Machine learning techniques for robust classification of partial discharges in oil–paper insulation systems

Ageing power systems infrastructure and concerns about climate change have increased interest in the next generation of grid infrastructure, known as the smart grid (SG). This study studies a particularly critical SG application: intelligent monitoring of power transformers for the early detection of insulation failure. Specifically, the focus is on the use of machine learning algorithms to distinguish between different types of partial discharges, which are closely correlated with insulation failure. Measurements made using acoustic emission sensors are used to train and test different classification algorithms. In an earlier study, high classification accuracies were achieved using training and test datasets collected under similar measurement conditions. However, under different conditions, classification accuracy was greatly reduced. Experiments using the latest classification techniques were performed, producing significant improvements in classification accuracy. A possible reason for these results could be a form of overfitting, and further experiments were conducted to test this hypothesis.