Article Details

Abstract

Aiming at the problems of knowledge fragmentation, low retrieval efficiency, serious hallucination, and lack of traceability in large language model question answering in the field of Railway Four Electrical Systems, a knowledge-enhanced retrieval system based on Retrieval-Augmented Generation (RAG) architecture is designed and implemented based on the Palm Electrification platform of China Railway Electrification Bureau Group. The system constructs a high-quality dataset through multi-source data collection and nine-level cleaning rules, and completes document structuring via MinerU. A dual-route collaborative framework of RAG + LoRA-based domain fine-tuning is proposed, which balances answer traceability and professional understanding capability. Engineering verification shows that the system achieves a Top-5 recall accuracy of 91% and a professional term recognition rate of 82.7%, with a single-query response time ≤ 2.6 s. In the trial operation of 18 subsidiaries, the system responds to 48 000 queries in total, reduces knowledge acquisition time from 8～12 minutes to 2.6 s, and reduces troubleshooting time from 20 minutes to 7 minutes. The system can effectively support the construction, operation, and maintenance of Railway Four Electrical Systems projects.

Keywords

railway four electrical systems; RAG; large model LoRA fine-tuning; knowledge retrieval

References

1. [1] China Railway Electrification Bureau Group Co., Ltd Technical Manual for Railway Four Electric Engineering (Communication Volume) [M]. Beijing: China Railway Press, 2022
2. [2] Touvron H，Lavril T，Izacard G，et al. LLaMA: Open and Efficient Foundation Language Models [PP/OL].（2023-02-27）.https://doi.org/10.48550/arXiv.2302.13971.
3. [3] Hong S，Kim S，Park J，et al. Haystack: An End-to-End Framework for Building Search Systems [J].Journal of Machine Learning Research，2022，23（283）：1-8.
4. [4] Zuo Jiming Research on Knowledge Graph Modeling Method Based on Railway Fault Text Data [D]. Beijing: Beijing Jiaotong University, 2022
5. [5] Guo Xiaoxiao, Li Xueshan, Li Junqi, etc Design and Implementation of Knowledge Preservation and Sharing Communication System for Railway Industry Institutions [J]. Railway Computer Applications, 2024, 33 (2): 38-42
6. [6] California Department of Transportation. Engineering Document QA System Based on LangChain [R].Sacramento：Caltrans，2024.
7. [7] Zhang Y，Li M，Wang H. A Survey on RAG Meeting LLMs: Towards Retrieval-Augmented Large Language Models [PP/OL].（2024-05-10）.https://doi.org/10.48550/arXiv.2405.06211.
8. [8] Hong Yingjie, Gao Yan, He Qing Construction method of railway multimodal knowledge base question answering system based on hybrid RAG [J]. Railway Computer Application, 2025, 34 (11): 24-31
9. [9] NetEase Youdao BCEmbedding project team BCE-RerankerBase-v1 Technical Manual and RAG Integration Guide [R]. Beijing: NetEase Youdao Information Technology (Beijing) Co., Ltd., 2025
10. [10] State Administration for Market Regulation, National Standardization Administration Artificial Intelligence Large Model Part 2: Evaluation Indicators and Methods: GB/T 45288.2-2025 [S]. Beijing: China Standard Press, 2025
11. [11] National Railway Administration Design Specification for Railway Communication: TB 10006-1999 [S]. Beijing: China Railway Press, 2016
12. [12] Wang Yujie Design and Implementation of a Literature Report Generation System for Railway Transportation [D]. Chengdu: University of Electronic Science and Technology of China, 2024
13. [13] OpenDataLab team MinerU 2.5 Document Analysis Model Technical Report [R]. Shanghai: Shanghai Artificial Intelligence Laboratory, 2025
14. [14] Alibaba Damo Institute. gte-qwen2-1.5b-instruction Vector Model Technology Report [R]. Hangzhou: Alibaba Group, 2024

Funding

Major Scientific Research Project of China Railway Electrification Bureau (2024 Major -18)

About the authors

Luo Jin (July 1988-), male, Mongolian, Beijing native, senior engineer, bachelor's degree, research direction: four electrical engineering, computer application technology.