Construction and Application Exploration of Railway Signal Intelligent Operation and Maintenance Large Model
-
摘要:
针对铁路信号领域设备种类繁多、数据分散的现状,探索利用人工智能大模型技术实现传统信号运营维护(简称:运维)智能化升级的路径。分析铁路信号运维在数据治理、故障诊断等方面的核心问题,选择DeepSeek-R1作为基础模型,通过接口扩展实现多源异构数据的标准化治理,构建统一的数据处理流程;采用分层学习机制与混合微调策略,结合增量学习、小样本学习等技术,提升基础模型对动态数据的适应性及罕见故障的诊断能力;设计设备智能诊断、智能问答助手和预防性维护等3个核心应用场景,推动大模型在铁路信号运维中的实际落地,降低人工运维成本。研究成果为铁路信号系统智能化升级提供了可行的技术方案与实践参考。
Abstract:In response to the diverse types of equipment and scattered data in the field of railway signaling, this paper explored the path of using artificial intelligence large model technology to implement intelligent upgrading of traditional signal operation and maintenance. The paper analyzed the core issues of railway signal operation and maintenance in data governance, fault diagnosis, and other aspects, chose DeepSeek-R1 as the basic model to implement standardized governance of multi-source heterogeneous data through interface extension, and built a unified data processing flow, adopted a hierarchical learning mechanism and a hybrid fine-tuning strategy, improved the adaptability of the basic model to dynamic data and the diagnostic ability for rare faults by combining techniques such as incremental learning and small sample learning. It designed three core application scenarios, including intelligent diagnosis of equipment, intelligent question and answer assistant, and preventive maintenance to promote the practical implementation of large models in railway signal operation and maintenance, and reduce manual operation and maintenance costs. The research results provide feasible technical solutions and practical references for the intelligent upgrade of railway signal system.
-
-
表 1 主流模型性能对比
模型特性 BERT GPT DeepSeek ChatGLM 架构类型 双向编码器 单向解码器 编码器-解码器 编码器-解码器 参数规模 110 M~330 M 1.5 B~175 B 7 B~671 B 6 B~130 B 训练数据类型 通用文本 通用文本 代码+文本 多语言文本 领域适配能力 强(双向上下文建模) 中(单向生成) 强(代码逻辑推理) 中(中文优化) 多模态支持 弱(需扩展) 弱(需扩展) 中(代码−文本对齐) 中(插件机制) 计算效率 高(编码器架构) 中(自回归生成) 高(量化支持) 高(量化优化) 工业迁移成本 低(预训练+微调) 中(指令微调) 低(代码领域优势) 中(中文生态) -
[1] Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[C]//The 31st International Conference on Neural Information Processing Systems, 4-9 December, 2017, Long Beach, USA. New York: ACM, 2017: 6000-6010.
[2] Rajapaksha P, Farahbakhsh R, Crespi N. BERT, XLNet or RoBERTa: The best transfer learning model to detect clickbaits[J]. IEEE Access, 2021(9): 154704-154716. DOI: 10.1109/ACCESS.2021.3128742
[3] 谈骏杰,许璟琳,欧金武,等. 基于大模型的建筑智慧运维技术探索与应用[C]//第十届全国BIM学术会议论文集, 北京:中国建筑工业出版社,2024:404-408. [4] 许璟琳,彭 阳,欧金武,等. 融合大模型和数字孪生的公共建筑智慧运维系统[J]. 图学学报,2024,45(6):1200-1206. [5] 吴精乙,景 峻,贺熠凡,等. 基于多模态大模型的高速公路场景交通异常事件分析方法[J]. 图学学报,2024,45(6):1266-1276. [6] 张云秋,殷 策.基于大模型的中文电子病历实体自动识别研究[J/OL].数据分析与知识发现,1-18[2025-03-10]. http://kns.cnki.net/kcms/detail/10.1478.g2.20241118.1735.004.html. [7] International Union of Railways. High-speed around the world[M]. Paris: UIC, 2023: 8-30.
[8] 陈建译. 电务大数据智能运维平台研究与应用[J]. 铁道通信信号,2019,55(S1):162-166. [9] 李凌健. 面向铁路智能运维的大模型技术应用研究[J]. 铁道技术标准(中英文),2025,7(1):70-75. [10] 滕 蕾,张卫军,屈 毅. 铁路通信智能运维方案及演进策略研究[J]. 铁道通信信号,2022,58(11):8-11. [11] Li L F, Guo L. Dynamic low-rank adaptation based pruning algorithm for large language models[C]//2024 7th International Conference on Pattern Recognition and Artificial Intelligence (PRAI), 15-17 August, 2024, Hangzhou, China. New York, USA: IEEE, 2024: 1094-1099.
[12] 姜栋瀚,林海涛. 云计算环境下的资源分配关键技术研究综述[J]. 中国电子科学研究院学报,2018,13(3):308-314. DOI: 10.3969/j.issn.1673-5692.2018.03.013 [13] 李 刚,徐长明,龚 翔,等. 基于掩码自编码器的小样本深度学习道岔故障诊断模型[J]. 中国铁道科学,2022,43(6):175-185. DOI: 10.3969/j.issn.1001-4632.2022.06.18 [14] Fleischer S, Hooke M. Aleatoric and epistemic uncertainty quantification in Bayesian dirichlet cost rules of thumb[C]//2023 IEEE Aerospace Conference, 4-11 March, 2023, Big Sky, USA. New York, USA: IEEE, 2023: 1-8.
-
期刊类型引用(0)
其他类型引用(2)