Research on Text Information Recognition and Mining Methods for Fault Records of Traction Power Supply Equipment

Like Pan¹Tong Xing^1*Haibo Zhang¹Yingxin Zhao¹Yuan Yuan¹Wenrui Dai¹ZhanHao Dong²

• ¹Standards & Metrology Research Institute, China Academy of Railway Sciences Corporation Limited, Beijing, China
• ²Southwest Jiaotong University, Chengdu, China

Traction power supply equipment is critical to the safe operation of electrified railways, yet the vast 12 volume of unstructured fault records accumulated in maintenance systems remains underutilized due 13 to the lack of automated text mining tools. To address this gap, this study proposes the first integrated 14 framework that combines deep learning-based named entity recognition with knowledge graph 15 reasoning for intelligent fault handling decision support.The core innovations include: A novel BERT-16 BiLSTM-CRF architecture that achieves state-of-the-art accuracy (94.66% F1-score) in extracting nine 17 key entities (e.g., fault type, cause, handling measures) from complex TPSE fault texts, significantly 18 outperforming rule-based and traditional machine learning methods; A domain-specific knowledge 19 graph constructed using Neo4j, which dynamically integrates extracted entities and their relationships 20 to represent Traction power supply equipment fault-handling expertise; An entity-similarity-driven 21 retrieval algorithm (leveraging Word2Vec embeddings and cosine similarity) that recommends 22 relevant historical cases beyond exact matches, enabling context-aware decision support. Validated on 23 912 real-world fault records, our framework reduces manual processing time by 98.5% while providing 24 maintenance personnel with actionable insights. Experimental results confirm its superiority in 25 precision and efficiency over baselines, demonstrating a scalable solution for transforming 26 unstructured railway data into operational intelligence.