Adaptive Neural Network Based Target Tracking

Design of nonlinear observers has received
considerable attention since the early development
of methods for state estimation. The most
popular approach is the extended Kalman filter (EKF)
that goes through significant degradation in the
presence of unmodeled nonlinearities. For uncertain
nonlinear systems, adaptive observers have been
introduced to estimate the unknown parameters where
no apriori information about the unknown parameters
is available. While establishing global results,
these approaches are only applicable to systems
transformable to output feedback form. Over the
recent years, neural network (NN) based
identification and estimation schemes have been
proposed that relax the assumptions on the
system at the price of sacrificing on the global
nature of the results. However, most of the NN based
adaptive observers in the literature require
knowledge of the full dimension of the system,
therefore may not be suitable for systems with
unmodeled dynamics. A novel approach to nonlinear
state estimation, robust to unmodeled dynamics, is
proposed from the perspective of augmenting an EKF
with an NN based adaptive element.

Autorentext

Venky Madyastha obtained his doctoral degree in the year 2005
from the school of aerospace engineering, Georgia Institute of
Technology, USA. His doctoral research focussed on adaptive
nonlinear state estimation for control of uncertain nonlinear
systems. He is currently with the General Electric Global
Research Center, Bangalore, India.

Klappentext

Design of nonlinear observers has received
considerable attention since the early development
of methods for state estimation. The most
popular approach is the extended Kalman filter (EKF)
that goes through significant degradation in the
presence of unmodeled nonlinearities. For uncertain
nonlinear systems, adaptive observers have been
introduced to estimate the unknown parameters where
no apriori information about the unknown parameters
is available. While establishing global results,
these approaches are only applicable to systems
transformable to output feedback form. Over the
recent years, neural network (NN) based
identification and estimation schemes have been
proposed that relax the assumptions on the
system at the price of sacrificing on the global
nature of the results. However, most of the NN based
adaptive observers in the literature require
knowledge of the full dimension of the system,
therefore may not be suitable for systems with
unmodeled dynamics. A novel approach to nonlinear
state estimation, robust to unmodeled dynamics, is
proposed from the perspective of augmenting an EKF
with an NN based adaptive element.