Core Concepts
This work presents a Kalman filter least squares (KFLS) cost function whose recursive minimizer gives the Kalman filter update equations. It is shown that various extensions of recursive least squares (RLS) from the literature are special cases of the Kalman filter, motivating the development of a new class of adaptive Kalman filters that incorporate forgetting from RLS extensions.
Abstract
The paper begins by deriving the Kalman filter least squares (KFLS) cost function, whose recursive minimizer gives the Kalman filter update equations. This is done by extending the generalized forgetting recursive least squares (GF-RLS) framework, which contains various RLS extensions as special cases.
The key insights are:
RLS extensions that are special cases of GF-RLS are also special cases of the Kalman filter, with a particular choice of the process noise covariance matrix.
This connection motivates the development of a new class of adaptive Kalman filters, where the prior covariance update equation is modified to incorporate forgetting from RLS extensions.
A numerical example is provided, showing that an adaptive Kalman filter using the robust variable forgetting factor algorithm can improve state estimation for a mass-spring-damper system with intermittent, unmodeled collisions.
The paper concludes that such adaptive Kalman filtering may provide potential benefits for systems with non-classical disturbances.