Graph-based Knowledge Acquisition with Convolutional Networks for Distribution Network Patrol Robots

With the popularization of smart grids, patrol robots have become critical devices in the distribution networks to check the state of equipment. In order to enrich the knowledge of patrol robots in this complex scenario, this paper presents the graph-based knowledge acquisition method with convolutional networks for distribution network patrol robots. The proposed method uses a graph convolutional network-based path-related embedding algorithm to complete the knowledge of the distribution network knowledge graph. The proposed algorithm generates the embeddings of entities and relations through aggregating the associated entities in the associated paths, instead of only the connected entities. The graph convolutional network consists of multiple graph convolution layers, and the message-passing process treats different entities discriminatorily according to the association strengths. For determining the plausibility of the knowledge, a scoring function is provided with the convolution operator. The experimental data sets are from a real grid company and contain four kinds of equipment. The experiments apply the proposed method to the equipment defects analysis, including the defect gradation, the coarse defect classification, and the fine defect classification. The proposed method is compared with some embedding methods. The experimental results verify that the proposed method outperforms the other methods for the real distribution network data sets, and the proposed method can analyze the defects effectively.