2023 IEEE International Conference on Communication, Networks and Satellite (COMNETSAT)

An Effective Cluster Head Selection

Algorithm using Machine Learning in IoNT
2023 IEEE International Conference on Communication, Networks and Satellite (COMNETSAT) | 979-8-3503-4110-2/23/$31.00 ©2023 IEEE | DOI: 10.1109/COMNETSAT59769.2023.10420624

Omer GULEC
Department of Management Information Systems
Pamukkale University
Denizli, Turkey
[email protected]

Abstract — Clustering is the first technique that comes to mind CHS techniques can be categorized as follows; clustering,
in order to achieve efficient routing in sensor networks. Instead artificial intelligence (AI) and low-energy adaptive
of transmitting the packets to each other, selecting a cluster clustering hierarchy (LEACH) approaches and optimization
head (CH) among the nodes is the best way to collect the packets
from the cluster members which leads to saving energy, reducing algorithms [6]. For this topic, Rami et al. (2023) [7]
network traffic, preventing packet loss and prolonging network designed a CHS model for Wireless Sensor Networks
lifetime. Cluster head selection (CHS) is a challenging process (WSNs) using improved Grey Wolf Optimization (GWO)
in a network therefore, CHS should be efficient and effective which reduces energy consumption and prolongs network
in Wireless Nano-Sensor Networks (WNSNs) due to the nano- lifetime. Yadav and Mahapatra (2021) [8] proposed a hybrid
domain characteristics. In this paper, an effective CHS algorithm energy-aware CHS model by hierarchical routing that
using Machine Learning (ML) is proposed for Wireless Nano-
minimizes delay and distance. Han et al. (2019) [9] proposed
Sensor Networks (WNSNs) and Internet of Nano-Things (IoNT)
applications. The proposed algorithm (PA) is compared with an a novel edge-computing supported CHS technique for
ordinary cluster head selection (OCHS) algorithm. According to Internet of Medical-Things (IoMT) applications in order to
the simulation results, PA provides nano-sensor node coverage provide effective communication that prolongs the network
on the network by 89.235% while it covers 20.355% more nano- lifetime.
nodes and spends 1.29 minutes less compared to OCHS on
average. Unfortunately, traditional sensor network approaches are
not suitable for WNSNs due to nano-domain
Keywords — Cluster Head Selection, Internet of Nano-Things, characteristics [10] hence, novel techniques should be
Machine Learning, Wireless Nano-Sensor Networks developed for IoNT applications. Therefore, in WNSNs,
I. INTRODUCTION there are several clustering-based routing protocols exist.
Gulec and Sahin (2023) [11] proposed a Red Deer Algorithm
Internet of Nano-Things (IoNT) concept is an emerging (RDA)-based clustering method to propose efficient routing
area for communication that enables the discovery of on the network. Gulec (2022) [5] proposed a clustering-based
molecular domains thanks to recent developments in routing protocol for IoNT which prolongs the network
nanotechnology and nano-sensors. IoNT applications have lifetime. Verma et al. (2018) [12] proposed an energy-
become more popular day by day in such areas as the military, balanced clustering scheme in order to achieve energy
industry, rescue operations, healthcare, agriculture, robotics, efficiency for routing in WNSNs. Bouchedjera et al. (2020)
energy and food industry [1]–[4]. The tiny nano-sensors in [13] proposed a cluster-based routing system for nano-
IoNT applications can communicate wirelessly with each networks. Pierobon et al. (2014) [14] proposed a hierarchical
other in a distributed manner which forms the Wireless Nano- cluster-based routing algorithm for WNSNs. Xu et al. (2022)
Sensor Networks (WNSNs) [5] which is the backbone of [15] proposed an energy-efficient routing scheme for
IoNT applications. WNSNs.
As in traditional sensor networks, one of the main Recently, Machine Learning (ML) methods have been
problems of WNSNs is routing due to the energy constraints quite helpful in solving WSN problems [16]. Dampage
of nano-sensors. In order to achieve efficient communication et al. (2022) [17] proposed an ML-supported model to detect
on WNSNs with low energy consumption and packet loss. forest fires at the initial stage. Srivastava and Bharti (2023)
On the other hand, clustering is a well-known technique that [18] hybridized Support Vector Machines (SVM) and
has been used for a long time for providing efficient routing Isolation Forest (IF) to detect anomalies in the data of
in sensor networks. In order to form a cluster, several node WSNs. Fan et al. (2023) [19] proposed a Particle Swarm
characteristics can be taken into account such as node Optimization (PSO) based ML scheme for fault detection in
locations, residual energy and the proximity to routers. WSNs. Gite et al. (2023) [20] proposed an underwater
Besides, among these cluster nodes, one of them needs to object-tracking and monitoring method using ML and
be selected to manage as the head of the cluster. Therefore, WSNs which enable the prediction of mobile nodes. ML
cluster head selection (CHS) is a crucial and challenging methods are useful not only in sensor networks but also in
process for clustering. In the literature, many CHS WNSNs. Therefore, Gulec (2023) [21] proposed a nano-
techniques have been developed for sensor networks. These

979-8-3503-4110-2/23/$31.00 ©2023 IEEE 672

Authorized licensed use limited to: VIT University. Downloaded on February 22,2024 at 01:46:16 UTC from IEEE Xplore. Restrictions apply.
 2023 IEEE International Conference on Communication, Networks and Satellite (COMNETSAT)

router localization algorithm using ML in WNSNs. Nieto energy loss. Therefore, optimal CHS algorithms should be
(2019) [22] designed an Internet of Bio-Nano-Things provided for WNSNs. In the proposed CHS model, the
(IoBNT) network to restore the flux of insulin in patients CHs are selected upon nodes’ selection factor (𝑠𝑓) which
using ML. Bakhshi and Shahid (2019) [23] proposed a is represented by (1).
security framework for IoBNT applications using ML to
𝑒 × 𝐶𝑝𝑟𝑜𝑏
identify anomalous operations. Similarly, Zafar et al. (2021) 𝑠𝑓 = ()
|Γ|×√𝑑
[24] proposed a model for securing the bio-cyber interface of
IoBNT using PSO and Artificial Neural Networks (ANN) as
The selection factor (𝑠𝑓) of a node is calculated by the
ML-based parameter profiling. Galal and Hesselbach (2022)
node’s residual energy (𝑒) times the probability parameter
[25] analyzed five supervised ML models for traffic
(𝐶𝑝𝑟𝑜𝑏 ) is divided by the node’s degree (|Γ|) times the square
classification in nano-networks that use electromagnetic
communication (EM). root of the distance between the node and the centroid (𝑑).
𝐶𝑝𝑟𝑜𝑏 is an initial probability parameter to be chosen as a CH
To this end, in this paper, an effective CHS algorithm is node. This parameter is inversely proportional to the relative
proposed for IoNT using ML method in order to provide distance to the central nano-router in a cluster in which
highly efficient routing on WNSNs. To the best of our it is maximum then the selection probability will be
knowledge, there exists no ML-supported effective CHS minimum. According to the equation, when a node’s
algorithm for IoNT applications in the literature. In contrast residual energy is maximum, besides the neighbor count and
to the studies mentioned above, not only does PA form the the distance to the centroid is minimum, therefore, it has
clusters and build routing paths but also it finds the effective more chance to be selected. After selecting the most
CHs using the ML techniques and the proposed efficient powerful CHs, the other nodes are bound to these CHs as
model in WNSNs. For measuring the efficiency, PA is shown in Fig. 1. b.
compared with an ordinary CH selection algorithm (OCHS)
that selects CHs according to the node degrees. The B. The Proposed Algorithm
contributions of PA are summarized as follows. The proposed algorithm has two parts that runs both on
• PA forms efficient clusters on WNSNs and IoNT the nano-sensor node and nano-gateway sides. Initially, all
applications using k-means as an ML method and nano-sensor nodes are distributed on the medium and they
finds optimal centroids of the clusters. Therefore, broadcast 𝐻𝐸𝐿𝐿𝑂_𝑀𝐸𝑆𝑆𝐴𝐺𝐸 to the medium to find out
these centroids can be used in the CHS process. their neighbor nodes. When a nano-sensor node receives this
• PA finds the most powerful candidate nodes as CHs message, it adds the sender node to its neighbor list. After
according to the selection factor related to the cluster this process ends, all nano-sensor nodes calculate their own
centroids. degree. In the following, they send their coordinates, residual
• PA assigns the selected nodes as CHs and binds cluster energy (𝑒𝑚 ) and the degree (|Γ𝑚 |) to the nano-gateway.
head members to these CHs. Therefore, it provides
In the second part of PA, the nano-gateway receives this
efficient routing from a cluster member node to the
info from all nano-nodes on the network by 𝑁𝑂𝐷𝐸_𝐼𝑁𝐹𝑂
CH all over the network instead of broadcasting. This
leads to consuming less energy and having a longer message, it calculates the distances (𝑑𝑚 ) between the
network lifespan. central nano-router and each nano-nodes. After this, the nano-
gateway clusters the nano-nodes according to their distances
The rest of the paper is organized as follows. The system using k-means clustering as an ML method. Besides, the nano-
model including the clustering head selection and the gateway calculates the selection factor (𝑠𝑓𝑚 ) of each nano-
proposed algorithm is summarized in Section II. Simulation sensor node and determines the CH nodes for each cluster.
setup and experimental results are discussed in Section III. This info is sent to nano-sensor nodes.
Finally, Section IV concludes the study.
In the last part of PA, when all nano-sensor nodes receive
II. SYSTEM MODEL the cluster information from the nano-gateway with
A. Clustering and Cluster Head Selection 𝐶𝐻𝑆_𝐼𝑁𝐹𝑂 message including CH node IDs, they change
their 𝑛𝑜𝑑𝑒_𝑡𝑦𝑝𝑒 upon they are CH or nonCH. If a nano-
In order to build efficient clusters, k-means clustering as sensor node’s 𝑛𝑜𝑑𝑒_𝑡𝑦𝑝𝑒 is CH then it forms its own cluster
an ML clustering technique is chosen for the proposed and waits for data from its cluster members. If a nano-sensor
model. According to the location of the nodes relative to node’s 𝑛𝑜𝑑𝑒_𝑡𝑦𝑝𝑒 is not CH then it binds itself to the CH
the central nano-router which is placed at [0, 0], PA forms and sends data to its CH where it belongs to. The flowchart of
clusters on the network as shown in Fig. 1. a. Afterward, PA PA is shown in Fig. 2.
determines the optimal centroids of each cluster and then it
starts to find candidate CH nodes of each cluster. In a III. SIMULATION, EXPERIMENTAL RESULTS AND
general approach, the closest nodes can be chosen as CHs DISCUSSION
to these centroids. However, the closest nodes may not In order to simulate the PA and OCHS, 50 different
have enough energy therefore, they will be offline after a network topologies are generated for each simulation set
short time and this will lead the clusters to be rebuilt. Due having 750, 1000 and 2000 nano-nodes. NS-3 [26] and Nano-
to the nano-domain constraints, these operations will cause

673
Authorized licensed use limited to: VIT University. Downloaded on February 22,2024 at 01:46:16 UTC from IEEE Xplore. Restrictions apply.
 2023 IEEE International Conference on Communication, Networks and Satellite (COMNETSAT)

FIG. 1. A) A WNSN HAVING 750 NANO-NODES, COLORED NODES FOR CLUSTERING AND CENTROIDS B) CLUSTERS AND THE COVERAGE OVER THE NETWORK

TABLE I. SIMULATION PARAMETERS

Simulation Tools NS-3 and Nano-Sim
Python Packages scikit-learn, Networkx, Pygal
Machine Learning Method k-means Clustering
Network Area 0.01 × 0.01 m2
Transmission Range of Nano-Nodes 0.0015 m
Initial Energy of a Nano-Node 800 nJ
Packet Size 100 bytes
Number of Topologies 50 on each simulation
Number of Nano-Sensor Nodes 750, 1000 and 2000

On the other hand, when a WNSN consists of 750 nano-

nodes, OCHS covers 519.6 nodes while 230.4 nodes remain
with a coverage percentage of 69.28% spending 0.467 mins
on average. If a WNSN has 1000 nodes, OCHS covers 686.7
nodes while 313.3 nodes remain with a coverage percentage
of 68.67% spending 1.111 mins on average. Finally, if there
are 2000 nodes on a WNSN, OCHS covers 1373.8 nodes
while 626.2 nodes remain with a coverage percentage of
68.69% spending 3.173 mins on average.
When these experimental results are analyzed, it is clearly
seen that PA provides more nano-sensor node coverage up
to 20.355% compared to OCHS while spending 1.29 minutes
less on average. Consequently, PA outperforms the OCHS
according to the node coverage and simulation time.
FIG. 2. FLOWCHART OF PA Therefore, using the PA for CH selection and clustering on a
Sim [27] are chosen as simulation tools due to their support WNSN will provide efficient routing on the network.
of nano-domains [28] and Python packages such as scikit- In order to illustrate the results of PA and OCHS, the
learn [29], Networkx [30] and Pygal [31] are used for ML and coverage percentage comparison results are given in Fig. 3,
visualizations. In the simulations, the packet size is 100 bytes covered and remaining node comparison results are given in
and the initial energy of nano-sensor nodes is 800 nJ. All of Fig. 4, total simulation time comparison results are given in
the simulation parameters are given in Table I. Fig. 5 and all experimental simulation results are given in
According to the simulation results, when a WNSN consists of Table II.
750 nano-nodes, PA covers 668.6 nodes while 81.4 nodes
TABLE II. AVERAGE SIMULATION RESULTS
remain with a coverage percentage of 89.146% spending
0.093 mins on average. If a WNSN has 1000 nodes, PA Nodes 750 1000 2000
covers 893.7 nodes while 106.3 nodes remain with a coverage Covered Node Count of PA 668.6 893.7 1783.8
percentage of 89.37% spending 0.144 mins on average. Covered Node Count of OCHS 519.6 686.7 1373.8
Remaining Node Count of PA 81.4 106.3 216.2
Finally, if there are 2000 nodes on a WNSN, PA covers
Remaining Node Count of OCHS 230.4 313.3 626.2
1783.8 nodes while 216.2 nodes remain with a coverage
Coverage Percentage of PA 89.146 89.37 89.189
percentage of 89.189% spending 0.644 mins on average.
Coverage Percentage of OCHS 69.28 68.67 68.69

674
Authorized licensed use limited to: VIT University. Downloaded on February 22,2024 at 01:46:16 UTC from IEEE Xplore. Restrictions apply.
 2023 IEEE International Conference on Communication, Networks and Satellite (COMNETSAT)

Total Simulation Time of PA (min) 0.093 0.144 0.644 energy-efficient CHS techniques should be proposed for IoNT
Total Simulation Time of OCHS (min) 0.467 1.111 3.173 applications.
In this paper, an ML-supported CHS algorithm is designed
for IoNT applications in order to build efficient clusters and
select powerful CHs that provide continuous data transmission
from cluster member nodes to these CHs. PA finds optimal
centroids, determines powerful CHs, binds the other nodes
to these CHs, forms clusters and achieves efficient network
traffic over these clusters.
With 50 different simulation sets having 750, 1000 and
2000 nodes for each topology, PA is compared with an
ordinary CHS algorithm. These both algorithms are
developed and tested under NS-3 and Nano-Sim simulation
tools and scikit-learn is used for ML processes. For
FIG. 3. COVERAGE PERCENTAGE RESULTS OF PA AND OCHS visualization of the networks, Networkx and Pygal are used
for each simulation.
According to the experimental simulation results, PA
successfully provides node coverage on the network with
cluster forming by up to 89.37% of the nodes. Compared to
OCHS, PA covers 20.355% more nano-nodes while it spends
1.29 minutes less on average. In brief, PA outperforms the
OCHS technique according to the node coverage and
simulation time. The use of k-means as an ML method played
an important role for the selection of CHs in this paper. In
addition, the use of k-means clustering method to determine
the locations of the clusters to be formed on the network will
take this study further.
However, the PA takes much time while the ML method
clusters the nodes and finds the centroids. Therefore,
FIG. 4. COVERED AND REMAINING NODE COUNT OF PA AND OCHS algorithm complexity should be reduced in future studies.
Additionally, PA should tested in dense networks and
compared with other CHS techniques. In contrast to the
general approach, more nano-routers will be placed in future
studies instead of central nano-router. In addition, for better
performance, optimization algorithms will be used for the
CHS process in future studies.
REFERENCES

[1] E. Sahin, O. Dagdeviren, and M. A. Akkas, “Terahertz channel model

for search and rescue nanonetwork applications in debris areas,” in
2021 International Balkan Conference on Communications and
Networking (BalkanCom). IEEE, 2021, pp. 66–70.
[2] O. Gulec, “Distributed routing and self-balancing algorithm for
medical IoNT,” Simulation Modelling Practice and Theory, p. 102833,
2023.
[3] L. Sellappan, S. Manoharan, A. Sanmugam, and N. T. Anh, “Role of
FIG. 5. TOTAL SIMULATION TIME RESULTS OF PA AND OCHS nanobiosensors and biosensors for plant virus detection,” in Nanosen-
sors for Smart Agriculture. Elsevier, pp. 493–506, 2022.
[4] O. Gulec, “Hybrid FFBAT optimized multi-hop routing in Internet of
Nano- Things,” Internet of Things, 2023.
IV. CONCLUSION [5] O. Gulec, “Extending lifetime of Wireless Nano-Sensor Networks: An
energy efficient distributed routing algorithm for Internet of Nano-
Clustering is one of the best ways in order to provide Things,” Future Generation Computer Systems, 2022.
efficient routing in WNSNs and IoNT. However, CHS is [6] R. Ramya and T. Brindha, “A comprehensive review on optimal cluster
a challenging issue due to circumstances such as energy head selection in wsn-iot,” Advances in Engineering Software, vol.
constraints. In order to select cluster heads for clusters, 171, p. 103170, 2022.
there exist several techniques developed for WSNs in the [7] M. Rami Reddy, M. Ravi Chandra, P. Venkatramana, and R. Dilli,
“Energy-efficient cluster head selection in wireless sensor networks
literature. Nonetheless, these techniques may not be suitable using an improved grey wolf optimization algorithm,” Computers, vol.
due to nano-domain characteristics. Therefore, optimal and 12, no. 2, p. 35, 2023.

675
Authorized licensed use limited to: VIT University. Downloaded on February 22,2024 at 01:46:16 UTC from IEEE Xplore. Restrictions apply.
 2023 IEEE International Conference on Communication, Networks and Satellite (COMNETSAT)

[8] R. K. Yadav and R. P. Mahapatra, “Energy aware optimized clustering Networks”, Wireless Personal Communications, 129(4), 2693-2710,
for hierarchical routing in wireless sensor network,” Computer Science 2023.
Review, vol. 41, p. 100417, 2021. [19] F. Fan, S.-C. Chu, J.-S. Pan, C. Lin, and H. Zhao, “An optimized
[9] T. Han, L. Zhang, S. Pirbhulal, W. Wu, and V. H. C. de Albuquerque, machine learning technology scheme and its application in fault
“A novel cluster head selection technique for edge-computing based detection in wireless sensor networks,” Journal of Applied Statistics,
iomt systems,” Computer Networks, vol. 158, pp. 114–122, 2019. vol. 50, no. 3, pp. 592–609, 2023.
[10] A. O. Balghusoon and S. Mahfoudh, “Routing protocols for wireless [20] P. Gite, A. Shrivastava, K. M. Krishna, G. Kusumadevi, R. Dilip, and
nanosensor networks and internet of nano things: A comprehensive R. M. Potdar, “Under water motion tracking and monitoring using
survey,” IEEE Access, vol. 8, pp. 200 724–200 748, 2020. wireless sensor network and machine learning,” Materials Today:
[11] O. Gulec and E. Sahin, “Red Deer Algorithm based nano-sensor node Proceedings, vol. 80, pp. 3511–3516, 2023.
clustering for IoNT,” Journal of Network and Computer Applications, [21] O. Gulec, “Machine learning supported nano-router localization in
p. 103591, 2023. WNSNs,” Sakarya University Journal of Science, vol. 27, no. 3, pp.
[12] S. Verma, S. Kaur, G. Dhiman, and A. Kaur, “Design of a novel energy 590 – 602, 2023.
efficient routing framework for wireless nanosensor networks,” in 2018 [22] H. Nieto-Chaupis, “Prospective end-to-end internet of bio-nano things
First International Conference on Secure Cyber Computing and network based in a machine learning algorithm to restore the flux of
Communication (ICSCCC). IEEE, 2018, pp. 532–536. insuline in prediabetic patients,” in 2019 IEEE CHILEAN Conference
[13] I. A. Bouchedjera, L. Louail, Z. Aliouat, and S. Harous, on Electrical, Electronics Engineering, Information and
“DCCORONA: Distributed cluster-based coordinate and routing Communication Technologies (CHILECON). IEEE, 2019, pp. 1–4.
system for nanonet- works,” in 2020 11th IEEE Annual Ubiquitous [23] T. Bakhshi and S. Shahid, “Securing internet of bio-nano things: Ml-
Computing, Electronics & Mobile Communication Conference enabled parameter profiling of bio-cyber interfaces,” in 2019 22nd
(UEMCON). IEEE, 2020, pp. 0939–0945. International Multitopic Conference (INMIC). IEEE, 2019, pp. 1–8.
[14] M. Pierobon, J. M. Jornet, N. Akkari, S. Almasri, and I. F. Akyildiz, [24] S. Zafar, M. Nazir, A. Sabah, and A. D. Jurcut, “Securing bio-cyber
“A routing framework for energy harvesting wireless nanosensor interface for the internet of bio-nano things using particle swarm
networks in the Terahertz Band,” Wireless networks, vol. 20, pp. 1169– optimization and artificial neural networks based parameter profiling,”
1183, 2014. Computers in Biology and Medicine, vol. 136, p. 104707, 2021.
[15] J. Xu, H. Huang, J. Kan, and R. Wang, “Energy-balanced routing [25] A. Galal and X. Hesselbach, “Machine learning models for traffic
protocol based on data priority for lung terahertz nanosensor classification in electromagnetic nano-networks,” IEEE Access, vol.
networks,” in 2022 IEEE 95th Vehicular Technology 10, pp. 38 089–38 103, 2022.
Conference:(VTC2022-Spring). IEEE, 2022, pp. 1–5. [26] NS-3, https://www.nsnam.org, accessed 20 September 2023.
[16] T. Kim, L. F. Vecchietti, K. Choi, S. Lee, and D. Har, “Machine [27] Nano-Sim, https://telematics.poliba.it, accessed 20 September 2023.
learning for advanced wireless sensor networks: A review,” IEEE
[28] E. Sahin, O. Dagdeviren, and M. A. Akkas, “An evaluation of internet
Sensors Journal, vol. 21, no. 11, pp. 12 379–12 397, 2020.
of nano-things simulators,” in 2021 6th International Conference on
[17] U. Dampage, L. Bandaranayake, R. Wanasinghe, K. Kottahachchi, and Computer Science and Engineering (UBMK). IEEE, 2021, pp. 670–
B. Jayasanka, “Forest fire detection system using wireless sensor 675.
networks and machine learning,” Scientific reports, vol. 12, no. 1, p.
[29] scikit-learn, https://scikit-learn.org, accessed 20 September 2023.
46, 2022.
[30] Networkx, https://networkx.org, accessed 20 September 2023.
[18] Srivastava, A., and Bharti, M. R., “Hybrid Machine Learning Model
for Anomaly Detection in Unlabelled Data of Wireless Sensor [31] Pygal, https://www.pygal.org, accessed 20 September 2023.

676
Authorized licensed use limited to: VIT University. Downloaded on February 22,2024 at 01:46:16 UTC from IEEE Xplore. Restrictions apply.