Genetic programming feature extraction with bootstrap for dissolved gas analysis of power transformers

This paper discusses a feature extraction technique with genetic programming (GP) and bootstrap to improve interpretation accuracy of dissolved gas analysis (DGA) fault classification in power transformers, dealing with highly versatile or noise corrupted data. Initial DGA data are preprocessed with bootstrap to equalize the sample numbers for different fault classes, thus improving subsequent extraction of classification features with GP for each fault class. The features extracted with GP are then used as the inputs to artificial neural network (ANN), support vector machine (SVM) and K-nearest neighbor (KNN) classifiers for fault classification. The test results indicate that the proposed preprocessing approach can significantly improve the accuracy of power transformer fault classification based on DGA data.