A manufacturing knowledge graph completion method based on a lightweight dual encoding model

Ensuring stable equipment operation is crucial for manufacturing. Intelligent maintenance decisions powered by manufacturing knowledge graphs can reduce reliance on manual maintenance and enhance efficiency. However, existing knowledge graphs face challenges such as sparse information and complex relationship modeling. Knowledge graph completion can predict missing relationships and entities to enrich the graph. Current completion methods neglect semantic information in entity descriptions, leading to incomplete data, while encoding triples and descriptions increases computational costs. Therefore, this paper proposes a Lightweight Dual Encoding Model (LDEM) for manufacturing knowledge graph completion. LDEM uses ALBERT to encode entity descriptions and captures rich semantics through precomputed embeddings. The graph attention module aggregates neighborhood information, and ConvKB decodes embeddings into predictions. The dataset used in this study comes from a vehicle welding workshop in Chongqing, China. Experiments show that LDEM outperforms state-of-the-art models in all metrics, achieving 80.1 points in Hits@10 and demonstrating superior ability to capture entity relationships and semantic information, thereby enhancing the completion of the manufacturing knowledge graph.