . . ., T

The overall functionality of FCM algorithm is summarized as


follows:

Algorithm 1. Proposed clustering FCM algorithm


Initialization:
1. Initialize cluster center points using (9);
2. Initialize min improvement threshold;
3. Initialize maximum iterations (qmax );
Input:
4. Node locations given in (8);
Output:
5. Clustered WSN
Main function:
6. For each iteration (q)
7. If q < qmax
8. Calculate OF using (10);
9. Improvement = absolute (OFq )—absolute (OFq−1 );
10. If improvement > min improvement threshold Fig. 6. (a) Simulation results for 200 nodes: Number of alive nodes vs. Rounds. (b)
11. Update Uij and yj using (11) and (12); Simulation results for 200 nodes: Remaining network energy vs. Rounds.
12. Else
13. Break;
14. End di is important because it determines probable value of node pro-
15. Else portionate to sensing range.
16. Break;
The estimation of perceived probability categorizes the sensor
17. End
18. End nodes into three zones as shown in Fig. 4. The nodes that are located
at a distance of (Ds − ˛) or less than (Ds − ˛) are the nearest to
3.2. CHN selection using perceived probability model the cluster center point. Thus, such nodes are said to be in max-
imum reachability zone and their P = 1. The nodes that are located
At the end of initial setup phase, clusters are created and each between the distances (Ds − ˛) and (Ds + ˛) constitute moderate
sensor node is associated to a single cluster. For every round, setup reachability zone and the nodes located beyond (Ds + ˛) are con-
phase and data transmission phase is executed consecutively. In sidered to be in minimum reachability zone.
setup phase, CHN is elected for each cluster based on perceived At the initial round of data transmission, the tie cases with
probability model. All the sensor nodes calculate their perceived respect to energy as well as perceived probability occur among
probability as given in (14). the nodes. All the sensor nodes are assumed to have same initial
⎧ energy. Thus, energy parameter cannot be considered for selec-
⎪ 1, d(xi , yj ) ≤ (Ds − ˛)
⎨ tion of first round CHNs. Also, a particular node cannot be selected
P(xi , yj ) = e−di , (Ds − ˛) < d(xi , yj ) ≤ (Ds + ˛) (14) as CHN based on (14) because nodes in the maximum reachabil-

⎩ ity zones have same value (P = 1). To resolve this stalemate and
0, d(xi , yj ) > (Ds − ˛) minimize random selection of CHN, PV and PE are proposed.
where P(xi , yj ) is the perceived probability of ith node with respect
to its cluster center point yj . d(xi , yj ) is the distance of the ith sensor 3.2.1. CHN selection for initial round
node from its cluster center. Ds is the sensing range of the sensor The energy levels of the sensor nodes in the initial round are
node. ˛ is the uncertainty factor of the sensor hardware unit.  is same and therefore CHN is elected based on only distance parame-
the exponent factor. di is the term used as the exponential variable. ter. A set of redundant nodes from each cluster are selected whose
It is calculated depending on distance d(xi , yj ). perceived probability is greater than 0.3 and less than 1. All the
redundant nodes lie in moderate reachability zone as per the above
di = d(xi , yj ) − (Ds − ˛) (15) mentioned condition and (14). All the nodes with perceived prob-
70 A. Rajput, V.B. Kumaravelu / Sustainable Computing: Informatics and Systems 22 (2019) 62–74

Fig. 7. (a) Simulation results for 300 nodes: Number of alive nodes vs. Rounds. (b) Fig. 8. (a) Simulation results for 400 nodes: Number of alive nodes vs. Rounds. (b)
Simulation results for 300 nodes: Remaining network energy vs. Rounds. Simulation results for 400 nodes: Remaining network energy vs. Rounds.

3.2.2. CHN selection for remaining rounds


ability equal to one will result in same PV which will cause a tie
For remaining rounds, CHNs are elected based on both energy
condition for CHN selection. The nodes with probability greater
and distance parameter. For each round, every node calculate its PE
than 0.3 is considered to allow more nodes to participate in CHN
value as,
selection. The experimentation results studied, indicates that num-
ber of nodes in moderate zone are comparatively less than in
maximum zone [25]. All these redundant nodes calculate their PV Eremaini
as, PE(xi , yj ) = P(xi , yj ) (17)
Eaveragej

P(xi , yj )
PV (xi , yj ) = C (16) where PE(xi , yj ) is the probable energy of ith node with respect to its
k=1
P(xi , yj ) cluster center point yj . Eremaini is the remaining energy of ith node in
jth cluster. Eaveragej is the total average energy of all the nodes in the
jth cluster. P(xi , yj ) is the perceived probability of the ith node. The
where PV(xi , yj ) is the perceived value of ith node with respect to PE in (17) is formulated in such a way that, a node which is more
its cluster center point yj . It is the ratio of individual probability to centrally located in the cluster and has high energy will result in
the total probability of the nodes in the corresponding cluster. P(xi , high PE value. Thus, a node with highest PE in the cluster is elected
yj ) is the perceived probability of the ith node. C is the total number CHN for the current round.
of nodes in the cluster. A node with maximum PV value is elected After completion of setup phase, each cluster has an elected
as CHN of the initial round. CHN. All the CHNs schedule time slots for data collection from
The nodes from the maximum reachability zone are not consid- the cluster members using Time Division Multiple Access (TDMA).
ered for initial CHN selection because the perceived probability of Cluster members transmit their sensed data to CHN in their allo-
the nodes in this zone is same and equal to one. The denominator cated time slots. When all the data is received, CHN performs data
in (16) is a constant for a given cluster. Thus, all the nodes will have aggregation process to form a single data packet. This aggregated
same PV, resulting in stalemate situation. data is then transmitted to sink.
A. Rajput, V.B. Kumaravelu / Sustainable Computing: Informatics and Systems 22 (2019) 62–74 71

Fig. 11. (a) Simulation results for 100 m × 100 m region size: Number of alive nodes
Fig. 9. (a) Simulation results for 500 nodes: Number of alive nodes vs. Rounds. (b) vs. Rounds. (b) Simulation results for 100 m × 100 m region size: Remaining network
Simulation results for 500 nodes: Remaining network energy vs. Rounds. energy vs. Rounds.

multiplications. The objective function also computes membership


matrix Uij and iterative new cluster center point yj . The matrix Uij
computation requires 4T − 1 additions and 2T + 1 multiplications.
The cluster center point yj computation requires 2(N − 1) additions
and N (2m − 1) + 1 multiplications (that is, Nm + N (m − 1) + 1). Step
9 of the Algorithm 1 requires 1 addition and step 10 requires 1
comparison. The overall computation required by FCM algorithm to
form clusters is tabulated in Table 1. Similar computation is applied
on perceived probability model. The perceived probability given
by (14) requires 2N additions and 4 comparisons. The exponential
variable, di requires N additions. The computation of PV given by
(16) requires C − 1 additions and 1 multiplication while PE given by
Fig. 10. Effect of node density on HND for simulated algorithms.
(17) requires C − 1 additions for calculating average energy value
of the cluster and 3 multiplications. The total computation of the
3.3. Computational complexity of the proposed algorithm probability model is also detailed in Table 1.

The proposed algorithm is also analyzed for computational


complexity as WSN is a resource constrained network and low pro- 4. Performance analysis
cessing operating system of sensor node hardware module should
be able to handle the complexity of the designed algorithm. The For fair evaluation, our proposed algorithm is compared with
mentioned Algorithm 1 has maximum qmax iterations. Every time LEACH [16], DCCEERS [24] and DCHSM [25] in terms of total remain-
the FCM may not execute for all qmax runs. Thus, the computa- ing network energy, number of alive nodes, first node dead (FND),
tion calculated for one iteration of FCM algorithm is showcased in HND and LND. LND specifies network lifetime and HND is used to
Table 1. Step 8 of the Algorithm 1 calculates objective function OF, calculate network life stability. The algorithm is simulated for mon-
where total number of sensor nodes are N and total number of clus- itoring application of IoT systems, where node density is supposed
ters formed are T. It requires (N − 1) (T − 1) additions and NT (m + 1) to be as large as possible. LEACH, DCCEERS, DCHSM and proposed
72 A. Rajput, V.B. Kumaravelu / Sustainable Computing: Informatics and Systems 22 (2019) 62–74

Fig. 13. (a) Simulation results for 200 m × 200 m region size: Number of alive nodes
Fig. 12. (a) Simulation results for 100 m × 150 m region size: Number of alive nodes
vs. Rounds. (b) Simulation results for 200 m × 200 m region size: Remaining network
vs. Rounds. (b) Simulation results for 100 m × 150 m region size: Remaining network
energy vs. Rounds.
energy vs. Rounds.

algorithms. This leads to better network stability. Though, the algo-


algorithm are implemented using MATLAB R2016a tool. The net- rithm has early dying of first node, it has negligible effect on IoT
work simulation parameters used are given in Table 2. network with thousands of sensor nodes deployed for farmland
Fig. 5 shows the effect of clustering techniques on cluster cen- monitoring. The stability of the network is accounted when maxi-
ter point estimation. The center points are marked with * in red mum number of the sensor nodes sustain for long period of time.
color. The clusters formed in Fig. 5(a) have center points that are This stability is achieved here. Fig. 10 represents effect of increasing
calculated using node locations. Thus, FCM generates center points number of nodes on HND of the algorithms. The proposed algorithm
approximately at the center positions of the respective groups of outperforms attaining maximum rounds than the three conven-
nodes. Ultimately, the distances of cluster members from their tional.
respective cluster center points in Fig. 5(a) are less than that of The algorithm is also tested for different monitoring area. For
in Fig. 5(b). this simulation scenario, 200 nodes are considered. Sink is located
To evaluate the performance of the proposed algorithm, sim- at (0,0) and fixed throughout the network simulation. Figs. 11–14
ulations are carried out for increased node density and size of shows number of alive nodes and remaining network energy for dif-
monitoring region. Figs. 6–9 shows the graphs of number of alive ferent area sizes of 100 m × 100 m, 100 m × 150 m, 200 m × 200 m
nodes and remaining network energy for 200, 300, 400 and 500 and 200 m × 250 m respectively. A five different network node
nodes respectively. In all the cases, proposed algorithm is more sta- deployments are used to calculate average FND, HND and LND
ble than the other algorithms. This is because for initial first round for given dimensions. The corresponding readings of each WSN
CHNs are selected based on distance probability as all nodes have deployment are tabulated in Tables 7–10. As area size increases, the
same energy. For later rounds, CHN rotations are done based on transmission distance also increases. This leads to more energy con-
energy probability with respect to cluster energy. An area size of sumption as per the explanation illustrated for the energy model
100 m × 100 m is considered with sink placed outside the sensor in Section 2.
network. Sink is located at (50,150). The effect of varying dimensions on the HND and LND of the
The randomly deployed nodes for ten different WSN networks simulated algorithms are represented in Figs. 15 and 16
are considered to evaluate performance of the algorithms in terms respectively. The proposed algorithm still holds better results
of FND, HND and LND. Tables 3–6 give statistics of the results for than the three conventional. This proves that the proposed algo-
increase in node density of the network. It shows that proposed rithm is energy efficient and scalable and can be used for precision
algorithm has significantly greater HND values compared to other agriculture.
A. Rajput, V.B. Kumaravelu / Sustainable Computing: Informatics and Systems 22 (2019) 62–74 73

5. Conclusions

In this work, an energy efficient and scalable clustering algo-


rithm is implemented for agricultural application of large scale
IoT systems. The main requirement of precision agriculture is
scalability, coverage and lifetime prolongation. FCM algorithm is
implemented to form optimal clusters in the network. Node loca-
tions and the number of clusters are the inputs to FCM algorithm.
It estimates cluster center points in the monitoring land for bet-
ter cluster structure. This reduces data transmission distances of
the nodes considerably. The results show that proposed algorithm
can form superior clusters to enhance network lifetime. A per-
ceived probability based on intra cluster distance and node energy
is modeled to select proper CHN. This enables proper CHN rota-
tions, which balances the network load among nodes. The load
balancing achieved thus results in stable death rate over long period
of time enhancing HND and LND of the network. The proposed
algorithm outperforms the conventional in terms of maximizing
network lifetime. Also, comparative results of varying node den-
sity and coverage area on HND prove the sustainability of the
network for scalable nodes and dimensions. Thus, the proposed
algorithm attains the requirement goals of precision agriculture
and can be preferred for implementation. In future, the algorithm
can be extended for mobile nodes and sink to gain more flexibility
in the various farming operations.

References

[1] J.A. Stankovic, Research directions for the Internet of things, IEEE Internet
Things J. 1 (1) (2014) 3–9.
[2] J. Gubbi, R. Buyya, S. Marusic, M. Palaniswami, Internet of things (IoT): a
vision, architectural elements, and future directions, Future Gener. Comput.
Syst. 29 (7) (2013) 1645–1660.
[3] L. Atzori, A. Iera, G. Morabito, The Internet of things: a survey, Comput. Netw.
54 (15) (2010) 2787–2805.
[4] M.U. Farooq, W. Muhammad, M. Sadia, K. Anjum, K. Talha, A review on
Internet of things, Int. J. Comput. Appl. 113 (1) (2015).
Fig. 14. (a) Simulation results for 200 m × 250 m region size: Number of alive nodes [5] O. Tamoghna, M. Sudip, S.G. Narendra, Wireless sensor networks for
vs. Rounds. (b) Simulation results for 200 m × 250 m region size: Remaining network agriculture: the state of the art in practice and future challenges, Comput.
energy vs. Rounds. Electron. Agric. 118 (2015) 66–84.
[6] S. Zhengguo, W. Hao, Y. Changchuan, H. Xiping, Y. Shusen, C.M. Victor,
Lightweight management of resource-constrained sensor devices in Internet
of things, IEEE Internet Things J. 2 (5) (2015).
[7] R. Tifenn, B. Abdelmadjid, C. Yacine, Energy efficiency in wireless sensor
networks: a top down survey, Comput. Netw. 67 (2014) 104–122.
[8] T. Dan, G. Laura, T. Nicolae, Radio transceiver consumption modelling for
multi-hop wireless sensor networks, UPB Sci. Bull. 75 (1) (2013).
[9] S.R. Theodare, Wireless Communication: Principles and Practice, Prentice
Hall, Upper Saddle River, NJ, 2001.
[10] M. Robert, J. Curry, S. Cole, A survey of optimization algorithms for wireless
sensor network lifetime maximization, Comput. Ind. Eng. 101 (2016)
145–166.
[11] A. Kumar, H. Shwe, K. Wong, P. Chong, Location-based routing protocols for
wireless sensor networks: a survey, Wirel. Sens. Netw. 9 (2017) 25–72.
[12] M. Hatamian, H. Barati, A. Movaghar, CGC: centralized genetic-based
clustering protocol for wireless sensor networks using onion approach,
Telecommun. Syst. 62 (4) (2016) 657–674.
[13] H.D. Tarigh, M. Sabaei, A new clustering method to prolong the lifetime of
Fig. 15. Effect of size of monitoring region on HND for simulated algorithms. WSN, Proceedings of International Conference on Computer Research and
Development (ICCRD) (2011) 143–148.
[14] F. Bajaber, I. Awan, Centralized dynamic clustering for wireless sensor
network, Proceedings of International Conference on Advanced Information
Networking and Applications Workshops (2009) 193–198.
[15] J.Y. Chang, H.J. Pei, An efficient cluster-based power saving scheme for
wireless sensor networks, EURASIP J. Wirel. Commun. Netw. (2012) 1–10.
[16] W.R. Heinzelman, A. Chandrakasan, H. Balakrishnan, Energy-efficient
communication protocol for wireless microsensor networks, Proceedings of
the 33rd Hawaii International Conference on System Sciences (2000) 1–10.
[17] M.J. Handy, M. Haase, D. Timmermann, Low energy adaptive clustering
hierarchy with deterministic cluster-head selection, Proceedings of Fourth
IEEE Conference on Mobile and Wireless Communications Network (2002)
368–372.
[18] P. Thulasiraman, K.A. White, Topology control of tactical wireless sensor
networks using energy efficient zone routing, Digital Commun. Netw. 2 (1)
(2016) 1–14.
[19] H. Lin, L. Wang, R. Kong, Energy efficient clustering protocol for large-scale
Fig. 16. Effect of size of monitoring region on LND for simulated algorithms. sensor networks, IEEE Sens. J. 15 (12) (2015) 7150–7160.
74 A. Rajput, V.B. Kumaravelu / Sustainable Computing: Informatics and Systems 22 (2019) 62–74

[20] A. Wang, D. Yang, D. Sun, A clustering algorithm based on energy information [30] J.S. Lee, W.L. Cheng, Fuzzy-logic-based clustering approach for wireless sensor
and cluster heads expectation for wireless sensor networks, Comput. Electr. networks using energy predication, IEEE Sens. J. 12 (9) (2012) 2891–2897.
Eng. 38 (3) (2012) 662–671. [31] R. Logambigai, A. Kannan, Fuzzy logic based unequal clustering for wireless
[21] J. Nadeem, B.R. Muhammad, I. Muhammad, G. Mohsen, A.K. Zahoor, A.A. Turki, sensor networks, Wirel. Netw. 22 (3) (2016) 945–957.
I. Manzoor, An energy-efficient distributed clustering algorithm for [32] M. Shokouhifar, F. Farokhi, An artificial bee colony optimization for feature
heterogeneous WSNs, EURASIP J. Wirel. Commun. Netw. 2015 (2015) 151. subset selection using supervised fuzzy c-means algorithm, Proceedings of
[22] V. Pal, G. Singh, R.P. Yadav, Balanced cluster size solution to extend lifetime of the International Conference on Information Security and Artificial
wireless sensor networks, IEEE Internet Things J. 2 (5) (2015) 399–401. intelligence (2010) 427–432.
[23] M. Tarhani, Y.S. Kavian, S. Siavoshi, SEECH: Scalable energy efficient clustering [33] P. Rajesh, T. Sachin, Energy aware fuzzy based multi-hop routing protocol
hierarchy protocol in wireless sensor networks, IEEE Sens. J. 14 (11) (2014) using unequal clustering, Wirel. Pers. Commun. 94 (3) (2017) 809–833.
3944–3954. [34] G. Anjana, D. Sonika, Performance analysis of various fuzzy clustering
[24] J.Y. Chang, A distributed cluster computing energy-efficient routing scheme algorithms: a review, Procedia Comput. Sci. 79 (2016) 100–111.
for Internet of things systems, Wirel. Pers. Commun. 82 (2) (2014) 757–776. [35] X. Shaoping, H. Lingyan, Y. Xiaohui, L. Xiaoping, A cluster number adaptive
[25] D. Jia, H. Zhu, S. Zou, P. Hu, Dynamic cluster head selection method for fuzzy c-means algorithm for image segmentation, Int. J. Signal Process. Image
wireless sensor network, IEEE Sens. J. 16 (8) (2016) 2746–2754. Process. Pattern Recognit. 6 (5) (2013) 191–204.
[26] C. Duo, C. Du-Wu, W. Chao-Xue, W. Zhu-Rong, A rough set-based hierarchical [36] M.A. Balafar, Fuzzy C-mean based brain MRI segmentation algorithms, Artif.
clustering algorithm for categorical data, Int. J. Inform. Technol. 12 (3) (2006). Intell. Rev. 41 (3) (2014) 441–449.
[27] D.C. Hoang, R. Kumar, S.K. Panda, Optimal data aggregation tree in wireless [37] C.H. Timothy, C.B. James, L. Christopher, O.H. Lawrence, P. Marimuthu, Fuzzy
sensor networks based on intelligent water drops algorithm, IET Wirel. Sens. c-means algorithms for very large data, IEEE Trans. Fuzzy Syst. 20 (6) (2012).
Syst. 2 (3) (2012) 282–292. [38] D.C. Hoang, R. Kumar, S.K. Panda, Realisation of a cluster-based protocol using
[28] K.M. Ajay, K. Rakesh, K. Vimal, S. Jitendra, A Grid-based approach to prolong fuzzy C-means algorithm for wireless sensor networks, IET Wirel. Sens. Syst. 3
lifetime of WSNs using fuzzy logic, in: Advances in Computational (3) (2013) 163–171.
Intelligence, 2017, pp. 11–22.
[29] P. Nayak, A. Devulapalli, A fuzzy logic-based clustering algorithm for WSN to
extend the network lifetime, IEEE Sens. J. 16 (1) (2016) 137–144.

You might also like