Transfer learning-based power transformer partial discharge diagnosis

Abstract

A diagnosis system is provided for diagnosing the partial discharge faults in a power transformer. The system includes a feature extractor, a fault classifier and a domain discriminator and whose parameters are optimized by minimizing a combination of feature loss, classifier loss and domain discrepancy discriminator loss. The feature loss is a weighted sum of cosine similarity distance loss, batch-based instance separation loss, and batch-based feature decorrelation loss. The measured applied voltages and partial discharge voltages are pre-processing using sliding window method to represent the each partial discharge event as a series of statistics based moments for a set of overlapped time blocks within the duration of fault event, wherein statistics based moments include an average value of applied voltage for representing the applied voltage magnitude and fault event occurring moment for each time block, and the mean, standard deviation, Kurtosis, and Skewness of partial discharge voltages for representing the variations of partial discharging within each time block.