铁路内燃机车故障维修知识图谱与决策推荐

Knowledge Graph and Decision Recommendation for Diesel Locomotive Fault Maintenance in Railway Systems

摘要: 针对铁路内燃机车维修领域数据分散、信息孤岛等痛点，文章提出了一种基于知识图谱构建与推理的智能辅助决策方法。首先，采用自顶向下策略，结合现场维修案例，通过自定义词典和正则规则抽取设备型号、故障部位；针对故障原因、现象和维修措施等语义复杂要素，引入BERT-CRF模型进行抽取，模型在构建的铁路故障语料库上实现了90.18%的Precision、90.91%的Recall与89.92%的F1值。将生成三元组数据由Neo4j图数据库存储。最后，基于TransE、TransH、TransR和TransD四种嵌入算法开展链接预测对比，TransE在MRR (0.880)、Hits@10 (0.944)等指标上表现最佳，选其作为推理核心。该方法能够有效整合维修数据，提升维修决策的准确性与效率。

Abstract: In response to the challenges of fragmented data and information silos in the maintenance domain of railway diesel locomotives, this paper proposes an intelligent decision-support approach based on knowledge graph construction and reasoning. First, a top-down strategy is employed, integrating on-site maintenance case studies to extract equipment models and fault locations using a custom dictionary and regular expression rules. For semantically complex elements such as fault causes, fault phenomena, and maintenance measures, a BERT-CRF model is introduced for sequence labeling, achieving a precision of 90.18%, a recall of 90.91%, and an F1 score of 89.92% on the constructed railway fault corpus. The resulting triplet data are stored in the Neo4j graph database. Finally, a comparative link prediction analysis is conducted using four embedding algorithms—TransE, TransH, TransR, and TransD—with TransE demonstrating superior performance in metrics such as MRR (0.880) and Hits@10 (0.944), and is thus chosen as the core inference model. This approach effectively integrates dispersed maintenance data, enhancing the accuracy and efficiency of decision-making and showing promising potential for practical application.