Version of Record: https://www.sciencedirect.

com/science/article/pii/S1570870516302359
79d5c1c91c29e199fb8b4753a2cbc124

Localization in Wireless Sensor Networks:

A Dempster-Shafer Evidence Theoretical Approach

Colin Elkin , Rajika Kumarasiri , Danda B. Rawat , Vijay Devabhaktuni

a a b,∗ a

a Department of Electrical Engineering and Computer Science, The University of Toledo,

Toledo OH 43606, USA.
b Department of Electrical Engineering and Computer Science, Howard University,
Washington, DC 20059, USA.

Abstract

This paper proposes a data fusion technique aimed at achieving highly ac-
curate localization in a wireless sensor network with low computational cost.
This is accomplished by fusing multiple types of sensor measurement data
including received signal strength and angle of arrival. The proposed method
incorporates a powerful data fusion technique, one that has never before been
used in low cost localization of a stationary node, known as Dempster-Shafer
Evidence Theory. Many useful functions of this theory, including sampling,
aggregation, and plausibility, are integrated into the localization method.
From there, the algorithm determines whether a set of given measurements
belong to a particular county. Motivated by the exible nature of Dempster-
Shafer Theory, a multitude of network setups and combinations of available
measurement features are tested to verify the performance of the proposed
method. Performance of the proposed approach is evaluated using numerical
results obtained from extensive simulations. When compared with the re-
sults of existing approaches in similarly constructed scenarios, the proposed
localization technique achieves up to 98% accuracy in less than a tenth of
the run-time required under presently established algorithms.
Keywords: Wireless Sensor Networks, Localization, Dempster-Shafer
Theory
∗
Corresponding Author Ph: +1(202)806-2209 and E-mail: [email protected]

Preprint submitted to Ad Hoc Networks September 8, 2016

© 2019. This manuscript version is made available under the Elsevier user license
https://www.elsevier.com/open-access/userlicense/1.0/
1. Introduction

Wireless sensor networks have been of great interest in today's world due
to their wide variety of practical applications and purposes. Practicality can
range from long distance navigational systems and localized contact points
for emergency services to industrial detection of carbon monoxide (Yang,
Zhou, Lv, Wei, and Song, 2015) and community detection of social clusters
(Li, Qin, and Song, 2016). In addition to emergencies and navigation, typical
uses of these networks can include tracking, security, and information (Lloret,
Tomas, Garcia, and Canovas, 2011).
While the demand for and scope of wireless sensor networks (or WSNs)
continue to grow, the need for identifying a node's location within such a
network becomes increasingly vital. However, achieving localization with a
steady balance of minimal power, low overall cost, and high accuracy (Mao,
Fidan, and Anderson, 2007) remains one of the greatest challenges in the
area of WSNs.
1.1. Localization Techniques
A brief overview of the most common unimodal localization techniques of
both low cost and high cost categories is presented in Fig. 1. In this scenario,
cost refers to the amount for signicant computational resources required to
achieve accurate localization. Some of the most common modern localiza-
tion techniques are well-known even outside of scientic research areas due to
2
 BT

GPS
High cost
INS

WLAN
Localization
algorithms
AOA

RSS
Low cost
TOA

TDOA

Figure 1: Classication of localization techniques.

their popularity in modern consumer electronics. These include global posi-

tioning systems (GPS), wireless local area networks (WLAN), and bluetooth
(BT). GPS is perhaps the most widely known method due to its frequent use
in navigational assistance through standalone devices and through embedded
navigational capabilities in smartphones and tablets. Bluetooth is primarily
appealing for short range solutions and thus can be ideal specically for local
positioning within a WSN (Shen, Chen, and Lu, 2008). WLAN, or Wi-Fi, is
another short range solution whose signicance lies within its high localiza-
tion accuracy throughout indoor WSN environments (Shen, Chen, and Lu,
2008). Inertial Navigation Systems (INS) are not quite as common as the
other three methods but are noteworthy due to their high potential for seam-
less integration with GPS systems (Bhatt, Aggarwal, Devabhaktuni, and
Bhattacharya, 2012). All four of these techniques share the common charac-
teristic of requiring a high computational cost to achieve a desirable level of
accuracy. For instance, a GPS-based solution requires sucient resources in
every node to receive a signal from a positioning satellite, resulting in long
3
 Monitor 1 (x1,y1)

Node Location (x,y)

Monitor 2 (x2,y2) Monitor 3 (x3,y3)

Figure 2: Depiction of RSS Trilateration in a three-monitor WSN Setup.

connection times, external signal outages, and shorter battery life (Bulusu,
Heidemann, and Estrin, 2000). Bluetooth and WLAN are also prone to faster
battery drain, as any smartphone or laptop user is quick to observe. One
potential resolution to this resource issue involves restricting measurement
capabilities to a small handful of nodes while reducing all other nodes to a
bare minimum level of communication capability. These specialized, higher
powered nodes are known as cluster heads, or anchor nodes (Mao, Fidan,
and Anderson, 2007). In order to utilize the remaining nodes, however, a
vital further step is to apply lower cost methods of a less computational and
more mathematical nature in order to achieve the full potential of eective
localization.
Many common low cost techniques include received signal strength (RSS),
angle of arrival (AOA), time of arrival (TOA), time dierence of arrival
(TDOA), and hop count. In the context of this research, RSS is dened
as the electric eld at the receiving node divided by the distance between
said node and the transmitting antenna. In outdoor environments, RSS can
be considered as a simple mathematical method of distance calculation, free
from the restrictions of walls, multipath eects, and humidity (Garcia, Mar-
tinez, Tomas, and Lloret, 2007; Garcia, Tomas, Boronat, and Lloret, 2009)
that are typically associated with indoor WSN environments. Due to RSS's
inverse relationship with squared distance (Xu, Lin, Lang, Zhang, and Wang,
4
 Node Location (x,y)

Monitor 1 (x1,y1) α1 Monitor 2 (x2,y2) α2

Figure 3: Depiction of AOA-based localization.

2010) in outdoor WSNs, highly accurate locations can be determined with

three or more RSS measurements from three unique points. This is accom-
plished by trilateration (Mazuelas et al, 2009), as can be seen in Fig. 2.
Trilateration is particularly benecial when three or more anchor nodes are
available but is ineective in WSN setups in which only two measurements
are available. That is, under two RSS measurements, only bilateration can
achieved, resulting in two possible intersecting points and thus a 50% chance
of achieving accurate localization. AOA is the arrival angle of the emitted
source signal observed at an anchor node (as shown in Fig. 3) and is a mea-
surement dependent on TDOA. More specically, AOA is calculated using
the dierences between arrival times of a transmitted signal (Patwari et al,
2005). TDOA is dependent on time dierences between nodes (Cakir, Kaya,
and Cakir, 2014) and is another eective means of localization by process
of multilateration (Pelant and Stejskal, 2011). TOA, on the other hand, is
simply the amount of time required for a signal to move from a transmitting
node to a receiving node (Patwari et al, 2005). Like RSS, AOA provides
a 50% chance of accurate localization using intersecting angles, but it also
requires expensive antenna arrays (Patwari et al, 2005). In addition, TOA
faces the drawback of having a dependence on source signal transmission
time to measure propogation time between sensors (Shen, Ding, Dasgupta,
and Zhao, 2008). TDOA measurements, on the other hand, strengthen noise,
5
leading to hindered performance (Shen, Ding, Dasgupta, and Zhao, 2008).
In spite of the aforementioned drawbacks, many of these techniques, such
as RSS and AOA, function adequately on their own. However, it has been
shown that combining multiple, varied types of data can enhance results and
understanding, eectively providing the best of all worlds in regards to each
data type's advantages (Xing et al, 2009; Tseng, Kuo, Lee, and Huang, 2004).
However, the very act of combining measurement typesdata fusionposes
a unique challenge.
1.2. Dempster-Shafer Theory and Its Existing Applications
Decision fusion is a form of data fusion in which the desired output is
a decision rather than a quantitative value. Dempster-Shafer (DS) The-
ory is a decision fusion technique that is based on Bayesian combinational
probability but undertakes a drastically dierent foundation with regards to
combinational factors (Nguyen and Walker, 1993). While Bayesian statistics
rely primarily upon combinations of internal probability factors within a sys-
tem (Chiodo and Mazzanti, 2006), typically through the use of propositions
or random variables, DS Theory is contingent upon external evidence fac-
tors (Limbourg, Savic, Petersen, and Kochs, 2007; Lipeng, Juan, and Laibin,
2011). Each factor consists of (i) a range of data in which the desired output
value can be found as well as (ii) a probability of condence that the data
range is in fact reliable. DS Theory is meant to model information that,
under more traditional Bayesian methods, would be considered incomplete
(Nguyen and Walker, 1993). Some localization problems have incorporated
DS Theory in the past, such as robot localization (Clerentin, 2000) and in-
door localization of a mobile user (Kasebzadeh, 2014). Prior to the validation
of this research, however, Dempster-Shafer Theory has never before been ap-
plied to low cost localization of a stationary node in a large outdoor WSN.
1.3. Research Motivations
The overall objective of this research is to develop an algorithm for local-
ization in a wireless sensor network that is both highly accurate and of low
computational cost. Time-critical localization of WSNs is a topic of increas-
ing concern in today's world, as a variety of civil and military applications
are heavily dependent on these two properties. For instance, many college
campuses and some urban areas have a network of emergency phones scat-
tered across town so that individuals in distress can reach emergency services
(Meyer, 2012). Using eective localization techniques such as that proposed
6
in this paper, emergency services can dispatch police, re, or medical services
to a caller's location at a signicantly improved response time. In addition,
the wireless aspect of these networks would ease any infrastructure-based
complexities that currently exist with current wired phone stations.
1.4. Overview of Subsequent Sections
The remainder of this paper is organized as follows. Section 2 expands
upon many of the previous subsections by providing an extensive theoreti-
cal background of DS Theory and WSN localization techniques. Section 3
provides an overview of how the research study was conducted, including
generation of the relevant WSN data, the properties and parameters of the
sensor network, and the determination of the method's accuracy and per-
formance. Section 4 presents the results of the experiment in a variety of
scenarios, a comparison of said results to those of existing localization tech-
niques, and a brief discussion of the impact thereof. Lastly, conclusions are
presented in Section 5.
2. Mathematical Background and Formulation

This section establishes three vital concepts that comprise the core of this
research study, provided in three subsections. The rst is Dempster-Shafer
Theory, or more specically the concepts and properties thereof. The second
contains the types of sensor node measurements needed for the localization
technique. Finally, the third subsection describes the technique itself, which
formulates the relationship between the two concepts to form a comprehen-
sive location-estimating algorithm.
2.1. Dempster-Shafer Theory
From a theoretical context, Dempster-Shafer Theory can be best de-
scribed as a dynamic generalization of Bayesian probability theory, form-
ing a divergence based on the concept of ignorance (Bloch, 1996). Typical
Bayesian approaches assume an ignorance quantity of zero, or equivalently
100% condence, for any probabilistic value, which is usually appropriate
when a probability factor is internal. Because DS Theory is based upon
external factors, however, this property can no longer by unconditionally as-
sumed. Hence, an array of evidence factors is used to determine all possible
outcomes in a given scenario (e.g. the distance between a sensor node and
an anchor node in a WSN). The nite set of all possible mutually exclusive
7
values is known as the frame of discernment and is denoted by Θ. From
there, the union of all subsets is given by the power set 2 . Θ

One of the most signicant fundamental concepts of DS Theory is the

basic probability assignment (BPA) function, which is essentially the equiv-
alent of the random variable in Bayesian probability theory. It is also known
as the belief variable (Auer, Luther, Rebner, and Limbourg, 2010) and is
denoted as m. Suppose A , . . . , A are all possible sets within a frame of
discernment, noting that A ∈ 2 . Then
1 n
Θ
i

∑
(1)
n
m : 2 → [0, 1],
Θ
m(A ) = 1, m(∅) = 0,
i
i=1

where ∅ denotes the empty set. Each set A takes the form of the interval
([x, x̄]), denoting the lower and upper bounds of an evidence factor, respec-
i

tively
¯
(Auer, Luther, Rebner, and Limbourg, 2010). Each BPA m consists
of three vital properties: a lower bound, an upper bound, and a degree of
condence µ that lies between the values of 0 and 1 (University of Duisenberg-
Essen, 2012). Thus, ignorance can be dened as simply 1 − µ.
2.1.1. Belief and Plausibility
Belief and Plausibility are two of the most vital core functions of an
aggregate BPA, as they represent upper and lower bounds in probability,
respectively. In a frame of discernment, they are dened as
∑
Bel(B) = m(A )
Ai ⊂B
i (2)
∑
P l(B) = 1 − m(A ),
Ai ∩B=∅
i (3)
where m(A ) denotes the BPA mass function for a set A .
i i

2.1.2. Set Representation

A BPA contains three vital properties: a lower bound ¯x, an upper bound
x̄, and a degree of condence µ. Hence, representation of a BPA as a set of
these three properties becomes a highly convenient organizational method.
In such a case, the resultant BPA m would be dened as
m = {x, x̄, µ}.
¯
(4)
Likewise, because calculation of belief and plausibility requires multiple BPAs,
or the CDF of an aggregate BPA, representation of the individual assignments
8
as a set of sets is also benecial. Suppose m , m , . . . , m represent all BPAs
u u u

in Θ such that   
{x , x̄ , µ }
u

1

u u
2

u
n

m 
 1 
 

   ¯1 1 1  
 mu2 
  {xu2 , x̄u2 , µu2 }
 
m=


..


=


¯


.. (5)

   
 u    u u u  
mn {xn , x̄n , µn }
and . The latter property is of particular importance, as
¯
x̄u1 < x̄u2 < · · · < x̄un
this sorted order of upper bounds can be used to dene belief from a more
specic, set-oriented context. Let the superscript u indicate that contents of
m are arranged in order of increasing upper bound. Also, recall that µ is the
degree of condence in each BPA and thus can be dened as µ , µ , . . . , µ . u u u

Then belief can be represented as 1 2 n

 u u 

 {x̄1 , µ1,new } 

 


 
{x̄ 2,new 
}
u u
 , µ 
 2

Bel = {x̄3 , µ3,new } ,
u u
(6)










..

 u u 

 
{x̄n , µn,new }

where
µu1,new = µu1 (7)
and ∑
(8)
n−1
µun,new = µui,new .
i=1

Now suppose m is rearranged and m , m , . . . , m

u
1
u
2
u
n are renamed such that
 l l l l 

 m1 

 {x1 , x̄1 , µ1 }
 

  ¯ 
 ml2 

 {xl2 , x̄l2 , µl2 }
 
m=

 

= ..


¯


, .. (9)

   
 l    l l l  
mn {xn , x̄n , µn }
¯
where ¯x < ¯x < · · · < ¯x and the superscript l indicates that contents of m
l l l

are now arranged in order of increasing lower bound. Hence, all conditions
1 2 n

9
are available to dene plausibility as
 l l 

 {x̄ , µ }
1,new 


1



 {x̄2 , µ2,new }
l l


 

{x̄ l
, µ l
} (10)
..
Pl = 3 3,new ,

 


 


 


 l l 

{x̄n , µn,new }
where
µl1,new = µl1 (11)
and ∑
(12)
n−1
µln,new = µli,new .

In other words, the top and bottom rows represent the least and most
i=1

believable outcomes, respectively. The latter is the row of greatest impor-

tance, where x̄ is the value of the most believable outcome and µ is
u u

the belief probability, which in this case is always 100% or 1, since it is the
n n,new

sum of all condence values. Similarly, the top and bottom rows of the be-
lief set respectively represent the least and most plausible outcomes. Also
likewise to belief, the most believable row is of most signicance, as ¯x sig-
l

nies the most plausible outcome and µ is another 100% probability.

l
n

Due to these powerful methods of outcome prediction, extensive analysis of

n,new

belief and plausibility serve as the initial foundations and formations of this
research.
2.2. Categories of Sensor Node Measurements
The proposed localization technique is dependent on a multitude of sensor
measurements, namely RSS, AOA, and standby. Before these values can be
mathematically dened, however, the distance between any sensor node and
any anchor node must also be evident. Suppose the coordinate [x , y ] refers
to the position of a sensor node and [x , y ] is the position of a monitor, or
n n

anchor node. Then, the distance d can be dened as

m m

dactual
√
= (x − x ) + (y − y ),
2
m
2
n
2
m
2
n (13)
where all aforementioned variables are measured in meters. For RSS cal-
culation in outdoor WSN environments, multiple formulas exist for exact
10
calculation but share the common property of having an inverse proportion-
ality with squared distance (Xu, Lin, Lang, Zhang, and Wang, 2010; Patwari
et al, 2005). Hence, in the context of this research, RSS is simply dened as
the inverse of distance squared, or RSS = 1/d . 2

Due to the generic nature of this interpretation, no specic unit is needed,

actual

as long as the generic unit contains the factor of m . AOA, however, is

−1

dened in a more specic context as


 −1 yn −ym
tan ( ) if y > y and x > x
) + 180 if y ̸= y and x < x (14)
xn −xm n m n m
AOA = tan ( −1 yn −ym

) + 360 if y < y and x > x



xn −xm
−1 yn −ym
n m n m
tan ( xn −xm n m n m

with resultant units in degrees. Standby is a unique type of distance mea-

surement in which a sensor node with this feature enabled is detected by
a dedicated standby node, which measures the distance thereof by inverse
square root RSS calculation. This is enabled primarily in network terrains
with smaller area, as each sensor node's signal strength must be within range
of the standby node. Let [x , y ] be the location of the standby node. Then
the standby distance is dened mathematically as
sb sb

√
d = (x − x ) + (y − y ).
SB
2
sb
2
n
2
sb
2
n (15)
2.3. Proposed Representation of Distance Range as BPA
In order to fuse any of the varied measurement types together, each mea-
surement must be converted into a common value. Such a value in this case
is best represented as the distance from a sensor node to an anchor node,
as determined by inverse square root RSS calculation. Due to this method
of distance calculation, the experimental setup will require RSS to be en-
abled at all times. If a distance or range of distances can be found for all
available cluster heads, then a node's location can be estimated by an en-
hanced process of trilateration, in which enhancement is based on the eects
of data fusion. The challenge associated with reaching such a goal becomes
the task of estimating such a distance range for each monitor. Thus, the
need for eective organization and evaluation of optimal distance values is
where Dempster-Shafer Theory becomes a mechanism of great interest.
Because external evidence factors are the primary means by which DS-
based estimation can occur, it becomes necessary to obtain distances and
measurement sets that are independent of those given as functional inputs.
11
This distinction will be explained at greater length in Subsection 3.2 by means
of test data and full data sets. Because such data is external, the amounts and
ranges of measurements must be reduced based on feature-specic criteria.
Suppose these range values, r, are dened as
r = {r ,r
RSS , r },
AOA SB (16)
where r > 1, 0 < r < 1, and r = −1, the latter value indicating that
there is no range for standby. If the inverse of a given feature measurement
RSS AOA SB

is less than the range r and the range is not −1, then the lower and upper
limits of such ranges are dened as |0, 1/r |. Otherwise, if the value is greater
i

than or equal to r , then the limits are | , s |, where s is the particular

i
si

feature measurement and i is the feature type. A range of −1, on the other
i ri si +1 i i

hand, matching feature measurements between both data sets. The next step
is to nd external distance values from external feature measurements that
lie within the corresponding range. After ltering of all such distance values,
the minimum and maximum of the remaining values become the lower and
upper bounds of the resultant BPA. Because all BPAs will aggregate in a
later step, the condence value can simply be 1 for the time being. Thus,
based on Equation 4, we have
m = {d , d , 1},
i min max (17)
where d denotes the minimum external distance value and d denotes
the maximum.
min max

Once a BPA is formed for each active feature, aggregation can then occur.
At certain times, manually adding new BPAs can be an exhaustive process,
in which case sampling from a distribution becomes an ideal solution. This
process, also known as discretization, involves generating n discrete samples
based on a lesser amount of initialized BPAs, which collectively form a cu-
mulative density function (CDF) of BPA structures (Tonon, 2004). Thus,
to smoothen the transition between BPAs, sampling each m by a consis-
tent function handle, such as inverse normalization, is recommended. Once
i

sampling has been applied, the aggregate BPAs can now be considered as
an adimensional, unit neutral entity, as required for mass assignments. Now
each BPA closely ts the parameters of Equation 4 and is written as
m = {x, x̄,
i
¯
1
fn
}, (18)
12
where denotes an evenly distributed condence probability in terms of f
1

active features and n BPAs. The resultant aggregate BPA for n samples thus
fn

becomes
λ = {λ ,λRSS , λ },AOA SB (19)
where
     

 m 1,RSS 
 
 m1,AOA 
 
 m1,SB 


     
 m2,RSS  
 m2,AOA 
 
 m2,SB 

λRSS =


.. 

, λAOA =


.. 

, λSB =


.. 

(20)

 
 
 
 
 

     
mn,RSS mn,AOA mn,SB

This is under the assumption that all three features are enabled. If one or
more features are unavailable, then their corresponding λ values are sim-
ply omitted. With all aggregation intact, the next step is the actual data
prediction process.
2.3.1. Formulation of Proposed Algorithm
Because Dempster-Shafer Theory is a new concept of keen interest in the
area of low cost WSN localization, a clear relation between actual distance
and believed or plausible distance was not available in any found literature.
Thus, some preliminary experimentation was needed to understand how Bel
and P l functions could potentially correlate measured values with calculated
values. The resultant analysis from belief and plausibility calculations showed
that the former method contained little to no eect in predicting a given
distance while the latter showed encouragingly high correlation. Thus, after
extensive exploration, we propose a completed high accuracy algorithm based
on plausibility, as presented in Fig. 4.
As a simple example of how the proposed method works, suppose there
exists a WSN of ve sensor nodes, two sensor nodes, a standby node, and a
fusion center, as depicted in Fig. 5 that encompasses a one square meter area
(details of these node types are explained further in Subsection 3.1). Here,
we wish to locate the sensor node located at [0.4, 0.6] meters, given only a
list of possible node locations and a set of signal measurements (RSS, AOA,
and standby, as measured by the standby node). For each anchor node, or
monitor, the technique (via the fusion center) conducts an external survey of
candidate distance values that satisfy feature-based constraints. Out of these
values, the minimum and maximum thereof encompass the bounds of a BPA.
13
Figure 4: Flowchart of proposed DS localization algorithm.

14
 1
Sensor Node
0.9 Anchor Node
Standby Node
0.8 Fusion Center
Latitudinal coordinate (m)

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Longitudinal coordinate (m)

Figure 5: WSN of 5 sensor nodes, 2 anchor nodes, standby node, and fusion center.

After the algorithm repeats this process for every feature type, it aggregates
all BPAs. Next, it executes the plausibility function on the distance BPAs
and selects the resultant distance value of greatest condence. This results
in a radius of plausibility for each monitor, as indicated by the lightly colored
circles of Fig. 5.
The end goal here is to generate a region of feasibility in which the target
node could lie. For this, the algorithm incorporates a pair of coecients,
determined by preliminary WSN conguration, which forms inner and outer
radii, as indicated by the dark circles in the gure. Finally, we inspect the
intersection of inner and outer circles to nd that only one node location
lies within the region and thus matches the given set of measurements. In
essence, this problem can be perceived as a multiple choice exam, and while
choosing from only ve nodes in this case does not seem too daunting, it is
in large scale WSNs with more than a hundred nodes that this technique can
truly ourish.

15
3. Simulation Setup

A variety of WSN simulation scenarios was created and analyzed using the
MATLAB scientic programming environment (The Mathworks, Inc., 2013).
The resultant software package includes all facets needed for experimental
simulation. This includes specic parameters and qualities of the WSN,
selectively randomized generation of data, and a complete implementation
of the technique proposed in Fig. 4. In addition, two auxiliary algorithms
designed to aid the execution of the core technique are also proposed as well
as multiple methods for calculating the algorithm's accuracy.
An overview of the default simulation parameters in the experimental
setup is as follows:
1. 100 nodes deployed throughout a rectangular region of 1000-by-1000
meters
2. Two or three anchor nodes (both options tested separately in identical
networks)
3. Anchor nodes in both two-monitor and three-monitor setups located
at [150,100] and [850,100] meters; third anchor node in three-monitor
setups located at [500,100] meters
4. Standby node placed at [500,800] meters; location of fusion center is
negligible
5. One-to-one mapping of counties and nodes
6. Range criteria of 20, 0.002, and -1 for RSS, AOA, and standby, respec-
tively
7. Setup tested for every combination of features in which RSS is active
(as RSS is needed to calculate distance from node to monitor)
8. Discretization consisting of 10 samples as an inverse normalized distri-
bution
9. Dierences of -1.5 meters or less for actual distance minus predicted
distance not considered in determining a minimum dierence between
distances
10. Distance range weights determined through iterative training (see Sub-
subsection 3.4.1) to be 1.3 and 0.95
11. Experimental results as the average of ten independent experiments in
which the WSN is regenerated upon every trial
12. Zero noise factor assumed
Details of the variables and terminologies above are explained in the pro-
ceeding subsections.
16
 1000
Sensor Node
900 Anchor Node
Standby Node
Fusion Center
800
Latitudinal coordinate (m)

700

600

500

400

300

200

100

0
0 100 200 300 400 500 600 700 800 900 1000
Longitudinal coordinate (m)

Figure 6: WSN of 100 sensor nodes, 3 anchor nodes, standby node, and fusion center.

3.1. WSN Setup

The sensor network in this simulation consists of four vital components:
the sensor nodes, the anchor nodes, the standby node, and the fusion center.
The sensor nodes comprise the vast majority of the components in this setup,
totaling 100 out of 105 nodes in a three-monitor setup. The anchor nodes,
also known as monitors or cluster heads, are placed strategically throughout
the WSN to measure the available properties of each and every sensor node,
such as RSS, AOA, and operating condition (standby). Such data is then
transferred from an anchor node to a fusion center node, which processes all
information using the algorithms and methods given in the following sub-
sections. In other words, the fusion center assumes the role of the entire lo-
calization software, minus the data generation stage. In addition, a standby
node measures the distance of any sensor nodes that are in standby mode.
The complete WSN setup is presented in Fig. 6.

17
3.1.1. Data Value Ranges
In this network setup, the origin of the coordinate system is positioned
at the lower left hand corner of the map, as indicated in Fig. 6. County
names and monitor names are given as numbers ranging from one to the
total number of counties and monitors available, respectively. All monitors
are horizontally aligned by having a common latitudinal coordinate of 100
meters as well as longitudinal coordinates from 150 to 850 meters. Ranges of
sensor node coordinates are based on the size of the WSN and can be placed
anywhere more than 30 meters from any border. Based on the criteria of
the latter two data types, distances between sensor nodes and monitors can
vary from 40 to 1360 meters. Because RSS is merely the inverse square root
of distance in the case of outdoor environments, it can range from 1/1360 2

to 1/40 . AOA can be any positive number of degrees up to 360, and the
2

operating condition, or standby, distance can range from 0 to 920 meters.

3.2. Data Generation
A multitude of sensor nodes, counties to which each node belongs, and
various measurements of nodes relative to various monitors have been written
to and read from a single comma separated value (CSV) le. This is known
as the full data set. Each sensor node encompasses multiple rows of data, in
which each row corresponds to a particular monitor observing the node and
its measurements. Each category of data, such as county number, distance,
or a type of measurement, takes the form of a column. The full data set
contains 100 samples of sensor node data for every one of the 100 counties in
the WSN, and each county is measured from three monitors, thereby totaling
30,000 rows overall.
A second CSV le known as the test data set contains the county lo-
cations and measurements that are to be tested against the full data set.
The test data contains only one sensor node per county. The location and
measurement values are dierent from any of the nodes of the corresponding
county in the full data. Hence, the test data set contains 300 total rows of
data for three monitor WSN setups and 200 rows for two monitor setups.
Each sample of data in each le contains ten columns, determined as
follows:
1. Anchor node number
2. X coordinate of anchor node
3. Y coordinate of anchor node
18
 4. County number
5. County x coordinate
6. County y coordinate
7. Distance from county to anchor node
8. RSS value
9. AOA value
10. Standby value
Worth noting is that the rst three values are intended to be treated inde-
pendently of measurement data in accordance with the objectives provided in
Subsection 3.5. In addition, the seventh column is optional to the algorithm,
as any distance value can be calculated directly from the corresponding RSS
value due to a consistent inverse proportionality between RSS and squared
distance.
3.3. IPP Toolbox
Aiding in the MATLAB implementation of DS Theory is the acquisition
of the open source IPP Toolbox (University of Duisenberg-Essen, 2012), re-
leased under the GNU General Public License (GPL). For the rst time in
WSN localization, this package has been utilized to execute all DS-specic
functions in the simulation, such as formation of BPAs, aggregation and
sampling of BPAs, and calculation of plausibility.
3.4. Algorithmic Implementation
The algorithm was programmed by following the owchart provided in
Fig. 4 and rening the technique in the form of the pseudocode provided
in Algorithm 1. The pseudocode was then converted into MATLAB-specic
syntax and thoroughly tested to ensure proper functionality and accurate
representation of the proposed algorithm.
An important distinction to note is that of the two data les presented
in Subsection 3.2. The localization algorithm uses the full data set only
for obtaining distance values for corresponding measurements that t given
range criteria in the distance collection stage. This essentially serves as an
external data set that helps to establish the lower and upper bounds of an
evidence factor, or BPA. The rest of the algorithm utilizes the internal test
data set, particularly for establishing input values for county locations and
measurement sets.
19
 1: test_data ← read measured data
2: full_data ← read potential data
3: range ← feature-specic range values
4: monitors
active features
for all do
5:
distances ← all distances from full data features satisfying range
for all do
6:
7: BPA_per_feat ← min_distance, max_distance, 1
8: apply sampling to and aggregate BPA
9:
pl_dist ← most plausible distance
end for
10:
11: min_accept, max_accept ← coecients of acceptable distance range
12: meas_dist ← measured distance from county to monitor
13: decision_per_mon ← measured distance is in acceptable range
14: dierence_per_mon ← measured - plausible distance
15:
decision ← all decisions per monitor are positive
end for
16:
17: dierence ← minimum dierence per monitor
18: return decision, dierence
Algorithm 1: Pseudocode of DS localization algorithm.

3.4.1. Determination of Acceptable Distance Values

As stated in Line 12 of Algorithm 2, a major factor in determining the de-
sired outputs in the next subsection is the formation of acceptable distance
values. Unlike articial neural networks and other conventional machine
learning techniques, DS Theory does not require an elaborate training stage.
Conversely, however, care must be taken to determine a precise distance range
to ensure optimal localization accuracy. Hence, a simple iterative approach
is taken to nd the most eective weights for the lower and upper bounds
of an acceptable distance range. The resultant approach is detailed in Algo-
rithm 3. For the sake of preserving the training-independent property of DS
Theory, and due to the highly intensive time requirement when compared to
Algorithm 2, this method is not intended to be executed upon every run of
the core algorithm. Instead, it is only to be run during the initial formation
of the WSN as well as signicant changes thereof. In other words, if a WSN
undergoes minor changes, such as the addition or removal of one anchor node
or the relocation of sensor nodes across short distances, the impact of the
20
 1: test_data ← read measured data
2: full_data ← read potential data
3: candidate upper bounds
candidate lower bounds
for all do
4:
combinations of available features
for all do
5:
generate WSN data
for all do
6:
7: run Alg 3 set number of times given bounds and features
8: decision_matrix, matching_county ← average outputs of Alg 3
9: longAcc ← long accuracy
10: shortAcc ← short accuracy
11: matchAcc ← match accuracy
12:
avgAcc ← (longAcc + shortAcc + matchAcc)
end for
∑
13: 1 1
3 no.f eat. f eatures
14: end for
15:
bestAcc ← maximum avgAcc
end for
16:
17: best_lower, best_upper ← bounds that result in best accuracy
18: combinations of available features
generate WSN data
for all do
19:
20: evaluate Alg 3 using best bounds and given features
21: end for

Algorithm 2: Pseudocode of distance range training algorithm.

acceptable distance ranges is too low to noticeably alter the accuracy of the
core algorithm. The three types of accuracy portrayed in Algorithm 2 as well
as the details of Algorithm 3 are provided in Subsection 3.6.
In machine learning techniques, the data used for training is intended
to contain many more training samples than validation samples (Krogh and
Vedelsby, 1995). Hence, in this approach, WSN data generation and Al-
gorithm 3 execution are run repeatedly and independently three times as
much as they are run in validation and training-independent stages. Due to
the high computational time and cost required for training, only a limited
number of weight combinations are considered. As will be indicated in the
experimental results in Subsection 4.1, the resultant lower and upper bound
coecients of the maximum plausibility value are ultimately chosen to be 1.3
and 0.95, respectively.

21
3.5. Desired Outputs
In order to address two distinctly dierent methods of application, the
proposed technique delivers two separate outputs that respectively address
the following questions:
1. Given a set of dierent measurement types and the location of a given
county, do the measurements belong to said county?
2. Given a set of dierent measurement types and the locations of all avail-
able counties, which county most closely matches the measurements?
3.6. Accuracy Calculation
Three dierent types of accuracy are established in calculating the eec-
tiveness of the two questions posed in the previous subsection: long accuracy,
short accuracy, and matching accuracy. The former two methods correspond
to the rst question, in which a single county location and a single list of
measurements are given as inputs and a binary decision of one or zero is the
sole output. To test long accuracy, every combination of county and mea-
surement set is run through the algorithm as the two inputs. From there,
each output is a square matrix whose side length is equal to both the num-
ber of counties and the number of measurement sets. In such a matrix, each
row corresponds to each measurement set and each column corresponds to
each county. Finally, long accuracy is calculated by comparing the resultant
matrix against an identity matrix of the same dimensions. The resultant
formula is as follows:
Accuracy = 100 ∗
cells
matching
cells
total
. (21)
Short accuracy follows a similar process except that only the matching coun-
ties and measurement sets are tested, e.g. county 1 with measurement set
1, county 2 with measurement set 2, etc. The results are recorded in an
output vector of length equal to the number of counties, rather than a two-
dimensional square matrix of the same side length. The vector is then com-
pared against a vector of ones of equal length, once again using Equation
21.
The third accuracy type is based on the second question, in which a single
list of measurements and a full list of all counties are the two inputs and the
number corresponding to the most likely county match is the output. To
test matching accuracy, every measurement set (in order of their matching
22
 1: test_data ← read measured data
2: full_data ← read potential data
3: range ← feature-specic range values
4: measurement sets
counties
for all do
5:
decision, dierence ← Alg 1 using test_data, full_data, range
for all do
6:
7:
index ← minimum dierence
end for
8:
9: matching_county ← county at index
10: end for

Algorithm 3: Pseudocode of DS testing algorithm.

county numbers) is tested against the list of counties to form a resultant

output vector of size similar to that in short accuracy. The contents of said
vector take the form of county numbers. This array is then compared against
the ideal vector, which is simply every county number in numerical order, on
the basis that the measurement sets are given in the same order. The two
matrices are then compared with the same formula as before for calculating
accuracy.
In order to thoroughly evaluate all three types of accuracy, a simple test-
ing algorithm is portrayed in the pseudocode of Algorithm 3, evaluating the
core localization algorithm in order to extract all desired distance dierence
and county match values.
4. Simulation Results and Discussion

This section provides the results for the MATLAB simulation setup and
parameters detailed in the previous section. This includes both the prelimi-
nary training-based approach intended only for initial WSN implementations
and the core training-independent approach. The latter of the two serves as
the primary focus of discussion due to its resultant combination of high ac-
curacy and low computational cost.
4.1. Results Under Distance Range Weight Training
20 possible combinations of desired values, as detailed in Subsubsection
3.4.1, served as the candidate inputs for the training-based portion of the
simulation. The results for all three accuracy types as well as the averages
23
 100 100

90 99

Short accuracy (%)

Full accuracy (%)

80 98

70 97

60 96

50 95
2 2
1 1
1.5 0.8 1.5 0.8
0.6 0.6
0.4 0.4
Upper bound weight 1 0.2 Upper bound weight 1 0.2
Lower bound weight Lower bound weight

(a) Full Accuracy (b) Short Accuracy

76.402 95
Matching accuracy (%)

Average accuracy (%)

76.4015 90

76.401 85

76.4005 80

76.4 75
2 2
1 1
1.5 0.8 1.5 0.8
0.6 0.6
0.4 0.4
Upper bound weight 1 0.2 Upper bound weight 1 0.2
Lower bound weight Lower bound weight

(c) Matching Accuracy (d) Average Accuracy

Figure 7: Plots of accuracy versus upper bound weight versus lower bound weight for
preliminary weight training.

24
thereof are presented in Fig. 7. Each result is the average of 30 indepen-
dent trials, that is, three times the amount of trials run in the proceeding
subsection. As the surface plots in the gure indicate, full accuracy reaches
optimal levels when the lower bound weight is maximized and the upper
bound weight is minimized. Short accuracy appears to be almost indepen-
dent of lower bound weight values and optimizes at the maximum upper
bound. Matching accuracy follows the most uniform distribution of accuracy
values, varying from 76.000% to 76.002%, optimizing at multiple possible
points. Eectively, this form of accuracy can be considered negligible in the
attempt to nd a single best weight combination. The average accuracy is
the primary decision factor of weight selection and optimizes upon a max-
imum lower bound of 0.95 and an upper bound of 1.3, located in a tightly
contained region in which average accuracy narrowly exceeds 90%. Hence,
the respective upper and lower weights of 1.3 and 0.95 have been veried
to produce the most accurate data and are thus used consistently under the
implementation detailed in the following subsection.
4.2. Training-Independent Results
The training-independent portion of the simulation was tested in 24 dif-
ferent scenarios resultant from four independent variables: the set of active
features, the number of monitors, the type of accuracy, and the trial-based
random generation of the WSN. These results reect the overall eective-
ness of Dempster-Shafer Theory, data fusion, and plausibility, as established
in Section 2. This simulation used two and three as the desired numbers of
monitors, as these amounts have been established minimums under unimodal
RSS and AOA based localization methods. For reliability and redundancy
purposes, the simulator regenerated the WSN prior to each of the ten exper-
imental trials for each of the above combinations of scenarios. The accuracy
results for all test cases are given in Table 1, and the runtime results are
provided in Table 2, where each value for each table is the average run time
for all trials.
4.2.1. Accuracy
As Table 1 indicates, the full accuracy test generates consistently high re-
sults in the steady area of 97.0-97.5% accuracy for all eight scenarios. Thanks
to the powerful core of the localization technique that is DS Theory, the algo-
rithm is proven to be consistently eective for both two-monitor and three-
monitor setups under all feasible combinations of active features. Results
25
 Accuracy Full (%) Short (%) Match (%) Average (%)
Monitors 2 3 2 3 2 3 2 3
RSS 97.04 97.49 97.20 98.50 56.17 66.64 83.47 87.54
RSS, SB 97.00 97.43 100.0 100.0 94.58 96.07 97.19 97.83
RSS, AOA 97.03 97.47 98.13 98.97 57.81 67.29 84.33 87.92
RSS, AOA, SB 97.01 97.44 100.0 100.0 92.52 94.21 96.51 97.22
Table 1: Simulation results for two-monitor and three-monitor WSNs.

of the short accuracy test climb even higher with a value range of 97-100%.
Under the matching accuracy test, results span a wider range of accuracy re-
sults from 56.1% to 94.6%. The average of all three accuracy types indicate a
somewhat less diverse range of 83.5% to 97.8%. Empirically, county match-
ing can be considered most eective when only RSS and operating condition
are enabled. The absence of operating condition results in the only accuracy
values lower than 94% as well as the only short accuracy values below 100%.
In other words, when a WSN is distributed across a small enough terrain that
a standby node is usable, average accuracy of over 95% is easily achieved. In
contrast, larger terrains in which standby measurements are unfeasible can
only reach average accuracy values of 80-90%.
The full accuracy test appears to be the only evaluation type that results
in a higher percentage in the absence of data fusion. While fusing multiple
types of data is intended to enhance an accuracy relative to that of a sin-
gle type of data, this test shows that full accuracy is most eective when
only RSS data is available. The other evaluation methods, however, clearly
demonstrate that fusion of multiple measurement types do in fact enhance
accuracy. Combining this observation with the fact that dierences in full
accuracies are minimal (all values for each monitor are within 0.06% of one
another) can easily infer that data fusion in this context is still an overall
success.
Also worth noting is that RSS as a sole feature forms the weakest accuracy
and may pale in comparison to traditional RSS-based localization when only
RSS is available. One should keep in mind, however, that all other remain-
ing feature combinations, that is, any combination of two or more features,
proves that fusion of multiple available types of data in this algorithm is sig-
26
 Runtime (ms) 2Mon 3Mon
Total 787478.00 1375261.00
Per Feature Set 196869.00 343815.20
Per Iteration 19.69 34.38
Per Node 0.197 0.344
Table 2: Runtime results for two-monitor and three-monitor WSNs.

nicantly more eective than any conventional unimodal technique. Thus,

RSS should be considered of lower importance, as the scope of this research
focuses primarily on the advantages and successes of data fusion.
Another successful outcome is the high accuracy involved in two-monitor
WSNs. Most established localization methods require three or more monitors
to observe a sensor node's measurements, but in this algorithm, two-monitor
WSNs can localize reliably with a less-than-0.5% decrease in accuracy. From
the standpoint of having RSS as the only enabled feature, achieving a high
accuracy with three monitors is trivial, as this can be achieved purely through
trilateration. Under two monitors, however, only bilateration is available,
meaning that if distance values calculated from RSS measurements are exact,
there is still only a 50% chance of obtaining the correct location. Thus,
the monumental incorporation of DS Theory can achieve highly accurate
localization with only two monitors just as easily as with three or more.
4.2.2. Computational Runtime
For each trial, we recorded the total runtime for an entire simulation under
all four available feature sets. From there, we averaged the calculated time
per feature set, then divided by all 100 = 10000 possible input combinations
2

to nd average runtime per iteration. Finally, we divided the latter runtime
by all 100 nodes to nd the average runtime per node. Table 2 provides the
average of each result for both two-monitor and three-monitor WSNs. The
program conducted each trial on consistent computer hardware and software,
with specications consisting of a standard Windows version of MATLAB on
a quad core 2.60 GHz processor and 8.0 GB of RAM. The signicance of these
results will be explained in greater detail over the next subsection.

27
 Algorithm PSO WSLA WSRA DS
Runtime (µs) 114570 7800 9700 344
Table 3: Comparison of computational runtime among dierent localization techniques.

4.3. Comparisons to Other Established Algorithms

To gain further perspective into the eectiveness of the proposed tech-
nique, we compared our results to those of four previously established local-
ization approaches. These consisted of particle swarm optimization (PSO),
maximum-likelihood estimation (MLE), and a weighted search-based local-
ization algorithm (WSLA). Also included is a weighted search-based rene-
ment algorithm combined with the latter technique that is referred by Yao
and Jiang as WSLA+WSRA but is reported here as simply WSRA. For
the compared methods, the simulated network consisted of 100 sensor nodes
with eight anchor nodes while each experiment contained 30 independent
trials averaged together.
4.3.1. Accuracy
The results are based on those presented in (Yao and Jiang, 2015) under
the assumptions of no positioning error on anchor nodes and no noise fac-
tor. Hence, we determined our error values under aforementioned conditions,
then subtracted each value from one and converted the result to a percent-
age. Because the authors reported their error values as a plot rather than
in exact numerical form, we had to estimate the error values by eyeing and
measuring the plots. Hence, to minimize any potential reporting error, we
indicated the accuracy of each algorithm was reported as a whole number
percentage. For the DS-based technique, we based our accuracy by the aver-
age of full accuracy values for three-monitor setups, rounded to the nearest
whole number percentage for consistency.
As Fig. 8 indicates, the proposed technique contains a noticeable edge in
accuracy in comparison to the other methods as well as a signicant lead over
PSO. Although some parameters had to be tweaked between this method and
the others, results are undeniably positive. Thanks to the novel capabilities
of Dempster-Shafer Theory, the proposed localization technique has a dis-
tinct advantage on par with, if not over, multiple established state-of-the-art
localization techniques.
28
 100

90

80

70
Accuracy (%)

60

50

40

30

20

10

0
PSO WSLA WSRA MLE DS
Algorithm

Figure 8: Comparison of accuracy among dierent localization techniques.

4.3.2. Computational Runtime

In addition, we compared the average runtime per node for a three-
monitor WSN setup with three of the four aforementioned techniques, based
on data given in (Yao and Jiang, 2015). The table provides the competing
data as the average CPU time per node with no noise factor. Hardware
specications and MLE runtime were not reported. All method runtimes are
provided in Table 3. Due to the unique and ecient nature of DS Theory, the
proposed technique has an overwhelmingly shorter CPU runtime when com-
pared to all reported competing methods, hence solidifying such an algorithm
as a novel low cost technique.
5. Conclusions and Future Work

This paper presented an ecient approach to localization in wireless sen-

sor networks. This approach was based heavily upon the innovative statistical
framework known as Dempster-Shafer Evidence Theory, a novel data fusion
29
technique that has never before been used in low cost WSN localization.
While DS Theory contains many unique and powerful properties that can
aid in data prediction and analysis, the concept of prediction by plausibility
is of the most vital levels of importance and interest. This is the result of
its consistently high accuracy, ranging up to 98%, under a variety of network
setups and feature sets as well as its overwhelmingly low computational cost
requirements when compared to other established techniques.
5.1. Future Work
Support vector machines (SVMs) have been of a high level of interest
in the eld of WSN localization, almost to the same extent as DS Theory,
due to its fast localization potential and ecient use of processing resources.
Thus, the most likely next step in this research will involve the fusion of
DS Theory with SVMs. As for more specic details of future work, training
of the upper and lower bound weights was done from a limited iterative
approach. Although this brief technique resulted in multiple high levels of
accuracy above 90%, accuracy could be even further enhanced through more
elaborate optimization methods. Possibilities include biologically inspired
methods, such as Genetic Algorithm or Ant Colony Optimization.
Acknowledgment

This research is supported partly by US National Science Foundation

(NSF) under Grant NSF CCSS grant # 1309658. The research of Danda
B. Rawat is supported by NSF CNS grant # 1658972 and NSF CNS grant
# 1650831. Any opinions, ndings, and conclusions or recommendations
expressed in this material are those of the authors and do not necessarily
reect the views of the National Science Foundation.
References

Auer, E., Luther, W., Rebner, G., and Limbourg, P. (2010) A Veried MAT-
LAB Toolbox for the Dempster-Shafer Theory. Workshop on the Theory
of Belief Functions.

Bhatt, D., Aggarwal, P., Devabhaktuni, V., and Bhattacharya, P. (2012)

Seamless Navigation via Dempster Shafer Theory Augmented by Support
Vector Machines. Proceedings of the 25th International Technical Meeting
of The Satellite Division of the Institute of Navigation, 98-104.

30
Bloch, I. (1996) Some aspects of Dempster-Shafer evidence theory for classi-
cation of multi-modality medical images taking partial volume eect into
account. Pattern Recognition Letters, 17, 905-919.
Bulusu, N., Heidemann, J., and Estrin, D. (2000) GPS-less low-cost outdoor
localization for very small devices. IEEE Personal Communications, 7(5),
28-34.
Cakir, O., Kaya, I., and Cakir, O. (2014) Investigation the eect of the
time dierence of arrival sets on the positioning accuracy for source local-
ization. Signal Processing and Communications Applications Conference
(SIU), 2249-2252.

Chiodo, E. and Mazzanti, G. (2006) Bayesian Reliability Estimation Based

on a Weibull Stress-Strength Model for Aged Power System Components
Subjected to Voltage Surges. IEEE Transactions on Dielectrics and Elec-
trical Insulation, 13(1), 146-159.

Clerentin, A. (2000) A localization method based on two omnidirectional per-

ception systems cooperation. IEEE International Conference on Robotics
and Automation, 2, 1219-1224.

Imprecise Probability Propogation Toolbox (IPP Toolbox) (2012) Essen, Ger-

many: University of Duisberg-Essen, Department of Information Logis-
tics, (https://www.uni-due.de/informationslogistik/ipptoolbox.php). Last
accessed March 7, 2016.
Garcia, M., Martinez, C., Tomas, J., and Lloret, J. (2007) Wireless Sensors
Self-Location in an Indoor WLAN Environment. SensorComm 2007. In-
ternational Conference on Sensor Technologies and Applications, 146-151.

Garcia, M., Tomas, J., Boronat, F., and Lloret, J. (2009) The Development of
Two Systems for Indoor Wireless Sensors Self-location. Ad Hoc and Sensor
Wireless Networks, 8(3-4), 235-258).

Kasebzadeh, P. Indoor localization via WLAN path-loss models and

Dempster-Shafer combining. 2014 International Conference on Localiza-
tion and GNSS (ICL-GNSS), 1-6.

Krogh, A. and Vedelsby, J. (1995) Neural Network Ensembles, Cross Valida-

tion, and Active Learning. In G. Tesauro, D. Touretsky, and T. Leen (eds.)
31
 Advances in Neural Information Processing Systems (volume 7) (231-238).
Cambridge, MA: MIT Press.
Li, S., Qin, Z., Song, H. (2016) A Temporal-Spatial Method for Group De-
tection, Locating and Tracking. IEEE Access, 1-10.
Limbourg, P., Savic, R., Petersen, J., and Kochs, H. (2007) Fault tree analysis
in an early design stage using the Dempster-Shafer theory of evidence.
In T. Aven and J. E. Vinnem (eds.) European Conference on Safety and
Reliability - ESREL 2007 (713-722). Taylor & Francis, Stavenger.

Lipeng, G., Juan, J., and Laibin, D. (2011) An Improved Fusion Algorithm of
Evidence Theory. Cross Strait Quad-Regional Radio Science and Wireless
Technology Conference, 1495-1498.

Lloret, J., Tomas, J., Garcia, M., and Canovas, A. (2009) A Hybrid Stochas-
tic Approach for Self-Location of Wireless Sensors in Indoor Environments.
Sensors, 9(5), 3695-3712.

Mao, G., Fidan, B., and Anderson, B. (2007) Wireless sensor network local-
ization techniques. Computer Networks, 51(10), 2529-2553.
MATLAB version 8.1 (2013) Natick, Massachusetts: The MathWorks, Inc.

Mazuelas, S., Bahillo, A., Lorenzo, R., Fernandez, P., Lago, F., Garcia,
E., Blas, J., and Abril, E. (2009) Robust Indoor Positioning Provided by
Real-Time RSSI Values in Unmodied WLAN Networks. IEEE Journal of
Selected Topics in Signal Processing, 3(5), 821-831.

Meyer, C. (2012) Mass Notication? There's an App for That. Security, 2-4.
Nguyen, H. and Walker, E. (1993) On Decision-Making using Belief Func-
tions. In Advances in the Dempster-Shafer Theory of Evidence, 311-330.
Patwari, N., Ash, J., Kyperountas, S., Hero, A., Moses, R., and Correal, N.
(2005) Locating the Nodes: Cooperative Localization in Wireless Sensor
Networks. IEEE Signal Processing Magazine, 22(4), 54-69.
Pelant, M., Stejskal, V. (2011) Multilateration System Time Synchronization
via Over-Determination of TDOA Measurements. 2011 Tyrrhenian Inter-
national Workshop on Digital Communications - Enhanced Surveillance of
Aircraft and Vehicles (TIWDC/ESAV), 12-14.

32
Shen, H., Ding, Z., Dasgupta, S., and Zhao, C. (2014) Multiple Source Lo-
calization in Wireless Sensor Networks on Time of Arrival Measurement.
IEEE Transactions on Signal Processing, 62(8), 1938-1949.

Shen, X., Chen, W., and Lu, M. (2008) Wireless Sensor Networks for Re-
sources Tracking at Building Construction Sites. Tsinghua Science and
Technology, 13(1), 78-83.

Tonon, F. (2004). Using random set theory to propagate epistemic uncer-

tainty through a mechanical system. Reliability Engineering and System
Safety, 85(1-3), 169-181.

Tseng, Y., Kuo, S., Lee, H., and Huang, C. (2004) Location Tracking in a
Wireless Sensor Network by Mobile Agaents and Its Data Fusion Strate-
gies. The Computer Journal, 47(4), 448-460.
Xing, G., Tan, R., Liu, B., Wang, J., Jia, X., and Ji, C. (2009) Data fusion
improves the coverage of wireless sensor networks. Proceedings of the 15th
annual international conference on Mobile computing and networking, 157-
168.
Xu, J., Liu, W., Lang, F., Zhang, Y., and Wang, C. (2010) Distance Mea-
surement Model Based on RSSI in WSN. Wireless Sensor Network, 2(8),
606-611.
Yang, J., Zhou, J., Lv, J., Wei, W., and Song, H. (2015) A Real-Time Mon-
itoring System of Industry Carbon Monoxide Based on Wireless Sensor
Networks. Sensors, 15(11), 29535-29546.
Yao, Y., Jiang, N. (2015) Distributed Wireless Sensor Network Localization
Based on Weighted Search. Computer Networks: The International Jour-
nal of Computer and Telecommunications Networking, 86, 57-75.

33