Long-term On-line Identification of Time-varying Systems

Abstract

This article discusses the possibility of deploying regression recursive continuous identification methods, which are using regularized exponential forgetting, during long-term monitoring of time-varying systems. The emphasis is placed on the long-term deployment since the advantage of this algorithm will only become evident with long-term deployment. These continuous identification algorithms are going to be applied on the mathematical model of the identified system for the needs of adaptive control. The article presents the implemented continuous identification techniques, where the non-informative data, that can possibly destabilize the numerical calculations of the identified system parameters, are weighted by the Dyadic reduction algorithm. These techniques are using an alternative covariance matrix for its “forgetting” in order to maintain the initial system dynamics in the mathematical model of the identified system. The matrix also suppresses the impact of high amounts of non-informative data received from the long-term operation the industrial systems with slow dynamics. The focus is mainly set on the non-informative data, which emerge especially during the long-term run of the tuned systems operating in between the bounds of the selected working point. A “Motor-Flywheel” laboratory model is used for the validation tests of the modified and the standard regression recursive algorithms.