Instrumental Variables based DREM for Online Asymptotic Identification of Perturbed Linear Systems
Anton Glushchenko, Konstantin Lastochkin
TL;DR
The paper tackles online asymptotic identification of linear systems subject to unknown bounded disturbances that may correlate with the regressor. It introduces an Instrumental-Variables based extension to Dynamic Regressor Extension and Mixing (IV-based DREM) with averaging, producing scalar regression equations and a gradient estimator that achieve unbiased, asymptotic convergence under weaker excitation and independence conditions than traditional methods. The authors prove conservative and relaxed convergence theorems, extend the approach to closed-loop identification, and validate the method with numerical simulations showing exact asymptotic convergence and improved transients. This work broadens the applicability of online identification in disturbance-rich, practical environments by relaxing stringent identifiability prerequisites.
Abstract
Existing online continuous-time parameter estimation laws provide exact (asymptotic/exponential or finite/fixed time) identification of dynamical linear/nonlinear systems parameters only if the external perturbations are equaled to zero or independent with the regressor of the system. However, in real systems the disturbances are almost always non-vanishing and dependent with the regressor. In the presence of perturbations with such properties the above-mentioned identification approaches ensure only boundedness of a parameter estimation error. The main goal of this study is to close this gap and develop a novel online continuous-time parameter estimator, which guarantees exact asymptotic identification of unknown parameters of linear systems in the presence of unknown but bounded perturbations and has relaxed convergence conditions. To achieve the aforementioned goal, it is proposed to augment the deeply investigated Dynamic Regressor Extension and Mixing (DREM) procedure with the novel Instrumental Variables (IV) based extension scheme with averaging. Such an approach allows one to obtain a set of scalar regression equations with asymptotically vanishing perturbation if the initial disturbance that affects the plant is bounded and independent not with the system regressor, but with the instrumental variable. It is rigorously proved that a gradient estimation law designed on the basis of such scalar regressions ensures online unbiased asymptotic identification of the parameters of the perturbed linear systems if some weak independence and excitation assumptions are met. Theoretical results are illustrated and supported with adequate numerical simulations.