Anda di halaman 1dari 11

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO.

1, JANUARY 2009 263


Nonorthogonal Transmission and Noncoherent
Fusion of Censored Decisions
Simon Yiu, Member, IEEE, and Robert Schober, Senior Member, IEEE
AbstractIn this paper, we consider the binary-distributed
hypothesis testing problem, where a wireless sensor network
(WSN) is used to determine the true state of nature. We propose
a novel signaling scheme where signature vectors (SVs) are used
to transmit censored sensor decisions to the fusion center (FC).
In particular, the length of the SVs is chosen to be smaller than
the number of sensors, resulting in nonorthogonal sensor chan-
nels and allowing a desirable tradeoff between bandwidth and
power efciency in WSNs. We consider deterministic SVs that
are assumed to be known to the FC and random SVs that are
unknown to the FC. For both cases, the optimum noncoherent
fusion rules are derived, assuming independent Ricean fading
sensorFC channels. Since the complexity of these optimum fusion
rules is prohibitively large, even for WSNs of moderate size, we
also consider a low-complexity suboptimum energy-based fusion
rule and show that it is equivalent to the optimum fusion rules
under certain conditions. Furthermore, we calculate the system
probabilities of detection and false alarm for the energy-based
fusion rule for both deterministic and random SVs. Numerical and
simulation results show that with the proposed nonorthogonal sig-
naling scheme, signicant improvements in bandwidth efciency
are possible at the expense of only a small loss in power efciency.
Index TermsData fusion, fading channels, nonorthogonal
signaling, wireless sensor networks (WSNs).
I. INTRODUCTION
R
ECENT advances in wireless communications and elec-
tronics have enabled the development of large-scale wire-
less sensor networks (WSNs). A typical WSN consists of
a fusion center (FC) and a number of low-cost, low-power
sensors. To keep computational and communication cost at a
minimum, local sensors typically make preliminary decisions
on their respective real-valued observations of the hypothesis
under test rst before they forward their local information to
the FC (cf., e.g., [1][5]). Further savings in communication
cost are possible if sensors are allowed to cooperate [6].
The FC combines the information received from the sensors
to obtain the nal decision, and the resulting decentralized
Manuscript received October 29, 2007; revised February 6, 2008 and
March 18, 2008. First published April 18, 2008; current version published
January 16, 2009. This work was supported by the National Science and Engi-
neering Research Council of Canada through a Canadian Graduate Scholarship.
The review of this paper was coordinated by Prof. H.-C. Wu.
S. Yiu was with the Department of Electrical and Computer Engineering,
University of British Columbia, Vancouver, BC V5G 1X9, Canada. He is now
with the School of Engineering and Applied Sciences, Harvard University,
Cambridge, MA 02138 USA (e-mail: simony@seas.harvard.edu).
R. Schober is with the Department of Electrical and Computer Engineering,
University of British Columbia, Vancouver, BC V5G 1X9, Canada (e-mail:
rschober@ece.ubc.ca).
Color versions of one or more of the gures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identier 10.1109/TVT.2008.923659
decision problem has received considerable attention in the lit-
erature. However, most of the work in this area assumed perfect
error-free communication channels between the sensors and the
FC for the derivation of fusion rules (cf., e.g., [1][5] and [7]).
Although this assumption is valid for systems that are working
in the high SNR regime and simplies the system design, it
may not be practical for WSNs, where unattended sensors have
limited power resources and are affected by fading.
The problem of fusing decisions transmitted over noisy
fading channels was investigated in [8] and [9], assuming
the availability of either full or partial (phase) channel state
information (CSI). However, phase synchronization may be
difcult to realize in low-cost WSNs employing cheap local
oscillators. This problem is overcome in [10], where sensors
employ censoring [11] and simple onoff keying (OOK), lead-
ing to a truly noncoherent fusion rule.
The aforementioned works assume orthogonal channels be-
tween the sensors and the FC, implying large system bandwidth
or large latencies for large-scale WSNs. Given the generally
limited available bandwidth and the possibly large number
of sensors, it may be challenging and wasteful to reserve an
orthogonal channel for each sensor in the network. This is
particularly true if censoring is applied since, in that case, only
a fraction of the sensors will transmit, even if the target is
present, and most channels will not be used most of the time.
Therefore, it may be more practical to assign nonorthogonal
channels to the sensors. Coherent nonorthogonal signaling
schemes for WSNs have recently been proposed in [12][17].
However, these coherent schemes require instantaneous CSI at
the receiver, and censoring was not considered in [12][17].
Motivated by the aforementioned prior work, we propose
and analyze in this paper a nonorthogonal signaling scheme for
the transmission of censored decisions and noncoherent fusion
rules, which do not require instantaneous CSI. In particular, we
claim the following contributions.
1) We consider the binary-distributed hypothesis testing
problem and propose a novel nonorthogonal signaling
scheme, where each sensor is assigned either a deter-
ministic or random signature vector (SV), and all sensors
whose likelihood ratios (LRs) exceed a predened thresh-
old concurrently transmit their SVs to the FC.
2) We derive the optimum noncoherent LR-based fusion
rules for deterministic and random SVs, respectively, and
a suboptimum energy-based fusion rule whose complex-
ity is independent of the number of sensors.
3) We analyze the performance of the energy-based fu-
sion rule for deterministic and random SVs. Based on
this analysis, we investigate the effect of censoring and
0018-9545/$25.00 2009 IEEE
264 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 1, JANUARY 2009
Fig. 1. Considered parallel fusion and signaling model with K sensors and
one FC.
the impact of various system parameters on the overall
performance.
Organization: In Section II, we introduce the proposed
transmission scheme and formulate the parallel fusion problem
under consideration. In Sections III and IV, the optimum and
suboptimum fusion rules for deterministic and random SVs
are derived, respectively. The performance of the energy-based
fusion rule is analyzed in Section V. Simulation and numerical
results are presented in Section VI, whereas some conclusions
are drawn in Section VII.
Notation: In this paper, bold upper case and lower case
letters denote matrices and vectors, respectively. []
H
, E{},

2
, and denote Hermitian transposition, statistical ex-
pectation, the L
2
-norm of a vector, and the union of two
sets, respectively. In addition, det(), tr{}, I
X
, 0
NM
, and
j

=

1 refer to the determinant of a matrix, the trace
of a matrix, the X X identity matrix, the N M all-
zeros matrix, and the imaginary unit, respectively. Further-
more, Q(x)

= (1/

2)
_

x
e
t
2
/2
dt, (x)

=
_

0
e
t
t
x1
dt,
(, x)

=
_

x
e
t
t
1
dt, I
n
(x), and (x) denote the Gaussian
Q-function, the Gamma function, the incomplete Gamma func-
tion, the nth-order modied Bessel function of the rst kind,
and the Dirac delta function, respectively.
1

2
means that
the fusion rules
1
and
2
are equivalent in the sense that
they will lead to the same decision. Finally, depending on the
context, we use | | to denote the cardinality of a set or the
absolute value of a scalar.
II. SYSTEM MODEL
In this paper, we consider the binary-distributed hypothesis
testing problem, where a set K

= {1, 2, . . . , K} of K sensors
is used to determine the true state of nature H as being H
0
(target absent) or H
1
(target present). The a priori probabilities
of the two hypotheses H
0
and H
1
are denoted as P(H
0
) and
P(H
1
), respectively. Fig. 1 illustrates the considered system
model, which will be discussed in detail in the following.
A. Sensor Processing of Local Observations
The sensors collect their own observations x
k
, process them,
and make a local decision, i.e., u
k
{0, 1}, k K. We assume
here that the sensor observations are described by
H
1
: x
k
=1 +n
k
, k K
H
0
: x
k
=n
k
, k K. (1)
Furthermore, we assume that the observations x
k
, k K, are
independent of each other conditioned on any of the two
hypotheses, that is, the noise samples n
k
, k K are mutually
statistically independent. For all numerical examples shown in
Section VI, we model n
k
, k K as independent real-valued
additive white Gaussian noise (AWGN) with zero mean and
variance
2
k
. After collecting its own observation, each sensor
computes its local LR
l
(x
k
) and makes a decision according
to the LR test, i.e.,

l
(x
k
) =
f(x
k
|H
1
)
f(x
k
|H
0
)

k
decide H
1
, set u
k
= 1
<
k
decide H
0
, set u
k
= 0
(2)
for k K, where f(|) denotes the conditional probability
density function (pdf) of x
k
, and
k
is the decision threshold.
For future reference, we dene the sets of sensors with u
k
= 1
and u
k
= 0 as S and O, respectively. The sensor performance
is characterized by the local probability of false alarm P
f
k

=
P(u
k
= 1|H
0
) and the local probability of detection P
d
k

=
P(u
k
= 1|H
1
), k K. For example, for the AWGN case, we
obtain P
f
k
= Q(
k
/
k
) and P
d
k
= Q((
k
1)/
k
), where

=
2
k
ln(
k
) + 0.5.
Only the sensors in S forward their decisions to the FC
using the signaling scheme introduced in Section II-B. There-
fore, the LR test presented in (2) essentially constitutes a
sensor censoring scheme [10], [11], where only sensors with
informative observation transmit; otherwise, they remain silent.
Censoring sensors are particularly attractive in WSNs where the
target-present hypothesis H
1
happens infrequently. In this case,
instead of wasting scarce energy resources to report a null event
H
0
to the FC for every time slot, the sensors in O simply stay
quiet [10], [11].
B. Signaling Scheme
Each sensor k K is assigned an SV g
k
of length N K.
After receiving a query by the FC, the sensors in S transmit the
elements of vector

Eg
k
in N chip intervals of duration T
c
to
the FC, where E denotes the average transmitted energy of the
sensors in S. Assuming that we keep the chip duration T
c
(and,
thus, the bandwidth) constant, small SV lengths N are desir-
able, as they correspond to shorter symbol duration T
c

= NT
c
and, therefore, to a more efcient use of the available band-
width. We note that although the proposed signaling scheme
is related to conventional direct-sequence code-division mul-
tiple access (DS-CDMA), there are important differences. In
DS-CDMA, different users transmit different data, and spread-
ing is performed to separate the data of the users at the receiver.
In the proposed scheme, all sensors transmit (ideally) the same
YIU AND SCHOBER: NONORTHOGONAL TRANSMISSION AND NONCOHERENT FUSION OF CENSORED DECISIONS 265
data, and the FC makes a joint decision where the SVs enable a
diversity gain.
In the following, we discuss the SVs and the censoring in the
proposed signaling scheme in more detail.
SVs: We consider the deterministic and random SVs. In the
deterministic case, the FC has full knowledge of the SVs, as
they can be preassigned to the sensors, whereas in the random
case, the FC does not know the SVs. In the latter case, the
sensors generate a new random SV in each symbol interval,
which reduces the protocol overhead if new sensors frequently
join or leave the WSN. For the energy of the SVs, we may
assume either an average constraint C1) E{g
k

2
2
} = 1, k K
or a stricter instantaneous constraint C2) g
k

2
2
= 1, k K.
Although C2) may be preferable from a practical point of view,
C1) leads to simpler analytical results, particularly for random
SVs, where we will mostly consider zero-mean Gaussian ran-
domSVs with E{g
k
g
H
k
} = (1/N)I
N
and E{g
k
g
H
i
} = 0
NN
,
k = i. Furthermore, for deterministic SVs, we will assume for
one of the proofs in Section III that the set of SVs forms
a tight frame [18]. The average transmitted energy per sym-
bol interval T
s
is given by E
s

= E{|S|}E, where E{|S|} =

K
k=1
P
d
k
P(H
1
) +

K
k=1
P
f
k
P(H
0
) is the average number
of sensors in S.
Censoring: Since only the sensors in S transmit, the deci-
sion threshold
k
constitutes a censoring mechanism. Small
k
leads to large P
f
k
and unreliable decisions. On the other hand,
very large
k
leads to small P
d
k
, and all sensors may miss the
target. The choice of
k
also inuences the energy that a sensor
can use for transmission if it detects a target. Larger
k
means
that each sensor will detect the target less frequently (because
of the smaller P
f
k
and P
d
k
) and can, therefore, use more energy
for transmission if it does detect the target. Therefore, the
thresholds
k
should be optimized to maximize performance
(cf. Section VI).
C. Channel Model
The signal samples received at the FC in N consecutive chip
intervals are collected into vector y. Assuming perfect timing
synchronization,
1
y can be modeled in equivalent complex
baseband representation as
y =

E

kS
g
k
h
k
+n =

EG
S
h
S
+n (3)
where h
k
and n denote the fading gain of sensor k and a
complex AWGN vector, respectively. G
S
is an N |S| matrix
whose columns are the SVs of the sensors in S, and vector
h
S
contains the corresponding fading gains h
k
, k S. We
model the channel gains as independent nonidentically distrib-
uted (i.n.d.) nonzero mean complex Gaussian random variables
(Ricean fading), i.e.,
h
k
=

h
k
+

h
k
, k K (4)
1
Synchronization issues in WSNs with concurrently transmitting sensors are
discussed in some detail in [19].
where

h
k

= E{h
k
} and

h
k

= h
k

h
k
denote the direct and
specular components, respectively. The variance and the Ricean
factor of h
k
are given by
2
h
k

= E{|h
k

h
k
|
2
} and

= |

h
k
|
2
/
2
h
k
, k K, respectively. We note that the Ricean
fading channel reduces to a Rayleigh fading and an unfaded
AWGN channel for
k
= 0 and
k
, respectively. Finally,
the elements of the noise vector n have variance
2
0
= N
0
,
where N
0
denotes the power spectral density of the underlying
continuous-time noise process.
We note that the scheme proposed in [10] may be viewed as
a special case of our scheme with N = K and orthogonal SVs
g
k
= e
k
, k K, where e
k
is a column vector of length K with
the kth element equal to 1 and all other elements equal to 0.
However, the main focus of this paper is the more interesting
bandwidth-efcient case N < K, where the SVs cannot be
made orthogonal.
Equation (3) shows that the N 1 effective channel vector is
h
e

= G
S

h
S
+G
S

h
S
, where

h
S

= E{h
S
}, and

h
S

= h
S

h
S
. The N N covariance matrix of h
e
is given by
C
e

= E
_
(h
e
G
S

h
S
)(h
e
G
S

h
S
)
H
_
= G
S

S
G
H
S
(5)
where
S

= E{

h
S

h
H
S
} is an |S| |S| diagonal matrix with

2
h
k
, k S in its main diagonal. C
e
has maximum rank
r(C
e
) = min{N, |S|}, which is also the maximum diversity
order of the channel. Since S and |S| are random and not
known at the FC, the full diversity gain of the communication
channel can only be enjoyed if any |S| N SVs are linearly
independent. Although the optimization of the set of SVs for
the maximization of the overall performance of the WSN does
not seem to be possible because of the complexity of the
problem, we will show in Section VI that a high diversity gain
is benecial if N and K are large. Therefore, in this paper,
we consider sets of SVs where any |S| N SVs are linearly
independent. This can be achieved with carefully designed
deterministic SVs or random SVs, cf., e.g., [20] and [21] for
suitable designs.
D. FC Processing
The FC combines the elements of the received vector y
according to some fusion rule (y) and compares (y) with
a predened threshold
0
. The FC outputs u
0
= 1 (corre-
sponding to H
1
) if (y)
0
, and u
0
= 0 (corresponding
to H
0
) otherwise. The resulting system probabilities of false
alarm and detection are denoted by P
f
0

= P(u
0
= 1|H
0
) and
P
d
0

= P(u
0
= 1|H
1
), respectively. Threshold
0
can be used
to achieve a desired probability of false alarm P
f
0
.
We only consider noncoherent fusion rules (y) since, due
to censoring, the elements of S may change on a symbol-by-
symbol basis, which makes the estimation of h
e
, as required
for coherent fusion, impossible.
III. FUSION RULES FOR DETERMINISTIC SVS
In this section, we derive the optimum and suboptimum
fusion rules for deterministic SVs, which are known to the FC.
266 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 1, JANUARY 2009
A. Optimum Fusion Rule
For the optimum fusion rule, we assume that the FC has
knowledge of the channel statistics, the local sensor perfor-
mance indexes P
d
k
and P
f
k
, k K, and the deterministic SVs
g
k
, k K. However, knowledge of the instantaneous channel
gains h
k
, k K, is not required. With these assumptions, the
optimum LR-based fusion rule is given by

opt,D
(y) =
f(y|H
1
)
f(y|H
0
)
=

u
f(y| u)P( u|H
1
)

u
f(y| u)P( u|H
0
)
(6)
where u

= [ u
1
, . . . , u
K
], and u
k
{0, 1}, k K. The condi-
tional pdf f(y| u) in (6) is given by (7), shown at the bottom
of the page, where

S denotes the set of sensors with u
k
= 1,
k K. For Rayleigh fading and unfaded channels, (7) can
be simplied by exploiting

h

S
= 0
|

S|1
and

S
= 0
|

S||

S|
,
respectively.
Denoting the set of sensors with u
k
= 0, k K by

O, the
conditional probabilities P( u|H
1
) and P( u|H
0
) in (6) can be
expressed as
P( u|H
1
) =
K

k=1
P( u
k
|H
1
) =

S
P
d
k

O
(1 P
d
k
) (8)
P( u|H
0
) =
K

k=1
P( u
k
|H
0
) =

S
P
f
k

O
(1 P
f
k
). (9)
Discussion: Since the local decision variables u
k
are binary
valued, the numerator and the denominator of (6) require 2
K
sum-of-product computations, that is, the complexity of the
optimum fusion rule
opt,D
(y) exponentially grows with K.
Therefore,
opt,D
(y) is only feasible for small K, and subop-
timum fusion rules have to be used for larger K.
B. Suboptimal Fusion Rules
Here, we show that the optimum fusion rule can be con-
siderably simplied under certain conditions, resulting in low-
complexity suboptimum fusion rules.
1) Reliable Sensor Decisions: If the local sensor decisions
are reliable, i.e., P
d
k
1 and P
f
k
0, k K, (6) can be
simplied. In particular, omitting all irrelevant terms, we obtain
the fusion rule in (10), shown at the bottom of the page. We note
that (10) is equivalent to
opt,D
(y) for P
d
k
1 and P
f
k
0,
k K.
Discussion: Since we assumed reliable sensor decisions for
its derivation,
rel,D
(y) only requires the knowledge of the
channel statistics

h
K
and
K
. Furthermore, the complexity of

rel,D
(y) is independent of K.
2) Energy-Based Fusion Rule: The energy-based fusion
rule is given by

ene
(y) = y
H
y (11)
and can be shown to be equivalent to
opt,D
(y) for the special
case of independent identically distributed (i.i.d.) Rayleigh
fading channels (i.e.,
k
= 0,
2
h
k
= 1, k K),
2
0
,
and identical sensors (i.e., P
d
k
= P
d
> P
f
k
= P
f
, k K) if
the SVs form a tight frame [18], i.e., G
K
G
H
K
= (K/N)I
N
(cf. Appendix A). For nite
2
0
, it is easy to see from (10) that

rel,D
(y) is also equivalent to
ene
(y) for i.i.d. Rayleigh fading
and G
K
G
H
K
= (K/N)I
N
.
Discussion: The energy-based fusion rule
ene
(y) is partic-
ularly simple, as it requires neither the knowledge of the chan-
nel statistics nor of the sensor performance indexes P
d
k
and
P
f
k
. Furthermore, its complexity is very low and independent
of K. We note that the energy-based fusion rule also generally
achieves high performance for
2
0
< (cf. Section VI). This
may be explained by the fact that an energy-based decision
rule is optimum for conventional OOK, and the proposed
nonorthogonal signaling scheme is basically an extension of
OOK. We also note that for the special case of orthogonal
channels with N = K, an energy-based fusion rule was already
considered in [10].
IV. FUSION RULES FOR RANDOM SVS
Here, we assume that the sensors employ random SVs, which
are not known at the FC. In particular, we assume that the
sensors use random Gaussian SVs since this facilitates the
derivation of optimum and suboptimum fusion rules.
A. Optimum Fusion Rule
Here, we derive the optimum fusion rule for Gaussian SVs.
Thereby, we rst consider general i.n.d. Ricean fading and
subsequently specialize to i.i.d. Ricean fading.
1) i.n.d. Ricean Fading: The optimum fusion rule

opt,R
(y) for random Gaussian SVs is also given by (6) but
f(y| u) =
exp
_
(y

EG

S
)
H
_
EG

S
G
H

S
+
2
0
I
N
_
1
(y

EG

S
)
_

N
det
_
EG

S
G
H

S
+
2
0
I
N
_ (7)

rel,D
(y) =
exp
_
(y

EG
K

h
K
)
H
_
EG
K

K
G
H
K
+
2
0
I
N
_
1
(y

EG
K

h
K
)
_
exp (y
H
y/
2
0
) det
_
EG
K

K
G
H
K
+
2
0
I
N
_ (10)
YIU AND SCHOBER: NONORTHOGONAL TRANSMISSION AND NONCOHERENT FUSION OF CENSORED DECISIONS 267
with a new denition of f(y| u). In particular, conditioned on
h

S
, the elements of y are independent Gaussian random
variables with zero mean and variance Ez

S
/N +
2
0
.
Furthermore, z

= h
H

S
h

S
is a noncentral chi-square random
variable with 2|

S| degrees of freedom and noncentrality


parameter a
2

S
=

2
=

k

S
|

h
k
|
2
, and we denote its pdf
by f
z

S
(z

S
). f(y| u) can be expressed as
f(y| u) =

_
0
f(y| u, z

S
)f
z

S
(z

S
)dz

S
(12)
where
f(y| u, z

S
) =
exp
_
y
H
y/
_
Ez

S
/N +
2
0
__

N
_
Ez

S
/N +
2
0
_
N
. (13)
The optimum fusion rule
opt,R
(y) can now be obtained by
combining (6), (8), (9), and (12). However, as mentioned be-
fore, the complexity of evaluating (6) exponentially grows with
K and is, therefore, prohibitive for large K.
2) i.i.d. Ricean Fading: Unlike the deterministic case, for
random SVs, the optimum fusion rule can be considerably
simplied if i.i.d. channels (i.e.,

h
k
=

h and
2
h
k
=
2
h
, k K)
and identical sensors (i.e., P
d
k
= P
d
, P
f
k
= P
f
, k K) are
assumed. In particular, for i.i.d. channels, the conditional pdf
of y depends only on the cardinality of the set

S and not on the
individual elements of

S. Therefore, (13) simplies to
f(y| u, z

S
) = f(y|i, z
i
) =
exp
_
y
H
y/
_
Ez
i
/N +
2
0
__

N
(Ez
i
/N +
2
0
)
N
(14)
where z
i

= h
H

S
h

S
, |

S| = i has pdf [22]


f
z
i
(z
i
) =
1

2
h
_
z
i
a
2
i
_
(i1)/2
exp
_

_
z
i
+a
2
i
_

2
h
_
I
i1
_
4a
2
i
z
i

4
h
_
, i > 0 (15)
with a
2
i
= i|

h|
2
. For i = 0, we obtain z
0
= 0, and the corre-
sponding pdf can be modeled as f
z
0
(z
0
) = (z
0
). For the spe-
cial cases of Rayleigh fading and an unfaded channel, f
z
i
(z
i
)
simplies to f
z
i
(z
i
) = z
i1
i
e
z
i
/(i), i > 0, and f
z
i
(z
i
) =
(z
i
i), respectively. The conditional pdf f(y|i) can now be
calculated via
f(y|i) =

_
0
f(y|i, z
i
)f
z
i
(z
i
)dz
i
. (16)
Assuming, furthermore, identical sensors, (6) can be simpli-
ed to

opt,R
(y) =
K

i=0
f(y|i)P (|S| = i|H
1
)
K

i=0
f(y|i)P (|S| = i|H
0
)
(17)
with the conditional probabilities P(|S| = i|H
0
) =
_
K
i
_
P
i
f
(1
P
f
)
Ki
and P(|S| = i|H
1
) =
_
K
i
_
P
i
d
(1 P
d
)
Ki
.
Discussion: Since P(|S| = i|H
0
) and P(|S| = i|H
1
) can
be precomputed, the complexity of (17) only linearly increases
with K. This is an advantage over the deterministic approach
considered in Section III. However, the integral in (16) has to
be evaluated for each i, 1 i K, which is computationally
expensive for
i
< . Therefore, also for random SVs, low-
complexity suboptimum fusion rules are desirable.
B. Suboptimum Fusion Rule
Similar to the deterministic case, we show in Appendices
B and C that the optimum fusion rule can be considerably
simplied for the special cases of reliable sensor decisions
(i.e., P
d
k
1 and P
f
k
0, k K) and identical sensors
(P
d
k
= P
d
> P
f
k
= P
f
, k K), with
2
0
, respectively.
For randomGaussian SVs, both cases lead to the simple energy-
based fusion rule
ene
(y), as given by (11).
Discussion: It is interesting to observe that for random SVs,
the energy-based fusion rule is optimum under more general
conditions than for deterministic SVs. For example, for
2
0

, we only have to assume identical sensors with P
d
> P
f
for
ene
(y) to be optimum in the random SV case, whereas
for deterministic SVs, we also have to assume i.i.d. Rayleigh
fading.
V. ANALYSIS OF ENERGY-BASED FUSION RULE
The optimum fusion rules
opt,D
(y) and
opt,R
(y) are dif-
cult to analyze. In addition, from a practical point of view, the
energy-based fusion rule
ene
(y) is more important, particu-
larly for large-scale WSNs with many sensors. Therefore, here,
we analyze the performance that is achievable with
ene
(y)
for deterministic and random SVs. In particular, we develop
closed-form expressions for the system probabilities of false
alarm P
f
0
and detection P
d
0
.
A. Deterministic SVs
P
f
0
and P
d
0
can be expressed as
P
f
0
=

u
P(u
0
= 1| u)P( u|H
0
) (18)
P
d
0
=

u
P(u
0
= 1| u)P( u|H
1
) (19)
where P( u|H
0
) and P( u|H
1
) are dened in (8) and (9), re-
spectively. Furthermore, P(u
0
= 1| u) denotes the probability
of u
0
= 1, given u and can be expressed as
P(u
0
= 1| u) = P (
ene
(y)
0
| u) = P(y
H
y
0
| u).
(20)
Since

= y
H
y is a quadratic form of Gaussian random
variables, for general i.n.d. Ricean fading channels, the Laplace
transform

(s) of the pdf of is given by (21), shown at the


bottom of the next page [23].
268 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 1, JANUARY 2009
Based on (20) and (21), P(u
0
= 1| u) can be expressed
as [24]
P(u
0
= 1| u) =
1
2j
c+j
_
cj

(s)e
s
0
s
ds (22)
where c is a small positive constant in the region of convergence
of the integral. The integral in (22) can be numerically evaluated
using, for example, GaussChebyshev quadrature rules [24] or
exactly using the residue theorem [24], [25]. P
f
0
and P
d
0
can
be calculated by combining (18) and (19) with (22).
Discussion: The numerical evaluation of (18) and (19) may
be problematic for large K since averaging of P(u
0
= 1| u)
over all 2
K
possible combinations of u is required. We will
show in Section V-B that this problem can be avoided for the
case of random SVs, identical sensors, and i.i.d. channels.
B. Random SVs
For the general case of i.n.d. channels and nonidentical
sensors, the complexity of evaluating P
f
0
and P
d
0
is still expo-
nential in K. Therefore, here, we concentrate on, practically, the
important case of i.i.d. fading channels and identical sensors,
where P
f
0
and P
d
0
can be expressed as
P
f
0
=
K

i=0
P(u
0
= 1|i)P (|S| = i|H
0
) (23)
P
d
0
=
K

i=0
P(u
0
= 1|i)P (|S| = i|H
1
) (24)
where P(u
0
= 1|i) denotes the probability that u
0
= 1, given
|S| = i, and that P(|S| = i|H
0
) and P(|S| = i|H
1
) were de-
ned in Section IV-A2.
For the calculation of P(u
0
= 1|i), it is useful to recall
from Section IV-A1 that, conditioned on h
S
, the elements of
y are independent Gaussian random variables with zero mean
and variance Ez
i
/N +
2
0
. Therefore, x

= y
H
y is central chi-
square-distributed with 2N degrees of freedom, and its condi-
tional pdf is given by
f
x
(x|z
i
) =
x
N1
exp
_
x/
_
Ez
i
/N +
2
0
__
(N) (Ez
i
/N +
2
0
)
N
(25)
and the pdf of z
i
is given in Section IV-A2. Hence, we can
express P(u
0
= 1|i) as
P(u
0
= 1|i) =

0
f
x
(x)dx (26)
where the pdf of x is given by f
x
(x) =
_

0
f
x
(x|z
i
)f
z
i
(z
i
)dz
i
.
In the following, we evaluate P(u
0
= 1|i) for Ricean fading,
Rayleigh fading, and unfaded channels, respectively.
i.i.d. Ricean fading: For i > 0, z
i
is a noncentral chi-square
random variable with 2i degrees of freedom and noncentrality
parameter a
2
i
= i|

h|
2
[cf. (15)]. After some manipulations,
P(u
0
= 1|i) can be expressed as
P(u
0
=1|i)=
1
(N)
2
h

_
0
_
z
i
a
2
i
_
(i1)/2
exp
_

_
z
i
+a
2
i
_

2
h
_
I
i1
_
4a
2
i
z
i

4
h
_

_
N,

0
Ez
i
/N +
2
0
_
dz
i
, i > 0. (27)
For the special case of i = 0, z
i
= 0 holds, and we obtain
P(u
0
= 1|i = 0) =

0
x
N1
exp
_
x/
2
0
_
(N)
2N
0
dx
=
_
N,

0

2
0
_
/(N). (28)
We note that (28) is also valid for Rayleigh fading and unfaded
channels.
i.i.d. Rayleigh fading: For i > 0 and i.i.d. Rayleigh fading
channels,

h = 0,
2
h
= 1, and z
i
simplies to a central chi-
square random variable with 2i degrees of freedom. Therefore,
P(u
0
= 1|i) can now be written as
P(u
0
= 1|i) =
1
(N)(i)

_
0
z
i1
i
e
z
i

_
N,

0
Ez
i
/N +
2
0
_
dz
i
, i > 0. (29)
Identical unfaded channels: In this case, z
i
= i, and i 0,
and we obtain
P(u
0
= 1|i) =

0
x
N1
exp
_
x/
_
Ez
i
/N +
2
0
__
(N) (Ez
i
/N +
2
0
)
N
dx
=
_
N,

0
Ei/N +
2
0
_
/(N). (30)
Discussion: Although the upper limit of the integrals in (27)
and (29) is innity, the integrals can be evaluated efciently,
as the integrands quickly decrease for increasing z
i
. Therefore,
(23) and (24) and (27)(30) can be used for the efcient

(s) =
exp
_
sE(G

S
)
H
_
I
N
s
_
EG

S
G
H

S
+
2
0
I
N
__
1
G

S
_
det
_
I
N
s
_
EG

S
G
H

S
+
2
0
I
N
__ (21)
YIU AND SCHOBER: NONORTHOGONAL TRANSMISSION AND NONCOHERENT FUSION OF CENSORED DECISIONS 269
Fig. 2. System probability of detection P
d
0
versus channel SNR (E
s
/N
0
)
for optimum and energy-based fusion rules. Gaussian and normalized Gaussian
(g
k

2
2
= 1, k K) SVs with N = 2, i.i.d. Rayleigh fading channels,
10 log
10
(SNR) = 5 dB at sensors, and P
f
0
= 0.01.

markers indicate
simulation points.
performance evaluation of the energy-based fusion rule

ene
(y) for random SVs. Note that the number of terms in (23)
and (24) increases only linearly with K.
VI. NUMERICAL AND SIMULATION RESULTS
Here, we present numerical and simulation results for the
proposed nonorthogonal signaling scheme and fusion rules.
Throughout this section, we assume that the sensor observations
are conditionally i.i.d. across the sensors, i.e., identical sensors
employing the same local decision thresholds
k
= , k K,
and P(H
0
) = P(H
1
) = 0.5. Unless stated otherwise, we also
assume a local SNR of 5 dB at each sensor, i.i.d. channels, and
a xed system probability of false detection of P
f
0
= 0.01 (P
f
0
is adjusted via
0
).
P
d
0
versus E
s
/N
0
: In Fig. 2, we show the system probability
of detection P
d
0
as a function of the channel SNR (E
s
/N
0
) for
WSNs with K = 8 and K = 30 sensors, respectively. We as-
sume i.i.d. Rayleigh fading and N = 2. For each channel SNR
value, the local decision threshold was numerically optimized
for the maximization of P
d
0
, assuming the energy-based fusion
rule. For complexity reasons, the same value for was then
also used for the optimum fusion rules. For the optimization
of , using the numerical methods outlined in Section V, we
generated P
d
0
versus P
f
0
curves by varying
0
(cf. Fig. 5)
for a large set of different values. Subsequently, we selected
from the set that achieved the largest P
d
0
for the desired
P
f
0
. This search entails comparatively high complexity but
can be performed ofine. In Fig. 2, we consider the Gaussian
and normalized Gaussian SVs. In the normalized Gaussian
case, we assume that the FC knows the SVs, which allows
the application of
opt,D
(y).
2
The curves for the optimum
fusion rules were generated by computer simulation, whereas
2
Strictly speaking, the normalized Gaussian SVs are still random, of course.
However, adopting random SVs also for the evaluation of
opt,D
(y) has the
advantage that our results do not depend on a particular set of SVs.
Fig. 3. System probability of detection P
d
0
versus the sensor SNR for
optimum and energy-based fusion rules. Gaussian and normalized Gaussian
(g
k

2
2
= 1, k K) SVs with N = 2, i.i.d. Rayleigh fading channels,
10 log
10
(E
s
/N
0
) = 15 dB, and P
f
0
= 0.01.
the curves for the energy-based fusion rules were calculated
using the results from Section V and conrmed by simula-
tions for K = 30 (

markers). We observe from Fig. 2 that


1) normalized Gaussian SVs perform better than unnormalized
Gaussian SVs and that 2) for K = 8, the optimum fusion
rules lead to better performance than the energy-based fusion
rules. However, in both cases, the performance differences are
small. We note that we have made very similar observations
for non-i.i.d. channels and scenarios, where the local sensors
had different SNRs. For K = 30, we were not able to simulate
the optimum fusion rules due to their high computational com-
plexity. For high E
s
/N
0
values, the P
d
0
curves become inde-
pendent of E
s
/N
0
, and P
d
0
is limited by local sensor decision
errors.
P
d
0
versus SNR: In Fig. 3, we keep the channel SNR con-
stant [10 log
10
(E
s
/N
0
) = 15 dB] and show P
d
0
as a function
of the sensor SNR. Otherwise, the same parameters as for Fig. 2
were adopted, and was optimized for each sensor SNR value,
assuming the energy-based fusion rule. We observe from Fig. 3
that, as expected from Sections III-B2 and IV-B, for high-
sensor SNRs (corresponding to reliable sensor decisions, i.e.,
P
f
0 and P
d
1), the energy-based fusion rule approaches
the performance of the respective optimum fusion rule for the
random and deterministic cases. We note that, even for very
high sensor SNRs, P
d
0
< 1 holds since the nite-channel SNR
causes transmission errors.
Since Figs. 2 and 3 suggest that the performance loss that is
incurred by unnormalized SVs is small and that the complexity
of the optimum fusion rules is prohibitive for moderate-to-
large values of K, we will concentrate, in the following, on
unnormalized Gaussian SVs and the energy-based fusion rule.
P
d
0
versus : Fig. 4 shows P
d
0
as a function of the local
sensor decision threshold for K = 8 and K = 30 and dif-
ferent values of N. i.i.d. Rayleigh fading, 10 log
10
(E
s
/N
0
) =
15 dB, and P
f
0
= 0.01 were adopted. Fig. 4 clearly shows
that has a large effect on the performance and, thus, in
270 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 1, JANUARY 2009
Fig. 4. System probability of detection P
d
0
as a function of the local
sensor decision threshold for WSNs with K = 8 and K = 30 sensors and
various N. Energy-based fusion rule, Gaussian SVs, i.i.d. Rayleigh fading
channels, 10 log
10
(E
s
/N
0
) = 15 dB, 10 log
10
(SNR) = 5 dB at sensors, and
P
f
0
= 0.01.
general, should be optimized.
3
For K = 8 sensors, 40 is
optimum for all considered N, which corresponds to an average
number of E{|SH
1
} 0.95 sensors in S if H = H
1
. This
implies that if the target is present most of the time, only one
sensor will transmit its information to the FC, whereas the
other sensors will be silent. This also means that for K = 8,
letting one sensor transmit a very reliable decision (because
of large ) with maximum power is preferable over having
multiple sensors transmit less-reliable decisions (because of
smaller ) at lower power but with a possible diversity gain. At
the optimum operating point for K = 8, increasing N has no
signicant effect on the performance since a diversity gain is
not achievable when only one sensor transmits. The situation
is different for the WSN with K = 30 sensors. In this case,
for N 4, the best performance can be achieved for relatively
small values of , implying that multiple sensors concurrently
transmit, enabling a diversity gain. For example, for = 8,
which is optimum for N = 4, on average, E{|SH
1
} 11.7
sensors transmit their decisions to the FC. However, for N = 1
and N = 2, where the maximum diversity gain is limited to one
and two, respectively, the optimum operating point is achieved
for 80, implying that very few sensors transmit highly
reliable decisions at relatively high power.
P
d
0
versus P
f
0
: In Fig. 5, we show the receiver operating
characteristic (ROC) for a WSN with K = 30 sensors. The
same parameters as in Fig. 4 are used, and the sensor decision
threshold is optimized to maximize P
d
0
for P
f
0
= 0.01
(cf. Fig. 4). In addition, for N = 1 and N = 2, we also show
the results for the suboptimum choice = 8. It is interesting to
note that the ROC curves for N = 1 and N = 2 with optimum
(optimum for P
f
0
= 0.01) have a completely different shape
than the curves for N 4. In particular, for N = 1 and N = 2,
3
We note that Fig. 4 shows that P
d
0
has local maxima for certain values of
N and K. Therefore, the analytical optimization of using, for example, the
well-known tools from convex optimization does not seem to be possible.
Fig. 5. ROC for WSNs with K = 30 sensors. Energy-based fusion rule,
Gaussian SVs, i.i.d. Rayleigh fading channels, 10 log
10
(E
s
/N
0
) = 15 dB,
and 10 log
10
(SNR) = 5 dB at sensors.
Fig. 6. System probability of detection P
d
0
versus N for a network with
K = 30 sensors. Energy-based fusion rule, Gaussian SVs, i.i.d. Ricean fading
channels, channel SNR 10 log
10
(E
s
/N
0
) = 15 dB, 10 log
10
(SNR) = 5 dB
at sensors, and P
f
0
= 0.01.
the performance quickly degrades if P
f
0
deviates from the
design value of 0.01. On the other hand, the curves for N = 1
and N = 2 obtained for the suboptimum = 8 show more
robust behavior with respect to deviations from P
f
0
= 0.01.
The reason for this increased robustness is that for = 8,
multiple sensors participate in the transmission, whereas for
the optimum 80, only one or two sensors transmit their
decisions to the FC most of the time.
P
d
0
versus N> Fig. 6 shows P
d
0
as a function of N for
a WSN with K = 30 sensors and i.i.d. Ricean fading with
different Ricean factors . The cases of optimized and = 8
are considered. Optimizing leads to noticeable gains only for
N < 4. For N 4, a xed threshold of = 8 leads to close-
to-optimum performance (cf. Fig. 4). As expected, performance
improves with increasing Ricean factor . More interestingly,
for all considered Ricean factors, increasing N beyond N = 10
leads only to small performance improvements. In general,
YIU AND SCHOBER: NONORTHOGONAL TRANSMISSION AND NONCOHERENT FUSION OF CENSORED DECISIONS 271
Fig. 7. System probability of detection P
d
0
versus K. Energy-based
fusion rule, Gaussian SVs with N = 2, i.i.d. Ricean fading channels,
10 log
10
(E
s
/N
0
) = 15 dB, 10 log
10
(SNR) = 5 dB at sensors, and
P
f
0
= 0.01.
decreasing N leads to some loss in power efciency (e.g., a
smaller P
d
0
for given P
f
0
) but also to improved bandwidth
efciency (due to the smaller symbol duration T
s
= NT
c
for
given chip duration T
c
). Fig. 6 nicely illustrates that N can
be used to nd a good tradeoff between power and bandwidth
efciency in WSNs. As mentioned before, the special case N =
K and orthogonal SVs was considered in [10]. We have veried
that orthogonal SVs g
k
= e
k
, k K achieve practically the
same performance as that shown in Fig. 6 for the random SVs
with N = K. Thus, Fig. 6 shows that the bandwidth efciency
of the scheme in [10] can be substantially improved by the
proposed scheme using N < K at the expense of a small loss
in power efciency.
P
d
0
versus K: In Fig. 7, we show P
d
0
as a function of the
WSN size K, assuming N = 2 and i.i.d. Ricean fading. Al-
though the average channel SNR E
s
/N
0
is independent of K,
increasing K signicantly improves P
d
0
. This clearly shows the
advantage of large WSNs over small ones. In addition, for the
considered case, optimizing yields signicant performance
gains due to the small value of N.
i.n.d. Rayleigh fading. In Fig. 8, we show P
d
0
as a function
of N for a WSN with K = 30 nodes. The nodes are uniformly
distributed in a circle with radius r, and the distance of the
center of the circle to the FC is d. We assume i.n.d. Rayleigh
fading, where the average received power decreases as d
3
k
, and
d
k
is the distance between sensor k and the FC. Local decision
threshold was optimized for all curves and N shown. Fig. 8
shows that r/d = 0 (corresponding to i.i.d. fading) yields the
best performance, and P
d
0
decreases as the area to be covered
by the WSN increases. We note that the optimal fusion rule
requires the knowledge of the sensor-to-FC channel variance of
all sensors. Due to the dynamic nature of the sensor nodes, this
information is usually not available at the FC, and therefore, we
did not include the results for the optimal fusion rule.
VII. CONCLUSION
In this paper, we have proposed a novel bandwidth-efcient
signaling scheme for WSNs, where nonorthogonal SVs are
Fig. 8. System probability of detection P
d
0
versus N for a network with
K = 30 sensors. Energy-based fusion rule, Gaussian SVs, i.n.d. Rayleigh fad-
ing channels, 10 log
10
(E
s
/N
0
) = 15 dB, 10 log
10
(SNR) = 5 dB at sensors,
and P
f
0
= 0.01.
used to transmit censored binary decisions to the FC. As a
consequence of the censoring and the practical limitations of
large-scale WSNs, noncoherent data fusion is preferable, and
we have derived optimum LR-based fusion rules for determin-
istic and random SVs, respectively. In the former case, it has
been assumed that the FC knows the SVs, whereas in the latter
case, this knowledge is not required. Since the complexity of
the optimum fusion rules rapidly increases with the WSN size,
we have derived a low-complexity suboptimum energy-based
fusion rule, which is also amenable to analysis. Our simulation
and analytical results show the following.
1) The performance loss suffered by the energy-based fusion
rule compared with the optimum fusion rules is relatively
small.
2) Normalized SVs perform only slightly better than unnor-
malized SVs.
3) Proper censoring is important, particularly if short SVs
are used.
4) The proposed nonorthogonal signaling scheme enables
large savings in bandwidth at the expense of a small loss
in power efciency.
Therefore, we consider the proposed scheme a promising solu-
tion for cooperative communication in large-scale WSNs.
APPENDIX
ENERGY-BASED FUSION RULE
Here, we show that
ene
(y) is equivalent to the optimum
fusion rules
opt,D
(y) and
opt,R
(y) under certain conditions.
A. Deterministic SVs
Here, we assume i.i.d. Rayleigh fading,
2
0
, P
d
k
=
P
d
> P
f
k
= P
f
, k K, and G
K
G
H
K
= (K/N)I
N
. Exploiting
272 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 1, JANUARY 2009
the i.i.d. Rayleigh fading assumption and neglecting all irrele-
vant terms, (7) can be simplied to
f(y| u)
exp
_
y
H
_
EG

S
G
H

S
+
2
0
I
N
_
1
y
_
det
_
EG

S
G
H

S
+
2
0
I
N
_ . (31)
For
2
0
, (EG

S
G
H

S
+
2
0
I
N
)
1
(I
N
G

S
G
H

S
)/

2
0
, and det(EG

S
G
H

S
+
2
0
I
N
)
2N
0
(1 +tr{G

S
G
H

S
}) =

2N
0
(1 +|

S|), where

= E/
2
0
, and (31) can be written as
f(y| u)
exp
_
y
H
y/
2
0
_
exp
_
y
H
G

S
G
H

S
y/
2
0
_

2N
0
_
1 +|

S|
_ . (32)
Dropping all irrelevant terms and using exp(x) 1 +x and
1/(1 +x) 1 x for x 0, (32) becomes
f(y| u)
_
1 +y
H
G

S
G
H

S
y/
2
0
_
_
1 |

S|
_
. (33)
Using (33) in (6) and neglecting all higher order terms of
result in

opt,D
(y)

u
_
1 +y
H
G

S
G
H

S
y/
2
0
|

S|
_
P( u|H
1
)

u
_
1 +y
H
G

S
G
H

S
y/
2
0
|

S|
_
P( u|H
0
)
=
1 +

u
_
y
H
G

S
G
H

S
y/
2
0
|

S|
_
P( u|H
1
)
1 +

u
_
y
H
G

S
G
H

S
y/
2
0
|

S|
_
P( u|H
0
)
.
The above equation can be further simplied by exploiting
ln[(1 +x
1
)/(1 +x
2
)] x
1
x
2
for x
1
, x
2
0, and the
monotonicity property of the logarithm. This leads to

opt,D
(y)

u
c

S
_
y
H
G

S
G
H

S
y/
2
0
|

S|
_
y
H
My
(34)
where c

= P( u|H
1
) P( u|H
0
) and M

=

u
c

S
G

S
G
H

S
,
and we used the fact that |

S| is independent of y. For identical


sensors (i.e., P
d
k
= P
d
, P
f
k
= P
f
, k K), M can be reformu-
lated as
M =
K

i=0
c
i

S
|

S|=i
G

S
G
H

S
(35)
with c
i
= P
i
d
(1 P
d
)
Ki
P
i
f
(1 P
f
)
Ki
. With the iden-
tity


S
|

S|=i
G

S
G
H

S
=
_
K
i
_
(i/K)G
K
G
H
K
and the assumption
that the set of SVs forms a tight frame, i.e., G
K
G
H
K
=
(K/N)I
N
, after tedious but straightforward operations, we
obtain M = I
N
, where

=

K
i=0
_
K
i
_
(i c
i
/N) = K(P
d

P
f
)/N. This shows that the energy-based fusion rule
ene
(y)
is equivalent to
opt,D
(y) if the assumptions that are outlined
above are fullled.
B. Random SVs: P
d
k
1 and P
f
k
0, k K
Assuming P
d
k
1 and P
f
k
0 and neglecting all irrele-
vant terms, it can be shown that [cf. (6), (8), (9), and (12)]

opt,R
(y)

_
0
exp
_
y
H
y
_
1/
2
0
1/
_
Ez
K
/N +
2
0
__
(Ez
K
/N +
2
0
)
N
f
z
K
(z
K
)dz
K
(36)
which is strictly monotonic in y
H
y. Therefore,
opt,R
(y) is
equivalent to
ene
(y) in this case.
C. Random SVs:
2
0
and P
d
k
= P
d
> P
f
k
= P
f
, k K
Using similar approximations as in the deterministic case
(cf. Appendix A), we can rewrite f(y| u, z

S
) [cf. (13)] for

2
0
as
f(y| u, z

S
) 1 +z

S
y
H
y/
2
0
z

S
N (37)
where we have neglected all higher order terms of , which is
now dened as

= E/(N
2
0
). Based on (37), f(y| u) can be
obtained as
f(y| u)

= E
_
f(y| u, z

S
)
_
= 1 +b
2

S
_
y
H
y/
2
0
N
_
(38)
with b
2

= E{z

S
} = a
2

S
+

S

2
h
k
and a
2

S
, as dened in
Section IV-A1. With (38), we obtain for the optimum fu-
sion rule

opt,R
(y)

u
_
1 +b
2

S
_
y
H
y/
2
0
N
_
_
P( u|H
1
)

u
_
1 +b
2

S
(y
H
y/
2
0
N)
_
P( u|H
0
)

u
c

S
b
2

S
_
y
H
y/
2
0
N
_
(39)
where we have used, again, the same approximations as in
the deterministic case (cf. Appendix A). Equation (39) shows
that
ene
(y) is equivalent to
opt,R
(y) for
2
0
if

=

u
c

S
b
2

S
> 0. Assuming identical sensors and using a similar
approach as in Appendix A, it can be shown that is given by
= (P
d
P
f
)

kK
(|

h
k
|
2
+
2
k
), that is,
ene
(y) is equiva-
lent to
opt,R
(y) for
2
0
if P
d
> P
f
.
REFERENCES
[1] R. Tenney and N. Sandell, Jr., Detection with distributed sensors, IEEE
Trans. Aerosp. Electron. Syst., vol. AES-17, no. 4, pp. 501510, Jul. 1981.
[2] J. N. Tsitsiklis, Decentralized detection, in Advances in Statistical
Signal Processing, vol. 2. Greenwich, CT: JAI, 1993, pp. 297344.
[3] R. Viswanathan and P. K. Varshney, Distributed detection with multiple
sensors: Part IFundamentals, Proc. IEEE, vol. 85, no. 1, pp. 5463,
Jan. 1997.
[4] R. S. Blum, S. A. Kassam, and H. V. Poor, Distributed detection with
multiple sensors: Part IIAdvanced topics, Proc. IEEE, vol. 85, no. 1,
pp. 6479, Jan. 1997.
[5] P. K. Varshney, Distributed Detection and Data Fusion. New York:
Springer-Verlag, 1997.
[6] T. Q. S. Quek, D. Dardari, and M. Z. Win, Energy efciency of dense
wireless sensor networks: To cooperate or not to cooperate, IEEE J. Sel.
Areas Commun., vol. 25, no. 2, pp. 459470, Feb. 2007.
YIU AND SCHOBER: NONORTHOGONAL TRANSMISSION AND NONCOHERENT FUSION OF CENSORED DECISIONS 273
[7] J.-J. Xiao and Z.-Q. Luo, Universal decentralized detection in a
bandwidth-constrained sensor network, IEEE Trans. Signal Process.,
vol. 53, no. 8, pp. 26172624, Aug. 2005.
[8] B. Chen, R. Jiang, T. Kasetkasem, and P. K. Varshney, Channel aware de-
cision fusion in wireless sensor networks, IEEE Trans. Signal Process.,
vol. 52, no. 12, pp. 34543458, Dec. 2004.
[9] R. Niu, B. Chen, and P. K. Varshney, Fusion of decisions transmitted
over Rayleigh fading channels in wireless sensor networks, IEEE Trans.
Signal Process., vol. 54, no. 3, pp. 10181027, Mar. 2006.
[10] R. Jiang and B. Chen, Fusion of censored decisions in wireless sensor
networks, IEEE Trans. Wireless Commun., vol. 4, no. 6, pp. 26682673,
Nov. 2005.
[11] C. Rago, P. K. Willett, and Y. Bar-Shalom, Censoring sensors: A low-
communications-rate scheme for distributed detection, IEEE Trans.
Aerosp. Electron. Syst., vol. 32, no. 2, pp. 554568, Apr. 1996.
[12] S. K. Jayaweera, Large sensor system performance of decentralized
detection in noisy, bandlimited channels, in Proc. IEEE VTCSpring,
Stockholm, Sweden, May 2005, pp. 10961100.
[13] S. K. Jayaweera, Optimal Bayesian data fusion and low-complexity
approximations for distributed DS-CDMA wireless sensor networks in
Rayleigh fading, in Proc. IEEE Int. Conf. Intell. Sens. Inf. Process.,
Jan. 2005, pp. 1924.
[14] J. Dyer, B. Natarajan, and S. K. Jayaweera, Optimal Bayesian
data fusion and low-complexity approximations for asynchronous DS-
CDMA wireless sensor networks in Rayleigh fading, in Proc. GLOBE-
COM, Dec. 2006, pp. 16.
[15] S. K. Jayaweera, Bayesian fusion performance and system optimization
for distributed stochastic Gaussian signal detection under communication
constraints, IEEE Trans. Signal Process., vol. 55, no. 4, pp. 12381250,
Apr. 2007.
[16] W. Li and H. Dai, Distributed detection in wireless sensor networks using
a multiple access channel, IEEE Trans. Signal Process., vol. 55, no. 3,
pp. 822833, Mar. 2007.
[17] J.-J. Xiao, A. Ribeiro, Z.-Q. Luo, and G. B. Giannakis, Distributed
compression-estimation using wireless sensor networks, IEEE Signal
Process. Mag., vol. 23, no. 4, pp. 2741, Jul. 2006.
[18] J. A. Tropp, I. S. Dhillon, R. W. Heath, and T. Strohmer, Designing
structured tight frames via an alternating projection method, IEEE Trans.
Inf. Theory, vol. 51, no. 1, pp. 188209, Jan. 2005.
[19] X. Li, M. Chen, and W. Liu, Application of STBC-encoded cooperative
transmissions in wireless sensor networks, IEEE Signal Process. Lett.,
vol. 12, no. 2, pp. 134137, Feb. 2005.
[20] S. Yiu, R. Schober, and L. Lampe, Distributed space-time block coding,
IEEE Trans. Commun., vol. 54, no. 7, pp. 11951206, Jul. 2006.
[21] B. S. Mergen and A. Scaglione, Randomized space-time coding for
distributed cooperative communication, in Proc. ICC, Istanbul, Turkey,
Jun. 2006, pp. 45014506.
[22] M. K. Simon, Probability distributions involving Gaussian random vari-
ables, in A Handbook for Engineers and Scientists, 1st ed. Norwell,
MA: Kluwer, 2002.
[23] M. Schwartz, W. Bennett, and S. Stein, Communication Systems and
Techniques. New York: McGraw-Hill, 1966.
[24] E. Biglieri, G. Caire, G. Taricco, and J. Ventura-Traveset, Computing
error probabilities over fading channels: A unied approach, Eur. Trans.
Telecommun., vol. 9, no. 1, pp. 1525, Jan./Feb. 1998.
[25] J. K. Cavers and P. Ho, Analysis of the error performance of trellis-
coded modulations in Rayleigh-fading channels, IEEE Trans. Commun.,
vol. 40, no. 1, pp. 7483, Jan. 1992.
Simon Yiu (S04M08) was born in Hong Kong in
1980. He received the Ph.D. degree in electrical and
computer engineering from the University of British
Columbia, Vancouver, BC, Canada, in 2007.
Since January 2008, he has been a Postdoctoral
Research Fellow with the School of Engineering and
Applied Sciences, Harvard University, Cambridge,
MA. His research interests span the broad area
of wireless communication systems, with particular
emphasis in cognitive networks, cooperative diver-
sity systems, MIMO systems, and space-time coding
theory.
Robert Schober (M01SM07) was born in
Neuendettelsau, Germany, in 1971. He received the
Diplom (Univ.) and Ph.D. degrees in electrical engi-
neering from the University of Erlangen-Nrnberg,
Erlangen, Germany, in 1997 and 2000, respectively.
From May 2001 to April 2002, he was a Postdoc-
toral Fellow with the University of Toronto, Toronto,
ON, Canada, sponsored by the German Academic
Exchange Service (DAAD). Since May 2002, he has
been with the University of British Columbia (UBC),
Vancouver, BC, Canada, where he is currently an
Associate Professor and the Canada Research Chair (Tier II) in Wireless
Communications. His research interests include cooperative diversity systems,
equalization, UWB, MIMO systems, and space-time processing and coding.
Dr. Schober is the Area Editor for Modulation and Signal Design for the
IEEE TRANSACTIONS ON COMMUNICATIONS. He was the recipient of the
2002 Heinz Maier-Leibnitz Award of the German Science Foundation (DFG),
the 2004 Innovations Award of the Vodafone Foundation for Research in
Mobile Communications, the 2006 UBC Killam Research Prize, and the 2007
Friedrich Wilhelm Bessel Research Award of the Alexander von Humboldt
Foundation. In addition, he was the recipient of the best paper awards from
the German Information Technology Society (ITG), the European Association
for Signal, Speech, and Image Processing, the 2006 IEEE International Con-
ference on Ultra-Wideband, the International Zurich Seminar on Broadband
Communications, and European Wireless 2000.

Anda mungkin juga menyukai