Online function approximation with streaming data for electric drives

Aaron Raffeiner

Publikation: AbschlussarbeitMasterarbeit


The accurate control of inverter-fed Permanent-Magnet Synchronous-Motors is essential for many industrial applications such as robotics and process automation. In order to improve to control performance, an accurate model of the system nonlinearities is required. This thesis investigates methods for online static function approximation from streaming data and evaluates them based on typical application scenarios form the electric drive domain. Kernel methods and Bayesian regression methods are investigated in detail since these models can be trained efficiently using convex optimization and allow to incorporate prior knowledge. Furthermore, the Bayesian framework provides a prediction uncertainty and systematic way for model selection. Simulation experiments are performed to evaluate the performance of the online approximators with artificially generated and measurement data. The robustness of the approximators against constant and heteroscedastic noise is tested. Additionally, the performance of the approximators is evaluated with a slowly time varying function. It is found that the Bayesian-Kernel-Recursive-Least-Squares (B-KRLS) is best suited for the considered scenarios since it obtained the most accurate approximations. Furthermore, the computational complexity of the B-KRLS can be defined a priori and it is able to approximate slowly time varying functions. Additionally, it is robust against heteroscedastic noise and it provides a predictive variance which can be used to evaluate the reliability of the prediction.
Gradverleihende Hochschule
  • TU Wien
Betreuer/-in / Berater/-in
  • Pock, Thomas, Betreuer:in, Externe Person
  • Kemmetmüller, Wolfgang, Betreuer:in, Externe Person
  • Gurtner, Markus, Betreuer:in
Datum der Bewilligung27 Apr. 2023
PublikationsstatusVeröffentlicht - 2023

Research Field

  • Complex Dynamical Systems


Untersuchen Sie die Forschungsthemen von „Online function approximation with streaming data for electric drives“. Zusammen bilden sie einen einzigartigen Fingerprint.

Diese Publikation zitieren