A new intelligent fault diagnosis framework for CNC systems, based on LLMS and KG, has been proposed by researchers from Huazhong University of Science and Technology, Center for Strategic Studies of the Chinese Academy of Engineering and Tsinghua University. The study, published in Engineering, aims to address the challenges faced by typical expert systems based on symbolic reasoning, significantly improving the efficiency and accuracy of fault diagnosis in CNC systems.
As a core component of modern manufacturing, the fault diagnosis capability of CNC systems directly impacts production efficiency and product quality. traditional fault diagnosis systems mainly rely on expert systems, which face three major issues: (1) ineffective organization of fault diagnosis knowledge; (2) lack of adaptability between static knowledge frameworks and dynamic engineering environments, and (3) difficulties in integrating expert knowledge with real-time data streams. These issues limit the ability of typical methods to address uncertainty problems.
Although existing LLMs have demonstrated strong capability and performance in understanding and logical reasoning, content generation, human–computer interaction, and other areas, CNC systems require extremely high reliability, where any errors or hallucinated are not acceptable. In addition, the integration of domain-specific knowledge such as technical documentation, operational data, forward design principles, and expert experience remains a complex task. To address these challenges, the researchers explored key technologies for integrating LLMs with domain KG.
The researchers proposed a retrieval-augmented generation (RAG) method based on LLMs and domain KG, along with a dynamic learning mechanism based on LLMs and expert input. This framework not only resolved the inherent issues in traditional RAG frameworks but also better supported multi-turn dialogues by mapping these dialogues to traversals within the KG. The mechanism enables the system to continuously learn and optimize through user interactions in real-world usage, allowing it to update its KG and improve its diagnostic capabilities.
The researchers designed a domain-specific LLM benchmark to address the issue of model adaptation to the domain. Experimental results demonstrated the system's performance has surpassed the diagnostic capabilities of engineers with two years of experience. Currently，the system has been integrated into the CNC Cloud Manager APP by Huazhong CNC.
The study innovatively designed an intelligent fault diagnosis system for CNC systems based on LLMs and KG. It addresses the challenges of symbolic reasoning in typical expert systems and provides a standardized framework for the application of LLMs in the industrial sector.Future research will focus on further exploring prompt design, fine-tuning strategies, and even the pre-training process of LLMs to achieve broader industrial applications and higher system performance
The paper "Intelligent Fault Diagnosis for CNC Through the Integration of Large Language Models and Domain Knowledge Graphs," is authored by Yuhan Liu, Yuan Zhou, Yufei Liu, Zhen Xu, Yixin He. Full text of the open access paper: https://doi.org/10.1016/j.eng.2025.04.003
