Application of Pulse Sequence Partial Discharge Based Convolutional Neural Network in Pattern Recognition for Underground Cable Joints

Abstract

In this study, a pulse sequence analysis (PSA) method is presented, which does not measure the voltage signal for the defect-type recognition of power cable joints based on partial discharges (PDs) using convolutional neural networks (CNNs). The PSA showed effective results in comparison to a conventional phase-resolved PD (PRPD) pattern, which is a combination of a PD and voltage signals. The optimal drawing formats of the PSA patterns, including the size, type, color, and marker size of the images, which influence the recognition accuracy, were evaluated. In total, 12 sets of PD data, consisting of three types of cable joint defects, were trained and tested using CNN-based models. The results show that the performance of the PSA-based CNN model is as good as that of a PRPD even when the voltage signal is lost, whereas PRPD mostly fails under such cases. The total recognition accuracy obtained from a PSA-based CNN was 95.3%, whereas that of the PRPD-based CNN was 90.6%. In addition, an advanced and useful application of a PSA pattern is discussed herein.