Scribd
Upload
8 views

Classification of sensor networks

Uploaded by

quang pham
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
8 views

Classification of sensor networks

Uploaded by

quang pham
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224754088

Classification in sensor networks

Conference Paper · August 2004


DOI: 10.1109/ISIT.2004.1365289 · Source: IEEE Xplore

CITATIONS READS
11 158

4 authors:

Venkatesh Saligrama Murat Alanyali


Boston University TOBB University of Economics and Technology
308 PUBLICATIONS 9,714 CITATIONS 73 PUBLICATIONS 1,193 CITATIONS

SEE PROFILE SEE PROFILE

Onur Savas Shuchin Aeron


Intelligent Automation, Inc. Tufts University
13 PUBLICATIONS 346 CITATIONS 167 PUBLICATIONS 3,427 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Venkatesh Saligrama on 10 September 2015.

The user has requested enhancement of the downloaded file.


ISIT 2004, Chicago, USA, June 27 – July 2, 2004

Classification in Sensor Networks


Saligrama Venkatesh, Murat Alanyali, Onur Savas, and Shuchin Aeron
Dept. of Electrical and Computer Engineering
Boston University, Boston, MA 02215
{srv, alanyali, savas, shuchin}@bu.edu

Abstract — We consider the problem of classifying Moreover, the network topologies for which these results hold
among a set of M hypothesis with N distributed noisy are restrictive.
sensors. The N sensors can collaborate over a finite To overcome these issues we develop a data-centric as op-
link-capacity network. The task is to arrive at a con- posed to decision-centric perspective. We consider an arbi-
sensus about the event after exchanging such mes- trary network of N sensors that must classify among a set
sages. In contrast to the conventional decentralized of M different hypothesis. Our problem focuses on deriving
detection approach, wherein the bit rates for each conditions for arriving at a consensus at all the sensors and
link is explicitly constrained, our approach is based situations where the consensus is the centralized MAP esti-
on high-rate limit perspective. We apply a variant mate. A natural idea for collaboration is to exchange a vector
of belief propagation as a strategy for collaboration of individual sensor beliefs (probabilities) for different hypoth-
to arrive at a solution to the distributed classification esis between linked sensors at any instant of time. This idea is
problem. We show that the message evolution can be formalized in the “so called” belief propagation algorithm [2].
re-formulated as the evolution of a linear dynamical A description is shown in Figure 2 where sensor nodes send
system, which is primarily characterized by network a vector of likelihoods for each hypothesis at any instant of
connectivity. It turns out that consensus is almost time. These likelihoods can be dynamically updated based on
always reached by the sensors for any arbitrary net-
work. We then derive conditions under which the
consensus is the centralized MAP estimate and show
that this is achieved with O(M log2 N ) bits.
The standard decentralized detection setup deals with N
sensors having communication links to a data fusion center as
shown in Fig. 1. The data from each sensor is compressed on to

Figure 2: Schematics of Distributed Detection

information received by the sensor in the past. Evidently, the


algorithm overcomes the centralization issue alluded to earlier.
In this setup, we neither have a fusion center nor does each
sensor need to know models for adjacent sensors. However,
belief propagation is known to work generally for non-loopy
Figure 1: Schematics of Decentralized Detection network topologies, a situation that is quite restrictive and
difficult to impose in a sensor network. Furthermore, on ac-
count of finite link capacity, it is unclear as to how to deal
a message taking values over a finite alphabet. In this respect, with attendant effects of quantization. Nevertheless, we show
the finite link capacity constraint is explicitly modelled in the that for the class of problems where all sensors are engaged in
setup. The objective is to find optimal fusion rule and the the same classification task, consensus can indeed be attained.
associated decision rules at each of the individual sensors to Furthermore, this consensus is generically the MAP estimate
minimize the error probability. The principle drawbacks of asymptotically for random graph configurations. The finite
the approach are well-known and has been documented in [1]. link capacity is dealt with by showing that the algorithm is
We point out some of these here for the sake of exposition. robust to quantization errors. Consequently, superior energy
It can be shown that (if the hypothesis were conditionally efficiencies as well as scalability with increasing network size
independent) the decision rules at each sensor reduce to a can be realized over existing schemes.
likelihood ratio threshold test. Nevertheless, the decision rules
are coupled in that thresholds have to be solved jointly for all References
the sensors. This not only has computational implications
[1] J. N. Tsitsiklis, “Decentralized detection,” in Advances in Sta-
but also assumes centralized knowledge of the sensor models. tistical Signal Processing, H. V. Poor and J. B. Thomas Eds,
vol. 2. JAI Press, 1993
1 This work was supported by ONR Young Investigator Award
[2] J. S. Yedidida, W. T. Freeman, Y. Weiss, “Generalized Belief
N00014-02-100362 and NSF CAREER Program under grant ANI- Propagation,” NIPS 13, 2001,
0238397.

‹,((( 

View publication stats

You might also like