Fusion of Decisions Transmitted Over Fading Channels in Wireless

Sensor Networks
Biao Chen, Ruixiang Jiang, Teerasit Kasetkasem, and Pramod K. Varshney
Syracuse University, Department of EECS, Syracuse, NY 13244
Abstract
Information fusion by utilizing multiple distributed
sensors is studied in this work. We derive the optimal
likelihood based fusion statistic for a parallel decision
fusion problem with fading channel assumption. This
optimum fusion rule, however, requires perfect knowledge of the local decision performance indices as well
as the fading channel. Several alternatives are presented that alleviate these requirements. At low SNR,
the likelihood based fusion rule reduces to a form analogous to a maximum ratio combining statistic; while
at high SNR, it leads to a two-stage approach using the
well known Chair-Varshney fusion rule. A third alternative in the form of an equal gain combiner is also
proposed that requires the least amount of information
regarding the sensor/channel. Simulation shows that
the two-stage approach, which considers the communication and decision fusion as two independent stages,
suers performance loss compared with the other two
alternatives for practical SNR range.

Introduction

Wireless sensor networks (WSN) have generated

intensive interest from the research community with
current and future envisioned applications ranging
from battleeld surveillance, environment and structure monitoring, to telemedicine. Much of current research eort on WSN is focused on the development
of energy ecient routing protocols, distributed data
compression and transmission schemes, and collaborative signal processing algorithms as documented in [1]
and references therein. The fact that locally processed
information is transmitted through a fading channel
has not attracted much attention. While channel fades
may be treated purely as a communication issue and
thus be dealt with exclusively through transceiver design, we contend in this paper that a more cohesive
approach that integrate the communication with the
information processing should be adopted.
Consider a distributed detection problem in the context of WSN where a number of sensors are used to
detect the possible presence of a target. The problem of distributed detection, and in particular, decision fusion, has been studied extensively in the past
decades. Indeed, in the absence of fading channel consideration, optimum fusion rules have been obtained
for both binary and multibit (soft) local sensor output

under the conditional independence assumption [2, 3].

Fusion rules with statistically correlated observations
have also been investigated though the results are considerably more involved [4, 5]. Decision fusion with
uncertainty has also been investigated and a Bayesian
sampling approach has been proposed to address this
issue [6]. Decision fusion under a communication constraint has also been considered [7, 8, 9, 10, 11]. The
constraint, however, is often in the form of the total
number of bits allowed [9, 10, 11]. The actual transmission is still idealized, i.e., the information sent from
local sensors is assumed intact at the fusion center.
While this assumption may be reasonable for some applications, it may not be realistic for many wireless
sensor networks where the transmitted informaiton has
to endure both channel fading and noise/interference.
Decision fusion with non-ideal channels has been investigated by Thomopoulos and Zhang [12]. The local
decisions are transmitted over noisy channels so that
they may not be correctly received at the fusion center. Yet the channel model is simplied as a binary
channel thus does not allow a full integration of signal
transmission into information processing.
In this paper, we study the practical WSN scenario
where decisions are transmitted over fading channels.
The emphasis is on the integration of the communication and information fusion. The organization of the
paper is as follows. In the next section we formulate
the parallel fusion problem with noisy and fading communication channel layer and derive the optimal LR
based fusion rule with binary local decisions. Section
3 provides three alternative fusion rules: a two-stage
approach using the Chair-Varshney fusion rule, a maximum ratio combining (MRC) fusion statistic, and an
equal gain combiner (EGC) fusion statistic. Performance analysis is given in Section 4 where we nd that
under the Rayleigh fading assumption, the last alternative, namely the EGC statistic, is the most robust
fusion alternative even though it requires minimum information. We conclude in Section 5.

2
2.1

Decision fusion under fading channel

assumption
Problem formulation

Fig. 1 depicts a typical parallel fusion structure

where a number of sensors observe data generated according to either H0 or H1 , the two hypotheses un-

der test. Each sensor processes its observations and

makes a preliminary decision about the hypothesis before sending it to a fusion center. In the conventional
parallel fusion paradigm, the fading channel layer is not
considered and the information sent from individual
sensors is assumed to be received intact at the fusion
center. For WSN with limited resources, the eect of
channel fade and noise renders the information received
at the fusion center to be unreliable. While increased
transmission power along with channel coding may be
implemented, they may not be enough (or desirable)
for many WSN application with limited resources.
Incorrectly received local decisions will lead to performance loss. This loss can be minimized by properly considering and including the eect of channel impairments in the derivation of the decision fusion rule.
Extending the classical parallel fusion problem by incorporating the fading channel layer, we consider the
decision fusion problem as described in Fig. 1. The following assumptions are made to simplify our analysis
and allow us to gain better insights from the results.
1. The k th local sensor make a binary decision uk
{+1, 1}, with false alarm and detection probabilities Pf k and Pdk respectively. That is, P [uk =
1|H0 ] = Pf k and P [uk = 1H1 ] = Pdk . Notice that
these performance indices for local sensors may
vary from sensor to sensor.
2. Each local decision uk is transmitted through a
fading channel and the output of the channel (or
input to the fusion center) for the k th sensor is
y k = hk u k + nk
(1)
where hk is some non-negative number and nk is
zero mean Gaussian with variance 2 . This at
fading channel model is reasonable for many lowpower WSN operating at short ranges (hence small
delay spread) with low bit rate.
Our goal now is to derive a fusion rule based on yk ,
for k = 1, , K that balances the performance with
its requirement on a priori information.

2.2

Optimal LR based fusion rule

We derive, using the model specied in the previous section, the optimal likelihood ratio (LR) based
fusion rule by assuming complete knowledge regarding
the fading channel and the local sensor performance
indices, i.e., the Pf k and Pdk values. Assuming conditional independence of observations at the sensors and
that each local sensor makes a binary decision, the nal LR test statistic can be derived in a straightforward
manner as
K

Pdk e

(yk hk )2
22

k=1

Pf k e

(yk hk )2
22

(y) =

+ (1 Pdk )e
+ (1 Pf k )e

(yk +hk )2
22

(yk +hk )2
22

(2)

where y = [y1 , , yK ]T is a vector containing observations received from all K sensors and 2 is the variance
of additive white Gaussian noise for all channels. An

H0 /H1
Pd2 /Pf 2

Pd1 /Pf 1
Sensor 1

Sensor 2

u1

n1

Sensor K
uK
hK

n2
y1

u2
h2

h1

PdK /Pf K

nK
yK

y2

Fusion
Center
u0
Figure 1: Parallel fusion model in the presence of fading and noisy channel between local sensors and the
fusion center.
implicit assumption is that all the channel outputs are
co-phased. This assumption allows us to deal exclusively with real observations. While the form of the
LR based fusion rule is straightforward to implement,
it does need both the local sensor performance indices
and complete channel knowledge. Suboptimum fusion
rules that relieve the above requirements are more desirable.

Suboptimum fusion rules

The rst two alternatives are proposed as high and

low SNR approximations (or equivalent, to be more
precise) to the LR fusion rule that partially relieve
some of the requirements associated with the LR based
fusion rule. Further, motivated by the form of the second alternative, we propose another alternative that
requires minimum a priori information, and, as it turns
out, enjoys the most robust performance among the
three suboptimal approaches for a wide range of channel SNR values.

3.1

A two stage approximation using the

Chair-Varshney fusion rule

The fusion rule specied in (2) jointly considers the

eects of the fading channel and the local sensor output
to achieve optimal performance. A direct alternative
is to separate this into a two stage process rst yk
is used to infer about uk , and then, the optimum fusion rule based on uk (assuming that the estimates are
reliable), as derived in [2] can be applied. Given (1)
and with uk {1, 1} and hk > 0, the maximum likelihood (ML) estimate for uk is simply uk = sign(yk ).
Applying the fusion rule derived in [2], herein termed

the Chair-Varshney fusion rule, we obtain

Pdk
1 Pdk
1 =
log
+
log
(3)
Pf k
1 Pf k
sign(yk )=1

sign(yk )=1

Clearly, this statistic is a good approximation for for

large channel SNR when the ML estimate on uk tends
to be most likely correct. Indeed, 1 is mathematically
equivalent to for the large SNR case.
Proposition 1 As the channel noise variance 2
0, i.e., SN R , the logarithm of dened in (2)
reduces to 1 dened in (3), i.e.,
lim
log = 1
2
0

Notice that 1 does not require any knowledge regarding the channel gain but does require Pdk and Pf k
for all k. Further, this two-stage approach falls into
the conventional thinking that separates the communication and signal processing aspects. We show later
through numerical examples that this two-stage approach suers signicant performance loss at low to
moderate channel SNR.

3.2

Fusion rule using a maximum ratio

combining statistic

At low SNR, we have the following result.

Proposition 2 As 2 , in (2) reduces to
K

2 =

(Pdk Pf k )hk yk
k=1

Further, if the local sensors are identical, i.e., Pdk and

Pf k are the same for all ks, then further reduces to
a form analogous to a maximum ratio combiner [13]:
2 =

1
K

hk y k

(4)

k=1

1
in 2 does not aect the detection perThe factor K
formance but is introduced for the convenience of performance analysis. Notice that the form of 2 in (4)
does not require the knowledge of Pdk and Pf k provided
Pdk Pf k > 0, i.e., the local detectors are unbiased.
Knowledge of the channel gain is, however, required.

3.3

Fusion rule using an equal gain combining statistic

2 (or 2 with identical sensors) reWhile 1 and

lieve some of the requirements compared with the optimal likelihood ratio based fusion rule, they still need
some information either about the local sensors or the
channel statistics. Further, we note that these fusion
statistics, 1 and 2 , as approximations to the optimal
LR based fusion rule at high and low SNR cases, may
suer performance loss for SNR outside those ranges.
It would be very important to investigate other robust
alternatives that operate well for the non-extreme SNR
range while requiring the same or even less amount of
information regarding the channel and/or the sensors.

Motivated by the fact that 2 resembles a MRC

statistic for diversity combining, we propose a third alternative in the simple form of an equal gain combiner
(EGC) that requires minimum amount of information:
3 =

1
K

yk
k=1

Since yk s are assumed to be phase coherent outputs

of each channel, we still require the phase of the fading
channel but no other information regarding the channel/sensor is needed. While this heuristic and the simple fusion rule in the form of an EGC statistic does
relieve most of the requirements compared with the
optimal LR based fusion rule, its usefulness largely depends on its performance compared with the optimum
fusion rule as well as the rst two alternatives.

Performance evaluation

While it is clear that the LR based fusion rule provides the best detection performance, it is interesting
to see how much performance degradation the other
three simple alternatives suer, and among these three,
which one provides the best and most robust detection performance. While analytical results are most
desirable, the problem is, in general, intractable. The
MRC and EGC fusion rules, however, are amenable to
asymptotic analysis because of their simple expression
in the form of a sum of some random variables that are
independent of each other. In the case of identical sensors and fading statistics, these independent random
variables are also identical to each other, which leads
to the direct application of the central limit theorem
(CLT) for asymptotic analysis. We emphasize that the
EGC and MRC are perhaps more desirable because of
their performance advantage compared with the twostage approach for low to medium SNR values. Most
WSN operating using on-board battery supply are energy limited. Given that RF communication is the
most energy consuming function of a sensor node, it is,
therefore, imperative to use as little power as possible
for data transmission, which usually results in modest
SNR values at the fusion center receiver.
Throughout this section, we will assume a Rayleigh
fading channel for both analysis and numerical simulation. Other fading types, such as Ricean fading, can
be used instead though the analysis is more involved.

4.1

Asymptotic analysis of MRC/EGC

While it is well known that MRC is optimal in output SNR, it relies on an assumption that is taken
for granted in wireless communications, that is, the
sources for multiple independently faded channels are
identical to each other. Under this condition, MRC
achieves maximum output SNR as it involves full coherent combining. In the context of sensor networks,
this is not necessarily the case the local sensors
are prone to make decision errors due to the nature
of the problem. Without identical input to the multiple fading channels, there is no guarantee that MRC is

M RC0
2
M
RC0
M RC1
2
M
RC1

H0
H1

=
=
=
=

MRC
2Pf 1
1
[1 + 2 + 4Pf (1 Pf )]
K
2Pd 1
1
[1 + 2 + 4Pd (1 Pd )]
K

EGC0
2
EGC0
EGC1
2
EGC1

=
=
=
=

EGC

(2Pf 1)
2
1
4
+ 2
K
4

(2Pd 1)
2
1
4
+ 2
K
4

+ Pf (1 Pf )
+ Pd (1 Pd )

Table 1: Mean and variance of MRC and EGC under H0 and H1 with K sensors.

0.9
0.8

Probability of detection

We can therefore obtain the asymptotic relative entropy as a function of channel SNR for both MRC and
EGC statistics by plugging in the corresponding mean
and variance from Table 1. Plotted in Fig. 3 are the
results for both MRC and EGC for the same parameter setting. While the MRC has slight advantage over
the EGC statistic for very low SNR, the EGC is a better statistic for a wide range of SNR values that are of
practical importance.
8 sensors with sensor level Pf=0.05 and Pd=0.5
3

2.5

0.6
0.5
0.4
Monte Carlo for MRC
Approximation for MRC
Monte Carlo for EGC
Approximation for EGC

0.2
0.1

10

10

10

1.5

KullbackLeibler divergence for MRC

KullbackLeibler divergence for EGC

0.5

0
10

10

15

20

25

30

SNR (dB)

4.2

0.7

0.3

Figure 3: Kullback Leibler distance (relative entropy)

between the two hypotheses for both MRC and EGC
using Gaussian approximations.

8 sensors with sensor level Pf=0.05 and Pd=0.5, channel SNR=5dB

1

0 4
10

can be calculated readily as

1
( 2 12 ) + (0 1 )2
+ 0
D(P1 P2 ) = log
0
212

KullbackLeibler Divergence

still preferable compared with other alternatives such

as EGC. In the following, we call the performance at
the network level as system level detection probability
and false alarm rate, denoted by Pd0 and Pf 0 , to distinguish them from sensor level Pdk and Pf k . We assume
in this section that the sensors are identical to each
other, thus Pf k = Pf and Pdk = Pd for all k. Therefore, both MRC and EGC fusion statistics are sums of
independent and identically distributed (i.i.d.) random
variables which allows direct application of the CLT.
This converts the decision fusion problem into hypothesizing between two Gaussian distributions which can
lead to a lot of insight. In order to use the CLT, we
need the rst and second order statistics which are derived and summarized in Table 1.
Given the above statistics, the probabilities of detection and false alarm can be easily obtained using the Q() function, dened as the complimentary
distribution function of standard Gaussian. Fig. 2
presents the receiver operating characteristic (ROC)
curves obtained both by Monte Carlo simulation and
numerical approximation using CLT. In this example, the total number of sensors is 8 with sensor level
Pf = 0.05 and Pd = 0.5 and channel SNR equal to 5dB.
While some discrepancy exists, the approximations using CLT match relatively well to the corresponding
simulation results,

10

Probability of false alarm

Figure 2: ROC curves for MRC and EGC obtained by

simulation and numerical approximation using CLT.
Application of CLT also allows more intuitive explanation and analysis. From Steins lemma [14], the relative entropy (Kullback-Leibler distance) between the
two distributions under test is directly related to the
detection performance in an asymptotic regime. The
relative entropy between two Gaussian distributions

Simulation results

We assume Rayleigh fading with unit power (i.e.,

E[h2k ] = 1) for easy SNR calculation. Binary decision
uk {+1, 1} is made at the local sensors and we
assume that the sensors have identical performance.
Specically, the sensor level false alarm rate is assumed to be Pf = 0.05 while the detection probability is Pd = 0.5 with the total number of sensors
xed at 8. Fig. 4 gives the probability of detection
as a function of channel SNR for a xed system false
alarm at Pf 0 = 0.01. The parameter setting is identical to the above example. From this gure, it is easy
to see that at very low and high SNR, MRC and the
Chair-Varshney statistic can approach the LR performance quite well, while the EGC fusion rule provides
the most robust performance among the three suboptimal approaches. The tradeo between detection performance and the requirement on a priori information
for each of the fusion schemes is summarized in Table
2.

Fusion rule
(LR)
1 (Chair-Varshney)
2 (MRC)
3 (EGC)

a priori information required

Channel SNR and sensor performance indices
Sensor performance indices
Channel SNR
None

Performance
Optimum
Near-optimal for large SNR
Near-optimal for low SNR
Robust for most SNR range

Table 2: Comparison among the four dierent fusion rules for binary decisions transmitted through fading channels.

Conclusions

Fusion of binary decisions transmitted over fading

channels in the context of WSN was studied in this
paper. Based on a parallel fusion structure that incorporates the fading channel, a likelihood ratio based
fusion rule has been derived. In the absence of prior
knowledge regarding the local sensors and/or fading
channels, several alternatives were proposed. The twostage implementation using the Chair-Varshney fusion
rule provides high SNR approximation to the LR fusion
rule, while the statistic in the form of a MRC gives low
SNR approximation. Another heuristic scheme in the
form of an EGC was proposed and we demonstrated
that it performs better than both the Chair-Varshney
approximation and MRC for a wide range of SNR values. Asymptotic analysis and numerical simulation are
carried out for performance comparison.

References
[1] S. Kumar, F. Zhao, and D. Shepherd edts., Special issue on collaborative signal and information
processing in microsensor networks, IEEE Signal
Processing Magazine, vol. 19, Mar. 2002.
[2] Z. Chair and P.K. Varshney, Optimal data fusion in multiple sensor detection systems, IEEE
Trans. Aerospace Electron. Sys., vol. 22, pp. 98
101, Jan. 1986.
[3] Pramod K. Varshney, Distributed Detection and
Data Fusion, Springer, New York, 1997.

[8] C. Rago, P.K. Willett, and Y. Bar-Shalom, Censoring sensors: a low-communication-rate scheme
for distributed detection, IEEE Trans. AES, vol.
32, pp. 554568, Apr. 1996.
[9] J. Hu and R. Blum, On the optimality of
nite-level quantization for distributed signal detection, IEEE Trans. Information Theory, vol.
47, pp. 16651671, May 2001.
[10] J. Chamberland and V.V. Veeravalli, Decentralized detection in wireless sensor networks, in
Proc. of the 36th Annual Conference on Information Science and Systems, Princeton, NJ, Mar.
2002.
[11] Q. Cheng, P.K. Varshney, K. Mehrotra, and C.K.
Mohan, Optimal bandwidth assignment for distributed sequential detection, in Proc. 5th International Conference on Information Fusion, Annapolis, MD, July 2002.
[12] S.C.A. Thomopoulos and L. Zhang, Distributed
decision fusion with networking delays and channel errors, Information Science, vol. 66, pp. 91
118, Dec. 1992.
[13] G.L. St
uber, Principles of Mobile Communication, Kluwer, Boston, MA, 2nd edition, 2001.
[14] T.M. Cover and J.A. Thomas, Elements of Information Theory, Wiley, New York, 1991.

[4] E. Drakopoulos and C.C. Lee, Optimum multisensor fusion of correlated local decisions, IEEE
Trans. on Aerospace and Elect. Syst., vol. 27, no.
4, pp. 593605, July 1991.

[6] B. Chen and P. Varshney, A Bayesian sampling approach for decision fusion using hierarchical models, IEEE Trans. Signal Processing, vol.
50, pp. 18091818, Aug. 2002.
[7] F. Gini, F. Lombardini, and L. Verrazzani, Decentralised detection stratigies under communication constraints, IEE Proceedings, Part F:
Radar, Sonar, and Navigation, vol. 145, pp. 199
208, Aug 1998.

f0

0.8
0.7

Probability of detection

[5] M. Kam, Q. Zhu, and W.S. Gray, Optimal data

fusion of correlated local decisions in multiple sensor detection systems, IEEE Trans. Aerospace
Elect. Syst., vol. 28, pp. 916920, July 1992.

System level P =0.01, 8 sensors with sensor level P =0.05 and P =0.5
0.9

0.6
0.5
0.4
0.3

LR
ChairVarshney
MRC
EGC

0.2
0.1
0
10

10

15

20

SNR (dB)

Figure 4: Probability of detection as a function of channel SNR for Rayleigh fading channels with 8 sensors.
The system false alarm rate is xed at Pf 0 = 0.01.