Knowledge Graph Completion for Industrial Robots Based on Graph Attention Network

The current state of the knowledge graph for industrial robots is confronted with several significant challenges. One issue is the insufficient fault data of industrial robots, which results in an inadequate number of nodes and edges in the knowledge graph and a pervasive problem of data sparsity. Additionally, industrial robots encompass a wide array of diverse and complex fault types, leading to intricate correlations and interactions among different fault types. In response to these challenges, this study proposes a method called BAKGAT for complementing the knowledge graph of industrial robot faults. BAKGAT leverages the semantic understanding and feature extraction capabilities of BERT, in conjunction with the multi-hop neighbor node information aggregation proficiency of KBGAT, thereby enhancing node and relation representation and capturing complex semantic relationship information more effectively. Furthermore, BAKGAT introduces an attention attenuation mechanism to mitigate the impact of distant nodes and amplify the attention weight of neighboring nodes, thereby facilitating the more precise capture of local structural information. To validate the efficacy of the proposed method, a series of comparative experiments was performed on two benchmark datasets, WN18RR and FB15k-237, as well as the industrial robot fault dataset IRF9K. The experimental results demonstrated that BAKGAT surpasses the performance of the baseline models, highlighting its superior effectiveness in addressing sparse and heterogeneous data.