【摘要】 电力变压器是整个电力系统的核心设备，其故障诊断对于保障系统的安全稳定运行和可靠供电有重要意义。然而现有的变压器诊断方法在分析故障性质、故障元件以及给出后续检测建议等方面存在明显欠缺，其原因是现有基于神经网络的变压器诊断方法无法利用案例信息完成多项推理任务。针对这个问题，提出了一种基于知识图谱的变压器故障初始诊断和后续检测项目引导方法。首先，在C-MPNN(Conditional message passing neural network)的基础上，引入基于增强的福曼–里奇曲率AFRC(Augmented Forman Ricci curvature)，形成了AFRC-C-MPNN知识图谱新表示方法；借助AFRC-C-MPNN表示，得到了变压器案例诊断联合图谱的实体、关系的嵌入向量，进而实现了离散知识向结构化语义表示的转变。然后，将嵌入向量作为故障性质识别、故障元件识别以及后续检测项目三层感知器的输入，实现三项任务的分类。使用不同知识推理方法对800个变压器样本进行了对比实验，实验结果表明：所提出的故障诊断和后续检测项目引导（分类）方法在故障性质识别、故障元件识别准确率方面优于先前方法，其对后续检测项目建议引导是独到的。更多还原

【Abstract】 Power transformers are the core equipment of the entire power system, and the fault diagnosis is of great significance for ensuring the safe and stable operation of the system and reliable power supply. However, existing transformer diagnosis methods have significant shortcomings in analyzing the nature of faults, faulty components, and providing subsequent detection recommendations, which is attributed to the fact that the existing neural network-based transformer diagnosis methods cannot utilize case information to complete multiple inference tasks. In response to this issue, a knowledge graph-based method for fault initial diagnosis and subsequent detection items guidance of transformers is proposed. Firstly, based on the conditional message passing neural network(C-MPNN), an augmented Forman Ricci curvature(AFRC) is introduced to form a new representation method with AFRC-C-MPNN knowledge graph. With the help of AFRC-C-MPNN representation, the embedding vectors of entities and relationships in the transformer case diagnosis joint graph are obtained, thereby realizing the transformation from discrete knowledge to structur semantic representation. Then, the embedded vector is used as input for the three-layer perception of fault property recognition, fault component recognition, and subsequent detection items, so as to achieve classification of the three tasks. A comparative experiment is conducted on 800 transformer samples using different knowledge reasoning methods. The experimental results showed that the proposed method for fault diagnosis and subsequent detection item guidance(classification) have better accuracy in identifying fault properties and fault components than previous methods, and it is unique in its suggestion and guidance of subsequent detection items.更多还原