A Partial Discharge Diagnosis Method for GIS Based on a Semi-Supervised Classification Framework and Density Peak Clustering Algorithm

The existing methods for diagnosing partial discharge (PD) insulation defects in gas-insulated switchgear (GIS) can only achieve a high identification accuracy (ACC) if sufficient labeled data are available. However, the GIS equipment was not disassembled to determine the type of insulation defect after PD occurred, resulting in a small number of PD samples and mostly unlabeled data. Semi-supervised learning can still have excellent performance with limited labeled data, but its performance is limited by the number of initial labeled data. Therefore, this article proposes a semi-supervised classification algorithm based on density peak clustering (DPC) and density decay graph [self-training classification algorithm based on DPC and density decay graph (STDP2DG)]. First, a parameter-free DPC method [DPC to construct density decay graph (DP2DG)] is proposed to construct the density decay graph and the real structure space. Then, the density decay graph constructed by DP2DG is used to generate boundary samples of sparse regions in the self-training iteration process to help train the classifier better. Finally, the final strong classifier is output for the classification of PD patterns and their severity in GIS. The experimental results show that STDP2DG has better recognition ACC than other semi-supervised algorithms in the case of limited labeled data. When the proportion of labeled data is only 5%, the accuracy still reaches 56.5%.