Machine Learning Based Fault Diagnosis for Multiphase Drives

Abstract

This thesis discusses the application of data-driven methods for fault diagnosis of multiphase drives, specifically the five-phase induction machine. Multiphase drives have numerous advantages over three-phase drive systems; one of them is the higher degrees of freedom or phase redundancy, making them more reliable as they require only three phases to produce a rotating magneto motive force. The main focus is on detecting and localizing the inverter side faults of the drive system which can be open-switch or open-phase faults. These diagnosis techniques are then used for derating the machine and reconfiguring the control for optimum operation.
The development of a machine learning-based algorithm for fault diagnosis in multiphase drives, utilizing data collected from MATLAB Simulink simulations and a five-phase Volts/Hz and field-oriented vector control machine, is discussed comprehensively. This study primarily focuses on open phase and open switch faults, with the data undergoing preprocessing, resampling, and the addition of Gaussian noise to simulate real-world conditions and improve generalizability to new data. Employing Random Forest, XGBoost, and Neural Networks, the models are initially trained on simulation data, followed by retraining with a minimal dataset of experimental data to enhance the accuracy on the training set and to reduce overfitting on the simulated data.
The performance of these models is benchmarked against existing literature, revealing that neural networks exhibit superior accuracy, particularly when using fixed sample points and Gaussian noise with both the training and experimental data.

Date of Award	16 May 2024
Original language	American English
Supervisor	Omar Alzaabi (Supervisor)