电网设备缺陷部位识别是设备故障分析的关键环节。该文提出一种基于预训练语言模型双向Transformers偏码表示(Bidirectional encoder representation from transformers,BERT)的电网设备缺陷文本分类方法。基于BERT预训练语言模型对电...电网设备缺陷部位识别是设备故障分析的关键环节。该文提出一种基于预训练语言模型双向Transformers偏码表示(Bidirectional encoder representation from transformers,BERT)的电网设备缺陷文本分类方法。基于BERT预训练语言模型对电网设备缺陷部位文本进行预训练生成具有上下文特征的词嵌入(Word embedding)向量作为模型输入,然后,利用双向长短时记忆(Bi-directional long short-term memory)网络对输入的电网设备缺陷文本向量进行双向编码提取表征缺陷文本的语义表征,并通过注意力机制增强电网设备缺陷文本中与缺陷部位相关的领域词汇的语义特征权重,进而得到有助于电网设备缺陷部位分类的语义特征向量。通过模型的归一化层实现电网设备缺陷部位文本分类。在主变压器、SF6真空断路器这两种设备缺陷文本数据集上实验结果表明,提出的方法比基于BiLSTM-Attention模型的F1值分别提升了2.77%和2.95%。展开更多
This is an extended version of the same titled paper presented at the 21st CIRED. It discusses a new technique for identification and location of defective insulator strings in power lines based on the analysis of hig...This is an extended version of the same titled paper presented at the 21st CIRED. It discusses a new technique for identification and location of defective insulator strings in power lines based on the analysis of high frequency signals generated by corona effect. Damaged insulator strings may lead to loss of insulation and hence to the corona effect, in other words, to partial discharges. These partial discharges can be detected by a system composed of a capacitive coupling device （region between the phase and the metal body of a current transformer）, a data acquisition board and a computer. Analyzing the waveform of these partial discharges through a neural network based software, it is possible to identify and locate the defective insulator string. This paper discusses how this software analysis works and why its technique is suitable for this application. Hence the results of key tests performed along the development are discussed, pointing out the main factors that affect their performance.展开更多
文摘电网设备缺陷部位识别是设备故障分析的关键环节。该文提出一种基于预训练语言模型双向Transformers偏码表示(Bidirectional encoder representation from transformers,BERT)的电网设备缺陷文本分类方法。基于BERT预训练语言模型对电网设备缺陷部位文本进行预训练生成具有上下文特征的词嵌入(Word embedding)向量作为模型输入,然后,利用双向长短时记忆(Bi-directional long short-term memory)网络对输入的电网设备缺陷文本向量进行双向编码提取表征缺陷文本的语义表征,并通过注意力机制增强电网设备缺陷文本中与缺陷部位相关的领域词汇的语义特征权重,进而得到有助于电网设备缺陷部位分类的语义特征向量。通过模型的归一化层实现电网设备缺陷部位文本分类。在主变压器、SF6真空断路器这两种设备缺陷文本数据集上实验结果表明,提出的方法比基于BiLSTM-Attention模型的F1值分别提升了2.77%和2.95%。
文摘This is an extended version of the same titled paper presented at the 21st CIRED. It discusses a new technique for identification and location of defective insulator strings in power lines based on the analysis of high frequency signals generated by corona effect. Damaged insulator strings may lead to loss of insulation and hence to the corona effect, in other words, to partial discharges. These partial discharges can be detected by a system composed of a capacitive coupling device （region between the phase and the metal body of a current transformer）, a data acquisition board and a computer. Analyzing the waveform of these partial discharges through a neural network based software, it is possible to identify and locate the defective insulator string. This paper discusses how this software analysis works and why its technique is suitable for this application. Hence the results of key tests performed along the development are discussed, pointing out the main factors that affect their performance.
