Performance Comparison of Various Classifiers in Detecting Incipient Fault Type Using DGA Data

Dissolved Gas Analysis (DGA) is a popular technique for incipient fault detection of power transformers. Machine learning (ML) algorithms are reported to be efficient in incipient fault detection using DGA data. Selection of input features and data balancing plays a crucial role in determining the fault classification accuracy of various ML classifiers. In the present work, Density-Based Clustering is first used to obtain a Euclidean distance feature using the gas concentration values of DGA data. The present work compares the performance of popular ML classifiers using input features constituted by gas concentration, gas ratios and the Euclidean distances of the fault points using DGA data. Various performance indicators are calculated to compare the performance of the tested ML classifiers.