3s PDF

This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSM.2020.2988994, IEEE
Transactions on Network and Service Management
Fault-Resilient Distributed Detection and Estimation

over a SW-WSN Using LCMV Beamforming
Om Jee Pandey, Member, IEEE, Ved Gautam, Ha H. Nguyen, Senior Member, IEEE, Mahendra K. Shukla, and
Rajesh M. Hegde, Senior Member, IEEE
Abstract—Recent technology advancement has resulted in opti- Many applications of WSNs involve distributed signal pro-
mistic view toward the practicability of wireless sensor networks cessing [8], [9], including soil moisture estimation in a farm,
(WSNs) in the context of Internet of Things (IoT) and Cyber forest fire detection, industrial monitoring, wireless multi-
Physical Systems (CPS). However, to realize their full benefits in
a broad range of commercial applications, there are still many media transmission, and noise level estimation in a dense
technical hitches that need to be overcome. In this paper, we populated area. Therefore, distributed signal processing using
address three vital technical issues in a WSN: (1) distributed WSNs has received much attention in recent years [10],
event detection, (2) distributed parameter estimation, and (3) [11]. The two main divisions of distributed processing are
network’s robustness. We make use of a recent development in distributed detection and distributed estimation [12]–[14]. Dis-
social networks called small world characteristics and propose
novel fault-resilient distributed detection and estimation methods tributed detection [15], [16] involves consensus in the network
over a small world WSN (SW-WSN). In particular, a small for an event detection in a distributed manner. It utilizes the
world WSN has been developed by mounting antenna arrays on data of neighboring nodes to reach consensus in the network.
sensor nodes for the purpose of beamforming. A low-complexity Similarly, distributed estimation [17], [18] involves estimation
optimization problem for beamforming is formulated by intro- of a vector of parameters at every node using measurements
ducing a new parameter Flow between node pairs. Additionally,
a new beamforming algorithm is also proposed which optimizes of the neighboring nodes. Distributed estimation schemes are
this flow, leading to optimal beam parameters. The proposed classified into incremental and diffusion methods. The require-
method yields a lower average path length and a higher average ment of cyclic paths in the network renders the incremental
clustering coefficient of the network. Experiments are conducted method in many practical scenarios useless. The diffusion
using simulations and real node deployments over a WSN testbed. method, on the other hand, allows communication of a node
Analysis and experimental results obtained demonstrate that the
proposed SW-WSN model achieves faster convergence rates for with its neighbors, thus practically more feasible. This paper
both distributed detection and distributed estimation while being is concerned only with the diffusion estimation method.
resilient to node failures when compared to results obtained using Along with distributed processing, reliability is another im-
state-of-the-art methods. portant aspect of WSNs [13], [19], [20]. Critical applications
Index Terms—Small world wireless sensor networks (SW- of WSNs demand the networks to be reliable. Reliability of
WSNs), small world characteristics (SWC), beamforming, dis- a network greatly depends on the robustness of the network
tributed event detection, parameter estimation, robustness. to node failure. Different metrics such as k−connectivity and
partial k−connectivity have been proposed in the literature to
asses robustness of a network [21], [22].
I. INTRODUCTION Numerous methods have been proposed to address the afore-
IRELESSSensor Networks (WSNs) are responsible for mentioned problems in WSNs. In [23], iterative distributed
W various sensing and control tasks in Internet of Things
(IoT), Cyber Physical Systems (CPS), and context-aware
detection is proposed to improve consensus for event detection.
In [24]–[29], it has been found that the network topology has a
pervasive systems [1], [2]. WSNs involve measurement and major influence on the performance of distributed processing.
computational units distributed geographically to monitor the The more “well connected” the network is, the better its
environment in their proximity. WSNs serve as a bridge consensus performance can be achieved. In [24], [30], [31],
between physical and cyber worlds and find applications in a systematic study is performed to analyze performance of
industrial automation [2], smart homes [3], smart agriculture the network topology in distributed estimation. These methods
[4], video surveillance [4], traffic monitoring [5], smart health- involve introduction of small world phenomena to improve the
care [6], and smart energy grids [7]. network performance. Small world phenomena were discov-
ered in social networks by Stanley Milgram [32]. It revealed
O. J. Pandey is with the Department of Electronics and Communication that the human society is a small world network having short
Engineering, School of Engineering and Applied Sciences, SRM University path lengths. Watts and Strogatz in [33] demonstrated how
AP, Andhra Pradesh 522502, India (e-mail: omjee.p@srmap.edu.in). rewiring a few links in a regular graph can transform the
V. Gautam is with the Urvija AI Private Limited, Bangalore 560061,
Karnataka, India (e-mail: vedg96gautam@gmail.com). graph into a small world graph. This procedure is widely
H. H. Nguyen and M. K. Shukla are with the Department of Electrical used in wireless networks [34], [35] to improve the network
and Computer Engineering, University of Saskatchewan, Saskatoon, SK S7N performance on robustness [19], distributed processing [24],
5A9, Canada (e-mail: {ha.nguyen; m.shukla}@usask.ca).
R. M. Hegde is with the Department of Electrical Engineering, Indian Insti- [25], latency [36], [37], localization [38], energy-efficiency
tute of Technology Kanpur, Kanpur 208016, India (e-mail: rhegde@iitk.ac.in). [39], [40], time synchronization [41], data gathering [42], and
1932-4537 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Durham. Downloaded on May 16,2020 at 04:22:07 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSM.2020.2988994, IEEE
required bandwidth [35]. network robustness over a real WSN testbed is also carried out.
A. Motivation Detailed comparison with the state-of-the-art methods is also
presented to highlight the advantages of the proposed methods.
Centralized event detection and parameter estimation
present several challenges over a WSN. For instance, the
exchange of a large amount of information in a multi-hop C. Notations
fashion leads to reduced network lifetime and poor bandwidth The parameters and symbols used in this paper are listed
utilization. It also reduces the network performance with and explained in Table I.
respect to decision latency and robustness to the node failure.
Distributed detection and estimation over a WSN improves Table I: List of used parameters/symbols
the network performance in various contexts by only utilizing Parameters/Symbols Meaning
the data of neighboring nodes [11], [16]. The performance N and K Number of sensor nodes and minimum number
of neighbors to each node in the network
of distributed detection and estimation highly depends on the ki , ei , and di Total number of neighbors of node i, total num-
network topology [24]–[31]. ber of links between neighbors of node i, and
Small world characteristics (SWC) can be introduced in a degree of node i
Ł, Ci , and C̄ Probability of link addition, clustering coeffi-
WSN, leading to a small world WSN (SW-WSN). A SW- cient (CC) of a node i, and average CC
WSN is a well connected network with reduced hop counts N , E, P, B, and L Sets of sensor nodes, conventional links, re-
between node pairs. The well-connected property of a SW- ceived signal strength (RSS) between node-
pairs, beamforming nodes, and set of new links
WSN leads to spectral phase transition phenomena and results introduced in the network
in improved distributed detection and estimation performance sN and γij N th node and RSS between ith and j th nodes
[24]. In the literature, most of the works utilize random links H0 , H1 , π0 , and π1 Null hypothesis, alternative hypothesis, prior
to create a SW-WSN without considering its practicability probabilities of H0 , and H1
H and H̃ True state and global decision
in real time applications. Introduction of random links also (·)? , k · kF , and E| · | Transpose, Frobenius norm, and Expectation
results in poor network performance. To address this issue, (·)∗ , and %, η Optimal parameter, and measures of SWness
the authors in [43], [44] have recently developed SW-WSNs on , Λn , and R Observation, log likelihood ratio (LLR) at nth
sensor node, and global LLR
via different beamforming techniques, which are more realistic zn , µn , and Pe Gaussian noise N (0, σ 2 ), signal mean at nth
than adding random links. Motivated by the works in [43], sensor node, and average probability of error
[44], in this paper we propose a novel ranking scheme based vn [k] and δ Observation value stored at nth sensor node at
on the structural importance of the sensor nodes to develop discrete time instant k and fusion threshold
ci,j , di,j , and li,j Weight corresponding to the link (i, j), length
SW-WSNs using Linearly-Constrained Minimum Variance of the shortest path between nodes i, j, element
(LCMV) beamforming [45], [46]. LCMV beamforming is a of Laplacian matrix between nodes i and j
technique in which the radiation pattern of antenna elements p(0) Vector parameter need to be estimated
Sn [k] and rn,k Scalar measurement and regression vector at nth
mounted on a sensor node can be made directional. In general, sensor node at time instant k
the use of beamforming methods leads to reduced neigh- ψn [k] and φn [k] Unbiased estimate and aggregate estimate of
borhood connectivity within the network. It also increases vector parameter at nth node at time instant k
r, ∆θ, and α Beam length, beam width, and beam direction
the development complexity of a SW-WSN and reduces the A and A0 Total radio area covered by a sensor node and
network performance in terms of robustness to node failure. minimum radio area required by the sensor node
However, the use of LCMV beamforming results in desired to maintain neighborhood connectivity
τ and σ Measures of shortest path length and number of
side lobes, leading to higher network connectivity. Hence, the shortest path between selected node-pair
primary objective of this work is to develop and analyze a low- ρ and λ Path loss exponent and wavelength of the signal
complexity SW-WSN using LCMV beamforming for fault- ζn and θs Correlation index and half power beamwidth of
side lobes
resilient distributed detection and estimation. β and β0 Power attenuation and corresponding threshold
B. Contributions r0 and Ae Omnidirectional radio range of a sensor node
and effective aperture of an antenna
The main contributions of this paper are as follows. A
novel method for the development of a SW-WSN using
D. Organization of the Paper
LCMV beamforming is first proposed. The method controls
the antenna gains in various directions leading to the optimal The rest of the paper is organized as follows. Section
SWC. A method to compute optimal beamforming parameters II discusses related research. Section III presents concepts
for individual nodes is also developed utilizing traffic flow and terminologies involved in describing a SW-WSN. Section
optimization. The method of traffic flow optimization uses IV describes the network model and problem formulations.
the parameter Flow, which defines the data flow between two The proposed methods of distributed detection and estimation
selected nodes within the network. The computation of Flow along with corresponding algorithm development are described
between node pairs is performed using nodes betweenness in Section V. Section VI presents performance evaluation.
and centrality measures. The parameter Flow is further used Concluding remarks and future work are given in Section VII.
to reduce the complexity of SW-WSN development. Subse-
quently, novel distributed detection and estimation algorithms II. R ELATED R ESEARCH
are proposed over the developed SW-WSN. Extensive investi- In recent years, with the emergence of wireless sensor
gation into the distributed detection, parameter estimation, and networks and ad hoc networks, research on distributed signal
processing algorithms has drawn significant attention (see, engines which allows fast convergence for the distributed
e.g., [14], [16] and references therein). In an earlier work [47], detection algorithm. In [26], the authors use the theory from
the problem of finding the maximum likelihood estimator of a small world networks in designing the network topology. The
commonly observed model based on data collected by a sensor results illustrate that the consensus is obtained with minimal
network is addressed. First, the authors propose an iterative communications and processing cost. In [27], the authors
algorithm to relax the need of complete data sharing. Then, a present the optimal topology design problem for distributed
method is proposed where the local data of a sensor is used to inference in sensor networks. A topology specifies the graph
obtain a sub-optimal estimate. In [48], bandwidth-constrained that decides the connection established among the sensors. In
distributed estimation is considered where quantization is an [29], the authors show that a more fast learning rate can be
essential part of the estimation process. Efficient distributed obtained in a well-connected network.
parameter estimation methods are developed in [49] based Robustness toward node failure is another challenging prob-
on the distributed sequential Bayesian estimation method. In lem in a WSN. Strong robustness enhances the network
[50], distributed parameter estimation in WSNs is performed lifetime and quality of service. In order to address the problem
with a total bit rate constraint. In [51], distributed estimation of robustness, various methods have also been investigated in
in large WSNs via a locally optimum approach is proposed. the literature. In [21], a method to place redundant sensor
Subsequently, in [52] the authors examine new versions of the nodes to establish multi-connectivity is proposed. This work
diffusion least-mean squares (LMS) algorithm, namely Adapt- explores how to add as few nodes as possible to a sensor
then-Combine (ATC) and Combine-then-Adapt (CTA). network such that the resulting network is k-connected or
In [28], the authors investigate the effect of network topol- partially k-connected. In [22], the authors propose a resilience
ogy on performance of the ATC diffusion LMS estimation al- factor to measure the network level of robustness and pro-
gorithm. Results obtained illustrate that the estimation perfor- tection against targeted attacks. In addition, they proposed
mance greatly depends on the network topological properties. strategies to improve resilience by simple alterations in the
Further, in [53] the authors illustrate that inferring the structure network topology. Recently, in [20] the authors propose a
of the graph characterizing the statistical dependencies among robust wireless sensor and actuator network to maintain the
the observed data can provide vital information for sensor control stability of multiple plants over the spatial-temporal
network topology development. Moreover, the authors find changes of wireless networks.
a network topology that minimizes the energy requirement Apart from the methods discussed above which deal with
to achieve the consensus. They illustrate that the network estimation and detection problems separately, various methods
topology has a very crucial role in the design of an efficient are proposed in literature which solve these problems jointly.
sensor network. In a recent work [54], the authors investigate For instance, in [8] the authors proposed a hidden Markov
the problem of non-fragile estimation for a class of complex random field (HMRF) framework for distributed detection and
networks with switching topologies and quantization effects. estimation in sensor networks. In this work, the authors pro-
In the context of distributed detection, various works have posed novel algorithms for distributed detection and estimation
also been carried out. In an earlier work [55], the detection of the hidden random field from the noisy measurements. The
problem is addressed jointly with system-resource constraints. proposed methods are less complex and applicable to several
In [24], an algorithm for ultrafast consensus is proposed sensing environments, where the nodes communicate only
which studies the convergence of the differential equation with their neighbors. However, the proposed methods are too
ẋ(t) = Lx(t), where x(t) is a decision variable and L is simplistic for many applications and therefore development of
the Laplacian of a communication graph of agents. In [26], more general process models need to be investigated. In [53],
an iterative detection procedure is proposed. The goal of [59], the authors illustrated that the topology of a network
the iterative algorithm is for every sensor state to converge plays a significant role in determining performance of dis-
to the global average local log likelihood ratios (LLRs). In tributed algorithms in terms of energy expenditure and latency.
[56], a consensus about the event is obtained via collaborative They proposed a method to optimize the network topology in
exchange of data among the sensors over a communication order to minimize energy consumption or to match the graph
network. In [57], opposite to the conventional method which describing the statistical dependencies among the variables
utilizes a bank of fixed parallel access channels for sending observed by the nodes. Recently, in [16] a joint problem for
the observed data to the fusion center, the authors explore the detection, estimation, and tracking has been investigated in
possibility of employing a common multiple access channel, the context of CPS. In this work, a distributed consensus
which drastically reduces the required bandwidth and detection algorithm for sparse events detection in CPS is proposed. The
delay. Further, the detection algorithms proposed in [58] use devices in a CPS compute the LLR from local observation via
the diffusion LMS and recursive least squares (RLS) algo- a consensus approach to iteratively optimize the consensus
rithms to obtain a parameter estimate which is further used to LLRs for the whole CPS system. However, the complexity of
execute a hypothesis test. the method is high due to the iterative process used to obtain
Regarding detection problems in WSNs, the network topol- the global decision.
ogy also plays an important role. For instance, in [25], the au- III. S MALL W ORLD - W IRELESS S ENSOR N ETWORKS :
thors address the issue of graph connectivity. They specify the C HARACTERISTICS AND D EFINITIONS
connectivity pattern among sensor nodes leading to efficient Small world phenomena were first observed in [32] in social
detection. For this, the authors utilize small world network connectivity of people. This work provides a theory of “six
degrees of separation”. In [33], small world phenomena have A. Network Model

been investigated in graphs. In [34], [37]–[39], [42]–[44], In this work, a WSN is considered as a graph represented by
small world characteristics (SWC) are observed in wireless G = (N , E, P, B, L). Here, N represents the total number of
networks. In general, small world WSNs (SW-WSNs) are char- sensor nodes {s1 , s2 , . . . , sN } in the network. E indicates a set
acterized by low Average Path Length (APL) and high Average of M conventional links {lm = (im , jm ), m = 1, 2, . . . , M },
Clustering Coefficient (ACC). For an undirected WSN with where, (i, j) ∈ E whenever sensor node si can commu-
N nodes, representing
PN a regular network, the APL is given by nicate with sensor node sj . P is the set of received sig-
LWSN = N (N1−1) i6=j di,j , where the sum is over all pairs nal strengths (RSSs) γ12 , γ13 , . . . , γN (N −1) between node
of distinct nodes and di,j is the length of the shortest path pairs. The node pairs can communicate with each other
between nodes i and j. On the other hand, in a WSN having when the corresponding RSS, γij , lies above or equal to
SWC the APL, LSW−WSN , between two randomly selected a certain threshold γ, i.e., γij ≥ γ. B denotes the set
nodes grows proportionally to the logarithm of the number of of B nodes {s1 , s2 , . . . , sB } used for beamforming for the
nodes in the network [33], i.e., LSW−WSN ∝ ln(N ). The APL development of a SW-WSN. L represents the set of T new
for a network with random link addition [60] is given by, links {lt = (it , jt ), t = 1, 2, . . . , T }, introduced in the net-
work, where L ∈ / E. In the case of a conventional WSN, B
N 1/d
Lrand (N, Ł) ∼ f (N KŁ), (1) and L are null sets {∅}. In this work, we have mainly focused
K on a static network, however, the network graph G can be
where f is a universal scaling function and is given by considered as both static and dynamic networks. In case of
a dynamic network, graph G can be recognized as snap-shot
const, if a 1 series of static graphs G1 , G2 , G3 , . . . , GT , at corresponding
f (a) =
ln(a)/a, if a 1 time instants [61]. Moreover, it is assumed that the change
and each node has at least K neighbors. The probability of in network topology is slow enough so that the time-varying
link addition [33], [60], denoted by Ł, is given by sensor network graph G can be recognized as snap-shot
series of static graphs at corresponding time instants for the
# links added development of a SW-WSN. All sensor nodes are uniformly
Ł= . (2)
# possible links − # existing links distributed over a 2-dimensional geographical area of length
L and width W . Sensor nodes provide information on sensed
From the above equations it is noted that if Ł is small (Ł →
data, observed phenomena, list of its neighbors, and distances
0), the network will have the property of a regular network.
to adjacent nodes at every time instant. An equal cost for all
On the other hand, if Ł is large (Ł → 1), then the network
communication links, i.e., c(li ) = c(lj ), ∀ li , lj ∈ {E, L} is
will have random network characteristics. A general SW-WSN
considered. All nodes are mounted with antenna arrays and are
corresponds to Ł in between 0 and 1.
capable of forming a desired beam pattern. Initially, all nodes
WSNs are spatial graphs, where links are created us-
transmit using omnidirectional beams with the radio range r0 .
ing radio connectivity. There is a limit on the radio range
Subsequently, a set of B nodes is selected for beamforming,
of a sensor node, hence in such networks the long-range
which forms long directional links.
connections are generally absent. Thus, these networks are
clustered but they do not experience small world phenomena. B. Problem Formulations
In [60], it is illustrated that the average distance between
In this section, problem formulation of distributed detec-
nodes, l(N, Ł), decreases drastically as long as Ł is non-
tion is discussed first. Subsequently, problem formulation for
zero. For Ł = 0, a linear chain of sites is obtained so that
distributed parameter estimation is described.
l(N, 0) = N4K(N (N +2K−2)
−1) ∼ 4KN
grows like N . For Ł = 1,
1) Detection Problem: The problem of distributed detec-
l(N, 1) grows like ln(N )/ln(2K − 1).
tion considered in this paper is a binary hypothesis testing
A common property of small world networks is that cliques
problem [23], [25]. The state of the environment takes one of
form, representing circles of acquaintances in which every
two possible hypotheses, null hypothesis H0 and alternative
node knows every other node. This inherent tendency to cluster
hypothesis H1 . Conventionally, H0 indicates that the target
is quantified by the clustering coefficient (CC). To be more
is absent, while H1 indicates that the target is present. The
specific, CC for a node i in the network is defined as [60]
true state H of the environment is observed by N nodes,
Ci = ki (k2ei −1)
i
, where ki is the total number of neighbors of
which collect N real observations, o = (o1 , o2 , . . . , oN ), as
node i and ei is the total links between PNneighbors of node illustrated in Fig. 1. Node observations are considered to be
i. ACC of WSNs is given by C̄ = N1 i=1 Ci . Therefore, a independent and identically distributed with known conditional
general SW-WSN exhibits low APL and high ACC. density f (o|Hi ), i = 0, 1. The Gaussian shift-in-mean models
for the observations at sensor node sn under H0 and H1 are
IV. N ETWORK M ODEL AND P ROBLEM F ORMULATION given by (
on = µ1 + zn : H0
In this section, the network model used for the development sn = (3)
on = µ2 + zn : H1 .
of small world wireless sensor networks is described first.
Subsequently, problem formulations for distributed detection In the above equation, µ1 and µ2 are the signal means under
and distributed estimation are presented. H0 and H1 , respectively, zn is Gaussian noise with zero mean
where, ci,j is the weight corresponding to the link (i, j). Note
that, ci,j = 0, if (i, j) ∈
/ {E, L}. A compact expression in a
matrix form can be written as,
v[k] = Cv[k − 1]. (9)
In Equation (9), v[k] is a N × 1 vector of all current states
and C is a weight matrix corresponding to the existing links
in the network. v[k − 1] is a N × 1 vector of all previous
states. The iterative process continues until vn [k] of all nodes
converges to the global average LLR, R/N . The final decision
is then obtained locally at each node according to its current
state given by,
H̃n =1
vn [k] ≷ δ/N, (10)
H̃n =0
where Hn denotes the decision of node sn . The updating rule

in terms of initial states can be written as,
Fig. 1: The model of a WSN consisting of multiple-antennas
mounted sensor nodes. Sensor nodes sense the environment v[k] = C(k) v[0]. (11)
within their proximity. Conventional links are red. New links
created using directional beamforming are illustrated in green. It is clear that the weight matrix C needs to be designed in
such a way that the iterative process converges fast. Equation
(11) shows that this can be achieved when C(k) converges to
and variance σ 2 , i.e., zn ∼ N (0, σ 2 ). The goal of distributed all-ones matrix, i.e., lim C(k) = N1 1N ×N . Hence, the dis-
k→∞
detection is for all nodes to reach a global common decision H̃ tributed detection problem can be seen as a optimal topology
about the true state H. It is considered that the statistics stored design through the optimal weight matrix. Mathematically, the
at each sensor node are updated in an iterative manner to reach distributed detection problem can be formulated as,
a consensus about the global statistics based on information 2
∗ ∗ ∗
(k) 1
exchange over the network. The detection performance is C (α , L ) = arg min C (α, L) − 1N ×N

, (12)
measured by the average probability of error, expressed as, C(α,L) N F
where the weight matrix C contains elements
Pe = Pr(H̃ 6= H) = π0 Pe,0 + π1 Pe,1 , where, (4) 
α,
 if (i, j) ∈ {E, L}
Pe,0 = Pr(H̃ = H1 |H0 ), (5)
ci,j = 1 − αdi , if i = j (13)
Pe,1 = Pr(H̃ = H0 |H1 ), (6)

0, otherwise,

and π0 and π1 are the prior probabilities of H0 and H1 , where di is the degree of node i. This can be written as C =
respectively. In a centralized network all observations are sent IN − αL, where L is the Laplacian matrix corresponding to
to the fusion center. The optimum fusion rule can be written graph G. The elements of matrix L are given by,
as a function of the sum of local log likelihood ratios (LLRs), 
because the observations are independent [23], [25]. Therefore, −1, if (i, j) ∈ {E, L}

li,j = di , if i = j (14)
N
H̃n =1 π0 
0, otherwise.
X 
R= Λn ≷ log = δ, (7)
n=1 H̃n =0 π1
2) Estimation Problem: The distributed estimation problem
is the global LLR. In Equation (7), δ is the fusion threshold and tries to solve an unknown vector parameter by using only
n |H1 )
Λn is the local LLR at the nth node, given by log Pr(o Pr(on |H0 ) .
measurements from neighboring nodes [28], [30], [62]. At
In distributed detection, the statistics stored at each node every discrete time instant k, the nth node takes a scalar
are updated in an iterative manner to reach a consensus about measurement Sn [k] and a 1 × M dimension regression vector
the global statistics. This updating is performed by sharing the rn,k . The objective is to estimate an M × 1 unknown vector
current statistics only with neighboring nodes. Initially, nodes p(0) using the data Sn [k] and rn,k . Each node i has access
perform observations o1 , o2 , . . . , oN , and then every node to time realizations {Si [k], ri,k }, i = 1, 2, . . . , N , as shown in
Pr(on |H1 )
sn , n = 1, 2, . . . , N , computes the LLR, Λn = log Pr(o Fig. 1. The vector parameter follows standard model given by,
n |H0 )
of its measurement on . These LLR values serve as the initial
Sn [k] = rn,k p(0) + zn [k], (15)
states for the decision making process. If at a given discrete
time instant k, the value stored in the node sn is given by 2
where zn [k] is zero-mean Gaussian noise with variance σv,n
vn [k], then, and independent over space and time.
N In a centralized estimation method the problem can be
expressed as a L2 -norm minimization problem. Let Rc =
X
vn [k] = cn,n vn [k − 1] + cn,i vi [k − 1], ∀n (8)
i=1,i6=n
col{r1 , r2 , . . . , rN }, be an (N × M ) global matrix and S =
col{S1 , S2 , . . . , SN }, be an (N ×1) data vector. Then we seek where Rr = diag{Rr,1 , Rr,2 , . . . , Rr,N } is block diagonal. To
an M × 1 vector p(0) that solves ensure the convergence of error vector E[ψ̃ ψ [k]], it requires
2 that λ(GB) < 1 with B = IN M − DRr . Thus, in an adaptive
min E S − Rc p(0) . (16)

network the convergence of distributed estimation is dependent
p(0) 2
on the space-time data statistics (represented by B) and net-
In the case of distributed estimation, the task is to devise work topology (represented by G). Therefore, ultimately the
a robust distributed method that can approach the solution problem is related to efficient network topology development
p(0) and provides a good estimate of that vector to each leading to fast convergence of parameter estimation. The final
node in the network. By utilizing the diffusion LMS Adapt- problem formulation is expressed as,
Then-Combine (ATC) algorithm [52], [62], the problem of 2
C∗ = arg min [GB][k] − ON M ×N M (24)

distributed estimation can be reformulated as
C F
ψn [k] = φn [k − 1] + ?
µn rn,k (Sn [k] − rn,k φn [k − 1]) (17) subject to: CqN = qN , (25)
X where, qN = col{1, 1, . . . , 1} and O is a matrix of zeros.

φn [k − 1] = cn,l ψl [k − 1]. (18)
l∈Nn,[k−1] V. FAULT-R ESILIENT D ISTRIBUTED D ETECTION AND
In the above equations, at any given time [k − 1], it is E STIMATION OVER S MALL W ORLD WSN S
assumed that node n has access to a set of unbiased estimates In this section, the introduction of small world characteris-
ψl [k − 1] from its neighborhood Nn . The local estimates are tics (SWC) using directional beamforming is discussed first.
fused at the nth node using local combiners cn,l resulting in Subsequently, LCMV beamforming leading to higher neigh-
φn [k − 1]. cn,l is the set of local combiners which contains borhood connectivity for sensor nodes is described. LCMV
information about the network topology and satisfies the beamforming exercises a radio area constraint on sensor nodes
following conditions, leading to power efficient small world WSN (SW-WSN) devel-
X opment. In addition, a traffic flow optimization algorithm for
cn,l = 1, and cn,l = 0, ∀l ∈
/ Nn . (19) SW-WSN development is also discussed in this section. The
l∈Nn traffic flow optimization algorithm leads to a low-complexity
SW-WSN development with optimal SWC. Finally, the algo-
After obtaining an aggregate estimate φn [k − 1] for
rithm development for fault-resilient distributed detection and
p(0) , the fusion of φn [k − 1] into the local adaptive
estimation over SW-WSNs is presented.
process is performed to update it to ψn [k]. Here, µn
is the local step size. In terms of global quantities we
define the following, zk = col{z1 [k], z2 [k], . . . , zN [k]}, A. Introduction of SWC using Directional Beamforming
ψ [k] = col{ψ ψ 1 [k], ψ 2 [k], . . . , ψ N [k]}, φ [k − 1] = Various strategies involving introduction of SWC in WSNs
col{φφ1 [k − 1], φ 2 [k − 1], . . . , φ N [k − 1]}, Rk = can be found in literature. Some of them involves usage of
col{r1,k , r2,k , . . . , rN,k }, Sk = col{S1 [k], S2 [k], . . . , SN [k]}, additional wires [35], [39], mobile sensor nodes [63], and high
and p(0) = col{p(0) , p(0) , . . . , p(0) } (N times). Therefore, the capacity heterogeneous sensor nodes [36], [40]. Introduction
model in Equation (15) can be rewritten as of SWC using these methods has several disadvantages [44].
In terms of the risk of having wired shortcuts, a WSN is not
Sk = Rk p(0) + zk . (20)
practical in applications such as battlefield surveillance and
Also, with these relations and expressions, Equations (17) and accident detection. Breakage of wired shortcuts will lead to
(18) can also be written in a matrix form as, loss of SWC in the network. Additionally, to compute the fixed
( length of long-range links a prior in mobile ad hoc networks
φ [k − 1] = G ψ [k − 1] is also not feasible. Utilization of mobile sensor nodes results
ψ [k] = φ[k − 1] + DR?k (Sk − Rk φ[k − 1]), in increased data latency and network deployment cost. In
addition, it yields several unwanted shortcuts in the network,
where G = C ⊗ IM is a N M × N M transition matrix and resulting in the formation of a random network. In order
C is a N × N diffusion combination matrix with entries to address these issues, in this work, a novel directional
[cn,l ]. Here, D = diag{µ1 IM , µ2 IM , . . . , µN IM } is a diagonal beamforming technique is used for the introduction of SWC.
matrix collecting the local step sizes. Now the global error Directional beamforming involves mounting an antenna
vector is given by, array on top of sensor nodes to act as a directional antenna.
ψ [k] = p(0) − ψ [k].
ψ̃ (21) Beamforming is achieved by combining the elements of an-
tenna array in such a way that signals transmitted experience
Utilizing the previous equations, the global error is written as, constructive interference at particular angles. Various methods
[45], [46] have been developed for transmitting a signal of
ψ [k] = G(IN M − DR?k Rk )ψ̃
ψ̃ ψ [k − 1] − GDR?k zk . (22) interest toward specific directions while suppressing its energy
In Equation (22), taking expectation on both sides results in in other directions. Beamforming leads to the formation of
a radiation pattern in the form of lobes at various angles.
ψ [k]] = G(IN M − DRr )E[ψ̃
E[ψ̃ ψ [k − 1]], (23) The direction in which there is larger field strength than in
others is called the main lobe and others are called side lobes.
These lobes can also be characterized on the basis of their
beam length (r), beam width (∆θ), and direction in which the
beam is directed (α). In this work, these parameters for an
individual node are computed to yield a low-complexity SW-
WSN development with optimal SWC. The selection of a set
of nodes used for beamforming and the corresponding optimal
beamforming patterns are obtained utilizing the parameter
Flow between node pairs. The beamforming pattern minimizes
the APL of the network while maintaining high ACC.
B. LCMV Beamforming for Neighborhood Connectivity

It has been shown in [44] that with beamforming, reduction
in network APL is accompanied by loss in connectivity, which
is exhibited by a decrease in ACC of the network. This
decrease in ACC is due to the fact that beamforming node Fig. 2: An illustration of LCMV beamforming performed by
“loses” links to closely located neighbors that are not within sensor nodes. Sensor nodes are denoted using black dots.
the main lobe or side lobes of the radiation pattern. Hence, The omnidirectional radio areas covered by sensor nodes
for maintaining a high clustering coefficient, neighboring are represented using blue circles. Directional beamforming
connectivity of beamforming nodes should be maintained. consisting of the main lobe and side lobes is shown in red.
Two popular directional antenna models, namely sector model
[44], and the keyhole model [43] have been proposed to
simplify the radiation pattern. However, both models fail desired radiation pattern for antennas based on the received
to capture most important features of a directional antenna, signal strength (RSS) measurement model.
i.e., side lobes and nulling capability. In recent literature, a Energy efficiency has always been a critical issue in WSNs
more realistic model called Iris antenna model [45], has been due to the limited powers at sensor nodes. The primary
proposed which can approximate the realistic antenna model motivation behind the use of LCMV beamforming is that
into a simplified model while maintaining its properties. In the new links between node pairs can be introduced while
this paper, we use Iris antenna model [45] to realize beam using the same amount of power as with omnidirectional
pattern of a beamforming node and utilizes a popular spatial sensor nodes. In this work, LCMV beamforming is used with
filtering technique called LCMV beamforming to ensure that a constraint on radio transmission area covered by a sensor
connectivity to the neighborhood nodes is maintained. node. The radio area constraint during signal transmission is
LCMV beamforming is specifically used for controlling as follows. Consider a node i with a directional antenna having
antenna gains in various directions, whose mathematical model l omnidirectional neighbors (omnidirectional radio range r0 ).
is given by, It operates in the mode of directional transmission and omni-
directional reception, i.e., the receiving gain is unity (Gr = 1).
min wH xxH w For maintaining neighborhood connectivity, the signal between
w
subject to: VH w = c. a node and its omnidirectional neighbors should be received
successfully, i.e., the power attenuation β is no greater than
Here, x is the received signal, which is a combination of the threshold β0 [45]. Therefore, the effective aperture (Ae )
transmitted signal (s) and noise (n). If a transmitter is sending of an antenna in the direction of a neighborhood node can be
a signal s, which is directed in some direction specified by expressed in terms of omnidirectional range and is given by,
the vector v, then x = vs + n, where, n is Gaussian noise, λ2 r0ρ
w is the weight vector for a linear combination of the signals Ae = , (27)
4πβ0
from different antennas. (·)H means the Hermitian operation
on (·). In the above mathematical model, the variance of the where ρ is the path loss exponent and λ is the wavelength of
received signal is minimized so that the noise component of the signal. In Equation (27), the shadowing effect is considered
the received signal is reduced. V is a Q × C constraint matrix, to be zero. Hence, the half power beam width of the side lobes
where Q represents the number of antennas mounted on the (θs ) is given by,
antenna array, and C represents the number of constraints. Vec-
 s 
ρ−2
tor c keeps the information about the gain of a beamforming λ r
θs = tan−1  0 . (28)
node. In this work, the weight vector c is evaluated using the 2π β0

LSW (SB×1 (ri , ∆θi , αi )) CSW (SB×1 (ri , ∆θi , αi ))
S∗B×1 (ri , ∆θi , αi ) = arg max − (26)
SB×1 (ri ,∆θi ,αi ) LW (0) CW (0)
Therefore, if a node has K neighborhood nodes then the mini- separated by more than one hop. Direct Flow from a node u
mum radio area required by the node to maintain neighborhood is given by,
connectivity is given by, X
FlowD (u) = Flow(u, v), (31)
θs
A0 = K ×× r02 . (29) v∈Nu
2 where Nu denotes the set of neighborhood nodes of node u.
The remaining area A−A0 , can be used to model the main lobe The parameter Flow between a node pair illustrates the
of the radiation pattern. Here, A is the total radio area covered importance of the link existing between them for the purpose
by the sensor node. Fig. 2 depicts the radio area A and A0 cov- of data transmission across the network. It represents the
ered by a sensor node: directional radio transmission utilizing number of shortest paths between other pair of nodes that
LCMV beamforming is illustrated in red, whereas radio area passes through the given node pair. This implies that the
covered using omnidirectional transmission is shown in blue. paramenter Flow plays an important role in governing the APL
After addressing the challenges of neighboring connectivity of the network. Therefore, a decrease in distance (number of
using LCMV beamforming, optimal beamforming parameters hops) between a node pair having higher Flow also results in
are obtained by maximizing the difference between normalized a decrease in the APL of the entire network. Hence, in this
APL and ACC. Let the total sensor nodes used for beamform- work the beamforming is performed in such a manner that
ing be B = N × ν, where, N is the total number of sensor the nodes having higher Flow values have shorter distances
nodes in the network and ν is the percentage of nodes used between them. Direct Flow for node u is given by,
for beamforming. Let, LSW (SB×1 (ri , ∆θi , αi )), ∀i ∈ B, and
LW (0) denote the APLs of a SW-WSN and a regular WSN FlowD (u) = mru + l∆θu + cαu , (32)
respectively. Similarly, CSW (SB×1 (ri , ∆θi , αi )), ∀i ∈ B and where, ru , ∆θu , and αu are the beam length, beam width, and
CW (0) denote the ACCs of a SW-WSN and a regular WSN, beam direction, respectively, corresponding to the uth node.
respectively. Let SB×1 (ri , ∆θi , αi ) be a B × 1 vector whose We focus our attention toward finding the optimal values for
element Si (ri , ∆θi , αi ) is a 2-dimensional representation of these parameters for every node such that the node pair with a
radiation pattern corresponding to the ith beamforming node higher Flow gets connected leading to maximization of Direct
(1 ≤ i ≤ B). Then, the computation of optimal beamforming Flow for every node. This is illustrated in Fig. 3.
pattern (S∗B×1 (ri , ∆θi , αi )) is carried out using Equation (26). Fig. 3(a) illustrates the beamforming by the uth node. The
Finding the solution of the optimization problem given in sensor nodes in green are falling in the main lode and side
Equation (26) is quite challenging as the relationship be- lobes of the uth node. The nodes inside the main lobe are
tween quantities is quite complex. Hence, we develop a low- shown within yellow beam and leads to Direct Flow. The
complexity technique to find an optimal radiation pattern for nodes which are neither in the main lobe nor in a side lobe
every node. To this end, we resort on SW-WSNs with traffic are indicated in red and cyan. These nodes lead to Indirect
flow optimization. This is discussed in the next section. Flow. The sensor nodes which are within the beam width of
the main lobe while out of range of the main lobe are in cyan.
C. A Low-Complexity Small World WSN Development using Fig. 3(b) illustrates the radio area covered by the main lobe
Traffic Flow Optimization of the uth sensor node with a yellow rectangular box. The
parameters corresponding to this box result in optimal beam
This section describes the development of a SW-WSN
parameters.
using a low-complexity traffic flow optimization method.
Let τ (u, v) be the length of the shortest path between
The method results in optimal beamforming parameters for
nodes u and v. Distances from node u to all other nodes are
individual nodes. Optimal parameters are computed utilizing
represented in a vector Ru = col[du (1), du (2), . . . , du (N )].
traffic flow between node pairs. The parameter Flow is used to
measure the traffic between a selected node pair. Specifically,
Flow between any two nodes u and v of the network is given
by [64],
X σ(s, t)u,v
Flow(u, v) = , (30)
σ(s, t)
s,t∈N t = Const.
Range (r)
FlowD(u,t)
where σ(s, t)u,v is the number of shortest paths between nodes
s and t, which pass through nodes u and v, and σ(s, t) is v
FlowID (u,v)
the total number of shortest paths between nodes s and t. u
Similar to betweenness centrality [64], the Flow between two
nodes is high if there is a larger number of times both nodes
00 3600
lie between the shortest path of randomly selected node pairs (a)
Angle (Dq )
(b)
in the network. Flow through any node can be divided into
two different categories, “Direct Flow”, (FlowD ) and “Indirect Fig. 3: An illustration of (a) Direct Flow and Indirect Flow
Flow”, (FlowID ), as illustrated in Fig. 3. Direct Flow exists between sensor nodes (b) radio area covered by the main lobe
between nodes which are directly connected to each other. On leading to optimal beam parameters is shown using yellow
the other hand, Indirect Flow exists between nodes which are rectangular box.
Algorithm 1 Computation of optimal parameters ru∗ , ∆θu∗ , αu∗ Algorithm 2 Fault-Resilient Distributed Detection and Esti-
1. Initialize Flow(u, v) = 0; mation over a Small World WSN
2. for s, t ∈ N − {u, v} do 1. Initialization: Consider a randomly deployed conventional WSN
3. Use breadth-first search [65] to compute τ (s, t) and σ(s, t); as, G(0) = (N , E, P, B, L), where the set of new links L, and set
4. if τ (s, t)u,v = τ (s, u) + τ (u, v) + τ (v, t) do of beamforming nodes B, are both null sets {∅}.
5. σ(s, t)u,v = σ(s, u) × σ(u, v) × σ(v, t); 2. Compute optimal beam parameters for each node: For all
6. Flow(u, v) = Flow(u, v) + σ(s, t)u,v /σ(s, t); existing nodes in the network, compute the optimal beamforming
7. end parameters by solving the traffic flow optimization:
8. end 
9. Compute Flow(u, v), ∀{u, v} ∈PN using steps 1 to 8. Su∗ (ru∗ , ∆θu∗ , αu∗ ) = arg max FlowD (u)
S(ru ,∆θu ,αu )
10. Initialize FlowD (u) = Flow(u, v) with ru∗ = 0,  Subject to: A − A0 = 2
ru
v∈Nu 2
∆θu = constant
∆θu∗ = 0, αu∗ = 0.
11. Sort Ru , Θu , Flowu in the increasing order of θu (i). 3. Selection of nodes for beamforming: Beamforming nodes are
12. Initialize FlowD (u)j = 0; selected from the network in order to maximize the difference
13. for i ∈ [1, N − 1] do between APL and ACC utilizing the Flow parameter:
14. ∆θmax = 2(A − A0 )/(du (i))2
15. Υ = θu (i) − ∆θmax LSW (SB×1 ) CSW (SB×1 )
S∗B×1 = arg max − (33)
16. $ = θu (i) + ∆θmax SB×1 LW (0) CW (0)
17. for j | θu (j) ∈ [Υ, $] do
4. Small world WSN development using LCMV beamforming:
18. k | θu (k) = θu (j)
P + ∆θmax After the selection of nodes for beamforming, use the LCMV filtering
19. FlowD (u)j∗ = Ω∈[j,k] Flow(u, Ω)
method to ensure neighborhood connectivity for each node based on
20. if FlowD (u)j∗ >= FlowD (u)j directional links:
21. FlowD (u)new = FlowD (u)j∗ + FlowD (u)
22. α∗ = (θu (k) + θu (j))/2 min wH xxH w
w
23. ru∗ = du (i)
24. ∆θ∗ = ∆θmax subject to: VH w = c.
25. end 5. Obtain new network topology: Compute the new network
26. end topology G(N , E, P, B, L) with L, B 6= {∅}.
27. end 6. Analysis of small worldness: The scalar measures [33], [60] of
27. Return ru∗ , ∆θu∗ , αu∗
small worldness [66] are given by, % = LLSW (B)
W (0)
< 1, η = CCSW (B)
W (0)
≈
1, where LW (0) is the APL for a WSN and LSW (B) is the APL
for the corresponding SW-WSN. Similarly, CW (0) is the ACC for a
Angles between node u and other nodes are collected in WSN and CSW (B) is the ACC for the corresponding SW-WSN.
a vector Θu = col[θu (1), θu (2), . . . , θu (N )], where θu (i) 7. Optimal small world characteristics introduction: Repeat steps
2 to 6 in an iterative manner for selection of the best B nodes leading
denotes the azimuthal angle between node u and node i.
to the optimal SWC.
Flows from node u to all other nodes in the network are 8. Performance analysis:
Flowu = col[Flow(u, 1), Flow(u, 2), . . . , Flow(u, N )]. Note
that the sizes of all vectors Ru , Θu , Flowu are (N −1)×1. As Distributed detection Convergence rate: TT(B)
(0)
the parameters of the main lobe vary, neighborhood nodes of Distributed estimation MSD: N1
PN
i=1 Ekp (0)
− ψn [k]k
node u also change and hence FlowD (u) changes. Therefore, Fault-resilience k-Connectivity check ∀k
the optimal parameters are obtained by maximizing FlowD (u),
which is given by, 9. Fault-resilient distributed detection and estimation over a
 SW-WSN: Developed network results in fault-resilient distributed
Su∗ (ru∗ , ∆θu∗ , αu∗ ) = arg max FlowD (u) detection and estimation over a SW-WSN.
S(ru ,∆θu ,αu )
2
 Subject to: A − A0 = ru
2 ∆θu = constant
where, recall that, A and A0 are the total radio area and In these methods, first a network with randomly deployed
radio area covered by side lobes of node u, respectively. sensor nodes is considered. The data traffic load on an
The algorithmic development of the proposed traffic flow individual link is assumed to be known. Data traffic load
optimization method yielding optimal beam parameters is computation is performed using Dijkstra’s algorithm. SWC
given in Algorithm 1. In Algorithm 1, first the number of are introduced in this network at various time instants based
shortest paths passing through the selected node pair, (u, v), on node and network characteristics. Introduction of SWC is
is calculated using the breadth-first search [65]. Subsequently, carried out using beamforming from a selected set of nodes.
flow from node u to other nodes is computed by varying node Nodes used for the beamforming are obtained using traffic flow
v. Finally, optimal parameters ru∗ , ∆θu∗ , αu∗ , corresponding to measurement between node-pairs. Algorithm 2 enumerates the
node u are obtained by maximizing its direct flow. steps for the distributed detection and estimation over a SW-
WSN. The flow diagram corresponding to Algorithm 2 is
presented in Fig. 4. In Fig. 4, the procedure for converting
D. Algorithm Development for Fault-Resilient Distributed De- a regular WSN into a SW-WSN is indicated by beamforming
tection and Estimation over Small World WSNs toward SW-WSN development block. The steps adapted for
This section discusses the proposed methods of fault- performance analysis of the proposed method are indicated
resilient distributed detection and estimation over SW-WSNs. by distributed processing over the developed SW-WSN block.
10
Beamforming toward SW-WSN Development

Input random Select ‘B’
Compute ‘Flow’ Obtain optimal Use LCMV
WSN, initialize nodes leading Obtain the new
between node- beam parameters beamforming
by G(0) = to introduction network topology
pairs within for nodes using to generate
(N , E, P, B, L)
of small world
the network ‘Flow’ metric SW-WSN
characteristics
Are scalar
measures
No
[60], % and
η satisfied?
Distributed Processing over the Developed SW-WSN Yes

Fault-resilient Use performance Select a random Stop
node over which Exchange the data
distributed measures of beamforming for
distributed with neighboring
detection and distributed the introduction
processing needs nodes, collected
estimation over processing of small world
to be performed from sensor nodes
a SW-WSN and robustness characteristics
Fig. 4: Block diagram illustrating the fault-resilient distributed detection and estimation over a small world WSN.
VI. P ERFORMANCE E VALUATION distance between NI mote-pairs. To compute the RSS at the
Performance evaluation of the proposed methods is carried receiver node (Pr ), a total of 10 readings are taken between
out using simulation and real node deployments over a WSN transmitter and receiver. The values of Gt and Gr are 1.5 dBi,
testbed. Detection and estimation performance along with λ is 0.12491 m, and Pt is 10 dBm. The conventional links have
network robustness performance are evaluated over both the a length of 7 m.
1.1 1.2
data sets. Resilience of the network is analyzed through k-
connectivity to highlight the applicability of the proposed 1
1
Normalized APL
Normalized ACC
methods in practice. Performance of the proposed methods 0.9
0.8
is compared with the existing betweenness centrality method APL BC
0.8 APL Flow
[44], and bio-inspired clustering method [43]. In addition, the APL Clustering 0.6
0.7 ACC BC
complexity of the proposed methods is also investigated to ACC Flow
0.4
illustrate their significance in real time applications. 0.6 ACC Clustering
0.5 0.2
0 0.2 0.4 0.6 0.8 1
A. Experimental Setups Fraction of Nodes used for Beamforming
The experimental setups used in the performance evaluation Fig. 5: Variations of APL and ACC with varying fraction of
are discussed in this section. beamforming nodes. The variations are compared among the
1) Simulated WSN: A network of dimension 40 m × 40 m proposed, betweenness centrality, and clustering methods.
is considered for simulation. The total number of sensor nodes
used for simulation is 100. The sensor nodes are uniformly B. Analysis of Small Worldness
distributed over the deployed area. Initially all nodes are Small world characteristics of SW-WSN are often deter-
performing omnidirectional transmission and reception. The mined using APL and ACC. However, analysis of small
conventional radio range of sensor nodes is 8 m. Nodes are worldness [66] of a SW-WSN can be performed by defining
capable of beamforming using directional antennas. Path loss some scalar measures [33], [60]. The scalar measures of small
exponent (α) is 2 and the power attenuation threshold is worldness are given by % = LLSW (B)
W (0)
< 1, η = CCSW (B)
W (0)
≈ 1.
considered to be 0.5. Wavelength of the transmitted signal is
0.125 m. 1.1
2) Real Node Deployment over a WSN Testbed: Perfor- 1 BC

mance evaluation is also conducted over a WSN testbed using 0.9 Flow
Clustering
real sensor node deployment at a residential site of IIT Kanpur.
T(B)/T(0)
0.8
The WSN testbed consists of 80 National Instruments (NI) 0.7
3230 wireless sensor motes. These motes are uniformly dis- 0.6
tributed over an open-space geographical area of dimension 33 0.5
m × 29 m. For RSS measurement, NI wireless gateway 9792
0.4
is used. All the data of real WSN testbed is collected over the 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fraction of Nodes used for Beamforming
NI gateway which is connected through a LAN. The antenna
gain for NI motes and gateway is 1.5 dBi corresponding to Fig. 6: Convergence speed of distributed detection with a
both the transmitter and receiver. The transmitted power is 10 varying fraction of beamforming nodes. The proposed method
dBm. Friis transmission model [38] is used to compute the is compared with betweenness and clustering methods.
11
Table II: Variations in the normalized average path length (N-APL), normalized average clustering coefficient (N-ACC),
distributed detection convergence speed (T (B)/T (0)), mean squared deviation (MSD), and k−connectivity check over a
real WSN testbed.
Real WSN Performance measures used and corresponding results obtained over a small world WSN
Testbed
% of Nodes T (B)/T (0) MSD MSD MSD
Method Beamforming N-APL N-ACC (#500) (#1000) (#1500) 2-C 3-C 4-C 5-C
0 1.0000 1.0000 1.0000 -147.48 -153.26 -153.14 0.0000 0.0000 9.7× 10−5 3.9× 10−4
20 0.6087 1.0365 0.2076 -150.16 -154.80 -155.37 0.0000 0.0000 9.7× 10−5 3.3× 10−4
Proposed 40 0.5358 1.0483 0.1579 -151.04 -158.76 -156.46 0.0000 0.0000 9.7× 10−5 2.0× 10−4
Method
60 0.4959 1.0279 0.1651 -150.66 -161.37 -157.14 0.0000 0.0000 4.6× 10−5 2.0× 10−4
80 0.4742 1.0204 0.1651 -67.02 -83.57 -87.09 0.0000 0.0000 4.1× 10−5 8.3× 10−5
100 0.4616 1.0038 0.1605 -150.15 -164.05 -161.71 0.0000 0.0000 2.4× 10−5 1.1× 10−5
0 1.0000 1.0000 1.0000 -147.81 -152.96 -151.49 0.0000 0.0000 9.7× 10−5 3.9× 10−4
Bio-inspired 20 0.6837 0.9282 0.3373 -151.61 -153.11 -156.56 0.0132 0.0263 0.0396 0.0509
Clustering 40 0.7228 0.9588 0.4286 -149.49 -157.33 -153.74 0.0132 0.0263 0.0395 0.0516
Method 60 0.7751 0.9965 0.4816 -149.42 -154.56 -154.17 0.0000 0.0000 6.1× 10−5 0.0989
80 0.6906 0.8831 0.4820 -147.50 -152.08 -152.42 0.0130 0.0260 0.0390 0.0516
100 0.6906 1.0000 0.4820 -148.14 -153.82 -152.57 0.0000 0.0000 9.7× 10−5 3.9× 10−4
0 1.0000 1.0000 1.0000 -148.89 -152.21 -150.90 0.0000 0.0000 9.7× 10−5 3.9× 10−4
Betweenness 20 0.8536 0.9022 0.6431 -147.26 -150.76 -152.16 0.0000 0.0018 0.0053 0.0084
Centrality 40 0.6907 0.6788 0.3354 -145.85 -150.23 -149.67 0.0395 0.0768 0.1123 0.1424
Method 60 0.6592 0.4299 0.3738 -148.82 -154.88 -155.93 0.0135 0.0274 0.0416 0.0513
80 0.6758 0.2690 0.3291 -100.66 -129.99 -136.17 0.0130 0.0263 0.0400 0.0493
100 0.6551 0.1473 0.3746 -147.14 -152.72 -151.82 0.0250 0.0494 0.0732 0.0821
Here LW (0) and LSW (B) are the APLs of WSN and SW- WSN can be seen in Fig. 5, which show that the APL of the
WSN, respectively. Similarly, CW (0) and CSW (B) quantify proposed method decreases as the fraction of beamforming
the ACCs of WSN and SW-WSN, respectively. The analysis nodes increases. In similar conditions the ACC is almost con-
of SWC is carried out using both the data sets. For observing stant. It is noted that with 20% of beamforming nodes, the APL
the variations in APL and ACC, the fraction of beamforming of the proposed method is 69.68% when compared to conven-
nodes is increased from 0 to 1. Variations in APL and ACC tional WSN. This reduction in APL is significantly lower when
using the proposed method are also compared with that of compared to betweenness centrality and clustering methods.
the existing methods. The results obtained over the simulated For the betweenness centrality and clustering methods this
-125 -125
Mean Square Deviation (dB)
-130 BC -130 BC
Flow Flow
-135 Clustering -135 Clustering
-140 -140
-145 -145
-150 -150
-155 -155
-160 -160
400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500
Number of Iterations Number of Iterations
(a) (b)
-125 -120
-130 -125
BC BC
Flow -130 Flow
-135
Clustering Clustering
-135
-140
-140
-145
-145
-150
-150
-155
-155
-160 -160
-165 -165
400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500
Number of Iterations Number of Iterations
(c) (d)
Fig. 7: Illustration of estimation performance analysis (MSD in dB) with a varying number of iterations. The results are
obtained with a varying fraction of nodes used for beamfoming (a) 0.1 fraction (b) 0.4 fraction (c) 0.8 fraction and (d) All
the nodes are used for beamforming. The results are obtained using the proposed method (violet), clustering method (yellow),
and betweenness centrality method (blue) over the simulated WSN.
12
reduction is 94.52% and 76.36%, respectively. The ACCs of methods. These results are noted when 20% of nodes perform
the proposed and clustering methods are almost equal to 1. beamforming in the network. It can also be seen that as the
For the betweenness centrality method, there is a reduction fraction of beamforming nodes increases, the betweenness
of 13.44% in ACC when compared to the conventional WSN. centrality method results in better performance compared to
These variations over a real WSN testbed can be seen in Table the other methods. However, it should be noted that in this
II, which clearly illustrates the significance of our proposed case the ACC of the network decreases by 50% compared
method. to a conventional WSN, leading to a random network. The
results over a real WSN testbed can also be seen from Table
C. Performance Analysis of Distributed Detection II. A 79.24% improvement in convergence speed over the
The analysis of distributed detection is performed utilizing real WSN testbed can be seen using the proposed method
the Bayesian model given in Equation (4), with a prior when only 20% of nodes perform beamforming. In similar
probabilities π0 = π1 = 0.5, which lead to the optimum conditions, the convergence speed increases by 35.69% and
fusion threshold of δ = 0. The signal means µ1 and µ2 66.27% respectively when using the betweenness centrality
are modeled as µ1 = −µ2 = 1, while the noise variance and clustering methods.
σ 2 = 0.501. The performance of distributed detection is
measured in terms of number of iterations required
2 by the

(k) 1
D. Performance Analysis of Distributed Estimation
weight matrix leading to C (α, L) − N 1N ×N → . In

F To measure performance of distributed estimation the re-
case of simulations, the error () is considered to be 0.01% of
that in the conventional WSN. Variation of the convergence gression data is generated by regressors with shifted structure
rate for distributed detection can be seen in Fig. 6. In case of size M , i.e., rn,k = col{rn (k), rn (k − 1), . . . , rn (k −
of a SW-WSN, a phase transition phenomenon in algebraic M + 1)} with rn (k) being a time series generated as follows:
connectivity is observed. Algebraic connectivity of a network rn (k) = ζn rn (k − 1) + ξn zn (k) for k > −∞, where
is the second smallest eigenvalue of its Laplacian matrix and is ζn ∈ [0, 1) is the correlation index, zn (k) is a spatially
a measure of speed of solving consensus problems in networks independent
q white Gaussian process with unit variance, and
ξn = σn,k 2 (1 − ζ 2 ). The scalar observation S [k] is obtained
[24]. In a SW-WSN, an increase in algebraic connectivity n n
utilizing addition of new links increases the speed of solving according to the model given in Equation (15) with unknown √
consensus problems. The convergence speed increases by M -dimensional vector set as p(0) = col{1, 1, . . . , 1}/ M .
47.32% using our proposed method when compared to a The performance analysis of distributed estimation is shown
conventional WSN. This increase is 28.99% and 11.22%, in Fig. 7, which illustrates a reduced mean squared deviation
respectively, using the clustering and betweenness centrality (MSD) using the proposed method when compared to a
0.025 0.2
Fraction of Disconnectivity in
BC
Flow
0.02
3 Connectivity Check
BC 0.15 Clustering
Flow
0.015 Clustering
0.1
0.01
0.05
0.005
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Fraction of Nodes used for Beamforming Fraction of Nodes used for Beamforming
(a) (b)
0.08 0.5
0.4 BC
0.06 Flow
BC 0.3 Clustering
Flow
0.04 Clustering 0.2
0.1
0.02
0
0 -0.1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Fraction of Nodes used for Beamforming Fraction of Nodes used for Beamforming
(c) (d)
Fig. 8: Illustration of robustness analysis (k−connectivity check) with varying fraction of nodes used for beamforming. The
results are obtained for different k−connectivity (a) 2−connectivity check (b) 3−connectivity check (c) 4−connectivity check
and (d) 5−connectivity check. The results are obtained using the proposed method (violet), clustering method (yellow), and
betweenness centrality method (blue) over the simulated WSN.
13
conventional WSN and other existing methods, namely clus- remains connected with its neighborhood nodes. Beamforming
tering, and betweeneess centrality methods. The performance only introduces new links in the network. The results of k-
is observed with a varying fraction of sensor nodes used for connectivity over a real WSN testbed can be seen in Table
beamforming. II. Improved performance of k-connectivity is also noted
Figs. 7(a), 7(b), 7(c), and 7(d) correspond to 10%, 40%, using the proposed method when compared to other existing
80%, and 100% nodes used for beamforming. Fig. 7(c) illus- methods over a real WSN testbed.
trates a reduced MSD as the number of iterations increases.
It can be seen that as the number of iterations increases F. Complexity Analysis of Traffic Flow Optimization Method
the MSD of the proposed method decreases to a minimum
of −160 dB. However for both the betweenness centrality The time complexity of the breadth-first search algorithm
and clustering methods the minimum MSD obtained is −153 [65] toward computation of both the shortest path and the
dB. The distributed estimation performance over a real WSN number of shortest paths is O(M + N ). Further, computation
testbed is shown in Table II. Improved performance using the of Flow between two nodes results in a complexity of order
proposed method is also observed for this case of real WSN O(M N 2 + N 3 ) or O(N 3 ) as the network is considered to be
testbed when compared to the existing methods. sparse. Therefore, the time complexity to compute Flow across
the network is O(N 5 ). The objective function developed for
maximization of Flow is a convex problem and can be solved
E. Network Resilience Analysis
in polynomial time. The algorithm involves maximization of
Node-failure resilience of the network is measured using the Direct Flow for every node leading to optimal beamform-
network responses to a perturbation. However, it is difficult to ing parameters for each node. In this context, sorting of θu (i)
see the quantitative assessment of network resilience. In this has a complexity of O(N × log N ). In addition, computation
paper, we use resilience factor (RF ) [21], [22] to measure the of optimal parameters, FlowD (u), r, ∆θ has time complexity
network resilience quantitatively, which is given by of O(a × N ). Here a is a constant which is obtained by the
PN
k(i) number of times the inner “for loop” runs. Hence the time
RF = i=2 . (34) complexity for computing optimal beamforming parameters is
N −2
O(N × (log N + a)).
In Equation (34), k(i) denotes the percentage of node com-
binations that guarantees partial i-connectivity. It is assumed
that all networks considered are 1-connected and n-connected, VII. C ONCLUSIONS AND F UTURE W ORK
thus these cases are excluded from observations. Even though In this paper, novel methods of distributed detection and dis-
the resilience factor gives us good information about network tributed estimation over a SW-WSN have been proposed. The
resilience, the challenging issue is it’s exponential complexity. methods utilize traffic flow between node pairs and result in
Therefore, rather than computing k(i) for all i’s, in this paper a robust and low-complexity SW-WSN development. A novel
for illustrating the network resilience performance we consider LCMV beamforming technique is used for the introduction
i ∈ {2, 3, 4, 5}. of SWC in the network. LCMV beamforming increases the
Fig. 8 illustrates the k-connectivity check, which is used as feasibility of the network utilization in real time applications.
a measure of robustness of the SW-WSN to the node failure. Experimental results are obtained over both computer sim-
The SW-WSN developed using the proposed method leads ulated WSN and real WSN testbed. Results obtained using
to a high algebraic connectivity [24] and results in a more the proposed methods are compared with existing methods in
robust network in the context of node failures. Connectivity literature, namely the betweenness centrality and bio-inspired
performance of the proposed method is also compared with clustering methods.
that of state-of-the-art methods. An important aspect to note In case of the real WSN testbed, distributed detection
here is that, with beamforming, directional nodes become speed increases by 84.21% using the proposed method when
significantly important than omnidirectional nodes. The failure compared to a conventional WSN. In this condition, only 40%
of such nodes will have a considerable impact on the network of the nodes are used for beamforming. This increase is by
since the network connectivity depends heavily on them. 57.14% and 66.46%, respectively, using the clustering and
Therefore, while doing k-connectivity check it can be noted betweenness centrality methods under similar conditions. The
that as the number of beamforming nodes increases in case of MSD of parameter estimation reduces to a minimum value
betweenness centrality and clustering methods, the failure of of −164.05 dB using the proposed method. This value for
these nodes results in higher disconnectivity in the network. the clustering and betweenness centrality methods is −157.33
Fig. 8(a) shows results for 2-connectivity check. In case of dB and −155.93 dB, respectively. The MSD in case of the
betweenness centrality and clustering methods the disconnec- conventional WSN is −126 dB. The 2-disconnectivity and
tivity of the network increases to 0.82% and 2.3% respectively 3-disconnectivity using the proposed method are zero in all
when compared to a conventional WSN. This result is obtained cases, while a maximum of 9.7× 10−5 and 8.3× 10−5 fraction
when 30% of the nodes perform beamforming. In case of of 4-disconnectivity and 5-disconnectivity are noted using
a conventional WSN the disconnectivity is 0.35%. However, the proposed method, respectively. The disconnectivity results
using the proposed method the disconnectivity in the network obtained using the proposed method illustrate a significant im-
remains almost equal to that of the conventional WSN. This provement when compared to the clustering and betweenness
is because the proposed method makes sure that the node centrality methods. To conclude, significant improvements
14
over distributed detection, distributed estimation, and network [17] S. Kar, J. M. Moura, and K. Ramanan, “Distributed parameter estimation
robustness are obtained utilizing the proposed methods when in sensor networks: Nonlinear observation models and imperfect com-
munication,” IEEE Trans. Inf. Theory, vol. 58, no. 6, pp. 3575–3605,
compared to two existing methods. 2012.
Mounting an antenna array on a node generally increases [18] X. Yin, Z. Li, L. Zhang, and M. Han, “Distributed state estimation
the hardware complexity of the node. Therefore, collaborative of sensor-network systems subject to markovian channel switching with
application to a chemical process,” IEEE Trans. Syst., Man, and Cybern.,
beamforming from the group of sensor nodes is interesting and Syst., vol. 48, no. 6, pp. 864–874, 2018.
shall be investigated for SW-WSN development in a future [19] D. L. Guidoni, R. A. Mini, and A. A. Loureiro, “On the design of
work. resilient heterogeneous wireless sensor networks based on small world
concepts,” Computer Networks, vol. 54, no. 8, pp. 1266–1281, 2010.
[20] B. Park, J. Nah, J.-Y. Choi, I.-J. Yoon, and P. Park, “Robust wireless
ACKNOWLEDGEMENTS sensor and actuator networks for networked control systems,” Sensors,
vol. 19, no. 7, p. 1535, 2019.
The authors would like to thank the Editor and the anony- [21] J. Pu, Z. Xiong, and X. Lu, “Fault-tolerant deployment with k-
mous reviewers for constructive comments which greatly connectivity and partial k-connectivity in sensor networks,” Wireless
Commun. and Mobile Comput., vol. 9, no. 7, pp. 909–919, 2009.
helped to improve the quality and clarity of this paper. [22] R. M. Salles and D. A. Marino, “Strategies and metric for resilience
in computer networks,” The Computer J., vol. 55, no. 6, pp. 728–739,
2012.
R EFERENCES [23] S. Kar and J. M. Moura, “Sensor networks with random links: Topology
design for distributed consensus,” IEEE Trans. Signal Process., vol. 56,
[1] Z. Chu, F. Zhou, Z. Zhu, R. Q. Hu, and P. Xiao, “Wireless powered
no. 7, pp. 3315–3326, 2008.
sensor networks for internet of things: Maximum throughput and optimal
[24] R. Olfati-Saber, “Ultrafast consensus in small-world networks,” in Pro-
power allocation,” IEEE Internet Things J., vol. 5, no. 1, pp. 310–321,
ceedings of the 2005, American Control Conference, 2005. IEEE, 2005,
2018.
pp. 2371–2378.
[2] C. Chen, J. Yan, N. Lu, Y. Wang, X. Yang, and X. Guan, “Ubiquitous
[25] S. A. Aldosari and J. M. Moura, “Distributed detection in sensor
monitoring for industrial cyber-physical systems over relay-assisted
networks: Connectivity graph and small world networks,” in Conference
wireless sensor networks,” IEEE Trans. on Emerg. Topics in Comput.,
Record of the Thirty-Ninth Asilomar Conference onSignals, Systems and
vol. 3, no. 3, pp. 352–362, 2015.
Computers, 2005. IEEE, 2005, pp. 230–234.
[3] M. Li and H.-J. Lin, “Design and implementation of smart home
[26] S. A. Aldosari and J. M. Moura, “Topology of sensor networks in
control systems based on wireless sensor networks and power line
distributed detection,” in Int. Conf. on Acoustics Speech and Signal
communications,” IEEE Trans. Ind. Electron., vol. 62, no. 7, pp. 4430–
Process. IEEE, 2006, pp. V–V.
4442, 2015.
[4] A.-J. Garcia-Sanchez, F. Garcia-Sanchez, and J. Garcia-Haro, “Wireless [27] S. Kar, S. Aldosari, and J. M. Moura, “Topology for distributed inference
sensor network deployment for integrating video-surveillance and data- on graphs,” IEEE Trans. Signal Process., vol. 56, no. 6, pp. 2609–2613,
monitoring in precision agriculture over distributed crops,” Computers 2008.
and Electronics in Agriculture, vol. 75, no. 2, pp. 288–303, 2011. [28] Y. Liu, C. Li, W. K. Tang, and Z. Zhang, “Distributed estimation over
[5] S. Gandham, R. C. Prasad, A. R. Singh, N. Yadav, K. Deen, and V. S. complex networks,” Information Sciences, vol. 197, pp. 91–104, 2012.
Malhotra, “High availability of collectors of traffic reported by network [29] S. Shahrampour, A. Rakhlin, and A. Jadbabaie, “Distributed detection:
sensors,” Jan. 15 2019, US Patent App. 10/181,987. Finite-time analysis and impact of network topology,” IEEE Trans.
[6] O. Salem, A. Serhrouchni, A. Mehaoua, and R. Boutaba, “Event Autom. Control, vol. 61, no. 11, pp. 3256–3268, 2016.
detection in wireless body area networks using kalman filter and power [30] Y. Liu, C. Yang, W. K. Tang, and C. Li, “Optimal topological design
divergence,” IEEE Trans. on Netw. and Service Manag., vol. 15, no. 3, for distributed estimation over sensor networks,” Information Sciences,
pp. 1018–1034, 2018. vol. 254, pp. 83–97, 2014.
[7] M. Erol-Kantarci and H. T. Mouftah, “Wireless sensor networks for [31] J. Yan, C. Chen, X. Luo, H. Liang, X. Guan, and X. Yang, “Topology
cost-efficient residential energy management in the smart grid,” IEEE optimisation-based distributed estimation in relay assisted wireless sen-
Trans. Smart Grid, vol. 2, no. 2, pp. 314–325, 2011. sor networks,” IET Control Theory & Applications, vol. 8, no. 18, pp.
[8] A. Dogandzic and B. Zhang, “Distributed estimation and detection 2219–2229, 2014.
for sensor networks using hidden markov random field models,” IEEE [32] S. Milgram, “The small world problem,” Psychology today, vol. 2, no. 1,
Trans. Signal Process., vol. 54, no. 8, pp. 3200–3215, 2006. pp. 60–67, 1967.
[9] A. Bertrand, “Distributed signal processing for wireless eeg sensor [33] D. J. Watts and S. H. Strogatz, “Collective dynamics of small world
networks,” IEEE Trans. Neural Syst. and Rehabil. Eng., vol. 23, no. 6, networks,” Nature, vol. 393, no. 6684, pp. 440–442, 1998.
pp. 923–935, 2015. [34] A. Helmy, “Small worlds in wireless networks,” IEEE Commun. Lett.,
[10] J. Guo, U. Rogers, X. Li, and H. Chen, “Secrecy constrained distributed vol. 7, no. 10, pp. 490–492, 2003.
detection in sensor networks,” IEEE Trans. Signal and Inf. Process. [35] C. K. Verma, B. R. Tamma, B. Manoj, and R. Rao, “A realistic small-
Netw., vol. 4, no. 2, pp. 378–391, 2018. world model for wireless mesh networks,” IEEE Commun. Lett., vol. 15,
[11] N.-T. Nguyen, B.-H. Liu, S.-I. Chu, and H.-Z. Weng, “Challenges, no. 4, pp. 455–457, 2011.
designs, and performances of a distributed algorithm for minimum- [36] D. L. Guidoni, R. A. Mini, and A. A. Loureiro, “Applying the small
latency of data-aggregation in multi-channel wsns,” IEEE Trans. on world concepts in the design of heterogeneous wireless sensor net-
Netw. and Service Manag., vol. 16, no. 1, pp. 192–205, 2018. works,” IEEE Commun. Lett., vol. 16, no. 7, pp. 953–955, 2012.
[12] S. Kar and J. M. Moura, “Distributed consensus algorithms in sensor [37] O. J. Pandey and R. M. Hegde, “Low-latency and energy-balanced
networks: Quantized data and random link failures,” IEEE Trans. Signal data transmission over cognitive small world wsn,” IEEE Trans. Veh.
Process., vol. 58, no. 3, pp. 1383–1400, 2010. Technol., vol. 67, no. 8, pp. 7719–7733, 2018.
[13] Y. Chen, S. Kar, and J. M. Moura, “Resilient distributed estimation [38] O. J. Pandey and R. M. Hegde, “Node localization over small world wsns
through adversary detection,” IEEE Trans. Signal Process., vol. 66, no. 9, using constrained average path length reduction,” Ad Hoc Networks,
pp. 2455–2469, 2018. vol. 67, pp. 87–102, 2017.
[14] C. Zhou, C.-K. Tham, and M. Motani, “Finding decomposable models [39] G. Sharma and R. R. Mazumdar, “A case for hybrid sensor networks,”
for efficient distributed inference over sensor networks,” IEEE Trans. IEEE/ACM Trans. on Netw., vol. 16, no. 5, pp. 1121–1132, 2008.
Mobile Comput., vol. 18, no. 1, pp. 70–83, 2019. [40] T. Zhang, J. Cao, Y. Chen, L. Cuthbert, and M. Elkashlan, “A small
[15] I. Nevat, G. W. Peters, and I. B. Collings, “Distributed detection in world network model for energy efficient wireless networks,” IEEE
sensor networks over fading channels with multiple antennas at the Commun. Lett., vol. 17, no. 10, pp. 1928–1931, 2013.
fusion centre,” IEEE Trans. signal process., vol. 62, no. 3, pp. 671– [41] H. Hong, M.-Y. Choi, and B. J. Kim, “Synchronization on small-world
683, 2014. networks,” Physical Review E, vol. 65, no. 2, p. 026139, 2002.
[16] S. Li, S. Zhao, P. Yang, P. Andriotis, L. Xu, and Q. Sun, “Distributed [42] O. J. Pandey, A. Mahajan, and R. M. Hegde, “Joint localization and data
consensus algorithm for events detection in cyber physical systems,” gathering over a small-world wsn with optimal data mule allocation,”
IEEE Internet Things J., 2019. IEEE Trans. Veh. Technol., vol. 67, no. 7, pp. 6518–6532, 2018.
15
[43] R. Agarwal, A. Banerjee, V. Gauthier, M. Becker, C. K. Yeo, and B. S. Om Jee Pandey (S’16, M’19) received the B.Tech.
Lee, “Achieving small-world properties using bio-inspired techniques in degree in Electronics and Communication Engineer-
wireless networks,” The Computer J., vol. 55, no. 8, pp. 909–931, 2012. ing from Uttar Pradesh Technical University, Luc-
[44] A. Banerjee, R. Agarwal, V. Gauthier, C. K. Yeo, H. Afifi, and F. B.-S. know, India, in 2008 and the M.Tech. degree in Dig-
Lee, “A self-organization framework for wireless ad hoc networks as ital Communications from ABV-Indian Institute of
small worlds,” IEEE Trans. on Veh. Technol., vol. 61, no. 6, pp. 2659– Information Technology and Management, Gwalior,
2673, 2012. India, in 2013. He completed his Ph.D degree from
[45] Q. Wang, H.-N. Dai, Z. Zheng, M. Imran, and A. Vasilakos, “On the Department of Electrical Engineering, Indian In-
connectivity of wireless sensor networks with directional antennas,” stitute of Technology Kanpur (IITK), Kanpur, India
Sensors, vol. 17, no. 1, p. 134, 2017. in 2019. From 2008 to 11, he worked with Escorts
[46] A. Bertrand and M. Moonen, “Distributed node-specific lcmv beamform- Limited and FANUC India Private Limited. He was
ing in wireless sensor networks,” IEEE Trans. Signal Process., vol. 60, a Senior Lecturer with Jaipur Engineering College and Research Center, Jaipur
no. 1, pp. 233–246, 2012. during 2013–14. He also worked as a Postdoctoral Fellow with Commu-
[47] D. Blatt and A. Hero, “Distributed maximum likelihood estimation for nications Theories Research Group (CTRG) Department of Electrical and
sensor networks,” in 2004 IEEE International Conference on Acoustics, Computer Engineering Engineering, University of Saskatchewan, Saskatoon,
Speech, and Signal Processing, vol. 3. IEEE, 2004, pp. iii–929. Canada after completion of his Ph.D program. Currently, he is working as an
[48] A. Ribeiro and G. B. Giannakis, “Bandwidth-constrained distributed Assistant Professor with Department of Electronics and Communication En-
estimation for wireless sensor networks-part i: Gaussian case,” IEEE gineering SRM University AP, Andhra Pradesh. His research interest includes
Trans. Signal Process., vol. 54, no. 3, pp. 1131–1143, 2006. the signal processing for wireless networks with a specific focus on robust
[49] T. Zhao and A. Nehorai, “Distributed sequential bayesian estimation sensor node localization and tracking over wireless ad hoc networks. He also
of a diffusive source in wireless sensor networks,” IEEE Trans. Signal works on related areas such as low-latency data transmission, data aggregation,
Process., vol. 55, no. 4, pp. 1511–1524, 2007. and distributed detection and estimation in wireless sensor networks. He is a
[50] J. Li and G. AlRegib, “Rate-constrained distributed estimation in wire- member of IEEE and also works as a reviewer for various reputed journals
less sensor networks,” IEEE Trans. Signal Process., vol. 55, no. 5, pp. of IEEE including IEEE TVT, IEEE Access, IEEE IoT and IEEE TCOM.
1634–1643, 2007.
[51] S. Marano, V. Matta, and P. Willett, “Distributed estimation in large
wireless sensor networks via a locally optimum approach,” IEEE Trans.
Signal Process., vol. 56, no. 2, pp. 748–756, 2008.
[52] F. S. Cattivelli and A. H. Sayed, “Diffusion lms strategies for distributed
estimation,” IEEE Trans. Signal Process., vol. 58, no. 3, pp. 1035–1048,
2010.
[53] S. Barbarossa, S. Sardellitti, and P. Di Lorenzo, “Distributed detection
and estimation in wireless sensor networks,” in Academic Press Library
in Signal Processing. Elsevier, 2014, vol. 2, pp. 329–408.
[54] Z.-G. Wu, Z. Xu, P. Shi, M. Z. Chen, and H. Su, “Nonfragile state Ved Gautam currently working at Urvija AI re-
estimation of quantized complex networks with switching topologies,” ceived his bachelor’s degree in Electrical Engineer-
IEEE Trans. Neural Netw. Learn. Syst., no. 99, pp. 1–11, 2018. ing from the Indian Institute of Technology Kanpur
[55] S. Appadwedula, V. V. Veeravalli, and D. L. Jones, “Energy-efficient (IITK). From 2017-19, he had worked as an under-
detection in sensor networks,” IEEE J. on Sel. Areas in Commun., graduate researcher in the Sensor Networks Lab of
vol. 23, no. 4, pp. 693–702, 2005. Electrical Engineering Department at IIT Kanpur.
[56] V. Saligrama, M. Alanyali, and O. Savas, “Distributed detection in sensor His current research interests include distributed
networks with packet losses and finite capacity links,” IEEE Trans. computation, use of machine learning algorithms
Signal Process., vol. 54, no. 11, pp. 4118–4132, 2006. for enhancement of network performance and dis-
[57] W. Li and H. Dai, “Distributed detection in wireless sensor networks tributed learning in Wireless Sensor Networks.
using a multiple access channel,” IEEE Trans. Signal Process., vol. 55,
no. 3, pp. 822–833, 2007.
[58] F. S. Cattivelli and A. H. Sayed, “Distributed detection over adaptive
networks using diffusion adaptation,” IEEE Trans. Signal Process.,
vol. 59, no. 5, pp. 1917–1932, 2011.
[59] M. Fanaei, Distributed Detection and Estimation in Wireless Sensor
Networks. West Virginia University, 2016.
[60] R. Albert and A.-L. Barabási, “Statistical mechanics of complex net-
works,” Reviews of modern physics, vol. 74, no. 1, p. 47, 2002.
[61] J. Tang, S. Scellato, M. Musolesi, C. Mascolo, and V. Latora, “Small-
world behavior in time-varying graphs,” Physical Review E, vol. 81,
no. 5, p. 055101, 2010.
[62] C. G. Lopes and A. H. Sayed, “Diffusion least-mean squares over Ha H. Nguyen received the B.Eng. degree from
adaptive networks: Formulation and performance analysis,” IEEE Trans. Hanoi University of Technology (HUT), Hanoi,
on Signal Process., vol. 56, no. 7, pp. 3122–3136, 2008. Vietnam, in 1995, the M.Eng. degree from the Asian
[63] C.-J. Jiang, C. Chen, J.-W. Chang, R.-H. Jan, and T. C. Chiang, “Con- Institute of Technology (AIT), Bangkok, Thailand,
struct small worlds in wireless networks using data mules,” in Sensor in 1997, and the Ph.D. degree from the University
Networks, Ubiquitous and Trustworthy Computing, 2008. SUTC’08. of Manitoba, Winnipeg, MB, Canada, in 2001, all in
IEEE International Conference on. IEEE, 2008, pp. 28–35. electrical engineering. He joined the Department of
[64] U. Brandes, “On variants of shortest-path betweenness centrality and Electrical and Computer Engineering, University of
their generic computation,” Social Networks, vol. 30, no. 2, pp. 136– Saskatchewan, Saskatoon, SK, Canada, in 2001, and
145, 2008. became a full Professor in 2007. He currently holds
[65] A. Yoo, E. Chow, K. Henderson, W. McLendon, B. Hendrickson, and the position of NSERC/Cisco Industrial Research
U. Catalyurek, “A scalable distributed parallel breadth-first search algo- Chair in Low-Power Wireless Access for Sensor Networks. His research
rithm on bluegene/l,” in Proceedings of the 2005 ACM/IEEE conference interests fall into broad areas of Communication Theory, Wireless Communi-
on Supercomputing. IEEE Computer Society, 2005, p. 25. cations, and Statistical Signal Processing. Dr. Nguyen was an Associate Editor
[66] D. S. Bassett and E. Bullmore, “Small-world brain networks,” The for the IEEE Transactions on Wireless Communications and IEEE Wireless
neuroscientist, vol. 12, no. 6, pp. 512–523, 2006. Communications Letters during 2007-2011 and 2011-2016, respectively. He
currently serves as an Associate Editor for the IEEE Transactions on Vehicular
Technology. He is a coauthor, with Ed Shwedyk, of the textbook ”A First
Course in Digital Communications” (published by Cambridge University
Press). Dr. Nguyen is a Fellow of the Engineering Institute of Canada (EIC)
and a Registered Member of the Association of Professional Engineers and
Geoscientists of Saskatchewan (APEGS).
16
Mahendra K. Shukla received the Ph.D. degree

in electronics and communication engineering from
the Indian Institute of Information Technology Al-
lahabad, Prayagraj, India, in 2018. Currently, he is
working as a Postdoctoral Fellow with Communica-
tions Theories Research Group (CTRG), Department
of Electrical and Computer Engineering, University
of Saskatchewan, Saskatoon, Canada. He is a regular
reviewer of premier journals including the IEEE
TRANSACTIONS ON VEHICULAR TECHNOL-
OGY, IEEE TRANSACTIONS ON NETWORK
SCIENCE AND ENGINEERING, IEEE SYSTEMS JOURNAL, IEEE COM-
MUNICATIONS LETTERS, IET COMMUNICATIONS. His research in-
terests include cooperative relaying, physical layer security, nonorthogonal
multiple access, energy harvesting, and signal processing.
Rajesh M. Hegde (M’09, SM’16) joined IIT Kanpur

as an Assistant Professor of EE in May 2008 after
obtaining a Ph.D degree in Computer Science and
Engineering from the Indian Institute of Technology
Madras in 2005. He became an Associate Profes-
sor in 2012 and a Full Professor of EE in July
2016. He currently holds the Umang Gupta Chair
position at IIT Kanpur. He was also awarded the
P.K. Kelkar Research Fellowship between 2009-13.
Between 2005 to 2008, He worked as a Researcher
at the California Institute of Telecommunication and
Information Technology (CALIT2) and concurrently as a lecturer (2007) at
the Department of Electrical Engineering at the University of California San
Diego, USA. Prior to 2005, He was involved with teaching undergraduate
engineering courses at various levels in India. He was elected as a Senior
Member of IEEE in 2017 and is an affiliate member of IEEE-AASP
technical committee. He has established two research Laboratories at IIT
Kanpur namely, Multimodal Information Systems Lab and Wireless Sensor
Networks Lab with funding obtained from BSNL, DST, MietY, LG Soft,
Samsung Research and Indian space research organization. He has published
prolifically at several international conferences and journals in the area of
signal processing, communication and networks. He is also a member of the
National working groups of ITU-T (NWG-16 and NSG-6) on developing
multimedia applications. He actively teaches both undergraduate and post
graduate courses related to electronics, digital signal processing, statistical
signal processing, array signal processing, wireless sensor networks, and
digital speech processing. Additional biographic information can be found
at the URL: http://home.iitk.ac.in/ rhegde.

3s PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

3s PDF

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in a future issue of this journal, but has not been

Fault-Resilient Distributed Detection and Estimation

degrees of separation”. In [33], small world phenomena have A. Network Model

where Hn denotes the decision of node sn . The updating rule

X where, qN = col{1, 1, . . . , 1} and O is a matrix of zeros.

B. LCMV Beamforming for Neighborhood Connectivity

Beamforming toward SW-WSN Development

Distributed Processing over the Developed SW-WSN Yes

2) Real Node Deployment over a WSN Testbed: Perfor- 1 BC

Mahendra K. Shukla received the Ph.D. degree

Rajesh M. Hegde (M’09, SM’16) joined IIT Kanpur

You might also like