摘要
为提升铁路信号设备故障分析的智能化水平,提出一种针对铁路信号设备故障分析任务构建的领域大语言模型微调方法。采用低秩微调(Low-Rank Adaptation,LoRA)技术实现参数高效微调,降低训练成本;融入拓扑思维链(CoT)构建故障因果推理框架,增强模型思维逻辑可解释性;外挂故障知识库并通过检索增强机制,提升术语识别与知识适配性。实验表明,该方法使故障板卡信息提取准确率较基线模型提升11.2%,推理时间减少52%;在推理时间相当的情况下,该方法准确率较传统微调方法提升2.5%;跨域迁移场景下模型准确率提升42.6%,展现强泛化能力。通过技术融合与知识增强,该方法有效解决传统方法的效率与泛化瓶颈,显著提升故障分析的准确率、效率及领域适应性,为铁路信号智能运维提供技术支撑,推动人工智能技术在轨道交通场景的工程化落地,具有突出的应用价值与创新突破。
To enhance the intelligent analysis of railway signaling equipment failures,this study proposes a domain-specific large language model fine-tuning approach established for the failure diagnosis task of railway signaling equipment.The proposed method employs Low-Rank Adaptation(LoRA)technology to achieve parameter-efficient fine-tuning,significantly reducing training costs.The topological Chain-of-Thought(CoT)technology is integrated to establish the causal reasoning framework for fault diagnosis,improving the model's logical interpretability.By incorporating an external fault knowledge base with retrieval-augmented mechanisms,the proposed system demonstrates enhanced terminology recognition and knowledge adaptation capabilities.Experimental results indicate that this approach achieves an 11.2%accuracy improvement in faulty board information extraction compared to baseline models,while reducing inference time by 52%.When the reasoning time is comparable,the accuracy of this method is 2.5%higher compared with that of the traditional fine-tuning method.In cross-domain migration scenarios,the proposed model exhibits robust generalization capabilities with a 42.6%accuracy enhancement.Through technical integration and knowledge enhancement,this method effectively addresses the efficiency and generalization bottlenecks inherent in conventional approaches,substantially improving diagnostic accuracy,operational efficiency,and domain adaptability.The proposed system provides critical technical support for intelligent railway signaling maintenance,accelerating the practical implementation of AI technologies in rail transportation scenarios.This research represents significant innovation with prominent application value.
作者
孙超
李涵蕊
丁子焕
Sun Chao;Li Hanrui;Ding Zihuan(CRSC Research&Design Institute Group Co.,Ltd.,Beijing 100070,China;Engineering Research Center of Railway Industry of Intelligent and Autonomous Train Control,Beijing 100070,China)
出处
《铁路通信信号工程技术》
2025年第7期18-26,87,共10页
Railway Signalling & Communication Engineering
基金
中国国家铁路集团有限公司科技研究开发计划重点课题项目(N2023G081)。
关键词
人工智能
大语言模型微调
思维链
知识检索增强
铁路信号
artificial intelligence
large language model fine-tuning
chain-of-thought
retrieval-augmented knowledge enhancement
railway signaling
作者简介
第一作者:孙超(1982-),男,高级工程师,硕士,主要研究方向:轨道交通装备RAMS设计、列车运行控制及系统健康管理(PHM),邮箱:sunchao@crscd.com.cn。
