ArticlePDF Available

Abstract and Figures

In WSNs, the most critical issue is energy consumption as sensor nodes have limited resources. The sensors collect data from the environment where they can fail due to variations in pressure, temperature, and electromagnetic noise. All these can result in misleading readings and measurements where a lot of energy is consumed. Therefore, data fusion is used to overcome these challenges as it assures the accuracy and the efficiency of gathered data, and eliminates data redundancy which results in saving power, thus improving the overall network performance. This paper provides a survey of research related to the data fusion domain to explore many aspects of data fusion in terms of architecture, taxonomy, and techniques and methods. It also evaluates and compares these techniques as it investigates the advantages and the drawbacks of each, and emphasizes the applicability of these techniques in the WSN domain. Finally, it presents the data fusion challenges in WSNs.
Content may be subject to copyright.
International Journal of Scientific & Engineering Research, Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
Data Fusion in WSNs: Architecture, Taxonomy,
Evaluation of Techniques, and Challenges
Marwah Almasri and Khaled Elleithy
AbstractIn WSNs, the most critical issue is energy consumption as sensor nodes have limited resources. The sensors collect data from
the environment where they can fail due to variations in pressure, temperature, and electromagnetic noise. All these can result in
misleading readings and measurements where a lot of energy is consumed. Therefore, data fusion is used to overcome these challenges
as it assures the accuracy and the efficiency of gathered data, and eliminates data redundancy which results in saving power, thus
improving the overall network performance. This paper provides a survey of research related to the data fusion dom ain to explore many
aspects of data fusion in terms of architecture, taxonomy, and techniques and methods. It also evaluates and compares these t echniques
as it investigates the advantages and the drawbacks of each, and emphasizes the applicability of these techniques in the WSN domain.
Finally, it presents the data fusion challenges in WSNs.
Index TermsWireless Sensor Networks (WSNs), Data Fusion, Data Fusion Architecture, Data Fusion Techniques, Data Fusion
Taxonomy, Data Fusion Challenges.
—————————— ——————————
he Wireless Sensor Network (WSN) is a network that is
composed of a large number of sensors. These sensors are
used to sense and observe the surrounding environment.
Subsequently, measurements and readings are collected in
order to be sent to the sink node. WSNs have gained a central
attention in latest research trends. However, many issues
should be considered as these sensors have a limited computa-
tional capability as well as limited energy.
In WSN, sometimes sensors fail to collect accurate data
from the environment due to pressure and temperature. In
other cases, this failure can be attributed to electromagnetic
noise or radiation. Therefore, all readings and measurement
would be inaccurate and inefficient. In order to overcome
these problems, data fusion which is a technique to combine
data from several sources to be more accurate and complete, is
used. Data fusion is applied in centralized systems as well as
in distributed systems [1]. It extends the lifetime of the net-
work, which is a challenging research aspect of WSNs [1]. Da-
ta fusion can eliminate redundant data and thus save energy,
which results in an improved network performance [2].
Data fusion has been used in many detection applications
such as robotics [3]. Recently, new applications such as Denial
of Service (DoS) detection deploy the data fusion concept suc-
cessfully [4]. Another example is intrusion detection [5]. In
WSNs, data fusion is applied in order to enhance the estima-
tions of sensor nodes’ locations [6].
In relation to the importance of data fusion especially in
WSNs, this paper highlights the different architectures of data
fusion and provides detailed information about various data
fusion taxonomy where all existing taxonomy are combined to
give the reader a wider overview. It also presents many tech-
niques that have been applied in WSNs and sensor based sys-
tems in general. Our goal is to analyze each technique and
evaluate the advantages and the disadvantages of each in or-
der to comprehend the best usability of these techniques in
different applications especially in WSNs. In addition, this
survey indicates the challenges of data fusion in WSNs.
This paper is organized as follows: section 2, provides the
data fusion architectures. Section 3, presents several data fu-
sion taxonomies. Section 4, discusses in detail different data
fusion techniques. Section 5, evaluates these techniques and
concludes the advantages and the limitations of each. It also
highlights the best and suitable techniques to be applied in
WSNs. Section 6, states the data fusion challenges in WSNs.
Finally, section 7, concludes our final remarks of the data fu-
sion domain and its applicability in WSNs.
This section presents the different data fusion architectures
applied in WSNs. There are centralized, decentralized, and
hierarchical architecture. Each one has its advantages and dis-
advantages as discussed in the following sub-sections.
2.1 Centralized Architecture
Centralized architecture is the traditional and the simplest
architecture in WSNs. In this architecture, there is one central
node which is called central processor fusion that receives the
sensed data from all other nodes. The central node is also re-
sponsible for fusing all reports gathered by the sensing nodes
[7]. The advantage of the centralized architecture is that it is
simple and optimal. Another advantage is that faulty reports
can easily be detected. On the other hand, this architecture
requires more resources for data processing as it needs higher
Marwah Almasri is currently pursuing PhD degree program in computer
science and engineering at University of Bridgeport, Bridgeport, CT 06604,
USA. E-Mail:
Prof. Khaled Elleithy is the Associate Dean for Graduate Studies in the
School of Engineering at the University of Bridgeport, Bridgeport, CT
06604, USA. E-Mail:
International Journal of Scientific & Engineering Research Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
bandwidth for transmitting data from all sensing nodes to the
central processor fusion [8]. Fig. 1, shows the centralized archi-
tecture of WSNs.
Fig. 1. The Centralized data fusion architecture.
2.2 Decentralized Architecture
Unlike the centralized architecture, the decentralized architec-
ture has no single central node. However, data fusion is im-
plemented locally at each node in the network based on the
observations from neighbor nodes. The advantages of this ar-
chitecture are as follows: the support of any dynamic changes
in the network, scalability, and tolerance [7]. This architecture
has a lighter processing load and a lower communication load
since data are sent to multiple nodes instead of being sent to
the central node. In addition, the user can access the fusion
results faster due to less communication delay [8]. Fig. 2,
shows the decentralized architecture of WSNs.
Fig. 2. The Decentralized data fusion architecture.
2.3 Hierarchical Architecture
The hierarchical architecture is a combination of the central-
ized and the decentralized data fusion architectures. The mo-
tivation of using the centralized architecture is to have better
accuracy where as using decentralized architecture is useful to
decrease computational workload and communication delay
[9], [10]. As shown in Fig. 3, all sensor nodes are partitioned
into a hierarchical level. At each level, many sensor nodes
send data to the fusion node using suitable routing algorithm
to reduce the transmission power. Therefore, the workload is
balanced among all nodes in the network [7].
Fig. 3. The Hierarchical data fusion architecture.
Data fusion can be categorized into three general taxonomy
types, which are: the "relationship among the sources", the
"levels of abstraction", and "input and output" [11]. This sec-
tion presents all data fusion taxonomies and combines the old
and the new taxonomies as shown in Fig. 4.
Fig. 4. All data fusion taxonomies.
3.1 Taxonomy Based on Relationship Among the
In this section, data fusion is divided into "complementary",
"redundant", or "cooperative" [12]. Fig. 5, shows the taxonomy
based on the relationship among the sources.
- Complementary fusion: fuse data from all sensor nodes in
order to reach more general information [13], [14].
International Journal of Scientific & Engineering Research Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
- Redundant fusion: data is fused in order to obtain high quali-
ty information and thus eliminate transmitting redundant data
- Cooperative fusion: data from independent sources is fused
to obtain new data or information such as finding the target
location by using angle and distance [1].
Fig. 5. Taxonomy based on the relationship among the sources.
3.2 Taxonomy Based on Levels of Abstraction
The taxonomy based on levels of abstraction is categorized
into Low Level fusion, Medium Level fusion, High Level fu-
sion, and Multilevel fusion. The details of each level are as
follows [15]:
- Low level fusion: it is also called a signal or a measurement
level fusion. Raw data is input which is combined to get more
accurate data as compared to the individual input and thus
reduce noise.
- Medium level fusion: also called feature/attribute level fu-
sion. The attributes and features of an object are fused in order
to provide a feature map that is used for various purposes
such as segmentation.
- High level fusion: it is also called "symbol or decision level
fusion" [11]. This level of fusion takes symbols as input and
further combines them in order to provide a more accurate
global decision.
- Multi-level fusion: at this level of fusion, the input and the
output of the data fusion system is one of previous levels. To
illustrate this, a decision can be the output of fusing a meas-
urement with a feature [15].
3.3 Taxonomy Based on Input and Output
There are five categories of data fusion based on the input and
the output of data as Dasarathy stated [16]. These categories
are as follows [16]:
- Data in data out (DAI-DAO): raw data is an input to the
data fusion system. The output is a raw data as well but with
more reliable data [11] .
- Data in feature out (DAI-FEO): raw data is the input of
the data fusion system. The extracted feature or attribute of an
entity such as object or situation is the output.
- Feature in feature out (FEI-FEO): the data fusion takes a
feature or attribute as an input to get an improved feature or
extracts new features and attributes.
- Feature in decision out (FEI-DEO): Data fusion input a
group of features into the system in order to generate deci-
sions [1].
- Decision in decision out (DEI-DEO): data fusion takes
decisions as inputs and fuses them to provide new decisions
as outputs.
3.4 Other Taxonomy of Data Fusion
Zhao and Wang [17] have also introduced a new taxonomy of
data fusion in WSNs based on data level, data type, and user’s
3.4.1 Data Level Fusion
Since data in many applications are fused at various levels, the
data fusion is divided into three different levels which are
"raw data level, feature level, and decision level" fusion [11].
Examples of applications at raw data level fusion are image
enhancement and image compression. At feature level fusion,
all characters and attributes of an entity or objects are extract-
ed for further processing. At decision level fusion, the result is
derived to make decisions [17]. Fig. 6, represents the data level
Figure 6. The data level fusion.
3.4.2 Data Type Fusion
Based on the data type, there are three types of data fusion.
These are as follows: "temporal fusion, spatial fusion and tem-
poralspatial fusion"[11]. The temporal fusion means fusing
the data in various time frames but from the same source
whereas spatial fusion means fusing the data at the same time
but from different sources [17], [11]. Finally, temporalspatial
fusion means fusing data continuously from different nodes
over a period of time [17], [11].
3.4.3 Data Fusion based on User’s Requirements
There are three types of data fusion based on user’s require-
ment. Sometimes the user needs a single information about a
concrete place which can be obtained by a single sensor or the
user might need new information regarding a certain area. In
addition, the user might need complete information about the
overall network [17].
International Journal of Scientific & Engineering Research Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
Based on the purpose of the method, data fusion techniques
can be implemented for a variety of "objectives such as infer-
ence, estimation, classification, feature maps, abstract sensors,
aggregation, and compression" [15]. In this section, many
techniques used in data fusion are discussed along with their
applications in WSNs. Fig. 7, shows all data fusion techniques
used in WSNs.
Fig. 7. Data fusion techniques in WSN.
4.1 Inference Methods
Inference method is mostly used in decision fusion where a
decision is generated depending on the perceived situational
knowledge. "Classical inference methods are based on Bayesi-
an inference and Dempster-Shafer Belief Accumulation theo-
ry" [15],[18]. Other inference methods such as fuzzy logic,
neural networks, abductive reasoning, and semantic data fu-
sion are also highlighted.
4.1.1 Bayesian Inference
Depending on the probability theory, Bayesian Inference
merge all evidences where the uncertainty in Bayesian Infer-
ence describes the belief. It assumes the value of 0 for absolute
disbelief and 1 for absolute belief. Bayesian inference is basi-
cally based on the "Bayes’ rule" [19], [15], which is represented
in Equation (1):
Pr(B | A ) = (Pr(A | B ) * Pr(B )) / ( Pr(A)) (1)
Where, Pr(A | B ) is the belief of hypothesis B given the in-
formation A, Pr(A | B ) is the probability of receiving A, given
that B is true, Pr(B ) is the prior probability, and Pr(A) is the
normalizing constant.
The critical issue in Bayesian Inference is that the probabili-
ties Pr (A) and Pr (A|B) should be estimated because they are
unknown. The neural network approach has been used to
guess the conditional probabilities for the decision-making
process in Bayesian inference module [20]. In addition, Cou´E
et al. [21] used Bayesian programming in fusing data from
various sensors such as laser and video in order to obtain
more reliable and accurate data. In WSNs, Krishnamachari
and Iyengar [22] uses Bayesian Inference method for event
detection. The inference algorithm in [23] uses Bayesian Infer-
ence to detect the missing data from sleep nodes within a sens-
ing period.
4.1.2 Dempster-Shafer Inference
This method is based on the "Dempster-Shafer Belief", which
generalizes the Bayesian theory. Dempster-Shafer Belief was
proposed by both Dempster [24] and Shafer [25]. Dempster-
Shafer Inference introduces a formalism that is applied for
incomplete knowledge and evidence combination [26]. An
important factor in Dempster-Shafer method is the set of all
possible states which further demonstrate the system. This set
is called the ‘frame of discernment’. The elements of the power
set of possible states are called hypotheses. Each hypothesis
has its assigned probability. In addition, the belief function
which is called ‘bel’ is defined by Dempster-Shafer and also
the degree of doubt ‘dou’ that is based on the belief function
are [27].
In [28], the authors provided an implementation of both
the "Dempster-Shafer" and the "Bayesian inference" into one
algorithm. The "Dempster-Shafer inference" was used to pro-
vide battlefields' dynamic pictures in a WSN that consists of
"Unmanned Aerial Vehicle (UAV)" as sensor nodes for evalua-
tion purposes where in fact the fusion challenges in a mobile
network were not evaluated [29]. "Data Service Middleware
(DSWare)" in WSNs, by [30], uses this technique where each
decision is assigned to a confidence value. This value is calcu-
lated by the predetermined confidence function.
4.1.3 Semantic Data Fusion
Semantic data fusion is done as an in-network inference. The
semantic data fusion method is composed of two important
phases. The first phase is called knowledge base construction,
which collects the "knowledge abstractions" into a form of se-
mantic data. The second phase is called pattern matching (in-
ference), which uses the semantic data provided by the previ-
ous phase to fuse relevant attributes for pattern matching [31].
This method was first introduced by Friedlander and Phoha
[31] for target classification. Friedlander [32] explains many
techniques that extract semantic data from sensors by convert-
ing sensor data into formal languages. He applies these tech-
niques for the recognition of the robots’ behavior and for sav-
ing resources. In [33], users can formulate queries based on
semantic values without the knowledge of which data or op-
erations are used.
4.1.4 Fuzzy Logic
Fuzzy logic deals with "approximate reasoning" in order to
obtain "conclusions from imprecise premises" [34], [1]. Zadeh
[35] has introduced the concept of fuzzy sets which later guid-
ed him to the fuzzy logic theory. The data fusion algorithm
based on fuzzy logic theory has four main phases: "fuzzifica-
International Journal of Scientific & Engineering Research Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
tion", "rule evaluation", "combination" or "aggregation of
rules", and "deffuzification" [36]. In the second phase which is
the rule evaluation, the implications or rules are used to pro-
cess the fuzzified inputs. These rules are in the form of “if A
then B”, where A is a conditional statement. Sometimes more
than two conditional statements are used which is called com-
plex implications. When applying complex implications, fuzzy
operators are used for computing the final result [37]. The
most common fuzzy logic inference operators used are shown
in Equations (2), (3), (4), (5), (6), (7), (8), and (9) as follows [37]:
xy = yx
xy = min{1,1-x+y}
xy = min {x,y}
xy = max { 1-x,y}
xy = 1-x+xy
In Equation (4), the Mamdani inference operator is present-
ed. It finds the minimum degree of the membership (x, y).
Both Mamdani and Tsukamoto-Sugeno inference methods are
based on fuzzy logic [38]. However, the Mamdani method is
considered a better method since it ensures an efficient data
fusion, extends the sensor lifetime, and reduces delay com-
pared to Tsukamoto method.
In [39], authors use fuzzy logic control and an intelligent
sensor network for autonomous navigational robotic vehicle
which has the ability of avoiding obstacles. Cui et al. [40] use
position algorithm based on a fuzzy logic to deal with the un-
certain data that the sensors gathered. Moreover, a fuzzy op-
timization algorithm is used to update the location of each
node. [41], uses fuzzy reasoning to find the best cluster-heads
in a WSN. Another implementation of fuzzy logic is for effi-
cient routing that minimizes energy usage [42]. Wallace et al.
[43] introduced the Medium Access Control (MAC) protocols
based on fuzzy logic concept in two stages. The purpose is to
extend the network lifetime. The first stage has several inputs
such as the current transmit queue size, collision of the previ-
ous packages, and remaining battery. The second stage uses
the same inputs used in the first stage but with a priority.
4.1.5 Neural Networks
The Neural network is applied in "learning systems" with
fuzzy logic to manage its "learning rate" [44], [45], [1]. In the
data fusion domain, neural networks have been applied for
"Automatic Target Recognition (ATR)" [46]. Neural Networks
have been applied in many applications. Lewis and Powers
[47] fused audio-visual information using neural networks for
audio-visual speech recognition.
4.1.6 Abductive Reasoning
Abductive Reasoning is the best hypothesis for explaining
observed evidence [48]. Fig. 8. shows the deduction and ab-
duction example. The abductive inference finds the maximum
a posteriori probability [49]. Abduction was used in machine
learning problems [50] and diagnosis problems [51].
Fig. 8. The deduction and abduction example
4.2 Estimation Methods
Estimation methods are derived from the control and the
probability theories in order to calculate a process vector from
a series of measurement vectors [52]. Examples of Estimation
methods are Maximum A Posteriori (MAP), Particle filter,
Least Squares, Kalman filter, Maximum Likelihood (ML), and
Moving Average filter. The details of each method are pre-
sented in this section.
4.2.1 Maximum A Posteriori (MAP)
This technique is based on Bayesian theory. Given that ‘a’, is
the state to estimate, where ‘b’= {b(1),b(2),..,b(n)} is a set of n
observations of ‘a’, the MAP estimator is used to figure out a
value of ‘a’ in order to maximize the posterior distribution
function [53] as in Equation (10).
(n)=argmaxa pdf(a|b) (10)
where pdf is the probability density function.
MAP estimator was used by Schmitt et al. [54] in a known
environment to locate the joint positions of mobile robots. An-
other implementation of MAP estimator was by Yuan and
Kam [55] in the collision resolution algorithm. The algorithm’s
purpose is to control the traffic between the fusion node and
the source, where MAP estimator figures out the number of
nodes that are being transmitted. Therefore, the retransmis-
sion probability of these nodes needs to be updated according-
International Journal of Scientific & Engineering Research Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
4.2.2 Particle Filter
These filters are recursive processes of the "sequential Monte
Carlo methods (SMC)" [56]. They are suitable for applications
that implement a non-Gaussian noise [57]. They use a large
number of random measures which are composed of particles
(samples) that are driven from distributions and weights of
the particles. The random measures are helpful in calculating
all kinds of unknown estimates such as minimum mean
square error (MMSE) and maximum a posteriori (MAP). The
Particle filter technique represents significant densities by par-
ticles and weights. It then computes the integrals by Monte
Carlo methods. There are three important operations of the
Particle filters: sample step which generates particles, im-
portance step which computes the particle weights which are
later normalized, and the resampling step. The resampling is
important as it eliminates the trajectories with small weights
and highlights the ones that are dominating [58].
Filters have been used in target tracking problems within
WSNs, such as [59] where Particle filters are used in a tracking
algorithm along with binary detection model. Wong et al. [60]
also used Particle filters in a collaborative data fusion scheme
to fuse information from different sensors for tracking targets.
Hu and Evans [61] used this technique in a mobile network to
find the nodes’ locations. They argue that mobility enhances
accuracy and thus decreases localization costs.
4.2.3 Least Squares
The "Least Squares method is a mathematical optimization
technique that searches for a function that best fits a set of in-
put measurements. This is accomplished by minimizing the
sum of the square error between points generated by the func-
tion and the input measurements" [1]. Unlike the "Maximum
A Posteriori Probability", this the Least Square does not use
any previous probability. Therefore, it works in a determinis-
tic manner [15]. The Least Squares method tries to find the
value of x [53] as in Equation (11).
Where h is the sensor model for a sequence of 1 ≤ i n ob-
The "Huber Loss function" [62], the "ordinary squared er-
ror" [53], and the "root mean squared error" [63] are various
Square Error metrics. An advantage of using the Least Squares
method is reducing the communication between the source
node and the sink. This is achieved by sharing the sensor data
through the linear regression instead of transmitting the actual
data [63]. In addition, these filters were implemented in the
sink node as well as in the source node to avoid sending all
the data from the source to sink. This is done in a dual predic-
tion scheme where the data will be transmitted to the sink
node if the predicted and the actual values have a difference
more than a given error [64].
4.2.4 Kalman Filter
The Kalman filter is invented by Kalman [65] and it gained
popularity as a technique used for data fusion in WSNs. The
Kalman filter is shown in Fig. 9. Based on some measurement
y(n) which is shown in Equation (12), and the system parame-
ters (which are known in advance), the estimate of x(n), and
the prediction of x(n + 1) are presented in Equations (13), and
(14) respectively.
y(n) = H(n) x(n) + r(n) (12)
Where: H(n) is the measurement matrix r is a random variable
that follows the zero-mean Gaussian laws.
(n)= (n | n-1)+K(n)[y(n)-H(n) (n | n-1)] (13)
Where K is the Kalman filter gain.
(n + 1 | n) = Ts (n) (t | t) + Ti (n)I(n) (14)
Where: Ts(n) is the state transition matrix, Ti (n) is the input
transition matrix, and I (n) is the input vector.
The Kalman filter technique works well in a linear model
where it retrieves optimal estimates recursively [66]. On the
other hand, in a nonlinear model, other methods should be
used such as "Extended Kalman filter (EKF)" [67], and the
"Unscented Kalman Filter (UKF)" [68]. In WSNs, data loss is an
issue due to unreliable communication links. [69] evaluated
this method's performance based on many observations where
they found that at some point the Kalman filter becomes un-
The Kalman filter has also been applied for the purpose of
source localization [53]. It is also used to track different
sources [70]. Others used a "dual Kalman Filter" method in
order to forecast the sensed data. Therefore, when the sink
node forecasting is inaccurate, the source node can send data
in this situation [71]. In addition, the Kalman filter used in the
SCAR routing algorithm to forecast some valuable infor-
mation about the nodes’ neighbors. After that, the SCAR rout-
ing algorithm would choose the routing path and the best
neighbor depending on these predictions [72].
Fig. 9. Kalman filter block diagram
4.2.5 Maximum Likelihood (ML)
To estimate a state a as an example, where b’=
{b(1),b(2),..,b(n)} is a set of n observations of a’, the likelihood
function is defined as follows:
International Journal of Scientific & Engineering Research Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
λ(a) = p (b |a)
where p is the probability density function.
The Maximum Likelihood estimator (MLE) is used to fig-
ure out a value of a’ in order to maximize the likelihood func-
tion [53] as in Equation (16).
A new distributed and localized MLE was proposed by
Xiao et al. [73] with more robustness, where each node can
compute a "local unbiased estimate" to eventually reach "the
global Maximum Likelihood solution" [15]. This method was
further developed by Xiao et al. [74] in order to deliver meas-
urements in a timely manner.
Other implementations of MLE that were helpful to reduce
the necessity of sharing all data are the "Decentralized Expec-
tation Maximization (EM) algorithm" [75], and the "Local Max-
imum Likelihood Estimator" [76]. The MLE is very helpful in
location discovery problems such as, to compute distance, di-
rection or angle to know the exact location of nodes or targets.
In the case of finding the node location, an example is the
"Knowledge-Based Positioning System (KPS)" [77] which has a
predefined value of the pdf of the node so that each node es-
timates its location using the MLE. Another example is using
MLE to find the source location which is provided by Chen et
al. [78], where the authors use the bird monitoring application.
In the network tomography, MLE was used throughout the
aggregation and reporting process for estimating per-node
loss rates which has a great impact on routing algorithms es-
pecially for robust fault-tolerant protocols [79].
4.2.6 Moving Average Filter
The moving average filter is mainly used in "digital signal
processing (DSP) solutions" [15]. It has many advantages such
that it is easy to use as it reduces "random white noise" while
maintaining a "sharp step response" [15]. For this reasons it is
an optimal filter in the time domain for processing encoded
signals [80]. The true signal x = ( (1), (2), . .) is estimated by
Equation (17).
Where z=(z(1), z(2), . . .), is the input digital signal, w is the
filter’s window that indicates the number of input observa-
tions for every n w.
In addition, w refers to the number of steps needed for the
filter to identify the signal level's variance. As the value of w
increases, the signal becomes cleaner. In contrast, as the value
of w decreases, the step edge becomes sharper. The Moving
Average filter is able to decline √w of the white noise variance
[80]. Yang et al. [81] have used this technique in target loca-
tions which in turn reduces the chances of inaccuracy of track-
ing applications in WSNs. Other types of Moving Average
filters in WSNs are "Weighted Moving Average" and "Expo-
nentially Weighted Moving Average" (EWMA) filters. The
EWMA filter has been used to determine noise in MAC proto-
cols [82]. It has other helpful uses in WSNs such as in localiza-
tion [83], in detection and classification [84], and local clock
synchronization [85].
4.3 Compression
Compression methods are applied in WSN through spatially
correlating all sensor nodes with no additional communication
cost. This can be obtained by providing two nodes with corre-
lated observations [86]. Several compression methods are dis-
cussed in this section.
4.3.1 Distributed Source Coding (DSC)
Distributed Source Coding (DSC) [87], is "the compression of
multiple correlated sources, physically separated, that do not
communicate with each other "[88]. One of the most popular
data compression methods in WSNs is the "Distributed Source
Coding Using Syndromes" (DISCUS) framework [89]. In DIS-
CUS, assuming we have a node X which wants to transmit its
observation to node Y. In order to code X’s observation, X can
send only an index. There is one requirement which is the
Hamming distance between X and Y which is at most one.
This means that, the difference of X and Y can be only one bit.
Suppose that a sensor observation can be any value of the set
S={000, 001, 010, 011, 100, 101, 110, 111}. X and Y have four
cosets {000, 111}, {001, 110}, {010, 101}, {100, 011}. As shown in
Fig. 10, node X sends the index of 10 which corresponds to the
coset of {010, 101}. Y now can decode the index along with its
own observation of (100). Since the Hammimg distance should
be at most one between the two, Y knows that the value pro-
vided by X should be 101 [15].
Critescu et al [90] applied Slepian-Wolf coding which is
based on distributed source coding. It is a kind of distributed
source coding technique that eliminates redundant data due to
the spatially correlated observations in WSNs [91]. Marco and
Neuhoff [92] applied Slepian-Wolf coding locally within each
cluster. The result was efficient as it mitigates the node’s fail-
ure when the data is reconstructed at the sink node.
4.3.2 Coding by Ordering
This technique was first introduced in Petrovic et al. [93]. In
this technique, each node sends the data to the border node.
The border nodes are responsible for sending what is called a
supper-packet, which is a group of all packets, to the sink
node. Table 1, gives an example of coding by order. As shown
in Table 1, we have four nodes that each of them provides an
observation of the value from 0 to 5: X,Y,Z, and W. As shown
in Table 1, the border node can suppress all values by W. The
ordering is 3! which means that we have 6 possible orderings
of the three remaining nodes: X, Y, and Z. For example, if the
observation value for node W is 1, the packet order is {X,Z,Y}
where it can be {Z,X,Y} if the observation value for node W is 4
and so on [15].
In addition, there are other data compression techniques
that are applied in WSNs. Ju and Cui [94] introduced a com-
pression technique called The Easinet Packet Compression
(EasiPC) which focuses on the transmitted packet and discov-
ers the redundancy within that packet. Recently, researchers
have focused on joint data compression. Pattem et al. [95] ar-
gue that a static clustering scheme offers near-optimal perfor-
mance for spatial correlations.
International Journal of Scientific & Engineering Research Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
Fig. 10. An example of DISCUS data compression in WSNs.
Packet Ordering
Observation Value (W)
4.4 Aggregation
According to Kulik et al. [96], data aggregations is defined as a
technique that is used for solving two kinds of problems: im-
plosion, which occurs when the data sensed is duplicated by
the same node because of the strategy used in routing and
overlap, which occurs when two different nodes broadcast the
same data (redundant sensors) [15]. Redundancy has a nega-
tive effect on the network as it wastes the network's energy as
well as its bandwidth. Therefore, data aggregation and data
fusion are important to reduce energy consumption. For that
specific reason, data aggregation is applied for the purpose of
reducing redundancy in neighboring nodes [97], [98]. Instead
of the classical address-centric approach that was used in data
forwarding, a novel data-centric approach is currently used
[99]. Each time the sensor node receives information from a
neighbor node, it needs to determine whether this information
is worth forwarding to other sensor nodes; otherwise it will be
a waste of resources. Using data fusion techniques can de-
crease the number of packets needed to be transmitted by pro-
cessing data locally and then send only a digest to the sink
node which in return saves energy and bandwidth. To illus-
trate this, the centralized approach takes O (n3/2) bit-hops,
where when applying data fusion techniques it takes only O
(n) bithops for data transmission [62].
In WSNs, data aggregation proved its benefits to save en-
ergy consumption. Krishnamachari et al. [100] have discussed
the results of the aggregation tree creation. They analyzed the
costs and the delay of data aggregation, and the complexity of
optimal data aggregation. In addition, the tradeoff between
accuracy and energy consumption have been studied while
using aggregation functions in WSNs [101]. Several aggrega-
tion functions are used in WSNs such as suppression [97],
which discards duplicates and thus eliminates data redundan-
cy. Another aggregation function is called packaging [102].
This aggregation function uses a single packet for all observa-
tions which reduces the overhead of the MAC protocol every
time a packet is sent. Moreover, the greedy aggregation ap-
proach outperforms the opportunistic approach in terms of
energy savings especially in a network with a high node den-
sity [103].
In-network data aggregation algorithms have gained a lot
of attention recently since they require coordination among
nodes when they are distributed in the network to assure high
performance which is basically a complex functionality. In-
network aggregation can be defined as collecting and routing
data within a "multi-hop network" where it processes data at
intermediate nodes in order to decrease energy consumption
and thus increase the network’s lifetime [14]. Regarding in-
network aggregation, there are two approaches which are as
follows: In-network aggregation with size reduction or with-
out size reduction. In the first approach, data from different
sources are combined and compressed and further sent over to
the network which decreases the information to be sent but
reduces the accuracy of the aggregated information at the sink
as well. The second approach merges all packets from various
sources into one packet with no data processing which keeps
the original information and thus ensures high accuracy at the
sink node [14].
4.5 An Information Theory Approach
Using multiple sensors instead of a single sensor in any net-
work can enhance data and observation reliability. Infor-
mation fusion based on multiple sensors are harder to esti-
mate in advance. This leads to probabilistic data collection and
processing which can be measured and analyzed by applying
the information theory principles [104]. In addition, the deci-
sion theory is another essential aspect in WSNs [105]. Both the
"Information" and "Detection" theories help in solving many
problems regarding data fusion. Ahmed and Pottie [106] have
used a Bayesian technique for fusion which uses different sen-
sor types along with different sensing capabilities. There are
interesting tradeoffs between information rate and the distor-
tion theory which can be found using entropies [107].
4.6. Reliable Abstract Sensors
This method was first proposed by Marzullo [108] which sug-
gests three different types of sensors: "concrete sensor", which
senses the environment by collecting samples of a physical
variable, "abstract sensor" which represents the observation in
a set of values depending on the concrete sensor, and "reliable
abstract sensor" which contain the real values of the physical
variable. This type of sensor is computed using a number of
abstract sensors. This fusion method has been applied in vari-
ous applications in time synchronization [109]. Many algo-
rithms and functions that are used with reliable abstract sen-
sors for time synchronization such as "Fault-Tolerant Averag-
ing" algorithm and "Fault-Tolerant Interval" (FTI) function.
International Journal of Scientific & Engineering Research Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
4.6.1 Fault-Tolerant Averaging
This algorithm is used in data fusion methods as it fuses a n
number of "abstract sensors" into correct "reliable abstract sen-
sors" even if there are incorrect sensors [108]. The algorithm
works as follows. Suppose we have L={I1, . . . , In} where Ii =
[xi , yi] by n abstract sensors at the same time and we have at
most f of n abstract sensors which are incorrect or faulty. The
"Fault-Tolerant Averaging" algorithm is shown in Equation
(18) which has a complexity of O(nlog n) [108].
Low refers to the smallest value in at least n − f intervals in L,
and High refers to the largest value in at least n − f intervals
in L.
Fig. 11, shows two different scenarios of applying the Fault-
Tolerant Averaging algorithm where there is one faulty sen-
sor. In Fig. 11 (a) Sen 2 and Sen 3 do not have any intersection;
therefore, one of them is the faulty sensor. (sen1,sen 2,sen
3 ,sen 4) has {Low,High}, where Low (the left edge of Sen 1)= n
− f = 4 − 1 = 3, and High (the right edge of Sen 4)= n − f = 4 − 1
= 3. However, in Fig.11 (b), the right edge of Sen 2 has moved
to the left and becomes Sen 2.
As a result, we have now (sen1,sen 2',sen 3 ,sen 4) which
indicates the instability of M. Consequently, the left edge of
the result is the left edge of Sen 3 (Low value) and the right
edge of the result is the right edge of Sen 4 (High value). This
algorithm was further extended by Chew and Marzullo [110]
where they fuse data from multidimensional sensors.
Fig. 11. Two different scenarios of applying the "Fault-Tolerant Averag-
ing" algorithm where there is a one faulty sensor.
4.6.2 The Fault-Tolerant Interval Function
This function was introduced by Schmid and Schossmaier
[111]. The Fault-Tolerant Interval (FTI) function is also used in
data fusion methods. Again, we have at most f of n abstract
sensors considered as incorrect or faulty sensors. FTI function
is shown in Equation (19).
(L)={Low,High} (19)
Low refers to the ( f + 1)th largest of the left edges {x1, . . . , xn}
High refers to the ( f + 1)th smallest of the right edges { y1, . . . ,
FTI function indicates that when there are few alterations
in the input intervals, unlike the Fault-Tolerant Averaging
algorithm, the result will include only few changes as well. As
a result, the FTI function is more robust as compared to the
Fault-Tolerant Averaging algorithm [111].
Fig. 12 shows the same example as Fig. 11, however the re-
sult is not that affected when Sen 2’ is moved (Fig. 12(b)).
Therefore, FTI obviously is less vulnerable to small alterations
in the input intervals as compared to the Fault-Tolerant Aver-
aging algorithm [111].
Fig. 12. Two different scenarios of applying The "Fault-Tolerant Interval"
(FTI) function
4.7 Feature Maps
Sometimes using raw sensory data is not sufficient especially
in guidance and resource management applications. As a re-
sult, some features that well describe the environment need to
International Journal of Scientific & Engineering Research Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
be extracted [18]. Many data fusion methods of inference and
estimation produce a feature map. There are two which are
occupancy grid and network scans.
4.7.1 Occupancy Grid
Occupancy maps define a 2D/3D representation of the space
which is organized in square cells where every cell has an es-
timate that indicates its probabilistic occupancy [112]. This
probability is calculated by using multiple types of sensors
and various data fusion techniques [113]. Occupancy maps are
used in many applications such as robot perception [114], the
location's estimation [115], and navigation [116].
4.7.2 Network Scans
Network Scans are kinds of activity maps for WSNs. They also
give an overview of the resource distribution in the network
[117]. One of the most popular network scans is called eScan
[117] which provides information about the remaining energy
in the network. The algorithm forms an aggregation tree
where each node calculates its local eScan and then sends it to
the sink. If two or more eScans are received at the same node,
an aggregation process is involved to identify the remaining
energy of nodes in a specific region. Finally a map is generated
5. Evaluation and Comparison of Data Fusion
This section evaluates all the data fusion techniques presented
in this paper and draws a conclusion about which technique is
most suitable and reliable to be applied in WSNs.
Both the "Bayesian Inference" and the "Dempster-Shafer"
theory are well-known Inference methods. Dempster-Shafer
method generalizes Bayesian Inference. However, "Dempster-
Shafer theory" is a more flexible method than "Bayesian Infer-
ence" due to its capability to fuse data from various types of
sensors unlike Bayesian Inference [14]. Another difference
between these two techniques is that Dempster-Shafer theory
does not require assigning apriori probabilities to unknown
propositions [18]. In contrast, Dempster-Shafer involves long-
er calculations [118]. In addition, Fuzzy logic method is best
suitable for decision making with uncertain information from
multiple sensor nodes. It also improves the quality of infor-
mation and thus can be implemented effectively in data fusion
in WSN [37]. On the other hand, fuzzy logic cannot solve
problems without the knowledge of an expert as it does not
have the learning membership function either during solving
the problem or after the problem has been solved [119].
Applying neural network in WSNs has many advantages.
In neural network, data fusion is done closely to the source
node which results in enhancing its performance. The algo-
rithm used in neural network draws the important features of
data and can be adjusted to meet the requirement of various
applications [120]. It also provides robustness to handle many
issues like noise [121]. It identifies various signals and reduces
the errors and false alarm rate of the sensors in an efficient
manner [122]. However, many issues need to be considered
during the implementation of a neural network such as the
problem of local extremum, misclassification due to data di-
mension increase, and convergence speed of the training [123].
Abductive Reasoning is another technique which works for
pattern reasoning more than a data fusion method. It has not
been formally used in WSNs but it is used successfully in fault
diagnosis and event detection [15].
The semantic data fusion technique has the ability to im-
prove resource utilization especially when collecting and pro-
cessing data in WSNs [15]. This method also reduces transmis-
sion cost because the nodes transmit formal language struc-
ture without the need of transmitting raw data. On the other
hand, this technique requires in some scenarios a known set of
behaviors in advance, which is a difficult process in specific
situations [124].
Moreover, when the state that needs to be estimated is not
based on some random variables, the Maximum Likelihood
(ML) technique is suitable to be applied. It also finds the value
of this state and assumes it is fixed. In contrast, the "Maximum
A Posteriori" (MAP) technique does not consider that the
state’s value is fixed. On the other hand, it takes it as the result
of some random variables with known prior pdf [53]. In addi-
tion, the "Least Squares" technique is more accurate and suita-
ble to be applied where the state is fixed. This technique does
not use any previous probability as compared to the Maxi-
mum A Posteriori (MAP) technique [15]. The Moving Average
Filter technique can be used to decrease the random white
noise. It has also been used in WSNs to reduce the errors
caused by tracking applications [81]. The downside of this
technique is that an old value will have the same impact as the
most recent measurement which will affect the final result
Kalman filter is an important and powerful technique as it
can estimate past, present, and future states [67]. However,
when used in WSNs, it needs clock synchronization which can
impact its performance [126]. The Kalman filter can be unsta-
ble due to the "critical value for the arrival rate of the observa-
tions" [69].
Furthermore, Particle filter is an excellent technique used
to overcome some difficult problems such as signal pro-
cessing, navigation, communications, and computer vision.
On the other hand, it has some drawbacks as it is considered a
complex technique that has a computational intensity [58].
In addition, even though Occupancy grids show only a re-
stricted class of maps which indicate incorrect independence
assumptions in prior and posterior distributions, they also
have the advantage of being simply applied [127]. The net-
work scan technique can be helpful in describing network re-
sources and activity. In particular, eScan can guide designers
as to where to deploy new sensors since it presents low energy
regions [117]. Moreover, the Fault-Tolerant Averaging tech-
nique can successfully fuse n number of abstract sensors into
correct reliable abstract sensors where in fact there are incor-
rect original sensors [108]. However, few alterations in the
input intervals can affect the performance of the "Fault-
Tolerant Averaging" algorithm [108]. On the other hand, the
Fault-Tolerant Interval Function is more robust due to the fact
that few alterations in the input intervals will lead to only few
alterations in the output [111].
The aggregation technique helps to eliminate redundancy
and traffic load which saves energy in the network. However,
International Journal of Scientific & Engineering Research Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
by using this technique, the fusion node can be compromised
by malicious attackers which affect the correctness of the fu-
sion data. Another disadvantage of this technique is that there
might be multiple copies of the same fusion results at the sink
node which increases the energy consumption at the sink node
Distributed Source Coding (DSC) has the advantage of
making the coding decisions process works efficiently sepa-
rated from the routing process. On the other hand, it requires
more computational complexity. It also needs to collect some
data from joint statistics which is not an easy task [14]. The
Code by Ordering technique is simple but does not present all
possible correlations between sensor nodes [15]. Finally, the
information theory approach is suitable for analyzing many
problems regarding data collection and processing by multiple
sensors [104].
Table 2, summaries the advantages and the disadvantages
of all data fusion techniques. Based on previous findings, we
evaluate the various data fusion techniques discussed in this
paper and draw a closure. To conclude, there are various data
fusion techniques that have been applied. However, in WSNs,
some of these techniques do not concern the specific require-
ments of this type of network such as low energy consumption
and flexibility. Therefore, for the best applicability of data fu-
sion in WSNs, some techniques outweigh others as follows:
- The Dempster-Shafer is a good technique as it fuses data
sensed by different types of sensors which are needed in many
- The Fuzzy logic technique performs very well in the decision
making process and has better data quality.
- Neural networks enhance the process of data fusion which is
an advantage in WSNs as it saves power consumption.
- The Semantic data fusion technique saves resources in
- The Least Squares technique has high accuracy in WSNs.
- The Moving Average Filter technique can be used in WSNs
to decrease the chances of errors which also saves a lot of en-
ergy and thus increases the performance of the network.
- The Network scan (eScan) can show low power regions in
order to fill in with new full energy sensors.
- The aggregation technique eliminates redundant data and
thus saves energy.
6. Data Fusion Challenges in Wireless Sensor
There are many challenges that need to be considered while
applying data fusion in WSNs. However, it is a challenging
task to try to handle all these issues in one data fusion algo-
rithm. These issues are as follows:
A. Security:
Although data fusion in WSNs saves power consumption as it
eliminates redundant data and thus enhances the overall per-
formance of the network, it risks the security of the network as
well. It makes the network easily attacked by data intercep-
tion, data falsification, data tampering and data repeated at-
tacks. Any attacker can reach security information such as
keys by capturing a single node; therefore, all data fusion al-
gorithms should guarantee the security of these information
even in case of one of the nodes is captured [128].
B. Data Imperfection:
Sometimes the data collected by sensors contain uncertain or
imprecise measurements. Hence, data fusion algorithms
should handle this issue by eliminating data redundancy ef-
fectively [129].
C. Data Correlation:
In WSNs, sensor nodes might be exposed to an external noise
which in turn affect the measurements. The data fusion algo-
rithm should consider data dependencies otherwise it experi-
ence over/under confidence in results [130].
D. Data Dimensionality:
Data collected can be preprocessed at every sensor node (lo-
cally) or at the fusion center (globally) and compressed in or-
der to lower the dimensional data. This is helpful in reducing
the power consumption as saving the communication band-
width [131].
E. Conflicting Data:
Since in the data fusion system various sources are used,
conflicting data can be occurred due to incomplete data, out-
of-date data, or by erroneous data [132].Therefore, a special
care is needed when dealing with conflicting data in any data
fusion algorithm.
7 Conclusion
With the revolution of WSNs and the size, redundancy, inac-
curacy of the collected data, researchers have focused on the
data fusion field. Data fusion plays a key role in WSNs as it
reduces power consumption and improves the efficiency of
the gathered data. Therefore, this paper provides a compre-
hensive survey of data fusion in WSNs. Our aim is to focus on
the evaluation and the comparison between various data fu-
sion techniques. However, some limitations of these tech-
niques which have been found need to be considered. Apply-
ing data fusion architecture in the WSNs context can face some
problems since they are not network-based. However, it can
be applied in specific applications in WSNs. There are some
challenges need to be handled when developing data fusion
algorithms in WSNs.
In future works, we would like to investigate and analyze
further challenges such as the assurance of temporal and spa-
tial correlation while applying data fusion and transmission
International Journal of Scientific & Engineering Research Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
Data Fusion Technique
Bayesian Inference
More accurate than Dempster-Shafer tech-
Does not fuse data from various types of sensors
Needs to assign apriori probabilities to unknown
Generalizes Bayesian Inference technique
Flexible technique because it has the ability
to fuse data from various types of sensors
Does not assign apriori probabilities to un-
known propositions
Less accurate technique as compared to Bayesian
Longer calculations involved
Fuzzy Logic
Effective data fusion technique to be ap-
plied in WSNs due to its ability of enhanc-
ing the data quality.
Needs the knowledge of an expert to solve the
Learning the membership function is difficult
during or after solving the problem
Neural Network
Enhance the performance of data fusion
because it is done closely to the source
The neural network’s algorithm is adjusta-
ble to the application requirements.
Efficiently decreases the errors and false
alarm rate of the sensors
Many issues need to be solved such as local ex-
tremum, misclassification, and convergence
speed of the training.
Abductive Reasoning
Successfully used in fault diagnosis and
event detection
Not been formally used in WSNs
Semantic Data Fusion
Improves resource utilization in WSNs
Reduces transmission cost
Requires a known set of behaviors in advance,
which is a difficult process in specific situations.
Maximum Likelihood (ML)
Suitable when the state is not a random
Does not require the sharing of all data
Maximum A Posteriori
The state’s value is the result of some ran-
dom variables with known prior pdf
Least Squares
Does not use any prior probability as com-
pared to the Maximum A Posteriori (MAP)
Moving Average Filter
Decreases the random white noise
Reduces the errors caused by tracking ap-
plications in WSNs.
The final result can be easily affected as the old
value will have the same impact as the most re-
cent measurement.
Kalman Filter
Estimates past, present, and future states.
It needs clock synchronization which can impact
its performance
Unstable due to the critical value found for the
arrival rate of the observations
Particle Filter
Can solve some difficult problems such as
signal processing, navigation, communica-
tions, and computer vision.
A complex technique that has a computational
Occupancy Grids
Can be simply applied
Shows only a restricted class of maps which pre-
sents incorrect independence assumptions.
Network Scan
Describes the network resources and activi-
eScan can guide designers as to where to
deploy new sensors as it demonstrates low
energy regions
If two or more eScans are received at the same
node, an aggregation process is required in order
to determine the remaining energy of the nodes.
Fault-Tolerant Averaging
Fuses several abstract sensors into correct
The performance can be affected by few altera-
International Journal of Scientific & Engineering Research Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
reliable abstract sensors where in fact these
abstract sensors are incorrect original sen-
tions in the input intervals
Fault-Tolerant Interval Func-
More robust than the Fault-Tolerant Aver-
aging technique because few alterations in
the input intervals will result in few altera-
tions in the output
Eliminates redundancy and traffic load
Saves energy in the network.
The fusion node can be compromised by mali-
cious attackers which affect the correctness of the
fusion data.
Multiple copies of the same fusion results at the
sink node lead to an increase in the energy level
at the sink node.
Distributed Source Coding
making the coding decisions process works
efficiently separated from the routing pro-
Requires more computational complexity.
Collects some data from joint statistics which is
not an easy task
Code by ordering
Simple technique
Does not present all possible correlations be-
tween sensor nodes
Information Theory Ap-
Analyzes problems in data collection and
processing by multiple sensors .
[1] A Abdelgawad, M Bayoumi, Resource-Aware Data Fusion Al-
gorithms for Wireless Sensor Networks. (Springer US, Boston,
MA, 2012).
[2] S V, C Chandraseka, Energy Efficient Multipath Data Fusion
Technique for Wireless Sensor Networks. ACEEE International
Journal on Network Security. 3(8), (2012).
[3] R. Brooks and S. Iyengar, Multi-Sensor Fusion: Fundamentals
and Applications with Software.(1998).
[4] C Siaterlis, B Maglaris, Towards multisensor data fusion for
DoS detection. In Proceedings of the 2004 ACM Symposium on
Applied Computing. ACM Press, Nicosia, Cyprus. 439446
[5] T Bass, Intrusion detection systems and multisensor data fusion.
Commun. ACM 2000. 43, 99-105 (2000).
[6] A Savvides, H Park, H, M Srivastava, The n-hop multilateration
primitive for node localization problems. Mob. Netw. 8(4), 443-
[7] K Maraiya, K Kant, N Gupta, Study of Data fusion in Wireless
Sensor Network. In Proceedings of the International Conference
on Advanced Computing and Communication Technologies
(ACCT 2011), Rohtak, 535-539(2011).
[8] M Liggins, C Chong, I Kadar, M Alford, V Vannicola, S Thomo-
poulos, Distributed fusion architectures and algorithms for tar-
get tracking. In Proceedings of the IEEE. 85(1), 95-107(1997).
[9] H Mitchell,. Multi-Sensor Data Fusion: An Introduction.
(Springer, NY, 2007), pp. 3844.
[10] J Raol, Multi-Sensor Data Fusion with MATLAB. (CRC Press,
New York, 2010), pp. 2122.
[11] P Neves, J Rodrigues, K Lin, Data fusion on wireless sensor and
actuator networks powered by the zensens system. Communi-
cations, IET 2011. 5(12), 1661-1668(2011).
[12] H Durrant-Whyte, Sensor models and multisensor integration.
Int. J. Rob. Res. 7(6), 97-113(1988).
[13] R Willett, A Martin, R Nowak. Backcasting: adaptive sampling
for sensor networks. In Proceedings of the 3rd international
symposium on Information processing in sensor networks,
Berkeley, California, USA, (2004).
[14] V Gupta, R Pandey, Data fusion and topology control in wire-
less sensor networks. WSEAS Trans. Sig. Proc. 4(4), 150-
[15] E Nakamura, A Loureiro, A Frery, Information fusion for wire-
less sensor networks: methods, models, and classifications.
ACM Comput. Surv. 39(3), (2007).
[16] B Dasarathy, Sensor fusion potential exploitation-innovative ar-
chitectures and illustrative applications. Proc. IEEE. 85, 24
[17] C Zhao,Y Wang, A new classification method on information fu-
sion of wireless sensor networks. Int. Conf. on Software and
Systems Symposia (ICESS 1008), Sichuan, China, 231236(2008).
[18] V Borges, W Jeberson, Survey of Context Information Fusion
for Sensor Networks based Ubiquitous Systems. In J. Sens. Ac-
tuator Netw. (2013).
[19] T Bayes, R Price, An essay towards solving a problem in the
doctrine of chances. Philosophical Transactions of the Royal So-
ciety. 53, 370418(1763).
[20] H Pan, Z Liang, T Anastasio, T Huang, A hybrid NN-Bayesian
architecture for information fusion. In Proceedings of the 1998
International Conference on Image Processing (ICIP’98), Chica-
go, IL. 1, 368371(1998).
[21] C Cou´E, T Fraichard, P Bessiere, E Mazer, Multi-sensor data fu-
sion using Bayesian programming: An automotive application.
In IEEE/RSJ International Conference on Intelligent Robots and
System, Lausanne, Switzerland. 141146(2002).
[22] B Krishnamachari, S Iyengar, Distributed Bayesian Algorithms
for Fault-Tolerant Event Region Detection in Wireless Sensor
Networks. IEEE Transactions on Computers. 53(3), 241-
[23] G Hartl, B Li, infer: A Bayesian Inference Approach towards En-
ergy Efficient Data Collection in Dense Sensor Networks. In
Proceedings of the 25th IEEE International Conference on Dis-
tributed Computing Systems. 371-380(2005).
[24] A Dempster, A generalization of Bayesian inference. J. Royal Stat.
Soc., Series B. 205247 (1968).
[25] G Shafer, A Mathematical Theory of Evidence. Princeton Uni-
versity Press. ( Princeton, NJ, 1976).
International Journal of Scientific & Engineering Research Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
[26] G Provan, A logic-based analysis of Dempster-Shafer theory. In-
ternational Journal of Approximate Reasoning. 4(5), 451
[27] T Garvey, J Lowrance, M Fischler, An inference technique for
integrating knowledge from disparate sources. In Proceedings of
the 7th international joint conference on Artificial intelligence. (Mor-
gan Kaufmann Publishers Inc., San Francisco, CA, USA, 1981),
pp. 319-325.
[28] A Pinto, J Stochero, J Rezende, Aggregation-aware routing on
wireless sensor networks. In Proceedings of the IFIP TC6 9th In-
ternational Conference on Personal Wireless Communications
(PWC’04). Lecture Notes in Computer Science. Delft, The Neth-
erlands. 238247 (2004).
[29] BYu, K Sycara, J Giampapa, S Owens, Uncertain information fu-
sion for force aggregation and classification in airborne sensor
networks. In AAAI-04 Workshop on Sensor Networks. (AAAI
Press, San Jose, CA, 2004).
[30] S Li, S Son, J Stankovic, Event detection services using data ser-
vice middleware in distributed sensor networks. In Proceedings
of the 2nd international conference on Information processing in
sensor networks (IPSN'03); Zhao, F.; Guibas, L., Eds.; .Springer-
Verlag: Berlin, Heidelberg. 502-517(2003).
[31] D Friedlander, S Phoha, Semantic information fusion for coor-
dinated signal processing in mobile sensor networks. Int. J.
High Perf. Comput. 16, 235241(2002).
[32] D Friedlander, Semantic information extraction. In Distributed
Sensor Networks.;S. S. Iyengar, S.S.; Brooks, R.R., Eds. (CRC
Press, Boca Raton, 2005), pp. 409417.
[33] K Whitehouse, F Zhao, J Liu, Semantic streams: a framework
for composable semantic interpretation of sensor data. In Pro-
ceedings of the Third European conference on Wireless Sensor
Networks, Zurich, Switzerland. (2006).
[34] V Novák, I Perfilieva, J Močkoř, Mathematical Principles of
Fuzzy Logic. The International Series in Engineering and Com-
puter Science. (Kluwer Academic Publishers, Norwell, MA,
[35] L Zadeh, Fyzzy Algorithms. Fuzzy sets, fuzzy logic, and fuzzy
systems: selected papers by Lotfi A. Zadeh. (World Scientific
Publishing Co., Inc., NJ, 1996), pp. 94102.
[36] J Jang, C Sun, E Mizutani, Neuro-Fuzzy and Soft Computing.
(Prentice Hall, New Jersey, 1996)
[37] W Su, T Bougiouklis, Modeling of data fusion algorithms in
cluster-based Wireless Sensor Networks. Signals, Systems and
Computers, 2008 42nd Asilomar Conference on, Pacific Grove,
CA. 868-872(2008).
[38] W Su, T Bougiouklis, Data fusion algorithms in cluster-based
wireless sensor networks using fuzzy logic theory. In Proceed-
ings of the 11th Conference on 11th WSEAS International Con-
ference on Communications (ICCOM'07), World Scientific and
Engineering Academy and Society (WSEAS), Stevens Point,
Wisconsin, USA, 2007; Mastorakis, N.E.; Kartalopoulos, S.; Sim-
ian, D.; Varonides, A.; Mladenov, V.; Bojkovic, Z.; Antonidakis,
E., Eds. 291-299(2007).
[39] W Chan Yet, U Qidwai, Intelligent sensor network for obstacle
avoidance strategy. In Proceedings of IEEE Conference on Sen-
sors. 405408(2005).
[40] X Cui, T Hardin, R Ragade, A Elmaghraby, A swarm-based
fuzzy logic control mobile sensor network for hazardous con-
taminants localization. In Proceedings of the 1st IEEE Interna-
tional Conference on Mobile Ad-hoc and Sensor Systems
(MASS’04). IEEE, Fort Lauderdale. 194–203(2004).
[41] I Gupta, D Riordan, S Sampalli, Cluster-head election using
fuzzy logic for wireless sensor networks. In Proceedings of the
3rd Annual Communication Networks and Services Research
Conference (CNSR’05). IEEE, Halifax, Canada. 255260(2005).
[42] M Yusuf, T Haider, Energy-aware fuzzy routing for wireless
sensor networks. In IEEE International Conference on Emerging
Technologies (ICET’05). IEEE, Islamiabad, Pakistan. 63–69(2005).
[43] J Wallace, D Pesch, S Rea, J Irvine, Fuzzy logic optimisation of
MAC parameters and sleeping duty-cycles in wireless sensor
networks. In 62nd Vehicular Technology Conference, VTC-2005-
Fall.. IEEE, Dallas, TX. 18241828(2005).
[44] L Zadeh, Fuzzy logic: computing with words. Fuzzy Systems,
IEEE Transactions., 4(2), 103111(1996).
[45] P Bonissone, Soft computing: The convergence of emerging rea-
soning technologies. Soft Comput. 1(1), 6-18(1997).
[46] M Roth, Survey of neural network technology for automatic tar-
get recognition. Trans. Neur. Netw. 1(1), 28-43(1990).
[47] T Lewis, D Powers, Audio-visual speech recognition using red
exclusion and neural networks. In Proceedings of the twenty-
fifth Australasian conference on Computer science, Melbourne,
Victoria, Australia.149-156(2002).
[48] C Peirce, Abduction and induction. In Philosophical Writings of
Peirce. (Peirce, C. S.; Buchler, J., Eds.; Dover, New York, 1955),
pp. 150156.
[49] L de Campos, J Gamez, S Moral, Partial abductive inference in
Bayesian belief networks an evolutionary computation ap-
proach by using problem-specific genetic operators. IEEE
Transactions on Evolutionary Computation 2002. 6, 105-131(2002).
[50] R Mooney, Integrating abduction and induction in machine
learning. In Abduction and Induction, Essays on their Relation and
Integration,; Flach, P.A.; Kakas, A.C., Eds.; (Applied Logic Series.
Kluwer: New York, 2000), 336p.
[51] J Aguero, A Vargas, Inference of operative configuration of dis-
tribution networks using fuzzy logic techniquespart II: Ex-
tended real-time model. IEEE Trans. Power Syst. 20(3), 1562
[52] B Bracio, W Hom, D Moller, Sensor fusion in biomedical sys-
tems. In Proceedings of the 19th Annual International Confer-
ence of the IEEE Engineering in Medicine and Biology Society.
IEEE, Chicago, IL. 3, 13871390(1997).
[53] C Brown, H Durrant-Whyte, J Leonard, B Rao, B Steer, Distrib-
uted data fusion using Kalman filtering: A robotics application.
In Data Fusion in Robotics and Machine Intelligence, San Diego,
CA, Abidi, M.A.; Gonzalez, R.C. Eds.( 1992), pp. 267309.
[54] T Schmitt, R Hanek, M Beetz, S Buck, B Radig, Cooperative
probabilistic state estimation for vision-based autonomous mo-
bile robots. IEEE Trans. Robotics Autom. 18(5), 670684(2002).
[55] Y Yuan, M Kam, Distributed decision fusion with a random-
access channel for sensor network applications. IEEE Trans. In-
str. Meas., 53(4), 1339-1344?( 2004).
[56] D Crisan, A Doucet, A survey of convergence results on particle
filtering methods for practitioners. IEEE Transactions on Signal
Processing. 50(3), 736-746(2002).
[57] P Nordlund, F Gunnarsson, F Gustafsson, Particle filters for po-
sitioning in wireless networks. In Proceedings of the XI Europe-
International Journal of Scientific & Engineering Research Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
an Signal Processing Conference (EURSIPCO'02). TeSA, Tou-
louse, France, 311-314(2002).
[58] M Bolic, Architectures for Efficient Implementation of Particle
Filters. Ph.D. Dissertation, State University of New York at
Stony Brook, Stony Brook, NY, USA, Advisor(s) Petar M. Djuric.
AAI3149104.( 2004)
[59] J Aslam, Z Butler, F Constantin, V Crespi, G Cybenko, D Rus,
Tracking a moving object with a binary sensor network. In Pro-
ceedings of the 1st international conference on Embedded net-
worked sensor systems, Los Angeles, California, USA. (2003).
[60] Y Wong, J Wu, L Ngoh, W Wong, Collaborative Data Fusion
Tracking in Sensor Networks using Monte Carlo Methods. In
Proceedings of the 29th Annual IEEE International Conference
on Local Computer Networks. 563-564(2004).
[61] L Hu, D Evans, Localization for mobile sensor networks. In Pro-
ceedings of the 10th annual international conference on Mobile
computing and networking, Philadelphia, PA, USA, (2004).
[62] M Rabbat, R Nowak, Distributed optimization in sensor net-
works. In Proceedings of the 3rd international symposium on
Information processing in sensor networks, Berkeley, Califor-
nia, USA. (2004).
[63] C Guestrin, P Bodik, R Thibaux, M Paskin, S Madden, Distrib-
uted regression: an efficient framework for modeling sensor
network data. In Proceedings of the 3rd international symposi-
um on Information processing in sensor networks. Berkeley,
California, USA. (2004).
[64] S Santini, K R¨omer, An adaptive strategy for quality-based data
reduction in wireless sensor networks. In Proceedings of the 3rd
International Conference on Networked Sensing Systems
(INSS). TRF, Chicago, IL. 29-36(2006).
[65] R Kalman, A new approach to linear filtering and prediction
problems. Trans. ASME J. Basic Engin. 82, 35-45(1960).
[66] R Luo, M Kay, Data fusion and sensor integration: State-of-the-
art 1990s. In Data Fusion in Robotics and Machine Intelligence,;
Abidi, M.A.; Gonzalez, R.C., Eds. (Academic Press, Inc., San Di-
ego, CA, 1992), pp. 7-135.
[67] G Welch, G Bishop, An Introduction to the Kalman Filter. Uni-
versity of North Carolina at Chapel Hill, Chapel Hill, NC,
[68] S Julier, J Uhlmann, A new extension of the Kalman filter to non-
linear systems. In Signal Processing, Sensor Fusion, and Target
Recognition VI. SPIE, San Diego. 182193(1997).
[69] B Sinopoli, L Schenato, M Franceschetti, K Poolla, M Jordan, S
Sastry, Kalman filtering with intermittent observations. IEEE
Trans. Autom. Cont.. 49, 1453-1464(2004).
[70] T Li, A Ekpenyong, Y Huang, Source Localization and Tracking
Using Distributed Asynchronous Sensors. IEEE Transactions on
Signal Processing. 54, 3991-4003(2006).
[71] A Jain, E Chang, Y Wang, Adaptive stream resource manage-
ment using Kalman Filters. In Proceedings of the 2004 ACM
SIGMOD international conference on Management of data. Par-
is, France. (2004).
[72] C Mascolo, M Musolesi, SCAR: context-aware adaptive routing
in delay tolerant mobile sensor networks. In Proceedings of the
2006 international conference on Wireless communications and
mobile computing. Vancouver, British Columbia, Canada.
[73] L Xiao, S Boyd, S Lall, A scheme for robust distributed sensor
fusion based on average consensus. In Proceedings of the 4th in-
ternational symposium on Information processing in sensor
networks, Los Angeles, California. (2005).
[74] L Xiao, S Boyd, SLall, A space-time diffusion scheme for peer-to-
peer least-squares estimation. In Proceedings of the 5th interna-
tional conference on Information processing in sensor networks,
Nashville, Tennessee, USA. (2006).
[75] R Nowak, Distributed em algorithms for density estimation and
clustering in sensor networks. IEEE Trans. Sig. Proc. 51, 2245-
[76] D Blatt, A Hero, Distributed maximum likelihood estimation
for sensor networks. In Proceedings of the IEEE International
Conference on Acoustics, Speech, and Signal Processing
(ICASSP’04). IEEE, Montreal, Canada. 929–932(2004).
[77] L Fang, W Du, P Ning, A beacon-less location discovery scheme
for wireless sensor networks. In Proceedings of the 24th Annual
Joint Conference of the IEEE Computer and Communications
Societies (INFOCOM). 161171(2005).
[78] C Chen, A Ali, H Wang, Design and testing of robust acoustic
arrays for localization and enhancement of several bird sources.
In Proceedings of the 5th international conference on Infor-
mation processing in sensor networks, Nashville, Tennessee,
USA. (2006).
[79] G Hartl, B Li, Loss inference in wireless sensor networks based
on data aggregation. In Proceedings of the 3rd international
symposium on Information processing in sensor networks,
Berkeley, California, USA. (2004).
[80] S Smith, The scientist and engineer's guide to digital signal pro-
cessing. (California Technical Publishing, San Diego, CA, 1997).
[81] C Yang, S Bagchi, W Chappell, Location tracking with direction-
al antennas in wireless sensor networks. In 2005 IEEE MTT-S In-
ternational Microwave Symposium Digest. IEEE, Long Beach,
CA. (2005).
[82] J Polastre, J Hill, D Culler, Versatile low power media access for
wireless sensor networks. In Proceedings of the 2nd interna-
tional conference on Embedded networked sensor systems. Bal-
timore, MD, USA. (2004).
[83] J Blumenthal, D Timmermann, C Buschmann, S Fischer, J Ko-
berstein, N Luttenberger, Minimal transmission power as dis-
tance estimation for precise localization in sensor networks. In
Proceedings of the 2006 international conference on Wireless
communications and mobile computing, Vancouver, British Co-
lumbia, Canada. (2006).
[84] L Gu, D Jia, P Vicaire, T Yan, L Luo, A Tirumala, Q Cao, T He, J
Stankovic, T Abdelzaher, B Krogh, Lightweight detection and
classification for wireless sensor networks in realistic environ-
ments. In Proceedings of the 3rd international conference on
Embedded networked sensor systems, San Diego, California,
USA. (2005).
[85] I Rhee, A Warrier, M Aia, J Min, Z-MAC: a hybrid MAC for
wireless sensor networks. In Proceedings of the 3rd internation-
al conference on Embedded networked sensor systems. San Di-
ego, California, USA. (2005).
[86] A Hoang, M Motani, Collaborative Broadcasting And Compres-
sion In Cluster-Based Wireless Sensor networks. In Proceedings
of the Second European Workshop on Wireless Sensor Net-
works (EWSN'05). IEEE. Istanbul, Turkey. 197-206(2005).
[87] Z Xiong, A Liveris, S Cheng, Distributed source coding for sen-
International Journal of Scientific & Engineering Research Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
sor networks. IEEE Sig. Proc. Mag. 21(5), 80-94(2004).
[88] E Nakamura, A Loureiro, Information fusion in wireless sensor
networks. In Proceedings of the 2008 ACM SIGMOD interna-
tional conference on Management of data (SIGMOD '08). ACM,
New York, NY, USA. 1365-1372(2008).
[89] S Pradhan, K Ramchandran, Distributed source coding using
syndromes (DISCUS): design and construction. IEEE Transac-
tions on Information Theory. 49, 626-643(2003).
[90] R Critescu et al. Networked slepian-wolf: theory, algorithms,
and scaling laws. IEEE Trans. on Inform. Theory, 51, 4057-
[91] D Slepian, J Wolf, Noiseless encoding of correlated information
sources. IEEE Trans. Information Theory IT-19. 471-480(1973).
[92] D Marco, D Neuhoff, Reliability vs. efficiency in distributed
source coding for field-gathering sensor networks. In Proceed-
ings of the 3rd international symposium on Information pro-
cessing in sensor networks, Berkeley, California, USA. (2004).
[93] D Petrovic, R Shah, K Ramchandran, J Rabaey, Data funneling:
routing with aggregation and compression for wireless sensor
networks. in Sensor Network Protocols and Applications, 2003. Pro-
ceedings of the First IEEE 2003 IEEE International Workshop on.
[94] H Ju, L Cui, EasiPC: A Packet Compression Mechanism for Em-
bedded WSN. In Proceedings of the 11th IEEE International
Conference on Embedded and Real-Time Computing Systems
and Applications, Hong Kong, China. 394-399(2005).
[95] S Pattem, B Krishnamachari, R Govindan, The impact of spatial
correlation on routing with compression in wireless sensor net-
works. In Proceedings of the 3rd international symposium on
Information processing in sensor networks, Berkeley, California,
USA. (2004).
[96] J Kulik, W Heinzelman, H Balakrishnan, Negotiation-based
protocols for disseminating information in wireless sensor net-
works. Wireless Networks. 169-185(2002).
[97] C Intanagonwiwat, R Govindan, D Estrin, Directed diffusion: a
scalable and robust communication paradigm for sensor net-
works. In Proceedings of the 6th annual international confer-
ence on Mobile computing and networking. Boston, Massachu-
setts, USA. 56-67(2000).
[98] C Intanagonwiwat, R Govindan, D Estrin, J Heidemann, F Silva,
Directed diffusion for wireless sensor networking. IEEE/ACM
Transactions on Networking (TON). 11(1), 2-16(2003).
[99] A Woo, T Tong, D Culler, Taming the underlying challenges of
reliable multihop routing in sensor networks. In Proceedings of
the 1st international conference on Embedded networked sensor
systems. Los Angeles, California, USA. (2003).
[100] B Krishnamachari, D Estrin, S Wicker, The Impact of Data Ag-
gregation in Wireless Sensor Networks. In Proceedings of the
22nd International Conference on Distributed Computing Sys-
tems. 575-578(2002).
[101] A Boulis, S Ganeriwal, M Srivastava, Aggregation in sensor
networks: An energy-accuracy trade-off. Ad Hoc Networks. 1,
[102] T He, B Blum, J Stankovic, T Abdelzaher, AIDA: Adaptive ap-
plication independent data aggregation in wireless sensor net-
work. ACM Trans. Embed. Comput. Syst. 3(2), 426457(2004). Spe-
cial issue on Dynamically Adaptable Embedded Systems.
[103] C Intanagonwiwat, D Estrin, R Govindan, J Heidemann, Impact
of Network Density on Data Aggregation in Wireless Sensor
Networks. In Proceedings of the 22 nd International Conference
on Distributed Computing Systems (ICDCS'02). 457-458(2002).
[104] P Varshney, Distributed Detection and Data Fusion. (Springer-
Verlag New York, Inc., Secaucus, NJ, 1996).
[105] H Poor, An introduction to signal detection and estimation , 2nd
ed. (Springer-Verlag New York, Inc., New York, NY, 1994).
[106] M Ahmed, G Pottie, Fusion in the context of information theory.
In Distributed Sensor Networks, ; Iyengar, S.S.; Brooks, R.R., Eds.;
CRC Press: Boca Raton. 419-436(2005).
[107] R Gallager, Information Theory and Reliable Communication.
(John Wiley & Sons, Inc., New York, NY, 1968).
[108] K Marzullo, Tolerating failures of continuous-valued sensors.
ACM Transactions on Computer Systems (TOCS). 8(4), 284-
[109] K R¨Omer, P Blum, L Meier, Time synchronization and calibra-
tion in wireless sensor networks. In Handbook of Sensor Net-
works: Algorithms and Architectures, I. Stojmenovic, Ed.( John
Wiley & Sons, Hoboken, NJ, 2005), pp. 199--237.
[110] P Chew, K Marzullo, Masking failures of multidimentional sen-
sors. In Proceedings of the 10th Symposium on Reliable Distrib-
uted Systems. IEEE, Pisa, Italy. 32-41(1991).
[111] U Schmid, K Schossmaier, How to reconcile fault-tolerant inter-
val intersection with the Lipschitz condition. Distributed Compu-
ting. 14, 101-111(2001).
[112] A Elfes, Using Occupancy Grids for Mobile Robot Perception
and Navigation. Computer. 22(6), 46-57(1989).
[113] M Ribo, A Pinz, A comparison of three uncertainty calculi for
building sonar-based occupancy grids. Robotics and Autonomous
Systems. 35, 201209(2001).
[114] A Hoover, B Oslen,. Sensor network perception for mobile ro-
botics. In Proceedings of the IEEE International Conference on
Robotics and Automation. San Fransico, California, USA. 342-
[115] C Wongngamnit, D Angluin, Robot localization in a grid. In-
formation Processing Letters. 77, 5-6(2001).
[116] D Pagac, E Nebot, H Durrant-Whyte, An evidential approach to
map-building for autonomous vehicles. IEEE Trans. Robotics Au-
tom. 14(4), 623629(1998).
[117] Y Zhao, R Govindan, D Estrin, Residual energy scans for moni-
toring wireless sensor networks. In Proceedings of the IEEE
Wireless Communications and Networking Conference
(WCNC'02), IEEE, Orlando, FL. 356362(2002).
[118] D Koks, S Challa, An introduction to bayesian and dempster-
shafer data fusion. Defence Science and Tech Org. (2003).
[119] A Ibrahim, Fuzzy Logic for Embedded Systems Applications.
(Butterworth-Heinemann: Newton, MA, USA. 2003).
[120] W Sung, C Hsiao, IHPG algorithm for efficient information fu-
sion in multi-sensor network via smoothing parameter optimi-
zation. Informatica. 24(2), 219-230(2013).
[121] P Castelaz, Neural networks in defense applications. In Pro-
ceedings of the IEEE International Conference on Neural Net-
works, IEEE, San Diego, CA. 473-480(1988).
[122] W Sung, M Tsai, Multi-sensor wireless signal aggregation for
environmental monitoring system via multi-bit data fusion. Ap-
plied Mathematics&Information Sciences. 5(3), 589-603(2011).
[123] Y Zeng, J Zhang, J Genderen, Comparison and analysis of re-
mote sensing data fusion techniques at feature and decision lev-
International Journal of Scientific & Engineering Research Volume 6, Issue 4, April-2015
ISSN 2229-5518
IJSER © 2015
els. In ISPRS 2006 : ISPRS mid-term symposium 2006 remote
sensing : from pixels to processes. (2006).
[124] F Castanedo, A Review of Data Fusion Techniques. The Scientific
World Journal. Article ID 704504. 19 pages(2013).
[125] M Ghahroudi, R Sabzevari, Multisensor Data Fusion Strategies
for Advanced Driver Assistance Systems. In Sensor Data Fusion;
I-Tech Education and Publishing KG. 141-166(2009).
[126] S Ganeriwal, R Kumar, M Srivastava, Timing-sync protocol for
sensor networks. In Proceedings of the 1st international confer-
ence on Embedded networked sensor systems, Los Angeles,
California, USA. (2003).
[127] M Paskin, S Thrun, Robotic mapping with polygonal random
fields. In Proceedings of the Conference on Uncertainty in Arti-
ficial Intelligence. 450458(2005).
[128] L Li, F Bai, Analysis of Data Fusion in Wireless Sensor Net-
works. In Proceedings of the 2011 international conference on
Electronics, Communications and Control (ICECC). 2547-
[129] B Khaleghi, A Khamis, F Karray, S Razavi, Multisensor data fu-
sion: A review of the state-of-the-art. Inf. Fusion. 14(1), 28-
[130] B Khaleghi, S Razavi, A Khamis, F Karray, M Kamel, Multisen-
sor data fusion: Antecedents and directions. Signals, Circuits
and Systems (SCS), 2009 3rd International Conference on,
Medenine. 6-8(2009).
[131] Y Zhu, E Song, J Zhou, Z You, Optimal dimensionality reduc-
tion of sensor data in multisensor estimation fusion. IEEE Trans-
actions on Signal Processing. 53(5), 1631-1639(2005).
[132] L Dong, F Naumann, Data fusion: resolving data conflicts for
integration. Proc. VLDB Endow. 2(2), 1654-1655(2009).
... Nevertheless, the fundamental principle is that the final prediction should be generated based on the perceived situational knowledge [11]. If multiple forecasts are fully exploited, then more valuable predictions can be obtained at the decision level [12]. ...
Full-text available
Stock market prediction based on machine or deep learning is an essential topic in the financial community. Typically, models with different structures or initializations provide different forecasts of the same response variable. In such cases, better prediction is often achieved by combining forecasts from multiple models rather than using a single model in isolation. This combination of forecasts from the base learners is known as decision fusion. Furthermore, although decision fusion is typical and essential for making the best possible use of multiple forecasts, few studies have systematically summarized the studies that apply this technique. Therefore, there is an urgent need for a literature review reflecting the application of decision fusion in this field. To this end, this study systematically reviewed research related to decision fusion for stock market prediction, focusing on the characteristics of base learners and decision fusion methods. Specifically, the research trend on this topic, which has shifted over the past two decades, is discussed. This review also presents future directions in applying decision fusion to stock market prediction, such as the fusion of forecasts with different data types, using new algorithms as base learners, and integrating sentiment analysis with decision fusion techniques.
... As demonstrated by the relevant literature, real case studies, and applications, the latter are widely adopted for IoT data fusion, especially in the presence of the Big Data challenges. Other relevant data fusion taxonomies and classifications [1] are briefly summarised in Table 2. ...
Full-text available
Abstract The Internet of Things (IoT) facilitates creation of smart spaces by converting existing environments into sensor-rich data-centric cyber-physical systems with an increasing degree of automation, giving rise to Industry 4.0. When adopted in commercial/industrial contexts, this trend is revolutionising many aspects of our everyday life, including the way people access and receive healthcare services. As we move towards Healthcare Industry 4.0, the underlying IoT systems of Smart Healthcare spaces are growing in size and complexity, making it important to ensure that extreme amounts of collected data are properly processed to provide valuable insights and decisions according to requirements in place. This paper focuses on the Smart Healthcare domain and addresses the issue of data fusion in the context of IoT networks, consisting of edge devices, network and communications units, and Cloud platforms. We propose a distributed hierarchical data fusion architecture, in which different data sources are combined at each level of the IoT taxonomy to produce timely and accurate results. This way, mission-critical decisions, as demonstrated by the presented Smart Healthcare scenario, are taken with minimum time delay, as soon as necessary information is generated and collected. The proposed approach was implemented using the Complex Event Processing technology, which natively supports the hierarchical processing model and specifically focuses on handling streaming data ‘on the fly’—a key requirement for storage-limited IoT devices and time-critical application domains. Initial experiments demonstrate that the proposed approach enables fine-grained decision taking at different data fusion levels and, as a result, improves the overall performance and reaction time of public healthcare services, thus promoting the adoption of the IoT technologies in Healthcare Industry 4.0.
... A data fusion architecture for the IoT ecosystem can be naturally identified by matching IoT infrastructure capabilities with the existing data fusion models [1]. On this premise, it is therefore natural to think on how to exploit edge devices and network nodes in supporting the IoT data fusion tasks. ...
Conference Paper
The Internet of Things (IoT) has embraced a 'vertical' off-loading model, where avalanches of raw data generated by numerous edge devices are continuously pushed through the network to a remote processing location, such as a datacenter or a cloud. In this rather unbalanced architecture, edge devices are typically not expected to perform sophisticated data processing and analytics, and data fusion takes place remotely from the original source of data. As a result, the underlying network and the remote datacenter have to handle increased amounts of unstructured raw data, which, in turn, may affect the overall performance and decrease reaction times. As a potential solution to these shortcomings, this paper introduces a distributed hierarchical data fusion architecture for the IoT networks, consisting of edge devices, network and communications units, and cloud platforms. According to the proposed approach, different data sources are combined at each level of the IoT hierarchy to produce timely and accurate results by utilising computational capabilities of intermediate nodes. This way, mission-critical decisions, as demonstrated by the presented smart healthcare scenario, are taken with minimum time delay, as soon as necessary information is generated and collected. The initial evaluation suggests that the proposed approach enables fine-grained decision taking at different data fusion levels, and, as a result, improves the overall performance and reaction time.
... Following data related challenges [22,23] needs to be addressed,  Data imperfection: The data is acquired from multiple sources, which leads to a rise in the imperfections in the measurement procedure. This can be avoided by reducing the data redundancy. ...
Full-text available
The primary objective of this research is to make use of diverse information such as textual, audio, and visual to learn a model. Effective interaction between these approaches often leads to a better-performing system. In addition, a new fusion taxonomy based on a data perspective introduces a fusion-based classification model that can neglect the context domain and focus on several features of the input data. The proposed fusion taxonomy is constructed based on the data perspective not on context perspective. This interpretation relies on data modality types, the importance of extracted features, the data noise, and the streaming time. It improves an interpretation of the hybrid fusion technique. The fundamental challenge of the smart environment is illustrated by infusing big sensory data extracted from IoT sensors and devices to support the main objective. Also, an adaptive smart environment multi-modal system is proposed as a solution for the modality fusion challenge to improve classification and prediction in various contexts, whether one or more data modality types. We have proposed a new adaptive smart environment multi-modal system (ASEMMS) for improving the classifications accuracy results. It relies on the common characteristics of smart applications such as smart health for monitoring patients remotely and smart military for improving the hyper spectral in night mode. The adaptive smart environment multi-modal system (ASEMMS) is designed based on constructing five layers, a software-defined fusion layer, pre-processing layer, dynamic classification layer, hybrid fusion layer, and evaluation layer. A software-defined fusion layer is considered a controller for managing data types, model types, and noisy data. A hybrid fusion layer is designed based on a tailored neural network for making a combination between Dempster-Shafer and Concatenation fusion techniques for getting bigger number of features. It measures the accuracy and optimization for the classification results. For validation, we make two comparisons between the proposed adaptive system and the baseline of Dempster-Shafer fusion technique and the baseline of concatenation fusion technique.
this paper presents a new taxonomy for the context-awareness problem in data fusion. It interprets fusing data extracted from multiple sensory datatypes like images, videos, or text. Any constructed smart environment generates big data with various datatypes which are extracted from multiple sensors. This big data requires to fuse with expert people due to the context-awareness problem. Each smart environment has specific characteristics, conditions, and roles that need to expert human in each context. The proposed taxonomy tries to cure this problem by focusing on three dimensions classes for data fusion, types of generated data, data properties as reduction or noisy data, or challenges. It neglects the context domain and introduces solutions for fusing big data through classes in the proposed taxonomy. This taxonomy is presented from studying sixty-six research papers in various types of fusion, and different properties of data fusion. This paper presents new research challenges of multi-data fusion.
Initially, the context is established for considering distributed networks as efficient information processing entities (Section 22.2). Next, in Section 22.3, the approaches taken towards analyzing such systems and the path leading towards the modern information theoretic framework for information processing are discussed. The details of the mathematical method are highlighted in Section 22.4, and applied specifically for the case of multi-sensor systems in Section 22.5. Finally, our conclusions are presented in Section 22.6.
Sensor networks generally observe systems that are too complex to be simulated by computer models based directly on their physics. We therefore use a semi-empirical model based on time series measurements. These systems can be stochastic, but not necessarily stationary. In this case, we make a simplifying assumption, that the time scale for changes in the equations of motion is much longer than the time scale for changes in the dynamical variables. The system is then in semi-equilibrium, so its dynamics can be determined from a data sample whose duration is long compared with changes in the dynamical variables but short compared with changes in the dynamical equations.
Communication Systems and Information Theory. A Measure of Information. Coding for Discrete Sources. Discrete Memoryless Channels and Capacity. The Noisy-Channel Coding Theorem. Techniques for Coding and Decoding. Memoryless Channels with Discrete Time. Waveform Channels. Source Coding with a Fidelity Criterion. Index.
Automatic speech recognition (ASR) performs well under restricted conditions, but performance degrades in noisy environments. Audio-Visual Speech Recognition (AVSR) combats this by incorporating a visual signal into the recognition. This paper briefly reviews the contribution of psycholinguistics to this endeavour and the recent advances in machine AVSR. An important first step in AVSR is that of feature extraction from the mouth region and a promising new technique is presented. This paper examines how useful this extraction technique in combination with several integration architectures and compares it with competing techniques, demonstrates that vision does in fact assist speech recognition when used in a linguistically guided fashion, and gives insight into remaining issues.
As its name suggests, computing with words, CW, is a methodology in which words are used in place of numbers for computing and reasoning. The point of this note is that fuzzy logic plays a pivotal role in CW and vice-versa. Thus, as an approximation, fuzzy logic may be equated to CW. There are two major imperatives for computing with words. First, computing with words is a necessity when the available information is too imprecise to justify the use of numbers. And second, when there is a tolerance for imprecision which can be exploited to achieve tractability, robustness, low solution cost and better rapport with reality. Exploitation of the tolerance for imprecision is an issue of central importance in CW. In CW, a words is viewed as a label of a granule, that is, a fuzzy set of points drawn together by similarity, with the fuzzy set playing the role of a fuzzy constraint on a variable. The premises are assumed to be expressed as propositions in a natural language. For purposes of computation, the propositions are expressed as canonical forms which serve to place in evidence the fuzzy constraints that are implicit in the premises. Then, the rules of inference in fuzzy logic are employed to propagate the constraints from premises to conclusions. At this juncture, the techniques of computing with words underlie -- in one way or another -- almost all applications of fuzzy logic. In coming years, computing with words is likely to evolve into a basic methodology in its own right with wide-ranging ramifications on both basic and applied levels.
Fuzzy Logic for Embedded Systems Applications, by a recognized expert in the field, covers all the basic theory relevant to electronics design, with particular emphasis on embedded systems, and shows how the techniques can be applied to shorten design cycles and handle logic problems that are tough to solve using conventional linear techniques. All the latest advances in the field aree discussed and practical circuit design examples presented.Fuzzy logic has been found to be particularly suitable for many embedded control applications. The intuitive nature of the fuzzy-based system design saves engineers time and reduces costs by shortening product development cycles and making system maintenance and adjustments easier. Yet despite its wide acceptanceâ??and perhaps because of its nameâ?it is still misunderstood and feared by many engineers. There is a need for embedded systems designersâ??both hardware and softwareâ?to get up to speed on the principles and applications of fuzzy logic in order to ascertain when and how to use them appropriately. Fuzzy Logic for Embedded Systems Applications provides practical guidelines for designing electronic circuits and devices for embedded systems using fuzzy-based logic. It covers both theory and applications with design examples.