Abstract—
In this paper, we address the problem of nonlinear
sensor dynamic compensation that will be performed on board
wireless sensor network nodes. To this aim, we design suitable
reduced-complexity learning-from-example algorithms and implement
them on resource-constrained devices, namely, 8-bit
microcontrollers. The proposed approach is validated with simulations
on different examples of nonlinear sensor models.
Index Terms—Dynamic compensation, low power, microcontroller
implementation, reduced-set methods, support vector machines
for regression (SVRs), wireless sensor networks (WSNs).
I. INTRODUCTION
WIRELESS sensor networks (WSNs) are becoming a
widely used technology [1]. Because of the new “distributed”
way of acquiring and processing data, which has
been made possible by the explosion of the microelectronics
process, an appealing paradigm of computing, called pervasive
computing [2], has been introduced, where each node
receives/transmits information and executes some local
processing to save transmission band. In this sense, new intelligent
techniques for local classification or signal detection in
WSNs have been proposed [3].
One of the main aspects that has to be faced when dealing
with sensor networks concerns the intrinsic behavior of the
sensor itself, because nonlinear distortions could modify the
expected output. This is why compensation techniques that are
based on inverse modeling criteria are usually employed to reconstruct
a given input signal on the basis of the corresponding
output.
Although several classical techniques have been employed to
solve these kinds of problems [4]–[7], artificial neural networks
(ANNs) have recently become the state-of-the-art approach for
solving this class of problems [8], [9]. The main advantage of
using ANNs with respect to traditional methods consists of the
fact that no a priori information on the model of the system is
required; instead, only a set of input–output measures is used to
infer a general rule that models the given sensor.
The use of a different well-known learning-from-example
approach, namely support vector machines for regression
(SVRs), has been proposed in [10] for the dynamic com-
Manuscript received July 5, 2007; revised December 13, 2007.
A. Marconato, A. Boni, and D. Petri are with the Department of Information
and Communication Technology, University of Trento, 38100 Trento, Italy.
M. Hu is with the Institute of Computing Technology, Chinese Academy of
Sciences, Beijing 100080, China.
Digital Object Identifier 10.1109/TIM.2008.922074
Fig. 1. General scheme of the proposed approach.
pensation of sensors based on inverse modeling. This choice
is motivated by the fact that SVRs have shown interesting
properties when compared to ANNs; in particular, they do not
suffer from local minima problems during the optimization
process, and they have proved to be theoretically sound, with
their foundation being Vapnik’s statistical learning theory [11].
In this paper, we first address the problem of designing suitable
SVR-based methodologies to compensate sensor nonidealities
by studying two different nonlinear sensor models. As a
further step, we discuss the implementation of such algorithms
on resource-constrained devices like simple 8-bit microcontrollers
that usually equip WSNs. In this context, modification
of the original SVR is proposed to reduce the complexity of the
model, thus fitting the hardware constraints that are given by
the available platform. A new SVR-like method called extended
reduced SVR (ERSVR) is thus introduced.
This paper is organized as follows. In Section II, we briefly
describe the proposed methodology and provide the details of
the SVR algorithms that are addressed in this paper. Several
microcontroller implementation issues that are closely related
to limits of the fixed-point representation are discussed in
Section III, whereas experimental settings and results are provided
in Section IV. Some final comments are summarized in
Section V.


Download full report
http://ieeexplore.ieeeiel5/19/4567685/04...er=4512331