Classification of Fault and Stray Gassing in Transformer by Using Duval Pentagon and Machine Learning Algorithms

An oil-filled transformer should be able to operate for a long time with proper maintenance. One of the best diagnostic methods for oil-immersed transformer condition is dissolved gas analysis (DGA). However, there are times where the produce of stray gassing event might lead to fault indication in the transformer. Machine learning algorithms are used to classify the DGA data into normal condition and corresponding faults based on IEEE limits and Duval pentagon method. The algorithms that will be used include boosted trees, RUS boosted trees and subspace KNN, which belongs to the same ensemble group. Data resampling technique (SMOTETomek) is applied and shows further improvement on the accuracy of predictions by machine learning algorithms when deal with imbalance data. The algorithms are able to achieve the accuracy of 82.6% (boosted trees), 81.2% (RUS boosted trees) and 72.5% (subspace KNN), respectively, when validated with actual transformer condition.