You are on page 1of 16

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/TWC.2021.3062568, IEEE
Transactions on Wireless Communications

Energy-efficient Edge Computing Framework for


Decentralized Sensing in WSN-assisted IoT
Vini Gupta and Swades De

Abstract—This paper addresses the problem of decentral- attention to energy efficiency. A greedy selection approach in
ized sensor selection in an energy-constrained wireless sen- [7] aimed at improved energy efficiency. The aspect of sensing
sor network-based Internet-of-Things, for monitoring a spatio- quality has gained recent attention. For instance, in [8] a sensor
temporally varying process. To do so, an adaptive edge computing
framework and its variants are proposed which distributedly op- selection scheme was presented for a linear measurement
timize a critical trade-off between sensing quality and remaining model using a well known D-optimal performance criterion.
energy of the sensor nodes (SNs). Unlike the existing distributed In [9] the Cramér-Rao lower bound performance measure
sensing approaches, the proposed one aims to maintain energy was employed as sensor selection criterion for a non-linear
balance among the SNs. The original sensor selection problem is measurement model. For monitoring a temporally-correlated
decoupled into multiple sub-problems, each solvable at an edge
node elected as head of a coverage region containing a set of SNs. source, in [10] a sensor selection scheme was developed that
The sub-problem in each coverage region is adapted to variations optimizes the sensing quality subject to a prescribed power
of the underlying process. In each region, the process is estimated budget for the energy harvesting (EH) system. Subsequently,
using PCA-SBL (principal component analysis-sparse Bayesian in [11] a sensor selection scheme was proposed that minimizes
learning) on noisy signal measured by the respective active SNs. the estimation error at the FC subject to energy and spectral
Further, to correctly adapt to the process and estimate the signal,
a novel logic is designed that indicates requirement of network budgets of an EH IoT network. The work in [12] optimizes
retraining in the next measurement cycle. The results from both sensing quality and energy efficiency while selecting a
extensive simulation studies illustrate improved energy efficiency fixed number of SNs. To address the shortcoming of sensing
and network energy balance of the proposed framework over a dynamic process by activating a fixed number of SNs,
the existing closest competitive centralized and decentralized centralized sensing schemes in [13]–[16] adapted the number
approaches. The proposed framework is tested on synthetic as
well as real data-sets of a sensor network. of active sensors.
The centralized sensor selection approaches involve high
Index Terms—Wireless sensor networks, adaptive and decen- communication and energy overheads and processing com-
tralized sensor selection , sparse Bayesian learning, principal
component analysis, network residual energy plexity. To overcome these limitations in large-scale WSNs,
decentralized strategies have been developed in recent studies.
I. I NTRODUCTION An offline iterative distributed sensor selection scheme was
Rapid growth of smart wireless sensors deployment for proposed in [17], [18], employing dual subgradient method.
Internet-of-Things (IoT) applications is expected to exponen- Likewise, the study in [19] suggested an optimization problem
tially increase the energy consumption and delay [1]. In this for decentralized sensor selection and solved using a fast
context, intelligent sensing and transmission at the field sensor empirically converging iterative alternating direction method
nodes (SNs) can prolong their life span [2], [3]. In a densely- of multipliers. Distributed implementation of the approach in
deployed wireless sensor network (WSN) for monitoring a [8] was formulated in [20], in a network with two leader
spatio-temporal process, inherent redundancy in the signal can nodes. Further, considering a Bayesian framework a low-
be exploited to perform parsimonious sensing without sacrific- complexity greedy decentralized algorithm for heterogeneous
ing the sensing quality [4]. Further, decentralized processing sensing environment was developed in [21].
at the edge nodes can overcome the scalability issues. Iterative nature of the existing distributed sensor selection
A. Prior art and motivation approaches [17]–[21] requires local information exchange
The existing works on sensing mainly focus on centralized among the SNs in each iteration, which significantly increases
decisioning wherein the fusion center (FC) receives informa- the energy and processing overheads. These works assume
tion (noisy measurements) from the SNs, takes sensing deci- regression coefficients to be known a priori, which may not
sion, and conveys it back to the SNs. The approaches in [5], be realistic. Moreover, these approaches did not account for
[6] randomly activate a fixed number of SNs without paying the critical aspect of unequal remaining energy of the SNs.
Consequently, a few SNs may be repeatedly selected over
This work has been supported in parts by the Department of Science and the others, which may result in a network coverage outage.
Technology, International Bilateral Cooperation Division, under the grant no. Also, importantly, these approaches are non-adaptive to the
INT/UK/P-153/2017, the Science and Engineering Research Board, under the
grant no. CRG/2019/002293, the Department of Telecommunications, under dynamics of the sensed process; they fix one of the two
the grant no. 4-23/5G test bed/2017-NT for building end to end 5G test-bed, performance measures (sensing quality and the number of
and TCS RSP Fellowship. active SNs) and optimize the other. Intuitively, optimizing
The authors are with the Department of Electrical Engineering and Bharti
School of Telecommunications, Indian Institute of Technology Delhi, New both the measures as per the process dynamics and minimum
Delhi 110016, India (e-mail: vini.gupta@ee.iitd.ac.in; swadesd@ee.iitd.ac.in). required sensing quality may yield a better performance.

1536-1276 (c) 2021 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: California State University Fresno. Downloaded on June 23,2021 at 11:40:23 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/TWC.2021.3062568, IEEE
Transactions on Wireless Communications

B. Novelty and Contribution TABLE I: Symbols and abbreviations


Notation Description
To overcome the above-mentioned lacuna in the literature, R, Rr , Nr = |Rr | Number of coverage regions, set containing SNs of region r,
total number of SNs in region r
this work proposes an adaptive edge computing framework A(k) (k)
r , Mr = |A(k)
r | Set containing active SNs of region r in kth cycle, number
for decentralized sensor selection that jointly optimizes the of region r in kth cycle
of active SNs 
sensing quality of a slowly-varying process and WSN energy z(k)
r ,yer(k) Spatial signal ∈ RNr×1 across SNs of region r in kth cycle,
 
(k)
efficiency. To the best of our knowledge this framework is signal measured by active SNs in set A(k)
r ∈ RMr ×1
first to provide energy balance among the SNs by accounting 
(k)

A(k) (k)
r , Br Sensing matrix ∈ RMr ×Nr , sparsification matrix
for their unequal remaining energy in the sensor selection  
problem. The existing works considered distributed sensor ∈ RNr ×Nr of region r in kth cycle
Ec(k) (n), Es , Erem
(k)
(n) Energy consumed by nth SN in transmission and sensing (in
selection and remaining energy in isolation, without factoring
J), remaining energy of nth SN in kth cycle (in J)
that the SNs can be selected for sensing only if they have (k)
Eproc (·) , Etx (dt−r ) Energy consumed by SN in processing tasks and transmission
sufficient remaining energy. In contrast to the prior art, the of b bits of information at a distance dt−r (in J)
Erx , Eelec , η Energy spent by SN to receive b information bits (J), consumed
proposed framework accounts for energy consumption at each
by radio electronics (J/bit), amplifier energy (J/bit/m2 )
step and does not require iterative local information exchange [α, β] Lower and upper limits of BCRB
in each cycle. Another important aspect is that the proposed Mr(k+1|k) Number of active SNs of region r predicted for (k + 1)th
cycle based on signal estimated during current cycle k
framework learns the regression matrix online using previously (k)
−i
Ak|M r Pruned active set obtained by removing i SNs with the lowest
estimated signals to better adapt to spatio-temporally evolving r
signal amplitude |zr(k) (·)| from A(k)
r
process, unlike a random Gaussian matrix considered in the (k)
−i
bk|M
xr
r ∈ RNr ×1 Sparse signal estimate obtained using measurements from SNs
literature. To keep a check on accumulated estimation error, a (k)
−i
∈ Ak|M
r
r
retraining logic is also developed. In an edge computing frame- (k) (k)
BCRBk|M r −i , b
(k) BCRB corresponding to SNs ∈ Ak|M r −i
, euclidean dis-
r r r
work, entire processing is done by the edge nodes in contrast (k)
tance between b(k)
x and bk|M
x r −i
to the cloud computing where the central entity performs all b(k)
δ
r r
Heuristic giving an idea of process variability with respect to
r
such computations. In network’s context, the centralized sensor (k − 1)th cycle
selection is analogous to the cloud computing. Similarly, t1 , t2 Time required to compute the matrix B(k) r using the PCA
scheme and run update step of the adaptive mechanism
for decentralized sensor selection, this paper considers edge t3 , t4 Time required in SBL-based recovery of x(k) and prediction
r
computing framework that is realized by dividing the WSN step of adaptive mechanism and to run retraining logic
field in coverage regions/clusters with an elected cluster head BCRBr , SSr , CHr , FC Bayesian Cramér-Rao bound, sensor selection function, cluster
head (or edge node) for region r, fusion center
(CH) in each. The CHs perform the role of edge nodes by Tr{A}, kak, (·)T , diag (·) Trace of matrix A, l2 -norm of vector a, transpose of the vector
taking sensing decisions and sharing them with the cloud (FC) /matrix, diagonalization operation on a vector
and the field SNs. Thus, the novel ideas are decoupling the work. It is also integrated with the closest existing
sensing quality and energy efficiency-based sensor selection approaches [15], [18], and [21] for fair performance
problem into various sub-problems, integration of energy con- comparison.
sumption, non-iterative decentralized sensing decisions, and 6) Relative performance of the proposed framework with
data-driven retraining of the network/coverage regions(s). The the closest approaches in [18], [21] demonstrates the
contributions of this work are as follows. achieved gain in energy efficiency (up to ≈ 84%) and
1) A decentralized sensor selection framework and its four network energy balance. Performance of the proposed
variants are proposed to jointly optimize the sensing framework is verified on real WSN data as well.
quality and energy efficiency.
2) A closed-loop adaptive mechanism is developed which C. Organization
runs on the edge nodes to predict the number of active The WSN data acquisition model, sparse signal represen-
SNs in the next measurement cycle in the respective tation, and recovery are presented in Section II. Section III
regions. describes the proposed decentralized sensing problem. Energy
3) The sensor selection and estimation processes rely on cost and coverage regions are discussed in Section IV. Sections
sparsification matrices constructed using previously es- V, VI, and VII respectively presents the proposed decentral-
timated signals. To limit the accumulation of estimation ized sensing framework and its variants, simulation results,
errors over the measurement cycles, a retraining logic is and concluding remarks. The main symbols and abbreviations
designed to indicate the next retraining instants. used are listed in Table I.
4) For monitoring a spatio-temporally dynamic process, a II. S YSTEM M ODEL , S PARSIFICATION , AND R ECOVERY
sparsification matrix is updated in each measurement
A. WSN system model and its distributed counterpart
cycle using principal component analysis (PCA) which
utilizes w past instances of signal samples estimated via A homogeneous WSN is considered where energy-
sparse Bayesian learning (SBL). The window size w constrained SNs are deployed for monitoring a slowly-varying
is chosen by addressing a trade-off between retraining spatio-temporal process (e.g., air pollution shown in Fig.
overhead and cluster head-to-cluster head (CH-to-CH) 1). Due to dense deployment, there exists spatial correlation
communication overhead. among signals across the SNs. This can be used for energy-
5) Energy consumption in sensing, processing, and trans- efficient sensing by activating a subset of SNs while the
mission is inherently integrated into the proposed frame- remaining ones can sleep.

1536-1276 (c) 2021 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: California State University Fresno. Downloaded on June 23,2021 at 11:40:23 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/TWC.2021.3062568, IEEE
Transactions on Wireless Communications

for next cycle), and broadcasts A(k+1) to the SNs. In a


decentralized setting, adaptive sensor selection (proposed in
Sections III-B, III-C) and signal estimation are carried out at
the node level. The WSN field is divided into R coverage
regions/clusters, each containing Nr SNs given by the set
Rr = {1, 2, . . . , Nr } , ∀ 1 ≤ r ≤ R, such that Nr = |Rr |.
PR that Rr , ∀r, are the sets of non-overlapping SNs and
Note
r=1 Nr = N . One SN from each region is elected as the
cluster head (CHr ), called edge node, which receives the
(k) (k)
measurement signal vector y er ∈ RMr ×1 from the set of
(k)
SNs Ar that are activated from the respective region r. The
(k) (k)
Fig. 1: Decentralized WSN system. total number of active SNs in region r is Mr = Ar ≤ Nr ,
SR (k) (k)
Let the signal corresponding to the random process observed and r=1 Ar = A(k) . The CHr processes y er to estimate
across N SNs in the k th measurement cycle be denoted by the spatial signal across all the SNs of that region denoted
T (k)
by zr ∈ RNr ×1 , which is followed by computation of the

the spatial signal vector z(k) = z (k) (1) , · · · , z (k) (N ) ∈
(k+1)
RN ×1 . Let A(k) ⊆ {1, · · · , N } represent the set of SNs region’s active set Ar for the next measurement cycle.
activated for sensing in the k th cycle (called active set); the Subsequently, each CHr sends the estimate of the signal vec-
(k) (k+1)
total number of active SNs is given by M (k) = A(k) ≤ N . tor zr to the FC for archival purpose and broadcasts Ar
The active SNs are selected using a sensor selection problem in their respective regions. To formulate the distributed sensor
(Section III-A) solved by the FC in a centralized setting. Let selection problem, the system model in (2) is represented
the activity status of the SNs be captured in a binary sensing as in (3), by placing the rows of the sensing matrix A(k)
(k)
matrix A(k) ∈ RM ×N . Each row of A(k) corresponds to a corresponding to the active SNs of each region Rr one after
(k) (k)
distinct active SN. Hence, there will be M (k) unique rows in the other and denoting the submatrix as Ǎr ∈ RMr ×N .
the matrix A(k) . If the mth row represents the ith active SN, 
(k)
 
(k)

(k)
 
(k)

then its entries are given by, ye1 Ǎ1 z1 n1
 .   ..  .   . 

1, n = i s.t. i ∈ A(k)  .  =   .  +  . , (3)
(k)
A (m, n) = (1)  .   .  .   . 
0, n ∈ {1, · · · , N } \ {i} . y
eR
(k)
ǍR
(k)
zR
(k)
nR
(k)

Likewise, if another row (say l) represents a different active PR (k) (k) (k)
SN (say j), then A(k) (l, j) = 1 and A(k) (l, n) = 0, (n 6= with M (k) = r=1 Mr and nr ∈ RMr ×1 , ∀r, denoting
(k) (k)
j). The measurement vector y e (k) ∈ RM ×1 contains signals the noise across the SNs belonging to Ar . The measurement
(k)
measured by the SNs ∈ A(k) . Similar to [15], [21], the system th
vector of the r region y er has no contribution from the
model in the centralized setting is given by (k)
signal across the SNs of the other regions, i.e., zj , ∀ 1 ≤ j ≤
(k)
e (k) = A(k) z(k) + n(k) ,
y (2) R, j 6= r. Thus, the sensing submatrix Ǎr can be truncated
where n (k)
∈ R M (k) ×1
is the additive white Gaussian by considering the columns corresponding to the SNs ∈ Rr
(k) (k)
noise vector with independent and identically distributed (IID) only. Denoting the truncated matrix as Ar ∈ RMr ×Nr , the
components of zero mean and variance σ 2 . It comprises system model for Rr , ∀r, is given by
(k)
both measurement and  channel noises, distributed as nm ∼ er(k) = A(k)
y (k) (k)
r zr + nr . (4)
2 (k) 2
N 0M (k) , σm IM (k) and nch ∼ N 0M (k) , σch IM (k) , re- Also, the correspondence between and
(k)
Ar
is given by,
(k)
Ar
spectively. These noises are independent of each other, as 
(k)
one is due to sensing/measurement impairment while the A(k) 1, n = i s.t. i ∈ Ar
r (m, n) = (5)
other is due to signal reception over the wireless channel. 0, n ∈ {1, · · · , Nr } \ {i} .
(k) (k)
Thus, the overall noise n(k) = nch + nm with variance  The system model in (4) is used in the decentralized sensor
2 2 2 (k)
σ = σch + σm , i.e., n ∼ N 0M (k) , σ 2 IM (k) . Note selection problem proposed in Section III-B. A detailed discus-
that the spatial correlation among the process signal across sion on finding the number of regions R and their edge nodes
different SNs due to their dense deployment is captured by is provided in Section IV-B. These edge nodes are re-elected,
the spatial signal vector z(k) . It is not related to the noise. as mentioned in the proposed variants of the decentralized
In general, the scenario where ambient noise corrupts the framework (Section V). The considered homogeneous network
SNs’ measurements may result in colored measurement noise, setting enables any SN to serve as a candidate edge node for
(k) (k)
distributed as nm ∼ N (0M (k) , Q), where Q ∈ RM ×M is the region in which it is located.
the colored covariance matrix.  Then, n(k) ∼ N (0M (k) , Σn ) ,
2 M (k) ×M (k) B. PCA-SBL based signal sparsification and recovery
where Σn = σch IM (k) + Q ∈ R . For simplicity,
the case with white noise is considered throughout the paper. The unknown signals from all SNs (vectors z(k) in (2)
(k)
In a centralized setting [15], the FC receives components and zr , ∀r, in (4)) are required to be reconstructed using
of the vector ye (k) from the currently active SNs, processes it the measurements from a few active SNs (vectors y e (k) in (2)
(k)
to estimate the unknown spatial vector z(k) (i.e., the signal and yer , ∀r, in (4)) in ill-posed estimation scenarios. This
across the entire WSN field), computes A(k+1) (active set is possible using sparse signal recovery frameworks, such as

1536-1276 (c) 2021 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: California State University Fresno. Downloaded on June 23,2021 at 11:40:23 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/TWC.2021.3062568, IEEE
Transactions on Wireless Communications

SBL and compressed sensing [4], that enable estimation of parameterized Gaussian distribution to the unknown sparse
sparse signals using the under-sampled measurement vectors vector and uses evidence-maximization criterion to estimate
within certain accuracy level. However, this requires a priori those hyperparameters. This work uses SBL to obtain estimate
(k) (k)
information of the sparsification matrices that can make the of the unknown vector xr denoted by x br ∈ RNr ×1 from
(k)
observed signals (z(k) , zr(k) ) sparse. which the spatial signal vector estimate b zr is obtained.
(k) (k) (k)
Since there exists spatial correlation among the signals It employs a dictionary matrix Θr = Ar Br and an
(k)
across the SNs, the vectors z(k) and zr , ∀r, can be repre- (k)
equivalent measurement vector yr during each k th cycle
sented in sparse form using PCA-based sparsification matrices in the rth region, ∀1 ≤ r ≤ R, considering the regional
similar to [13], [15]. PCA-based sparsification is known to system model (7). SBL-based signal recovery is explained in
work well with real signals compared to other existing trans- Appendix B. In presence of colored noise, CoNo-SBL scheme
formation matrices [22]. Thereafter, using the SBL framework, [24] is used for signal estimation.
the sparse vectors can be estimated respectively from the The sparsification and recovery techniques presented here
(k)
under-sampled measurement vectors y(k) and yr , ∀r, of the are used in the formulation of the decentralized sensing
sparse system model, that is discussed next. problem in the next section.
PCA-based sparsification: For the system model (2), let
B(k) ∈ RN ×N be an orthogonal sparsification matrix obtained III. P ROPOSED D ECENTRALIZED S ENSING P ROBLEM AND
using PCA for representing the vector z(k) as a sparse vector A DAPTIVE S OLUTION
x(k) ∈ RN ×1 . Then, the system model (2) becomes A. Sensor selection problem formulation
(k) (k) (k) (k)
y =Θ x +n , (6) A practical sensing scheme needs to optimize a critical
(k)
where Θ = A B ∈ R (k) (k) M (k) ×N
denotes the measurement trade-off between sensing quality and energy efficiency. Sens-
(k)
matrix and y(k) ∈ RM ×1 is the equivalent measurement ing quality is measured by the mean-squared error (MSE)
(k)
vector. The details on PCA-based sparse representation are in then estimation of
(k) (k)
2 othe unknown vector x : MSE =
presented in Appendix A. Similarly, sparse representation Ex(k) x − x b . However, the MSE in the underlying
of the regional system model (4) is given by (7), with ill-posed scenario (6) may not have a tractable closed-form
(k) (k) Mr(k) ×Nr (k) expression [12]. Also, its dependence on the actual estimator
Θ(k)
r =Ar Br ∈R and Br ∈RNr ×Nr being regional
measurement and sparsification matrices, respectively. b(k) and x(k) makes it practically unsuitable since the sensor
x
selection task is followed by the estimation of x(k) . Therefore,
yr(k) = Θ(k) (k) (k)
r xr + nr , ∀r, (7) similar to [15], [21], the current work exploits a theoretical
In the cycle immediately after training/retraining, the matrix lower bound, called Bayesian Cramér-Rao bound (BCRB), to
B(k) is computed using Ktr instances of the vector z(·) . In measure sensing quality. It enables sensor selection before data
other cycles, this matrix cannot be computed as presented acquisition and does not mandate availability of either x(k) or
in Appendix A due to the unknown vector z(k) . Exploiting y(k) (measurement vector). The BCRB in the k th cycle is
the fact that the considered densely-deployed WSN monitors expressed in (8); for the derivation and convexity proof, see
a slowly-varying process, the matrix B(k) is anticipated to Appendix C and [15] respectively.
change slowly. This fact along with the availability of previous (
−1−1
)
signal estimates allow online computation of B(k) using w (k) 1  (k)T  
(k) (k)

(k)
BCRB = Tr B diag a B +Γ ,
σ2
e
most recent instances of the estimated signal vector in the
z(k−w) , · · · , b
learning set, i.e., T = {b z(k−2) , b
z(k−1) }, such that (8)
where a e(k) ∈ RN ×1 contains binary diagonal elements
|T | = w. This adapts the sparsification matrix to the varying e (k) , i.e., diag a

of the diagonal matrix A e(k) = A e (k) =
process better and avoids ensuing loss in sensing quality due T
A(k) A(k) ∈ RN ×N . These diagonal elements have a

(k)
to the usage of a stale estimate of B(k) and  hence Θ in
recovering the unknown vector x (k)
or z(k)
. A similar argu- one-to-one correspondence to the SNs; e a(k) (n) = 1 implies
th (k) e (k) (n, n) =
(k)
ment goes for regional sparsification matrices Br , ∀r, where  active state of the n SN, i.e., e
an a (n) = A
(k−w) (k−1) 1, n ∈ A(k) (k)
the region-wise sets Tr ={b zr ,··· ,b zr }, ∀r, are used. 0, n∈/ A(k) . The parameter Γ ∈ RN ×N represents a
The parameter w, referred to as the window size, is chosen diagonal matrix with variance of the components of sparse
by suitably addressing a trade-off observed in variations of vector x(k) as its diagonal elements. These elements are set
retraining overheads and CH-to-CH transmission overheads, as the eigen values of the matrix Σ
b T , computed using PCA
as detailed in Section VI-B. A larger value of w compared to in Appendix A, as done in [15]. For the energy efficiency of
Ktr is considered in simulations (Section VI) because error the WSN, consider a weighted sensor selection (SS) function:
due to the use of previously estimated signals is absent in the  
latter case which uses the exact signal. N (k)
X Ec (n) + Es
SBL-based recovery: A basic requirement of the adaptive SS(k) = (k)
a(k) (n) ,
e (9)
sensing framework is a reliable estimate of the sensed signal n=1 Erem (n)
(k) (k)
to update the sparsification matrix and accurately provide where Ec (n) , Es , and Erem (n) respectively denote the
feedback on the process dynamics for sensing and recovery energy consumed by the nth SN in signal transmission,
purposes. SBL is known to give improved sparse signal sensing and its remaining energy in the k th cycle detailed
estimate in ill-posed estimation scenarios [23]. It assigns a in Section IV-A.

1536-1276 (c) 2021 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: California State University Fresno. Downloaded on June 23,2021 at 11:40:23 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/TWC.2021.3062568, IEEE
Transactions on Wireless Communications

n o  −1  5
(k) T 2 T T 2 T T
BCRB = Tr BΓB − Tr σ IM + ABΓB A ABΓ B A (11)
n o   −1 
T 2 T T 2 T T
= Tr BΓB − Tr AB σ IN + ΓB A AB Γ B A (12)
  −1 
= Tr {Γ} − Tr BT ABe σ 2 IN + ΓBT ABe Γ2 (13)
R
!
X X Γr (n, n)2
= Tr {Γr } −
r=1 n∈Rr
σ 2 + Γr (n, n)
R  n 
X −1 2 o
= Tr {Γr } − Tr σ 2 INr + Γr Γr (14)
r=1
R 
X   −1  
= Tr {Γr } − Tr BTr A
e r Br σ 2 INr + Γr BTr A
e r Br Γ2r . (15)
r=1

−1 −1
The sensor selection problem needs to simultaneously min- (I + PQ) P = P (I + QP) [26], the BCRB simpli-
imize the BCRB (expression (8)) and SS function (expression fies to (12). Using the notation AT A = A e (mentioned
(9)). Minimizing the BCRB minimizes the MSE [9], [12]. along with (8)), (12) is rewritten as (13). If all the SNs
Minimizing the SS function gives more weight  to the activa- arenactive, i.e. A e = IN , the second term in (13) becomes
(k)
 −1 2 o
tion of the SNs with a low energy consumption Ec + Es Tr B B σ IN + ΓBT B
T 2
Γ . Further, using orthogo-
  n −1 2 o
(k)
and high remaining energy Erem . This ensures efficient nality property of B, it reduces to Tr σ 2 IN + Γ Γ =
energy utilization and aids in energy load balancing among PN Γ(n,n)2 th
n=1 Γ(n,n)+σ 2 . Thus, the contribution of the n SN to the
SNs. Basic constraints that are required to be handled by 
Γ(n,n)2

a practical sensing problem of a WSN BCRB is − Γ(n,n)+σ2 . Segregating the SNs’ contributions
 are- (1) active/sleep
state of the SNs e a(k) (n) = 1/0, ∀n , (2) SNs with zero region-wise, (13) can be rewritten as in (14) with the diagonal
elements of Γ(k) ∈ RNr ×Nr equal to Γ(k) (n, n) , n ∈ Rr .

(k) r
remaining energy Erem (n) = 0 should not participate in
Further, employing the orthogonality property of Br and
sensing, and (3) coverage constraint (activates one SN from
the assumption of all active SNs (A e r = IN ), the BCRB
r
each designated regions) to enable better monitoring, ensure
expression (14) can be generalized as in (15). Following the
network energy balance, and prevent the creation of coverage
intermediate steps (11) to (12) in reverse order and applying
holes. The binary constraint (1) makes the sensing problem
the dropped subscript (k) of the measurement cycle in (15),
non-convex; solving it using exhaustive search is infeasible (k)
one obtains BCRB(k) as the sum of regional BCRBr func-
for large number of nodes (N ). Hence, similarly as in [8],
tions:
[25], a convex sensing problem is realized by considering 
relaxed convex box constraint e a(k) (n) ∈ [0, 1] , ∀n . Thus,
the relaxed convex multi-objective sensor selection problem R
(
−1−1
)
[15] is given by (10). Its decentralized formulation is proposed (k)
X 1  (k)T  
(k) (k)

(k)
BCRB = Tr B diag a
er Br + Γr ,
in the next section. r=1
σ2 r
h i | {z }
minimize BCRB(k) , SS(k) (10a) BCRBr
(k)
a(k) (n) ∀n
contains diagonal elements of (16)
(k) Nr ×1
where a ∈ R the
e
er
s.t. a(k)(n) ∈ [0, 1] , n = 1, · · · , N,
e (10b) e (k)
diagonal matrix Ar which have a one-to-one correspondence
n o  
(k) (k) (k) e (k)
a (n)=0, n∈ i|Erem (i) = 0, 1 ≤ i ≤ N , (10c)
e with SNs ∈ Rr , i.e., diag aer =A r . Note that, (16) holds
X true when all the SNs are active; while for the other cases it is
a(k) (j) ≥ 1, r = 1, · · · , R. (10d) PR (k)
approximated as BCRB(k) < r=1 BCRBr . Validity of this
e
j∈Rr
approximation is verified via simulations, in Section VI-D.
Further, the SS function can be rewritten as a sum of regional
B. Decentralized sensor selection: Edge computing paradigm (k)
SSr functions,
In decentralized sensing, (10) is decoupled into R sub-
problems, solvable at the edge nodes of the clusters. The  
R X (k)
BCRB (8) is decoupled as follows. Using the orthogonality X Ec (n) + Es
SS(k) = a(k)
T
property of the PCA-based matrix B(k) , i.e. B(k) B(k) = (k)
er (n) . (17)
T r=1 n∈Rr Erem (n)
B(k) B(k) = IN , cyclic property of trace operator, | {z }
(k)
i.e. Tr{PQR} = Tr{RPQ} [26], and dropping the sub- SSr

n (k) for brevity, expression


script
−1 o
(8) becomes BCRB =
−1 T T 1
Tr BΓ B + A σ2 IM A . Further, employing the
matrix inversion lemma [26] and property Tr {C + D} = Employing (16), (17) in (10) and interchanging minimiza-
Tr {C} + Tr {D} [26], the BCRB is given in (11). Denoting tion and summation, the relaxed sensor selection problem (10)
P = AB and Q = ΓBT AT in (11) and using identity is decoupled into R sub-problems as:

1536-1276 (c) 2021 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: California State University Fresno. Downloaded on June 23,2021 at 11:40:23 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/TWC.2021.3062568, IEEE
Transactions on Wireless Communications

h i
minimize BCRB(k) (k)
r , SSr the sensing application. The energy information is sent by the
(k)
ar (n), n∈Rr
e SNs to the respective CHr s along with the measured signal
(k)
s.t. ar (n) ∈ [0, 1] , n
e n ∈ Rr , (18)
o
(k)
yr (·).
(k) (k) (k)
ar (n) = 0, n∈ i|Erem(i) = 0, i∈Rr ,
e Prediction step: In k th cycle, a heuristic δbr (defined
P (k)
ar (n) ≥
. 1 in (22)) is computed to estimate the process variability with
n∈Rr e th (k)
| {z } respect to the (k − 1) cycle. If the value of δbr is greater
R sub-problems th
than a suitably chosen threshold  δr , then  the predicted value
(k) (k) (k+1|k) (k)
The number of active SNs Mr and the active set Ar in is increased as Mr = Mr + 1 .

Rr are obtained from the solution of (18) as: (k)
zr − b
b zr
(k−1)

Nr
! δbr(k) = √ . (22)
X Nr
Mr(k) = round ar(k)
e (n) , (19)
(k) (k+1|k)
n=1 When δbr ≤ δrth , the value of Mr is decreased
A(k) = Indices of Mr(k) largest elements of ordered set: (k|Mr(k) −i)
r
n o as follows. Consider the pruned active set Ar by
a(k) ar(k)(n) ≥ e
r (n) |e ar(k)(m) , ∀n, m ∈ {1, . . . , Nr } (20)
. removing i SNs (starting i = 1) with the lowest signal
(k|M (k) −i)
e
(k) (k)
(k)
amplitude |bzr (·)| from the set Ar . Employ Ar r -
ar (n)
The solution can also be obtained by first rounding e (k|Mr(k) −i)
(k) (k) based signal estimate x br to compute another heuris-
to the nearest integer and then computing Mr and Ar , or (k) (k)
tic b r defined in (23). If b r ≤ th r (an appropriate
using randomized rounding [9]. Note that the aim of current (k|Mr(k) −i) (k+1|k)
work is not to compare the decentralized and centralized threshold) and BCRBr
 ∈ [α, β], set Mr =
(k)
sensing problems, but to propose and solve a decentralized Mr − i ; repeat the process with unit increment of i until
sensing problem with components similar to the centralized (k|M (k) −i)
BCRBr r  ∈ / [α, β] or non-selection of any SN from
one (Section III-A). (k)
the region i.e. Mr − i < 1. The reason is to explore the
C. Proposed adaptive solution for decentralized sensing possibility of maximum achievable decrease in Mr for the
(k)

The R multi-objective optimization problems for regional th (k|M (k) −i) (k)
br r
(k + 1) cycle such that the estimates x and x
br
sensor selection (18) are solved at the respective edge nodes (k|Mr(k) −i)
by using the scalarization method [27], with the difference (current one) are similar and BCRBr ∈ [α, β].
 
(k)

that, here a mechanism is proposed to adapt the scalar weights k|Mr −i
(k)
(k)
(k) r = x
r − x r
. (23)
λr , ∀r, in the k th cycle. The scalarized sub-problem in (21)
b b b
(k+1)
is solved using the CVX solver [28] due to its convex nature Update step: The scalar λr and active set for the
(k+1) (k+1|k)
(similar to the sensor selection problem in [15]). next cycle Ar with Mr active nodes, are computed
    by considering energy resources in the network, the updated
minimize 1 − λ(k) r BCRB(k)
r + λr
(k)
SS(k)
r (21a) Br
(k+1)
, and employing modified binary search (MBS) (Alg.
(k)
ar (n), n∈Rr
e (k+1)
1) algorithm [15]. The matrix Br is oupdated
n using the
a(k)
s.t. er (n) ∈ [0, 1] ,n n ∈ Rr , (21b) n
(k−w+1) (k)
o
o updated learning set Tr = Tr \ b zr ∪ bzr as dis-
a(k) (k)
r (n) = 0, n∈ i|Erem(i) = 0, i∈Rr , (21c) cussed in Section II-B. The MBS  minimizes the Usub-problem
e
X (21) on a search space λL , λU
r ∈ [0, 1] with λr − λr < ∆
L
a(k)
er (n) ≥ 1. (21d) P (k+1)
r
(k+1|k)
n∈Rr or n∈Rr e ar (n) = Mr as the stopping criteria and
∆ being a small positive number. The active set is extracted
The motivation behind the adaptation method is two-fold. (k+1) (k+1)
from the output a er . If the corresponding BCRBr
First, it is intuitive that a higher number of active SNs (k+1)
belongs to [α, β], then the SNs ∈ Ar are used for data
is required to sense a process with a faster dynamics, so (k+1|k)
as to obtain the same sensing accuracy [29]. Second, it is acquisition.
 n Otherwise M ro is increased
n or decreased
o
(k+1|k) (k+1|k)
observed that associating a fixed scalar weight λr gives more i.e. min Mr + 1, Nr or max Mr − 1, 1 for
(·) (k+1)
importance to the decreasing energy based function SSr over the cases with BCRBr > β or < α. The update step
(·) (k+1)
the BCRBr as the measurement cycle progresses. This may reiterates until BCRBr ∈ [α, β] is achieved.
(·)
keep increasing the value of BCRBr and the corresponding The adaptation mechanism stops when the energy of all SNs
(k+1)
MSE.
P
of a region is exhausted, i.e., n∈Rr Erem (n) = 0 or when
A feedback-based adaptation mechanism (run by each it is not possible to satisfy the BCRB requirement with all alive
CHr ) is proposed that predicts the number of active SNs SNs i.e. λU L
r − λr < ∆. At this juncture, it is worth noting
(k+1|k) th
Mr for the (k + 1) cycle based on signal variations that despite the unavailability of the true MSE, the adaptation
(k) (k) (k+1)
estimated using current estimates of x br and bzr . This value mechanism ensures from the beginning that BCRBr ∈
is further updated considering the SNs’ updated energies [α, β] , ∀k. Thus, the proposed approach is practically useful.
(k+1) (k+1)
( i.e., Erem (n) , Ec (n) + Es , ∀n ∈ Rr ) while keeping the A stepwise description of the proposed adaptive mechanism
estimation error within a tolerable range. For this, it is ensured that runs at each edge node is given under block 2 head of
(k)
that the BCRBr ∈ [α, β] during each cycle, with α and β Algorithm 2. The energy update equations and the other blocks
respectively being lower and upper BCRB limits required by included in Algorithm 2 are explained in Section V.

1536-1276 (c) 2021 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: California State University Fresno. Downloaded on June 23,2021 at 11:40:23 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/TWC.2021.3062568, IEEE
Transactions on Wireless Communications

Algorithm 1 Modified Binary Search (MBS) Radio energy cost: To account for the radio energy con-
(k+1) (k+1) (k+1) (k+1|k)
Input: Es , Ec (n) , Erem (n) , ∀n∈Rr , Br , Mr ; sumption, the model in [32] is used. Energy spent by a SN to
Initialization: λL U
 
r , λr ← [0, 1]. transmit and receive a b bits long packet at a distance dt−r ,
while λU L
r − λr ≥∆ do  denoted respectively by Etx (dt−r ) and Erx , are given by,
(k+1) λL r +λr
U
(k+1)
Set λr = and obtain a
er by solving (21) using ν
2 Etx (dt−r ) = bEelec + bη (dt−r ) and Erx = bEelec , (25)
(k+1) (k+1) (k+1) (k+1)
λr , Br , Es , E c (n) , Erem (n) , ∀ n ∈ Rr . where Eelec (J/bit) is the energy consumed by the radio
(k+1) (k+1|k) (k+1)
then λL electronics, η (J/bit/mν ) characterizes the amplifier energy, and
P
if n∈Rre ar (n) > Mr r ← λr .
(k+1) (k+1|k) U (k+1)
ν is the path loss exponent considered as 2. It is assumed that
P
else if n∈Rre ar (n) < Mr thenλr ← λr .
else break. each SN knows its distance to the other SNs in its cluster. In
end if
end while
the proposed framework, communication takes place between
(k+1) (k+1) the SNs and their respective CHs, CHs and FC, and among
Output: a
er , λr .
Note that, the temporal correlation (or variations) of the CHs; thus, dt−r and Etx vary accordingly. The communication
(k) among CHs happens when a new CH is elected because the
process is used to adapt the number of active SNs (Mr )
existing CH’s remaining energy is on the verge of exhaustion.
in a region in each cycle, while the spatial correlation is
(k) The old CH transfers control information to enable the new
used in the MOP (18) to select these Mr SNs out of total
CH to execute the decentralized sensor selection task. This
Nr SNs. This MoP ensures that the same SNs may not be
is explained in block 1 (control operations) of the proposed
selected in each cycle. Thus, this mechanism implicitly adjusts
framework in Section V.
the inter-sample time of each SN (i.e., gap between cycles
in which a SN is activated for sensing). The spatio-temporal B. Number of coverage regions and their formation
modeling of synthetically generated process signal is discussed
in simulation section VI. In the proposed framework, the number of coverage regions
R is determined during the initial stage (before the start
IV. E NERGY C OST AND C OVERAGE R EGIONS of any sensing operation). It is calculated by considering
A. Operational energy costs in WSN the energy consumed in communication per measurement
The energy stored in a SN is used for transmission, recep- cycle, as noted in [32]. However, the energy consumed in
tion, sensing, sleeping, and processing. Unlike the existing lit- processing and CH-to-CH communication are not considered
erature on distributed sensor selection, this work incorporates because they are unknown at the initial stage. The frequency
the energy information to improve energy efficiency. of occurrence of CH-to-CH communication depends on the
Sensing cost: The energy consumed per sensing operation remaining energies of the SNs and CHs left (based on the
depends on the process to be sensed and the type of sensors outcome of the decentralized sensor selection problem) and
that are used. The measurement time varies with the sensors, to-be-chosen new CHs. Eproc is tracked during the execution
so does their sensing energy Es [30]. It can be obtained from of the decentralized framework (block 2, 3 of Section V).
the data-sheet of the used sensor. Consider a SN deployment density in the ith region of radius
Processing cost: In the proposed framework, an edge node ri as ρi with probability distribution fρ , the average number
R∞
(i.e., CH in that cycle) performs computation for taking of SNs in the region is then given by N i = 0 ρi ri2 fρ (x) dx.
sensing decision of the SNs in the cluster. The processing For simplicity, we consider the same radius for each region,
energy consumption of the CH in the k th cycle is computed i.e. ri = r, ∀i. Accordingly, N i = N
(k)
 
(k) (k) R (constant), where R
as Eproc tf = Vf If tf [31], where Vf is the supply is the number of equal sized coverage regions. Thus, with N R
voltage of the processor at operating frequency f , If is its nodes per region on average, energy dissipated per cycle of a
(k) non-CH node (Enon−CH) and a CH node (ECH) are given by,
average current, and tf is the computation time required in
the k th cycle for making sensing decision for the next cycle. Enon−CH = bEelec + bη (dSN −CH ) +
2
bEelec ,
(k) (k)
tf is expressed in terms of number of clock cycles  Nclk 
| {z }
Etx : SN to CH comm.
| {z }
Erx : CH to SN comm.
(k)
th (k) Nclk (26)
needed to execute the task in k cycle as = . tf
 (k) 
f ECH = (N/R − 1) bEelec + bEelec + bη (dCH−SN )2 +
  | {z } | {z }
(k) (k) Nclk Erx : SN to CH comm. Etx : CH to SN comm.
Thus, Eproc tf = Vf If f . To approximately obtain
bEelec + bη (dCH−F C )2 . (27)
(k)
Nclk of the sensor board’s processor, the task is run on a | {z }
Etx : CH to FC comm.
(k)
computer’s processor (e.g., Intel i7-6700 CPU). The time tproc
required to run current cycle’s sub-problems on the computer It is assumed that every SN packs information in payload part
(k) t(k) of the b bits long packet, with zero padding if the information
is measured and N is then set to proc where fe is the
clk 1/fe is less. Since the number of active SNs in a cycle is not known
processor’s operating frequency. Thus, the energy consumed at the initial stage, consider all the SNs to be active in a
by the on-board processor is approximated! as: cycle; the total communication energy consumption per cycle
(k)  
(k)

(k)
 t proc 1 is Etotal ≈ R [ECH + (N/R − 1) Enon−CH ] . The number of
Eproc tproc = Vf If · . (24)
1/f ˜ f coverage regions R in a F × F m2 field that minimizes Etotal
Since a non-CH does not carry out any computation, its is obtained by setting its derivative with respect to R to zero
(k) (k)
processing energy is Eproc tproc = 0 J. [32]:

1536-1276 (c) 2021 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: California State University Fresno. Downloaded on June 23,2021 at 11:40:23 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/TWC.2021.3062568, IEEE
Transactions on Wireless Communications

v
(k)
u NF2 For the case Erem (CH r ) > 0, the current CHr continues
R≈t . (28)
u 
2π (dCH−F C ) −
2 Eelec to remain the edge node. Further, if retraining is not needed,
η
the region’s operations continue without any overheads. Oth-
Out of N SNs, R SNs with the highest remaining energy are erwise, the region is retrained without changing CH and the
chosen as edge nodes (designated as CHr , ∀ r) whenever cov- corresponding energy update is given by (30).
erage regions are formed/reformed. However, in simulations Block 2: This block comprises the proposed adaptive solu-
(Section VI), initially (when sensing starts, i.e. with k = 0) tion for decentralized sensing and SBL-based recovery (Sec-
all the SNs are considered to have the same remaining (initial) tion III-C) that is run in parallel by the CHs in all regions. The
energy. To avoid skewed selection of positions of the CHs SNs ∈ Rr are informed by CHr about their active/sleep status
in the network, κ-medoid [33] algorithm is used. It ensures by running the sub-problem (21). Thereafter, the measured
(k)
a good distribution of the CHs over the WSN field. In the data yr is sent by the active SNs to the respective CHr .
κ-medoid algorithm, spatial distance-based cost function is The remaining energies of CHr and the active/sleeping SNs
used. Since communication energy Etx depends on the spatial are updated as,
(k) (k−1)
(CHr)−Erx Mr(k) −Etx d¯CHr −SN

distance, using κ-medoid energy load is evenly distributed Erem (CHr ) = Erem
among all the SNs. This avoids any overly-utilized SN that (k)
−Eproc (t1 + t2 ) , (31a)
may run out of energy before the others. Each remaining (k) (k−1)
Erem (n) = Erem (n)−Erx −Etx(dn−CHr )−Es
SN chooses to connect with one CHr that requires minimum
communication energy. [for active SNs, n ∈ A(k)
r ], (31b)
(k) (k−1)
Erem (m) = Erem (m) − Erx − Esl
V. P ROPOSED D ECENTRALIZED S ENSING F RAMEWORK
[for sleeping SNs, m ∈ Rr − A(k)
r ], (31c)
This section presents the proposed edge computing frame-
where t1 and t2 are respectively the time required to compute
work for decentralized sensor selection and its variants. The (k)
the matrix Br and to run the update step of the adaptive
variants are proposed to get insights on the network perfor-
solution (Section III-C) for decentralized sensor selection,
mance provided by them which could help in deciding their
d¯CHr −SN is the average distance of the CHr to any SN
suitability for different application scenarios. The computa-
∈ Rr , and Esl denotes the energy needed by a SN to sustain
tional complexity of the framework has also been discussed. In
network operations without collecting data in sleep mode. The
the schematic shown in Fig. 2, block 1 is customized according (k)
vector x br is estimated at CHr using the SBL approach
to the variant, while blocks 2 and 3 are generic for all the
(Section II-B) and then the estimated spatial signal vector
variants, and thus explained only once in Section V-A. (k) (k+1|k)
zr is transmitted to the FC. Further, the value Mr
b and
(k+1)
hence λr are predicted using the mechanism described in
Section III-C. The remaining energy of CHr is then updated
as in (32) with t3 being the processing time required in the
(k)
Fig. 2: Edge computing framework for decentralized sensing. SBL-based recovery of xr and prediction step of the adaptive
mechanism.
A. Variant 1 (k)
Erem (k−1)
(CHr) = Erem (k)
(CHr)−Etx(dCHr −F C)−Eproc (t3) . (32)
In this variant, the coverage regions remain fixed.
Block 1: This block carries out control/training operations Block 3: This block runs the retraining logic at CHr at
in all the regions. When the remaining energy of the cur- the end of every measurement cycle. The remaining energy is
rent edge node of a region is on the verge of depletion updated as in (33) with t4 being processing time to run the
(k)
i.e., Erem (CH r ) ≈ 0, another SN from the same region with logic.
(k) (k−1) (k)
the highest remaining energy is elected as the new CHr . If Erem (CHr ) = Erem (CHr ) − Eproc (t4 ) . (33)
the region does not require retraining, i.e., if retrain flag = 0,
Proposed retraining logic: As the measurement cycles
this new CH election process incurs CH-to-CH communication (k)
progress, the recovery process of xbr can become increasingly
(hence energy) overhead due to transmission of w instances of
(·) erroneous due to the usage of overcomplete dictionary Θ(k) r
recently estimated signal vectors bzr from the previous CH to (k)
(k) which is based on Br . Since a few past instances of the
the new one so that the required matrix Br can be computed (·) (k)
estimated signal bzr are used to estimate the matrix Br , the
at the new CH. The remaining energies are updated as follows:
estimation error accumulates over cycles. This in turn affects
(k) sensor selection and signal recovery accuracy, producing non-
CHr(old) =E(rem
 k−1)  
Erem CHr(old) −wEtx dCHr(old) −CHr(new)
(29a), (k) (k)
sparse estimate xbr of otherwise-unknown sparse signal xr .
(k)
 (k−1) 
Erem CHr(new) =Erem CH r(new) −wErx . (29b) This calls for an error detection logic and retraining in the
region r.
If the region requires retraining as well (retrain flag = 1), the (k)
As the true signal vector xr is unknown, the retraining
energy is consumed only due to retraining (using Ktr instances logic is proposed based on the current reconstructed signal
(·)
of the signal zr ) and not CH-to-CH communication. Thus, vector x
(k)
br itself. The idea is to approximately find the number
the remaining energy is computed as: (k)
(k) (k−1) of non-sparse components N d C r in the reconstructed signal
Erem (n) = Erem (n)−Ktr Es , 1 ≤ k ≤ K, n ∈ Rr . (30) (k)
br and compare it against a suitable threshold value N Crth .
x

1536-1276 (c) 2021 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: California State University Fresno. Downloaded on June 23,2021 at 11:40:23 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/TWC.2021.3062568, IEEE
Transactions on Wireless Communications

Algorithm 2 Adaptive decentralized sensing framework is wrongly considered as sparse. Also, the knowledge of N Crth
can be roughly obtained from instances of the training signal.
n o
(−K +1) (−1) (0)
1: Input: Tr = zr tr , . . . , zr , zr , Rr , α, β, δrth , th
r .
(1) (0) (1) The logic runs in parallel at all the R edge nodes during each
2: Initialization: k = 1, Ec (·), Erem(·), λr , exitf lag = 0.
3: -Compute z̄Tr , Σ b T , & B(0) (k)
using PCA (Section II-B); set Br ≈
cycle.
r r
(k−1)
Br ; calculate t1 .
(1) (1)
4: -Solve (21) and construct Mr , Ar using (19), (20). B. Variant 2 (Block 1)
(1)
5: Obtain Ar using (5); calculate t2 . This variant keeps the coverage regions fixed while changes
6: while exitf lag == 0 do
7: Block 1: Execute operations; update energy (29,30).
the CHr s in every measurement cycle. The SN in a region
8:
(k) (k) (k+1)
Block 2: Broadcast Ar , collect yr , Erem (·) from active SNs. with the highest remaining energy is elected as the new CH.
9: Update remaining energy using (31). If the region does not require retraining (retrain flag = 0),
(k) (k) (k)
10: Start clock to measure time t3 . Set Θr = Ar Br . the remaining energies left after the transmission of overheads
(k)
11: Compute x br using SBL scheme (Section II-B). are given by (29). Else (retrain flag = 1), the region requires
(k) T
 
(k) (k)
12: zr = z̄Tr + Br
Set b x
br & transmit to the FC. retraining; the energy is consumed in retraining overheads, and
P (k+1)
13: if n∈RrErem (n)>0 then (Prediction step) the remaining energy is given by (30).
(k+1|k) (k) (k)
14: Initialize Mr ←Mr ; Calc. δbr using (22).
(k) (k+1|k) (k)
15: if δbr > δrth then Mr ← Mr + 1. C. Variant 3 (Block 1)
16: else
17:
(k)
for i = 1, . . . , Mr − 1 do In this variant, the coverage regions and their CHs are
(k) (k)
18: Construct Ar
(k|Mr −i) (k|Mr
,Ar
−i)
(eq. 5). re-formed (using the procedure outlined in Section IV-B)
(k) (k)
19: Set Θr
(k|Mr −i)
= Ar
(k|Mr −i) (k)
Br .
when the remaining energy of any CH is exhausted, i.e.
(k)
20: Obtain x
br
(k|Mr
(k)
−i)
using SBL (Section II-B).
Erem (CHr ) ≈ 0, 1 ≤ r ≤ R, and the entire network
(k)
(k|Mr −i) (k) (all regions) is retrained (irrespective of retrain flag status).
21: Calc. BCRBr ,b
r using (16, 23).
(k) th
(k)
(k|Mr −i) Otherwise, the regions and their CHs remain fixed and region-
22: r ≤r and BCRBr
if b ∈[α,β] then
(k+1|k) (k) wise retraining is carried out only when respective retrain flag
23: Mr ← Mr − i.
(k)
(k|Mr −i) is set 1. In both cases, the updated remaining energy of SNs
24: else if BCRBr ∈[α,
/ β] then break.
25: end if is given by (30). When the regions are re-formed, their N Crth ,
(k+1|k)
26: end for δrth are changed and Mr are also updated such that
27: end if th
BCRBr ∈ [α, β] , ∀r, for the (k + 1) cycle.
28: n & calc. t3 ;oupdate
Stop clock n energy
o using (32).
(k−w+1) S (k)
29: Tr = Tr \b zr zr
b (Update step).
b T , Br(k+1) (Section II-B); calc. t1 .
D. Variant 4 (Block 1)
30: Compute z̄Tr , Σ r
31: do This variant re-forms the coverage regions and the CHs
32: Start clock to measure t2 . Call Alg. 1- MBS. whenever any region requires retraining (i.e., retrain flag = 1).
33:
P
if n∈Rr A e r(k+1) (n, n) == Mr(k+1|k) then
(k+1) (k+1)
The updated remaining energy is given by (30). Like variant
34: Obtain Ar , Ar using (20), (5). (k+1|k)
(k+1) (k+1) (k+1) 3, Mr , N Crth , and δrth are changed for all the R
35: Calc.BCRBr usingAr ,Br in (16).
(k+1) regions. However, if retrain flag = 0 for all the regions and
36: if BCRBr < α then (k)
Erem (CHr ) ≈ 0 for any region r, then a SN from that
n o
(k+1|k) (k+1|k)
37: Mr ←max Mr − 1, 1 .
38: else if BCRBr
(k+1)
>nβ then
region with the highest remaining energy is elected as the
new CH and the updated remaining energies are given by
o
(k+1|k) (k+1|k)
39: Mr ←min Mr + 1, Nr .
40: end if (29). Otherwise, the regions continue their operations when
(k)
41: else exitf lag ← 1; break. Erem (CHr ) > 0, ∀r.
42: end if
(k+1)
43: while BCRBr ∈/ [α, β].
44: Set Mr
(k+1)
←Mr
(k+1|k)
. Stop clock; calc. t2 .
E. Complexity analysis
45: else exitf lag ← 1. The CHs of R regions run the proposed framework and the
46: end if complexity is computed as noted in [15]. Let the complexity of
47: Block 3: Run retraining logic; calc. t4 .
48: Update remaining energy using (33). Set k ← k + 1. the recovery scheme executed by the CHs be denoted by ORec .
49: end while Complexity of block 1 and 3 are respectively ≈ O (1) and
(k) O (Nr ). The overall complexity
N
d C r > N Crth hints the need of retraining the region r.  of the prediction
 step
 of block

However, keeping in mind the uncertainty that the logic is 2 run by each CHr is ≈ O (Nr ) + O (Mr(k) )3 + ORec .
based on the reconstructed signal x
(k)
br , which usually has some The complexity
 of update
 step in block 2 is given by 
 3     
(k+1|k) (k+1|k) 1 1
error, the threshold violation is checked for a few consecutive ≈ O Mr + Mr log2 ∆
O Ner log
ν
,
measurement cycles (3 cycles considered for simulation results
(·) where N er represents the number of variables after converting
in Section VI). If for all cycles N d C r > N Crth , the network problem (21) to a standard form by the CVX and ν is the
(k)
is retrained. Note that a component x br (·) is considered as precision accuracy (10−8 by default). Note that N er is a func-
non-sparse if it constitutes more than 0.1% of total energy of tion of N
r . The overall complexity of the proposed framework
2
(xb(k)
r (·))
  3
(k)
   
signal vector x br , i.e., P 2 > 0.1 . The 0.1% is ≈ O max Nr , Mr(k) (k+1|k) 3
+ ORec + O (Mr ) +
100

(k)
n∈Rr x
br (n)  
(k) (k+1|k) 1
  1

energy criteria ensures that no non-sparse component of x
br Mr log2 ∆
O Ner log
ν
.

1536-1276 (c) 2021 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: California State University Fresno. Downloaded on June 23,2021 at 11:40:23 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/TWC.2021.3062568, IEEE
Transactions on Wireless Communications

10

8000 8000

7000 7000

6000 6000

Network residual energy

Network residual energy


5000 5000 Variant 1
Variant 1
4000 4000 Variant 4
Variant 4
3000 3000
Variant 3 Variant 3
2000 2000
Variant 2 Variant 2
1000 1000

0 0
0 50 100 150 200 250 0 50 100 150 200 250 300 350 400 450
Measurement cycle Measurement cycle

(a) Es = 0.48 J (b) Es = 0.0048 J


Fig. 3: Performance of network residual energy of the proposed variants of decentralized sensing framework.

VI. S IMULATION R ESULTS (28), the value of R is obtained as 4. Initial value of the
hyperparameter estimates and convergence accuracy of the
Energy efficiency of the proposed adaptive decentralized  (0)
(k)
sensing framework is studied over the competitive decentral- SBL scheme are respectively set as γr (n) = 1, ∀n, ∀k
 
(l) (l+1)
ized approaches in [21], [18] and the centralized one in [15]. ≤ 10−7 . Performance measures
(k)

and
γr − γ (k)
r
The results show comparable sensing accuracy achieved due

2 2
to adaptive sensor selection and retraining logic. are, sensing error (MSE) = x(k) − x
b(k) = z(k) − b
z(k)

Consider N = 80 sensors equipped with non-rechargeable


PN (k)
and network residual energy = n=1 Erem (n). SBL-based
batteries and deployed in a 2-D WSN √ √ F ×F =
field of size signal estimation is averaged over 300 Monte-Carlo iterations
100 × 100 m2 , which is divided into d N e × d N e = 9 × 9 in each measurement cycle. The objective functions BCRBr
(k)

square areas of size ≈ 11×11 m2 each roughly containing one (k)


and SSr of (21) are ( normalized using their respective
randomly deployed SN [34]. The FC is located at a distance
)
 −1
(k) T (k) −1
     
1 (k) (k)
Tr Br diag aer Br + Γr
σ2
110 m above the center of the field. Consider a total S = 1 smallest values as, ( −1 )
(k) T (k) −1
   
1 (k)
pollution source located at the origin of the field. The pollution Tr
σ2
Br INr Br + Γr

signal at the source s, 1 ≤ s ≤ S, during k = 1 cycle


   
(k)

P Ec (n)+Es 

(k)
(f (k) (s)) is generated as an independent real Gaussian [23], 

n∈R r (k)
Erem (n)
a
er (n)


[35] with zero mean and variance σf2 (s) = 1. Modeling the and 
 Ec(k) (n)+Es

 


. CVX [28] is used to solve
(k)
temporal correlation using an AR(1) process [15], the temporal min
 (k)
Erem (n)
a
e r (n),n∈R r

samples for (k > 1) are generated as f (k)(s) = φf(k−1) (s)+
 

p the decoupled sensing problem (21). The parameter ∆ in the


1 − φ2 w(k) (s), where w(k) (s) ∼ N 0, σf2 (s) and φ is is set as 10−4 and the BCRB window is [α, β] =
MBS
temporal correlation coefficient between two consecutive sam- 10 , 10−4 . To effectively capture the process dynamics,
−5
−5
ples. Further, the spatial correlation is modeled using power the threshold thr is set as 10 and δrth is the average of
exponential model. During the k th cycle, the pollution signal threshold value of signal variation observed at SNs of that
PS −dn-s −dn-s
across SN n is obtained as z (k) (n) = s=1 e θ f (k) (s) region, i.e. N1r n∈Rr (e θ × 0.11), where 0.11 is the
P
[34], where dn-s is the distance between SN n and the threshold value of signal variation. The default window size
source s and θ is the spatial correlation parameter of the (k)
for Br estimate is set as w=6 as explained in Section VI-B.
source. Spatio-temporal dynamics of the pollution process Though the competitive approaches do not account for SN’s
are set as (θ, φ) = (100, 0.99). To guarantee a reasonable energy, for comparison energy consumption of the SNs in
sensing
 performance
 in severely ill-posed estimation scenarios these approaches are considered similar to the proposed one.
Nr /Mr  2 , the noise variance is fixed as σ 2 = 10−5 .
(k)

Ktr = 5 cycles are used for training/retraining purpose. A. Performance of the proposed CH selection variants
The existing works [12], [21] generate the sparse signal x(k) Network residual energy of all the variants versus measure-
with known sparsity and Gaussian distributed matrix B(k) ment cycles are plotted for two different values of the sensing
separately. In contrast, herein synthetic data, that considers energy Es in Figs. 3(a)-(b). Variant 3 offer the highest network
the effect of process dynamics, is used to generate the matrix lifetime. A higher network lifetime of variant 3 compared to
(k) (k) variant 1 is attributed to the fact that it re-forms coverage
Br and obtain the sparse signal xr without assuming any
prior knowledge of sparsity. Thus, the sensing framework regions and elects new CHs when the energy of the CH
evaluated on this synthetic data is readily applicable to real of any region depletes. This ensures that the SNs always
spatio-temporal process, as illustrated in simulation Section communicate with the CHs that require the least energy, which
VI-G. The diagonal elements of Γ(k) r are approximately set as
is not the case with variant 1. Lesser network lifetime of
(k) (k) the variant 2 is due to the CH-to-CH transmission overheads
the eigenvalues of ΣTr [15]. The sparsification matrix Br
(k) incurred in each cycle.
is set as eigen vectors of ΣTr . The initial energy of SNs

(0) Table II shows the network lifetime (total number of mea-
is set to 100 J i.e. Erem (·) = 100 and the other energy surement cycles executed (K)) and normalized average node
parameters are: Eelec = 50 nJ/bit, η = 10−10 J/bit/m2 [32], lifetime (normalized by K) in different CH selection variants.
b = 50, ν = 2, Vf = 0.8 V, If = 0.05 A, f = 60 MHz Note that, the terms normalized average node lifetime and
[36], and f˜ = 3.4 GHz as mentioned in Section IV-A. Using node lifetime are used interchangeably in this work. It can

1536-1276 (c) 2021 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: California State University Fresno. Downloaded on June 23,2021 at 11:40:23 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/TWC.2021.3062568, IEEE
Transactions on Wireless Communications

11

TABLE II: Network lifetime (in cycles) and normalized average C. Performance comparison with state-of-the-art
node lifetime of the proposed variants
Es = 0.48 J Es = 0.0048 J
Performance of the competitive variants of the proposed
Variant Network lifetime Node lifetime Network lifetime Node lifetime decentralized sensing framework are now compared with the
1 85 0.7065 71 0.7375 closest decentralized approach developed by Hwang et al. [21]
2 105 0.8725 95 0.9144
3 213 0.6806 438 0.5739
and centralized framework [15] in Figs. 5 and 6 respectively.
4 91 0.6506 159 0.6164 It can be observed from Figs. 5(a)-(c) that the proposed
be observed that, despite being less energy-efficient than the variants with standout performance (variant 3 with maximum
other variants, variant 2 has a higher node lifetime which can network lifetime, variant 2 with maximum node lifetime) offer
help in preventing network coverage outage. significant improvement in energy efficiency and node lifetime
Remark 1. Variant 3 offers a higher network lifetime, variant (or lower outage of network coverage) without compromis-
2 offers a higher node lifetime. ing the sensing quality. Energy consumption per cycle in
Hwang’s method and that with variants 2 and 3 are respec-
4 12 tively 223.274, 42.12, and 33.909 J/cycle. Thus, the variants
Total energy spent in overheads (in J)

Total energy spent in overheads (in J)

3.5
10 2 and 3 are respectively 81.135% and 84.813% more energy
3
8
efficient compared to the Hwang’s method. The gain in energy
2.5

2 6
efficiency is due to non-iterative information exchange by the
1.5
SNs in a measurement cycle and the adaptation mechanism
4
1 in decentralized sensor selection. On the other hand, node-
0.5
2
level energy saving is due to the consideration of unequal
0
3 4 5 6 7 8
0
3 4 5 6 7 8 9
energy consumption and remaining energy information in the
Window size (w) Window size (w)
proposed framework. The discontinuities in Fig. 5(c) show
Fig. 4: Variation of total overhead with window size. η = that, whenever the sensing quality of a region becomes poor,
10−7 J/bit/m2 and Es = 0.0048 J.
the proposed retraining logic detects it, and retraining is
B. Choice of window size w conducted in that region. Due to space constraint, MSE plot of
the estimated signal of only one region is shown. Further, these
Using a higher number of instances w of the estimated
(·) (k) discontinuities occur during different measurement cycles in
signal bzr increases the accuracy of Br computation. It
different regions; thus, plotting the MSE summed over all the
decreases the overhead of frequent retraining, but increases the
(·) regions is not possible at these points.
communication overhead, as w instances of b zr are transmitted From Fig. 6(a) it can be observed that the proposed variants
from a CH to the chosen new CH. Thus, the chosen w needs to 2 and 3 outperform the centralized approach in [15] when the
balance between the retraining and CH-to-CH communication power amplification component η ≥ 2.5 × 10−5 . Compared to
overheads. Retraining overhead is calculated as the sum of the centralized approach, these variants have reduced energy
sensing energy of all the SNs in all retraining phases; CH-to- consumed in direct SN-FC communication. For the case
CH communication overhead is the sum of transmission and η < 2.5 × 10−5 , only the variant 3 performs better than the
(·)
reception energy consumed in sending w instances of b zr from centralized one. The reason for the variant 2 not performing
one CH to another. From Figs. 4(a)-(b) it is observed that, with better in this case is that, the energy consumed in CH-to-
the considered network settings the total energy spent in the CH communication overheads incurred in each cycle outweigh
overheads is minimum when w = 6 in both the variants 1 and the energy gain achieved via reduced SN-FC communication
2. A similar behavior is seen for variant 4 as well. In variant overhead. However, the component of energy consumed in the
3, there is only retraining overhead. However, for consistency SN-FC communication overhead in the proposed decentralized
in performance comparison among all the variants in Section case are still lesser than that in the centralized case [15]. Note
VI-A, w is set as 6. that, loss in sensing accuracy in some measurement cycles in
Since the sensor selection and process estimation tasks are Fig. 6(b) is addressed by the proposed retraining logic and
performed region-wise to monitor the same spatio-temporal is shown as discontinuities in the MSE curve. It can also be
process, the parameter w is kept the same for all regions to observed that the frequency of retraining is high in the later
(·)
maintain the same computation accuracy of Br , ∀r, across the cycles. Due to low network residual energy in these cycles,
regions. Different choices of w would be desirable for multi- there is a lesser flexibility in sensor selection from all the SNs
sensing scenario [37] where each region monitors a different in Rr , as some of them may not participate owing to insuf-
process. ficient remaining energy. Although the proposed framework
(·)
Note that w instances of the estimated signal b zr are used in may select more number of SNs to satisfy BCRBr ∈ [α, β],
(·)
computing the matrix Br in every measurement cycle except the resulting sensing quality may still be poor which results
for the one that follows the training/retraining phase. For the in frequent retraining of the region.
cycle immediately after training/retraining, Ktr instances of An acceptable range of sensing quality (see [38]) achieved
(·)
the actual signal zr are used. To mitigate the effect of error in the proposed framework validates suitability of the used
due to the use of previously estimated signal vectors in the BCRB. The BCRB window [α, β] is set as per the mini-
(·)
Br matrix computation, it is also ensured that w > Ktr . mum and maximum BCRB values achieved in the respec-
(1)
Remark 2. The choice of w is required to limit the retrain- tive comparative approaches. λr , ∀r, in (21) are set such
ing overheads without significantly increasing the CH-to-CH that the minimum number of SNs are activated and the
communication overheads.
1536-1276 (c) 2021 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: California State University Fresno. Downloaded on June 23,2021 at 11:40:23 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/TWC.2021.3062568, IEEE
Transactions on Wireless Communications

12

4
× 10
4 80

Number of SNs with 0 remaining energy


Hwang's Variant 3
-2
3.5 70 method 10 (Region 2) Variant 2
100
(Region 2)
Hwang's method 10
-4
60
Network residual energy

3
Variant 3 120 160 200
2.5 50

MSE
Variant 2 10-2
2 40

1.5 30
Variant 3
1 20 Variant 2
Hwang's method 10
-4

0.5 10

0 0
0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000
Measurement cycles Measurement cycle Measurement cycles
(a) (b) (c)
Fig. 5: Comparison of (a) network residual energy, (b) number of SNs with 0 remaining energy, and (c) MSE of the proposed framework
(0)
with Hwang’s method [21]. Es = 0.0048 J, η = 10−7 J/bit/m2 , and Erem (·) = 500 J. The best performing variants are considered: variant
3 (maximum network life), variant 2 (maximum node life).
200
Centralized
Variant 3
Variant 3 approach
10 0 (η = 10 -5) -5
160 (η = 10 )
Network lifetime (K)

Variant 2
120 -5
(η = 10 )

MSE
10 -2
80
Centralized
approach Variant 2

40
10 -4 Coverage regions & edge nodes
re-formation and retraining

0
1 2 3 4 5 6 7 8 9 10 0 40 80 120 160 200
η × 10-5 Measurement cycle
(a) (b)
Fig. 6: Comparison of (a) network lifetime versus power amplification parameter η and (b) MSE of the best performing variants: variant 3
(maximum network lifetime), variant 2 (maximum node lifetime) of the proposed decentralized framework and that of centralized framework
(0)
[15]. Parameters values: Es=0.48 J and Erem (·)=100 J.
TABLE III: Performance gap: centralized versus decentralized BCRB (R = 4)
M (k) 12 16 20 24 28 32 36 40
Centralized avg BCRB(k) (×10−5 ) 3.287 2.671 2.326 2.157 2.096 2.062 2.0370 2.012
avg MSE(k) (×10−4 ) 11.019 9.258 7.440 6.291 5.653 5.270 4.978 4.786
Econs
avg Etotal (%) 46.277 48.105 50.234 52.654 55.355 58.362 61.656 65.243

(×10−5 )
PR (k)
avg r=1 BCRBr 15.961 12.424 10.223 8.977 8.233 7.572 7.177 6.834
Decentralized
(×10−4 )
PR (k)
avg r=1 MSEr 196 190 8.2 6.747 6.007 5.547 5.167 4.937
Econs
avg E (%) 46.177 47.392 48.405 49.205 49.856 50.576 51.352 52.181
total

(1)
achieved BCRBr ∈ [α, β] , ∀r. The usage of sparsification approximation is tabulated in Table III. It can be observed
matrix, computed using w instances of previously estimated by comparing respective BCRBs and MSEs that the per-
signal, in current cycle’s signal estimation in the decentralized formance loss decreases with increasing number of active
framework results in MSE variations (seen in Figs. 5(c) and SNs i.e., M (k) . Further, the energy consumption percentage
Econs
6(b)) owing to ensuing error which gets accumulated over (E total
%) of the centralized approach increases at a faster pace
cycles. These variations are absent in the Hwang’s method compared to the decentralized approach. Trade-off between
and the centralized scheme because of the considered strong the sensing quality and energy consumption can also be seen
assumption of known sparsification matrix and the use of as the number of active SNs changes. The weight λ(k) of
matrix computed using all the estimated signal instances centralized M (k) is used to set the weights of decentralized
(k)
available from the beginning of network operation respectively. problem as λr = λ(k) , ∀r. Further, it was observed that
(k)
PR (k)
To address the degradation of sensing quality, a retraining M need not necessarily be equal to r=1 Mr . The
logic is proposed. Further, these variations can also be reduced average BCRB values are computed after running centralized
by tightening the BCRB window at the expense of reduced and decentralized problems for 80 different sets of remaining
network lifetime. energy of the SNs in k = 1 cycle.
Remark 3. Thus, the proposed edge computing framework of-
Table IV summarizes the effect of number of clusters (R)
fers energy-efficient decentralized sensing of a slowly varying
on centralized and decentralized sensing performance, keeping
process, thereby extending the network lifetime.
M (k) same. It can be observed that as R increases (or cluster
PR (k)
size decreases), the MSE of the centralized approach (10)
D. BCRB(k) versus r=1 BCRBr decreases (because of improved feature selection) with no
Assuming all the SNs active, the BCRB is decoupled as significant change
PR (k) PRin energy consumption.
PR In the decentralized
BCRB(k) = r=1 BCRBr (16). For the cases when only approach (18), r=1 BCRBr and r=1 MSEr increase with
a few SNs are active, it is approximated as BCRB(k) < increase in R.PThe gap between BCRB of the centralized
PR (k) R
r=1 BCRBr . The performance degradation caused by this approach and r=1 BCRBr of the decentralized approach

1536-1276 (c) 2021 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: California State University Fresno. Downloaded on June 23,2021 at 11:40:23 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/TWC.2021.3062568, IEEE
Transactions on Wireless Communications

13

TABLE IV: Impact of R on performance of centralized and decentralized selection


Centralized Decentralized
R Econs Econs
M (k) avg BCRB(k) avg MSE(k)
PR (k) PR (k)
avg Etotal (%) avg r=1 BCRBr r=1 MSEr avg Etotal (%)
2 12 3.203 × 10−5 0.00586 46.2209 13.525 × 10 −5
0.0991 45.9833
4 12 3.287 × 10−5 0.0011 46.2773 15.961 × 10−5 0.0196 46.1773
6 12 3.820 × 10−5 0.00066 46.9012 27.309 × 10−5 0.0573 46.6825

PR with decreasing R. Similarly, the gap between MSE


decreases compared with δrth ) evaluates to lower values as mentioned in
and r=1 MSEr decreases. This is because the approxima- the work [15] as well for centralized sensing case. Increasing
PR
tion BCRB ≤ r=1 BCRBr becomes tight as R decreases. th
(·) would allow larger decrease in value of Mr which may
Further, though the decentralized problem consumes lesser result in increased network lifetime and MSE.
Econs
energy ( E total
%) compared to the centralized approach, the On increasing the number of coverage regions (or the degree
gap between two decreases with increasing R. This is because, of decentralization), the sensing quality deteriorates faster
the number of CHs increases with increased R, which results which invokes frequent retraining. This results in an increase
in an increased number of CH-FC communication. These com- in number of retraining phases as illustrated in Table VI for
munications consume significant energy due to long distances variant- 2. However, as R increases, the energy consumption
of the CHs from the FC. per cycle (network lifetime) first decreases (increases) and then
increases (decreases). Due to this, a criterion is developed in
E. Factors governing retraining in the network Section IV-B to find appropriate R such that minimum energy
Table V captures the number of times retraining occurs, is consumed per cycle. It may be noted that comment on the
using CH selection variant 4 for different values of R and w. MSE is made based on retraining count because different re-
By comparing the case {R, w} = {4, 6} with the cases {2, 8} gions undergo retraining during different measurement cycles
and {6, 4} it is observed that the retraining count decreases due to which computing overall MSE was not possible.
with the increase in window size w and/or decrease in the
number of coverage regions R, and vice-versa. Further, by
simultaneously increasing R and w (e.g., {2, 4} to {6, 8}) or G. Performance study using real WSN data-set
decreasing them (e.g., {6, 8} to {3, 5}) no significant change
in the retraining count is observed. The same trend is seen This section presents comparative performance results
with the other variants. of the proposed variants and Hwang’s method [21] and
Remark 4. In the proposed framework, retraining count DiSparSense [18] using humidity data collected by a WSN
decreases as the reconstruction accuracy/window size w in- deployed in Intel Berkeley research lab [39]. Signals sensed
creases and degree of decentralization of WSN decreases. by a set of 32 Mica2Dot sensors spanning over 25 × 32
m2 lab area are considered. The FC is considered to be
TABLE V: Retraining count for various {R, w} values in variant located at 40 m above the center of the considered field span.
(0) (0)
4. Erem (·) = 500 J, Es = 4.8 J Parameters [α, β] , δrth , th
r ∀r, R, w, Es , η, Eelec , and Erem (·)
{R, w} {4, 6} {2, 8} {6, 4} {2, 4} {6, 8} {3, 5}
2.1628 × 10−5 , 0.0522 ,

Retraining count 15 1 41 7 8 10
are respectively set as
[0.0709, 0.4892, 0.8305, 0.0532] , 0.5∀r, 4, 6, 4.8 J, 10−10
J/bit/m2 , 140 nJ/bit, 50 J. Performance metric used in the
F. Impact of various parameters on WSN’s performance kz(k)
r −b r k
z(k)
2

simulation plots is NMSE = (k) 2


. It gauges the
An increase in value of β may increase network lifetime zr
and decrease sensing quality because the constraint BCRBr ∈ error’s energy against the signal energy. Note that it has not
(k)
[α, β] will be satisfied with a lower number of active SNs (i.e., been directly used in the framework, as the WSN signal zr is
lower Mr ). This effect was demonstrated for the centralized unknown. It can be observed from Table VII that the variants 2
sensor selection proposed in a prior study [15, Figs. 2(c) and (3) of the proposed framework provide 58.9424% (62.9614%)
3(c)]. Similar effect can be seen for α as well. and 72.1997% (74.9210%) energy efficiency compared to
The parameters δ th and th quantify increase/decrease in Hwang’s method (Gaussian B, M = 30) and DiSparSense,
the number of active SNs by effectively capturing the process respectively. Sensing quality of the proposed variants is
variations. For a coverage region r, decreasing δrth may result significantly better than the competitive schemes. Practical
in frequent increase in the value of Mr which leads to decrease significance of the data-driven PCA-based sparsification
in sensing error and network lifetime. Intuitively, a process matrix B of the proposed framework can be seen from the
with high temporal correlation (or slower variations) could be improvement achieved in the sensing quality when this B
effectively monitored using a lesser Mr . A higher value of is used in Hwang’s method instead of Gaussian B matrix.
th
δ(·) would be suitable for this. Further, the parameter th (·) aid Further, it can be verified that the variant 3 (respectively,
in quantifying the extent up to which the number of active 2) offers better network (respectively, node) lifetime. Unlike
SNs can be decreased for sensing in subsequent measurement the competitive schemes (Hwang’s and DiSparSense), a
cycle such that the resulting signal estimate will not be much higher gain of decentralization (i.e., energy efficiency) is
different from the prior available signal estimates. It also gives achieved by paying lower price (i.e., sensing quality) when
an approximate idea of spatial correlation of the process as the proposed framework is compared with the centralized
(k)
for high spatial correlation case, the heuristic δr (which is exhaustive scheme (with all active SNs).

1536-1276 (c) 2021 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: California State University Fresno. Downloaded on June 23,2021 at 11:40:23 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/TWC.2021.3062568, IEEE
Transactions on Wireless Communications

14

TABLE VI: Impact of number of coverage regions


R 2 3 4 5 6 7
Network Lifetime 173 188 194 192 179 175
Energy consumption per cycle (in J) 229.1995 218.7119 205.3688 208.0227 223.4185 228.1437
Retraining Count 4 13 15 24 27 38

TABLE VII: Performance comparison with state-of-the-art approaches


Hwang’s method: Gaussian random B (PCA-based B) Variant 2 Variant 3 DiSparSense Exhaustive
M 12 17 25 30 − − − 32
Average NMSE 1.8039 (2.7813 × 10−4 ) 1.7697 0.3963 0.3695 2.4886 × 10−4 2.0254 × 10−4 0.5405 1.9295 × 10−5
Econs 71.3274 (110.1692) 92.2521 164.2230 213.7741 87.7705 79.1789 315.7181 132.6000
Network lifetime 18 (13) 10 8 7 18 19 5 12
Node lifetime 10 (9) 10 8 7 12 9 4 11

VII. C ONCLUSION eigenvalues form the diagonal elements of the diagonal matrix
(k)
In this paper, a novel edge computing framework for Γr ∈ RNr ×Nr . Following the same procedure and denoting
(k) (k) (k)
decentralized sensing of a spatio-temporal process has been yr = y er −Ar z̄Tr , the sparse representation of the system
(k) (k) (k) (k) (k)
proposed. For this, a critical trade-off between the sensing model (4) is given by yr = Ar Br xr + nr , ∀r.
quality and energy efficiency has been distributedly optimized
at the dynamically selected edge nodes. Residual energy of
the SNs has been incorporated in the formulation which was A PPENDIX B
otherwise ignored in the literature. The framework adapts SBL A PPROACH FOR S IGNAL R ECOVERY
the number of active SNs in a coverage region according
to the process dynamics and keeps the sensing error within The SBL framework assumes a Gaussian likelihood model
a given limit. To limit the accumulation of estimation error for the estimation of the underlying monitored phenomenon
over the measurement cycles, a logic to detect the need [23]. During each cycle k, the SBL scheme assigns a param-
(k)
for retraining the network coverage regions has also been eterized Gaussian prior to the unknown weight vector xr as
proposed. Extensive simulation results have demonstrated that,  2
(k)
compared to the closest competitive decentralized approach   N 
Y −1/2 −
xr (n)
(k)
the proposed framework provides up to 84% higher energy p x(k) (k)
r ; γr = 2πγr(k) (n) e 2γr (n) . (34)
efficiency (network lifetime) and improves energy balance n=1

among the SNs without impacting the sensing quality. The (k)
energy efficiency of the proposed decentralized framework has The component
h γr (n) of the iT hyperparameter vector
(k) (k) (k)
been verified by executing it on a real WSN data-set as well. γr = γr (1) , . . . , γr (Nr ) ∈ RNr ×1 corresponds
(k)
to the variance of component xr (·) of the weight vector.
A PPENDIX A Due to the intractability of the maximum likelihood
 estimate

(k) k (k) (k)
S IGNAL S PARSIFICATION : A PCA- BASED A PPROACH of γ r , i.e. γ b r = arg maxγ (k) 0 log p yr ; γ r [23],
r
At the onset, Ktr training instances of signal vector z(·) , an iterative expectation maximization (EM) algorithm
(k)
forming a learning set T = z(k−Ktr +1) , . . . , z(k−1) , z(k) , is employed to estimate γ (k) r considering xr as a
are P considered to compute the sample mean vector z̄T = hidden variable. Let the estimate of the hyperparameter
(k)
1
|T | z(·) ∈T z
(·)
∈ RN ×1 and the sample covariance matrix vector γ r in the lth EM iteration be denoted as
T  (l)
bT = 1 ∈ RN ×N . The b (k) . The E-step in the lth iteration evaluates the

z(·) − z̄T z(·) − z̄T γ
P
Σ |T | (·)z ∈T r
sparsification matrix B(k) ∈ RN ×N is formed using orthog-
  (l) 
likelihood as L γ (k)
r | γ (k)
r = E (k) (k)  (k) (l) {
onal eigen vectors of the matrix Σ b T , while its eigenvalues  o xr |yr ; γ r

form the diagonal elements of a diagonal matrix Γ(k) ∈ (k) (k)


log p yr , xr ; γ (k)
r . The posterior distribution
RN ×N (used later), i.e., Σ b T = B(k) Γ(k) B(k) T . Owing to
   (l)  
(k) (k) (k) (k) (l) (l)
of xr is p xr |yr ; γ r ∼ N µ (k) , Σ (k)
the correlation present in elements of the set T , projection T
xr xr

of z(k) onto vector space R B(k) gives the corresponding



(l) (l) (k) (k) (l)
with µ (k) = σ −2 Σ (k) Θr yr ∈ RNr ×1 , Σ (k) =
T (k)  xr xr xr
sparse vector x(k) as x(k) = B(k) z − z̄T . Substitut-  (l) −1
!−1
 (k)(l)
(k) T b (k)
 
(k)
ing the sparse representation z(k) = z̄T + B(k) x(k) in (2) and σ −2 Θr Θr + Γ r ∈ RNr ×Nr , Γ
b
r =
denoting y(k) = y e (k) − A(k) z̄T ∈ RMk ×1 , the system model 
(k)
(l) 
(k)
(l) 
becomes y = A(k) B(k) x(k) + n(k) .
(k) diag γ
br (1) ,··· , γbr (Nr ) . A less
Likewise, for the regional system model (4), (l)
complex expression of Σ (k) is obtained using
compute z̄Tr and Σ bT
r
using the learning set xr
(l)
Tr = {zr
(k−Ktr +1) (k−1) (k)
, . . . , zr , zr }. Note that Σ b T is the matrix inversion lemma [26] as Σ (k) =
r xr
 (k) (l)  (k) (l)  T  −1  (k) (l)
a full rank matrix irrespective of whether Nr is greater Γr
b − Γr
b Θr
(k) (l)
Σ (k) Θr
(k) b
Γr with
than, less than, or equal to Ktr . The eigen vectors of Σ bT
r  y
 (k) (l) r T 
(k)
give a unique regional sparsification matrix Br ; while the
(l)
Σ (k) = σ 2 I (k) + Θr
Mr
(k) b
Γr Θr
(k)
∈ RMr ×Mr .
yr

1536-1276 (c) 2021 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: California State University Fresno. Downloaded on June 23,2021 at 11:40:23 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/TWC.2021.3062568, IEEE
Transactions on Wireless Communications

15

 (l+1)
Thereafter, the M-step estimates γb r(k) as: [6] T. Xue, X. Dong, and Y. Shi, “Multiple access and data reconstruction
 (l+1) n  o in wireless sensor networks based on compressed sensing,” IEEE Trans.
b r(k)(n)
γ = arg max E (k) (k)
 (k) (k) (k)
(k) (l) log p yr , xr ; γ r
 Wireless Commun., vol. 12, no. 7, pp. 3399–3411, Jul. 2013.
(k) xr |yr ; γ r
γ r (n) [7] Y. Chen and Q. Zhao, “On the lifetime of wireless sensor networks,”
(l)

(l)
2 IEEE Commun. Lett., vol. 9, no. 11, pp. 976–978, Nov. 2005.
= Σ (k) (n, n) + µ (k) (n) . (35) [8] S. Joshi and S. Boyd, “Sensor selection via convex optimization,” IEEE
xr xr
After repeating the E and M-step for LEM iterations, the final Trans. Signal Process., vol. 57, no. 2, pp. 451–462, Feb. 2009.
(k) (LEM ) [9] S. P. Chepuri and G. Leus, “Sparsity-promoting sensor selection for
sparse estimate is obtained as x br = µ (k) . Without prior non-linear measurement models,” IEEE Trans. Signal Process., vol. 63,
xr
knowledge of sparsity,the SBL-EM scheme provides a maxi- no. 3, pp. 684–698, Feb. 2015.

(k)
(l) 
(k)
(l)  [10] O. M. Bushnaq, A. Chaaban, and T. Al-Naffouri, “Joint sensor loca-
mally sparse solution γbr (n) →0 ⇒ x br (n) →0 . tion/power rating optimization for temporally-correlated source estima-
(k) (k) (k) tion,” in Proc. IEEE Intl. Wksp. Signal Process. Advances Wireless
The estimate b zr is given by z̄Tr + Br x br . Commun. (SPAWC). Sapporo, Japan, 2017, pp. 1–5.
[11] O. M. Bushnaq, A. Chaaban, S. P. Chepuri, G. Leus, and T. Y. Al-
A PPENDIX C Naffouri, “Sensor placement and resource allocation for energy harvest-
ing IoT networks,” Elsevier Digital Signal Process., vol. 105, pp. 1–14,
BAYESIAN C RAM ÉR -R AO BOUND (BCRB) 2020.
BCRB characterizes the MSE in the estimate of [12] W. Chen and I. J. Wassell, “Optimized node selection for compressive
(k)
the unknown vector  x in the signal model (6). sleeping wireless sensor networks,” IEEE Trans. Veh. Technol., vol. 65,
BCRB(k) = Tr J−1 B , with the Bayesian Fisher no. 2, pp. 827–836, Feb. 2016.
informationmatrix JB ∈ RN×N given below [15]:JB = [13] G. Quer, R. Masiero, G. Pillonetto, M. Rossi, and M. Zorzi, “Sensing,
  compression, and recovery for WSNs: Sparse signal modeling and
∂ 2L y(k) |x(k) ; Γ(k)  ∂ 2L x(k) ;Γ(k) 
monitoring framework,” IEEE Trans. Wireless Commun., vol. 11, no. 10,
−E(y(k) ,x(k)) −E (x(k)) (k) .
 ∂x(k) ∂x(k) T  ∂x ∂x(k) 
T pp. 3447–3461, Oct. 2012.
| {z }| {z } [14] J. Hao, B. Zhang, Z. Jiao, and S. Mao, “Adaptive compressive sensing
JD    JP  based sample scheduling mechanism for wireless sensor networks,”
Pervasive and Mobile Comput., vol. 22, pp. 113–125, Sep. 2015.
The terms L y(k) |x(k) ; Γ(k) , L x(k) ; Γ(k) , and [15] V. Gupta and S. De, “SBL-based adaptive sensing framework for WSN-
JD , JP ∈ RN ×N respectively denote the log-likelihood assisted IoT applications,” IEEE Internet Things J., vol. 5, no. 6, pp.
functions of vectors y(k) , x(k) parameterized by Γ(k) , and 4598–4612, Dec. 2018.
(k) (k) [16] V. Gupta and S. De, “Adaptive multi-sensing in EH-WSN for smart
FIMs with respect to y , x . The prior distribution of environment,” in Proc. IEEE Global Commun. Conf. (GLOBECOM).
x(k) (34) is used to compute the log-likelihood  function as: IEEE, 2019, pp. 1–6.
   T  −1
L x(k) ; Γ(k) = e k − 12 x(k) Γ(k) x(k) , where e k is [17] H. Jamali-Rad, A. Simonetto, and G. Leus, “Sparsity-aware sensor
selection: Centralized and distributed algorithms,” IEEE Signal Process.
a parametric constant. Further, its second order derivative Lett., vol. 21, no. 2, pp. 217–220, Feb. 2014.
2 (k) −1 [18] H. Jamali-Rad, A. Simonetto, X. Ma, and G. Leus, “Distributed sparsity-
∂ L(x )
with respect to x(k) evaluates to (k) (k) T = Γ(k) . aware sensor selection,” IEEE Trans. Signal Process., vol. 63, no. 22,
∂x (∂x )
 −1 pp. 5951–5964, Nov. 2015.
(k) [19] S. Liu, S. P. Chepuri, G. Leus, and A. O. Hero, “Distributed sensor
Thus, JP = Γ . Similarly, after ignoring the selection for field estimation,” in Proc. IEEE Int. Conf. Acoust., Speech,
(k) (k)

constant terms, the log-likelihood function L y ; x is Signal Process. (ICASSP). New Orleans, LA, USA, Mar. 2017, pp.
  2 4257–4261.
(k) (k) (k) 1 (k) (k) (k) (k)
obtained as: L y |x ; Γ = 2σ2 y − A B x . [20] F. Altenbach, S. Corroy, G. Böcherer, and R. Mathar, “Strategies for
T T distributed sensor selection using convex optimization,” in Proc. IEEE
Thus, JD = σ12 B(k) A(k) A(k) B(k) . The BCRB is Global Commun. Conf. (GLOBECOM). Anaheim, CA, USA, Dec.
expressed as: ( ) 2012, pp. 2367–2372.
1  (k)T  (k)T (k) (k)  (k)−1 −1 [21] S. Hwang, R. Ran, J. Yang, and D. K. Kim, “Multivariated Bayesian

(k)
BCRB = Tr B A A B + Γ . (36) compressive sensing in wireless sensor networks,” IEEE Sensors J.,
σ2
(k) = vol. 16, no. 7, pp. 2196–2206, Apr. 2015.
For
(
colored noise case, BCRB evaluates −1)
to: BCRB [22] R. Masiero, G. Quer, D. Munaretto, M. Rossi, J. Widmer, and M. Zorzi,
T 1 1
 −1
Tr B(k) (Σn ) 2 diag a e(k) (Σn ) 2 B(k)+ Γ(k)

. “Data acquisition through joint compressive sensing and principal com-
ponent analysis,” in Proc. IEEE Global Telecommun. Conf. Honolulu,
HI, USA, Nov. 2009, pp. 1–6.
ACKNOWLEDGMENT [23] D. P. Wipf and B. D. Rao, “Sparse Bayesian learning for basis selection,”
The authors would like to thank the anonymous reviewers and the editors IEEE Trans. Signal Process., vol. 52, no. 8, pp. 2153–2164, Aug. 2004.
for their pertinent comments and constructive criticisms that have helped [24] A. Mishra, V. Gupta, S. Dwivedi, A. K. Jagannatham, and P. K.
improve the quality of the work. Varshney, “Sparse bayesian learning-based target imaging and parameter
estimation for monostatic MIMO radar systems,” IEEE Access, vol. 6,
R EFERENCES pp. 68 545–68 559, 2018.
[25] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge univ.
[1] X. Cao, L. Liu, Y. Cheng, and X. Shen, “Towards energy-efficient press, 2004.
wireless networking in the big data era: A survey,” IEEE Commun. [26] K. B. Petersen, M. S. Pedersen et al., “The matrix cookbook,” Tech.
Surveys Tut., 1st Quart. 2017. Univ. Denmark, vol. 7, p. 15, Nov. 2008.
[2] V. Gupta, S. Tripathi, and S. De, “Green sensing and communication: [27] Z. Fei, B. Li, S. Yang, C. Xing, H. Chen, and L. Hanzo, “A survey
A step towards sustainable IoT systems,” Springer J. Indian Inst. Sci., of multi-objective optimization in wireless sensor networks: metrics,
pp. 1–16, 2020. algorithms, and Open Problems,” IEEE Commun. Surveys Tut., vol. 19,
[3] G. Anastasi, M. Conti, M. Di Francesco, and A. Passarella, “Energy no. 1, pp. 550–586, 1st Quart. 2017.
conservation in wireless sensor networks: A survey,” Ad hoc netw., vol. 7, [28] M. Grant and S. Boyd, “CVX: Matlab software for disciplined convex
no. 3, pp. 537–568, May 2009. programming, version 2.1,” http://cvxr.com/cvx, Mar. 2014.
[4] M. Hooshmand, M. Rossi, D. Zordan, and M. Zorzi, “Covariogram- [29] A. D. Marbini and L. E. Sacks, “Adaptive sampling mechanisms in
based compressive sensing for environmental wireless sensor networks,” sensor networks,” in Proc. London Commun. Symp., vol. 174. London,
IEEE Sensors J., vol. 16, no. 6, pp. 1716–1729, Mar. 2016. UK, Sep. 2003.
[5] Q. Ling and Z. Tian, “Decentralized sparse signal recovery for compres- [30] M. A. Razzaque and S. Dobson, “Energy-efficient sensing in wireless
sive sleeping wireless sensor networks,” IEEE Trans. Signal Process., sensor networks using compressed sensing,” Sensors, vol. 14, no. 2, pp.
vol. 58, no. 7, pp. 3816–3827, Jul. 2010. 2822–2859, 2014.

1536-1276 (c) 2021 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: California State University Fresno. Downloaded on June 23,2021 at 11:40:23 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/TWC.2021.3062568, IEEE
Transactions on Wireless Communications

16

[31] V. Tiwari, S. Malik, and A. Wolfe, “Power analysis of embedded


software: a first step towards software power minimization,” IEEE Trans.
Very Large Scale Integr. (VLSI) Syst., vol. 2, no. 4, pp. 437–445, Dec.
1994.
[32] W. B. Heinzelman, A. P. Chandrakasan, and H. Balakrishnan, “An
application-specific protocol architecture for wireless microsensor net-
works,” IEEE Trans. Wireless Commun., vol. 1, no. 4, pp. 660–670, Oct.
2002.
[33] H.-S. Park and C.-H. Jun, “A simple and fast algorithm for k-medoids
clustering,” Expert syst. appl., vol. 36, no. 2, pp. 3336–3341, 2009.
[34] M. Leinonen, M. Codreanu, and M. Juntti, “Sequential compressed sens-
ing with progressive signal reconstruction in wireless sensor networks,”
IEEE Trans. Wireless Commun., vol. 14, no. 3, pp. 1622–1635, Mar.
2015.
[35] J. V. Zidek, W. Sun, and N. D. Le, “Designing and integrating composite
networks for monitoring multivariate Gaussian pollution fields,” J. Royal
Statist. Soc.: Series C (Appl. Statist.), vol. 49, no. 1, pp. 63–79, 2000.
[36] A. Sinha and A. P. Chandrakasan, “JouleTrack: a web based tool for
software energy profiling,” in Proc. Des. Autom. Conf. ACM, Jun.
2001, pp. 220–225.
[37] V. Gupta and S. De, “Collaborative Multi-sensing in Energy Harvesting
Wireless Sensor Networks,” IEEE Trans. Signal Inform. Process. over
Netw., vol. 6, pp. 426–441, May 2020.
[38] D. Hooper, J. Coughlan, and M. Mullen, “Structural equation modelling:
Guidelines for determining model fit,” The Electron. J. Bus. Res.
Methods, vol. 6, pp. 53–60, Jan. 2008.
[39] P. Bodik, W. Hong, C. Guestrin, S. Madden, M. Paskin, and R. Thibaux,
“Intel lab data,” Online dataset, Feb. 2004.

Vini Gupta received the B.Tech. degree in Elec-


tronics and Communication Engineering from Indira
Gandhi Institute of Technology, Guru Gobind Singh
Indraprastha University, New Delhi, India, in 2013
and the M.Tech. degree in Signal Processing and
Communication from the Department of Electrical
Engineering, IIT Kanpur, Uttar Pradesh, India, in
2016. She worked in Tata Consultancy Services Lim-
ited, New Delhi, as an Assistant System Engineer,
from 2013 to 2014. She is currently working toward
the Ph.D. degree in the Department of Electrical En-
gineering, IIT Delhi, New Delhi, India. Her research interests include design
of centralized and distributed green sensing framework for wireless sensor
networks-based IoT applications, application of sparse signal processing and
Bayesian learning for networked sensing, and radar signal processing. She
was a recipient of TCS RSP Fellowship (2016-2020).

Swades De (S’02-M’04-SM’14) received the


B.Tech. degree in Radiophysics and Electronics from
the University of Calcutta in 1993, the M.Tech. de-
gree in Optoelectronics and Optical communication
from IIT Delhi in 1998, and the Ph.D. degree in
Electrical Engineering from the State University of
New York at Buffalo in 2004.
Dr. De is currently a Professor with the Depart-
ment of Electrical Engineering, IIT Delhi. Before
moving to IIT Delhi in 2007, he was a Tenure-Track
Assistant Professor with the Department of ECE,
New Jersey Institute of Technology, Newark, NJ, USA, from 2004–2007.
He worked as an ERCIM Post-doctoral Researcher at ISTI-CNR, Pisa, Italy
(2004), and has nearly five years of industry experience in India on telecom
hardware and software development, from 1993–1997, 1999. His research
interests are broadly in communication networks, with emphasis on perfor-
mance modeling and analysis. Current directions include energy harvesting
sensor networks, broadband wireless access and routing, cognitive/white-
space access networks, smart grid networks, and IoT communications. Dr. De
currently serves as an Area Editor of IEEE COMMUNICATIONS LETTERS
and Elsevier Computer Communications, and an Associate Editor of IEEE
TRANSACTIONS ON VEHICULAR TECHNOLOGY, IEEE WIRELESS
COMMUNICATIONS LETTERS, and IEEE NETWORKING LETTERS.

1536-1276 (c) 2021 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: California State University Fresno. Downloaded on June 23,2021 at 11:40:23 UTC from IEEE Xplore. Restrictions apply.

You might also like