Automatic construction of asset knowledge graph with large language model

Lifecycle engineering is a critical concept for fostering environmentally sustainable practices within the manufacturing sector. An essential component of lifecycle management for achieving sustainability is reliability-centered maintenance, which enhances various key performance indicators (KPIs), including machine availability and environmental impact. Effective and reliable maintenance necessitates expert knowledge of the equipment. For instance, determining which components and failure modes should be addressed through condition-based maintenance requires insights derived from failure mode and effect analysis (FMEA). However, constructing expert knowledge is labor-intensive, and ensuring its quality presents significant challenges. This study proposes a method for the automated construction of expert knowledge related to maintenance, along with a corresponding tool designed to reduce construction costs and enhance knowledge quality. The proposed method leverages a large language model (LLM) to automatically generate asset knowledge graphs based on FMEA. By combining general knowledge about equipment derived from the pre-trained LLM with specialized information extracted from technical documents, the tool creates knowledge structures such as component trees and failure modes. Subject matter experts can then iteratively refine and validate this knowledge. To evaluate the proposed approach, we assessed the accuracy and coverage of the knowledge generated by the tool in two case studies involving specific types of equipment. The results indicated that the LLM-generated output contained 4.98 times more items than those manually created, with precision ranging from 0.490 to 0.662 and recall ranging from 0.481 to 0.810.