Symmetry Index Analysis for Inter-turn Short Circuit Fault Detection in Electrical Machines

Electric motors are a pivotal part of the ongoing shift in the transportation industry towards electrification. Nonetheless, electric motors are subject to faults. Among all faults, interturn short-circuit can be the most damaging and drastically shorten the motor’s life. It is admissible that if a motor is healthy, the distribution of the magnetic field will be symmetric around the motor. Therefore, it will be demonstrated that by capturing the magnetic signature from the end winding of the motor and processing this data, the symmetry index can be used as proof of the health of an electric machine. This thesis uses the concept of symmetry index analysis to present a fault diagnosis method to study the effects of faults specifically inter-turn short-circuits on two different types of electrical machines with two different winding arrangements. First, the study was conducted on an induction machine (IM) with a distributed winding, where the winding is customdesigned in such a way that inter-turn short circuits of 1%, 5%, and 10% can be manually applied to phase A winding. The second motor under this study is a switched reluctance motor (SRM) with concentrated winding, where 6% (1 turn out of 16 turns in phase A) and 43% (7-turn inter-turn short circuit of 16 turn in phase A) can be applied to the motor’s phase A winding. In order to collect the data, two integrated sensor boards are designed and installed at the proximity of the end winding, where the magnetic data can be captured and processed. It will be visually demonstrated that when the motor is healthy, the magnetic data will be distributed in a linearly symmetric manner. Finally, A series of machine learning (ML) methods will be applied and compared to this data to classify them. These machine learning methods are Decision trees (DT), Support Vector Machine (SVM), Gradient Boosting (GB), Random Forest (RF), and Logistic Regression (LR). Two key factors to compare the methods which are the accuracy and the execution time that the data will fit to the machine learning model are compared between these methods and then are reported using bar diagrams.