Научная визуализация, 2021, том 13, номер 5, страницы 105 - 112, DOI: 10.26583/sv.13.5.09

Visual Based Tuning of Regularized Kalman Filter for System Identification Problem

Автор: M.D. Grebenkin¹

National Research Nuclear University MEPhI (Moscow Engineering Physics Institute)

¹ ORCID: 0000-0002-3644-5912, mdgrebenkin@mephi.ru

 

Аннотация

System identification problem for linear and non-linear systems utilizes a large set of algorithms for estimating a vector of model parameters, relying on measurements and system dynamics. In particular one can use a family of Kalman filter adaptive algorithms. In situation when system of interest is ill-conditioned it is proper to use regularized modification of Kalman filter. In comparison to standard algorithm, properly tuned RKF is significantly more stable to ill-conditioned problems, which frequently arise in the field of system identification due to limited observability or controllability of systems of interest.

This paper shows an approach for preliminary tuning and analyzing regularized Kalman filter algorithm (RKF) for parameter identification of a vector meter unit using visualization of its crucial values on a computer model. Visual based approach to RKF tuning on a computer model allows for simple and intuitive way to find suboptimal regularization strength and set it at initialization stage avoiding the necessity to include computationally expensive methods of real-time tuning in algorithm loop. It is shown that regularization strength value, found using this approach, yielded a better estimation accuracy not only in comparison with standard Kalman filter but in comparison with other possible regularization strength values as well.

 

Ключевые слова: regularized Kalman filter, ill-conditioned problem, Tikhonov regularization, system dynamics visualization.