基于注意力机制的混合神经网络电力设备缺陷文本挖掘方法

Hybrid Neural Network Text Mining Method for Power Equipment Defects Based on Attention Mechanism

电网在运行过程中会产生大量的设备缺陷文本记录,针对变电设备缺陷文本的特点,文章提出了基于注意力机制的混合神经网络(hybrid neural network based on attention mechanism,HNNA)电力设备缺陷文本挖掘方法。首先在总结电力设备缺陷文本特点的基础上,参考中文文本分类的一般流程,结合自主编写的词典和停用词表对缺陷文本进行预处理;利用Word2vec模型将词语映射到高维空间;使用卷积神经网络(convolution neural network,CNN)和双向长短期记忆网络(bidirectional long short term memory,BiLSTM)提取文本局部特征和上下文特征;将提取的特征进行融合,最后采用Attention实现特征权重的分配,增强关键特征对分类效果的影响,并从多个评价维度与传统机器学习模型、深度学习模型对比。算例结果表明,提出的模型具有更好的分类效果,可以实现电力设备缺陷等级的高效准确划分。

A large number of equipment defect text records will be generated during the operation of power grid.According to the characteristics of substation equipment defect text,a hybrid neural network based on attention mechanism(HNNA)text mining method for power equipment defects is proposed.Firstly,on the basis of summarizing the characteristics of power equipment defect text,referring to the general process of chinese text classification,combined with the self-written dictionary and stop list,the defect text is preprocessed.Then,Word2vec model is used to map words to high-dimensional space.Convolution neural network(CNN)and bidirectional long short term memory(BiLSTM)are used to extract text local features and context features.The extracted features are fused.Finally,attention is used to realize the distribution of feature weight,enhance the impact of key features on classification effect,and compared with traditional machine learning model and deep learning model from multiple evaluation dimensions.The example results show that the proposed model has better classification effect and can realize the efficient and accurate classification of defect grade of power equipment.