Large Language Models Empower Rail Transit Locomotive Maintenance: A Graph Topology Constrained Retrieval-Augmented Reasoning Framework

This article introduces a knowledge graph dataset for rail transit locomotive maintenance, which supports a graph topology constrained retrieval-augmented reasoning framework based on large language models, providing a new technical path for intelligent maintenance of industrial equipment. This project combines the advantages of knowledge graphs and large language models to solve the problem that traditional maintenance records are unstructured and difficult to utilize, and has important application value.

The operation of rail transit locomotives relies on regular maintenance, but traditional maintenance records are mostly unstructured text, which is difficult for computers to effectively understand and utilize. Knowledge graphs can store entities and relationships in a structured way and support complex query reasoning; large language models are good at extracting knowledge from unstructured text and generating natural language answers. The combination of the two provides new ideas for solving maintenance problems.

This open-source dataset includes three core components: 1. Maintenance record sample data: de-identified table samples containing fields such as original text, involved components, and fault types; 2. Knowledge graph relationship data: entities and relationships such as locomotive components and fault phenomena stored in JSON format, supporting complex queries; 3. Prompt template collection: prompt templates covering entity extraction, relationship extraction, entity linking optimization, answer generation, and other links.

The core technical framework is the "Graph Topology Constrained Retrieval-Augmented Reasoning Framework", with a three-stage process: 1. Knowledge graph construction: using large language models to extract entities and relationships from unstructured maintenance records, reducing construction costs while maintaining accuracy; 2. Retrieval augmentation: when users ask questions, using the graph's topology to retrieve relevant entities and relationship paths; 3. Reasoning and answer generation: inputting the retrieved graph information as context into the large language model to generate answers, alleviating the hallucination problem.

This dataset and framework can be applied to: 1. Intelligent maintenance Q&A: maintenance personnel obtain diagnostic suggestions and maintenance plans through natural language queries; 2. Knowledge discovery and mining: discovering fault correlation patterns and optimizing maintenance resource allocation through graph analysis; 3. Training and knowledge inheritance: storing maintenance knowledge in a structured manner to facilitate new employees' learning and help enterprises accumulate and preserve knowledge.

This project demonstrates a typical paradigm of combining knowledge graphs and large language models: knowledge graphs provide a structured factual foundation and interpretable reasoning paths, while large language models provide semantic understanding and generation capabilities. Their combination ensures the accuracy and interpretability of outputs. This technical route has reference significance for industrial knowledge management. In the future, multimodal large models and graph neural network technologies will promote more intelligent and efficient industrial knowledge graphs.