Underground cable systems play an important role in distribution networks. The insulation failure of underground cables, especially in cable joints, is mostly attributed to defects resulting from imperfect manufacturing or improper installation practices. Partial discharge (PD) is a well- known diagnostic indicator for detecting flaws in cable joints, which is further expected to provide advice on the maintenance of cable accessories. The experimental procedure was performed in a laboratory on 25 kV cross-linked polyethylene cable joints with 3 different types of artificial defects to simulate field-poor installation. In this paper, a convolutional neural network (CNN) is introduced to recognize PD sources by using 3 different types of phase-resolved partial discharge (PRPD) input patterns. The key factors that influence CNN-based pattern recognition accuracy are discussed, including the number of network layers, convolutional kernel size, and activation function. The results show that different types of PRPD input images affect recognition accuracy. In this study, the adopted 2-D PRPD input map (n-Φ-q) presents an average recognition accuracy of 92%, which is 5% superior to the types of q-Φ-t.
展开▼
