Machine learning implementation in electrical machine modelling and fault diagnosis

Machine-learning (ML) particularly deep neural networks (NN) have been developed rapidly over the past decade, enabling powerful new tools for modelling, design, prediction, and fault diagnosis in electrical machines. This thesis explores how ML can enhance the modelling and fault diagnosis of permanent-magnet synchronous machines (PMSMs).
Firstly, this thesis presents a high-fidelity PMSM model that combines multiple lightweight NN blocks with established physical relationships. The framework captures magnetic saturation, spatial harmonics, temperature effects and iron losses without relying on large look-up tables (LUT) or simplifying assumptions that limit operating range. Compared with conventional finite-element or LUT-based approaches, the proposed model achieves comparable accuracy while reducing memory requirement and computational load.
The second part of the work focuses on open-circuit (OC) fault diagnosis in PMSM drives. Traditional model-based or signal-analysis methods are either parameter-sensitive or computationally intensive. This thesis introduces ‘normalised space-vector current sorting’ to convert three-phase currents into distinctive sequence-like patterns, which are then classified by a compact one-dimensional convolutional neural network (1D-CNN). Extensive offline tests, robustness analyses and real-time experiments confirm high diagnostic accuracy across a wide operating envelope.
Given that labelled fault data are scarce and simulation data suffer from domain shift, two simulation data-based strategies are investigated. (i) Data augmentation + multiscale 1D-CNN: prior knowledge based synthetic perturbations broaden the training set, and receptive-field scaling helps the network learn features that generalise to unseen real-world signals. (ii) Metric-learning with TripletNet, a neural-network framework that learns discriminative feature embeddings by comparing triplets of samples (anchor, positive, and negative): this structured embedding space improves separability even with few labels. Both approaches are validated in target experiment dataset.
Finally, this thesis also investigates the bearing impedance modelling to aid the prediction of high-frequency bearing currents in voltage source inverter-fed machines. A probabilistic ML framework is developed to model bearing impedance and its transition from predominantly capacitive to resistive behaviour. The proposed bearing impedance model is a conner stone for an accurate bearing current prediction, ultimately improving the bearing lifetime of electrical machines.