Current research in mechanical fault inference diagnosis primarily relies on knowledge graph networks (KGN) for reasoning. Compared to traditional diagnostic methods, KGN offers advantages in knowledge representation, but it suffers from insufficient causal mining and limited reasoning capability. To this end, a fault inference diagnosis method (CN-LLM-KGN) that integrates causal networks (CN), large language models (LLMs), and KGN is proposed. Firstly, the causal structure and key causal paths of fault propagation are mined from the monitoring data through CN. Domain knowledge constraints are introduced to correct and optimize the traditional Peter-Clark algorithm. Secondly, the LLMs are utilized to enhance the KGN, including entity/relation completion of the knowledge graph, semantic alignment, and optimization of reasoning rules, to address the problems of knowledge sparsity and reasoning rigidity in traditional knowledge graphs. Finally, the constructed causal inference module, via an attention-weighted fusion strategy, drives the LLM to achieve causal traceability inference, knowledge-based fault diagnosis, and precise root cause localization (RCL). The experiment showed that the RCL accuracy of the CN-LLM-KGN method reached 81.12%, providing a new path for the intelligent diagnosis of rotating machinery systems. This study proposes a novel intelligent diagnostic architecture that possesses strong causal reasoning capabilities, enabling effective fault traceability and precise localization, thereby forming an efficient and reliable hybrid intelligent diagnostic system.
Guo et al. (Sat,) studied this question.