Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com
Available online at www.sciencedirect.com

ScienceDirect
Procedia Computer Science 00 (2019) 000–000
Procedia
Procedia Computer
Computer Science
Science 00 (2019)
171 (2020) 000–000
2242–2251 www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia

Third International Conference on Computing and Network Communications (CoCoNet’19)

Third International Conference on Computing and Network Communications (CoCoNet’19)
Data
Data Aggregation
Aggregation using
using Compressive
Compressive Sensing
Sensing for
for Energy
Energy Efficient
Efficient
Routing Strategy
Routing Strategy
Deepa Puneetha,b,∗, Muralidhar Kulkarnia,c
Deepa Puneetha,b,∗, Muralidhar Kulkarnia,c
a National Institute of Technology Karnataka, Surathkal, Mangalore-575025, India.
a National Institute of Technology Karnataka, Surathkal, Mangalore-575025, India.
b deepapuneethk@gmail.com
b deepapuneethk@gmail.com
c mkul@nitk.edu.in
c mkul@nitk.edu.in

Abstract
Abstract
Data gathering in an energy efficient way, is one of the crucial concerns in Wireless Sensor Networks (WSNs). With incorporation
Data gathering inSensing
of Compressive an energy efficient
(CS), way,
as a data is one of the
aggregation crucial
scheme theconcerns in Wireless
information contained Sensor
in theNetworks (WSNs).
signal, is safely With incorporation
maintained through its
of Compressive Sensing (CS), as a data aggregation scheme the information contained in the signal, is safely
projections, which can be reconstructed later. Inclusion of CS, for an energy efficient routing technique further enhances the maintained through its
lifetime
projections, which
of the network. To can be reconstructed
improve later. Inclusion
network lifetime, of CS,
CS has been for an energy
employed at levelefficient routing
1 (at the technique
leaf nodes). Thefurther enhancesreduction
dimensionality the lifetimeat
of
thethe network.isTodone,
transmitter improve
usingnetwork lifetime, matrix.
a measurement CS has At
been
theemployed at level
receiver, data 1 (at the leaf
is recovered nodes).
using The
greedy dimensionality
based reduction
methods. Greedy basedat
the transmitter is done, using a measurement matrix. At the receiver, data is recovered using greedy based
method offers low complexity, and low implementation cost. The success rate of greedy method, depends on the sparsity of the methods. Greedy based
method offers low evaluation
data. Performance complexity,ofand low based
greedy implementation
method is cost. The success
analyzed, rate of greedy
by considering varyingmethod,
sparsity depends
and plottedon the sparsity
against numberof the
of
data. Performance
measurements, evaluation
required of greedythe
to reconstruct based method
signal. is analyzed,
At the same time, byreconstruction
considering varying
error ofsparsity and plotted
every method against
has been number by
evaluated of
measurements,
varying sparsityrequired to reconstruct
of the signal. The samethe signal. Atbythe
is analyzed same time,real
considering reconstruction
temperature,error of every data
and humidity method
sets.has been evaluated by
varying sparsity of the signal. The same is analyzed by considering real temperature, and humidity data sets.
c 2020

© 2020 The
The Authors.
Authors. Published
Published by
by Elsevier
Elsevier B.V.
B.V.
c 2020
 The Authors. Published by Elsevier B.V.
This is
This is an
an open
open access
access article
article under
under the
the CC
CC BY-NC-ND
BY-NC-ND license
license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
This is an open
Peer-review access
under article under
responsibility of the
theCC BY-NC-ND
scientific licenseof(http://creativecommons.org/licenses/by-nc-nd/4.0/)
committee the Third International Conference on Computing and Network
Peer-review
Communications
Communications under(CoCoNet’19).
responsibility of the scientific committee of the Third International Conference on Computing and Network
(CoCoNet’19).
Communications (CoCoNet’19).
Keywords: Wireless Sensor Network’s (WSN’s); Compressive Sensing (CS); greedy algorithm; data aggregation, multi-path routing algorithm;
Keywords: Wireless Sensor Network’s (WSN’s); Compressive Sensing (CS); greedy algorithm; data aggregation, multi-path routing algorithm;
network life-time;
network life-time;

1. Introduction
1. Introduction
The main objective of sensor is to sense the physical changes in the area of application, which may include envi-
The main
ronmental objective of
monitoring, sensorapplications,
security is to sense the physical
health changesetc.
monitoring in the
Theseareasensors
of application, which may
have to transmit include envi-
the observations,
ronmental monitoring, security applications, health monitoring etc. These sensors have to transmit the observations,
to the control unit usually the Base Station (BS) in order to perform necessary action. As these sensors are randomly
to the control
deployed, fewunit usually
nearby the Base
sensors Station
may pick up (BS)
sameinobservations
order to perform
whichnecessary action.
is redundant. As these
These sensorscauses
redundancy are randomly
burden
deployed, few nearby sensors may pick up same observations which is redundant. These redundancy
to the entire network, which finally affects the network lifetime. The energy resources, and communication rangecauses burden
of a
to the entire network, which finally affects the network lifetime. The energy resources, and communication range of a

∗ Corresponding author.
∗ Corresponding
E-mail address:author.
deepapuneethk@gmail.com
E-mail address: deepapuneethk@gmail.com
1877-0509
1877-0509  c 2020
© 2020 The Authors. Published
The Authors. Published by
by Elsevier
Elsevier B.V.
B.V.
1877-0509
This is an c 2020

open Thearticle
access Authors. Published
under the CC by Elsevier B.V.
BY-NC-ND licenselicense
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
This is an open access article under the CC BY-NC-ND (http://creativecommons.org/licenses/by-nc-nd/4.0/)
This is an open access article under
of the
Peer-review
Peer-reviewunder
underresponsibility
responsibility ofCC
the theBY-NC-ND
scientific
scientific license
committee of (http://creativecommons.org/licenses/by-nc-nd/4.0/)
the Third
committee International
of the Conference on
Third International Computingon
Conference andComputing
Network Communications
and Network
Peer-review under responsibility of the scientific committee of the Third International Conference on Computing and Network Communications
(CoCoNet’19).
Communications (CoCoNet’19).
(CoCoNet’19).
10.1016/j.procs.2020.04.242
 Deepa Puneeth et al. / Procedia Computer Science 171 (2020) 2242–2251 2243
2 Deepa Puneeth / Procedia Computer Science 00 (2019) 000–000

sensor node is limited. The expensive data transmission costs in general, can be reduced by using a suitable data com-
pression technique. CS is an emerging field, in which simultaneous sensing and compression offers a promising result,
particularly in large scale Wireless Sensor Network(WSN). It greatly reduces sampling and computational costs.
Multi path routing, facilitates more than one paths between source and sink, which helps to enhance the reliable data
delivery by transmitting multiple copies of data, via different routes. Multi hop communication between source and
BS is usually more energy-efficient, than single hop. CH (Cluster Head) closer to the BS relay more messages, than the
other CH, which leads to hot-spot problem or energy-hole problem. Therefore, how to select the best forwarding CH,
relative to the load imbalance becomes important point of concern. In the literature, we found several papers which
works on hierarchical routing protocols, which are multi-hop and multi-path but hot-spot or energy-hole problem is
not addressed properly for CS based hierarchical routing protocol.
Several CS Recovery algorithms can be found in the literature. Greedy based recovery algorithms capture special
interest, because of their simplicity and low implementation costs. We incorporate, greedy based recovery procedure
in order to retrieve the data samples from the compressed measurements.

Notation

x : The signal vector of length ’N’ which is to be compressed.

y : Compressed signal of length ’M’, M<N.
x̂ : Estimated signal from the compressed version.
Θ : Measurement/Projection matrix.

1.1. Related Work

Authors in [1] proposed, Data Aggregation for WSN’s based on LEACH-CS. In order to aggregate the data they
have employed Total Variation(TV) method. Simulations are conducted considering synthetic data, not experimented
with real data set. Authors of [2] proposed, a minimum-energy compressed data gathering problem (MECDA). Inves-
tigation of minimum-energy CDA problem providing both an exact solution (for small networks) and approximate so-
lutions (for large networks). A CS–oriented data aggregation technique for the multi–hop topology has been presented
in [3]. The proposed Data Aggregation and Recovery in WSN’s using CS, which addresses Quality of Service(QOS)
issues, by considering packet loss and energy dissipation. Authors [4, 5, 6] provides, the survey of the available greedy
methods employed in the field of WSN. The ease of implementation and reduced complexity of greedy based methods
provides an attractive way to implement data aggregation for WSN. Motivated by which, we employ greedy methods
in order to retrieve the data which is compressed using CS.

1.2. Compressive sensing for data aggregation

In CS, compression of the signal is done during sensing itself, and no additional processing is required in order to
compress the data, after acquisition. But we have to deal with costly recovery procedures, which involves obtaining
an unique solution from an under-determined system. Acquisition based on CS is shown in Fig.1

1.3. Signal acquisition and method of reconstruction in CS based system.

Fig. 1: Signal/Image acquisition in CS.

2244 Deepa Puneeth et al. / Procedia Computer Science 171 (2020) 2242–2251
Deepa Puneeth / Procedia Computer Science 00 (2019) 000–000 3

Let x = (x1 , x2 , x3 , .......xN )T is the signal of dimension N, and x is ’k’ sparse, when it consists of only ’k’ nonzero
values in the acquired signal either in the domain, in which it is acquired or in the transform domain. Consider the
Fig.2, which depicts the humidity values captured by the sensor. If we consider the signal in time domain, it cannot
be compressed as the signal is not sparse. CS can be employed, provided the signal is sparse. We perform the DCT of
the signal, then we can have larger and smaller component of the signal which can be considered as a sparse signal as
shown Fig.2.

3
Amplitude

2
1
00 2 4 6 8 101214
Samples

Fig. 2: Humidity signal in acquired and transformed domain.

We can define x, is k sparse in ψ if there are an orthonormal basis denoted by (ψ1 , ψ2 , ψ3 , .......ψN ) as in equation 1.

n

x= ki Ψi or x = Ψk (1)
i=1

CS theory affirms that, k sparse signal x of N dimension can be confined into y using M (M  N) linear projections
with the help of a M×N matrix as in equation 2,

y = Φx = ΦΨk = Θk
(2)
Where Φ is the projection or measurement matrix.

In order to preserve the information contained in x, the M×N matrix Φ must maintain the inherent properties of a k
sparse signal during the transformation of x ∈ RN to y ∈ R M . The ultimate goal is to estimate x from y. Measurement
vector in equation 2, has fewer equations, than unknowns. It is an under-determined system, thus it does not possess
an unique solution. CS theory suggests that, if x is a sparse vector it can have a unique solution, even though it’s an
under-determined system. The recovery of N dimensional signal x, from the M dimensional data values can be dealt
through combinatorial optimization.

mink∈RN kl0 such that y = Φx or x = Ψk (3)

 Deepa Puneeth et al. / Procedia Computer Science 171 (2020) 2242–2251 2245
4 Deepa Puneeth / Procedia Computer Science 00 (2019) 000–000

But the above solution 3 is NP-complete problem and considered to be numerically unsolvable. Surprisingly op-
timization based on L1 norm minimization gives the same result as that of L0 norm under certain conditions. The
solution to L1 minimization problem can be solved using linear programming [7] [8].

mink∈RN kl1 such that y = Φx or x = Ψk (4)

The recovery of data samples is guaranteed based on the measurement matrix. The matrix must possess certain
property. In that case even from an under determined system, it is possible to obtain good estimation of the data at
the receiving end. The same is verified by the Restricted Isometric Property (RIP) [8] of the matrix which is used for
the dimensionality reduction. Restricted isometric constant of the measurement matrix Φ, for a k sparse vector is the
smallest ∆k > 0 such that equation 5 holds good.

(1 − ∆k )x22 ≤ Φx22 ≤ (1 + ∆k )x22 (5)

Measurement matrix Φ, must satisfy the RIP. Successful recovery of the data depends on RIP. If ∆k = 0 for all k
≤ N, it implies that Φ is orthonormal. But when the signal measurement M is greater than original acquired signal
dimension N, the value of ∆k will be no longer zero. If the value is nearly zero, means that the measurement matrix is
nearly orthonormal, which increases the estimation accuracy of x.

2. CS Recovery Algorithms

The major fact of CS is to recover the signal from an under-determined system, which has infinitely many solu-
tions. The sparse recovery algorithms must be able to identify the subset of columns from the sensing matrix, which
contribute to the signal during compression. Once the correct subspace in which the signal projections are confined
are estimated, then non zero coefficients are evaluated using pseudo-inversion process.

2.1. Greedy iterative pursuit

This class of reconstruction algorithms, solves the recovery process by evaluating the solution in steps by an
iterative strategy. In each iteration the current evaluation for the solution vector x is refined. Halting criteria of the
algorithm varies for different matching pursuits, which may involve the number of iterations or if the residual has
a smaller magnitude etc. Most commonly used greedy algorithms includes, Orthogonal Matching Pursuit (OMP)
[9], one among the earliest method of sparse signal recovery. For several applications OMP may not offer ample
performance, which led to the development of improved pursuit methods, which works better and yields substantially
optimal results. Several greedy based algorithms have been listed in the literature to address a few variations of OMP
i.e regularized OMP [10], Residual Minimization Pursuit [11] and Stage-wise OMP [12], CoSaMP [13], Subspace
Pursuit [14]. The attractive feature of this greedy pursuits is the low implementation costs and the speed of recovery
but if the sparsity is low, implementation cost turns to be higher.

2.1.1. Orthogonal Matching Pursuit (OMP)

OMP [9] ia an improved method over Matching pursuit (MP). In MP we need to approximate, the best matching
projection of the data from the dictionary or to be precise from the sensing matrix in order to approximate the sparse
solution. The process of orthogonality is an extra addition in OMP. In each iteration of the algorithm, it ensures the
orthogonal direction of projection.
2246 Deepa Puneeth et al. / Procedia Computer Science 171 (2020) 2242–2251
Deepa Puneeth / Procedia Computer Science 00 (2019) 000–000 5

2.1.2. Residual Minimization pursuit (RMP)

In RMP, the algorithm considers, minimization of the residual, rather than correlation maximization as it is done
in OMP. OMP algorithm first finds the columns of the measurement matrix, which is highly correlated with residual
signal to estimate one active element at each iteration. In paper [11] target localization problem is considered analo-
gous to sparse recovery problem. In order to achieve the same, they used RMP algorithm, which provides a suitable
platform for target localization alternate to OMP based method. When the sensing matrix has orthonormal rows, RMP
converges to OMP.

2.1.3. St-OMP
The above mentioned sparse recovery algorithms OMP and RMP, takes ’k’ iterations (sparsity of the signal). St-
OMP [12] runs faster than, OMP and RMP. In OMP based method only one coefficient enters at each iteration,
but in St-OMP many coefficients can enter. Rather than selecting the largest component of the vector selection is
done, based on a threshold and several coefficients are selected, if it is above specified threshold. Further residual is
calculated based on least square problem. The algorithm runs only for fixed number of stages. St-OMP is faster than
OMP, as it can take many coefficients rather than one as in OMP. St-OMP converges to OMP, when the threshold is
set such a way that only one term enters in each stage.

2.1.4. Subspace Pursuit (SP)

The algorithm SP [14] is similar to St-OMP, in the sense many coefficients can enter the algorithm, rather than one
as in OMP. But the strategy of finding the k columns is different, than StOMP. Initially it chooses k columns from the
sensing matrix, then it keeps refining the subset of k columns at every iteration. In the current iteration of estimating
x (sparse signal), if y does not exist in the estimated subspace then the estimate list of k columns is updated, by
holding the reliable candidates and releasing the unreliable ones. The process continues by adding the same number
of candidates. In contrast to the other revised versions of OMP, in SP there is a simple procedure of re-evaluation of
the reliability of the candidates at every iteration.

3. Multi path Routing using adjustable Transmission Power levels

Reliability can be achieved through multi-path routing. The data can be routed through more than one path which
improves the accuracy of data reception at the BS.

3.1. Impact of using adjustable transmit power levels

The transmission power levels directly influence, the transmission range, communication energy drain, and con-
nectivity in the sensor network. The transmission range in multi hop communication can also be considered as shorter
hops (smaller transmission range between the hops) and longer hops (larger transmission range between the hops).
There have been attempts in the past to decide which is better in multi hop communication from different perspectives
but no consensus has been reached [15]. As a result, we have used a combination of shorter and longer hops of differ-
ent transmission ranges in multi-hop communication and verified its effects on network lifetime (considering only the
communication energy drain).
Most of the time fixed smaller transmission ranges cause a bottle necking problem with less nodes near the BS to
access the BS for multi hop communication (unless more nodes are deployed near the BS that can reach the BS with
that particular transmission range). Most often the nodes near the BS die faster than other nodes because they must
carry all the data traffic from the rest of the nodes to reach the BS, thus leading to early network partition. From the
perspective of network lifetime, the biggest advantage of nodes capable of adjusting to different transmission power
levels is that a large number of nodes can reach the BS and can help to achieve a better overall network lifetime.

3.2. Cluster based Node-disjoint Multi path routing

In multi hop communication, the importance of nodes near the BS often determines the factors affecting it. The
main factors would be the deployment strategies of sensor nodes and the BS or the BSs.[16]. There is a significant
 Deepa Puneeth et al. / Procedia Computer Science 171 (2020) 2242–2251 2247
6 Deepa Puneeth / Procedia Computer Science 00 (2019) 000–000

amount of literature available on the deployment strategies in sensor networks and forms another area of research
direction itself [17]. Routing in sensor systems can be extensively grouped into two significant classifications: BS
selecting and initiating the routing and the other is distributed [18]. We opted for the BS selecting the routes and
initiating the multi-path. This was due to possibility of many applications in the field of sensor networks where
the deployment of sensor nodes is predetermined such as secure monitoring of protected areas or disputed areas or
even surveillance applications. There could also be many applications which require nodes to be location aware, be
equipped with GPS units and convey their location information to the BS. We are targeting only those applications of
sensor networks where node deployment positions are either predetermined or determined after being deployed to the
BS. Therefore, the BS selects and conveys the information regarding routing process to all nodes.
The BS selects proper intra transmission and inter transmission power levels associated with each multi path. The
transmit power which is able to provide the minimum node degree will be decided as intra transmit power level, and
above intra transmit are the inter transmit power levels. Selection of CHs are selected based on the highest degree of
a node among the nodes available in the intra cluster area. During the process of CH election there may be few nodes
which does not plunge under the territory of any CH, those nodes need to have higher power level in order to get
connected with the nearest CH which is termed as Hnode . Selected CH has to bear the data traffic of the network, thus
there is a need to account the number of hops related with individual elected CH as defined in [18] and given in 6.

NCH
I=1HI
HInd = (6)
NCH

Where HInd is the counter for Average number of Hops of the network.
NCH is the total number of nodes selected as CHs in that particular combination of power levels
HI accounts for number of hops required by individual CH to arrive at the BS.
Average end to end communication energy drain is as given in 7

Edrain = HInd × ICH pd × DT dur (7)

Where E(drain) in milli joules accounts for communication energy drain.

ICH pd is the inter CH communication drain in milli watts.
DTdur is the data packet transmission duration in seconds.
HInd is the counter for Average number of Hops of the network.

The choice of proper inter transmit power levels with elected intra transmit power levels for different paths in node
disjoint multi-path routing is achieved by minimizing the energy drain of the network.

Minimize(Edrain ) (8)

Among the available inter transmission power levels, the power level that satisfies the above objective function is
selected as the inter transmission power level and its corresponding path will be one of the node-disjoint path in the
multi path routing.

3.2.1. Experimental results of CS recovery based on Greedy Algorithms.

To evaluate the performance, we have considered the data N=512 and the measurements varying in steps for dif-
ferent sparsity levels. The measurement matrix is independent and identically distributed (i.i.d) Gaussian matrix. The
2248 Deepa Puneeth et al. / Procedia Computer Science 171 (2020) 2242–2251
Deepa Puneeth / Procedia Computer Science 00 (2019) 000–000 7

OMP Recovery relative error in l2-norm

Percentage of Signals Recovered using Omp 5
100 s=5
s=5 s = 12
Percentage of Signals Recovered (%)

s = 12 s = 20
s = 20 4
s = 25
80 s = 25 s = 30
s = 30

Relative error
3
60

2
40

1
20

0
0 0 50 100 150 200 250
0 50 100 150 200 250
Compressed vector length(m)
Compressed vector length(m)

Fig. 3: CS recovery using OMP.

Fig. 4: Relative error using OMP.
Percentage of Signals Recovered using Rmp RMP recovery relative error in l2-norm
100 5
s=8 s=8
Percentage of Signals Recovered (%)

s = 12 s = 12
s = 20 s = 20
80 4
s = 25 s = 25
s = 30 s = 30
Relative error

60 3

40 2

20 1

0 0
0 50 100 150 200 250 0 50 100 150 200 250
Compressed vector length(m) Compressed vector length(m)

Fig. 5: CS recovery using RMP. Fig. 6: Relative error using RMP.

Percentage of Signals Recovered using Stomp Stomp recovery relative error in l2-norm
100 5
s=8 s=8
Percentage of Signals Recovered (%)

s = 12 s = 12
s = 20 s = 20
80 4
s = 25 s = 25
s = 30 s = 30
Relative error

60 3

40 2

20 1

0 0
0 50 100 150 200 250 0 50 100 150 200 250
Compressed vector length(m) Compressed vector length(m)

Fig. 7: CS recovery using StOMP. Fig. 8: Relative error using StOMP.

 Deepa Puneeth et al. / Procedia Computer Science 171 (2020) 2242–2251 2249
8 Deepa Puneeth / Procedia Computer Science 00 (2019) 000–000

platform for simulation is MATLAB. Simulations are done based on greedy algorithms (OMP, RMP, StOMP, SP), and
the plots depicts percentage of recovery and the error. Relative error is calculated as below 9. The tolerance limit for
perfect recovery is 1e-6. The figures below 4-10 shows recovery and error rate of the greedy methods(OMP, RMP,
StOMP, SP). The performance of greedy methods, for varying sparsity(k) and measurements(M) can be viewed in
Fig.11. When the sparsity is less, almost all the methods convergence rate is similar except OMP. The time required to
recoup the data depends on the type of iteration, the algorithm performs. When the sparsity is less all greedy methods
convergence rate is almost same, but as k increases, we can make a clear distinction of the convergence rate of various
greedy methods.
Simulations are carried out by considering the data set from Intel lab [19] which comprises of temperature and hu-
midity data set. The signals are not sparse when they are acquired, we used DCT in order to sparsify the data.

Percentage of Signals Recovered using SP SP recovery relative error in l2-norm

100 10
s=8 s=8
Percentage of Signals Recovered (%)

s = 12 s = 12
s = 20 s = 20
80 8
s = 25 s = 25
s = 30 s = 30

Relative error
60 6

40 4

20 2

0 0
0 50 100 150 200 250 0 50 100 150 200 250
Compressed vector length(m) Compressed vector length(m)

Fig. 9: CS recovery using SP. Fig. 10: Relative error using SP.
Percentage of Signals Recovered Vs Number of measurements greedy based recovery relative error in l2-norm
100 5
Percentage of Signals Recovered (%)

80 4

OMP @ s = 16
Relative error

60 OMP @ s = 16 3
RMP @ s = 16 RMP @ s = 16
SP @ s = 16 SP @ s = 16
StOMP @ s = 16 StOMP @ s = 16
OMP @ s = 32
40 OMP @ s = 32 2
RMP @ s = 32 RMP @ s = 32
SP @ s = 32 SP @ s = 32
StOMP @ s = 32 StOMP @ s = 32
20 OMP @ s = 48 1 OMP @ s = 48
RMP @ s = 48 RMP @ s = 48
SP @ s = 48 SP @ s = 48
StOMP @ s = 48 StOMP @ s = 48
0 0
0 50 100 150 200 250 0 50 100 150 200 250
Compressed vector length(m) Compressed vector length(m)

Fig. 11: CS recovery using greedy methods. Fig. 12: Relative error using greedy methods.

As shown in Fig.13 for temperature and humidity, SNR of all the methods are almost same(nearly 40db), but
it differs with the rate of convergence. In order to meet the same SNR the iterations taken by the greedy methods
are different. WSN generate huge amount of data, it has to be processed and propagated to the BS in order to take
necessary action. Further certain applications need speedy processing and few may not. Thus, based on the nature of
2250 Deepa Puneeth et al. / Procedia Computer Science 171 (2020) 2242–2251
Deepa Puneeth / Procedia Computer Science 00 (2019) 000–000 9

Temperature data Humidity data

(a) (b)

Fig. 13: CS recovery using greedy methods for temperature and humidity data with N=8500 and M=2500

application suitable greedy algorithms can be selected.

x − x̂2
relativeerror = (9)
x2

If we consider CS as a data aggregation scheme along with the multi path routing strategy as described previously
network lifetime can be enhanced. With CS the number of packets which has to be transmitted can significantly
reduce. Since the nodes are deployed randomly the data emitting will be correlated either spatially or temporally. If
the source nodes use lower intra transmission power levels to get connected to one hop CH, then we can significantly
reduce the power consumption of the network.

4. Conclusion

In order to improve network lifetime under multi-path strategy, we propose routing strategy with tuneable power
levels along with data aggregation using CS. In order to achieve better network lifetime under node-disjoint multi
path routing, sensor nodes adjust themselves to the best possible transmission power levels as per requirements. Since
CS is used along with an efficient routing protocol, it is possible to reconstruct lost packets which gives an added
advantage. CS reconstruction using Greedy algorithm is found to be better than other reconstruction methods, in
terms of computation time and complexity. Simulations prove that, Recovery of synthetic and real data using greedy
methods diminishes data stream into the network and in turn enhances the lifetime of the network.

References

[1] Y. Liu, W. Zhao, L. Zhu, B. Ci, S. Chen, The method of data aggregation for wireless sensor networks based on leach-cs, in: China Conference
on Wireless Sensor Networks, Springer, 2014, pp. 489–498.
[2] L. Xiang, J. Luo, C. Rosenberg, Compressed data aggregation: Energy-efficient and high-fidelity data collection, IEEE/ACM Transactions on
Networking (TON) 21 (6) (2013) 1722–1735.
[3] G. Cao, P. Jung, S. Stanczak, F. Yu, Data aggregation and recovery in wireless sensor networks using compressed sensing., IJSNet 22 (4)
(2016) 209–219.
[4] S. B. Meenakshi, A survey of compressive sensing based greedy pursuit reconstruction algorithms, International Journal of Image, Graphics
and Signal Processing 7 (10) (2015) 1–10.
[5] Z. Zhang, Y. Xu, J. Yang, X. Li, D. Zhang, A survey of sparse representation: algorithms and applications, IEEE access 3 (2015) 490–530.
[6] L. Du, R. Wang, W. Wan, X. Q. Yu, S. Yu, Analysis on greedy reconstruction algorithms based on compressed sensing, in: 2012 International
Conference on Audio, Language and Image Processing, IEEE, 2012, pp. 783–789.
 Deepa Puneeth et al. / Procedia Computer Science 171 (2020) 2242–2251 2251
10 Deepa Puneeth / Procedia Computer Science 00 (2019) 000–000

[7] J. Emmanuel, Candes and terence tao. decoding by linear programming, IEEE Transactions on Information Theory 51 (12) (2005) 4203–4215.
[8] D. L. Donoho, Compressed sensing, IEEE Transactions on information theory 52 (4) (2006) 1289–1306.
[9] J. A. Tropp, A. C. Gilbert, Signal recovery from random measurements via orthogonal matching pursuits, IEEE Transactions on information
theory 53 (12) (2007) 4655–4666.
[10] D. Needell, R. Vershynin, Uniform uncertainty principle and signal recovery via regularized orthogonal matching pursuit, Foundations of
computational mathematics 9 (3) (2009) 317–334.
[11] H. Song, G. Wang, Y. Zhan, Sparse target localization in rf sensor networks using compressed sensing, in: 2013 25th Chinese Control and
Decision Conference (CCDC), IEEE, 2013, pp. 3999–4004.
[12] D. L. Donoho, Y. Tsaig, I. Drori, J.-L. Starck, Sparse solution of underdetermined systems of linear equations by stagewise orthogonal matching
pursuit, IEEE Transactions on Information Theory 58 (2) (2012) 1094–1121.
[13] D. Needell, J. A. Tropp, Cosamp: Iterative signal recovery from incomplete and inaccurate samples, Applied and computational harmonic
analysis 26 (3) (2009) 301–321.
[14] W. Dai, O. Milenkovic, Subspace pursuit for compressive sensing signal reconstruction, IEEE transactions on Information Theory 55 (5) (2009)
2230–2249.
[15] M. Haenggi, D. Puccinelli, Routing in ad hoc networks: a case for long hops, IEEE Communications Magazine 43 (10) (2005) 93–101.
[16] W. B. Heinzelman, A. P. Chandrakasan, H. Balakrishnan, et al., An application-specific protocol architecture for wireless microsensor networks,
IEEE Transactions on wireless communications 1 (4) (2002) 660–670.
[17] X. Xing, D. Xie, G. Wang, Energy-balanced data gathering and aggregating in wsns: a compressed sensing scheme, International Journal of
Distributed Sensor Networks 11 (10) (2015) 585191.
[18] J. D. Nishant, K. Shivaprakasha, M. Kulkarni, Multi threshold adaptive range clustering (m-trac) algorithm for energy balancing in wireless
sensor networks, in: 2012 Ninth International Conference on Wireless and Optical Communications Networks (WOCN), IEEE, 2012, pp. 1–6.
[19] F. Koushanfar, N. Taft, M. Potkonjak, Sleeping coordination for comprehensive sensing using isotonic regression and domatic partitions (2006).