Knowledge Graph Completion for Rail Transit Maintenance Based on Directed Subgraphs and Dynamic Information Flow

In order to deal with sparsity and incompleteness issues in the knowledge graph (KG) of urban rail transit operation and maintenance (O&M), this paper introduces a dynamic information flow based directed subgraph-based knowledge graph completion (KGC) method. Adding ontology constraints and semantic similarity calculations, the dynamic directed subgraph of new entities is constructed, enabling precise candidate entity and relation set selection, and successfully capturing contextually relevant domain information. Next, an embedding generation model with a dynamically updated information flow is constructed, integrating multi-layer message passing and self-attention mechanism to progressively obtain semantic features and structural dependencies from the subgraph and generate context-aware embeddings for entities and relations. Finally, the ConvE model acts as a decoder to learn higher-order entity and relation interactions in the triples and generate correct triple scores for efficient prediction and completion of missing relations. Experiments demonstrate that the proposed method outperforms state-of-the-art baselines on Mean Reciprocal Rank (MRR) and Hit Rates (Hits@1, Hits@3), confirming its effectiveness for KGC in urban rail transit O&M.