Performance Evaluation of AI-Based Partial Discharge Classifier using Correlation between PRPD Pattern Distance and Identification Accuracy

This study examines the identification performance of a convolutional neural network (CNN)-based partial discharge (PD) source classifier when applied to data obtained under measurement conditions not included in the training data. The results obtained under the present experimental conditions indicate that classification accuracy tends to decrease as the distance between the average PRPD patterns of the training and target data increases. Among the evaluated distance metrics, the Bray-Curtis distance showed relatively good correspondence with the decrease in classification accuracy. These results suggest that the distance between PRPD patterns can be used as an empirical indicator for evaluating the applicability of an AI-based PD classifier to previously unseen measurement conditions.