Machine Learning for Real-Time Temperature Monitoring of Electric Machines

Abstract

This thesis investigates the application of machine learning techniques to the real-time thermal modelling of electric machines. Traditional methods like high-order lumped parameter models and finite element analysis, while effective, are computationally expensive and often unsuitable for real-time applications. This work seeks to address these limitations by developing data-driven machine learning-based models that are fast, easy to train, generalisable, and accurate.

The thesis begins with a comprehensive review of conventional thermal monitoring techniques and the state-of-the-art in machine learning methods for real-time modelling. The review highlights the limitations of physical temperature sensors and conventional finite element methods. Through the collection of detailed and extensive experimental data this thesis expands on the literature by increasing the available datasets for training thermal machine learning models, allowing new techniques and their generalisability to be analysed.

For the first time, this thesis analyses and compares an approach that integrates known physical principles with data-driven models to enhance model performance. Physics-informed feature engineering is explored to embed domain knowledge, demonstrating improved prediction accuracy and a reduction in required training data. When tested, the physics- informed features reduced modelling errors by up to 85.2%, resulting in the lowest prediction errors available in the literature. The proposed methodology allows simple machine learning models to meet real-world requirements, particularly in terms of speed and accuracy, making them suitable for embedded applications.

This thesis also extends data-driven machine learning models to predict future temperatures when given an expected duty cycle. The results demonstrate that these models, when combined with appropriate pre-processing and domain knowledge, can effectively estimate electric machine temperatures hours into the future without the prediction error drifting. A comparison of two future prediction models is presented. The first directly predicts temperature from the expected torque and speed duty cycle, whilst the second proposed model integrates a novel multistage pipeline. The proposed model reduces the prediction error by 1 to 2 orders of magnitude.

The thesis concludes by investigating transfer learning, exploring how trends learnt from one electric machine can be transferred to another. This significantly reduces the amount of data required for model training, enabling useful models from as little as 1 hour of training data. The combination of findings from this thesis will aid the widespread deployment of real-time thermal models for electric machines, improving machine utilisation and safety.

Date of Award	25 Jun 2025
Original language	English
Awarding Institution	University of Bath
Sponsors	AVL
Supervisor	Christopher Vagg (Supervisor), Xiaoze Pei (Supervisor), Thomas Haidinger (Supervisor) & Nic Zhang (Supervisor)