Title:
Neural Network Augmented Kalman Filtering in the Presence of Unknown System Inputs
Neural Network Augmented Kalman Filtering in the Presence of Unknown System Inputs
Authors
Sattigeri, Ramachandra J.
Calise, Anthony J.
Calise, Anthony J.
Authors
Advisors
Advisors
Associated Organizations
Organizational Unit
Organizational Unit
Organizational Unit
Series
Collections
Supplementary to
Permanent Link
Abstract
We present an approach for augmenting a linear, time-varying Kalman filter with an adaptive neural
network (NN) for the state estimation of systems with linear process models acted upon by unknown inputs.
The application is to the problem of tracking maneuvering targets. The unknown system inputs represent the
effect of unmodeled disturbances acting on the system and are assumed to be continuous and bounded. The
NN is trained online to estimate the unknown inputs. The training signal for the NN consists of two error
signals. The first error signal is the residual of the Kalman filter that is augmented with the NN output. The
second error signal is obtained after deriving a linear parameterization model of available system signals in
terms of the ideal, unknown NN weights that linearly parameterize the unknown system inputs. The
combination of two different sources of error signals to train the NN represents a composite adaptation type
approach to adaptive state estimation. The approach is applied in a vision-based formation flight simulation
of a leader and a follower unmanned aerial vehicle (UAV). The adaptive estimator onboard the follower UAV
estimates the range, azimuth angle, and elevation angle to the leader UAV, the derivatives of these LOS
variables, and the unknown leader aircraft acceleration along the axes of the Cartesian coordinate inertial
frame. Simulation results with the presented approach are greatly improved when compared to those
obtained with just a linear, time-varying Kalman filter and a particular adaptive state estimation method that
utilizes just one source of error signals to train the NN [17].
Sponsor
Date Issued
2006-08
Extent
Resource Type
Text
Resource Subtype
Proceedings