Cognitive Power Electronics for Smart Drives in Unmanned Aerial Vehicles

Conference: PCIM Europe 2022 - International Exhibition and Conference for Power Electronics, Intelligent Motion, Renewable Energy and Energy Management
05/10/2022 - 05/12/2022 at Nürnberg, Germany

doi:10.30420/565822007

Proceedings: PCIM Europe 2022

Pages: 8Language: englishTyp: PDF

Authors:
Huf, Tobias; Roeder, Georg; Schellenberger, Martin; Lorentz, Vincent R. H. (Fraunhofer Institute for Integrated Systems and Device Technology IISB, Germany)
Steinmetz, Harm-Friedrich (mdGroup Germany GmbH, Germany)

Abstract:
For Unmanned Aerial Vehicles (UAV) with electric propulsion, the motors and propellers are key for safe operation. The propulsion system is exposed to varying loads and harsh ambient conditions, typically reducing the service interval of the whole UAV. A regular estimation of bearing condition enables a more predictable maintenance and more cost-efficient maintenance planning and efforts. In this paper, we propose an interpretable manifold learning approach towards the development of smart drives for UAVs. The approach enables visualization of the bearing condition based on the analysis of motor phase cur-rents of the drone motors as a preceding result used in the following anomaly detection. We applied a sequence of machine learning algorithms to study and compare several healthy and damaged drone motors. The bearing condition was first examined by comparing specific peaks of the motor phase current in the frequency domain. A lower resolution of the spectra is chosen to simplify the visual analysis and enables the interpretation of the type of bearing wear. For use in machine learning, spectra with higher resolution are compressed with a kernel principal component analysis and allow a later optical separability. We show that the visual inspection of the spectra enables human interpretation, whereas highly resolved spectra will improve the separability of motor states in the machine learning process. The feasibility of using anomaly detection with a support-vector machine in a real application is discussed as a further step.