Titelaufnahme
Titelaufnahme
- TitelUncertainty quantification for data-driven motor drive temperature estimation models / by Shubham Gupta ; first examiner: Prof. Dr.-Ing. Jakub Kucka, second examiner: Prof. Dr.-Ing. Erdal Kayacan
- Autor
- Gutachter
- Erschienen
- Umfang1 Online-Ressource (viii, 98 Seiten) : Diagramme
- HochschulschriftUniversität Paderborn, Masterarbeit, 2026
- AnmerkungTag der Abgabe: 17.05.2026
- Datum der Abgabe17.5.2026
- SpracheEnglisch
- DokumenttypMasterarbeit
- Schlagwörter (GND)
- URN
- DOI
Links
- Social MediaShare
- Nachweis
- IIIF
Dateien
Klassifikation
Abstract
Permanent magnet synchronous motors dominate many industries due to their superior power and torque density. However, temperature sensitivity remains a key vulnerability, as high temperatures can damage components like permanent magnets, stator windings, etc. Monitoring the temperature inside these motors with physical sensors comes with its own set of challenges, such as difficulty in measuring temperatures in a spinning rotor, sensor drift, etc. In addition to that, multiple sensors are needed to obtain a complete thermal profile, which poses another challenge in cost-prohibitive industries, such as automotive. To solve these problems, alternative methods of temperature estimation have been a focus of research. Data-driven temperature estimation models that integrate physical knowledge, such as thermal neural networks (TNNs), have shown great potential here. But their black-box nature and deterministic point predictions make them challenging to deploy in safety-critical applications. This work intends to address this challenge by quantifying the uncertainty in data-driven temperature estimation models. It evaluates Gaussian processes, dropouts, deep ensembles, evidential deep learning (EDL) optimized with negative log-likelihood (NLL), and EDL optimized with the Continuous Ranked Probability Score (CRPS) on TNNs. To maintain the baseline TNN's gray-box physical interpretability, novel architectural solutions are proposed. Furthermore, this work demonstrates that CRPS is a viable loss function for EDL, an approach that has not previously been explored in the literature. The performance of these uncertainty quantification (UQ) methods has been rigorously compared using a wide range of metrics, including CRPS, NLL, reliability diagrams, miscalibration area, inference time, and model storage requirements. Based on the analysis, Gaussian processes emerged as the winner, with the best overall uncertainty calibration performance. Dropout achieved the best accuracy on the MSE metric. EDL-based methods showed good performance, with negligible computational overhead. This work also demonstrated the unsuitability of deep ensembles for UQ in TNNs.
Inhalt
Statistik
- Das PDF-Dokument wurde 7 mal heruntergeladen.
Lizenz-/Rechtehinweis
