In view of the problems of poor field applicability of emergency plans, heavy reliance on human experience and low efficiency of knowledge index in the complex and changeable operation environment of railway system, an intelligent generation method of railway emergency disposal strategy based on knowledge retrieval enhancement is proposed. Firstly, the multi-layer classification architecture model based on semantic understanding is used to identify the scene characteristics such as type, location and severity of emergencies, which can be used as the retrieval index of emergency disposal knowledge base. Secondly, a hybrid retrieval strategy combining dense vector retrieval and keyword matching is used to accurately locate the relevant information of a specific scene. Finally, a dual-role prompt mechanism of "system global constraints + user scenario adaptation" is established to drive the generation of standardized disposal strategies that conform to the habits of staff. The results show that the fidelity and correlation index of the disposal strategy generated by the proposed method are 0.880 and 0.966, which are better than the benchmark method. In the complex event scenario, the correlation index of this method is 0.901, which is slightly lower than that of the single event scenario, maintaining good stability. The average time of generating emergency disposal strategy is 9.610 s, which is significantly lower than that of the Graph RAG benchmark method. This method can effectively meet the accuracy, universality and real-time requirements of emergency disposal, and provide reliable technical support for improving the emergency decision-making ability in the context of railway complex emergencies.
近年来,大语言模型(Large Language Model,LLM)在自然语言处理领域取得了突破性进展。凭借其超大规模的参数设置以及海量的训练数据,这些模型展现出卓越的自然语言理解与生成能力[8],GPT-o1和Deepseek等模型也呈现出较好的逻辑推理能力[9],推动了大模型在各行业的垂直应用[10-14]。然而,铁路行业由于其专业性、复杂性与安全性等特点,包含铁路应急预案、行车技术规范、历史事件处置案例在内的行业专业知识并未纳入通用大语言模型的训练中,导致既有大模型在领域专业任务上表现不佳。尽管如此,大语言模型卓越的上下文理解与动态生成能力,为应急处置策略智能生成领域提供了新思路:通过语义解析[15-17]和知识向量表征与转化技术[18],从预案与案例中检索到专业知识作为大模型的决策依据,该方法能够克服传统方法依赖固化规则或完备案例库的弊端,根据当前事态进行灵活推理,从而在应急处置策略生成方面展现出巨大潜力。
JIALimin, CHENXiyuan, MAXiaoping, et al. Architecture of Autonomous Disaster Prevention System of High-Speed Railway Based on Collaborative Cloud and Edge Mechanism [J]. China Railway Science, 2022, 43 (5): 165-176. in Chinese
[3]
程晓卿,贾利民,秦勇,等.铁路应急管理研究[J].铁道学报,2012,34(3):7-13.
[4]
CHENGXiaoqing, JIALimin, QINYong, et al. Research on Railway Emergency Management [J]. Journal of the China Railway Society, 2012, 34 (3): 7-13. in Chinese
JINGHaiquan, ZHONGRendong, HEXuhui, et al. Multi-Level Early Warning Method for Gale Based on LSTM-GMM Model [J]. China Railway Science, 2023, 44 (3): 221-228. in Chinese
[7]
NIW J, LIUT, CHENS Y. Rapid Generation of Emergency Response Plans for Unconventional Emergencies [J]. IEEE Access, 2020, 8: 181036-181048.
WANGPu, LIPing, ARUNA, et al. Method of Digitalization of High-Speed Railway Emergency Plan Integrating Ontology and Deep Learning [J]. Journal of the China Railway Society, 2020, 42 (8): 29-36. in Chinese
[10]
WUH T, ZHONGB T, MEDJDOUBB, et al. An Ontological Metro Accident Case Retrieval Using CBR and NLP [J]. Applied Sciences, 2020, 10 (15): 5298.
[11]
SUNJ P, CAOH T, GENGB, et al. Demand Prediction of Railway Emergency Resources Based on Case-Based Reasoning [J]. Journal of Advanced Transportation, 2021, 2021: 6666631.
[12]
WEIJ, TAY Y, BOMMASANIR, et al. Emergent Abilities of Large Language Models [EB/OL]. arXiv:2206.07682 [cs.CL], 2022 (2022-10-26) [2025-08-13].
[13]
SHAOZ H, WANGP Y, ZHUQ H, et al. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models [EB/OL]. arXiv:2402.03300 [cs.CL], 2024 (2024-05-27) [2025-08-13].
[14]
ZHANGK P, ZHOUF, WUL, et al. Semantic Understanding and Prompt Engineering for Large-Scale Traffic Data Imputation [J]. Information Fusion, 2024, 102: 102038.
QISiyang, HUHuiyun, LIHongbing, et al. Domain-Specific Question Answering System Construction Approach Integrated with Large Language Model [J]. Journal of Beijing University of Posts and Telecommunications, 2024, 47 (4): 50-56. in Chinese
HUJiahui, LIJiao, YAOKuanda, et al. Research on the Construction of Medical Q&A System Integrating Large Language Model and Knowledge Graph [J]. China Digital Medicine, 2024, 19 (6): 91-95. in Chinese
LIMinzhe, YINJibin. TCM Named Entity Recognition Model Combining BERT Model and Lexical Enhancement [J]. Computer Science, 2024, 51 (6A): 230900030-6. in Chinese
CHENWenjia, YANGLin, LIJinlin. Semantic Classification Model for Healthcare Q&A Texts with Embedded Intent Recognition [J]. Data Analysis and Knowledge Discovery, 2025, 9 (2): 26-38. in Chinese
[23]
MACKENZIEJ, TROTMANA, LINJ. Efficient Document-at-a-Time and Score-at-a-Time Query Evaluation for Learned Sparse Representations [J]. ACM Transactions on Information Systems, 2023, 41 (4): 1-28.
WEIWei, DINGXiangxiang, GUOMengxing, et al. Review of Text Similarity Calculation Methods [J]. Computer Engineering, 2024, 50 (9): 18-32. in Chinese
[26]
CUIY M, CHEW X, LIUT, et al. Pre-Training with Whole Word Masking for Chinese BERT [J]. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2021, 29: 3504-3514.
[27]
CHENJ L, XIAOS T, ZHANGP T, et al. BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings through Self-Knowledge Distillation [EB/OL]. arXiv:2402.03216 [cs.CL], 2024 (2024-06-28) [2025-08-13].
[28]
ZHANGY, WALLACEB. A Sensitivity Analysis of (and Practitioners’ Guide to) Convolutional Neural Networks for Sentence Classification [EB/OL]. arXiv:1510.03820 [cs.CL], 2015 (2016-05-06) [2025-08-13].
[29]
LEWISP, PEREZE, PIKTUSA, et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks [C]// NIPS’20: Proceedings of the 34th International Conference on Neural Information Processing Systems. USA: ACM Digital Library, 2020, 793: 9459-9474.
[30]
EDGED, TRINHH, CHENGN, et al. From Local to Global: a Graph RAG Approach to Query-Focused Summarization [EB/OL]. arXiv:2404.16130 [cs.CL], 2024 (2025-02-19) [2025-08-13].
[31]
LIANGL, BOZ P, GUIZ K, et al. KAG: Boosting LLMs in Professional Domains via Knowledge Augmented Generation [C]// WWW’25: Companion Proceedings of the ACM on Web Conference 2025. USA: ACM Digital Library, 2025: 334-343.