Exploring bridge maintenance knowledge graph by leveraging GrapshSAGE and text encoding

Knowledge graphs (KGs) are crucial in documenting bridge maintenance expertise. However, existing KG schemas lack integration of bridge design and practical inspection insights. Meanwhile, traditional methods for node feature initialization, relying on meticulous manual encoding or word embeddings, are inadequate for real-world maintenance textual data. To address these challenges, this paper introduces a bridge maintenance-oriented KG (BMKG) schema and approaches for graph data mining, including node-layer classification and link prediction. These methods leverage large language model (LLM)-based text encoding combined with GraphSAGE, demonstrating excellent performance in semantic enrichment and KG completion on deficient BMKGs. Additionally, ablation studies reveal the superiority of the pre-trained BERT text encoder and the L2 distance pairwise scoring calculator. Furthermore, a practical implementation framework integrating these approaches is developed for routine bridge maintenance, which can facilitate various practical applications, such as maintenance planning, and has the potential to enhance the efficiency of engineers' documentation work.