Knowledge-Driven Recognition Methodology of Partial Discharge Patterns in GIS

Recognition of partial discharge (PD) patterns in gas insulated switchgear (GIS), as a basis of fault diagnosis, provides essential information for the condition assessment of GIS. However, traditional recognition methods are limited to handcrafted feature extraction or are extremely sensitive to the quantity and quality of datasets. Therefore, this article proposes a knowledge-driven algorithm, composed of the feature space and the knowledge space, to automatically extract PD features and improve the performance on noised, insufficient, and imbalanced datasets. First, the deep residual network (ResNet) in feature space extracts features from raw signals. Second, in knowledge space, the graph convolutional network (GCN) extracts additional information from the knowledge graph, which compensates for the missing information of original datasets. Finally, the algorithm recognizes patterns by ranking similarities between feature vectors and knowledge vectors. Verified by the comparison experiment, the proposed algorithm outperforms traditional methods with the accuracy of 99.58% on the experimental dataset and 95.67% on the online-detected dataset. Moreover, the accuracy of the proposed algorithm achieves 88.79% and 70.37% on noised and insufficient datasets, respectively, while the F-measure is higher than those of the comparison methods by 12.95% \sim 18.69% on imbalanced datasets.