k-nearest neighbor algorithm based classification and localization of seven different types of disc-to-disc impulse insulation failures in power transformer

Identification and localization of seven different types of disc-to-disc insulation failure in power transformer using K - Nearest Neighbor classifier (kNN) has been proposed. Proper identification of insulation failure in power transformer is vital to take appropriate corrective measure. In the present work, seven types of series insulation failures are simulated in EMTP based digital model of 33 kV winding of 3 MVA power transformer. Resultant winding currents of each insulation failures are acquired following the tank current method for the time span of 0-500 μs. By correlating healthy and faulty winding currents of 3 MVA transformer, significant time-frequency domain features are extracted by employing cross-wavelet transform. Using extracted features, the k - Nearest Neighbor classifier has successfully identified and localized all seven different types of disc-to-disc insulation failures within ± 9% winding length with acceptable accuracy. Simulation of insulation failures, feature extraction and fault identification are explained.