You are on page 1of 5

CHANNEL AWARE SENSOR SELECTION IN DISTRIBUTED DETECTION SYSTEMS

Hamid R. Ahmadi and Azadeh Vosoughi


ECE Department, University of Rochester, Rochester, NY 14627
ABSTRACT
We propose a novel censoring scheme for the distributed de-
tection problem in a wireless sensor network (WSNs) with N
sensors, where the channels between the sensors and the fu-
sion center (FC) is subject to fading and noise. To achieve the
best tradeoff between energy efciency and detection reliabil-
ity, the FC forms the maximum ratio combing (MRC) fusion
rule by integrating the partial knowledge of fading channel
state information (CSI) and the local sensor performance in-
dices, nds the best set of K (K < N) sensors that maxi-
mizes the total detection probability in the Neyman-Pearson
(NP) sense, and informs the selected sensors via one bit feed-
back. The FC learns the Rayleigh at fading channels, utiliz-
ing training symbols sent by the sensors, via applying mini-
mum mean square error (MMSE) channel estimator. Assum-
ing the sensors employ BPSK to modulate their binary local
decisions, we derive the MRC fusion rule that depends on the
channel estimates and the sensors performance indices, and
incorporates the effect of channel estimation error. Simula-
tion results are provided to support the analytical derivations.
Index Terms Censoring, channel estimation error, dis-
tributed detection, partial channel state information.
1. INTRODUCTION
In deploying large-scale sensor networks, consisting of typ-
ically low-cost battery-powered devices, for the purposes of
monitoring or detection, we often need to address a funda-
mental tradeoff existing between energy efciency and de-
tection reliability. Censoring networks have been introduced
in [1][2][3], as an alternative formulation of the original dis-
tributed detection (DD) problem proposed in [4]. In [1][2][3]
the authors raise the interesting question of deciding which
sensors should transmit their observations to the fusion center
(FC), assuming that the average number of transmitting sen-
sors is constrained. The sensors make local censoring deci-
sion and the uncensored sensors transmit a real valued func-
tion of their measurements (i.e., the likelihood ratio value),
with innite precision, to the FC [1][2][3]. The references
[5][6][7] consider the asymptotic performance of these cen-
soring networks. Under Neyman-Pearson (NP) framework,
[8][9] propose an extreme censoring scheme with an on/off
local sensor signaling and investigate the optimal binary local
sensor quantizer designs (i.e., the optimal threshold for the
local LRT). Fusion of local decisions received from uncen-
sored sensors transmitted over wireless fading channels are
considered in [10], where the FC has noncoherent reception
and the uncensored sensors send one information bit, indi-
cating that their local LRT exceeds a threshold. There is a
similarity between the censoring sensors problems consid-
ered in the literature: the censoring decisions are made locally
at the sensors, based on local sensor performance indices and
does not consider the quality of the wireless links between the
sensors and the FC. When censoring is implemented locally
at the sensors, any channel aware censoring requires feeding
back the CSI (available side information at the sensors [7])
from the FC to the local sensors with high precision.
In this paper we consider the problem of censoring sen-
sors from a different perspective. Consider a network of N
sensors where each sensor has a different performance index
(i.e., the kth sensor has detection probability P
d
k
and false
alarm probability P
f
k
). Furthermore, the fading channel co-
efcient between the k th sensor and the FC (denoted as h
k
) is
different across sensors. The FC employs coherent reception
to fuse the binary local decisions received from the uncen-
sored sensors. We raise the question: given we wish to select
only K out of the N sensors, what is the best set of K sen-
sors, in terms of maximizing the detection probability? The
question does not have a trivial answer: a sensor with an infor-
mative observation may experience a poor quality link, while
a sensor with less informative measurement may have a bet-
ter quality link. The difculty of the sensor selection problem
increases as we relax the perfect channel state information
(CSI) assumption at the FC. Estimation error due to imperfect
CSI affects the effective noise of the DD system, and hence
degrades the system detection performance. To address the
raised question we rst derive a maximum ratio combining
(MRC) fusion rule, assuming that the sensors employ BPSK
modulation to send data. The transmission period consists of
two phases: in phase I all N sensors send a training symbol
to facilitate channel estimation at the FC. Given the minimum
mean square error (MMSE) estimates of the complex fading
channels and the sensors performance indices, the FC selects
the set of K sensors that would have maximized the detec-
tion performance, and informs the selected sensors via one
bit feedback. In phase II the uncensored sensors send their
modulated binary local decisions. Our implicit assumption is
978-1-4244-3696-5/09/$25.00 2009 IEEE 71
that the fading channel is quasi-static, i.e., the fading coef-
cient is constant during the two phases of transmission.
The organization of the paper is as follows: in Section 2
we introduce our system model and the MMSE channel esti-
mation. In Section 3 we derive the MRC fusion rule and the
total detection probability at the FC, under the NP framework.
Section 4 includes the numerical examples for a specic DD
system and our conclusive remarks.
2. SYSTEM MODEL AND MMSE CHANNEL
ESTIMATION
Suppose there are two hypotheses H
1
and H
0
under test. Our
DDsystemconsists of N sensors, where the kth sensor has lo-
cal performance indices P
f
k
(the false alarm probability) and
P
d
k
(detection probability). In general the pairs (P
f
k
, P
d
k
)
need not be identical and are functions of the thresholds and
observation noise at the sensors. We assume that the observa-
tions at the sensors conditioned on each hypothesis are in-
dependent. In phase I of transmission all N sensors send
one training symbol to enable the FC estimating all N fad-
ing channels coefcients
1
. The received training symbol at
the FC from the kth (k = 1, 2, ..., N) sensor is y
k,t
= s
k
h
k
+
n
k,t
. For simplicity of presentation we assume the training
symbol is s
k
=

E
b
. The fading channel coefcient h
k
=

k
e
j
k
is modeled as zero mean complex Gaussian random
variable with unit variance h
k
CN(0, 1),
k
is the channel
amplitude with Rayleigh distribution and
k
is the channel
phase with uniform distribution
k
[, +). We assume
that the FC forms the MMSE estimate of h
k
using y
k,t
as
[11]:

h
k
= E(h
k
|y
k,t
) =

E
b
E
b
+
2
n
y
k,t
(1)
We model the channel estimation error as the difference be-
tween the actual and the estimate channel

h
k
= h
k


h
k
.
We assume that the estimation error is a zero mean complex
Gaussian variable and has a variance [11]:

h
= (
E
b

2
n
+ 1)
1
(2)
We note that for the linear Gaussian model we have consid-
ered, the MMSE estimate of h
k
given y
k,t
is equivalent to the
LMMSE estimate of h
k
[11].
In phase II of transmission the K uncensored sensors
send their local binary BPSK modulated decisions over K
channels. We model each of N channels as quasi-static at
Rayleigh fading where the fading coefcients are xed dur-
ing the two phases of transmission. The signals at the FC
received from the kth uncensored sensor is:
y
k,d
= u
k
h
k
+n
k,d
k {1, 2, ..., N} (3)
1
The accuracy of channel estimation can be enhanced by increasing the
number of training symbols, at the cost of consuming more energy for learn-
ing the channels.
The noises n
k,t
, n
k,d
, and h
k
s are all mutually independent
and n
k,t
, n
k,d
CN(0,
2
n
), and u
k
{

E
b
} is the dis-
crete time modulated binary decisions.
3. CHANNEL AWARE SENSOR SELECTION
To extend the lifetime of a DD system, consisting of battery-
powered sensors with limited communication capabilities,
that operates in a time varying wireless medium, one should
design adaptive censoring schemes that select the sensors
with greater impact on the total detection performance to
transmit their data, and turn off the sensors with insignicant
effect. The contribution of each sensor to the total detec-
tion performance depends on the local performance indices
(P
d
k
, P
f
k
) (which are related to the observation noise at the
sensors), as well as the fading channel between the sensor and
the FC. Given the pairs (P
d
k
, P
f
k
) and the channel estimates

h
k
for k = 1, 2, ..., N the FC selects the best K sensors and
informs the uncensored sensors, via one bit feedback.
In this section, we derive the optimal LRT fusion rule

LRT
and the suboptimal MRC fusion rule , which is the
low SNR approximation of
LRT
. The MRC rule depends
on the pairs (P
d
k
, P
f
k
) and the channel estimates

h
k
for k =
1, 2, ..., N. Next, we nd the total probability of detection
P
D
, by approximating the MRC rule as a Gaussian random
variable. We derive P
D
as if all N sensors would have trans-
mitted their modulated decisions during phase II. We use the
P
D
expression as a cost function whose maximization leads
us to the optimal set of sensors. For a given K, we select a set,
among all the possible choices (
N
K
), that provides the largest
P
D
, under the constraint that the total false alarm probability
P
F
at the FC is always smaller than or equal to a predeter-
mined value . Only these selected K sensors will actually
send their modulated decisions to the FC during phase II.
3.1. LRT and MRC FUSION RULES
To capture the effect of channel estimation error we substitute
h
k
=

h
k
+

h
k
in (3), where

h
k
is the MMSE channel estimate
given in (1). We can rewrite y
k,d
as:
y
k,d
= u
k

h
k
+w
k,d
k = 1, 2, ..., N (4)
where
w
k,d
= u
k

h
k
+n
k,d
is the new noise term that is uncorrelated with the data symbol
u
k
and combines the effect of channel estimation error and
the AWGN. Recalling that the estimation error

h
k
and n
k,d
are uncorrelated we nd that the noise w
k,d
is a zero mean
complex Gaussian variable with variance:

2
w
= E
b

h
+
2
n
=
E
b

2
n
E
b
+
2
n
+
2
n
(5)
Given

h
k
and u
k
we nd that y
k,d
CN(u
k

h
k
,
2
w
).
72
Hence, we can write the LRT as the following:

LRT
=
N

k=1
f(y
k,d
|H
1
)
f(y
k,d
|H
0
)
=
N

k=1
P
d
k
f(y
k,d
|u
k
=

E
b
) + (1 P
d
k
)f(y
k,d
|u
k
=

E
b
)
P
f
k
f(y
k,d
|u
k
=

E
b
) + (1 P
f
k
)f(y
k,d
|u
k
=

E
b
)
=
N

k=1
P
d
k
e

|y
k,d

h
k

E
b
|
2

2
w
+ (1 P
d
k
)e

|y
k,d
+

h
k

E
b
|
2

2
w
P
f
k
e

|y
k,d

h
k

E
b
|
2

2
w
+ (1 P
f
k
)e

|y
k,d
+

h
k

E
b
|
2

2
w
=
N

k=1
P
d
k
+ (1 P
d
k
)e
4

E
b

2
w
Re(y
k,d

k
)
P
f
k
+ (1 P
f
k
)e
4

E
b

2
w
Re(y
k,d

k
)
(6)
In low SNR regime as
2
n
, we observe that
2
w

. By taking a logarithm from the LRT in (6) and using the
approximations e
x
1 x and log(1 + x) x for small
values of x, we can simplify the above rule and reach to:
=
K

k=1
P
k
Re(y
k,d

k
) (7)
where P
k
= P
d
k
P
f
k
. Given

h
k
we note that is a lin-
ear combination of Gaussian random variables and thus has a
Gaussian distribution.
3.2. TOTAL PROBABILITY DETECTION
For a given threshold value the total probability of false
alarm P
F
at the FC can be written as:
P
F
= Pr( > |H
0
) = Q(

|H
0

|H
0
)
where Q(x) =

x
1

2
e

t
2
2
dt. Assuming P
F
= as the
NP constraint, can be derived in terms of and then detec-
tion probability P
D
is written as:
P
D
= Pr( > |H
1
)
= Q(
Q
1
()
|H
0
+
|H
0

|H
1

|H
1
) (8)
where Q
1
(.) is the inverse of Q(.) function. To nd P
D
we
need to calculate
|H
0
,
|H
1
,
|H
0
,
|H
1
. Let z
k
=
Re(y
k,d

k
). Considering (7) and using the fact that z
k
s are
mutually independent, we nd that:

|H
i
=
K

k=1
P
k
E(z
k
|H
i
)

|H
i
=
K

k=1
P
2
k

2
z
k
|H
i
i = 0, 1 (9)
Under hypothesis H
0
mean of z
k
would be equal to:
E(z
k
|H
0
) = |h
k
|
2
(2P
f
k
1)

E
b
(10)
To derive its variance under hypothesis H
0
, we can write:
E(z
2
k
) = E
b
|

h
k
|
4
+E

(Re(w
k,d

k
))
2

(11)
As Re(w
k,d

k
) is a zero mean Gaussian randomvariable with
variance

2
w
|

h
k
|
2
2
, (11) can be rewritten as:
E(z
2
k
) = E
b
|

h
k
|
4
+

2
w
|

h
k
|
2
2
which leads to:

2
z
k
|H
0
= 4E
b
|

h
k
|
4
P
f
k
(1 P
f
k
) +

2
w
|

h
k
|
2
2
(12)
To nd E(z
k
|H
1
) and
2
z
k
|H
1
we should just substitute P
f
k
with P
d
k
in E(z
k
|H
0
) and
2
z
k
|H
0
. Therefore, the P
D
expres-
sion in (8) can be evaluated, given (P
d
k
, P
f
k
) and the channel
estimates

h
k
, as:
P
D
= Q

Q
1
()

K
k=1
P
2
k

2
z
k
|H
0
2

E
b

K
k=1
P
2
k
|

h
k
|
2

K
k=1
P
2
k

2
z
k
|H
1

(13)
4. NUMERICAL RESULTS AND CONCLUSIONS
In this section we consider two algorithms to nd the set of
K sensors that provides the largest detection probability P
D
between all possible sets with K sensors. There are (
N
K
)
sets with K sensors that we can choose from N sensors. The
rst algorithm is brute force algorithm (so-called B.F. Alg.)
that examines all these sets to nd the set that results in the
Table 1. Local performance indices of sensors and their corresponding channel coefcients. (
2
h
= 1)
- S1 S2 S3 S4 S5 S6 S7 S8 S9 S10
P
d
k
.95 .9 .85 .8 .75 .7 .65 .6 .55 .5
P
f
k
.05 .05 .05 .05 .05 .05 .05 .05 .05 .05
h
k
0.05e
j2.95
0.6e
j1.25
0.78e
j2.2
0.8e
j2.17
0.96e
j2.47
1.1e
j0.99
1.14e
j0.7
1.15e
j1.43
1.38e
j2.8
1.4e
j1.2
73
1 2 3 4 5 6 7 8 9 10
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Number of Selected Sensors (K)
P
r
o
b
a
b
i
l
i
t
y

o
f

D
e
t
e
c
t
i
o
n


SNR =1dB, B.F. Alg.
SNR =1dB, So. Alg.
SNR =1dB, B.F. Alg.
SNR =1dB, So. Alg.
SNR =3dB, B.F. Alg.
SNR =3dB, So. Alg.
Fig. 1. Total probability of detection P
D
versus number of selected
sensors K, using brute force (B.F. Alg.) and sorting (So. Alg.) algo-
rithms. Perfect CSI is assumed at the FC, P
F
= 0.01 and N = 10.
maximum P
D
. To nd the maximum P
D
we need to perform
(
N
K
) 1 comparisons between the detection probabilities.
The second algorithm is a sorting one (so-called So. Alg.)
that rst evaluates the detection probability P
D
at the FC for
each individual sensor and nds the set containing the K sen-
sors with highest individual P
D
values. It is easy to show that
this sorting algorithm needs to perform at most K(N K)
comparisons to nd the desired set, i.e., sorting method has
less complexity compared with that of brute force one. Sup-
pose we want to choose K = 10 sensors out of N = 50 sen-
sors. The rst algorithm requires 1.03 10
10
comparisons
while the second one requires only 400 comparisons. How-
ever, the sorting algorithm nds the sensors for which their
individual P
D
values are good, that do not necessarily result
in the maximum value of P
D
when working with other sen-
sors. So the brute force algorithm is the optimal one with high
complexity, while sorting algorithm is a suboptimal one with
less complexity.
For the numerical examples we assume that there are N =
10 sensors with the local performance indices and channel co-
efcients given in Table 1(the phases are in radian). We as-
sume that the complex channel has unit variance and SNR =
E
b
/
2
n
. Also we assume that the total false alarm probabil-
ity at the FC is xed at = 0.01 and sensors have the same
false alarm probabilities P
f
k
= 0.05 for k = 1, ..., N. Sen-
sors in the eld are indexed such that as the sensors indices
increase, the channel amplitudes |h
k
| increase while sensor
detection probabilities P
d
k
decrease, as it is shown in Table
1. If the channel amplitudes |h
k
| and detection probabilities
P
d
k
were both increasing the answer to selection problem was
trivial and both algorithms would choose the same set of sen-
sors: the set of sensors with larger |h
k
| or P
d
k
values would
have been chosen. But in scenarios where |h
k
| is ascending
1 2 3 4 5 6 7 8 9 10
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Number of Selected Sensors (K)
P
r
o
b
a
b
i
l
i
t
y

o
f

D
e
t
e
c
t
i
o
n


SNR =1dB, B.F. Alg.
SNR =1dB, So. Alg.
SNR =1dB, B.F. Alg.
SNR =1dB, So. Alg.
SNR =3dB, B.F. Alg.
SNR =3dB, So. Alg.
Fig. 2. Total probability of detection P
D
versus number of selected
sensors K, using brute force (B.F. Alg.) and sorting (So. Alg.) al-
gorithms. Imperfect CSI is assumed at the FC, P
F
= 0.01 and
N = 10.
and P
d
k
is descending (e.g., Table 1), it is not clear which set
would be selected.
Figures 1 and 2 show P
D
in (13) versus the number of se-
lected sensors K for both perfect and imperfect CSI in the FC.
The sets of sensors selected by each algorithm are different,
and consequently their detection probabilities P
D
are differ-
ent. As we expected the performance of the selected set by
sorting algorithm is never greater than that of the brute force
algorithm. Sets of chosen sensors for K = 3 and K = 6
at SNR = 1dB, when we have perfect CSI at the FC are
shown in Table 2. As it is shown different sets are selected
by different algorithms. Figures 3 and 4 illustrate detection
probability P
D
versus SNR for K = 3 and K = 6 values for
both perfect and imperfect CSI at the FC. We observe that at
low and high SNR values the algorithms are approaching to
each other and their selected sets are expected to be the same.
Table 2. selected sensors at SNR = 1dB.
K selected sensors
3 brute force: S6, S5, S3
sorting: S6, S9, S7
6 brute force: S3, S4, S5, S6, S7, S8
sorting: S9, S10, S5, S6, S7, S8
In summary, we have proposed a novel censoring scheme
for the distributed detection problem in a WSNs with N sen-
sors, where the channels between the sensors and the FC is
subject to fading and noise, and the sensors employ BPSK for
data modulation. Under the communication constraint that
only K out of the N sensors can transmit their local deci-
sions to the FC, we have addressed the question of selecting
74
10 5 0 5 10 15
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
SNR(dB)
P
r
o
b
a
b
i
l
i
t
y

o
f

D
e
t
e
c
t
i
o
n


K=3, B.F. Alg.
K=3, So. Alg.
K=6, B.F. Alg.
K=6, So. Alg.
Fig. 3. Total probability of detection P
D
versus SNR, using brute
force (B.F. Alg.) and sorting (So. Alg.) algorithms. Perfect CSI is
assumed at the FC, P
F
= 0.01 and N = 10.
which K sensors should transmit, by integrating the local sen-
sors performance indices as well as the MMSE estimates of
the fading channels between the sensors and the FC into the
MRC fusion rule. We have relied on numerical evaluations
to select the set from all (
N
K
) possible choices. We have
used brute force and sorting algorithms to nd the best set of
K sensors, where sorting algorithm is a suboptimal method
with less complexity, compared with brute force method. Fu-
ture work includes investigating heuristics, based on convex
optimization, for approximately solving this sensor selection
problem.
5. REFERENCES
[1] C. Rago, P. Willett, and Y. Bar-Shalom, Censoring sen-
sors: a low-communication-rate scheme for ditributed
detection, IEEE Trans. Aerospace Elec. Syst., vol. 17,
pp. 554567, April 1996.
[2] S. Appadwedula, V. V. Veeravalli, and D. L. Jones, Ro-
bust and locally-optimum decentralized detection with
censoring sensors, in Proceedings of the Fifth Interna-
tional Conference on Information Fusion (ICIF), July
2002, vol. 1, pp. 5663.
[3] S. Appadwedula, V. V. Veeravalli, and D. L. Jones,
Energy-efcient detection in sensor networks, IEEE
Journal on Selected Areas in Communications, vol. 23,
no. 4, pp. 693702, April 2005.
[4] R.R. Tenney and N.R. Sandell Jr., Detection with dis-
tributed sensors, IEEE Trans. Aerospace Elec. Syst.,
vol. 17, pp. 501510, July 1981.
10 5 0 5 10 15
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
SNR(dB)
P
r
o
b
a
b
i
l
i
t
y

o
f

D
e
t
e
c
t
i
o
n


K=3, B.F. Alg.
K=3, So. Alg.
K=6, B.F. Alg.
K=6, So. Alg.
Fig. 4. Total probability of detection P
D
versus SNR, using brute
force (B.F. Alg.) and sorting (So. Alg.) algorithms. Imperfect CSI
is assumed at the FC, P
F
= 0.01 and N = 10.
[5] J. F. Chamberland and V. V. Veeravalli, Asymptotic
results for decentralized detection in power constrained
wireless sensor networks, IEEE Journal on Selected
Areas in Communications, vol. 22, no. 6, pp. 1007
1015, August 2004.
[6] J. F. Chamberland and V. V. Veeravalli, Decentralized
detection in sensor networks, IEEE Trans. on Signal
Processing, vol. 51, no. 2, pp. 407416, February 2003.
[7] W. P. Tay, J. N. Tsitsiklis, and M. Z. Win, Asymp-
totic performance of a censoring sensor network, IEEE
Trans. on Information Theory, vol. 53, no. 11, 2007.
[8] F. Gini, F. Lombardini, and L. Verrazzani, Decen-
tralised detection strategies under communication con-
straints, IEE Proceedings, Radar, Sonar and Naviga-
tion, vol. 145, no. 4, pp. 199 208, August 1998.
[9] Y. Lin, B. Chen, P. Willett, and B. Suter, Distributed
binary quantizers for communication constrained large-
scale sensor networks, in Proceedings of the Ninth In-
ternational Conference on Information Fusion (ICIF),
2006.
[10] R. Jiang and B. Chen, Fusion of censored decisions
in wireless sensor networks, IEEE Trans. on Wireless
Communications, vol. 4, no. 6, pp. 2668 2673, Novem-
ber 2005.
[11] S. M. Kay, Fundamentals of Statistical Signal Pro-
cessing I: Estimation Theory, Prentice-Hall, Englewood
Cliffs, NJ, 1993.
75

You might also like