Multi-label text mining method for electrical equipment defects based on improved CNN-SVM
Zhao Qi, Yang Lv, Tian Jiang, Hongen Ding, Jiachen Chen, Chun Li, Shan Gao · 2022 Power System and Green Energy Conference (PSGEC) · 2022
A large number of descriptive transmission and transformation equipment defect records have been accumulated in the work of power grid informatization, and the category imbalance and label symbiosis phenomenon directly affect the multi-label classification effect of the documents. There is a large amount of complex information in the existing power equipment defect record text, and the same text information corresponds to multiple data labels, so the data is highly correlated. It is necessary to comprehensively extract the text information features and to mine the correlation between the global tags from the graph structure data. To this end, this paper proposes a multi-label text classification model (CS-GAT) that fuses convolutional neural network-self-attention mechanism (CNN-SAM) and graph attention network (GAT). The scheme first considers the professional word classification method in the field of power system, and performs text preprocessing on the sentences in the defect record text. The word2vec model is used to map words to a high-dimensional feature space to form distributed word vectors; and meanwhile, the correlation between different text labels is transformed into an edge-weighted graph with global information, and the multi-layer graph attention mechanism is used to automatically learn the relationship between different labels. And then interact with the contextual semantic information of the text to obtain the global label information with the semantic relationship of the text. The example analysis shows that the multilabel classification scheme proposed in this paper has better classification effect, can achieve automatic accurate and efficient classification of power equipment defect texts, and can effectively improve the automatic classification performance of power texts.