基于长短期记忆网络-卷积神经网络的电力设备缺陷文本归口研究  

Study on Defect Text Attribution of Power Equipment Based on LSTM-CNN

作　　者：王璇 曹靖 韩培洁 WANG Xuan;CAO Jing;HAN Peijie(Technology Training Center of State Grid Shanxi Electric Power Company,Taiyuan,Shanxi 030020,China;State Grid Shanxi Electric Power Company,Taiyuan,Shanxi 030021,China)

机构地区：[1]国网山西省电力公司技能培训中心,山西太原030020 [2]国网山西省电力公司,山西太原030021

出　　处：《山西电力》2025年第1期10-14,共5页Shanxi Electric Power

摘　　要：随着电网企业发展,电网生产运营中会产生大量电力设备缺陷文本,其中蕴含着电力设备维护与检修的重要信息。由于缺陷文本是非结构化数据,其价值的挖掘依赖于归口,为提升文本利用效率,提出了一种基于长短期记忆网络-卷积神经网络的电力设备缺陷文本自动归口模型。以变压器缺陷文本为例开展研究,模型采用长短期记忆网络对词的权重进行学习,卷积神经网络对带权重的词进行特征提取,用softmax进行分类,最终得到文本归口。通过算例分析,证明该模型在准确度、召回率等方面均优于卷积神经网络等常规方法。With the development of power grid enterprises,a large number of power equipment defect texts will be generated in power grid production and operation,which contains important information about power equipment maintenance and overhaul.Since the defect text is unstructured data,its value mining depends on the attribution.In order to improve the utilization efficiency of the text,an automatic attribution model for the defect text of power equipment based on LSTM-CNN(Long Short Term Memory-convolutional neural network)is proposed.Taking transformer defect text as an example,the model adopts LSTM to learn the weight of words,CNN to extract the features of weighted words,and softmax to classify them,and finally the text attribution is obtained.Through comparison and analysis of examples,it is proved that the model is superior to CNN and other conventional methods in accuracy and recall rate.

关 键 词：电力设备缺陷文本 文本分类 长短期记忆网络 卷积神经网络 

分 类 号：TM507[电气工程—电器]