This thesis addresses the development of a thermal digital twin of a TEFC squirrel-cage induction motor through a hybrid approach that combines electromagnetic and thermal finite element simulations with machine learning techniques. Finite element analyses are used to estimate losses and temperature distribution within the motor; to this end, 2D models are employed in order to reduce computational cost. A reduced-order model, built from these simulations, enables the reproduction of the system’s dynamic behaviour with a computational cost compatible with real-time applications. Due to the simplifications introduced in the simulations and the approximation of the reduced-order model, a discrepancy arises with respect to experimental data. For this reason, a regression model based on Gaussian Process is implemented to learn the difference between simulated and measured temperatures. The machine learning model is then used to correct the outputs of the reduced-order model, improving prediction accuracy and providing an associated uncertainty quantification. The results show that the proposed approach significantly reduces the prediction error compared to the standalone reduced-order model, while maintaining low computational cost.
Questo lavoro di tesi affronta lo sviluppo di un digital twin termico di un motore a induzione a gabbia di scoiattolo di tipo TEFC attraverso un approccio ibrido che combina simulazioni agli elementi finiti elettromagnetiche e termiche con tecniche di machine learning. Le analisi agli elementi finiti sono utilizzate per la stima delle perdite e della distribuzione di temperatura nel motore; a tal fine vengono impiegati modelli 2D, così da ridurre il costo computazionale. Un modello a ordine ridotto, costruito a partire da tali simulazioni, consente di riprodurre il comportamento dinamico del sistema con un costo computazionale compatibile con applicazioni in tempo reale. A causa delle semplificazioni introdotte nelle simulazioni e dell’approssimazione del modello ridotto, emerge una discrepanza rispetto ai dati sperimentali. Per questo motivo, viene implementato un modello di regressione basato su Gaussian Process, utilizzato per apprendere la differenza tra le temperature simulate e quelle misurate. Il modello di machine learning viene quindi utilizzato per correggere le uscite del modello a ordine ridotto, migliorando l’accuratezza delle previsioni e fornendo una quantificazione dell’incertezza associata. I risultati mostrano che l’approccio proposto consente di ridurre significativamente l’errore di previsione rispetto al solo modello a ordine ridotto, mantenendo al contempo un basso costo computazionale.
Thermal behavior prediction in totally enclosed fan-cooled induction motors using a hybrid digital twin
GHIOTTO, NICOLÒ
2025/2026
Abstract
This thesis addresses the development of a thermal digital twin of a TEFC squirrel-cage induction motor through a hybrid approach that combines electromagnetic and thermal finite element simulations with machine learning techniques. Finite element analyses are used to estimate losses and temperature distribution within the motor; to this end, 2D models are employed in order to reduce computational cost. A reduced-order model, built from these simulations, enables the reproduction of the system’s dynamic behaviour with a computational cost compatible with real-time applications. Due to the simplifications introduced in the simulations and the approximation of the reduced-order model, a discrepancy arises with respect to experimental data. For this reason, a regression model based on Gaussian Process is implemented to learn the difference between simulated and measured temperatures. The machine learning model is then used to correct the outputs of the reduced-order model, improving prediction accuracy and providing an associated uncertainty quantification. The results show that the proposed approach significantly reduces the prediction error compared to the standalone reduced-order model, while maintaining low computational cost.| File | Dimensione | Formato | |
|---|---|---|---|
|
Ghiotto_Nicolò.pdf
Accesso riservato
Dimensione
9.44 MB
Formato
Adobe PDF
|
9.44 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/110169