Iterative Search-Based RAG: Using Iterative Retrieval for Efficient Large Language Model Enhancement in Industrial Fault Diagnosis

In industrial systems, diagnostic documents, such as fault logs recorded during routine operation and maintenance, contain a wealth of domain-specific knowledge. Leveraging these document collections, retrieval-augmented generation (RAG) methods can retrieve information relevant to specific queries, enabling general-domain large language models to assist in fault diagnosis within industrial systems. However, existing RAG methods typically assess the relevance of queries to individual documents in isolation, neglecting potential relationships between the retrieved documents, which could influence the relevance and accuracy of the results. To address this limitation, this paper proposes an iterative searched-based retrieval-augmented generation method. Specifically, the proposed method preprocesses the fault logs and vectorizes them using a pre-trained model. It then constructs a composite, associative dual-tower structure, which performs iterative retrieval across multiple cycles to refine fault diagnosis document retrieval. In each cycle, the relevance of a given document is evaluated by integrating information from both the query and previously retrieved documents. The subsequent experimental section, which utilizes fault logs from Type 1 and Type D heavy-duty trains collected over the course of one year, demonstrates the effectiveness of the proposed method through performance metrics including precision and mean position.