You are on page 1of 9

Determination of Optimized Data Fusion

Algorithms for Radars Network by Ordered


Weighted Averaging
Ali Naseri and Omid Azmoon
AbstractDecision making is a fundamental aim of radars network and it is depends on data fusion algorithm. Probability of
detection is an important parameter, because of uncertainty, for data fusion algorithm and it is depends on radar gating and
swerling of targets. In this paper data of 2-cell and 4-cell network are fused by custom algorithms in radars network, Averaging,
Bayesian and Dempster-Shafer. Results of simulations are compared by convergence, precision, influence of fluctuations, time
of running and complexity of implementation and are evaluated by fuzzy logic operator, OWA. Simulations are declare
Dempster-Shafer algorithm is optimized for 2-cell network, and if radars outputs are mass function, centralized topology by
Dempster-Shafer algorithm will be optimized in 4-cell network, else if radars outputs are probability, because of inverse Pignistic
transform in input and Pignistic transform in output of every cells, centralized topology by Bayesian algorithm will be optimized.
Index Terms Averaging, Bayesian, Data Fusion, Dempster-Shafer, OWA, Radars Network.



1 INTRODUCTION
IM of radars networking is information fusion and
refuse unwanted parameters [1]. The received data
from radars fuse by weighted coefficients that come
from uncertainty [2]. Ambient come from radars
specifications, environmental parameters and targets
characteristics [3]. Some algorithms have presented for
fusion of information in sensor networks [2]. In radars
network, Statistical, Bayesian and Evidence reasoning are
custom algorithms that usually use for fusion of data, that
is depends on demands and limits [1]. Convergence,
precision, influence of fluctuations, time of running and
complexity of implementation are important criterions for
selecting algorithm [4]. For determination of optimized
algorithm, all criterions must be attended. Some
algorithms have presented for evaluation [5]. In this
article, OWA (Ordered Weighting Averaging) operator,
based on fuzzy logic, is used. OWA converts qualitative
elements to quantitative parameters by weighting the
elements and advantages of algorithms have more effect
and weaknesses of them have also paid attention [6]. In
continue, important concepts of fusion and data
extraction in radar have described. In section 2 algorithms
and composite of them have expressed, and simulated in
section 3. Results and evaluation by OWA have presented
in section 4.
1.1 Data Fusion
Sensor relations and fusion methods have categorized in
different context. Four fusion types have classified, across
sensors, attributes, domains, and time. In addition, three
configuration methods for sensors management have
classified; complementary, cooperative and competitive.
There are three models of topologies; centralized,
decentralized and hierarchical, these are illustrated in Fig
1. Some models of data fusion have also developed, that
commonly have three levels, signal or pixel, feature and
decision. Choosing model depends on processing power
and aim of networking [2, 7].




(a)





(b)







(c)
Fig. 1. Topologies of Data Fusion, (a) Centralized, (b)
Decentralized, (c) Hierarchical.
1.2 Detection in Radar
Backscattered signals, these are measured by SNR (Signal
to Noise Ratio), contain white noises (false alarm
probability (p
Ia
)) and noise from clatter (constant false
alarm ratio (p
CFAR
)), that are minimized or removed by
some methods like integration of received pulses [3].

Ali Naseri is with the Department of Information and Communications
Technology, IHU University, Tehran, Iran.
Omid Azmoon is with the Department of Information and
Communications Technology, IHU University, Tehran, Iran.

A
Radar1
Data Fusion
Data
Fusion
in Second
Layer
Data Fusion in First Layer
Radar2
Radar3
Radar4
Radar1 Radar2
Radar4
Radar3
Radar1
Radar2
Radar3
Radar4
Data Fusion in First Layer
JOURNAL OF COMPUTING, VOLUME 4, ISSUE 1, JANUARY 2012, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG 12

Color noises generate in network too, that attempts have
made to remove them like special Kalman filters [8].
Nevertheless, there is difference between posterior and
measured data. If the measured distance and angle are R
m

and
m
, and the estimated distance and angle are R
p

and
p
, and
R
2
and

2
are variances of (R
p
- R
m
) and
(
p
-
m
) respectively, normalized gating will be:
u
2
= (R
p
- R
m
)
2

R
2
+(
p
-
m
)
2

2
, (1)
If Maximum of normalized gating is G, it must be d<G. If
coordinates have M dimensions and every components
change by Gaussian probability density function, sum of
M components will equal to Chi-square probability
density function by M degrees of freedom (
M
2
).
According [8] probability of observation of targets in
ellipsoidal gate, when coordinates have two dimensions,
is equal to:
p
G
= 1 -e
-G 2
(2)
Ambiguous in data also depends on fluctuations of radar
cross section (RCS) and angle of aspect of targets. Four
models for swerling of targets are introduced that are
described in Table 1. In addition, the non-fluctuating
model is called swerling 0. Swerling depends on PRF
(pulse repetition frequency), SNR, p
Ia
and p
CFAR
[3, 9, 10].
TABLE 1
Models of Swerling (SW=Swerling)
Probability Density
Function
Fluctuations
of RCS
Receive Pulse
Independence
SW
1
c
(-c c
c
) ,
o
u
,
High /
Slowly
Scan to Scan
SW
2
(no 2o
av
2
) c
(-nc
2
4 c
av
2
) ,

High /
Rapidly
Pulse to Pulse
SW
3
(4o o
av
2
) c
(-2c c
c
)

Low / Slowly Scan to Scan
SW
4
(o
4
o
3
2o
av
4
, )c
(-u
2
c
2
2c
av
2
) ,

Low /
Rapidly
Pulse to Pulse
1.3 OWA Operator
Vector X=( x
1
,, x
n
) is considered to be real numbers and
every member is one of the criterions. If the members of X
have sorted in ascending to descending order and
member in new position have named x
1
i
, OWA function
is:
0WA(X) = w
I
x
I
i n
I=1
,
w
I
= 1
n
I=1
w
I
= Q[
I
n
-Q[
I-1
n
, i = 1, , n
Q(t) =
] (-y+1)dy
t
0
] (-y+1)dy
1
0
(3)
w
i
is the weight factor and Q is the quantifier that can be
other functions by defined [6].
1.4 Related Works
Comparison between fusion methods depends on aim
and level of model. In [11], at pixel level and for satellite
pictures, fusion methods have compared. In [12],
Bayesian and Dempster-Shafer algorithms have
compared and author has shown results of algorithms for
decision level that is nearby similar. In [13], authors
compare two custom algorithms at feature level and in
human activity, and in [14] at decision level and in
landmine detection. In [15], at feature and decision levels
for remote sensing, three algorithms, Bayesian, Dempster-
Shafer and Neural Network have compared and survey
their advantages and limits.
Comparison in radars network depends on level, too.
Bieker in [16], compare fusion algorithms, Bayesian,
Dempster-Shafer and voting, at feature level. Another
research is in [17], at feature level and in target
identification two custom algorithms have compared. In
these works, methods have compared but have not
proposed method for evaluation or determination of
optimized algorithm.
In research before this paper [18], three fusion method,
Statistical, Bayesian and Evidence reasoning, have
simulated, compared and evaluated in signal level. For
evaluation and determination of optimized algorithm,
average of results of comparisons have calculated and it is
declared Dempater-Shafer algorithm is optimized for 2-
cell network and in 4-cell network hierarchical topology
with Dempater-Shafer algorithm in both layers will be
optimize, when radars outputs are mass functions
directly. When radars outputs are probability,
hierarchical topology with Average algorithm in first
layer and Bayesian algorithm in second layer will be
optimize. In this paper, data fused in decision level and
OWA method has proposed for evaluation and
determination of optimized algorithms.
2 DATA FUSION ALGORITHMS
Three algorithms usually use in radars network;
Statistical, Bayesian and Dempster-Shafer reasoning [1].
These algorithms have described for 2-cell and 4-cell
network with centralized and hierarchical topologies. In
this network, fusion type is sensory, configuration is
competitive, and data are fused in signal level. Probability
of detection of target is named E, reported probability in
nth step and ith radar is p
(n,I)
and result of fusion in nth
step is p
(n,I)
.
2.1 Statistical
Statistical methods and Boolean algebra, like majority
voting, have used for data fusion [19]. In this paper, data
are fused by averaging. Fusion of N radars in nth step is
equal to:
p
(n,I)
(E) =
p
(n,)
(E)
N
=1
N
(4)
2-cell fusion is named C and 4-cell fusion with centralized
topology is named C4.
2.2 Bayesian
Bayes equation is [20]:
p(A|B) =
p(B|A)p(A)
p(B|A

)p(A

)
n
=1
(5)
JOURNAL OF COMPUTING, VOLUME 4, ISSUE 1, JANUARY 2012, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG 13

If probability of detection of target is A and probability of
received data is B, probability of detection in fusion node
will be equal to:
P(detection of target|receive data)=
P(receive data|detection of target)P(detection of target)/
{ P(receive data|detection of target)P(detection of target)
+ P(receive data|none detection)P(none detection)} (6)
Fusion of N radars in nth step is equal to:
p
(n,I)
(E) =
p
|
(n,I)
(E)
p
|
(n,I)
(E)+p
|
(n,I)
(E
)
,
p
i
(n,I)
(E) =
{ p
(n,)
(E) ]
N
=1
p
(n-1,I)
(E)
p
(n-1,)
(E)
N
=1

,
p
i
(n,I)
(E

) =
{ p
(n,)
(E

) ]
N
=1
p
(n-1,I)
(E

)
p
(n-1,)
(E
)
N
=1

(7)
And value in first step is:
p
(0,I)
(E) =
p
N
=1 (0,)
(E)
n
, p
(0,I)
(E

) =
p
N
=1 (0,)
(E
)
n

2-cell fusion by Bayesian algorithm is named B and 4-cell
fusion with centralized topology is named B4.
2.3 Dempster-Shafer or Evidence
If A
1
A
2
A
n
are members of (frame of discernment),
has 2
n
subsets. A number between (0,1) is assigned to
them that is called mass function; m ({A
I
, , A
j
]), where
[20]:
m(A) = 1
A
, m(A) > u ,m()=u (8)
If two radars send data from one target, m
1
(A)and m
2
(A),
fused data is calculated by:
m
1
m
2
(A) =
m
1
(B)m
2
(C)
EC=A
1- m
1
(B)m
2
(C)
EC=

(9)
N mass functions are fused by:
m
1
uuum
n
(A) =
[ m
]
(A
]
)
n
]=1

A

=A
n
=1
1- [ m
]
(A
]
)
n
]=1

A

n
=1
=
(10)
For comparison between Bayesian and Dempster-Shafer
results, transformation is needed. Mass function transfers
to probability domain by pignistic transform [21]:
p(x
j
) =
m(A

)
|A

|

x
]
A

(|A
I
|=number of A
I
members) (11)
The custom method for probability to mass function
transformation is inverse pignistic transform [22]. For
= { A
1
, A
2
, , A
n
], monotonic possibility density function
is defined:
( A
k
) = min(p
k
, p
I
)
n
I=1
,
1 ( A
1
) ( A
2
) ( A
n
)
Mass functions are equal to:
m({A
1
]) = |( A
1
) - ( A
2
)|
m({A
1
, , A
I
]) = |( A
I
) -( A
I+1
)|
m() = ( A
n
) (12)
2-cell fusion by Evidence algorithm is named D and 4-cell
fusion with centralized topology is named D4.
2.4 Hierarchical Topologies
Another type of topology in radars network is
hierarchical. Composite algorithms are named by
algorithms these are used in layers. If Statistical is used in
both layers, it will be named CC. According to [23] results
of CC algorithm are equal to C4. If data is fused in both
layers by Bayesian, it will be named BB and by Dempster-
Shafer, it will be named DD. According to [23], if values
of BB and B4 in their first steps are equal, results of
algorithms will equal, and if mass functions in all radars
are simple mass functions, results of algorithms will equal
(if 0 has only one member, mass functions are called
simple). Difference between CC and C4, BB and B4, DD
and D4 is in running time and complexity of
implementation. If first and second layers fusion are
Statistical and Bayesian respectively, it will name CB.
Others will name similarly; for Bayesian-Statistical BC, for
Statistical-Evidence CD, for Evidence-Statistical DC, for
Bayesian-Evidence BD and for Evidence-Bayesian DB.
3 SIMULATION
Probability of detection of radars is simulated by
MATLAB according to radar gating and swerling of
targets [24]. Algorithms also are simulated. Simulated
scenario is surveillance radars signals in L band, range of
target is 150km, velocity is 200 m/s, height is 1000m, RCS
and SNR are 10 m2 and 7dB, false alarm with clatter is 10-
6, radars have 2 dimensions and G=10. Radars
arrangement is serial that means target move with
consistent distance from radars in simulation time.
Duration of simulation is 2 minutes and 25 scans. Every
scan is integration of 16 pulses. First for 2-cell network,
(R1, R2), target is assumed by swerling 0 and radars
detect trajectory of it. Then scenario is repeated by
swerling 1 to 4, respectively. It must be attend swerling 0,
1 and 3 are simulated for 25 scans that every scan is
integration of 16 pulses and are illustrated in figures 2, 3
and 4. Swerling 2 and 4 also are simulated for 400 pulses
and are illustrated in figures 5 and 6. For better view,
initial forty samples are illustrated. Simulation is repeated
for swerling 1 to 4 for five times and averages and
standard deviations of results of simulations are
summarized in table 2. Then scenario is repeated for 4-cell
network (R1, R2, R3, R4). Results of fusion are illustrated
in figures 7 to 11. Simulation is run for swerling 1 to 4 for
five times and averages and standard deviations of
results of simulations are summarized in table 3. Figures
have four parts, two parts in left illustrate probability of
detection of radars and result of fusion. Two parts in right
illustrate result of decision fusion. If probability of
detection is over 0.7 target has certainty detected, and is
below 0.3 has not detected. Else probability between these
borders is uncertainty. It must be attend, in algorithms
JOURNAL OF COMPUTING, VOLUME 4, ISSUE 1, JANUARY 2012, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG 14

with Dempster-Shafer equations, mass functions are
derived from inverse pignistic transform and then data
are fused, results of fusion are transferred to probability
domain by pignistic transform and are illustrated for
comparison.

Fig.2*.Data Fusion in 2-cell Network by Target with Swerling Model 0

Fig.3*.Data Fusion in 2-cell Network by Target with Swerling Model 1

Fig.4*.Data Fusion in 2-cell Network by Target with Swerling Model 2

Fig.5*.Data Fusion in 2-cell Network by Target with Swerling Model 3
JOURNAL OF COMPUTING, VOLUME 4, ISSUE 1, JANUARY 2012, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG 15


Fig.6*.Data Fusion in 2-cell Network by Target with Swerling Model 4

Fig.7**.Data Fusion in 4-cell Network by Target with Swerling Model0


Fig.8**.Data Fusion in 4-cell Network by Target with Swerling Model1

Fig.9**.Data Fusion in 4-cell Network by Target with Swerling Model2

JOURNAL OF COMPUTING, VOLUME 4, ISSUE 1, JANUARY 2012, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG 16


Fig.10**.Data Fusion in 4-cell Network by Target with Swerling
Model3

Fig.11**.Data Fusion in 4-cell Network by Target with Swerling
Model4
*: Left Up- Radars Probability of Detection, Left Down Fusion of
Probabilities of Detection by 2 Algorithms, Right Up Radars
Decision, Right Down Fusion of Decisions by 2 Algorithms
**: Left Up- Radars Probability of Detection, Left Down Fusion of
Probabilities of Detection by BB-BC-BD, Right Up Fusion of
Probabilities of Detection by CC-CB-CD, Right Down Fusion of
Probabilities of Detection by DD-DC-DB
TABLE 2
Results of Five Times Tests of Data Fusion for 2-cell Network
(SD= Standard Deviation, SW=Swerling)
Model Detection R1 R2 C B DS
SW0 Average 19 19 19 21 21
SW1
25Scan
Average 8.2 7.6 4.6 15.6 11.2
SD(%) 1.6 1.4 1.0 2.4 3.0
SW2
400Pulses
Average 170.6 169.6 147.4 332.0 273.4
SD(%) 5.99 14.19 8.98 50.76 8.09
SW3
25Scan
Average 0.0 0.2 0.2 15.8 8.0
SD(%) 0.0 0.4 0.4 7.93 1.27
SW4
400Pulses
Average 166.4 164.6 154.0 305.8 273.4
SD(%) 10.29 7.86 11.6 59.66 8.50
TABLE 3
Results of Five Times Tests of Data Fusion for 4-cell Network
(SD= Standard Deviation, SW=Swerling)
Model Detection R1 R2 R3 R4 CC
SW0 Average 19 19 19 19 19
SW1
25Scan
Average 2.8 7.6 8.4 8.4 7.6
SD(%) 1.94 2.87 2.50 2.06 2.42
SW2
400Pulses
Average 131.8 178.6 177.0 168.2 172.8
SD(%) 6.14 4.88 8.10 12.94 3.18
SW3
25Scan
Average 0 0 0.6 0 0.4
SD(%) 0 0 0.49 0 0.49
SW4
400Pulses
Average 155.0 177.8 168.8 173.6 177.8
SD(%) 7.54 8.16 9.52 7.23 5.34
CB CD BB BC BD DD DB DC
21 23 23 23 21 23 21 21
9.8 18.0 14.8 18.0 13.2 18.2 10.8 14.6
1.17 4.20 1.72 4.98 7.39 5.42 1.72 20.24
296.0 334.0 356.2 338.0 266.2 316.6 304.4 268.8
8.53 40.70 7.11 36.20 92.90 66.10 10.54 63.80
5.6 21.0 19.4 20.8 8.2 21.2 5.6 16.0
1.50 1.10 0.49 0.98 10.09 1.33 1.50 7.54
297.4 383.2 352.6 382.8 224.6 387.6 304.8 360.8
3.61 6.24 2.65 6.52 60.40 7.39 7.03 16.71
4 EVALUATION AND DETERMINATION OF
OPTIMIZED ALGORITHM
Convergence, precious, influence of fluctuations, running
time and complexity of implementation are criterions for
evaluation. Then optimized algorithm is determined by
OWA operator.
4.1 Convergence
Convergence is calculated by:
C
AI
=
1
Sm
C
s

R
]
AI
5
I=1
m
j=1


C
s

R
]
AI
=
`
1
1
1
1
u foi
s

AI

s

R
]

AI
-
s

R
]

max
-
s

R
]
foi
s

R
]
<
s

AI

max
1 foi
max
<
s

AI



(13)
m is the number of radars,
s

R
]
is average of jth radar in ith
swerling,
s

AI
is average of data in ith swerling and
max
is
maximum acceptable average in radars that means if
JOURNAL OF COMPUTING, VOLUME 4, ISSUE 1, JANUARY 2012, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG 17

value is over it, target will detect certainly.
max
is
assumed 20 for swerling 0, 1, 3 and 320 for swerling 2, 4
[23].
4.2 Precious
Precious is calculated by:
P
AI
=
1
4m
P
s

R
]
AI
4
I=1
m
j=1



P
s

R
]
AI
=
`
1
1
1
1
1
1
10o
s

R
]
-2o
s

AI
10o
s

R
]
foi
s

AI

s

R
]
u.8
P
max
-o
s

AI
P
max
-o
s

R
]
foi
s

R
]
< o
s

AI
P
max
u foi P
max
< o
s

AI


14)

R
]
is standard deviation of jth radar in ith swerling,

AI
is standard deviation of data in ith swerling and P
max

is maximum acceptable standard deviation. P
max
is
assumed 5 times of
s

R
]
[23].
4.3 Influence of Fluctuations
Influence of fluctuations is calculated by [23]:

F
AI
=
1
4m
F
s

R
]
AI
4
I=1
m
j=1


Foi C
s

R
]
AI
= u

F
s

R
]
AI
= _
P
s

R
]
AI
C
s

R
]
AI
foi P
s

R
]
AI
, C
s

R
]
AI
1 foi P
s

R
]
AI
> C
s

R
]
AI


Foi C
s

R
]
AI
= u F
s

R
]
AI
= u (15)
4.4 Running Time
Suitable time for process in simulated scenario is 2 to 4
seconds [24]. In same condition of processing for all
algorithms, scenario run for swerling 0 and time is
measured. For evidence algorithm, time of pignistic and
inverse pignistic transform is calculated too. If radars
outputs can be mass functions directly, because of
evidence algorithm, time will be decrease. This condition
for 2-cell network is named Direct Dempster-Shafer or
DDS. For 4-cell network is named DD4 and DDD. Results
for 2-cell and 4-cell network with 10% error are
summarized in table 4 and 5. If i
AI
is running time, Imin
is minimum and Imax is maximum of acceptable time,
expression of running time in percentage is [23]:

I
AI
=
`
1
1
1
1
2I
mn
- I
AI
2I
mn
foi i
AI
I
mIn
u.S
I
max
- I
AI
I
max
-I
mn
foi I
mIn
< i
AI
I
max
u foi I
max
< i
AI



I
min
=2s , I
max
=4s (16)
TABLE 4
Running Time of Algorithms for 2-cell Network (Mili Second)
DDS DS B C
420 910 310 200
TABLE 5
Running Time of Algorithms for 4-cell Network (Mili Second)
BC BB B4 CD CB CC C4
860 720 1020 1550 890 790 460
DD4 DDD DC DB DD D4 BD
580 1260 2440 2310 3110 1910 1680
4.5 Complexity of Implementation
According [17] complexity of implementation is
expressed by sum of addition/subtraction,
multiplication/division, conditions and relations of sets.
Results are summarized in tables 6 and 7. Because of
conditions and relations of sets, calculations of
mathematical operations for evidence algorithms are
approximated. For 2-cell network, sum of row of table 6
are calculated and results are mapped from (0, 1000) to (0,
100%) and for 4-cell network sum of row of table 7 are
calculated and results are mapp`ed from (0, 2000) to (0,
100%). It must be attend running time is not similar with
complexity of implementation, because of relations of sets
for evidence algorithms [23].
TABLE 6
Complexity of Data Fusion Algorithms for 2-cell Network
Multipl-
ication
Divi-
sion
Add-
ition
Subtr-
action
sets Cond-
itions
C 0 25 25 0 0 0
B 144 73 145 125 0 0
DDS ~75 ~0 ~25 ~75 ~50 ~25
DS ~125 ~50 ~75 ~150 ~50 ~125
TABLE 7
Complexity of Data Fusion Algorithms for 4-cell Network
Multipl-
ication
Divi-
sion
Add-
ition
Subtr-
action
sets Cond-
itions
C4
0 28 75 0 0 0
CC
0 75 75 0 0 0
CB
144 27 27 120 0 0
CD
~125 ~100 ~75 ~150 ~50 ~100
B4
350 73 28 225 0 0
BB
432 219 75 375 0 0
BC
288 171 51 250 0 0
BD
~413 ~170 ~75 ~365 ~50 ~100
D4
~700 ~125 ~100 ~700 ~150 ~415
DD
~375 ~150 ~225 ~450 ~150 ~375
DB
~400 ~150 ~150 ~400 ~100 ~200
DC
~250 ~100 ~150 ~250 ~50 ~100
DD4
~550 ~100 ~75 ~550 ~150 ~350
DDD
~250 ~30 ~75 ~225 ~150 ~150
4.6 Evaluation by OWA
Vector X is generated by x1 for convergence, x2 for
precious, x3 for influence of fluctuations, x4 for running
time and x5 for complexity of implementation.
Coefficients are equal to:
w
1
=
9
2S
, w
2
=
7
2S
, w
3
=
S
2S
, w
4
=
S
2S
, w
5
=
1
2S

JOURNAL OF COMPUTING, VOLUME 4, ISSUE 1, JANUARY 2012, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG 18

Results of simulations are summarized in table 8 and 9.
For 2-cell network, Direct Dempster-Shafer (DDS) is
optimized. For 4-cell network, if radars outputs are
probability, B4 will optimized, and if radars outputs are
mass function directly, DD4 will optimized.
Networks with more cells and complicated conditions
must survey independently.
TABLE 8
Evaluation of Algorithms for 2-cell Network
Criterion (%) DDS DS B C
Convergence
61.13 61.13 86.66 0
Precious
59.95 59.95 20.52 69.81
Influence of
Fluctuations
86.58 86.58 28.56 0
Running Time
89.5 77.25 92.25 95
Complexity of
Implementation
75 47.3 51.3 95
OWA
81.20 74.11 71.98 74.76
TABLE 9
Evaluation of Algorithms for 4-cell Network
(CV=Convergence, P=Precious, F=Influence of Fluctuations,
I=Running Time, CI=Complexity of Implementation)
BC C4 CC CB CD B4 BB
CV 56.20 0 0 79.91 65.83 96.53 96.53
P 9.78 72.77 72.77 13.53 61.96 39.45 39.45
F 75.00 0 0 69.71 82.98 76.97 76.97
I 78.50 88.50 80.25 77.75 61.25 74.50 82.00
CI 62.00 95 92.5 84.10 70.00 66.20 44.95
OWA 68.80 73.54 70.32 77.12 72.52 80.72 80.08
BD D4 DD DB DC DDD DD4
CV 96.66 90.72 90.72 96.66 61.99 90.72 90.72
P 46.16 73.23 73.33 46.24 67.26 73.23 73.23
F 77.59 81.35 81.35 77.62 86.69 81.35 81.35
I 58.00 52.25 22.25 42.25 39.00 68.50 85.50
CI 41.35 0 13.75 30.00 55.00 56.00 11.25
OWA 75.32 76.35 73.30 71.25 70.60 80.54 82.11
5 CONCLUSION
In this paper, Averaging, Bayesian and Dempster-Shafer,
custom data fusion algorithms, are simulated. Probability
of detection have generated according to radar gating and
swerling of targets. Then the generated data have fused
by described algorithms in 2-cell network and 4-cell
network by various topologies. Results of simulation are
evaluated by five criterions, convergence, precious,
influence of fluctuations, running time and complexity of
implementation. Optimized algorithm is determined by
OWA operator that is method of fuzzy logic. Results of
OWA declare Dempater-Shafer algorithm is optimized
for 2-cell network. For 4-cell network, if radars outputs
are mass functions directly, centralized topology with
Dempater-Shafer algorithm (DD4) will be optimized, and
when radars outputs are only probability, pignistic
transform is needed for radars outputs and inverse
pignistic transform for radars inputs, then centralized
topology with Bayesian algorithms will optimized (B4).
REFERENCES
[1] P.Valin, .Boss, A.Jouan, Airborne Application of Information Fusion
Algorithms to Classification, Defence R&D Canada - DRDC Valcartier
TR 2004-282, 2006.
[2] H.Mitchell, "Multi-Sensor Data Fusion, an Introduction", Part I, Chapter
3, Springer, 2007.
[3] M.Skolnik, Introduction to Radar Systems, McGraw-Hill, 1981.
[4] R.Gill, SAR Ice Classification Using Fuzzy Screening Method, Danish
Meteorological Institute Scientific Report, 2002.
[5] F.Smarandache, An In-Depth Look at Quantitative Information
Fusion Rules, Part 8 of "Advances and Applications of DSmT for
Information Fusion", Collected Works, Volume 2, F.Smarandache,
J.Dezert, American Research Press, 2006.
[6] R.Yager, On Ordered Weighted Averaging Aggregation Operators in
Multi-Criteria Decision Making, IEEE Transactions on Systems, 1988.
[7] A.Klausner, A.Tengg, B.Rinner, Distributed Multi level Data Fusion
for Networked Embedded Systems, IEEE Journal of Selected Topics in
Signal Processing, 2008.
[8] S.Blackman, Multi-Target Tracking with Radar Applications, Artech
House, 1986.
[9] S.Bocquet, Calculation of Radar Probability of Detection in K-
Distributed Sea Clutter and Noise, DSTO Australia, 2011.
[10] M.Richards, Generating Swerling Random Sequences, 2008.
[11] K.Khoshelhama, S.Nedkova, C.Nardinocchi, A Comparison of
Bayesian and Evidence-Based Fusion Methods for Automated Building
Detection in Aerial Data, The International Archives of the
Photogrammetry, Remote Sensing and Spatial Information Sciences.
Vol. XXXVII. Part B7. Beijing, 2008.
[12] B.Cobb, P.Shenoy, A Comparison of Bayesian and Belief Function
Reasoning, Information Systems Frontiers, Vol. 5, No. 4, pp. 345358.
Dec. 2003.
[13] A.Tolstikovi, et al. Comparison of fusion methods based on DST and
DBN in human activity recognition, J Control Theory Appl 2011 9(1)
1827, 2011.
[14] F.Cremer, et al. A Comparison of Decision Level Sensor Fusion
Methods for Anti-personnel Landmine Detection, Elsevier,
Information Fusion 2, 187-208, 2001.
[15] Y.Zeng, J.Zhang, J.Genderen, Comparison and Analysis of Remote
Sensing Data Fusion Techniques at Feature and Decision Levels,
International Journal of Remote Sensing, 2004.
[16] T.Bieker, Comparison of Fusion Methods for Successive Declarations
of Radar Range, IEEE International Conference on Information Fusion,
2005.
[17] D.Buede, P.Girardi, A Target Identification Comparison of Bayesian
and DempsterShafer Multi-Sensor Fusion, IEEE Transactions on
Systems, Man and Cybernetics, 1997.
[18] A.Naseri, O.Azmoon, Evaluation of Data Fusion in Radars Network
and Determination of Optimum Algorithm, International Journal of
UbiComp (IJU), Vol.2, No.4, October 2011.
[19] T.Bieker, Statistical Evaluation of Decision-Level Fusion Methods for
Non-Cooperative Target Id by Radar Signatures, IEEE 11th
International Conference on Information Fusion, 2008.
[20] D.Koks, S.Challa, An Introduction to Bayesian and Dempster Shafer
Data Fusion, DSTO Australia, 2005.
[21] P.Smets, R.Kennes, The Transfer Belief Model, In Symbolic and
Quantitative Approaches to Uncertainty ,Springer, 1991.
[22] A.Aregui, T.Denoeux, Constructing Consonant Belief Functions from
Sample Data using Condence Sets of Pignistic Probabilities, Centrede
Recherche de Royallieu, 2008.
[23] O.Azmoon, Survey of Radars Networking Algorithms and Propose
JOURNAL OF COMPUTING, VOLUME 4, ISSUE 1, JANUARY 2012, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG 19

Optimum Algorithm, Thesis for the Degree of M.S., Department of
Information and Communications Technology, IHU University, 2011.
[24] H.Versmissen, Study of the Performance Degradation of the Belgian S-
band Air Surveillance Radars due to the Interference of Upcoming 4G
Technologies, Test Report, IE-RPT-00357-008 Test Report Study BIPT -
Public.odt, Edition Date: 24 Jun 2011.
[25] J.Roux,J,Vuuren, Real-time Threat Evaluation in a Ground Based Air
Defense Environment, ORION Volume24(1), pp.75-101, 2008.

Ali Naseri is professor of Department of Information and
Communications Technology, and chair for Faculty of Information
Processing of IHU University. He has after twenty years research
experience in processing and computing fields and radar processing.
Dr Naseri is also an advisor of Information and Communications
Technology Ministry of Iran.

Omid Azmoon is master engineer of Department of Information and
Communications Technology of IHU University. He has after ten
years research experience. His researches have focused on data
fusion in radars network. He currently works in Iran Communications
Industries.

JOURNAL OF COMPUTING, VOLUME 4, ISSUE 1, JANUARY 2012, ISSN 2151-9617
https://sites.google.com/site/journalofcomputing
WWW.JOURNALOFCOMPUTING.ORG 20

You might also like