As large language models (LLMs) continue to demonstrate their potential in handling complex tasks, their value in knowledge-intensive industrial scenarios is becoming increasingly evident. Fault diagnosis, a critical domain in the industrial sector, has long faced the dual challenges of managing vast amounts of experiential knowledge and improving human–machine collaboration efficiency. Traditional fault diagnosis systems, which are primarily based on expert systems, suffer from three major limitations: 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 systemic shortcomings restrict the ability of conventional approaches to handle uncertainty. In this study, we proposed an intelligent computer numerical control (CNC) fault diagnosis system, integrating LLMs with knowledge graph (KG). First, we constructed a comprehensive KG that consolidated multi-source data for structured representation. Second, we designed a retrieval-augmented generation (RAG) framework leveraging the KG to support multi-turn interactive fault diagnosis while incorporating real-time engineering data into the decision-making process. Finally, we introduced a learning mechanism to facilitate dynamic knowledge updates. The experimental results demonstrated that our system significantly improved fault diagnosis accuracy, outperforming engineers with two years of professional experience on our constructed benchmark datasets. By integrating LLMs and KG, our framework surpassed the limitations of traditional expert systems rooted in symbolic reasoning, offering a novel approach to addressing the cognitive paradox of unstructured knowledge modeling and dynamic environment adaptation in industrial settings.