Abstract
The faults associated with lightning impulse testing of a power transformer are often destructive to the transformer internal structure which includes the windings and insulation structure. Diagnosing and locating the fault can also be destructive as the insulation might be damaged in trying to get to the source of the fault, therefore an exact method for identifying and locating the fault is needed in ensuring minimal damage is done to unaffected transformer parts. A method for using the resultant lightning impulse waveforms from a fault is proposed which uses machine learning algorithms to extract fault features from the waveform and classify the faults according to the waveform parameters.
The proposed machine learning algorithm integrates two primary algorithms: the Discrete Wavelet Transform (DWT) for extracting fault features from lightning impulse waveforms and a Support Vector Machine (SVM) for classifying the extracted features. Key waveform parameters such as rise time, tail time, and peak voltage are used to differentiate between fault types. The Daubechies wavelet is selected for feature extraction, and the SVM classifier with a Radial Basis Function (RBF) kernel is chosen for classification.
The proposed model demonstrated that lightning impulse faults could be detected and classified using the resultant waveforms, providing a reliable means for fault detection without the need for expert intervention.