Intelligent Data-Driven Diagnosis of Incipient Interturn Short Circuit Fault in Field Winding of Salient Pole Synchronous Generators

Abstract

This article examines if machine learning (ML) and signal processing can be used for online condition monitoring to reveal interturn short circuit (ITSC) fault in the field winding of salient pole synchronous generators (SPSG). This was done by creating several ML classifiers to detect ITSC faults. A dataset for ML was built using power spectral density of the air-gap magnetic field extracted by fast Fourier transform, discrete wavelet transform energies, and Time Series FeatuRe Extraction based on Scalable Hypothesis tests (TSFRESH) to extract features from measurements of SPSG operated under several different severities of ITSC fault. Using this dataset, a wide range of classifiers were trained to detect the presence of ITSC faults. The classifiers evaluated were logistic regression, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">K</i> -nearest neighbors, radial basis function support vector machine (SVM), linear SVM, XGBoost decision tree forest, multilayer perceptron, and a stacking ensemble classifier including all of the aforementioned. The classifiers were optimized using hyperparameter grid searches. In addition, some feature selection and reduction algorithms were assessed such as random forest feature selection, TSFRESH feature selection, and principal component analysis. This resulted in a classifier capable of detecting 84.5% of samples containing ITSC fault, with a 92.7% chance that fault detections are correct.