Event Detection in Clustered Wireless Sensor

Networks Using Dynamic Cell Structures Neural

Networks
M. Othmani1*, T. Ezzedine2, and Q. G. Wang3
1,2
University Tunis El Manar, National Engineering School of Tunis, Communication System Laboratory, Sys’Com, Campus Universitaire
Farhat Hached El Manar, BP 37, Le Belvedere 1002 Tunis, Tunisia.
3
Institute of Artificial Intelligence and Future Networks, Beijing Normal University at Zhuhai; BNU-HKBU United International College,
Zhuhai; Guangdong, PR China.
* Corresponding author: Mohsen. Othmani (e-mail: [email protected]).

ABSTRACT Localization of intrusions at country borders is a very important task for surveillance
applications. This task can be approached within the framework of Wireless Sensor Networks (WSN). In this
context, we investigate the localization of an intruder in a clustered WSN using Dynamic Cell Structures
Neural Networks (DCSNN). Sensors are randomly deployed according to a Poisson Point Process and the
Region of Interest (ROI) is considered as one, four, and nine clusters respectively in the simulations. We start
by presenting a novel scenario of the communication model, where we add a new condition to any Cluster
Head (CH) having its sensor nodes detecting, justified the true hypothesis test. Thus, only one CH is elected
by the Fusion Center (FC) to run the DCSNN algorithm for each intrusion. This can help us to avoid the false
alarm due to the additive noise sensing to the power of the signal. On other hand, we prove our choice of
DCSNN through a comparative study between it and the MLPNN. The results collected by the simulations
about localization and taking into account the variation of some important parameters like Probability of false
alarm (Pfa), the deployment density (𝜆), and the transmission power of sensors (P0), show that more the
region is divided into clusters, the smaller the localization mean error becomes in a short time.

INDEX TERMS Cluster, Distributed detection, Dynamic cell structures neural network, Localization,
Wireless sensor networks.

I. INTRODUCTION requirements, etc., these networks become well suited for

monitoring in harsh conditions, over large areas, where
Wireless Sensor Networks (WSN) are a promising research
applicators can work free of network wiring and network
platform. They represent the backbone of many systems:
service issues. However, some tasks are required to
surveillance system for railway traffic [1], border surveillance
accomplish this mission: sensing, design and deployment,
[2], smart agriculture [3], etc... In most cases, the areas in
classification, and essentially localization and tracking of the
which the WSN are deployed for surveillance application are
target. In this paper, we focus on localization and tracking.
dangerous or inaccessible to humans such as the northwest
border between Tunisia and Algeria. Despite the bifurcation By briefly going through the algorithms and techniques used
of the road and the presence of a large number of predators, for localization, we can cite (not exclusively) the following:
this area is often a station for surreptitious passage and One of the first solutions is to attach a global positioning
smuggling of commercial goods and even a favorite hiding system (GPS) to each sensor node. This solution is rejected
place for terrorists. The surveillance of cross-border activities because it is greedy in energy and it will be very expensive.
can be implemented in a very convenient and reliable manner Thus, we have to resort to other techniques [4]. WSN
using wireless sensor networks. In fact, due to its low cost, localization algorithms are divided into two categories,
ease of maintainability, no complicated equipment Range-Based and Range-Free:

1
  The first, dependent on distance measurements such distributed detection in a clustered wireless sensor network as
as Received Signal Strength Indication (RSSI) [5], an architecture. The structure of the WSN is three-level
Time of Arrival (ToA) [6], Time Difference of hierarchical (Sensor Nodes (SN), Cluster Head (CH), and
Arrival (TDoA) [7] or Angle of Arrival (AoA) [8]. Fusion Center (FC)). The Region of interest is considered as a
They are used to calculate the distances between the set of clusters. In each one, we use a stochastic process to
unknown nodes and beacon nodes but, they provide model the WSN, more precisely a stationary Poisson point
relatively accurate localization results. Excluding process [14]. Based on the signal strength measurements (the
the RSSI technique, we have usually required extra strength of the received signal decreases as the distance to the
hardware. target increases), the Sensor Nodes (SN) compute local
 In the second category (Range-Free), no additional decisions about the presence of the intruder and send them to
hardware is required. The localization algorithms the cluster heads (CHs). To avoid the false alarm due to the
such as APIT, MDS-MAP, and Centroid described additive noise sensing to the power of the signal, CHs
by [9], and DV-Hop [10] localize the unknown collaborate with the FC according to a novel scenario
nodes using proximity or connectivity information. described in the communication model (Section III.2) to elect
the cluster subject of intrusion and then run the Dynamic Cell
A large community of academics and researchers is moving
Structures Neural Network algorithm.
towards machine learning-based localization methods. Data
collected from anchor nodes (mainly RSSI values or Hop-
The main contributions of this paper are:
count information) feeds a training database for prediction
algorithm methods. For example, Neural Networks [11],  Present a new event detection strategy in a large-
Supports Vector Machine (SVM) [12], and Supports Vector scale area while respecting the constraints of limited
resources (energy and bandwidth): only sensors that
Regression (SVR) [13].
have detected the presence of an intruder will
On the other hand, the major challenge with sensor nodes is participate in the localization phase.
how to maintain a longer lifetime since they are powered by
batteries. Moreover, the communication between sensor nodes  Due to the imperfection of the channels that suffer
from Additive White Gaussian Noise (AWGN), we
is very limited by the wireless channel, mainly in poor study the case of false intrusion detections according
conditions. All these factors led to finding a collaborative to the probability of false alarm (Pfa) and the
solution between the sensor nodes to overcome these deployment density (𝜆): in addition to the true
constraints and succeed in their missions, (classification, hypothesis test which must be justified by the sensor
localization, tracking, etc...). Indeed, the objective is to nodes detecting, we add a new condition to the
combine data and knowledge from different sources to cluster head to be able to perform the localization
maximize useful information content. phase.
The motivation for the present work comes from our  Present a new kind of neural network (DCS) to
previous paper entitled "Border trespasser classification using exploit it in the localization phase in order to achieve
artificial intelligence" [2] where the Region of Interest (ROI) better accuracy in a short time. We are conducting a
is the border area between Tunisia and Algeria, represents the comparative study between the DCS network and
preferred path for those who want to penetrate and smuggle MLP network to show the superiority of DCS
from one country to another in an irregular way and especially performance in terms of mean error and localization
for terrorists. So, we have studied the kinematics model of execution time. Therefore, we justify our choice of
intruders (Person, Soldier, Animals for Contraband Purposes the conjunction of this kind of neural network with
(ACP)) based on speed, acceleration, and bearing, then the the WSN.
sensing phase, and finally we have achieved with a The simulations of the whole concept of this manuscript are
comparative study between six classification techniques: performed using the MATLAB R2018b software.
Naïve Bayes, SVM, Multilayer Perceptron (MLP), BF Tree,
LAD Tree, and J48. This is how we find ourselves in full The rest of this paper is organized as follows: Section II is
enthusiasm to study and propose a solution for the border devoted to related works. In section III, we present the sensor
trespasser event detection in WSN using DCSNN. network model. Then, through the communication model, we
In critical applications such as border surveillance, it is described our scenario for localization. In Section IV, we
crucial to detect the area of the event after an intruder is study false intrusion detection. Section V is reserved to present
detected. Thus, dividing the region of interest into an event the Dynamic Cell Structure Neural Networks algorithm and its
zone and a non-event zone (peace zone) can allow border comparative study with the Multi-Layer Perceptron neural
supervisors to have a general view via a dashboard that allows networks (MLPNN). Finally, the localization of the intruder is
them to distinguish between the different regions and track the simulated in section VI. At the end of this paper, we conclude
intruder in time from the point of entry to exit point or arrest with a discussion and conclusion.
point. Thus border supervisors can estimate the impact of each
event and make appropriate plans. Therefore, we will adopt a
2
 II. RELATED WORKS For instance, Singh and Agrawal [22] propose a time
difference of arrival based node localization using neural
The appropriate selection of the topology of a WSN is a
networks. They compared two models: the Back Propagation
fundamental pillar to facilitate the achievement of the
Network (BPN) and the Radial Basis Function (RBF). They
objectives for which the WSN has been deployed. Many prior
concluded the superiority of the RBF in term of root mean
surveillance applications using WSNs have adopted a perfect
square error. In [23], the authors used artificial neural
linear deployment as a barrier cover that requires fewer sensor
networks to establish the relationship between the location of
nodes but can experience radio disconnection due to sensor
the unknown mobile node and the experimentally obtained
failure and exhaustion [15]. The major inconvenience of this
values of RSSI and the Link Quality Indicator (LQI).
type of architecture is that it does not allow the tracking phase,
because once the intruders cross the sensors-barrier, they are Compared to existing systems, we present in this manuscript a
lost and can no longer be tracked. Clustering [16,17] is the new detection strategy that is distinguished by the following
most popular WSN topology management method. This features:
method is mainly based on dividing the randomly dispersed
 In addition to the true hypothesis test that must be
wireless sensors into independent groups. Therefore, the
justified by the sensing sensor nodes (equation 5) and
outputs and results obtained from each group will be
in order to avoid the false intrusion detection
subsequently combined and fused for achieving a better
phenomenon (detailed in Section 4), we add a new
understanding of a phenomenon of interest and leading to the
condition to the cluster head to be able to perform the
final result that reflects the whole WSN. In this context, K.
localization phase (equation 10). This allows us to
Pradeepa et al. explain in [18] how clustering is a technique
prevent the CH from transmitting unnecessary data
used to increase the various capabilities of a sensor network.
to the FC and overloading it, thus reducing the
On their part, A. Sharhraki et al. [19] surveyed the most
number of communications between them and
prominent clustering techniques, praised their advantages such
respecting the constraints of limited resources
as quality improvement and energy consumption
(energy and bandwidth).
improvement. They also study about 215 clustering techniques
by categorizing and ranking them and providing statistics  Define a simple election system at the FC that
based on some selected criteria. determines which CH is subject to intrusion and
therefore subject to localization. To do so, we opt for
In [20], T. Darmala, and A. Mehmood describe the data
Single-hop inter-cluster communication where all
processing (collected at the U.S. southwest border) performed
CHs justify the previous point, communicate directly
to determine the target type. The detections collected at each
with FCs and avoid communications between them
node will be fused using algorithms based on the Neyman-
(multi-hop inter-cluster communication). Indeed,
Pearson criteria, to determine the required false alarm rate.
cluster communication is one of the performance
Then they share the results with adjacent unattended ground
metrics of clustered WSN. In this context, we fully
sensors (UGSs) to track people. Finally, to detect and classify
agree with the authors of [24] who find that single-
targets, the authors use PIR, seismic and ultrasonic sensors in
hop inter-cluster communication performs well
each UGS system. In [21], the authors treat the distributed
compared to multi-hop inter-cluster communication
detection and decision fusion for a multi-sensor system. Under
in terms of energy depletion.
some assumptions like the perfect channel between sensors
and fusion center and the known of the local sensors’  Define a dashboard for the borderline supervisors
performance indices, they discuss the design for optimal that shows them the event areas (subject to intrusion)
decision fusion rule at the fusion center and the optimal local so that they can take anticipatory measures while
decision rules at sensors using both Bayesian and Neyman– waiting for the accurate localization.
Pearson criteria. In the same vein, several other lines of
 There is extensive literature that has studied the
research have been conducted, such as collaborative signal
problem of WSN clusters. In [16], the author
processing [18], distributed target detection [19], data fusion
surveyed the state of the research and classified the
and optimal fusion rules [20,21], Mobile Target Tracking with
various schemes. Localization has been discussed in
Multiple Objectives, etc.
various contexts such as distributed estimation in a
On other hand, neural networks have proved their
performance as adaptive systems able to learn quickly and clustered WSN using quantized data [25], the effect
respond in real-time. It has been used in several fields of of uniform and non-uniform clustering [26], etc... On
application, such as recognition (form, voice, writing, etc.), our side, we chose to go ahead with neural networks:
aircraft control, imagery, etc...So the conjunction between we establish a conjunction between the clustering
neural networks and wireless sensor networks in localization technique and DCSNN to predict and study the
goal has proved an accuracy improvement about intruders. intruder location in a clustered wireless sensor
network. In Section 5, we present the superiority of
the performance of this type of network compared to
3
 MLPNN. Essentially, we focus on the effect of the Under ideal conditions (without noise), a sensor node (SN)
number of clusters in WSN to improve the i located in a cluster at the position 𝑥𝑖 can detect the signal
localization accuracy. emitted by a target having a signal power 𝑃𝑡 and located in the
ROI. The sense signal is as follow:
III. SENSOR NETWORK AND COMMUNICATION 𝑎(𝑥𝑖 ) = √𝑃𝑡 ⁄𝑚𝑎𝑥 (𝑑0 , 𝑑𝑖 ) (1)
MODEL
where d0 is the reference distance and di is the Euclidian
In this section, we present the sensor network model. We distance to the target.
define how to compute the local decision of each SN taking
into account the additive white Gaussian noise. Next, we
describe the communication model of the sensor network
and detail the adopted event detection scenario.

1. SENSOR NETWORK MODEL

A WSN is composed of a large number of low-cost,
resource-limited sensors (energy and bandwidth) densely
deployed in an ROI. In this paper, we consider the architecture
of the sensor network as three-level hierarchical. The region
of interest is considered as a set of clusters. In each one, we
use a stochastic process to model the WSN, more precisely a
stationary Poisson Point Process (PPP) [14]. In fact, PPP is the
most widely used point process to model spatial locations. We
assume that the sensor network has the following properties
and capabilities:
FIGURE 1. The influence field of a target detected by sensors.
 The topology of the network remains unchanged over
time. In real conditions, any amplitude (equation 1) of any SN
 In this paper, we assume that the CH does not have an located at a position 𝑥𝑖 will be corrupted by white Gaussian
energy problem. In fact, there are many researches and noise (𝒵𝑖 ): we consider a mean equal to zero and a
studies [27, 28] that suggest different ways to manage variance 𝜎𝑠2 , 𝑠𝑜 𝒵~𝒩(0, 𝜎 2 ). We assume that the noise is
energy resources in order to maximize the lifetime of identically and independently distributed (i.i.d.). So, the SNR
the cluster head which is most important in the is equal to sensing signal power by the variance:
combination, aggregation, and transmission of 𝑃𝑡
received data. In [29] the authors described some 𝑆𝑁𝑅 = (2)
𝜎2
modern energy-efficient clustering approaches to
improve the lifetime of WSNs. They proposed four Thus, the ith SN collects data from its environment in order to
clustering methods: (i) fuzzy-logic-based cluster head decide on the presence or absence of the target. The following
election, (ii) efficient sleep duty cycle for sensor nodes, form (equations 3, 4) takes into account the effect or not of
(iii) hierarchical clustering, and (iv) estimated energy (𝒵𝑖 ):
harvesting. Then they compared the performance of 𝐻1 : 𝑆(𝑥𝑖 ) = 𝑎(𝑥𝑖 ) + 𝒵𝑖 (3)
these approaches with classical clustering approaches
such as low energy adaptive clustering hierarchy 𝐻0 : 𝑆(𝑥𝑖 ) = 𝒵𝑖 (4)
(LEACH). The first step is to respect the network bandwidth and
 We consider a homogeneous PPP, which means that energy constraints, we only keep the SNs that have detected
intensity λ is independent of the sensor's node location the intrusion and will thus communicate with the CHs. This is
xi. Hence, we have the same average number of SNs in why we proceed from the beginning to the calculation of the
each cluster. local decision of each SN by defining a local decision
threshold τ which will be the lower bound of each detection.
In practice, the influence field is the area(s) within which
the Signal-to-Noise Ratio (SNR) of a target is detectable by 𝐹𝑎𝑙𝑠𝑒, 𝑆(𝑥𝑖 ) < 𝜏,
𝐼(𝑥𝑖 ) = { (5)
the sensors and exceeds their minimum detectable threshold. 𝑇𝑟𝑢𝑒, 𝑆(𝑥𝑖 ) ≥ 𝜏,
The interval of this area varies between a minimum and a
We assume that the threshold τ is the same for all SNs.
maximum value due to the sensitivity of the sensor, the power
of the noise, and other nuisance parameters such as the target Therefore the probabilities of detection 𝑃𝑑 and false alarm 𝑃𝑓𝑎
orientation (Fig. 1). can be written as:
4
 −(𝑡2 −𝑎(𝑥𝑖 )) Also, we assume that all clusters (implicitly CHs) are
1 ∞ 2𝜎2 𝜏−𝑎(𝑥𝑖 )
𝑃𝑑 (𝑥𝑖 ) = ∫ 𝑒 𝑠 dt = 𝑄 ( ) (6) identified and known by the fusion center.
𝜎𝑠 √2𝜋 𝜏 𝜎𝑠

−𝑡2 The sensor network communication model revolves around

1 ∞ 2 𝜏
𝑃𝑓𝑎 = ∫𝜏 𝑒 2𝜎𝑠 dt = 𝑄 (𝜎 ) (7) the detection scenario that we will present below:
𝜎𝑠 √2𝜋 𝑠
where 𝑄(. ) represent the Gaussian Q-function defined by: When an event detection has occurred, it sometimes
happens that a sensor node in cluster i receives the signal
∞ 1 𝑡2
𝑄(𝑥) = ∫𝑥 𝑒 − 2 𝑑𝑡 (8) emitted by an intruder in cluster j whose signal strength is
√2𝜋
higher than the threshold (𝑆(𝑥𝑖 ) ≥ 𝜏). This phenomenon can
2. SENSOR NETWORK COMMUNICATION MODEL occur when the intruder location is near the borders of several
clusters but mainly due to noise sensing. In [33, 34] this
As described above, our ROI was divided into 𝛫 clusters, problem is addressed by using the scan statistic (SS) detector.
populated by sensor nodes. We assume that all the clusters are On our part, we study this phenomenon in the next section.
identical and that each has an area equal to the square of size Thus, our strategy is entrusted to the CHs: each CHi will count
L × L. Given the Poisson distribution of density λ, there are on the detecting nodes belonging to it and which justifies
average λL2 sensor nodes in each cluster. If N is the total equation 5 then pass the information to the FC. The latter, in
number of sensors to be deployed, we can therefore easily turn, records the information received in a table in the form of
deduce that the ROI must be divided into some clusters 𝛫 two columns (the number of detector nodes, CH identifier),
equal to N / λL2. The calculation of the optimal number of then applies a simple descending order to deduce therefrom
clusters [30, 31] remains an open and interesting field of the CHi having the greatest number of nodes detecting the
research: the dilemma is how to balance between the number intruder. (More details on this point of view will be explained
of clusters for a large region of interest and the minimum in the next section). The FC updates a preliminary dashboard
energy to be consumed by each cluster (hence the energy of which gives a general view of the state of the region of interest
the whole WSN). In fact, the number of inter-cluster divided by zone (event region, not-event region). Finally, to
communications increases with the number of clusters. On the further reduce the communication complexity, we opt for the
other hand, the amount of intra-cluster communications unicast (communication from FC to the Elected CHi) instead
increases dramatically as the number of clusters decreases of the multicast (communication from FC to all CHs). The FC
[32]. informs the elected CHi that it is subject to intrusion. With
In each cluster, one sensor node is elected to play the role regard to the others CHi, they will consider that the detecting
of the Cluster Head (CH). A CH will not be concerned by the sensors are in false alarm when an allowable waiting time is
detection of an intruder. However, it has the mission of exceeded. For its part, the chosen CHi executes the DCSNN
middleware between the FC and the SNs verifying the true algorithm in order to predict the position of the intruder and to
hypothesis in (equation 5) via the calculation of their local transmit it to the FC for tracking (Fig. 2).
decision. We assume that both SNs and CHs are synchronized.

5
 FIGURE 2. Network Communication model.

As described in the introduction, border surveillance IV. FALSE INTRUSION DETECTION

officials clearly distinguish border areas by the extent of their
In the following, we show the effect of the additive white
strategic and sovereign importance to the country. Thus, we
Gaussian noise on the sensing signal. Therefore we make a
present the idea of a dashboard that divides the ROI into an
comparison between the probability of true intrusion detection
event region (or cluster) and a non-event region (Fig. 3). This
(TID) and the false intrusion detection (FID) respectively for
helps supervisors take effective proactive actions and
a ROI considered as 4 clusters, and 9 clusters. Next, we study
decisions while waiting for the accurate localization of the
the simulated and Poisson cumulative distribution of the total
intruder.
number of detecting SNs (equation 11) for both H0 and H1.
The simulations are performed using the MATLAB R2018b
software.
By dividing the ROI into clusters, it sometimes happens that
sensor nodes receive the signal emitted by an intruder located
in another cluster. We label this as "False Intrusion Detection"
(FID). This is generally due to the additive noise sensing to the
power of the signal. This phenomenon can mislead the CH by
signaling the existence of an intrusion into its space (Fig. 4.).

FIGURE 3. Dashboard tracking per cluster in time. Red clusters are the
event region tracked in time. Green clusters are the non-event region.

6
 FIGURE 4. False intrusion detection: A=60m, 𝜆 = 0.3, P0=50, 𝑃𝑓𝑎 = Dashed lines: Probability of true detection in the cluster intruder, according
0.01, d0 = 1, 𝜎𝑠2 = 1, an intruder position = (3, 2). to the variation of the parameters: 𝜆 and Pfa.
Solid lines: False detection in other clusters.
In what follows, we study the evolution of false intrusion
detection compared to true intrusion detection: we mean by (a) 4 clusters.
true intrusion detection, all the SNs detecting the intruder and
belong to the same cluster. From this perspective, we thought
of looking for the effect of changing the number of clusters. In
figure 5, the ROI is divided into 4 clusters, we set the
deployment density equal to 2. Figure 5 shows that the
probability of detection in the cluster concerned by the
intruder is greater than the probability in the remaining cluster.
In figures 6.a-b., ROI is divided into 4 and 9 clusters
respectively. We simulate the probability of true intrusion
detection (𝑃𝑇𝐷 ) in each cluster according to the probability of
the false alarm (0.01 ≤ 𝑃𝑓𝑎 ≤ 0.1) and the deployment
density (0.2≤ λ ≤ 5):
Number of true intrusion detecting SN
𝑃𝑇𝐷 = (9)
∑k
i=1 number of detecting SN in cluster i

Dashed lines: Probability of true detection in the cluster intruder, according

to the variation of the parameters: 𝜆 and Pfa.
Solid lines: False detection in other clusters.
(b) 9 clusters
FIGURE 6. a. 4 clusters, b. 9 clusters. Probability of positive detection
with A=60, P0=60, 𝜆 ∊[1, 5], P𝑓𝑎 ∊ [0.01, 0.05].

As well for a ROI divided into 4 clusters as another divided

into 9 clusters, we notice that "False intrusion detection"
always persists and takes place in almost all the clusters not
concerned by the intruder. On the other hand, we note that the
probability of the true intrusion detection of this event
(probability of all the SNs that claim to have detected the
intruder in a given cluster) is very low in all the areas where
the intruder is not present. Moreover, this probability remains
low even by varying both the density of the deployment and
the probability of the false alarm: the simulation values are
between [0.1, 0.3] and [0.05, 0.2] for 4 and 9 clusters
FIGURE 5. Probability of detection for TID Vs FID. System respectively. Which is attractive through the results that we
parameters: P𝑓𝑎 ∊ [0.01, 0.05], 𝜆 = 2, ROI divided into 4 clusters obtained, is that this probability decreases when the number of
clusters increases.
7
On the other hand, and whatever the value of the deployment
density as well as the value of the probability of false alarm, it Start
is observed that the probability of the true intrusion detection
is always greater than 0.55 for a ROI divided into 4 or 9
clusters respectively. Each SNi calculate local decision and
alert the correspondent CH if 𝑆(𝑥𝑖 ) > 𝜏
By analyzing the results that we obtained and supported by
figures 6.a-b, we see that it is more worth adding another
condition to equation 5. In other words, a CH receiving the CH calculate the 𝑃𝑇𝐷
local decision of the SNs verifying that 𝐼(𝑥𝑖 ) = True, must
also verify that the probability of the true intrusion detection
(𝑃𝑇𝐷 ) is greater than a certain threshold 𝛽.
False CH check
Let Φ = {x1, x2, …, xN} is the sensor nodes in a given cluster.
𝑃𝑇𝐷 > β
the set of sensors accepted by a CH (𝛷𝐶𝐻 ) is:
𝛷𝐶𝐻 = {𝑥𝑖 ∈ Φ / 𝑆(𝑥𝑖 ) > 𝜏 𝑎𝑛𝑑 𝑃𝑇𝐷 (𝑥𝑖 ) > β} (10) True
Send the identity of the
Thus, we can propose a second alternative of the detection CH to the FC
scenario: when an intrusion event has occurred, the policy of
a given CHj is to (i) count the total number of detecting SNs FC update the status of the
in the cluster j: we assume that there is only one intruder at a True region in the dashboard
time in a ROI, (ii) check if 𝑃𝑇𝐷 > β is true, to inform the FC
by sending its identity: this makes it possible to set update the CHi run the DCSNN Algorithm and
dashboard to track the ROI by zone as a first step (Fig. 3), (iii) sends results to Fusion Center (FC)
finally, if 𝑃𝑇𝐷 > β is true, the CHj runs the DCSNN to predict
the localization of the intruder and then notify the FC (Fig. 7).
Final decision in Fusion Center (FC)
We find that this strategy favors the respect of the constraints
of limited resources (energy and bandwidth) by (i) minimizing
unnecessary communications between the CH and the FC
(false alert if 𝑃𝑇𝐷 < β), (ii) executing the localization of the End
intruder only if the two conditions mentioned above are true.
FIGURE 7. Detection flowchart.

1. DISTRIBUTION OF DETECTING SENSOR NODES

 Under H1, the probability of the nodes retained by the
Since we have opted for the Poisson point process (PPP) to CH for the localization is:
model the distribution of the nodes of the WSN, and since the 𝑃(𝑠(𝑥𝑖 ) > 𝜏) = 𝑃𝑑 (𝑥𝑖 ) (14)
number of these detecting SNs represents the key factor of the
subject of this paper, then we study in the following their Thus the mean of the total number of detecting SNs is given
distribution under influent factors. The WSN is deployed in by:
60x60 ROI. The reference distance d0 is fixed to 1 unit. The 𝜆1 = 𝜆 ∫𝐶𝐻 𝑃𝑑 (𝑥) 𝑑𝑥 (15)
values of power P0 belong to the interval [0.1, 1]. Figure 8.a-b Figure 8.a-b and figure 9.a-b highlight that the simulation for
is the results of a Monte Carlo simulation with 105 runs. It the H0 and H1 assumptions and different situations fit the exact
shows a comparison between the simulated and the theoretical cumulative Poisson distribution perfectly.
cumulative distribution of the total number of detecting SNs The purpose of these two situations is to follow the
(Λ) under the test hypothesis defined in (12) and (14) as evolution of the cumulative distribution of Λ. In fact, Λ
follows. increases when the probability and/or the intensity of
Λ = ∑𝑥𝑖 ∈ 𝛷𝐶𝐻 𝐼(𝑥𝑖 ) (11) deployment increases. The localization simulations with
DCSNN take into account the variation of these parameters
Therefore, Λ follows a Poisson distribution: besides the value of the power P0.
 Under H0, the probability of the nodes retained by the
CH for the localization is:
𝑃(𝑠(𝑥𝑖 ) > 𝜏) = 𝑃𝑓𝑎 (12)
Like this value is a constant, so the intensity is equal to:
𝜆0 = 𝜆𝑃𝑓𝑎 |𝐴| (13)

8
 (b) 𝐻1: 𝑃𝑓𝑎 = 0.01 λ ∈ {2, 5, 8, 10}
FIGURE 9. a-b. Cumulative Distribution of Λ. (a-b) variation of the
(a) 𝐻0: 𝑃𝑓𝑎 = 0.07 λ = 8 total number of detecting SNs

V. DYNAMIC CELLS STRUCTURE NEURAL NETWORKS

(DCSNN)
In this section, we present the DCSNN algorithm and how it
works. Then, we prove our choice of this type of neural
network through a comparative study with MLPNN. All
simulations are performed using the MATLAB R2018b
software.
To locate the intruder, we will use artificial intelligence. More
precisely, we will take advantage of DCSNN to obtain a more
precise localization. The operation of DCSNN is essentially
based on how to balance and adjust the connection strength
and weight representations of neurons in a dynamic way,
during its progress respecting learning data. To do this, firstly
DCSNN calculates for each candidate neuron to be inserted
during each epoch of training, the distance of the two neurons
(b) 𝐻1: 𝑃𝑓𝑎 = 0.07 λ = 8 which are closest to it: they are named respectively by Best
FIGURE 8. a-b. Cumulative Distribution of Λ: ROI= 60x60, d0 = 1, Matching Unit (BMU) and Second Best Matching Unit
and x0 = (20, 20). (SBU). The Euclidean distance metric is adopted to find both
units. After that, two rules representing the core of the DCSNN
algorithm will be applied:
 The lateral connection strength of any two neurons is
administered by the Hebbian rule [35]. Perfect
strength (set to 1) is assigned either to the two
neurons in the initialization phase or to the new
neuron to be inserted and its BMU. Otherwise, the
lateral connection strength from the closest neuron is
degraded by a forgetting constant α as shown in
equation (16), [36]. An existing connection will no
longer exist if it drops below a predefined threshold
(θ). With regard to the other side connections, they
remain the same.
1, 𝑖 ∈ {𝐵𝑀𝑈, 𝑆𝑀𝑈}, 𝐽 ∈ {𝐵𝑀𝑈, 𝑆𝑀𝑈}
𝛼. 𝐶𝑖𝑗 , 𝛼. 𝐶𝑖𝑗 > 𝜃,
𝐶𝑖𝑗 = (16)
(a) 𝐻0 : λ = 8 𝑃𝑓𝑎 ∈ {0.01, 0.05, 0.1, 0.2} 0, 𝛼. 𝐶𝑖𝑗 < 𝜃,
{ 0, 𝑖 = 𝑗. }

9
  The representation of weights between active
neurons is governed by Kohonen's learning rule [37].
The modification of the weight vectors is derived by

∆𝑤𝐵𝑀𝑈 = 𝜀𝐵𝑀𝑈 (𝑣 − 𝑤𝐵𝑀𝑈 ) (17)

𝑤𝑁 = 𝜀𝑁 (𝑣 − 𝑤𝑗 ) (18)
where 𝜀𝐵𝑀𝑈 is the BMU weight adjustment, while the BMU
neighborhood weight adjustment is performed by 𝜀𝑁 . (b) Step 1
In the following, the DCSNN algorithm [27] is presented.

DCSNN Algorithm
Initialization;
While a stopping criterion is not satisfied do
{
Repeat N times
(c) Step 2
{
Compute the BMU and SBU;
Regulate lateral connections;
Update the weights;
Adjust resource values;
}
Insert a new neuron (if needed);
Decrement Resource Values;
} (d) Step 3

The strong point of the DCSNN is that it avoids all the

restrictions of the network’s topology while being based
on a self-organized structure, by combining the two
learning rules of Hebb and Kohonen. In the following, we
illustrate an example of some steps to show how DCSNN
works (Fig. 10). In figures 10.a-e the black disc represents
a target, blue points are the reference points whose
coordinates supply a learning database. Red points and (e) Final step of building the DCSNN
green edges are the DCS network. We use a table entitled FIGURE 10. An example showing the stages of building DCSNN.
Hebb-table to record the strength of the lateral
connections between all DCS neurons. We note that the Hebb-table in Figure 10.e resume the final lateral
initialization phase (Figure 10-a) consists in randomly connection strength between all the neurons of the
choosing two neurons among the nodes to be trained, then DCSNN.
defining the lateral connection as being a perfect force
between them, and finally starting the dynamic
construction of the network. 1. DCSNN Vs MLPNN
To prove our choice of DCS, we carried out a comparative
study on the training performance between DCSNN and
MLP neural networks whose activation function is the
hyperbolic tangent which makes it the most popular
network in various applications. To do this, we have
submitted its learning data about some mathematical
functions with an increasing order complexity:
(a) Initialization phase sin(x)*cos(x), sin(2x)∗cos(2x), and sin(3x)∗cos(3x). We
try through learning simulations to know how well these
two types of neural networks can adapt to the target data.

10
 MLPNN DCSNN DCSNN
Blue lines represent the target function. Red circles show the possible functions approximated.
Green lines are the connections between nodes in DCSNN.

FIGURE 11. DCSNN Vs MLPNN. Function: sin(x)cos(x).

MLPNN DCSNN
Blue lines represent the target function. Red circles show the possible functions approximated.
Green lines are the connections between nodes in DCSNN.

FIGURE 12. DCSNN Vs MLPNN. Function: sin(2x)cos(2x).

11
 MLPNN DCSNN
Blue lines represent the target function. Red circles show the possible functions approximated.
Green lines are the connections between nodes in DCSNN.

FIGURE 13. DCSNN Vs MLPNN. Function: sin(3x)cos(3x).

From Figure 11 concerning the training of the sin(x)cos(x) objective is to show the performance of the DCSNN in terms
function, we notice that MLP could not get close to the of localization accuracy and speed execution. We start by
coverage of the mathematical function (represented by the applying DCSNN to localize an intruder in a ROI divided into
blue lines). The look obtained by MLP (the red lines) reflects four clusters with fixed parameters of 𝑃𝑓𝑎 , λ, and 𝑃𝑜 . In the
a failure of this network during the training. On the DCS side next step, we conduct a comparative study related to the
and after 10 epochs of learning, we notice that the network DCSNN localization performance obtained when changing
reaches almost the same representation of the function (the red the values of the parameters 𝑃𝑓𝑎 , λ, and 𝑃𝑜 respectively for
circles are the neurons of the DCS, while the green lines an ROI considered as 1, 4, and 9 clusters. All simulations are
represent the connections between them). At 20 epochs, the performed using the MATLAB R2018b software.
DCS has perfectly learned the mathematical function
sin(x)cos(x). The same scenario is repeated for the functions In the following simulation, we place ourselves in the context
sin(2x)cos(2x) (Figure 12) and sin(3x)cos(3x) (Figure 13) for of a region of interest divided into 4 clusters with the presence
the MLP and DCS networks respectively. As the mathematical of an intruder (represented by a red star). Figure 14 gives a
function becomes more and more complex, the reaction of the general overview of the detecting sensor nodes: the blue
MLP to learning becomes more and more difficult even after circles are the sensor nodes deployed throughout the region,
a long time of learning. On the other hand, the results obtained the red circles are the sensor nodes belonging to clusters
on the part of the DCS describe again a perfect behavior to (Cluster 2, cluster 3, and Cluster 4) not containing the intruder
adapt to the exact coverage of the mathematical functions in a and whose their received signal strengths are greater than the
minimal number of epochs of 40 and 60 respectively. DCS threshold 𝑆(𝑥𝑖 ) ≥ 𝜏 (False Intrusion Detection). The blue
has shown its merit by winning the comparison, which in turn disks with a red perimeter are the detecting sensor nodes
translates on the side of the cluster head into energy-saving belonging to the same cluster (Cluster 1) like the intruder and
and short execution time of the DCS algorithm and
whose received signal strengths are greater than the threshold
consequently into the speed of localization.
𝑆(𝑥𝑖 ) ≥ 𝜏 (True Detection Intrusion). Thus, only these latter
detecting sensor nodes are considered for the localization of
VI. LOCALIZATION
the intruder via the DCS network (Fig. 15).
In this section, we will begin the localization phase: we are at
step 4 of our communication model (Fig. 2) where the cluster
head decides to run the DCS algorithm. Basically, the

12
 FIGURE 14. Intruder detection in a ROI divided into clusters. Red star: intruder, blue circles: sensor nodes, blue disks in Cluster 2,3,4: False intrusion detection,
and blue disks with red perimeters (Cluster 1): True intrusion detection

FIGURE 15. Intruder localization with DCS network. Red star: intruder, blue disks: detecting sensor nodes, green circles: predicted locations with DCS
network.

In Fig. 15, simulation was carried out according to the what is the impact of dividing the ROI into a certain number
following parameter values 𝑃𝑓𝑎 = 0.05, 𝜆 = 0.5, and 𝑃𝑜 = of clusters? For this reason, we carried out a number of
0.8. There is a real adaptation of the DCS network which simulations, through which we tried to change the values of
results in the occupation of its nodes the same detecting sensor the 0.05 ≤ 𝑃𝑓𝑎 ≤ 0.09, 0.2 ≤ 𝑃𝑜 ≤ 1, and 0.5 ≤ λ ≤ 0.9. In
nodes coordinates. The extract from figure 15 under the addition, the same simulations were carried out on the region
caption "Quick adaptation" shows how the DCS network of interest considered respectively as a single cluster, 4
evolves dynamically during its learning by adding, into the clusters, and 9 clusters. Our main concern through the results
network and at each epoch, a neuron which converges more obtained was mainly based on the accuracy of the localization
and more towards the position of the detection nodes until the via the mean error values, and the time taken to achieve that.
total adaptation. This generates a very accurate localization
Figure 16 below, illustrates a comparative study of
(results show a mean error = 1.538m). In addition, this localization performance under the mean error criterion. The
adaptation is achieved above all in a minimal time (results objective is to deduce if the increase of the number of clusters
show an elapsed time = 0.569s). But what is the effect of helps to have a better precision of localization or not, without
changing the values of these parameters on the accuracy of the forgetting the impact of the variation of the deployment
localization? On the other hand, and for the same purpose, density which, obviously, reflects the number of sensor nodes.
13
 ROI considered as 1 Cluster, P0= 0.2 ROI divided into 4 Clusters, P0= 0.2 ROI divided into 9 Clusters, P0= 0.2

ROI considered as 1 Cluster, P0= 0.4 ROI divided into 4 Clusters, P0= 0.4 ROI divided into 9 Clusters, P0= 0.4

ROI considered as 1 Cluster, P0= 0.6 ROI divided into 4 Clusters, P0= 0.6 ROI divided into 9 Clusters, P0= 0.6

ROI considered as 1 Cluster, P0= 0.8 ROI divided into 4 Clusters, P0= 0.8 ROI divided into 9 Clusters, P0= 0.8

ROI considered as 1 Cluster, P0=1 ROI divided into 4 Clusters, P0=1 ROI divided into 9 Clusters, P0=1
FIGURE 16. Mean error of localization using DCSNN. System parameters are d0 = 1, x0 = (20, 20). ROI= 60x60. Each dashed line describe the mean
error of localization according to the values of 𝑃𝑓𝑎 ∈ [0.05, 0.09] and λ [0.5, 0.9].

14
 Each sub-figure is composed of five dashed lines. Each one λ=0.6, Pfa =0.07), (P0 = 0.8, λ = 0.7, Pfa =0.06), (P0 = 0.4, λ =
shows how the mean error of intruder localization varies as a 0.5, Pfa =0.07). knowing that the highest density value in the
function of the variation of the Pfa value for fixed values of P0 simulations is 0.9, we can deduce that the fact of thinking to
and deployment density, according to some cluster increase the number of sensors to be deployed per unit of an
distribution of the ROI. The results collected show that the area does not automatically mean to have a better localization
mean error range varies as follows: 4m <= error <= 5m, accuracy. In this context, we fully agree with the authors N.
1.6m<= error <= 2.3m, and 0.9m <= error <= 1.45m for a Assad, et al. in [38]. They show that the optimal number of
region of interest considered as 1 cluster, 4 clusters, and 9 nodes that can be deployed to efficiently cover a region of
clusters respectively. This leads to deducing that, the more the interest depends mainly on the value of the sensing range, and
region is divided into clusters, the smaller the localization vice versa. As for the value of Pfa, it is generally desirable to
error becomes. However, this precision can be further choose a fairly high value, especially for critical applications,
improved if the deployment density and the transmission (having a detection alert despite a high probability of false
power of the sensor nodes P0 are suitably chosen. alarms implies that the alert is certain). The best localization
Indeed, one notices that there exist several occurrences of accuracies obtained by the simulations are for Pfa values
minimum value of error localization (around 0.9m) for a between 0.06 and 0.07.
density of deployment close to 0.6. As examples, we observe
the following for the ROI divided into 9 clusters: (P0 = 1,

ROI considered as 1 Cluster, P0= 0.2 ROI divided into 4 Clusters, P0= 0.2 ROI divided into 9 Clusters, P0= 0.2

ROI considered as 1 Cluster, P0= 0.4 ROI divided into 4 Clusters, P0= 0.4 ROI divided into 9 Clusters, P0= 0.4

ROI considered as 1 Cluster, P0= 0.6 ROI divided into 4 Clusters, P0= 0.6 ROI divided into 9 Clusters, P0= 0.6

15
 ROI considered as 1 Cluster, P0= 0.8 ROI divided into 4 Clusters, P0= 0.8 ROI divided into 9 Clusters, P0= 0.8

ROI considered as 1 Cluster, P0=1 ROI divided into 4 Clusters, P0= 1 ROI divided into 9 Clusters, P0=1
FIGURE 17. Elapsed time for localization using DCSNN. System parameters are d0 = 1, x0 = (20, 20). ROI= 60x60. Each dashed line describe the
elapsed time of localization according to the values of 𝑃𝑓𝑎 ∈ [0.05, 0.09], and λ [0.5, 0.9].

The results collected show that the elapsed time range, for The last mission is performed on the fusion center which will
localization using DCS network, varies as follows: 1.9s <= collect from the CHi, the positions of the intruders and
time<= 4s, 0.7s <= time<= 1.6s, and 0.3s <= time<= 0.9s for therefore order them chronologically to track the movement of
a region of interest considered as 1 cluster, 4 clusters, and 9 intruders over time. The results we have obtained confirm
clusters respectively (Fig. 17). This confirms the results found once again the effectiveness of the DCSNN algorithm to
previously (Fig. 16). We note that the division of the region achieve the localization with the highest possible accuracy and
into a large number of clusters reinforces the idea of rapid in the shortest time. This is the main motivation that made us
detection. Indeed, the best performance of the localization think of running the DCSNN at the level of each HC, and then
time returns to the ROI divided into 9 clusters. As examples, sending the results to the FC.
we observe an elapsed time around 0.3s for the following
VII. DISCUSSION AND CONCLUSION
parameters: (P0 = 0.8, λ = 0.7, Pfa = 0.05), (P0 = 0.6, λ = 0.65,
Pfa = 0.07), (P0 = 0.4, λ = 0.55, Pfa = 0.07).
1. DISCUSSION
What is interesting, is the important difference that we gain in In this study, we presented a new scenario relating to the
terms of intruder’s localization accuracy as well as the speed detection of an intruder in a clustered wsn. This is translated
of localization. in the communication model section. During the presentation
of all steps of this work, the communication complexity was
considered as one of the most important metrics for measuring
the performance of a wsn. The main focus of this manuscript
is not the energy study, that's why we have not provided
numerical results in this regard. However, the communication
strategy adopted, between the sensor nodes and the CH, in
first-order then between the CH and the FC, in second-order,
saves the energy of the wsn and is in total agreement with the
literature: Choose a direct communication between the CHs
and the FC (and not make the CHs communicate with each
FIGURE 18. Performance comparison between 1 cluster, 4 cluster, and 9
clusters. other) to determine which CH has the maximum number of
sensors that have detected an intruder. According to [24]
The more we go from a region divided into a number of Single-hop, inter-cluster communication has good
clusters to another divided into a larger number of clusters, the performance over multi-hop inter-cluster communication in
more interesting the results will be (Fig. 18). terms of energy depletion, while multi-hop inter-cluster
16
communication has a better performance compared to single- uniform and non-uniform clustering). In particular, they show
hop inter-cluster communication in terms of network that the detection performance for a false alarm probability of
scalability. In addition, it is in the interest of having the CHs 10-2, is improved by 34.82% for a number of clusters N = 8
communicate directly with the FC so that the latter (i) ensure and a number of sensors node = 125 and by 67.23% for a
the election of the CH subject to intrusion, (ii) inform the CH number of clusters N = 100 and a number of sensors node =
elected by a unicast communication, (iii) keep an intrusion 10, compared to the results obtained during the detection with
history useful for tracking, and (iv) update the event-region a number of clusters N = 4 and a number of sensors node =
dashboard. Which will be expensive in terms of energy if we 250. These results come to further strengthen our current
limit ourselves to inter-CHs communication. work. Having an optimal localization in a large-scale region of
From an energy consumption point of view, Q. Wang et al. interest does not automatically mean thinking about deploying
[21] propose an analytical model to give the optimal number a large number of sensor nodes.
of clusters in wireless sensors networks while taking into Despite all the encouraging results that we get through most
account reaching the minimum value of the total energy of the research on the optimal number of clusters for a WSN,
consumption. most of this research has remained separate from each other so
On the localization side which is the core of our manuscript, that each one looks from a different angle than the other, be it
we focus on the cluster head which can perform the in terms of energy efficiency and prolonging network lifetime,
localization phase using DCSNN. The conjunction between localization, Secure Data Communication, Latency
this type of neural network and the WSN has given good Reduction, Efficient Quality of Services, Data
results in terms of performance (average error and execution Communication Assurance, high scalability, etc.
time). Indeed, in [39] the authors confirm the importance of This invites us to move forward to find a generalization that
DCSNN when they deployed it in the NASA F-15 flight balances the maximum possible of the different points of view.
control system case study, demonstrating that their techniques
effectively detect unusual events and provide validation 2. CONCLUSION
inferences in real-time. On the other hand, according to our In this paper, we studied the detection of border intrusion
simulations, the obtained localization results showed that in events in clustered wireless sensor networks. We described the
order to have better localization performance, we should sensor network model and defined how to compute the local
consider increasing the number of clusters. Therefore, we tried decision of each SN considering additive white Gaussian
to prove it by applying our simulations according to a ROI noise. Then, we detailed a new event detection scenario of the
considered as 1, 4 and 9 clusters. As shown in Figure 16, for communication model. Based on the "False intrusion
the following parameter values 0.05<= Pfa <=0.09, 0.5 <= λ detection" study, we refined the condition mentioned in
<= 0.9, and 0.2 <= P0 <= 1, increasing the number of clusters Equation 5 by adding another criterion that requires any
improves the detection performance of the DCS in terms of cluster with sensor nodes detecting an intrusion event to
localization accuracy by 56.67% and 73.88% for a number of additionally verify that its probability relative to true intrusion
clusters equal to 4 and 9, respectively, compared to a ROI detection (PTD) is greater than a certain threshold 𝛽. Once
considered as 1 cluster. Similarly, according to our simulations justified, and after being elected by the FC, the CH executes
(Figure 17), and for the same parameter values considered the dynamic cell structure network algorithm to locate the
above in Figure 16, increasing the number of clusters intrusion with high accuracy and minimal time. In this
improves the detection performance in terms of elapsed time framework, we introduced the DCSNN algorithm and its
by the DCS algorithm for localization by 61.01% and 79.66% operation. Then, we proved our choice for this type of neural
for a number of clusters equal to 4 and 9, respectively, network through a comparative study with the MLPNN. The
compared to a ROI considered as 1 cluster. But we think that DCS has a great superiority in adaptability and performance.
the real justice will only know if we evaluate the overall time On the other hand, we proceeded to the comparison of the
of the whole process (detection, classification and localization performances when the ROI is divided into
localization). In [40], Arora et al. define a summary of clusters. We find that the accuracy and elapsed time for
performance requirements: a Probability of detection PD localization improves more and more when the number of
>0.95, a Probability of false alarm PFA <0.10, a Latency of clusters increases. Contrary to those who think of deploying a
detection (s) TD <15, Position estimation error (m) ± (2.5, large number of sensor nodes. However, we hope to continue
2.5). our research by including the kinematic characteristics of the
On the other hand, Eritmen et al. investigate in [41] a intruders (acceleration, speed, and bearing already studied in
distributed decision fusion in hierarchical wireless sensor our paper [2]) as constraints in the localization process and
networks that are degraded by fading and noise. Numerical mainly considering some important parameters such as energy
results obtained through simulations show that the detection efficiency, secure data communication, quality of service, etc.
probability is sensitive for various parameters such as false
alarm probabilities, SNR regimes, number of clusters, and/or ACRONYMS
number of sensors per cluster and clustering types (i.e., The following acronyms are used in this paper:
17
AWGN Additive White Gaussian Noise [14] R.L. Streit. (2010) "The Poisson Point Process". In: Poisson Point
Processes. Springer, Boston, MA. https://doi.org/10.1007/978-1-
BMU Best Matching Unit 4419-6923-1_2
CH Cluster Head [15] Ramzi Bellazreg, Noureddine Boudriga, and Sunshin An , "Border
DCSNN Dynamic Cell Structures Neural Networks Surveillance using sensor based thick-lines," ICOIN 2013, 978-1-
4673-5742-5/13/$31.00 ©2013 IEEE.
FC Fusion Center [16] A. A. Abbasi, M. Younis, "A survey on clustering algorithms for
MLP Multi-Layer Perceptron wireless sensor networks," Computer Communications, Volume 30,
PPP Poisson Point Process Issues 14–15, 2007, Pages 2826-2841, ISSN 0140-3664,
SBU Second Best Matching Unit https://doi.org/10.1016/j.comcom.2007.05.024.
[17] N. Mazumdar, S. Roy and S. Nayak, "A Survey on Clustering
SN Sensor Nodes Approaches for Wireless Sensor Networks," 2018 2nd International
SNR Signal-to-Noise Ratio Conference on Data Science and Business Analytics (ICDSBA), pp.
ROI Region of Interest 236-240, 2018, doi: 10.1109/ICDSBA.2018.00049.
[18] K. Pradeepa, W. Regis Anne, S. Duraisamy, "Design and
WSN Wireless Sensor Networks Implementation Issues of Clustering in Wireless Sensor Networks,"
International Journal of Computer Applications (0975 – 888) Volume
REFERENCES 47– No.11, June 2012.
[19] A. Shahraki, A. Taherkordi, Ø. Haugen, and F. Eliassen, "Clustering
[1] P. Sikora, L. Malina, M. Kiac, Z. Martinasek, et al. "Artificial objectives in wireless sensor networks: A survey and research
Intelligence-based Surveillance System for Railway Crossing direction analysis," Computer Networks, Volume 180, 2020, 107376,
Traffic," in IEEE Sensors Journal, DOI: ISSN 1389-1286, https://doi.org/10.1016/j.comnet.2020.107376.
10.1109/JSEN.2020.3031861. [20] T. Damarla, and A. Mehmood, "Detection of targets using distributed
[2] M. Othmani, M. H. Jeridi, Q. G. Wang, and T. Ezzedine, "Border multi-modal sensors with correlated observations," SENSORS, 2013
Trespasser Classification Using Artificial Intelligence," in IEEE IEEE, 2013, pp. 1-4, DOI: 10.1109/ICSENS.2013.6688471.
Access, vol. 9, pp. 72284-72298, 2021, DOI: [21] Q. Cheng, R. Niu, A. Sundaresan, and Pramod K. Varshney,
10.1109/ACCESS.2021.3079702. "Distributed Detection and Decision Fusion with Applications to
[3] A. Atmaja, A. Hakim, A. Wibowo, and L. Pratama, "Communication Wireless Sensor Networks," In Integrated Tracking, Classification,
Systems of Smart Agriculture Based on Wireless Sensor Networks in and Sensor Management (eds M. Mallick, V. Krishnamurthy and B.‐
IoT," Journal of Robotics and Control (JRC), 2(4), 297-301. N. Vo). (2016). https://doi.org/10.1002/9781118450550.ch16.
doi:https://doi.org/10.18196/jrc.2495. [22] P. Singh, S. Agrawal.: "TDOA Based Node Localization using
[4] T. J. S. Chowdhury, C. Elkin, V. Devabhaktuni, D.B. Rawat, J. Neural Networks". International Conference on Communication
Oluoch, "Advances on localization techniques for wireless sensor Systems and Network Technologies". 2013.
networks: A survey," Computer Networks, (2016). 110, 284–305. [23] Nazish Irfan, Miodrag Bolic, Mustapha C.E. Yagoub.: "Neural-based
DOI: 10.1016/j.comnet.2016.10.006. approach for localization of sensors in indoor". Telecommun Syst
[5] D. Jinze, D. Jean-François, W. Yide, "An RSSI-based parameter (2010) 44: 149–158.
tracking strategy for constrained position localization," EURASIP [24] A. Zeb, A. Islam, M. Zareei, et al., "Clustering Analysis in Wireless
Journal on Advances in Signal Processing. 2017, 77. DOI: Sensor Networks: The Ambit of Performance Metrics and Schemes
10.1186/s13634-017-0512-x. Taxonomy," International Journal of Distributed Sensor Networks.
[6] S. Zhu, and Z. Ding, "Joint synchronization and localization using July 2016. doi:10.1177/155014774979142.
TOAs: A linearization based WLS solution," IEEE J Select Areas [25] R. Niu, and P.K., Varshney, "Target location estimation in sensor
Commun. (2010). Vol. 28 Issue: 7 Month: Sep. DOI: networks with quantized data," IEEE Transactions on Signal
10.1109/JSAC.2010.100906. Processing, 54(12), pp.4519-4528, 2006.
[7] P. Wu, S. Su, Z. Zuo et al., "Time Difference of Arrival (TDoA) [26] M. Martalo, and G. Ferrari, "A simple information-theoretic analysis
Localization Combining Weighted Least Squares and Firefly of clustered sensor networks with decentralized detection," in IEEE
Algorithm," in Sensors, (2019), 19, 2554; DOI:10.3390/s19112554. Communications Letters, vol. 14, no. 6, pp. 560-562, June 2010, doi:
[8] H. J. Shao, X. P. Zhang, and Z. Wang, "Efficient closed-form 10.1109/LCOMM.2010.06.092286.
algorithms for AOA based self-localization of sensor nodes using [27] L. Jin-Gu, C. Seyha, and P. Ho-Hyun, "Energy-Efficient Cluster-
auxiliary variables," IEEE Transactions on Signal Processing. Head Selection for Wireless Sensor Networks Using Sampling-Based
(2014). Vol. 62, No 10, 2580–2594. DOI: 10.1109/ Spider Monkey Optimization". Sensors. 2019; 19(23):5281.
TSP.2014.2314064. https://doi.org/10.3390/s19235281.
[9] A. Mesmoudi, M. Feham, and N. Labraoui, "Wireless Sensor [28] L. Zhao, S. Qu, and Y. Yi, "A modified cluster-head selection
Networks Localization Algorithms: A Comprehensive Survey,” in algorithm in wireless sensor networks based on LEACH". J Wireless
International Journal of Computer Networks & Communications Com Network 2018, 287 (2018). https://doi.org/10.1186/s13638-
(IJCNC), (2013), Vol.5, No.6. DOI: 10.5121/ijcnc.2013.5603. 018-1299-7.
[10] J. Mass-Sanchez1, E. Ruiz-Ibarra1, J. Cortez-Gonza´lez, "Weighted [29] I.S. Akila, S.V. Manisekaran, and R. Venkatesan (October 4th 2017).
Hyperbolic DV-Hop Positioning Node Localization Algorithm in "Modern Clustering Techniques in Wireless Sensor Networks,
WSNs," in Wireless Pers Commun, (2016), DOI: 10.1007/s11277- Wireless Sensor Networks - Insights and Innovations, Philip Sallis,
016-3727-5. IntechOpen, DOI: 10.5772/intechopen.70382. Available from:
[11] A. Chatterjee, "A Fletcher-Reeves Conjugate Gradient Neural- https://www.intechopen.com/chapters/56752.
Network-Based Localization Algorithm for Wireless Sensor [30] R.Tandon, "Determination Of Optimal Number Of Clusters In
Networks," IEEE Trans Veh Technol (2010). Volume: 59 Issue: 2. Wireless Sensor Networks," International Journal of Computer
First page: 823. Month: Feb. DOI: 10.1109/TVT.2009.2035132. Networks & Communications. 2012, 4. 10.5121/ijcnc.2012.4415.
[12] V. Feng, and S. Y. Chang, "Determination of Wireless Networks [31] A. S. Raghuvanshi, S. Tiwari, R. Tripathi, and N. Kishor, "Optimal
Parameters Through Parallel Hierarchical Support Vector Machines," number of clusters in wireless sensor networks: An FCM
IEEE Trans Parallel Distrib Syst. (2012). Volume: 23 Issue: 3. First approach," 2010 International Conference on Computer and
page: 505. Month: Mar. DOI: 10.1109/TPDS.2011.156. Communication Technology (ICCCT), 2010, pp. 817-823, doi:
[13] J. Lee, W. Chung, and E. Kim, "A new kernelized approach to 10.1109/ICCCT.2010.5640391.
wireless sensor network localization," Information Sciences. 2013. [32] Q. Wang, S. Guo, J. Hu, and Y. Yang, "Spectral partitioning and
Vol. 243 No. 10, pp. 20–38. DOI: 10.1016/j.ins.2013.04.024. fuzzy C-means based clustering algorithm for big data wireless
sensor networks". EURASIP Journal onWireless Communications

18
 and Networking (2018) 2018:54. https://doi.org/10.1186/s13638-
018-1067-8.
[33] X. Song, P. Willett, J. Glaz, and S. Zhou, "Active Detection With a
Barrier Sensor Network Using a Scan Statistic," IEEE Journal of
Oceanic Engineering, Vol. 37, No. 1, January, 2012.
[34] M Guerriero, P Willett, and J Glaz, "Distributed target detection in
sensor networks using scan statistics," IEEE Trans Signal Process.
2009. Vol. 57 No. 7. Month: Jul., 2629–2639.
[35] K. Ishthaq Ahamed, and A. Shaheda, "Survey on Artificial Neural
Network Learning Technique Algorithms," International Research
Journal of Engineering and Technology (IRJET). 2016. Vol. 03 Issue:
02. E-ISSN: 2395-0056. P-ISSN: 2395-0072.
[36] J. Bruske, and G. Sommer, "Dynamic Cell Structures," In: Neural
Information Processing Systems (NIPS1995). 1995, Vol. 7. pp. 497-
504.
[37] S. Hiregoudar, K. Manjunath, K.S. Patil, "Research summary on
neural networks," IJRET: International Journal of Research in
Engineering and Technology. 2014. Vol. 03 Special Issue: 03. eISSN:
2319-1163. pISSN: 2321-7308.
[38] Assad N., Elbhiri B., El Fkihi S., Faqihi M.A., Ouadou M.,
Aboutajdine D. (2014) Short: Intrusion Detection Quality Analysis
for Homogeneous Wireless Sensor Networks. In: Noubir G., Raynal
M. (eds) Networked Systems. NETYS 2014. Lecture Notes in
Computer Science, vol 8593. Springer, Cham.
https://doi.org/10.1007/978-3-319-09581-3_24.
[39] L. Yan, C. Bojan, and G. Srikanth, "Validating neural network-based
online adaptive systems: a case study," Springer Science, Business
Media, LLC ().2007. Issue: 3. Month: Sep. DOI: 10.1007/s11219-
007-9017-4, pp. 309 –326.
[40] A. Arora, P. Dutta, S. Bapat, V. Kulathumani, H. Zhang, V. Naik, V.
Mittal, H. Cao, M. Demirbas, M. Gouda et al., "A line in the sand: a
wireless sensor network for target detection, classification, and
tracking," Computer Networks. Vol. 46, No. 5, 605-634, 2004.
[41] Eritmen, Kayhan & Keskinoz, Mehmet. (2013). Distributed decision
fusion over fading channels in hierarchical wireless sensor networks.
Wireless Networks. 20. 987-1002. 10.1007/s11276-013-0649-y.

19