Study on Deep-Learning Model for Phase Resolved Partial Discharge Pattern Classification Based on Convolutional Neural Network Algorithm

Partial discharge is one of the major causes that accelerates the deterioration of insulation in high-voltage electrical equipment, leading to insulation breakdown and causing significant damage to power systems such as power outages and fires. Partial discharge can occur both inside electrical equipment and on its surfaces, with various types. This paper proposes an artificial intelligence model capable of classifying patterns of various partial discharges. To analyze the designed model, pattern classification training data for each type of partial discharge, generated through UHF sensors, were collected. These data were transformed into 2D data using the Phase Resolved Partial Discharge. The proposed models were individually designed based on deep learning algorithms, namely VGG and ResNet. Additionally, Grad-CAM was used to visualize the learning areas of the pattern classification algorithms. Experimental result shows that each model can effectively improve the accuracy of partial discharge pattern classification. For the VGG model, the classification accuracies for DI, FE, and PE patterns were 99%, respectively. Regarding ResNet, the classification accuracy for the Noise pattern was 93%. Especially, Because Grad CAM provides class-discriminative and high-resolution visualization, it can effectively prove the weight of training data.