Information Fusion 49 (2019) 1–25

Contents lists available at ScienceDirect

Information Fusion
journal homepage: www.elsevier.com/locate/inffus

Machine learning algorithms for wireless sensor networks: A survey

D. Praveen Kumar, Tarachand Amgoth∗, Chandra Sekhara Rao Annavarapu
Department of Computer Science and Engineering, Indian Institute of Technology (Indian School of Mines), Dhanbad, India

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

Keywords: Wireless sensor network (WSN) is one of the most promising technologies for some real-time applications be-
Wireless sensor networks cause of its size, cost-effective and easily deployable nature. Due to some external or internal factors, WSN may
Machine learning change dynamically and therefore it requires depreciating dispensable redesign of the network. The traditional
Energy efficiency
WSN approaches have been explicitly programmed which make the networks hard to respond dynamically. To
Network lifetime
overcome such scenarios, machine learning (ML) techniques can be applied to react accordingly. ML is the pro-
Data aggregation
cess of self-learning from the experiences and acts without human intervention or re-program. The survey of the
ML techniques for WSNs is presented in [1], covering period of 2002–2013. In this survey, we present various
ML-based algorithms for WSNs with their advantages, drawbacks, and parameters effecting the network lifetime,
covering the period from 2014–March 2018. In addition, we also discuss ML algorithms for synchronization,
congestion control, mobile sink scheduling and energy harvesting. Finally, we present a statistical analysis of the
survey, the reasons for selection of a particular ML techniques to address an issue in WSNs followed by some
discussion on the open issues.

1. Introduction chitecture that can learn to improve its performance. Because of the in-
terdisciplinary nature, it plays a pivotal role in various fields including
Wireless sensor network (WSN) is one of the most promising tech- engineering, medical, and computing. Recent advances in ML have been
nologies for some real-time applications because of its size, cost-effective applied to solve various issues in WSNs [1]. Applying ML not only im-
and easily deployable nature [2]. The job of WSN is to monitor a field proves the performance of WSNs and also limits the human intervention
of interest and gather certain information and transmit them to the base or re-program. Access vast amount of data collected by the sensors, and
station for post data analysis [3,4]. Some of the WSN applications con- extract the useful information from the data is not so easy without ML. It
sists of a large number of sensor nodes. Therefore managing such a large also uses to integrating Internet of things (IoT), cyber-physical systems
number of nodes requires a scalable and efficient algorithms. In addi- (CPS) and machine to machine (M2M) [1]. Some of the applications of
tion, due to the external causes or intended by the system designers, the ML in WSNs are:
WSNs may change dynamically. Therefore it may affect network rout-
ing strategies, localization, delay, cross-layer design [5], coverage, QoS,
link quality, fault detection, etc. [6]. Because of the highly dynamic na- - For target area coverage problem, deciding an optimal number of
ture, it may require depreciating dispensable redesign of the network, sensor nodes to cover the area is easily obtained by ML techniques.
but the traditional approaches for the WSNs are explicitly programmed, - Energy-harvesting provides a self-powered and long lasting mainte-
and as a result, the network does not work properly for the dynamic nance for the WSNs deployed in the harsh environment. ML algo-
environment. rithm improves the performance of WSNs to forecast the amount of
Machine Learning (ML) is the process that automatically improves energy to be harvested within a particular time slot.
or learns from the study or experience, and acts without being explic- - Sensor nodes may change their location due to some internal or ex-
itly programmed [7–9]. ML was making our computing processes more ternal factors. Accurate localization is smooth and rapid with the
efficient, reliable and cost-effective. ML produce models by analyzing help of ML algorithms.
even more complex data automatically, quickly and more accurately. - ML used to segregate the faulty sensor nodes from normal sensor
It is mainly classified into supervised learning, unsupervised learning, nodes and improve the efficiency of the network.
semi-supervised learning and reinforcement learning. The strength of - Routing data place a major role in improving the network lifetime.
ML lies in their ability to provide generalized solutions through an ar- The dynamic behavior of sensor network requires dynamic routing
mechanisms to enhance the system performance.

∗
Corresponding author.
E-mail address: tarachand@iitism.ac.in (T. Amgoth).

https://doi.org/10.1016/j.inffus.2018.09.013
Received 17 April 2018; Received in revised form 17 September 2018; Accepted 20 September 2018
Available online 21 September 2018
1566-2535/© 2018 Elsevier B.V. All rights reserved.
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Table 1
Abbreviations.

AD Anomaly Detection ABC Artificial bee colony

ACO Ant colony optimization ANN Artificial Neural Networks
ANOVA Analysis of variance AOGE Angle optimized global embedding
BCS Bayesian compressive sensing BSN Body sensor network
BTMS Bayesian-based trust management strategy CHs Cluster heads
CNN Conventional Neural Network CPS Cyber physical systems
DACR Distributed adoptive cooperative routing DBSCAN Density-based Spatial Clustering of Applications with Noise
DFRTP Dynamic 3D fuzzy routing-based on traffic probability DFTDT Distributed functional tangent decision tree
DL Deep learning DLRDG Distributed linear regression-based data gathering
DoS Denial-of-Service DT Decision trees
FCM Fuzzy c-means FCMTSR Fuzzy C-means training sample reduction
FDS Fault detection scheme FIS Fuzzy information system
FLI Fuzzy location indicator GPS Geographical position system
HMM Hidden Markov Model ICA Independent component analysis
IDS Intrusion detection system IoT Internet of Things
LEACH Low Energy Adaptive Clustering Hierarchy LQI Link quality indicator
MAC Medium Access Controller MDBN Multi-layer dynamic Bayesian network
ML Machine Learning MST Minimum spanning tree
NREL National renewable energy laboratory ODT Optimal deadline-based trajectory
PCA Principle Component Analysis PCRB Posterior CramerRao bound
PDR Pocket delivery ratio PPM Parts Per Million
PSO Particle Swarm Optimization QoS Quality of Service
RBF Radial Basis Function RF Random forests
RL Reinforcement learning RMSE Root mean square error
RPs Rendezvous points RSS Received Signal Strength
RSSI Received Signal Strength Indicator SHM Structural health monitoring
SLT Shallow light tree SMC Sequential Monte Carlo
SMO Sequential minimal optimization SVD Singular Value Decomposition
SVDD Support Vector Data Description SVM Support Vector Machine
THMSO Two-hop mass spring optimization VBEM Variational Bayesian expectation maximization
VOC Volatile organic compound WEH Wireless energy harvesting
WH Wormhole WNN Wavelet neural networks
WSNs Wireless Sensor Networks ZEEP Zone-based energy efficient routing

- Transmitting the entire data to the base station will lead to transmis- 2.1. Supervised learning
sion overhead in the network. ML also helps to reduce the dimen-
sionality of the data at the sensor or cluster head level. Supervised learning is one of the most important data processing ap-
proaches in ML. In supervised learning, we provide a set of input and
In [1], a detailed survey of the ML for WSNs has been covered for
outputs (datasets with labels), and it finds the relation between them
the period 2002–2013 in which various issues were discussed. The cov-
while training the system. At the end of the training process, we can find
ered issues are localization, object tracking, routing, clustering and data
a function from an input x with a best estimation of output y (f: x → y). A
aggregation, event detection, query processing, and MAC protocols as
major responsibilities of supervised learning algorithms are to generate
a functional challenge and whereas security, anomaly detection, fault
the model which represents relationships and dependency links between
node detection, and QoS as non-functional challenges. We present a sur-
input features and forecast objective outputs. Supervised learning solve
vey on various ML-based algorithms for WSNs with their advantages,
various challenges in WSNs such as localization [10–25], coverage prob-
drawbacks, and parameters effecting the network lifetime, covering the
lems [26–31], anomaly and fault detection [32–45], routing [46–53],
period from 2014–March 2018. In this survey, we also discuss ML algo-
MAC [54], data aggregation [55–67], synchronization[68–71], conges-
rithms for synchronization, congestion control, mobile sink scheduling
tion control [72–74], target tracking [75–78], event detection [79–81],
and energy harvesting for WSNs. Finally, we present a statistical analy-
and energy harvesting [82,83]. Supervised learning categorized into re-
sis of the survey, the reasons for selection of ML to solve various issues
gression and classification. Classification can be divided into logic-based
in WSNs, and some of the open issues to be addressed by ML techniques
(decision tree and random forest), perceptron based (ANN and deep
for WSNs.
learning), statistical learning (Bayesian and SVM) and instance-based
The rest of the paper organized as follows. For the readers conve-
(k-NN) algorithms.
nient, abbreviations used in this paper are listed in Table 1. The nec-
essary background of the ML approaches discussed in Section 2. In
Section 3, a brief survey of various applications of ML in WSNs issues 2.1.1. Regression
are discussed. In Section 4, statistical analysis and limitations are pre- Regression is a supervised learning method, and it will predict some
sented. In Section 5, we present open issues for ML-based algorithms for value (Y) based on a given set of features (X). The variables in the re-
WSNs. Finally, in Section 6, we conclude the paper. gression model are continuous or quantitative. Regression is very simple
ML approach and predicts accurate results with minimum errors. The
2. Machine learning techniques mathematical notation for linear regression [84] is shown in Eq. (1).

𝑌 = 𝑓 (𝑥) + 𝜀 (1)
In this section, we explore various ML techniques and their learn-
ing procedures that will helps to understand the later sections. We where Y is the dependent variable (output), x indicates independent
also given a brief description of evolutionary computing techniques for variable (input), f is a function that it makes the relation between x and
WSNs. Based on the learning styles, ML techniques have been catego- Y, and 𝜀 represents the possible random error. The working model of a
rized into supervised learning, unsupervised learning, semi-supervised simple linear regression is shown in Fig. 2. Regression is applied to solve
learning and reinforcement learning. Fig. 1 shows the taxonomy of ML various issues in WSNs such as localization [85], connectivity problem
techniques. [26,27], data aggregation [55–57], and energy harvesting [82,83].

2
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Fig. 1. Taxonomy of ML techniques.

Fig. 4. A simple ANN architecture with different layers.

Fig. 2. The simple linear regression model [86]. tection [33], data aggregation [59,60], and mobile sink path selection
[88].

2.1.3. Random forest

Random forest (RF) algorithm is a supervised ML technique with a
collection of trees and each tree in the forest gives a classification. RF
algorithm works in two stages, creation of random forest classifier and
prediction of results [89]. RF works efficiently for the larger datasets
and heterogeneous data. This approach accurately predicts the missing
values. The impact of randomly selecting a subset of training samples
and isolating variables at each tree node will produce a large number of
decision trees. Therefore, the sensitivity level of RF classifier is less with
appraising to other streamline ML classifiers because of the quality of
training samples and to over robust decision trees. Existing classification
methodologies are facing significant challenges due to a curse of dimen-
sionality and highly correlated data. RF classifier will be the best appro-
Fig. 3. Graphical representation of a decision tree [87]. priate method for classifying hyperspectral data [90]. RF algorithm has
been applied to solve various issues in WSNs such as coverage [30] and
MAC protocol [54].
2.1.2. Decision trees
Decision trees (DT) are a class of supervised ML approach for clas- 2.1.4. Artificial neural networks
sification based on a set of if-then rules to enhance the readability. A An artificial neural network (ANN) is a supervised ML technique
decision tree contains two types of nodes called as leaf nodes (final out- based on the model of a human neuron for classifying the data [91,92].
comes) and decision nodes (choice between alternatives) [87]. Decision ANN connected with a huge number of neurons (processing units) that
tree uses to predict a class or target by creating a training model based process information and produce accurate results. ANN typically oper-
on decision rules inferred from training data. An example graphical rep- ates on layers, these layers connected with nodes and each node asso-
resentation of a decision tree is shown in Fig. 3. The major advantages of ciated with an active function. Fig. 4 shows the basic layer structure
the decision tree are transparent, reduces ambiguity in decision-making, of an ANN. Each ANN contains three layers called input layer, one or
and allows for a comprehensive analysis. Decision trees are adopted to more hidden layer(s) and output layers. ANN classifies complex and non-
solve various issues in WSNs such as connectivity [29], anomaly de- linear data sets very easily, and there is no restriction for the inputs like

3
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

other classification methods. Several real-time WSN applications have anomaly detection and fault detection [32,45] and data aggregation
using ANN though it has higher computation requirement. ANN can be [58] approaches are used the k-NN algorithm.
applied to improve the efficiency of various issues in WSNs including lo- The comparisons of classification algorithms with respect to vari-
calization [10–15], detecting faulty sensor nodes [44], routing [46–49], ous parameters are tabulated in Table 2. The number 4 indicates the
data aggregation [61], and congestion control [72,73]. best performance whereas 1 indicates poor performanc, 3 indicates very
good and 2 can be considered as satisfactory.
2.1.5. Deep learning
Deep learning is a supervised ML approach used for classification, 2.2. Unsupervised learning
and it is a subcategory of ANN. Deep learning approaches are the data
learning representation methods with multi-layer representations (be- In unsupervised learning, there is no output (unlabeled) associated
tween the input layer and output layer). It compose with simple non- with the inputs; even the model try to extract the relationships from
linear modules that transforms the representation from lower layer to the data. Unsupervised learning approach used as classifying the set of
higher layer to achieve the best solution [93]. It is inspired by commu- similar patterns into clusters, dimensionality reduction, and anomaly
nication patterns and information processing in human nerve systems detection from the data. The major contributions of unsupervised learn-
[94]. The key benefits of deep learning are extracting high-level fea- ing in WSNs are to tackle various issues such as connectivity problem
tures from the data, work with or without labels, and it can be trained [110], anomaly detection [111], routing [112–115], and data aggrega-
to fulfill multiple objectives. It can be useful in various domains such tion [116–125]. Unsupervised learning further categorized into cluster-
as Bioinformatics, social network analysis, business intelligence, medi- ing (k-means, hierarchical and fuzzy-c-means) and dimensionality re-
cal image processing, speech recognition, handwriting recognition. The duction (PCA, ICA and SVD).
advantages of deep learning have attracted researchers of WSNs. Deep
learning have addressed various issues in WSNs such as anomaly and 2.2.1. k-means clustering
fault detection [95,96], routing [49], data quality estimation [97], and The k-means algorithm easily forms a certain number of clusters from
energy harvesting [98]. a given dataset [126]. Initially k number of random locations are con-
sidered and all the remaining points associated with the nearest cen-
2.1.6. Support vector machine ters. Once the clusters are formed by covering all the points from the
Support vector machine (SVM) is a supervised ML classifier which dataset, a new centroid from each cluster is re-calculated. The centroid
finds an optimal hyperplane to categorize the data. SVM performs the of the cluster change in each iteration, and repeat the algorithm until
best classification using hyperplane and coordinate individual observa- no more changes in the centroid of all clusters. The time complexity of
tion [99]. Most of the training data is redundant once a boundary es- the k-means algorithm is O(n∗ k∗ i∗ d), where n represents the number of
tablished and a set of points helps to identify the boundary. The points points, k indicates the number of centroids, i indicates a number of it-
which are used to find the boundary called as support vectors. SVM erations, and d represents the number of attributes. The minimization
provides the best classification from a given set of data. Therefore, the function for the sum of squares of errors [127] is presented in Eq. (2).
model complexity of an SVM is unaffected by the number of features
𝑘 ∑
∑ 𝑁
encountered in the training data. For this reason, SVMs are well suited
𝑚𝑖𝑛 𝑓 (𝑋) = ||𝑥𝑖 − 𝑦𝑗 ||2 (2)
to deal with learning tasks where the number of features is large with 𝑖=1 𝑗=1
respect to the number of training instances. Applying SVM for WSNs
have addressed issues in WSNs such as localization [15–20], connectiv- where ||𝑥𝑖 − 𝑦𝑗 || indicates the Euclidean distance between xi and yj , N
ity problem [28,29], fault detection [39,40,100,101], routing [50], and represents the number of data points from ith cluster. k-means clustering
congestion control [74]. is the simplest clustering and useful in WSNs to find optimal cluster
heads (CHs) for routing the data towards to base station [112–114]. This
2.1.7. Bayesian approach also useful to find the efficient rendezvous points for mobile
Bayesian is a supervised ML algorithm based on statistical learning sink [128].
approaches [102]. A Bayesian learning finds the relationships among the
datasets by learning the conditional independence using several statisti- 2.2.2. Hierarchical clustering
cal methods (example: Chi-square test). A set of inputs 𝑋1 , 𝑋2 , 𝑋3 … ., 𝑋𝑛 Hierarchical clustering technique groups the similar objects into
returns a label 𝜃 the probability 𝑝(𝜃|𝑋1 , 𝑋2 , 𝑋3 … ., 𝑋𝑛 ) to be maxi- clusters that have a predetermined top-down or bottom-up order. Top-
mize. Bayesian learning allows different probability functions for differ- down hierarchical clustering also called divisive clustering; in this clus-
ent variables of class nodes. Recently, several WSNs problems are solved tering, a large single partition split recursively until one cluster for each
based on the Bayesian learning strategies to improve the efficiency of the observation. Bottom-up hierarchical clustering also called as agglomer-
network. The issues are localization [21–25], coverage [31], anomaly & ative clustering; in this approach, each observation assigns to its cluster
fault detection [37,38,41–43], routing [51–53], data aggregation [63– based on density functions [129,130]. In the hierarchical clustering ap-
67], synchronization [71], target tracking [75–78,103], event detection proach, no prior information needed about the number of clusters and it
[104], and mobile sink path selection [105]. is easy to implement. The worst case time complexity of this clustering
method is O(n3 ) and the space complexity is O(n2 ) . The hierarchical
2.1.8. k-Nearest neighbor clustering used to solve various problems in WSNs, such as data aggre-
K-Nearest Neighbor (k-NN) is the most straightforward lazy, gation [131], synchronization [132], mobile sink [133,134], and energy
instance-based learning method in regression and classification. The k- harvesting [135].
nearest training set consider as an input from the feature space. K-NN
commonly classifies based on the distance between specified training 2.2.3. Fuzzy-c-means clustering
samples and the test sample. The K-NN method uses various distance Fuzzy-c-mean (FCM) clustering also called as soft clustering devel-
functions such as Euclidean distance, Hamming distance, Canberra dis- oped by Bezdek in 1981 using fuzzy set theory, which assigns the ob-
tance function, Manhattan distance, Minkowski distance and Chebychev servation to one or more clusters [136]. In this technique, clusters are
distance function. The complexity of the k-NN algorithm depends on identified based on the similarity measurements such as the intensity,
the size of input dataset and optimal performance if the same scale of distance or connectivity. Depends on the applications or data sets, the
the data. This approach finds the possible missing values from the fea- algorithms may considered for one or more similarity measures. The al-
ture space and also reduces the dimensionality [106–108]. In WSNs, gorithm iterates on the clusters to find the optimal cluster centers. FCM

4
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Table 2
Comparisons of ML techniques with various specifications [109].

Specifications DT RF ANN DL SVM Bayesian k-NN

Handling parameters 3 3 1 2 1 4 3
Accuracy 2 2 3 3 4 1 2
Learning speed 3 2 1 1 1 4 4
Classification speed 4 4 4 4 4 4 1
Handling missing values 3 2 1 2 2 4 2
Handling redundant variables 2 2 2 2 3 1 2
Dealing over-fitting 2 2 1 1 2 3 3
Handling noise 2 3 2 3 2 3 1
Handling highly independent attributes 2 2 3 3 3 1 1
Handling irrelevant variables 3 3 1 2 4 2 2

Table 3
Comparisons of various clustering algorithms.

Specifications k-means Hierarchical clustering Fuzzy-c-means

Accuracy Low High High

Speed of clustering Fast Fast Slow
Average prediction accuracy High Low Low
Performance with small observations in datasets High Moderate Moderate
Quality with huge datasets High Moderate Moderate
Results of randomness in the datasets Moderate Good Moderate
Sensitivity of noise data High Low Low

produce the optimal clustering as compared to k-means for the over- 2.2.6. Independent component analysis
lapped datasets. Like k-means clustering, it also requires prior knowl- Independent component analysis (ICA) finds a new basis for data rep-
edge about the number of clusters. The time complexity of the FCM is resentation and decomposes multivariate observations into additive sub-
higher than the other clustering approaches, and it mainly depends on components. Here the subcomponents are non-Gaussian observations
the number of clusters, dimensions, data points and iterations. This clus- [143]. ICA is a more powerful technique than PCA, in other words, it
tering approach used in various fields such as pattern recognition, image is an extended version of the PCA. ICA will remove the higher order
segmentation, Bioinformatics, and business intelligence, etc. FCM tech- dependencies, whereas PCA was unable to do. ICA analyzed data from
nique used to solve several issues in WSNs such as localization [16,137], various application fields such as web content, digital images, psycho-
connectivity [110], and mobile sink [138]. Comparisons of clustering metric measurements, business intelligence, and social networking, etc.
algorithm summarized in Table 3. In many application data, the observations are time series or set of par-
allel observations; to characterize these observations, the blind sources
2.2.4. Singular value decomposition separation method is used.
Singular value decomposition (SVD) is a matrix factorization method
which is used to reduce the dimensionality. Matrix factorization means 2.3. Semi-supervised learning
representing a matrix into a product of matrices. In Eq. (3), a m × n ma-
trix M’s SVD has been represented as. Most of the real-world application’s data is the combination of la-
∑ beled and unlabeled. The supervised learning algorithms work effi-
𝑀 =𝑈 𝑉∗ (3)
ciently on the labeled information, and unsupervised learning works ef-
where U is a m × m left unitary matrix, Σ is a m × n diagonal matrix (the ficiently on unlabeled data. The semi-supervised learning introduced to
diagonal values of Σ called as singular values of M), V is a n × n right work on the data with the combination of both labeled and unlabeled.
unitary matrix and V∗ is conjugate transpose of V. SVD can be used effi- It encompasses semi-supervised classification to perform classification
ciently for reducing the data dimensionality of the given feature space. on partially labeled data, constrained clustering to performs clustering
SVD guarantees the optimal low-rank representation of the data [139]. with both labeled and unlabeled data, regression with unlabeled data
SVD used in WSNs to address various issues like routing [115] and data and dimensionality reduction for labeled data [144,145]. There are two
aggregation [125]. distinct goals of semi-supervised learning that are to predict the labels
on unlabeled data in the training set and to predict the labels on future
2.2.5. Principle component analysis test data sets. Concerning these goals, semi-supervised learning divided
Principle component analysis (PCA) is a multivariate analysis feature into two categories: Transductive learning and inductive semi-supervised
extraction method for dimensionality reduction[140]. The PCA combine learning. Transductive learning is used to predict the exact labels for a
all the information and drops the least priority information from the fea- given unlabeled dataset, whereas the inductive semi-supervised learn-
ture space to reduce the dimensionality. The output of PCA is a linear ing learns a function f: X↦Y so that f expected to be a good predictor on
combination of observed variables (principal components). PCA some- future data. Semi-supervised learning fits with several real-time applica-
time used to detect anomalies from the data as well as in regression. In tions such as natural language processing, classifying the web content,
WSNs, sensors continuously gather information from the environments speech recognition, spam filtering, video surveillance, and protein se-
and transmitting to the base station. Applying PCA in WSNs can reduce quence classification, etc [144]. Recently, WSNs uses this learning pro-
the dimensionality of the data either at sensor level or at cluster head cess to solve localization [146–149] and fault detection [150] problems.
level to reduce the communication overheads. It reduce the buffer over-
flows at the sensor nodes or cluster heads in event-driven applications, 2.4. Reinforcement learning
which avoids the congestion problem. Several algorithms of WSNs such
as localization [141], fault detection [100,101], data aggregation [117– Reinforcement learning (RL) algorithm continuously learn by inter-
124], and target tracking [142] have adopted PCA. acting with the environment and gathers information to take certain

5
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Fig. 5. Visualization of reinforcement learning.

Fig. 7. Example of localization.

[155], coverage [156], routing [157,158], target tracking [159,160],

and mobile sink [161–164].

3. Machine learning algorithms for WSNs

In this section, we explore applications of ML in WSNs. In each sub-

Fig. 6. General scheme of evolutionary algorithms [165]. section, we discuss the advantages of selecting a ML technique to address
the issues in WSNs and also presents the features of existing approaches
in the tabular form.
actions. RL maximize the performance by determining the optimal re-
sult from the environment. Fig. 5 shows the functionality of the rein- 3.1. Localization
forcement learning. Q-learning techniques is one of the model-free rein-
forcement learning approach [151]. In Q-learning, each agent interact Localization in WSNs can be defined as recognizing the physi-
with environment and generate a sequence of observation as state-action- cal/geographical location of a sensor node. In many applications of
rewards (for example: < 𝑠0 , 𝑎0 , 𝑟1 , 𝑠1 , 𝑎1 , 𝑟2 , 𝑠2 , 𝑎2 , 𝑟3 , … . >) [152]. The re- WSNs, the sensor nodes placed in a field without knowing their positions
ward matrix R(S, A), where A and S indicate a set of actions and a set of and there is no sufficient infrastructure available to locate them after
states respectively. The actions of the agent in Q-learning shown in the the deployment. However, identifying the location of the sensor node is
matrix Q(S, A) form, it is equal to the size of R with initial values of ze- very important task. Location of sensor nodes can be known by means
ros. The rows and columns of the matrix Q are current state of agent and of manual assignment or geographical position system (GPS), some spe-
possible next state respectively. The transaction rule that update each cial sensor nodes known as Anchor node or Beacon node. An example
entry of the matrix Q with sum of the corresponding value in matrix R is shown in Fig. 7. Localization mainly categorized as proximity-based
and the learning parameter 𝛾, multiplied by the maximum value of Q localization, range based localization, angle and distance based localiza-
for all possible actions in the next state, as shown in Eq. (4). tion and known location based localization [166]. The position of sensor
nodes in the environment can change dynamically due to some external
𝑄(𝑠𝑖 , 𝑎𝑖 ) = 𝑅(𝑠𝑖 , 𝑎𝑖 ) + 𝛾 ∗ 𝑀𝑎𝑥[𝑄(𝑛𝑒𝑥𝑡_𝑠𝑡𝑎𝑡𝑒, 𝐴)] (4) causes. To handle such situations, we may require the reprogramming
or reconfiguration of the network, so applying ML techniques in such
2.5. Evolutionary computation scenarios improve the accuracy of location [167]. Along with this, ML
also provides some benefits:
Evolutionary computation is a problem-solving approach that uses
- ML algorithms can easily classify the anchor nodes and unknown
the computational models inspired from nature and biological evolu-
nodes in the network.
tion. Evolutionary computing is a subcategory of artificial intelligence
- ML algorithms are used to create clusters in the sensor networks, and
and it uses various combinatorial optimization techniques. In evolution-
each cluster is training separately to find the sensor node coordinates
ary computation, the solution of a particular problem generates over
rapidly.
iteration by iteration [153]. Initially it generates a random set of so-
- Mobile sensor nodes dynamically change their positions in WSNs, so
lutions, in every iteration it removes less fit solutions as per objective
to identify the accurate localization in such an environment is more
functions by the trail-and-error basis to achieve optimal results. Fig. 6
contented, and it is rapid with ML approaches.
shows the general scheme of an evolutionary algorithm in the form
- Table 4 summarizes the ML-based algorithm for localization in
of a flowchart. The dialects of evolutionary or nature inspired algo-
WSNs.
rithms include genetic algorithms, genetic programming, evolutionary
algorithms, evolutionary programming, ant colony optimization, parti- In [137], authors have introduced a method to improve the accuracy
cle swarm optimization, artificial bee colony, firefly algorithm, artifi- of localization using k-means algorithm and fuzzy c-means technique.
cial immune systems, memetic algorithms, and differential evolution, These two clustering are used only at the sink to train the system. The
etc. Evolutionary computation successfully implemented various appli- total field divided into clusters based on RSSI values, and each region
cations including WSNs. In [154], the authors provided the survey of dif- is trained separately to find the sensor node coordinates. This approach
ferent evolutionary approaches for WSNs. Recently, an evolutionary or requires RSSI data to train the system. A new PSO-based Neural Net-
nature inspired algorithms used to solve WSNs issues such as localization work (LPSONN) scheme has been developed using neural networks in

6
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Table 4
ML-based localization techniques for WSNs.

ML Technique Studies Complexity Environment Mobility Remarks

k-means + fuzzy-c-means [137] High Centralized static nodes Improved accuracy

ANN [10] Low Distributed static nodes Reduced error-rate
[11] Moderate Centralized static nodes Improved accuracy
[12] Moderate Centralized mobile nodes Improved accuracy
Fuzzy logic [13] Low Distributed static nodes Reduced time complexity
[14] High Centralized mobile nodes Improved accuracy
CNN+SVM [15] Moderate Centralized static nodes Improved accuracy
fuzzy-c-means + SVM [16] Low Centralized static nodes Improved accuracy and minimize time complexity
SVM [17] High Distributed static nodes Improved accuracy
[18] High Centralized static nodes Energy-efficient
[19] Moderate Distributed static nodes Improved accuracy
[20] Moderate Centralized static nodes Improved accuracy
Bayesian [21] Moderate Centralized static nodes Energy-efficient
[22] Low Centralized static nodes minimize time complexity
[23] High Centralized static or mobile nodes Energy-efficient
[24] Moderate Distributed static nodes Improved accuracy
[25] High Distributed static nodes Improved accuracy
PCA [141] Moderate Distributed static nodes Outlier detection
Regression [85] Moderate Distributed static nodes Improved accuracy
Semi-supervised [146] Moderate Distributed mobile nodes Improved accuracy
[147] Low Centralized mobile nodes Optimal time complexity
[148] Moderate Centralized mobile nodes Improved accuracy
[149] High Centralized mobile nodes Improved accuracy and reduced error-rate

[10] for WSNs. In this scheme, each anchor node finds its hop count with the system and SVM was used to find the unknown nodes in WSNs. A
other anchor nodes and sends the information to head anchor node. LP- range-free localization method has been developed using SVM classi-
SONN trains only the head anchor nodes using neural networks, and a fier, and polar coordinate system (PCS) referred as LSVM-PCS in [18].
PSO algorithm used to find the optimal number of hidden layers. The In LSVM-PCS, the region of WSNs divided into a finite number of po-
performance of the algorithm is good as compared to the other existing lar grids and each sensor node labeled into any one of the grids using
algorithms interns of error rate. A range-free localization with energy- SVM. The efficiency of the localization was improved by using a two-hop
efficient distance estimation using ANN has been developed in [11]. This mass-spring optimization (THMSO) approach. A fast-SVM based local-
algorithm used to solve the anisotropic signal attenuation interference ization algorithm has been presented in [19] for large-scale WSNs. The
for shadow and fading. This algorithm is robust in various parameters fast-SVM divide the feature space into support vector groups based on
including accuracy. Authors in [12] have presented a localization of mo- the maximum similarities. Because of the separate groups, it is easy to
bile sensor nodes using ANN and improved the performance of the pro- apply the SVM and improves the performance. Fast-SVM also overcomes
posed algorithm with PSO. Here the use of PSO was to find the optimal the coverage-hole problem and border problem. In [20], a device-free
number of neurons needed to find the accurate localization in WSNs. localization technique has been developed using multi-class SVMs. This
This algorithm improved the accuracy and performance as compared to algorithm uses a fingerprinting method along with SVM. Based on the
other traditional approaches. signal eigenvector at a receiver, an antenna array was used instead of re-
In [13], authors have used a fuzzy logic along with vector PSO to ceived signal strength (RSS). RSS is mainly affected temporal and spatial
range-free localization for a non-uniform node deployment in WSNs. fluctuations because of noise and multi-path fading. This method also
This algorithm mainly focuses on heterogeneous scenarios, and achieves improves the accuracy of the localization compared with conventional
lower complexity as compared to other localization techniques. Authors approaches.
in [14] have focused on fuzzy linguistic modeling for indoor localiza- In [21], authors have focused on the localization of unknown nodes
tion of mobile sensor nodes for WSNs. This approach mainly carries out by using the Bayesian approach. To estimate the source location, pos-
in two ways: initially to handle RSSI variations using an interval type terior CramRao bound (PCRB) was used, and sequential Monte Carlo
2 fuzzification and after that by considering fuzzy location indicator (SMC) identifies the unknown nodes which gathers the data from
(FLI) identification of geometric re-partitions. In [15], authors have pre- the sources. PCRB improves the efficiency of the SMC. A variational
sented a graph based localization algorithm using conventional neural Bayesian expectation-maximization (VBEM) approach has been pre-
networks (CNN) and SVM. The usage of heterogeneous dual classifiers sented in [22] to solve compressed sensing (CS) based localization.
in this approach is to analyze the target signals and improves the clas- VBEM iterate over two phases: VBE-phase and VBM-phase. In each iter-
sification accuracy. By enhancing signal strength, the graph-based algo- ation, VBE-phase updates the sparse signals whereas the VBM-phase up-
rithm used to identify the leakage location in the network and reduced dates the grid parameters. The efficiency of VBEM algorithm mainly de-
the error. Even if the leakage happens in the same position more than pends on the number of iteration. This algorithm is remarkable achieve-
once, it does diagnosis results and produces the accurate results. This ap- ment when the number of grids increases. In [23], a device-free localiza-
proach also overcomes the time synchronization errors. A localization tion method has been proposed using Bayesian approach. In this work,
method has been presented in [16] using SVM for large-scale WSNs. A sensor nodes send link state signals to the base station instead of trans-
fuzzy c-means training sample reduction (FCMTSR) method was used mitting the RSS measurement to reduce the data delivery rate. This ap-
to reduce the training overhead and learning calculation. FCMTSR-SVM proach work for either static or mobile target or sensor nodes.
algorithm improves the localization accuracy and reduces the training A Bayesian compressive sensing (BCS) framework has been intro-
samples time. This algorithm also effectively overcome the coverage- duced in [24], for adoptive localization in WSNs. BCS framework esti-
hole problem and border problem. mates the target locations as well as the noise in the environment. This
A range-free localization based on SVM (RFSVM) has been presented approach provides better accuracy than the traditional methods. In [25],
in [17]. In this approach, a transmit matrix was introduced to show the counting and localization techniques were presented based on the iter-
relation between hops and distances. Transmit matrix was used to train ative variational Bayesian interface. An accurate sparse approximation

7
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

method was implemented to reduce the faults caused by off-grid. In this

strategy, a three-level hierarchical prototype used to recover the faulty
prior information. A non-convex Robust PCA algorithm has been devel-
oped in [141] to eliminate the outliers. The primary goal of the PCA is
dimensionality reduction, the low-dimension gives efficient and accu-
rate results. This approach identifies the outliers between the original
data matrix and preprocessed data matrix. A localization method has
been developed using weighted linear regression in [85]. To improve
the accuracy of the localization, this approach increases the weights
of neighboring anchor nodes. This method outperforms the different
topologies of anisotropic networks. An artificial bee colony (ABC)-based
optimal relay node placement algorithm has been proposed in [155]. In
this algorithm, ABC uses to find the position where the relay node fits
to give optimal energy consumption while routing the information from
the source node to base station.
A semi-supervised learning based localization approach has been de-
veloped in [146] for mobile WSNs. This approach can easily incorpo-
rate the newly added sensor, or any environmental changes affects on
the network without any explicit or implicit changes in the algorithm.
This algorithm achieves the best accuracy even the low label data in Fig. 8. Example of coverage and connectivity.
the input. In [147], authors have presented a semi-supervised hidden
Markov model (HMM) based localization technique for mobile sensor
nodes in WSNs. This model works with various conditions, even the less In [26], end-to-end data delivery reliability (E2E-DDR) algorithm has
training data. This algorithm works for both indoor and outdoor en- been presented for link quality control in WSNs. E2E-DDR algorithm
vironment with the time complexity of O(n). A localization algorithm captures the mapping function between the background noise, packet
has been proposed in [148] based on the semi-supervised learning al- reception ratio, and RSS. These studies optimize the network quality
gorithm for mobile WSNs. This algorithm produces an accurate location and reliability in deployment of sensor nodes. The results of E2E-DDR
of the sensor nodes using large unlabeled data with the small amount method was tested using linear regression technique. In [27], a new
of labeled data collected from the signal and physical space in the net- regression-based accuracy-aware interface design has been developed
work. Authors in [149] have proposed a localization algorithm using for WSNs. This method has reduced the overhead of in-network aggre-
semi-supervised learning and support vector regression. Initially, they gation technique while measuring the connectivity. In [28], a fully dis-
used a semi-supervised learning algorithm for training the system with tributed SVM learning algorithm has been presented to a simple commu-
a small amount of labeled data and then they applied support vector nication mechanism for WSNs. In order to prove the global optimality,
regression approach to finding the target localization. Combination of the joint operations of convex hulls have used. This algorithm also im-
the algorithms provides accurate results and minimal error rate. This proves the efficiency of connectivity by reducing the communication
algorithm also utilizes less memory because of using a simple support complexity, and shorter message exchange with local nodes. In [29],
vector regression method. an SVM with decision tree based algorithm has been presented to esti-
mate a link quality in WSNs. In this method by considering estimation
parameters as RSSI and link quality indicator (LQI), SVM performs the
3.2. Coverage & connectivity classification to determine the communication quality. This method pro-
duced the best link quality accuracy, reduced energy consumption and
Apart from the energy efficiency, coverage and connectivity are also improved the network lifetime.
the challenging issue in WSNs. Coverage means how efficiently each In [30], RF-based algorithm has been used to monitor the WSNs. In
deployed sensors monitoring the area of interest. The deployment of the context of changing the quantity of the features randomly, RF pro-
sensor nodes in a network has either deterministically or randomly de- duces the best performance. This algorithm improves the accuracy of
pended on the application [168,169]. Most of the WSNs application ran- the coverage target area in WSNs. In [31], Naive Bayes Classifier has
dom deployment was feasible as compared to deterministic deployment. been used to track the location of the human in the sensor networks. In
Coverage mainly classified into two categories such as full coverage and this approach, the area of the WSNs divided into consistent static sub-
partial coverage [159,168]. Partial coverage further categorized into fo- regions. By using the Naive Bayes in the static regions, the algorithm
cused coverage, sweep coverage, target coverage and barrier coverage. tracks the initial location of the human. This approach has taken sub-
Connectivity means no isolated sensor node in the network; it means stantially less computational complexity to trace the target by covering
every node in the WSNs is sending its data to sink node directly or the area. In [110], a distributed k-means algorithm has been used along
through relay nodes. Fig. 8 demonstrates coverage and connectivity. with the fuzzy c-means algorithm for the WSNs with a strongly con-
In Fig. 8 node A and node B are isolated nodes (disconnected nodes), nected network. To classify the data of sensor nodes, k-means algorithm
and coverage hole in the network represented is in white. In WSNs, has been used. The fuzzy c-means algorithm used for partitioning the
several algorithms have been proposed to solve coverage and connect- workload depends on the different measurements.
ing problems[170]. trying to address static sensors in WSNs whereas, in A two-stage sleep scheduling approach [173] based on reinforcement
[169,171] trying to solve a mobile WSNs. Using ML techniques for cov- learning for area coverage has been presented. In this approach, sensors
erage [172] and connectivity problems have the following advantages: sparsely cover the area without knowing their geographical location. Q-
learning algorithm selects the appropriate nodes that cover the target
- To find a minimum number of sensor nodes to cover the target area area with a limited number of sensor nodes. This algorithm was pro-
can be obtained rapidly and dynamically in WSNs. viding the desired area coverage and enhances the network lifetime. In
- Classifying nodes either connected or disconnected in the network [174], the authors have presented a reinforcement learning-Probe (RL-
and changing routes dynamically without data loss. Probe), to measures the accurate quality of a link with minimal energy
- Table 5 summarizes the ML approaches for coverage and connectiv- wastage and overheads. To maintain the up-to-date information about
ity in WSNs. link quality, RL-Probe combines both synchronous and asynchronous

8
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Table 5
ML-based coverage & connectivity algorithms for WSNs.

ML Technique Studies Coverage or Connectivity Complexity Mobility Environment Remarks

Regression [26] Connectivity Low Static Centralized Optimize network quality and reliability
[27] Connectivity Moderate Static Centralized Improved accuracy
SVM [28] Connectivity Moderate Static Distributed Connection efficiency
SVM + Decision tree [29] Connectivity High Static Centralized Improved network lifetime
Random forest [30] Coverage Moderate Static Distributed Improved accuracy
Bayesian [31] Coverage Moderate Static Distributed Reduced time complexity
k-means + fuzzy-c-means [110] Connectivity Low Static Distributed Reduced work load
Reinforcement learning [173] Coverage Low Static Distributed Improved network lifetime
[174] Connectivity Moderate Static or mobile Centralized Improved network lifetime

monitoring mechanisms, and then it minimizes the measurement over- - For detecting the anomalies in non-stationary environments, ML ap-
head without interrupting the routing changes. RL-Probe reduces the proaches guarantee to handle faults, attacks, and outliers in WSNs.
energy consumption and provides the accurate link quality in WSNs. - For online anomaly detection, it is possible to adjust the parameters
The coverage control algorithm has been proposed in [156] for WSNs dynamically using the historical information.
using hybrid multi-objective evolutionary algorithms. In this, the Evo- - Table 6, summarizes the ML approaches for detecting an anomaly in
lutionary algorithm used to optimize the randomly generated weights WSNs.
and the search direction.
In [177], authors have proposed a method to detect hybrid attack
using a popular k-means clustering approach. In this approach, au-
3.3. Anomaly detection thors used Opnet modeler dataset for training and testing the system.
This algorithm works efficiently to detect malicious nodes automatically
An inconsistent observation appears significantly from the particular when blackhole and misdirection attacks happen in WSNs. Authors in
data readings are called as an anomaly [175–179]. In WSNs, misbehav- [32] have used a hypergrid k-NN algorithm for online anomaly detec-
ior identified either in measuring the sensor data or in traffic-related tion to monitor WSNs, which mainly protect from random faults and
attributes. Most of the applications in WSNs, sensors continuously gath- cyber-attacks. Using hypergrid over original k-NN improves and meets
ering the data from the environment and transmit it to the base station the specific requirements to detect anomalies in WSNs. In this approach
through relay nodes. While data transmission there is a possibility of to lower the computational and communication overhead, hypersphere
data loss because of abnormal attacks, therefore there is need for pro- detection region redefined to a hypercube detection region for online
tecting sensing data. Anomaly detection in WSNs minimizes the com- anomaly detection. In [33] authors have designed a new architecture
munication overhead while sharing the information between the sensor called intrusion detection systems (IDS) uses decision tree classification
nodes. The survey of intrusion detection for energy efficient WSNs have approach to ensure high detection rate. This approach mainly used in
presented in [180]. environmental monitoring.
Some of the possible attacks in WSNs are blackhole attack, misdi- Authors in [111] have used two ML-based algorithms such as least
rection Attack, wormhole attack, sinkhole attack, and hybrid anomaly squares-SVM along with sliding window and PCA to detect outlier in
[177]. Fig. 9(a) shows the general data flow of WSNs. In blackhole at- WSNs. In this work, modified RBF kernel has been used along with least
tack, a block receives the packets instead of forwarding it to the base squares-SVM to improve the performance of anomaly detection in non-
station as shown in Fig. 9(b). A misdirection attack is shown in Fig. 9(c), stationary time-series. PCA has been applied for tracking subspace re-
in this the attackers routes the packets with distinct nodes rather than cursively in WSNs. Support vector data description (SVDD) classifier has
its neighbor nodes. It may construct the longer route and reduces the been used in [34] to detect anomalies of node’s data. The primary goal
throughput. A wormhole (WH) attack, a WH tunnel is formed between of this approach is to minimize the complexity of training and testing
two distinct nodes and misapprehension that they are very close. This phases. A two-order approximation based sequential minimal optimiza-
WH can bypass or attract a number of packets from the network and at- tion (SMO) algorithm has been used to reduce the complexity of training
tackers perform the manipulations [181]. In Fig. 9d, a wormhole attack phase. In the testing phase, a fast decision-making technique was pre-
is shown. A particular node (Sinkhole) advertises an optimal route to sented using the center point of a hypersphere in kernel feature space.
its neighbor nodes to the base station. In sinkhole attack, it tampers the In [35], authors have provided a detailed analysis of various outlier
data and damages the network. We represent a sinkhole in Fig. 9e. In (event or error) detection methods using one-class SVM in harsh en-
hybrid attack, different types of attacks such as blackhole, misdirection, vironments. They also analyzed different planes like hyper-sphere, hy-
WH, and SH attack affects at a time. perplane, hyper-ellipsoidal and quarter-sphere they found that quarter-
ML approaches have used to detect anomalies in WSNs, protecting sphere formulations are low computational complexity and communi-
from various attacks and misapprehensions. In WSNs, several algorithms cation overhead as compared to others for outlier detection.
have used to protect from anomalies such as self-learning threshold and In [36], authors have proposed a DBSCAN algorithm for detecting
sleep scheduling approaches [182] for a heterogeneous sensor network anomalies by evaluating various features in WSNs. They also used a
based on geological sensor nodes, [183] for sinkhole attack as well as well-known classification technique of SVM to improve the detection
denial-of-service (DoS), by detecting inconsistent data [184] provides accuracy in DBSCAN. The efficient learning feature of SVM used to train
resilience and genuine information. The survey of anomaly detection the system with standard and general data. DBSCAN is accurate to detect
on non-stationary datasets using ML presented in [179]. Using ML for low-density regions in the network, and removes the anomaly. Detect-
anomaly detection in WSNs significantly improved as compared to other ing DoS attacks in distributed WSNs is an extremely challenging task
approaches, benefits listed as follows: with traditional approaches. In [185], a combination of game theoretic
method and reinforcement learning based fuzzy Q-learning approach
- A hybrid anomaly is the combination of various attacks, therefore has been used to detect malicious behavior in WSNs. Compared with
detecting the node which effects and type of anomaly are happen- Markovian game-theory, ML based approaches noted the best perfor-
ing. Employing clustering algorithms minimizes the complexity and mance in various parameters like accuracy, lifetime, energy consump-
communication overhead for the problem. tion, efficiency, and visibility. In [37], Bayes Classifiers have been used

9
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Fig. 9. Anomaly detection in WSNs (a) normal flow (b) blackhole attack (c) misdirection attack (d) wormhole attack (e) sinkhole attack.

Table 6
ML-based anomaly detection techniques for WSNs.

ML Technique Studies Anomaly Complexity Environment Remarks

k-Means [177] Hybrid anomaly Moderate Centralized Improved accuracy

k-NN [32] Random faults, Cyberattacks Low Distributed Reduced time complexity
Decision tree [33] Sinkhole Moderate Centralized Improved accuracy
SVM + PCA [111] Outliers detection Moderate Distributed Improved accuracy
SVM [34] Anomaly detection High Centralized Reduced time complexity
[35] Error detection Low Centralized or Distributed Reduced time complexity
[36] Intrusion detection High Centralized Improved accuracy
Q-learning [185] DoS High Distributed Improved network lifetime
Bayesian [37] outlier detection Moderate Distributed Improved accuracy
[38] Trust management High Distributed Improved accuracy
Regression [186] Anomaly detection Moderate Centralized Improved accuracy
Deep learning [95] Intrusion detection High Centralized Improved accuracy

to detect outliers for WSNs application. This method performs outlier training set data. The results of this algorithms show that the accuracy
detection in two levels; one at the sensor nodes level and the second is improved as compared to random forest, and SVM.
one at the gateway. This approach produces a high outlier detection
accuracy for WSNs. 3.4. Fault detection
Bayesian-based trust management strategy (BTMS) has been pro-
posed for WSNs in [38]. BTMS perform the trade-off between the energy WSNs have been deployed in a various application-specific environ-
consumption of sensor nodes and security. From the overall data gath- ment such as harsh, hostile, and unattended. Consequently, faults may
ered from the nodes, it evaluates the trust and identifies direct and indi- occur in WSNs such as communication failures, battery failures, hard-
rect trust information. In [186], authors have used SMO regression tech- ware failures, software failures, inefficient base station or topological
nique to detect an anomaly of sensor data for healthcare application. In changes [187]. Detecting the fault in WSNs is challenging because of
this technique, based on the historical data, the algorithm predicts false several reasons such as resource limitation, a different type of environ-
positive and true negatives. The detection rate of this approach reach ment (forest, indoor), changes in deployment, detection accuracy be-
100% accuracy rate and providing rapid results as compared to tradi- tween the normal and faulty nodes [39]. Several fault detection meth-
tional approaches. In [95], authors have proposed a method to detect ods [188–190] are presented by the researcher without using any ML
intrusion by combining a special clustering and deep neural network ap- approaches, whereas applying ML for detecting faults have the follow-
proaches. Initially this algorithm divide the entire data set into k-subset ing:
using special clustering method, and then deep neural network perform
the intrusion detection operation by comparing the testing set with the - Brisk detection and categorization of faults.
- Improves accuracy in detecting faults.

10
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Table 7
ML-based fault detection mechanisms for WSNs.

ML Technique Studies Fault detected Accuracy Complexity

SVM [39] Negative alerts 99% Low

[40] Fault nodes 98% High
SVM + PCA [100] Hardware faults 99.8% Moderate
[101] Online fault diagnosis 99% High
Bayesian [41] Body sensors -30% fault rate Moderate
[42] faulty sensor nodes 100% High
[43] faulty sensor nodes 98% Moderate
ANN [44] faulty sensors 98% High
Deep learning [96] Faulty data 99% High
k-NN [45] fault nodes 99% Moderate
Semi-supervised [150] Faulty node detection 99% Moderate

- Several ML-based fault detection mechanisms for WSNs are shown

in Table 7.

To detect the faults in WSNs, a well-known SVM classifier has been

proposed in [39]. A significant advantage of using SVM is to classify the
data as well as decision making. This approach detects faults in sensor
data very accurately and rapidly. In this approach, a kernel function
was used for detecting the faults in non-linear classification data and
the accuracy rate of detecting defects exceeds 99%. An error predic-
tion method has been developed in [40] using SVM and cuckoo search
algorithms. In this approach, SVM used to predict the errors in sensor
dynamically; however, it depends on the parameter comparisons. To
optimize the key parameter, this method adopted cuckoo search algo-
rithm to avoid local minimum value. In [100] authors have presented a
technique for diagnosis induction motor (signal processing) faults using
multi-class SVM classifier and PCA. Here, the multi-class SVM classifier
was used for classifying and training the various faults affected by the in- Fig. 10. Example of routing in WSNs.
duction motor in the sensor node. SVM increases the fault classification
accuracy and PCA was used to extract more dominant feature dimen-
sions. This approach achieves very high accuracy (99.80%) as compared In [96], a centralized solution has been developed by the authors
to other traditional strategies. to detect the faults in WSNs using deep neural networks. The objective
Authors in [101] have presented an online fault detection method for of this method is to improve the fault detection accuracy and reduce
real-time data streams using recursive PCA and multi-class SVDD clas- the power consumption. This approach achieves 99% of detection accu-
sification method. Recursive PCA method was used to detect the fault racy. In [150], authors have introduced an efficient label propagation
in WSNs, and it is lightweight. SVDD classifier used to identify the fault method for faulty node detection method using semi-supervised learn-
categories. Fault detection in body sensor network (BSN) causes a false ing approach. A kernel density based function used in this approach to
medical diagnosis, therefore it is very important to detect a fault in BSN. label the unlabeled data in the input data set.
In [41], Bayesian network based fault detection method has been pre-
sented for BSNs. A Bayesian network method was used to capture the 3.5. Routing
temporal and spatial correlation of body sensor. Based on an optimal
threshold value, sensor faults can be identified. Authors in [43] have Routing is one of the primary challenges in WSNs because of the lim-
proposed efficient malicious nodes identification for Smartphone net- ited power supply, low transmission bandwidth, less memory capacity,
work based on the Bayesian network model. To ease the security and a and processing capacity. In WSNs, sensors deployed randomly in the
trust computation proposed method has been implemented by the hier- environment, each sensor node collects the data from the environment
archical method. This method provides an accurate faulty node detec- and transmits to the base station for further processing. Fig. 10 shows
tion and energy-efficiency. the multi-hop transmission from sensor nodes to base station. In gen-
Authors in [42] have presented a fault detection scheme (FDS) which eral, the nodes which are near to the base station consume more energy
detect the faulty nodes in WSNs with their batteries and sensing infor- because they serve as a relay nodes. The goal of routing protocol de-
mation. In FDS, faulty nodes can be found in two-level verification. In sign is to reduce the energy consumption of sensor nodes and increase
the first phase, a Naive Bayesian classifier was used to detect the fault the network lifetime. Recently, several routing methods [191–193] are
inside the sensor node, while the second phase the cluster head or gate- developed for WSNs by the researcher using different approaches. But
way evaluated fault detection. This method shows the 100% accuracy employing the ML techniques for WSNs to develop routing protocols.
rate through simulation results. In [44], a fault node distribution and Some of the benefits of ML-based routing for WSNs are as follows:
management approach have been presented based on the fuzzy rules
in WSNs. This algorithm mainly concentrated on re-usability of faulty - To select an optimal set of CHs for routing in WSNs, ML algorithm
nodes by implementing the efficient route towards base station. It pro- adopts changes in the environment without re-programming.
vides a better QoS and network lifetime. A k-NN classifier was used to - ML has a wide range of applications in WSNs such as optimal routing,
separate fault sensor nodes from the normal sensor nodes based on the lowering communication overhead and delay-aware [194].
abnormal behavior of sensors [45]. This approach mainly focus on error - Several ML-based routing protocols for WSNs are summarized in
rate of the sensors to decide faulty sensor nodes. Table 8.

11
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Table 8
ML-based routing algorithms for WSNs.

ML Technique Studies Topology Complexity QoS Environment Mobility

ANN [46] Tree High No Centralized static nodes

[47] Tree Moderate Yes Distributed static nodes
[48] Tree Moderate No Distributed mobile nodes
Deep learning [49] Hybrid High No Centralized mobile nodes
SVM [50] Hybrid Moderate No Distributed static nodes
Bayesian [51] Tree Moderate No Distributed static nodes
[52] Hybrid Low No Centralized static nodes
[53] Hybrid Moderate No Centralized & decentralized mobile nodes
k-means [112] Hybrid Low Yes Distributed static nodes
[113] Tree Moderate No Distributed static nodes
[114] Hybrid Moderate No Centralized static nodes
SVD [115] Arbitrary Moderate No Distributed static nodes

In [46], authors have developed ELDC approach based on artificial provides a better packet delivery ratio, throughput, lowering the energy
neural networks (ANN) to determine energy-efficient and robust rout- consumption and controls the traffic overhead. In [113], authors have
ing for WSNs. In ELDC, the usage of ANN is to train the protocol. ANN proposed energy efficient clustering protocol using k-means (EECPK-
training the protocols using various parameters in WSNs such as resid- means) algorithm to find the optimal center point from the cluster
ual energy and the distances between nodes, cluster heads (CH), border from a random initial center point. It selects optimal CHs based on the
nodes and sink or base station. It generates a massive amount of train- Euclidean distance and residual energy of the sensor nodes in WSNs.
ing set, even ANN an efficient threshold values for selecting a group of EECPK-means algorithm finds the efficient multi-hop communication
reliable CH based on backpropagation. ELDC is highly energy-efficient path from the CHs to base stations. This algorithm avoids the data loss
and balances the energy consumption of the sensor nodes and avoids the and balances the energy consumption of the sensor nodes. An energy
data loss in the WSNs. Authors have presented a dynamic 3D fuzzy rout- efficient k-means technique (EKMT) has been presented in [114] to find
ing based on traffic probability (DFRTP) routing protocol in [47] using the optimal cluster heads (CH), which are near to the member nodes
the fuzzy system. The fuzzy system takes neighbor nodes and distances as well as the base station. EKMT measure sum of squared distances be-
as an input and produces routing probability as an output. One of the tween nodes and choose CH based on minimum distance. This technique
neighbor nodes of every node in the network which provide optimal improves the throughput and reduces the delay by re-selecting the CHs
(lowest) traffic probability is chosen as data forwarding node towards dynamically.
to sink. DFRTP increases the data delivery ratio and reduces the energy Authors in [115] have developed singular value decomposition
consumption of the sensor nodes. (SVD) based in-network routing protocol of WSNs with an arbitrary
The zone-based energy efficient routing (ZEEP) protocol has been topology where the shallow light tree (SLT) used along with SVD to
developed in [48] using a fuzzy system for mobile sensor networks. route the information towards to base station. This can implement es-
ZEEP selects a set of CHs using fuzzy inference system (FIS) by consider- pecially for smart cities and structural health monitoring (SHM) using
ing parameters such as mobility, density, distance, and energy. Finally, IoT devices. Due to a large number of sensor nodes, this is quite a high
a genetic algorithm was used to finalize the optimal set of CHs, from transmission overhead. In [195], a secure cluster based routing proto-
the collection of CHs nominated by FIS. ZEEP balance the energy con- col has been developed to enhance the network lifetime for WSNs. In
sumption by selecting the optimal set of CHs and enhance the network this approach, CHs selected based on their distances and residual en-
lifetime. In [49], a deep learning based routing protocol has been in- ergy. This algorithm mainly focuses on the isolated CH and edge node
troduced with the base station as an infrastructure. It means the route to balance the node energy consumption. In [157] authors have pre-
maintained, assigned and recovered by the base station. Proposed deep sented an efficient routing mechanism based on the transmission range
learning based algorithm adopts dynamic routing in a mobile sensor net- of the sensor nodes and the data forwarding load. In this mechanism,
work. The base station initially creates a list of virtual routing paths, and efficient clustering method was used which based on the particle swarm
from them it identifies the optimal route. This algorithm overcomes the optimization, to balances the load of the sensor nodes in the networks.
congestion and packets loss and power management. An efficient SVM An ACO-based routing algorithm has been presented in [158] for WSNs.
based routing protocol has been developed in [50] which balance the In order to find the optimal routing, they consider various parameters
energy consumption of the nodes by assigning the nodes to nearest CH. such as the residual energy of a node, transmission distance, transmis-
This algorithm achieves higher network lifetime as compared to LEACH sion path, and the shortest path between the source nodes to the base
protocol. station. This algorithm results in minimum energy consumption and pro-
A Naïve Bayes based optimal cluster heads selection for efficient longs the network lifetime.
routing has been presented in [51]. The optimal set of CHs always bal-
ances the energy consumption of sensor nodes and enhances the net- 3.6. MAC
work lifetime. Naïve Bayes guarantees that even new features added or
changed in the network dynamically. A new adaptive integrated rout- In WSNs, the network lifetime can maximize by developing energy
ing framework has been presented in [52] for data collection based on a efficient medium access control (MAC) protocols. MAC layer is a sub-
Bayesian technique. In the projected technique, an adaptive projection layer of data link layer with the primary functions are addressing, data
vector is constructed in each iteration of routing by introducing a new transfer from upper layers, channel allocation, predict errors, and frame
target node selection. In [53], a Bayesian learning method based opti- recognition. Consequently, designing energy efficient MAC protocols is
mal routing prediction model has been developed for both decentralized a challenging for WSNs because of the dynamic behavior of adding or
and centralized versions. This approach also performs the scheduling removing (maybe dying) sensor nodes, noise data transmission and so
approach while routing the data to balance energy consumption. This forth [196]. MAC protocols mainly classified as contention-based (no
algorithm is much more suitable for decentralized than centralized. need for central coordination) and schedule-based (each node communi-
Authors in [112] have used a well-known k-means classification al- cates during specific time slot) protocols. ML approaches for MAC layer
gorithm to find optimal clustering in WSNs for routing. This algorithm protocol design guarantees the energy efficient and avoids latency. The

12
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Table 9
ML-based MAC designs for WSNs.

ML Technique Studies Complexity Category Type Synchronization

Random forest [54] Low contention-based Hybrid No

Reinforcement learning [197] Moderate schedule-based Hybrid Yes
[198] Low contention-based ALOHA Yes
[199] High contention-based Hybrid Yes
[200] Moderate contention-based ALOHA/ CSMA-CA No
[201] Low schedule-based CSMA Yes

summary of ML based MAC protocols listed in Table 9 and benefits of posed in [203–205]. ML algorithms for data aggregation in WSNs are
ML-based MAC protocol designs are as follows: summarized in Table 10 and their benefits are listed below:

- Reduces the additional burden to reconfigure of any newly joined or - ML approaches select efficient CHs in the network for data aggrega-
died nodes in the network. tion that will significantly balance the energy of the sensor nodes.
- Improves the efficiency of self-learning process of the network and - ML algorithms helpful for dimensionality reduction of the data at the
reduces the end-to-end delay. sensor node level, therefore it reduces the communication overhead
in the network. The reduction may be performed at sensor nodes or
In [54], authors have proposed MAC address spoofing detection cluster heads to minimize the delay in transmission of data.
mechanism based on random forest. This method performs better re- - ML adopts the environment and work accordingly without reconfig-
garding prediction time, and accuracy compared to traditional cluster- uring or reprogramming in the context of data aggregation.
based methods. Authors have suggested the one-class SVM to train the
In [55], a distributed linear regression based data gathering (DL-
system without the whole network range. An optimal channel allocation
RDG) algorithm has been presented to enhance the CHs functionality
scheme has been proposed in [197] based on reinforcement learning
in WSNs. Based on the past data at each CH in WSNs, the regression
for sensor networks. This approach adopts the dynamic behavior of the
method predicts the actual monitoring measurements and broadcast to
network and considers various parameters to allocate the channel such
the base station. Linear regression model also chooses the faulty nodes
as availability of the channel, cost of sense, and impairment of sense.
in the network during the data gathering process. An energy efficient
This method improves the network lifetime by balancing the energy
multivariate data reduction model has been developed in [56] based on
of the sensor nodes and enhances the accuracy of channel allocation.
periodic data aggregation using polynomial regression functions. The re-
MAC layer protocols (ALOHA) extended using the Q-learning approach
duction of data in sensor level will reduce the communication overhead
in [198], in order to extend the channel performance and accuracy. This
during the transmission of the data between the nodes or CHs towards
protocol efficiently works for any new sensor nodes added into the net-
to base station. Euclidean distance only not sufficient in all cases of the
work and balances the energy consumption of the nodes in WSNs. The
WSNs, it is also required to determine a temporal and spatial correlation
ALOHA-Q reduces the packet loss and works dynamically according to
between the sensor nodes in the complex WSNs deployments. A linear
the changes happen in the network.
regression model has been used in [57] to estimate the non-random
In [199], reinforcement learning based cooperative approaches for
parameters during information gathering. Before transmitting the data
WSNs has been proposed. This method works for both the distributed
to the base station, the CHs compress the data. This method specially
and centralized sensor networks. The parameters of the system are dy-
designed for the decentralized network, with sufficient condition, cen-
namically activated or deactivated depends on the situation of the net-
tralized approach also achieve good performance.
work. Due to dynamic activation or deactivation policy, the performance
In [59], authors have focused on agriculture and health monitor-
of the network may be positive or negative. Reinforcement learning-
ing application. Heterogeneous WSNs are utilized to collect the data
based self-learning techniques for an optimal set of services have been
from the fields. For agriculture application, decision tree analysis was
presented in [200]. This method is highly dynamic in nature and pro-
used and for the health monitoring application, a honeybee soft com-
vides a well efficient compelling set of services from the network. This
puting method was used. The decision tree used to classify the data,
algorithm reduces the end-to-end delay of multi-hop transmission and
and the classification accuracy is 95.38% for different cases. Authors
improves the reliability of the network. In [201], a channel hopping
in [60] have developed a distributed data fusion framework for hetero-
method developed based on reinforcement learning approach to im-
geneous WSNs. This framework has the capability of a self-organizing
prove the performance of the network has been presented. This method
using decision tree for proving optimal computational scalability, data
efficiently works without any additional configuration even if any new
flow, data quality, running time and energy consumption. This frame-
node joins in the WSNs. This method controls the message exchange
work adopts the dynamic changes in the network automatically. In [61],
based on time synchronization.
a low-cost data gathering and decision making approach have been de-
veloped using fuzzy set theory for wireless body sensor networks. De-
3.7. Data aggregation cision making on the gathered data from the bio-sensors improves the
data quality and transmission overhead. In [62] authors have presented
The process of collecting and aggregate data from the sensor nodes a decision tree based imbalanced data classification and predict imbal-
called data aggregation. Data aggregation in WSNs affects various pa- anced data classes using Naïve Bayes to reduce imbalance class of data.
rameters such as power, memory, communication overhead and com- This approach also reduces the communication overhead in the network
putational units. In order to reduce the number of transmission and and balances the energy consumption of sensor nodes.
communication overhead, data aggregation places a significant role in Authors in [63] have proposed multi-sensor data fusion techniques
WSNs. An efficient data aggregation method balances the energy con- using the Bayesian system to collect and analyze the heterogeneous
sumption of the sensor nodes and enhances the network lifetime. There data. This technique self-organize or configure according to the dynamic
are several types of data aggregation methods depends on the network changes in the network. This method provides a context-aware, adap-
structure like cluster-based data aggregation, tree-based data aggre- tive system for efficient data fusion process for WSNs. In [64], sparse
gation, in-network data aggregation, and centralized data aggregation Bayesian-based compressive sensing approaches have been proposed to
[202]. Several approaches to data aggregation in WSNs have been pro- decrease the active sensor nodes while gathering data. This method also

13
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Table 10
ML-based data aggregation approaches for WSNs.

ML Technique Studies Mobility Environment Complexity Topology Remarks

Regression [55] static Distributed Low Tree Improved network lifetime

[56] static Distributed Low Tree Improved network lifetime
[57] static Distributed/ Centralized Moderate Hybrid Improved network lifetime
k-NN [58] static Distributed Moderate Hybrid
Decision tree [59] static Distributed Moderate Tree Improved accuracy
[60] static Distributed Low Tree Improved network lifetime
ANN [61] static Centralized High Hybrid Improved accuracy
[72] static Distributed High Tree Improved network lifetime
Naïve bayes + Decision tree [62] static Distributed High Tree Improved network lifetime
Bayesian [63] static or mobile Distributed Moderate Hybrid Improved accuracy
[64] static Centralized Low Hybrid Improved time complexity
[65] static Distributed Moderate Hybrid Reliable data transmission
[66] static Distributed Moderate Hybrid Improved network lifetime
[67] static Centralized Moderate Tree Improved accuracy
k-means [116] static Distributed Moderate Tree efficient redundancy elimination
Hierarchical clustering [131] static Distributed or centralized Moderate Hierarchical Reduce unnecessary transmissions
PCA [117] static Distributed Low Tree Improved network lifetime
[118] static Distributed Moderate Tree Improved network lifetime
[119] static Distributed High Tree Improved network lifetime
[120] static Distributed Moderate Tree Efficient drift findings
[121] static Distributed Moderate Tree Reduce unnecessary transmissions
[122] static Distributed Moderate Tree Improved network lifetime
[123] static Distributed Low Hybrid Improved network lifetime
[124] static Distributed Moderate Hybrid Improved network lifetime
SVD [125] static Centralized High Hybrid Reduce unnecessary transmissions
Genetic classifier [206] static or mobile Distributed Moderate Star Improved network lifetime

minimizes the estimation error while gathering the data from sensor tion overhead between sensor nodes, CHs and base station. In [120], a
nodes. It outperforms the traditional approaches and has lower compu- PCA-based drift detection technique and angle optimized global embed-
tational complexity. A Bayesian model has been presented in [65] for re- ding (AOGE) have been presented. PCA and AGOA analyze projection
liable data transmission between the node and base station. This method variance and projection angle respectively to determine principle com-
achieves channel-aware data transmission between the sensor nodes ei- ponent. This method efficiently finds the drift from data streams.
ther single-hop or multi-hop manner. In [66], a distributed Bayesian A PCA-based distributed adaptive algorithm has been developed in
network model has been used to compress the sensing data at each sen- [121] to find Q smallest and highest eigenvalues for the sensor network.
sor node level before transmitting to the CH. To avoid the non-existing This approach dynamically determines the eigenvectors without any ex-
packet paths, this algorithm adopts arithmetic coding. This algorithm plicit constructions. A controlled number of transmissions for producing
reduces the communication overhead and improves network lifetime. quality data using Compressive-Projections PCA has been developed in
Temporal sparse Bayesian learning based sensor drift estimation [122] to reconstruct data. This approach provides an energy efficient
method has been proposed in [67] from under-sampled observations quality data reconstruction. This method also constructs the clusters for
for WSNs. This method identifies the drifting sensor without dense de- in-network processing. A recursive-PCA (R-PCA) method has been devel-
ployment or prior knowledge of data models. Authors in [58] devel- oped in [123] for energy efficient data gathering and outlier detection.
oped a novel k-NN based missing data estimation method for WSNs by R-PCA algorithm runs in CHs to aggregate data as well as outlier detec-
determining the temporal and spatial correlation between the sensor tion. R-PCA performs the data aggregation with high recovery accuracy
nodes. For the optimal computational search and measuring the missing and efficient outlier detection. In [124], a novel PCA-based framework
data percentage, weighted k dimensional-tree data structure has been has been proposed for data recovery, prediction and compression. The
used. In [116], a k-mean and ANOVA-based clustering approach have primary goal of this approach is to reduce the communication over-
been proposed for efficient data gathering in underwater WSNs. This head. By compressing the data at CHs, this algorithm achieves lower
approach eliminates redundancies at each sensor node before transmit- communication overhead. This approach provides an accurate predic-
ting to CHs. Once the data set received by CH, it performs k-means and tion, recovery and efficient compression mechanism for enhancing the
ANOVA to identify the similar datasets before sending it to the base network lifetime.
station. In [125], SVD based numerical analysis has been performed to in-
In [117], PCA based data aggregation technique has been proposed. vestigate the missing information in an image. This method is very ac-
The primary goal of this approach is to reduce the energy consump- curate and also avoids noisy data from the image sensed by sensors but
tion of the sensor nodes, and it is achieved by data compression and inaccurate when the sensor nodes are less. Neural networks based data
signaling. In this method, the sensor nodes send their data to CHs, and compression technique have been proposed in [72] for avoiding the con-
the CHs take the compression responsibility before transmitting to the gestion in the network as well as balancing the energy consumption of
base station. Authors in [118] have devised a PCA-based data com- the sensor nodes in WSNs. This algorithm provides accurate results com-
pression model for stationary WSNs. The primary design goal of this pared with various traditional algorithms for spatial and temporal data
approach is to reduce the communication overhead among the sensor compressions. A novel power-aware hybrid data aggregation approach
nodes. This approach successfully achieves energy efficiency and max- investigated in [131] using compression sensing along with hierarchical
imizes the network lifetime. In [119], self-organizing algorithms using clustering approach for large scale WSNs. In this method, cluster sizes
distributed PCA for WSNs have been implemented to minimize unneces- vary at each leave with the help of different threshold values to optimize
sary transmission and improving network lifetime. This approach better the number of data transmissions in the network. This algorithm espe-
for any compression rate and able to solve principal component in CHs. cially reduces communication overhead on the cluster head. In [206],
This method achieved to improve the network lifetime and communica- authors have used genetic ML approach for parallel data gathering for

14
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Fig. 12. Congestion in WSNs (a) Node level congestion (b) Link level congestion
[213,214].

Fig. 11. Clock model of sensor nodes.

synchronization methods. A clustering task-scheduling mechanism has
been introduced in [132] for WSNs using hierarchical clustering. This
WSNs. This approach work dynamically according to the topological
method mainly reduces the energy consumption of the sensor nodes in
changes of the WSNs, save energy of the sensor nodes, and reduce the
the network for applying the unnecessary clustering process.
packet loss.

3.8. Synchronization 3.9. Congestion control

Clock synchronization is one of the major components of WSNs, In WSNs, congestion happens when a sensor node or communica-
it mainly used in protocols design. Synchronization was used in var- tion channel handles more data transmission than its capacity. There
ious parameters such as data aggregation, power management (sleep are several causes for the occurrence of congestion such as node buffer
scheduling), transmission scheduling, localization, security and target overflow, many-to-one data transmission scheme, transmission channel
tracking. In WSNs, each sensor node has a common time frame, and it contention; packet collision, dynamic time variation, and transmission
may be different from others as shown in Fig. 11. The clock rate in WSNs rate [213–215]. Fig. 12, illustrates the node level congestion and link
measured using PPM. The clock rate of a sensor node i is represented as level congestion. Node level congestion happens because of high packet
Ci (t) calculated [207,208] as in Eq. (5). arrival rate to a particular node (Figure. 12(a)) whereas link-level con-
gestion occurs because of the collision and lower bit transmission rate
𝐶𝑖 (𝑡) = 𝜃 + 𝑓 .𝑡 (5) between two nodes (Figure. 12(b)). Congestion affects various parame-
where the parameter 𝜃 indicates offset, and f means frequency difference ters of WSNs such as energy consumption, QoS, end to end delay, and
(clock skew). packet delivery ratio (PDR). Congestion control is one of the most sig-
Fundamental time synchronization methods are classified into three nificant challenges in WSNs. Recently, some of the control congestion
categories such as one-way messaging, two-way messaging and receiver- approaches have been proposed in [216–218] for WSNs to improve and
receiver synchronization. Synchronization achieved with various tech- energy-aware data routing. However, ML approaches more promising
niques without using ML [209–212]. However, there are several advan- for congestion control and have following benefits:
tages of using ML approaches for WSNs are listed as follows:
- ML approaches are more accurate to estimate the traffic and also
- ML improves the accuracy of the synchronization for WSNs. can find optimal path for minimized end-to-end delay between the
- ML adopts the changes in the environment and resynchronizes ac- nodes and base station.
cordingly. - Using ML techniques, transmission ranges may change dynamically
due to dynamic changes in the network.
A regression-based synchronization protocol has been developed in
- Table 11 briefs a comparison of ML approaches for congestion con-
[68] to analyze various factors affect the performance synchronization.
trol in WSNs.
The factors considered here are clock frequency noise, latency, and clock
drift. This method is particularly useful for the low-cost WSNs. In [69], Congestion affects directly to the pocket loss, energy consumption
a linear regression model-based long-term synchronization method has and end-to-end delay. Neural networks based data compression tech-
been proposed. Because of external factors, the WSNs will change dy- nique has been proposed in [72] for avoiding the congestion in the WSNs
namically, and the clock drift between the sensor nodes also can change as well as balancing the energy consumption of the sensor nodes. Trans-
dynamically it may affect synchronization in the network; therefore this mitting the compressed data between the nodes or CHs will reduce the
algorithm resynchronizes according to the variation. This approach also communication overhead. Therefore it reduces the energy consumption
reduces the synchronization error and produces the high accuracy rate rate and also protects from the congestion occurrence. This algorithm
as compared to other conventional methods. performs the better in congestion control as compared to various tradi-
In [70], regression-based synchronization method has been pre- tional algorithms for spatial and temporal data compressions.
sented for low-cost applications of WSNs. To improve the synchroniza- In [73], fuzzy logic based congestion detection and control mecha-
tion performance, this algorithm focuses mainly on clock drift, resolu- nism have been developed to minimize packet loss ratio using efficient
tion of clock, latency, clock wander and jitter. In [71], a Bayesian in- and active queue management. This approach performs in three phases.
terface along with linear regression methods used to improve the accu- Initially, it manages the queue with the fuzzy system used to detect
racy of synchronization in WSNs. The major constraints consider in this the congestion. In the second phase it adjusts the congestion control,
work are clock and power consumption. Because of using Bayesian inter- and finally, it recovers from the congestion by balancing the flow. In
face the algorithm improve 12% accuracy compared with conventional [74], authors have presented an SVM based congestion control method

15
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Table 11
ML-based congestion control strategies for WSNs.

ML Technique Studies Congestion control Parameter metric Data flow Control pattern

ANN [72] traffic control end-to-end delay, energy consumption continuous Hop-by-hop
Fuzzy logic [73] queue length packet loss ratio continuous Hop-by-hop
SVM [74] transmission rate throughput, latency, energy consumption continuous Hop-by-hop
Reinforcement learning [219] traffic control throughput, energy efficient continuous Hop-by-hop

Table 12
ML-based target or node tracking algorithms for WSNs.

ML Technique Studies #Targets #Sensors Mobility of target Mobility of sensor Remarks

Bayesian [75] single single static static Improved accuracy

[76] single single mobile mobile improved accuracy
[77] single multiple static static Reduced communication overhead
[78] multiple single/ multiple static static Improved network lifetime
Bayesian + RL [103] single multiple static static or mobile Improved network lifetime
PCA [142] single multiple static static Improved network lifetime
Q-learning [224] single multiple static static efficient task scheduling
Genitic algorith [159] single multiple static static Improved network lifetime
memetic algorithm [160] single single static static Improved network lifetime

for WSNs. In this method, transmission rates of each node are adjusted selects a subset of a sensor for target selection. This approach reduces
due to dynamic changes in traffic. As compared with other classifica- the communication overhead and efficient signal recovery approach. A
tion methods, SVM produces the accurate results. An energy-efficient multi-target tracking method has been proposed in [78] for WSNs us-
data gathering approach has been proposed in [219] using multi-agent ing Bayesian filtering algorithm. This method developed with two lay-
reinforcement learning approach. This approach controls traffic at the ers; in first layer handling information fusion problem and second layer
cluster-ring level and takes the optimal routing decision. This approach Bayesian approach adopted for tracking the target. Because of using
also adopts routing patterns for the dynamic changes in network topolo- two-layer approach, computational overhead of this approach will re-
gies. duce and balances the energy consumption of the sensor nodes in the
network.
3.10. Target tracking In [103], two ML approaches, Bayesian and reinforcement learning
are used to monitoring an event. A dynamic Bayesian network was used
Target tracking indicates detecting and monitoring a particular static to observe the event and reinforcement learning was used to apply the
or mobile phenomenon in the network. Target tracking in WSNs may optimal duration for sleep schedules to the sensor nodes. This approach
perform with single or multiple sensor nodes depends on the application. was more efficient regarding energy consumption of the sensor nodes
Using single node for tracking the target benefits consumes less energy and data accuracy. Authors in [142] have developed a system for ob-
as compared to multiple sensor nodes, whereas multiple sensor nodes serving a specific volatile organic compound (VOC) in the industrial
provide accurate results. The quality of target tracking invites some of environment using wireless sensing system. This method adopts PCA to
the challenges such as node failure, target missing, coverage and con- process the sensing information from the field to determine gaseous con-
nectivity, data aggregation, tracking latency, and energy consumption. ditions in the network. In [224], Q-learning-based task model for target
A comparative study of target tracking has been provided in [220]. Tar- tracking has been developed for dynamic WSNs. This method mainly
get tracking using divide-and-conquer approach solved in [221], target developed for task scheduling for cooperative sensor nodes.
tracking based on estimation presented in [222]and information fusion A genetic algorithm based target tracking approach with a k-
for target tracking in large-scale WSNs has been presented in [223]. Re- coverage model for WSNs has been presented in [159]. In this approach,
cently, several approaches are adopted ML techniques to improve the the genetic algorithm used to improve the network life and scheduling
accuracy of target tracking in WSNs; some of them are listed below: the sensor nodes to track the object. In [160], authors have proposed a
memetic algorithm based target coverage scheduling strategy for WSNs.
- ML approaches reduce the computational overhead to track the mo- The goal of this approach is to organize all the sensor nodes into disjoint
bile object with either stationary or mobile sensor nodes. subsets and allow one after another to cover the particular target.
- Dynamic nature of the targets in a sensor network, ML improves the
efficiency of target tracking.
- Table 12 shows ML-based target or node tracking algorithms for 3.11. Event detection
WSNs
In WSNs, sensor nodes continuously monitor the environment, ob-
In [75], authors have proposed a target tracking along with data fu- serve a phenomenon and process locally and make certain decisions. The
sion method using Bayesian posterior for WSNs. The accuracy of the feature of detecting the event or misbehavior from the data refereed as
algorithm verified based on computer-generated data as well as the event detection. The requirements to satisfy the event detection are syn-
real-time environment. This approach produces an accurate target track- chronization, low false alarm rate, and high true detection rate. Sensors
ing and data fusion for WSNs. In [76], a multi-layer dynamic Bayesian are of limited power, memory, and computational resources, Therefore
network (MDBN) has been proposed for mobile target tracking. The event detection is challenging. ML based approaches have the poten-
target tracking accuracy improved because of Bayesian statistics. This tial to address event detection problems. Some of the benefits are listed
approach also works online mobile tracking in nonlinear observation below
and time-varying RSS precision. This method promising optimal target
tracking approach compared to the conventional approaches. Authors - ML is very useful to detect an event from the complex forms of sensor
in [77] have presented a node selection algorithm for signal recovery data.
based on Bayesian learning approach for WSNs. This method efficiently - ML improves packet delivery ratio by achieving efficient duty cycles.

16
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

In [79], authors have focused to detect a different kind of events - ML for mobile sink will provide optimal rendezvous points or effi-
for variant WSN applications. The authors have suggested sophisticated cient cluster heads for data collection.
regression model to improve the accuracy of the event detection. k-NN - Mobile sink path selection place a significant role to avoid delay-
applied to extract information of interest from the raw sensor data ef- aware data collection, ML achieve delay aware routing of mobile
ficiently. In [80], k-NN based query processing approach has been pro- sink.
posed to extract information of interest from information storage and
in-network rather from the raw sensor data. This approach improves In [105] authors have determined optimal deadline-based trajectory
the query processing time and balances the energy consumption of the (ODT) algorithm for an efficient mobile sink scheduling based on dy-
sensor nodes. Authors in [81] have presented an event detection method namic rendezvous point selection and virtual structures in the WSNs us-
for sensor network using fuzzy logic. This method improves the accuracy ing decision trees. ODT performs the delay-aware data gathering from
of event detection by considering neighbor nodes data. This method also the network from active sensor node (AN) groups. The ODT obtain a fea-
applies rule-based approaches to speed up the event detection process. sible solution by running decision tree with polynomial time complex-
In [225], a human-activity recognition method has been proposed ity. This method improves the network lifetime and also delay-aware
using deep learning technique and PCA. The dimensionality has reduced path planning of mobile sink. In [88] authors have proposed a data
using PCA and deep learning approach used for training and testing the gathering methods using mobile sink based on Naïve Bayesian classifier.
Smartphone data. As compared to the traditional methods of human This method was efficiently gathering the information from the sensor
recognition approach, the algorithm provides 94.12% overall accuracy. nodes and outperforms compared to traditional approaches. Authors in
A deep learning-based projection-recovery network (PRNet) has been [232] find optimal rendezvous points and efficient mobile sink path se-
developed in [97] to estimate blind online calculations of sensor data. lection for WSNs. k-means algorithm used to perform the clusters and
PRNet approach initially focuses on drifted data in feature space; later RPs for data gathering. k-means useful to reduce the intra-cluster com-
deep learning recover drifted data from the sensor data. This approach munication. Efficient mobile sink path determined using a minimum
has achieved 2 × accurate than conventional methods. In [226], ANN- spanning tree (MST). This method provides optimal energy balance of
based groundwater quality estimation method has been proposed for sensor nodes and improves network lifetime. An energy efficient PSO
sensor networks. This method used in real-time water quality estimation based routing algorithm named EPMS has been proposed in [161] for
and provides the accurate result than traditional approaches. ANN was WSNs using a mobile sink. EPMS is the combination of virtual clustering
used to train the large set of sample to achieve accurate results. and mobile sink scheduling approach. PSO algorithm used in EPMS to
In [104], Bayesian learning-based motion detection method has been split the network into clusters and then finds the cluster heads.
presented for a moving object with different temporal dimensions in a In [162], authors have proposed an algorithm to determine the mo-
polygon region. This method also works concurrently to detect multiple bile sink path for non-uniform data constrained WSNs using Ant colony
moving objects in WSNs. This approach balance the energy consumption optimization (ACO) algorithm. In this algorithm, the authors have used
of the sensor nodes and reduces the latency. Authors in [227] have de- ACO to find the rendezvous points for collecting the data as well as
veloped a reinforcement learning-based sleep/wake-up scheduling ap- the path of the mobile sink. This approach produces optimal results
proach for energy efficient WSNs. The duty cycling approach maintained for energy consumption and network lifetime in both uniform and non-
by reinforcement learning and it has tradeoff with packet delivery delay. uniform data constraints. A mobile sink based data collection algorithm
This approach also achieves high packet delivery ratio and increases net- has been presented using discrete firefly algorithm for WSNs in [163].
work lifetime. In [228], authors have presented distributed functional The firefly algorithm has been used to find the order of visiting the sen-
tangent decision tree (DFTDT) for estimating the quality of water from sor nodes in the network to optimize the mobile sink tour. In [164],
the pond using sensors. In DFTDT, the routing determined by using ABC authors have presented an ABC based mobile sink path planning algo-
and decision trees. rithm for WSNs to collect the data efficiently. The ABC uses to find the
optimal mobile robot visiting points and minimum tour length. Authors
in [133,134] have proposed a mobile sink based data collection in WSNs
using hierarchical clustering approach. In this, initially, apply the clus-
3.12. Mobile sink tering method to identify the cluster heads and then efficient path plan-
ning algorithm designed for a mobile sink to collect the data from the
In WSNs, sensor nodes gather information from the environment and CHs. A fuzzy clustering approach has been proposed in [138] for col-
transmit the data to the base station directly or multi-hop manner. When lecting the data using mobile sink. In this approach, it selects super-CHs
the data transmits in a multi-hop manner, the node which is near to the to reduce the tour length of the mobile sink.
sink will die soon referred as energy-hole problem. To avoid the energy-
hole problem, a mobile sink concept has been introduced, a mobile sink 3.13. Energy harvesting
visits each sensor node in the network and collects information directly.
In large WSNs, visiting every node is difficult, so scheduling mobile sink Battery power is one of the significant current source of energy for
in an efficient delay-aware manner is a research issue. Therefore instead sensor nodes in WSNs, and the lifetime of the network depends on the
visiting each sensor node in the network, mobile sink visit only a few energy consumed by the sensor nodes. Most of the applications of WSNs
nodes or points in the network called rendezvous points (RPs) to collect require longer network lifetime ranges from months to years. To pro-
data, and all remaining nodes send their data to nearest RPs [229,230]. long the lifetime of WSNs, either we apply the energy-efficient proto-
We can also use multiple mobile sinks to avoid delay of mobile sink to cols or providing energy harvesting approaches for sensor nodes. Sev-
visit sensor nodes, but it is cost-effective. eral routing, sleep-scheduling, mobile sink, mobile charger etc., proto-
Authors in [231] have proposed multiple mobile sink concepts to cols have been introduced for energy-efficiency. However the energy
gather the information from the WSNs and store it into the cloud. In this, requirement is not filled because of high computational resources, un-
each mobile sink treated as a fog device, and it acts as the bridge be- reachable sensor nodes and additional maintenances. Energy harvest-
tween WSNs and cloud. This algorithm designed for parallel data gath- ing includes maintenance free, long lasting, and self-powered in WSNs.
ering process for minimizing the latency and maximizing the schedul- Energy harvesting gives uninterrupted energy for sensor nodes from an
ing efficiency. This approach balances the energy consumption and im- ambient surrounding such as radio frequency energy, wind energy, solar
proves the network lifetime. Recently, ML approaches have adopted for energy, thermal, mechanical and vibrations. Energy harvesting is cate-
WSNs to schedule mobile sink and to choose the optimal set of ren- gorized into two types, without energy storage (sensor nodes directly
dezvous points. This adoption will result from following benefits: used the power without any backup) as shown in Fig. 13(a), and pow-

17
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Table 13
ML-based energy harvesting techniques for WSNs.

ML Technique Studies Environment Complexity Energy source

Regression [82] Centralized High Solar

[83] Distributed/ centralized High Solar
Reinforcement learning [234] Centralized Moderate solar
[235] Centralized Moderate Solar
[236] Centralized Low Solar
Deep learning [98] Centralized High Wind
Hierarchical clustering [135] Distributed Low Solar/ wind

Fig. 13. Types of energy harvesting for sensor nodes (a) energy harvesting with-
out storage (b) energy harvesting with storage [233].
Fig. 14. QoS functionality for WSNs [193].

ered storage (rechargeable battery) as shown in Fig. 13(b). Several ML

vested WSNs. RLMan approach balances the energy consumption of the
based models have been used to track effective energy harvesting meth-
sensor nodes as well as maintains energy harvesting.
ods for WSNs. The summary of ML-based energy harvesting methods
A deep learning-based fault prediction and diagnose method for wind
listed in Table 13. Some of the benefits of ML-based energy harvesting
power generation in IoT has been developed in [98]. It produces high
are as follows:
accuracy rate and low error rate of fault prediction in any conditions
without human interaction. In [135], authors have focused on energy
- ML algorithm improves the performance of WSNs to forecast the
harvesting using a hierarchical clustering algorithm for heterogeneous
amount of energy to be harvested within a particular time slot [234].
WSNs. The algorithm found static cluster heads for data collection and
- ML approaches reduce the computational complexity to calculate the
these cluster heads are using renewable sources and all other sensor
amount of energy harvested and maintain appropriately by balanc-
nodes are non-renewable. The objective of this approach is to find the
ing the energy consumption.
optimal locations for the cluster headsto reduce the power consumption
Authors in [82] have investigated and evaluated solar irradiance pre- of the sensor nodes.
diction using ML-based approaches. This method performs various op-
erations on historical data sets and results out correlation coefficient, 3.14. QoS
forecast accuracy, and root mean square error (RMSE). Authors have
collected data sets from national renewable energy laboratory (NREL) The level of service given by WSNs to its users depends on the qual-
to perform the testing operations. The resultant of this algorithm pro- ity of service (QoS). QoS is mainly classified into two categories such as
vides best accuracy rate compared with other conventional methods. In network-specific and application-specific. The network-specific parame-
[83] authors have presented an indoor test methodology based on astro- ters for QoS are the energy consumption rate of sensor nodes and band-
nomical models and PV (photovoltaic) cell design principles for solar- width. The application-specific parameters are measurements of sensor
powered WSNs. The experiments were tested for both centralized and nodes, deployment and the number of sensor nodes is active in the net-
distributed networks using linear regression based methods. The result work. There are several challenges to maintain the QoS for WSNs are se-
of this algorithm shown to perform better for distributed approaches vere resource constraints, unbalanced traffic, data redundancy, dynamic
than the centralized strategies. network, energy balancing, scalability, multiple sinks, the difference in
A Q-learning-based solar energy prediction (Q-SEP) approach has traffic types [193]. Fig. 14 demonstrate the QoS functionality for WSNs.
been proposed in [234] for energy harvesting in WSNs. Q-leaning was The summary of ML-based approaches for QoS in WSNs is shown in
used efficiently for prediction based on past observations. The algorithm Table 14.
improves the performance of WSNs to forecast the amount of energy to In [237], authors have proposed an energy efficient method for data
be harvested within a particular time slot. In [235], a reinforcement fusion based on fuzzy logic to achieve QoS in WSNs. This method ag-
learning-based method has been presented for energy harvesting WSNs. gregate only true information, instead of gathering complete data from
The algorithm automatically adjusts duty cycles according to the learn- the WSNs. The algorithm is applicable for cluster-based sensor networks
ing process of reinforcement learning. This method outperforms as com- where each CH capture information from its members until an event
pared to conventional approaches. In [236], reinforcement learning- detects. The fuzzy logic controller used in every sensor node to find
based energy management (RLMan) have been proposed for energy har- the true information before it transmits to the nearest CHs. Authors in

18
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Table 14
Summary of ML-based QoS for WSNs.

ML Technique Studies Complexity Objective

ANN [44] Moderate Faulty node detection

[237] Moderate Data fusion and energy balancing
[238] Low Link-quality estimation
Reinforcement learning [239] Moderate Cross-layer communication framework
[240] Low Topology control and data dissemination protocol
[241] High Constraint-satisfied service composition
[242] Moderate Distributed adaptive cooperative routing

[238] have presented a wavelet neural-network-based link quality es- inputs, and it does not impose restrictions on input data. ANN is also
timation (WNN-LQE) mechanism to enhance the QoS requirements in handles the independent attributes efficiently as compared to others.
WSNs and smart grids. This method estimates different link quality pa- Reinforcement learning approach does not require any prior knowl-
rameters by using the signal to noise ratio (SNR). This method also im- edge. It can efficiently balance the exploration and exploitation, whereas
proves the QoS by reducing the communication overhead and balanc- most of the supervised learning algorithms fail. By interacting with the
ing the energy consumptions. A cross-layer communication protocol has environment, according to the things happen in the network, it takes
been presented in [239] based on multi-agent reinforcement learning for decisions accordingly. The computational complexity is low compare to
WSNs. most of the supervised learning approaches. This approach reduce the
In [240], a multi-agent RL and energy-aware topology control and energy consumption and improves the lifetime of the network. From the
data dissemination protocols have been developed for effective self- literature, we found that SVM also given priority to resolve the various
organization of WSNs. The multi-agent RL algorithm used to select the issues in WSNs. It attracts the researcher even for slow learning process;
active neighbor nodes to maintain the reliable topology. The connectiv- because it produces high accuracy rate and perform the classification
ity and coverage for the boundary nodes sustained through a convex-hull faster. SVM provides good results with labeled, semi-labeled or unla-
algorithm. This method provides the optimal QoS and improves the per- beled data, without any restrictions.
formance of the sensor network. A Q-leaning based mechanism has been Regression is attracting the WSNs researchers, because of its rapid
used in [241] to solve the uncertainty of QoS and dynamic behavior of a prediction strategy, efficient decision support, error correctness, and less
service for WSNs. This method achieves globally optimal for the safety computational requirements. It produces accurate results for both large
service. Authors in [242] have proposed distributed adaptive cooper- and small applications. The dimensionality reduction is the necessary
ative routing protocols (DACR) based on light-weighted reinforcement requirement, because it avoids unnecessary data transmissions in the
learning mechanism, to achieve reliable QoS for WSNs. The reinforce- networks. PCA and ICA are used to reduce the dimensionality of the
ment learning algorithm select optimal relay nodes in the network and data in the network for improving the network lifetime. The less effort
decides transmission mode at each node to maximize the reliability. had done with the Random forest approach in the literature; because
this algorithm fits and produce a better result to the large-scale WSNs.
4. Statistical analysis and limitations Because of the large data, it requires additional space to store the data
and consumes more time to computation. Simulation of large-scale net-
In this section, we discuss the statistical analysis of recent research works require high computational machines to prove the results. How-
topics for ML-based algorithms for WSNs and their limitations. ever, there are several benefits with Random forests such as quick train-
ing, implicit feature selections, less variance and highly accurate for the
large-scale WSNs.
4.1. Statistical analysis
4.2. Limitations
Here, we present statistical charts to show the overview of recent
research on ML-based algorithms for WSNs, as shown in Fig. 15. From
In spite of many benefits, there are some limitations of applying ML
Fig. 15(a) and Figure. 15(b), we notice that the most researchers focus
in WSNs:
on data aggregation, localization, routing, anomaly and fault detection,
event detection, coverage,and connectivity. In contrast, a little amount - ML techniques do not produce immediate accurate predictions, be-
of research has been done by using ML to solve MAC, target tracking, cause they require to learn from historical data. The performance of
QoS, energy harvesting, congestion control, synchronization and mobile the system depends on the amount of historical data. If the data size
sink scheduling. is large then energy consumed to process the data is also very high.
From Fig. 15(c), we estimate that most of the WSNs issues solved us- In other words, there is tradeoff between energy limitations of the
ing supervised learning algorithms. The supervised learning approaches WSNs and high computation complexities of the ML algorithm. To
are solved 67% of the WSNs issues in recent times. Unsupervised learn- overcome this tradeoff, ML algorithm are need to run centrally.
ing approaches have solved 18% of the WSNs problems and finally, rein- - Validating the predictions that produced by the ML algorithm in the
forcement learning approach has solved 15% of the problems of WSNs real-time environment is a cumbersome task.
in recent years. From Fig. 15(d), we see that most of the WSNs issue - Sometimes identifying a particular ML technique to address an issue
solved by using the Bayesian learning approach. The reasons are that in WSNs is very difficult.
the Bayesian learning efficiently combine the prior information with the
5. Open issues
current data. It produces solid decisions with more than 95% of confi-
dence interval. It also handle the parameters successfully, noise, missing
Several challenges still open and require further research in WSNs.
values effectively, and classifies the data rapidly. Bayesian learning ap-
Here we listed out some of the open research issues for WSNs that can
proach allows different probability functions specific to the problem in
be solvable using ML approaches. The summary of ML choice to be solve
WSNs. Table 2. Summarize the benefits of this approach with additional
a issue in WSNs are listed in Table 15.
specifications. Because of the faster classification and high accuracy rate
in the result, the ANN learning approach also attracted the researcher to 1. Localization: It is essential to design efficient path planning for the
address various in WSNs. It can find the unseen relationships from the beacon nodes in the context of mobile sensor nodes. To the best

19
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Fig. 15. Statistical charts (a) Issues in WSNs addressed by ML algorithms (2014–March 2018) (b) Year wise research articles (c) Classification of ML algorithms for
WSNs (d) ML-based algorithms for WSNs.

of our knowledge, there is no particular predefined path planning cur. Predicting such situations in the network and finding the reme-
strategy available for mobile anchor nodes. Further, ML may pro- dies require further research. Most of the current researchers have
vide efficient path planning approaches for each anchor nodes in developed coverage algorithms for two-dimensional space, however,
the sensor networks to achieve higher localization accuracy with three-dimensional coverage with optimal computational complexity
minimum energy consumption. Most of the real-time applications is still unexplored.
deploys sensor nodes in three-dimensional space, but most of the 3. Anomaly detection: Anomaly detection is one of the most promising
existing localization algorithms examined only for two-dimensional research issues in WSNs, and recently most of the researchers have
space. Therefore, there is a need to develop localization techniques developed various techniques to detect anomalies. Anomalies in
for three-dimensional spaces for static and mobile WSNs. WSNs effects communication overhead, transmission delays, or some
2. Coverage and Connectivity: A number of recent works have been pro- time it misleads the sensor nodes data. Authors in [32–38,111,175–
posed for coverage and connectivity, still a lot of problems and chal- 179,181–186] focused on anomaly detection, but detection of the
lenges need to be explored. Predicting the minimum number of sen- anomaly itself is not a complete solution, further research is needed,
sor nodes require to cover an area of interest or target, and finding its on what actions need to be taken once the anomaly detects, what
optimal placement of sensor nodes is also a challenging issue. Most of are the efforts required to reduce the damages. Anomaly detection
the real-time WSNs applications, the deployment of the sensor nodes method is application specific; therefore the selection of anomaly
are random. This random deployment may cause the coverage hole. detection algorithm for heterogeneous WSNs is a challenging task.
Finding such a coverage hole in the network is a challenging task. The detection algorithm needs to fulfill the accuracy and speed of
Due to the dynamic changes in the network, coverage holes may oc- the detection.

20
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

Table 15
ML techniques to solve various issues in WSNs.

Sno WSNs Issues ML techniques Remarks

1 Localization Reinforcement learning Prior knowledge not required and solution works even for the dynamic environment
k-NN Efficient distance estimation for range free localization
2 Coverage & Connectivity Decision tree Efficient classification of connected or isolated nodes in the network
Deep learning To find the minimum number of sensor
Evolutionary computation nodes for well covered area with optimal connectivity
3 Anomaly and fault detection Random forest To classify the faulty sensor nodes from the normal nodes
PCA To detect anomaly in the network
ICA
Deep learning Online anomaly or fault detection
4 Routing Decision tree To predict the optimal routing paths through dynamic alternative path selection to control the data traffic
Random forest
Evolutionary computation
5 MAC SVM Efficient channel assignment
Decision tree
Deep learning Reconfiguring newly joined sensor nodes and predict time slots
6 Data aggregation k-means To find optimal clusters heads in the network
SVM
Reinforcement learning To identify the optimal data routing paths within the network without prior knowledge
7 Synchronization Deep learning Predict the efficient time slots for channel allocation and resynchronize dynamically
8 Congestion control Reinforcement learning To predict the congestion locations in the network, and find the optimal alternative routing paths
Random forest Classifying the congestion nodes from the normal nodes in large-scale WSNs.
Decision tree
SVM
Evolutionary computation To find optimal dynamic alternative path selection for congestion avoidance
PCA Dimensionality reduction to control unnecessary data transmission
ICA
9 Target tracking Deep learning Efficient multiple target tracking for mobile WSNs
SVM To classify the targets in heterogeneous WSNs
Decision tree
10 Event detection PCA To detect an event from the complex forms of sensor data
ICA
Deep learning Efficient duty cycling management
Evolutionary computation
11 Mobile sink Evolutionary computation To select optimal mobile sink path between sensor nodes or rendezvous points.
Reinforcement learning Selecting optimal rendezvous points, and optimal tour selection
Random forest To find the data forwarding routes and optimal RPs selection for large-scale networks
12 Energy harvesting SVM To forecast the amount of energy to be harvested within a particular time slot
Deep learning
Evolutionary computation To predict the amount of energy to be harvested.

4. Routing: Most of the existing routing mechanisms are developed for decentralized strategies. Congestion control should adopt the self-
collecting data from a single source and transmitting it to a single learning approach to respond according to the dynamic changes in
destination. WSNs with multiple source and multiple destinations, the network. In self-adoption of congestion control mechanism, it
there may be a possibility of packet collision. Developing collision must respond to remove or add nodes while congestion detected
free cooperative routing protocols for a WSNs with multiple source in the network. There is a need to develop traffic estimation pro-
and multiple targets is an emerging research interest. In a mobile tocols to identify the rapid and dynamic change in route to avoid
WSNs, the position of the nodes changes dynamically. Sometimes congestion in the network. Further need for efficient mobile agent
due to the external causes the nodes in a WSNs may change their approaches to collect the data from sensor nodes rather transmitting
position. Therefore, there is a requirement to develop routing proto- data between the nodes.
cols that adopt the dynamic changes in the network. 7. Energy harvesting: In WSNs, the tiny sensor nodes deployed in the en-
5. Data aggregation: Most of the researchers interested in data aggre- vironment with a limited energy source (battery). Due to limited en-
gation methods for WSNs with uniform data rate sensor nodes. The ergy resources of the sensor nodes, it should need a low-cost, highly
further research is required in the WSNs with non-uniform data rate efficient, small wireless harvesting system (WHS) for operating the
sensor nodes. The data collection is more complex in mobile WSNs network fora long time. There is a need of efficient cross-layer wire-
with the non-uniform sensor nodes. Data gathering and energy effi- less energy harvesting protocols for WSNs. Majority of the existing
ciency can improve by introducing the mobile sink. Scheduling the routing protocols are energy efficient, therefore it should revisit the
mobile sink in a WSNs with non-uniform data is a challenging task. reliability routing protocols using ML. For the large-scale sensor net-
The major goals to consider the efficient data aggregation process works there should be a self-charging and discharging cycles accord-
with mobile sink are energy efficient, scalability and low-cost. ing to the dynamic changes in the environment. Efficient power al-
6. Congestion control and avoidance: Due to the dynamic behavior of location strategies needed to improve the network lifetime. So, there
WSNs operations, efficient and effective congestion control and is a need for synchronization between physical layer power control
avoidance approaches should require robustness and phenomenal and MAC layer to adjust the duty cycles.
survivability to internal disturbances and external stimuli or data 8. QoS: The goal of QoS in WSNs needs to meet the requirements of
loss [214]. Due to the limited energy and memory constraints of users and applications. QoS standards vary from different require-
WSNs, it should require to implement simple congestion control ments (such as sensor type, data rate, traffic handling methods, or
mechanisms at each node level and minimize the transmission rate data types) or application for WSNs. Therefore, defining the stan-
between the nodes. WSNs require a faster, efficient and effective dards of QoS for different requirement or application is challenging.
congestion control and avoidance mechanism for autonomous and Designing efficient cross-layer protocols may improve the QoS. For

21
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

a large heterogeneous mobile sensor network, developing QoS stan- [18] Z. Wang, H. Zhang, T. Lu, Y. Sun, X. Liu, A new range-free localisation in wireless
dards is a challenging issue. sensor networks using support vector machine, Int. J. Electron. 105 (2) (2018)
244–261.
[19] F. Zhu, J. Wei, Localization algorithm for large scale wireless sensor networks based
6. Conclusion on fast-SVM, Wireless Personal Commun. 95 (3) (2017) 1859–1875.
[20] J. Hong, T. Ohtsuki, Signal eigenvector-based device-free passive localization using
array sensor, IEEE Trans. Veh. Technol. 64 (4) (2015) 1354–1363.
In this survey, we have presented recently published ML-based algo- [21] T.L.T. Nguyen, F. Septier, H. Rajaona, G.W. Peters, I. Nevat, Y. Delignon, A bayesian
rithms for WSNs. We have discussed various ML algorithms briefly for perspective on multiple source localization in wireless sensor networks, IEEE Trans.
reader convenience. We have highlighted various issues in WSNs ad- Signal Process. 64 (7) (2016) 1684–1699.
[22] B. Sun, Y. Guo, N. Li, D. Fang, Multiple target counting and localization using vari-
dressed by ML techniques such as localization, anomaly detection, fault ational Bayesian EM algorithm in wireless sensor networks, IEEE Trans. Commun.
nodes detection, routing, data aggregation, MAC protocols, synchroniza- 65 (7) (2017) 2985–2998.
tion, congestion control, energy harvesting and mobile sink path deter- [23] Z. Wang, H. Liu, S. Xu, X. Bu, J. An, Bayesian device-free localization and tracking
in a binary RF sensor network, Sensors 17 (5) (2017) 1–21.
mination. In addition, we have discussed issues like synchronization, [24] Z. Xiahou, X. Zhang, Adaptive localization in wireless sensor network through
congestion control and energy harvesting which are not included in any bayesian compressive sensing, Int. J. Distrib. Sens. Netw. 2015 (2015) 1–15.
previous survey papers. Besides, we have compared and summarized the [25] Y. Guo, B. Sun, N. Li, D. Fang, Variational bayesian inference-based counting and
localization for off-grid targets with faulty prior information in wireless sensor
ML-based algorithm for WSNs in the tabular form. The statistical charts
networks, IEEE Trans. Commun. 66 (3) (2018) 1273–1283.
have summarized the impact of recent research on ML-based algorithm [26] W. Sun, X. Yuan, J. Wang, Q. Li, L. Chen, D. Mu, End-to-end data delivery reliabil-
for WSNs followed sugestions to chose a particular ML techniques to ity model for estimating and optimizing the link quality of industrial WSNs, IEEE
Trans. Autom. Sci. Eng. 15 (3) (2018) 1127–1137.
address a issues in WSNs. Finally, we have discussed some of the open
[27] X. Chang, J. Huang, S. Liu, G. Xing, H. Zhang, J. Wang, L. Huang, Y. Zhuang, Accu-
issues. racy-aware interference modeling and measurement in wireless sensor networks,
IEEE Trans. Mob. Comput. 15 (2) (2016) 278–291.
[28] W. Kim, M.S. Stankovi, K.H. Johansson, H.J. Kim, A distributed support vector
Acknowledgments
machine learning over wireless sensor networks, IEEE Trans. Cybern 45 (11) (2015)
2599–2611.
We appreciate the time and efforts made by the editor during re- [29] J. Shu, S. Liu, L. Liu, L. Zhan, G. Hu, Research on link quality estimation mechanism
viewing this manuscript. We pay our sincere thanks to the esteemed for wireless sensor networks based on support vector machine, Chin. J. Electr. 26
(2) (2017) 377–384.
reviewers for their valuable comments and suggestions to improve the [30] W. Elghazel, K. Medjaher, N. Zerhouni, J. Bahi, A. Farhat, C. Guyeux, M. Hakem,
quality of this survey paper. Random forests for industrial device functioning diagnostics using wireless sensor
networks, in: Aerospace Conference, 2015 IEEE, IEEE, 2015, pp. 1–9.
[31] B. Yang, Y. Lei, B. Yan, Distributed multi-human location algorithm using naive
Supplementary material bayes classifier for a binary pyroelectric infrared sensor tracking system, IEEE Sens.
J. 16 (1) (2016) 216–223.
Supplementary material associated with this article can be found, in [32] M. Xie, J. Hu, S. Han, H.-H. Chen, Scalable hypergrid k-NN-based online anomaly
detection in wireless sensor networks, IEEE Trans. Parallel Distrib. Syst. 24 (8)
the online version, at doi:10.1016/j.inffus.2018.09.013. (2013) 1661–1670.
[33] A. Garofalo, C. Di Sarno, V. Formicola, Enhancing intrusion detection in wireless
References sensor networks through decision trees, Dependable Comput. (2013) 1–15.
[34] Z. Feng, J. Fu, D. Du, F. Li, S. Sun, A new approach of anomaly detection in wireless
[1] M.A. Alsheikh, S. Lin, D. Niyato, H.-P. Tan, Machine learning in wireless sensor sensor networks using support vector data description, Int. J. Distrib. Sens. Netw.
networks: algorithms, strategies, and applications, IEEE Commun. Surv. Tutorials 13 (1) (2017) 1–14.
16 (4) (2014) 1996–2018. [35] N. Shahid, I.H. Naqvi, S.B. Qaisar, One-class support vector machines: analysis of
[2] P. Rawat, K.D. Singh, H. Chaouchi, J.M. Bonnin, Wireless sensor networks: a survey outlier detection for wireless sensor networks in harsh environments, Artif. Intell.
on recent developments and potential synergies, J. Supercomput 68 (1) (2014) Rev. 43 (4) (2015) 515–563.
1–48. [36] H.S. Emadi, S.M. Mazinani, A novel anomaly detection algorithm using DBSCAN
[3] I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, E. Cayirci, Wireless sensor networks: and SVM in wireless sensor networks, Wireless Personal Commun. 98 (2) (2018)
a survey, Comput. Networks 38 (4) (2002) 393–422. 2025–2035.
[4] J. Yick, B. Mukherjee, D. Ghosal, Wireless sensor network survey, Comput. Net- [37] C. Titouna, M. Aliouat, M. Gueroui, Outlier detection approach using bayes clas-
works 52 (12) (2008) 2292–2330. sifiers in wireless sensor networks, Wireless Personal Commun. 85 (3) (2015)
[5] D.K. Sah, T. Amgoth, Parametric survey on cross-layer designs for wireless sensor 1009–1023.
networks, Comput Sci. Rev. 27 (2018) 112–134. [38] R. Feng, X. Han, Q. Liu, N. Yu, A credible bayesian-based trust management scheme
[6] J.B. Predd, S.B. Kulkarni, H.V. Poor, Distributed learning in wireless sensor net- for wireless sensor networks, Int. J. Distrib. Sens. Netw. 11 (11) (2015) 1–9.
works, IEEE Signal Process. Mag. 23 (4) (2006) 56–69. [39] S. Zidi, T. Moulahi, B. Alaya, Fault detection in wireless sensor networks through
[7] T.M. Mitchell, Machine Learning, first ed., McGraw-Hill, Inc., New York, NY, USA, SVM classifier, IEEE Sens. J. 18 (1) (2018) 340–347.
1997. [40] M. Jiang, J. Luo, D. Jiang, J. Xiong, H. Song, J. Shen, A cuckoo search-support
[8] T.O. Ayodele, Introduction to Machine Learning, first ed, InTech, 2010. vector machine model for predicting dynamic measurement errors of sensors, IEEE
[9] P. Langley, H.A. Simon, Applications of machine learning and rule induction, Com- Access 4 (2016) 5030–5037.
mun. ACM 38 (11) (1995) 54–64. [41] H. Zhang, J. Liu, N. Kato, Threshold tuning-based wearable sensor fault detection
[10] S.S. Banihashemian, F. Adibnia, M.A. Sarram, A new range-free and storage-ef- for reliable medical monitoring using bayesian network model, IEEE Syst. J. 12 (2)
ficient localization algorithm using neural networks in wireless sensor networks, (2017) 1886–1896.
Wireless Personal Commun. 98 (1) (2018) 1547–1568. [42] C. Titouna, M. Aliouat, M. Gueroui, FDS: Fault detection scheme for wireless sensor
[11] A. El Assaf, S. Zaidi, S. Affes, N. Kandil, Robust ANNs-based WSN localization in networks, Wireless Personal Commun. 86 (2) (2016) 549–562.
the presence of anisotropic signal attenuation, IEEE Wireless Commun. Lett. 5 (5) [43] W. Meng, W. Li, Y. Xiang, K.-K.R. Choo, A bayesian inference-based detection mech-
(2016) 504–507. anism to defend medical smartphone networks against insider attacks, J. Netw.
[12] S.K. Gharghan, R. Nordin, M. Ismail, J.A. Ali, Accurate wireless sensor localiza- Comput. Appl. 78 (2017) 162–169.
tion technique based on hybrid PSO-ANN algorithm for indoor and outdoor track [44] P. Chanak, I. Banerjee, Fuzzy rule-based faulty node classification and manage-
cycling, IEEE Sens. J. 16 (2) (2016) 529–541. ment scheme for large scale wireless sensor networks, Expert Syst. Appl. 45 (2016)
[13] S. Phoemphon, C. So-In, D.T. Niyato, A hybrid model using fuzzy logic and an 307–321.
extreme learning machine with vector particle swarm optimization for wireless [45] W. Li, P. Yi, Y. Wu, L. Pan, J. Li, A new intrusion detection system based on KNN
sensor network localization, Appl. Soft Comput. 65 (2018) 101–120. classification algorithm in wireless sensor network, J. Electr. Comput. Eng. 2014
[14] N. Baccar, R. Bouallegue, Interval type 2 fuzzy localization for wireless sensor net- (2014) 1–9.
works, EURASIP J. Adv. Signal Process. 2016 (1) (2016) 1–13. [46] A. Mehmood, Z. Lv, J. Lloret, M.M. Umar, ELDC: An artificial neural network based
[15] J. Kang, Y.J. Park, J. Lee, S.H. Wang, D.S. Eom, Novel leakage detection by en- energy-efficient and robust routing scheme for pollution monitoring in WSNs, IEEE
semble CNN-SVM and graph-based localization in water distribution systems, IEEE Trans. Emerg. Top Comput. PP (99) (2017) 1–8.
Trans. Ind. Electron. 65 (5) (2018) 4279–4289. [47] M.S. Gharajeh, S. Khanmohammadi, DFRTP: Dynamic 3D fuzzy routing based on
[16] F. Zhu, J. Wei, Localization algorithm for large-scale wireless sensor networks traffic probability in wireless sensor networks, IET Wireless Sens. Syst. 6 (6) (2016)
based on FCMTSR–support vector machine, Int. J. Distrib. Sens. Netw. 12 (10) 211–219.
(2016) 1–12. [48] J.R. Srivastava, T. Sudarshan, A genetic fuzzy system based optimized zone based
[17] T. Tang, H. Liu, H. Song, B. Peng, Support vector machine based range-free lo- energy efficient routing protocol for mobile sensor networks (OZEEP), Appl. Soft
calization algorithm in wireless sensor network, in: International Conference on Comput. 37 (2015) 863–886.
Machine Learning and Intelligent Communications, Springer, 2016, pp. 150–158.

22
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

[49] Y. Lee, Classification of node degree based on deep learning and routing method query algorithm leveraging grid division routing in wireless sensor networks of
applied for virtual route assignment, Ad Hoc Netw. 58 (2017) 70–85. skewness distribution, Pers. Ubiquitous Comput. 18 (8) (2014) 1989–2001.
[50] F. Khan, S. Memon, S.H. Jokhio, Support vector machine based energy aware rout- [82] A. Sharma, A. Kakkar, Forecasting daily global solar irradiance generation using
ing in wireless sensor networks, in: Robotics and Artificial Intelligence (ICRAI), machine learning, Renewable Sustainable Energy Rev. 82 (2018) 2254–2269.
2016 2nd International Conference on, IEEE, 2016, pp. 1–4. [83] W.M. Tan, P. Sullivan, H. Watson, J. Slota-Newson, S.A. Jarvis, An indoor test
[51] V. Jafarizadeh, A. Keshavarzi, T. Derikvand, Efficient cluster head selection using methodology for solar-powered wireless sensor networks, ACM Trans. Embedded
naïve bayes classifier for wireless sensor networks, Wireless Netw. 23 (3) (2017) Comput. Syst. (TECS) 16 (3) (2017) 82.1–82.25.
779–785. [84] D.C. Montgomery, E.A. Peck, G.G. Vining, Introduction to Linear Regression Anal-
[52] Z. Liu, M. Zhang, J. Cui, An adaptive data collection algorithm based on a bayesian ysis, 821, John Wiley & Sons, 2012.
compressed sensing framework, Sensors 14 (5) (2014) 8330–8349. [85] W. Zhao, S. Su, F. Shao, Improved DV-hop algorithm using locally weighted linear
[53] F. Kazemeyni, O. Owe, E.B. Johnsen, I. Balasingham, Formal modeling and analysis regression in anisotropic wireless sensor networks, Wireless Personal Commun. 98
of learning-based routing in mobile wireless sensor networks, in: Integration of (4) (2018) 3335–3353.
Reusable Systems, Springer, 2014, pp. 127–150. [86] C.G. Emily Fox, The Simple Linear Regression Model, Machine Learning : Regres-
[54] B. Alotaibi, K. Elleithy, A new MAC address spoofing detection technique based on sion, Coursera, 2018.
random forests, Sensors 16 (3) (2016) 1–14. [87] J.R. Quinlan, Induction of decision trees, Mach. Learn 1 (1) (1986) 81–106.
[55] X. Song, C. Wang, J. Gao, X. Hu, DLRDG: distributed linear regression-based hi- [88] S. Kim, D.-Y. Kim, Efficient data-forwarding method in delay-tolerant P2P network-
erarchical data gathering framework in wireless sensor network, Neural Comput. ing for IoT services, in: Peer-to-Peer Networking and Applications, 2017, pp. 1–10.
Appl. 23 (7–8) (2013) 1999–2013. [89] L. Breiman, Random forests, Mach. Learn 45 (1) (2001) 5–32.
[56] I. Atoui, A. Makhoul, S. Tawbe, R. Couturier, A. Hijazi, Tree-based data aggrega- [90] M. Belgiu, L. Drăguţ, Random forest in remote sensing: a review of applications
tion approach in periodic sensor networks using correlation matrix and polynomial and future directions, ISPRS J. Photogramm. Remote Sens. 114 (2016) 24–31.
regression, in: Computational Science and Engineering (CSE) and IEEE Intl Confer- [91] S.S. Haykin, S.S. Haykin, S.S. Haykin, S.S. Haykin, Neural networks and learning
ence on Embedded and Ubiquitous Computing (EUC) and 15th Intl Symposium on machines, vol. 3, Pearson Upper Saddle River, NJ, USA, 2009.
Distributed Computing and Applications for Business Engineering (DCABES), 2016 [92] H. White, Learning in artificial neural networks: a statistical perspective, Neural
IEEE Intl Conference on, IEEE, 2016, pp. 716–723. Comput. 1 (4) (1989) 425–464.
[57] L. Gispan, A. Leshem, Y. Be’ery, Decentralized estimation of regression coefficients [93] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (7553) (2015) 436–444.
in sensor networks, Digit Signal Process. 68 (2017) 16–23. [94] A.H. Marblestone, G. Wayne, K.P. Kording, Toward an integration of deep learning
[58] Y. Li, L.E. Parker, Nearest neighbor imputation using spatial–temporal correlations and neuroscience, Front. Comput. Neurosci. 10 (2016) 1–41.
in wireless sensor networks, Inf. Fusion 15 (2014) 64–79. [95] T. Ma, F. Wang, J. Cheng, Y. Yu, X. Chen, A hybrid spectral clustering and deep neu-
[59] F. Edwards-Murphy, M. Magno, P.M. Whelan, J. OHalloran, E.M. Popovici, b+ ral network ensemble algorithm for intrusion detection in sensor networks, Sensors
WSN: smart beehive with preliminary decision tree analysis for agriculture and 16 (10) (2016) 1–23.
honey bee health monitoring, Comput. Electron. Agric. 124 (2016) 211–219. [96] C. Li, X. Xie, Y. Huang, H. Wang, C. Niu, Distributed data mining based on deep neu-
[60] H. He, Z. Zhu, E. Mäkinen, Task-oriented distributed data fusion in autonomous ral network for wireless sensor network, Int. J. Distrib. Sens. Netw. 11 (7) (2015)
wireless sensor networks, Soft Comput. 19 (8) (2015) 2305–2319. 1–7.
[61] C. Habib, A. Makhoul, R. Darazi, C. Salim, Self-adaptive data collection and fusion [97] Y. Wang, A. Yang, X. Chen, P. Wang, Y. Wang, H. Yang, A deep learning ap-
for health monitoring based on body sensor networks, IEEE Trans. Ind. Inf. 12 (6) proach for blind drift calibration of sensor networks, IEEE Sens. J. 17 (13) (2017)
(2016) 2342–2352. 4158–4171.
[62] H. Yang, S. Fong, R. Wong, G. Sun, Optimizing classification decision trees by using [98] F. Chen, Z. Fu, Z. Yang, Wind power generation fault diagnosis based on deep
weighted naïve bayes predictors to reduce the imbalanced class problem in wireless learning model in internet of things (IoT) with clusters, Cluster Comput. (2018)
sensor network, Int. J. Distrib. Sens. Netw. 9 (1) (2013) 1–15. 1–13.
[63] A. De Paola, P. Ferraro, S. Gaglio, G.L. Re, S.K. Das, An adaptive bayesian system [99] V.N. Vapnik, An overview of statistical learning theory, IEEE Trans. Neural Net-
for context-aware data fusion in smart environments, IEEE Trans. Mob. Comput. works 10 (5) (1999) 988–999.
16 (6) (2017) 1502–1515. [100] M.R. Islam, J. Uddin, J.-M. Kim, Acoustic emission sensor network based fault di-
[64] S. Hwang, R. Ran, J. Yang, D.K. Kim, Multivariated bayesian compressive sensing agnosis of induction motors using a gabor filter and multiclass support vector ma-
in wireless sensor networks, IEEE Sens. J. 16 (7) (2015) 2196–2206. chines., Adhoc Sensor Wireless Netw 34 (2016) 273–287.
[65] Y. Dingcheng, W. Zhenghai, X. Lin, Z. Tiankui, Online bayesian data fusion in en- [101] Q.-y. Sun, Y.-m. Sun, X.-j. Liu, Y.-x. Xie, X.-g. Chen, Study on fault diagnosis algo-
vironment monitoring sensor networks, Int. J. Distrib. Sens. Netw. 10 (4) (2014) rithm in WSN nodes based on RPCA model and SVDD for multi-class classification,
945894. Cluster Comput. (2018) 1–15.
[66] C. Wang, E. Bertino, Sensor network provenance compression using dynamic [102] F.V. Jensen, An Introduction to Bayesian Networks, 210, UCL Press London, 1996.
bayesian networks, ACM Trans. Sensor Netw. (TOSN) 13 (1) (2017) 5.1–5.32. [103] S.N. Das, S. Misra, B.E. Wolfinger, M.S. Obaidat, Temporal-correlation-aware dy-
[67] Y. Wang, A. Yang, Z. Li, X. Chen, P. Wang, H. Yang, Blind drift calibration of sensor namic self-management of wireless sensor networks, IEEE Trans. Ind. Inf. 12 (6)
networks using sparse bayesian learning, IEEE Sens. J. 16 (16) (2016) 6249–6260. (2016) 2127–2138.
[68] D. Capriglione, D. Casinelli, L. Ferrigno, Analysis of quantities influencing the per- [104] B. Avci, G. Trajcevski, R. Tamassia, P. Scheuermann, F. Zhou, Efficient detection
formance of time synchronization based on linear regression in low cost WSNs, of motion-trend predicates in wireless sensor networks, Comput. Commun. 101
Measurement 77 (Suppl. C) (2016) 105–116. (Supplement C) (2017) 26–43.
[69] J.J. Prez-Solano, S. Felici-Castell, Adaptive time window linear regression algo- [105] F. Tashtarian, M.Y. Moghaddam, K. Sohraby, S. Effati, ODT: optimal deadline-based
rithm for accurate time synchronization in wireless sensor networks, Ad Hoc Netw. trajectory for mobile sinks in WSN: a decision tree and dynamic programming ap-
24 (Part A) (2015) 92–108. proach, Comput. Networks 77 (2015) 128–143.
[70] G. Betta, D. Casinelli, L. Ferrigno, Some Notes on the Performance of Regres- [106] S.A. Dudani, The distance-weighted k-nearest-neighbor rule, IEEE Trans. Syst. Man
sion-based Time Synchronization Algorithms in Low Cost WSNs, Springer Inter- Cybern. SMC-6 (4) (1976) 325–327.
national Publishing, Cham, 2015. [107] L.E. Peterson, K-nearest neighbor, Scholarpedia 4 (2) (2009) 1883.
[71] J.J. Pérez-Solano, S. Felici-Castell, Improving time synchronization in wireless sen- [108] J.M. Keller, M.R. Gray, J.A. Givens, A fuzzy k-nearest neighbor algorithm, IEEE
sor networks using Bayesian inference, J. Netw. Comput. Appl. 82 (2017) 47–55. Trans. Syst. Man Cybern. SMC-15 (4) (1985) 580–585.
[72] M.A. Alsheikh, S. Lin, D. Niyato, H.P. Tan, Rate-distortion balanced data compres- [109] S.B. Kotsiantis, I. Zaharakis, P. Pintelas, Supervised machine learning: a review of
sion for wireless sensor networks, IEEE Sens. J. 16 (12) (2016) 5072–5083. classification techniques, Emerg. Artif. Intell. Appl. Comput. Eng. 160 (2007) 3–24.
[73] A.A. Rezaee, F. Pasandideh, A fuzzy congestion control protocol based on active [110] J. Qin, W. Fu, H. Gao, W.X. Zheng, Distributed k-means algorithm and fuzzy
queue management in wireless sensor networks with medical applications, Wireless c-means algorithm for sensor networks based on multiagent consensus theory, IEEE
Personal Commun. 98 (1) (2018) 815–842. Trans. Cybern. 47 (3) (2017) 772–783.
[74] M. Gholipour, A.T. Haghighat, M.R. Meybodi, Hop-by-Hop congestion avoidance [111] P. Gil, H. Martins, F. Januário, Outliers detection methods in wireless sensor net-
in wireless sensor networks based on genetic support vector machine, Neurocom- works, Artif. Intell. Rev. (2018) 1–26.
puting 223 (2017) 63–76. [112] R. El Mezouary, A. Choukri, A. Kobbane, M. El Koutbi, An energy-aware clustering
[75] P. Braca, P. Willett, K.D. LePage, S. Marano, V. Matta, Bayesian tracking in under- approach based on the K-means method for wireless sensor networks, in: Advances
water wireless sensor networks with port-starboard ambiguity., IEEE Trans. Signal in Ubiquitous Networking, Springer, 2016, pp. 325–337.
Process. 62 (7) (2014) 1864–1878. [113] A. Ray, D. De, Energy efficient clustering protocol based on k-means (EECP-
[76] B. Zhou, Q. Chen, T.J. Li, P. Xiao, Online variational bayesian filtering-based mobile K-means)-midpoint algorithm for enhanced network lifetime in wireless sensor net-
target tracking in wireless sensor networks, Sensors 14 (11) (2014) 21281–21315. work, IET Wireless Sens. Syst. 6 (6) (2016) 181–191.
[77] B. Xue, L. Zhang, W. Zhu, Y. Yu, A new sensor selection scheme for bayesian learn- [114] B. Jain, G. Brar, J. Malhotra, EKMT-k-means clustering algorithmic solution for
ing based sparse signal recovery in WSNs, J. Franklin Inst. (2017) 1–21. low energy consumption for wireless sensor networks based on minimum mean
[78] H. Chen, R. Wang, L. Cui, L. Zhang, EasiDSlT: a two-layer data association method distance from base station, in: Networking Communication and Data Knowledge
for multitarget tracking in wireless sensor networks, IEEE Trans. Ind. Electron. 62 Engineering, Springer, 2018, pp. 113–123.
(1) (2015) 434–443. [115] P. Guo, J. Cao, X. Liu, Lossless in-network processing in WSNs for domain-specific
[79] V.P. Illiano, E.C. Lupu, Detecting malicious data injections in event detection wire- monitoring applications, IEEE Trans. Ind. Inf. 13 (5) (2017) 2130–2139.
less sensor networks, IEEE Trans. Netw. Serv. Manage. 12 (3) (2015) 496–510. [116] H. Harb, A. Makhoul, R. Couturier, An enhanced k-means and ANOVA-based clus-
[80] Y. Li, H. Chen, M. Lv, Y. Li, Event-based k-nearest neighbors query processing over tering approach for similarity aggregation in underwater wireless sensor networks,
distributed sensory data using fuzzy sets, Soft Comput. (2017) 1–13. IEEE Sens. J. 15 (10) (2015) 5483–5493.
[81] Y. Han, J. Tang, Z. Zhou, M. Xiao, L. Sun, Q. Wang, Novel itinerary-based KNN [117] A. Morell, A. Correa, M. Barceló, J.L. Vicario, Data aggregation and principal com-
ponent analysis in WSNs, IEEE Trans. Wireless Commun. 15 (6) (2016) 3908–3919.

23
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

[118] C. Anagnostopoulos, S. Hadjiefthymiades, Advanced principal component-based [150] M. Zhao, T.W. Chow, Wireless sensor network fault detection via semi-supervised
compression schemes for wireless sensor networks, ACM Trans. Sensor Netw. local kernel density estimation, in: Industrial Technology (ICIT), 2015 IEEE Inter-
(TOSN) 11 (1) (2014) 7. national Conference on, IEEE, 2015, pp. 1495–1500.
[119] M.I. Chidean, E. Morgado, E. del Arco, J. Ramiro-Bargueo, A.J. Caamao, Scalable [151] C.J. Watkins, P. Dayan, Q-learning, Mach. Learn. 8 (3–4) (1992) 279–292.
data-coupled clustering for large scale WSN, IEEE Trans. Wireless Commun. 14 (9) [152] D.L. Poole, A.K. Mackworth, Artificial Intelligence: Foundations of Computational
(2015) 4681–4694. Agents, Cambridge University Press, 2010.
[120] S. Liu, L. Feng, J. Wu, G. Hou, G. Han, Concept drift detection for data stream learn- [153] A. Eiben, M. Schoenauer, Evolutionary computing, Inf. Process. Lett. 82 (1) (2002)
ing based on angle optimized global embedding and principal component analysis 1–6. Evolutionary Computation.
in sensor networks, Comput. Electr. Eng. 58 (2017) 327–336. [154] R.V. Kulkarni, A. Forster, G.K. Venayagamoorthy, Computational intelligence in
[121] A. Bertrand, M. Moonen, Distributed adaptive estimation of covariance matrix wireless sensor networks: a survey, IEEE Commun. Surv. Tutorials 13 (1) (2011)
eigenvectors in wireless sensor networks with application to distributed PCA, Sig- 68–96.
nal Process. 104 (2014) 120–135. [155] H.A. Hashim, B.O. Ayinde, M.A. Abido, Optimal placement of relay nodes in wire-
[122] M.I. Chidean, E. Morgado, M. Sanromn-Junquera, J. Ramiro-Bargueo, J. Ramos, less sensor network using artificial bee colony algorithm, J. Netw. Comput. Appl.
A.J. Caamao, Energy efficiency and quality of data reconstruction through data– 64 (2016) 239–248.
coupled clustering for self-organized large-scale WSNs, IEEE Sens. J. 16 (12) (2016) [156] Y. Xu, O. Ding, R. Qu, K. Li, Hybrid multi-objective evolutionary algorithms based
5010–5020. on decomposition for wireless sensor network coverage optimization, Appl. Soft
[123] T. Yu, X. Wang, A. Shami, Recursive principal component analysis based data out- Comput. 68 (2018) 268–282.
lier detection and sensor data aggregation in IoT systems, IEEE Internet Things J. [157] P. Kuila, P.K. Jana, Energy efficient clustering and routing algorithms for wireless
4 (6) (2017) 2207–2216. sensor networks: particle swarm optimization approach, Eng. Appl. Artif. Intell. 33
[124] M. Wu, L. Tan, N. Xiong, Data prediction, compression, and recovery in clustered (2014) 127–140.
wireless sensor networks for environmental monitoring applications, Inf. Sci. 329 [158] Y. Sun, W. Dong, Y. Chen, An improved routing algorithm based on ant colony
(2016) 800–818. optimization in wireless sensor networks, IEEE Commun. Lett. 21 (6) (2017)
[125] G. Gennarelli, F. Soldovieri, Performance analysis of incoherent RF tomography 1317–1320.
using wireless sensor networks, IEEE Trans. Geosci. Remote Sens. 54 (5) (2016) [159] M. Elhoseny, A. Tharwat, A. Farouk, A.E. Hassanien, k-coverage model based on
2722–2732. genetic algorithm to extend WSN lifetime, IEEE Sensors Lett. 1 (4) (2017) 1–4.
[126] T. Kanungo, D.M. Mount, N.S. Netanyahu, C.D. Piatko, R. Silverman, A.Y. Wu, An [160] C.-P. Chen, S.C. Mukhopadhyay, C.-L. Chuang, T.-S. Lin, M.-S. Liao, Y.-C. Wang,
efficient k-means clustering algorithm: analysis and implementation, IEEE Trans. J.-A. Jiang, A hybrid memetic framework for coverage optimization in wireless
Pattern Anal. Mach. Intell. 24 (7) (2002) 881–892. sensor networks, IEEE Trans. Cybern. 45 (10) (2015) 2309–2322.
[127] J.P. Ortega, M. Del, R.B. Rojas, M.J. Somodevilla, Research issues on k-means al- [161] J. Wang, Y. Cao, B. Li, H.-j. Kim, S. Lee, Particle swarm optimization based cluster-
gorithm: an experimental trial using matlab, in: CEUR Workshop Proceedings: Se- ing algorithm with mobile sink for WSNs, Future Gener. Comput. Syst. 76 (2017)
mantic Web and New Technologies, 2009, pp. 83–96. 452–457.
[128] K. Almi’ani, A. Viglas, L. Libman, Energy-efficient data gathering with tour length- [162] K.D. Praveen, T. Amgoth, C.S.R. Annavarapu, ACO-based mobile sink path deter-
-constrained mobile elements in wireless sensor networks, in: Local Computer Net- mination for wireless sensor networks under non-uniform data constraints, Appl.
works (LCN), 2010 IEEE 35th Conference on, IEEE, 2010, pp. 582–589. Soft Comput. 69 (2018) 528–540.
[129] S.C. Johnson, Hierarchical clustering schemes, Psychometrika 32 (3) (1967) [163] G. Yogarajan, T. Revathi, Nature inspired discrete firefly algorithm for optimal
241–254. mobile data gathering in wireless sensor networks, Wireless Networks (2017) 1–15.
[130] R.G. D’Andrade, U-statistic hierarchical clustering, Psychometrika 43 (1) (1978) [164] W.-L. Chang, D. Zeng, R.-C. Chen, S. Guo, An artificial bee colony algorithm for data
59–67. collection path planning in sparse wireless sensor networks, Int. J. Mach. Learn.
[131] X. Xu, R. Ansari, A. Khokhar, A.V. Vasilakos, Hierarchical data aggregation using Cybern. 6 (3) (2015) 375–383.
compressive sensing (HDACS) in WSNs, ACM Trans. Sensor Netw. (TOSN) 11 (3) [165] A.E. Eiben, J.E. Smith, et al., Introduction to Evolutionary Computing, vol, 53,
(2015) 45. Springer, 2003.
[132] P. Neamatollahi, S. Abrishami, M. Naghibzadeh, M.H.Y. Moghaddam, O. Younis, [166] J. Kuriakose, S. Joshi, R.V. Raju, A. Kilaru, A review on localization in wireless
Hierarchical clustering-task scheduling policy in cluster-based wireless sensor net- sensor networks, in: Advances in Signal Processing and Intelligent Recognition Sys-
works, IEEE Trans. Ind. Inf. 14 (5) (2018) 1876–1886. tems, Springer, 2014, pp. 599–610.
[133] R. Zhang, J. Pan, D. Xie, F. Wang, NDCMC: A hybrid data collection approach for [167] P. Cottone, S. Gaglio, G.L. Re, M. Ortolani, A machine learning approach for user lo-
large-scale WSNs using mobile element and hierarchical clustering, IEEE Internet calization exploiting connectivity data, Eng. Appl. Artif. Intell. 50 (2016) 125–134.
Things J. 3 (4) (2016) 533–543. [168] S.M. Mohamed, H.S. Hamza, I.A. Saroit, Coverage in mobile wireless sensor net-
[134] R. Zhang, J. Pan, J. Liu, D. Xie, A hybrid approach using mobile element and hier- works (M-WSN): a survey, Comput. Commun. 110 (C) (2017) 133–150.
archical clustering for data collection in WSNs, in: Wireless Communications and [169] W. Fang, X. Song, X. Wu, J. Sun, M. Hu, Novel efficient deployment schemes for
Networking Conference (WCNC), 2015 IEEE, IEEE, 2015, pp. 1566–1571. sensor coverage in mobile wireless sensor networks, Inf. Fusion 41 (2018) 25–36.
[135] S.W. Awan, S. Saleem, Hierarchical clustering algorithms for heterogeneous energy [170] H. Li, S. Wang, M. Gong, Q. Chen, L. Chen, IM2 DCA: Immune mechanism based
harvesting wireless sensor networks, in: Wireless Communication Systems (ISWCS), multipath decoupling connectivity algorithm with fault tolerance under coverage
2016 International Symposium on, IEEE, 2016, pp. 270–274. optimization in wireless sensor networks, Appl. Soft Comput. 58 (2017) 540–552.
[136] W. Peizhuang, Pattern recognition with fuzzy objective function algorithms (james [171] M. Abo-Zahhad, N. Sabor, S. Sasaki, S.M. Ahmed, A centralized immune-voronoi
c. bezdek), SIAM Rev. 25 (3) (1983) 442. deployment algorithm for coverage maximization and energy conservation in mo-
[137] M. Bernas, B. Płaczek, Fully connected neural networks ensemble with signal bile wireless sensor networks, Inf. Fusion 30 (2016) 36–51.
strength clustering for indoor localization in wireless sensor networks, Int. J. Dis- [172] A. Farhat, C. Guyeux, A. Makhoul, A. Jaber, R. Tawil, A. Hijazi, Impacts of wireless
trib. Sens. Netw. 11 (12) (2015) 1–10. sensor networks strategies and topologies on prognostics and health management,
[138] P. Nayak, A. Devulapalli, A fuzzy logic-based clustering algorithm for WSN to ex- J. Intell. Manuf. (2017) 1–27.
tend the network lifetime, IEEE Sens. J. 16 (1) (2016) 137–144. [173] H. Chen, X. Li, F. Zhao, A reinforcement learning-based sleep scheduling algorithm
[139] V. Klema, A. Laub, The singular value decomposition: its computation and some for desired area coverage in solar-powered wireless sensor networks, IEEE Sens. J.
applications, IEEE Trans. Automat. Contr. 25 (2) (1980) 164–176. 16 (8) (2016) 2763–2774.
[140] S. Wold, K. Esbensen, P. Geladi, Principal component analysis, Chemometrics Intell. [174] E. Ancillotti, C. Vallati, R. Bruno, E. Mingozzi, A reinforcement learning-based link
Lab. Syst. 2 (1–3) (1987) 37–52. quality estimation strategy for RPL and its impact on topology management, Com-
[141] X. Li, S. Ding, Y. Li, Outlier suppression via non-convex robust PCA for efficient put. Commun. 112 (2017) 1–13.
localization in wireless sensor networks, IEEE Sens. J. 17 (21) (2017) 7053–7063. [175] P. Oluwasanya, Anomaly detection in wireless sensor networks, arXiv:1708.08053
[142] P. Oikonomou, A. Botsialas, A. Olziersky, I. Kazas, I. Stratakos, S. Katsikas, D. Di- (2017).
mas, K. Mermikli, G. Sotiropoulos, D. Goustouridis, et al., A wireless sensing system [176] A.D. Paola, S. Gaglio, G.L. Re, F. Milazzo, M. Ortolani, Adaptive distributed outlier
for monitoring the workplace environment of an industrial installation, Sens. Ac- detection for WSNs, IEEE Trans. Cybern. 45 (5) (2015) 902–913.
tuators, B 224 (2016) 266–274. [177] M. Wazid, A.K. Das, An efficient hybrid anomaly detection scheme using K-means
[143] T.-W. Lee, Independent component analysis, in: Independent Component Analysis, clustering for wireless sensor networks, Wireless Personal Commun. 90 (4) (2016)
Springer, 1998, pp. 27–66. 1971–2000.
[144] X. Zhu, A.B. Goldberg, Introduction to semi-supervised learning, Synth. Lect. Artif. [178] H.H. Bosman, G. Iacca, A. Tejada, H.J. Wörtche, A. Liotta, Spatial anomaly de-
Intell. Mach. Learn. 3 (1) (2009) 1–130. tection in sensor networks using neighborhood information, Inf. Fusion 33 (2017)
[145] M.F.A. Hady, F. Schwenker, Semi-supervised learning, in: Handbook on Neural 41–56.
Information Processing, Springer, 2013, pp. 215–239. [179] C. O’Reilly, A. Gluhak, M.A. Imran, S. Rajasegarar, Anomaly detection in wireless
[146] J. Yoo, W. Kim, H.J. Kim, Distributed estimation using online semi-supervised sensor networks in a non-stationary environment, IEEE Commun. Surv. Tutorials
particle filter for mobile sensor networks, IET Control Theory Appl. 9 (3) (2015) 16 (3) (2014) 1413–1432.
418–427. [180] I. Butun, S.D. Morgera, R. Sankar, A survey of intrusion detection systems in wire-
[147] S. Kumar, S.N. Tiwari, R.M. Hegde, Sensor node tracking using semi-supervised less sensor networks, IEEE Commun. Surv. Tutorials 16 (1) (2014) 266–282.
hidden markov models, Ad Hoc Netw. 33 (2015) 55–70. [181] D. Dong, M. Li, Y. Liu, X.Y. Li, X. Liao, Topological detection on wormholes in
[148] B. Yang, J. Xu, J. Yang, M. Li, Localization algorithm in wireless sensor networks wireless ad hoc and sensor networks, IEEE/ACM Trans. Networking 19 (6) (2011)
based on semi-supervised manifold learning and its application, Cluster Comput. 1787–1796.
13 (4) (2010) 435–446. [182] Y. Wang, Z. Liu, D. Wang, Y. Li, J. Yan, Anomaly detection and visual perception
[149] J. Yoo, H.J. Kim, Target localization in wireless sensor networks using online semi– for landslide monitoring based on a heterogeneous sensor network, IEEE Sens. J.
supervised support vector regression, Sensors 15 (6) (2015) 12539–12559. 17 (13) (2017) 4248–4257.

24
D. Praveen Kumar et al. Information Fusion 49 (2019) 1–25

[183] M. Xie, J. Hu, S. Guo, A.Y. Zomaya, Distributed segment-based anomaly detec- [214] C. Sergiou, P. Antoniou, V. Vassiliou, A comprehensive survey of congestion control
tion with kullback -leibler divergence in wireless sensor networks, IEEE Trans. Inf. protocols in wireless sensor networks, IEEE Commun. Surv. Tutorials 16 (4) (2014)
Forensics Secur. 12 (1) (2017) 101–110. 1839–1859.
[184] V.P. Illiano, A. Paudice, L. Muñoz-González, E.C. Lupu, Determining resilience [215] M.A. Kafi, D. Djenouri, J. Ben-Othman, N. Badache, Congestion control protocols
gains from anomaly detection for event integrity in wireless sensor networks, ACM in wireless sensor networks: a survey, IEEE Commun. Surv. Tutorials 16 (3) (2014)
Trans. Sensor Netw. (TOSN) 14 (1) (2018) 5. 1369–1390.
[185] S. Shamshirband, A. Patel, N.B. Anuar, M.L.M. Kiah, A. Abraham, Cooperative game [216] C. Raman, V. James, FCC: Fast congestion control scheme for wireless sensor net-
theoretic approach using fuzzy Q-learning for detecting and preventing intrusions works using hybrid optimal routing algorithm, Cluster Comput. (2018) 1–11.
in wireless sensor networks, Eng. Appl. Artif. Intell. 32 (2014) 228–241. [217] H.A. Al-Kashoash, M. Hafeez, A.H. Kemp, Congestion control for 6LoWPAN net-
[186] S.A. Haque, M. Rahman, S.M. Aziz, Sensor anomaly detection in wireless sensor works: a game theoretic framework, IEEE Internet Things J. 4 (3) (2017) 760–771.
networks for healthcare, Sensors 15 (4) (2015) 8764–8786. [218] M.A. Kafi, J. Ben-Othman, A. Ouadjaout, M. Bagaa, N. Badache, REFIACC: Reli-
[187] T. Muhammed, R.A. Shaikh, An analysis of fault detection strategies in wireless able, efficient, fair and interference-aware congestion control protocol for wireless
sensor networks, J. Netw. Comput. Appl. 78 (2017) 267–287. sensor networks, Comput. Commun. 101 (2017) 1–11.
[188] H. Geng, Y. Liang, F. Yang, L. Xu, Q. Pan, Model-reduced fault detection for mul- [219] S.-H. Moon, S. Park, S.-j. Han, Energy efficient data collection in sink-centric
ti-rate sensor fusion with unknown inputs, Inf. Fusion 33 (2017) 1–14. wireless sensor networks: a cluster-ring approach, Comput. Commun. 101 (2017)
[189] R.R. Swain, P.M. Khilar, S.K. Bhoi, Heterogeneous fault diagnosis for wireless sen- 12–25.
sor networks, Ad Hoc Netw. 69 (2018) 15–37. [220] A. Ez-Zaidi, S. Rakrak, A comparative study of target tracking approaches in wire-
[190] R. Palanikumar, K. Ramasamy, Effective failure nodes detection using matrix cal- less sensor networks, J. Sensors 2016 (2016) 1–12.
culus algorithm in wireless sensor networks, Cluster Comput. (2018) 1–10. [221] K. Xiao, R. Wang, T. Fu, J. Li, P. Deng, Divide-and-conquer architecture based
[191] M. Hammoudeh, R. Newman, Adaptive routing in wireless sensor networks: QoS collaborative sensing for target monitoring in wireless sensor networks, Inf. Fusion
optimisation for enhanced application performance, Inf. Fusion 22 (2015) 3–15. 36 (2017) 162–171.
[192] X. Liu, Routing protocols based on Ant colony optimization in wireless sensor net- [222] S. Vasuhi, V. Vaidehi, Target tracking using interactive multiple model for wireless
works: a survey, IEEE Access 5 (2017) 26303–26317. sensor network, Inf. Fusion 27 (2016) 41–53.
[193] M. Asif, S. Khan, R. Ahmad, M. Sohail, D. Singh, Quality of service of routing [223] A. Abrardo, M. Martal, G. Ferrari, Information fusion for efficient target detection
protocols in wireless sensor networks: a review, IEEE Access 5 (2017) 1846–1871. in large-scale surveillance wireless sensor networks, Inf. Fusion 38 (2017) 55–64.
[194] K.-L.A. Yau, H.G. Goh, D. Chieng, K.H. Kwong, Application of reinforcement learn- [224] Z. Wei, Y. Zhang, X. Xu, L. Shi, L. Feng, A task scheduling algorithm based on
ing to wireless sensor networks: models and algorithms, Computing 97 (11) (2015) Q-learning and shared value function for WSNs, Comput. Networks 126 (2017)
1045–1075. 141–149.
[195] Shashikala, C. Kavitha, Secure Cluster-based Data Integrity Routing for Wireless [225] M.M. Hassan, M.Z. Uddin, A. Mohamed, A. Almogren, A robust human activity
Sensor Networks, Springer Singapore, Singapore, 2018. recognition system using smartphone sensors and deep learning, Future Gener.
[196] J. Kabara, M. Calle, MAC Protocols used by wireless sensor networks and a general Comput. Syst. 81 (2018) 307–313.
method of performance evaluation, Int. J. Distrib. Sens. Netw. 8 (1) (2012) 1–11. [226] Y. Kılıçaslan, G. Tuna, G. Gezer, K. Gulez, O. Arkoc, S.M. Potirakis, ANN-based
[197] I. Mustapha, B.M. Ali, A. Sali, M.F.A. Rasid, H. Mohamad, An energy efficient re- estimation of groundwater quality using a wireless water quality network, Int. J.
inforcement learning based cooperative channel sensing for cognitive radio sensor Distrib. Sens. Netw. 10 (4) (2014) 1–8.
networks, Pervasive Mob. Comput. 35 (2017) 165–184. [227] D. Ye, M. Zhang, A self-adaptive sleep/wake-up scheduling approach for wireless
[198] S. Kosunalp, Y. Chu, P.D. Mitchell, D. Grace, T. Clarke, Use of q-learning approaches sensor networks, IEEE Trans. Cybern PP (99) (2017) 1–14.
for practical medium access control in wireless sensor networks, Eng. Appl. Artif. [228] S.B. Chandanapalli, E.S. Reddy, D.R. Lakshmi, DFTDT: distributed functional tan-
Intell. 55 (2016) 146–154. gent decision tree for aqua status prediction in wireless sensor networks, Int. J.
[199] M. Rovcanin, E. De Poorter, I. Moerman, P. Demeester, A reinforcement learning Mach. Learn. Cybern. (2017) 1–16.
based solution for cognitive network cooperation between co-located, heteroge- [229] W. Wen, S. Zhao, C. Shang, C.-Y. Chang, EAPC: Energy-aware path construction for
neous wireless sensor networks, Ad Hoc Netw. 17 (2014) 98–113. data collection using mobile sink in wireless sensor networks, IEEE Sens. J. 18 (2)
[200] M. Rovcanin, E. De Poorter, D. van den Akker, I. Moerman, P. Demeester, C. Blon- (2018) 890–901.
dia, Experimental validation of a reinforcement learning based approach for a ser- [230] H. Salarian, K.-W. Chin, F. Naghdy, An energy-efficient mobile-sink path selection
vice-wise optimisation of heterogeneous wireless sensor networks, Wireless Netw. strategy for wireless sensor networks, IEEE Trans. Veh. Technol. 63 (5) (2014)
21 (3) (2015) 931–948. 2407–2419.
[201] K.-H. Phung, B. Lemmens, M. Goossens, A. Nowe, L. Tran, K. Steenhaut, Sched- [231] T. Wang, J. Zeng, Y. Lai, Y. Cai, H. Tian, Y. Chen, B. Wang, Data collection from
ule-based multi-channel communication in wireless sensor networks: a complete WSNs to the cloud based on mobile fog elements, Future Gener. Comput. Syst.
design and performance evaluation, Ad Hoc Netw. 26 (2015) 88–102. (2017), doi:10.1016/j.future.2017.07.031.
[202] M. Ambigavathi, D. Sridharan, Energy-aware data aggregation techniques in wire- [232] I. Ha, M. Djuraev, B. Ahn, An optimal data gathering method for mobile sinks in
less sensor network, in: Advances in Power Systems and Energy Management, WSNs, Wireless Personal Commun. 97 (1) (2017) 1401–1417.
Springer, 2018, pp. 165–173. [233] F.K. Shaikh, S. Zeadally, Energy harvesting in wireless sensor networks: a compre-
[203] k. xie, L. Wang, X. Wang, G. Xie, J. Wen, Low cost and high accuracy data gath- hensive review, Renewable Sustainable Energy Rev. 55 (2016) 1041–1054.
ering in WSNs with matrix completion, IEEE Trans. Mob. Comput. 17 (7) (2017) [234] S. Kosunalp, A new energy prediction algorithm for energy-harvesting wireless sen-
1595–1608. sor networks with Q-learning, IEEE Access 4 (2016) 5755–5763.
[204] H. Lin, D. Bai, Y. Liu, Maximum data collection rate routing for data gather trees [235] R.C. Hsu, C.T. Liu, H.L. Wang, A reinforcement learning-based ToD provisioning
with data aggregation in rechargeable wireless sensor networks, Cluster Comput. dynamic power management for sustainable operation of energy harvesting wire-
(2017) 1–11. less sensor node, IEEE Trans. Emerg. Top Comput. 2 (2) (2014) 181–191.
[205] E. Kanjo, E.M. Younis, N. Sherkat, Towards unravelling the relationship between [236] F.A. Aoudia, M. Gautier, O. Berder, RLMan: an energy manager based on reinforce-
on-body, environmental and emotion data using sensor information fusion ap- ment learning for energy harvesting wireless sensor networks, IEEE Trans. Green
proach, Inf. Fusion 40 (2018) 18–31. Commun. Networking (2018) 1–11.
[206] A. Pinto, C. Montez, G. Arajo, F. Vasques, P. Portugal, An approach to implement [237] M. Collotta, G. Pau, A.V. Bobovich, A fuzzy data fusion solution to enhance the
data fusion techniques in wireless sensor networks using genetic machine learning QoS and the energy consumption in wireless sensor networks, Wireless Commun.
algorithms, Inf. Fusion 15 (2014) 90–101. Mobile Comput. 2017 (2017) 1–10.
[207] Y.C. Wu, Q. Chaudhari, E. Serpedin, Clock synchronization of wireless sensor net- [238] W. Sun, W. Lu, Q. Li, L. Chen, D. Mu, X. Yuan, WNN-LQE: Wavelet-neural-net-
works, IEEE Signal Process. Mag. 28 (1) (2011) 124–138. work-based link quality estimation for smart grid WSNs, IEEE Access 5 (2017)
[208] D. Djenouri, M. Bagaa, Synchronization protocols and implementation issues in 12788–12797.
wireless sensor networks: a review, IEEE Syst. J. 10 (2) (2016) 617–627. [239] E.K. Lee, H. Viswanathan, D. Pompili, RescueNet: Reinforcement-learning-based
[209] W. Wang, S. Jiang, H. Zhou, M. Yang, Y. Ni, J. Ko, Time synchronization for accel- communication framework for emergency networking, Comput. Networks 98
eration measurement data of Jiangyin bridge subjected to a ship collision, Struct. (2016) 14–28.
Control Health Monit. 25 (1) (2018) 1–18. [240] A.P. Renold, S. Chandrakala, MRL-SCSO: Multi-agent reinforcement learning-based
[210] K.-P. Ng, C. Tsimenidis, W.L. Woo, C-Sync: Counter-based synchronization for du- self-configuration and self-optimization protocol for unattended wireless sensor
ty-cycled wireless sensor networks, Ad Hoc Netw. 61 (2017) 51–64. networks, Wireless Personal Commun. 96 (4) (2017) 5061–5079.
[211] Y.-P. Tian, Time synchronization in WSNs with random bounded communication [241] L. Ren, W. Wang, H. Xu, A reinforcement learning method for constraint-satisfied
delays, IEEE Trans. Automat. Contr. 62 (10) (2017) 5445–5450. services composition, IEEE Trans. Serv. Comput. PP (99) (2017) 1–14.
[212] C. Tan, S. Ji, Z. Gui, J. Shen, D.-S. Fu, J. Wang, An effective data fusion-based [242] M.A. Razzaque, M.H.U. Ahmed, C.S. Hong, S. Lee, QoS-aware distributed adaptive
routing algorithm with time synchronization support for vehicular wireless sensor cooperative routing in wireless sensor networks, Ad Hoc Netw 19 (Supplement C)
networks, J. Supercomput. (2017) 1–16. (2014) 28–42.
[213] A. Ghaffari, Congestion control mechanisms in wireless sensor networks: a survey,
J. Netw. Comput. Appl. 52 (2015) 101–115.

25