Figure 2
The illustration shows a system workflow diagram for fault diagnosis using a knowledge graph, document retrieval, re-ranking, and large language model generation. The diagram is arranged horizontally, showing the process from data sources to final diagnostic output. At the top left, two rectangular input sources inside a block are shown: “Fault Cases” and “Technical Manuals”, followed by ellipses indicating additional document sources. These sources feed into a process labeled “Knowledge Graph Construction”, represented by an arrow pointing toward a circular node network graphic representing the knowledge graph. From the knowledge graph, data is stored and accessed through “Milvus”, a vector database system shown on the top right with its logo. Dashed arrows connect the knowledge graph and Milvus to the document retrieval stage below. The main workflow begins at the bottom left with a user query, represented by a person icon and the label “Query”. An arrow leads to the “Query Understanding” stage. The “Query Understanding” stage contains a dashed rectangular box with two processes: “Intent Recognition” and “Keyword Extraction”. From this stage, the workflow moves to “Multi-route Retrieval”, represented by another dashed box containing two retrieval approaches: “Keyword-based Retrieval” and “Vector-based Retrieval”. The system then generates “Top-K Candidate Documents”, shown in a dashed box listing example outputs: “Candidate Document 1”, “Candidate Document 2”, ellipsis, and “Candidate Document K”. Next, the candidate documents are processed through a “Re-rank” stage labeled “Caps G C N-Rank”. In this step, the candidate documents are reordered based on relevance. The box shows a new ranking example: “Candidate Document 3”, “Candidate Document K”, ellipsis, and “Candidate Document 6”. Finally, the re-ranked results are sent to the “Large Language Model Generation” stage, represented by a dashed box on the far right. This stage produces the final diagnostic output, including “Failure Mode Root Cause” and “Maintenance Plan”.KG-CapsGCN-RAG framework