BERT-Based Dual-Channel Power Equipment Defect Text Assessment Model

Abstract

Accumulating a substantial amount of textual data on power equipment defects during maintenance and inspection stages presents a valuable problem of assessing and grading these text-based information. This paper proposes a dual-channel text feature extraction model based on the pre-trained BERT model, applied to the evaluation of power equipment defect levels in textual data. Firstly, a dataset of power equipment defect levels is established, followed by data augmentation and preprocessing. Then, a neural network model is constructed, utilizing the pre-trained BERT model for initial semantic information extraction from the text, further extracting features through two modules, Bi-LSTM and CNN, on top of BERT’s output. Finally, the obtained feature vectors are concatenated to generate the output. Comparative experiments with other algorithms demonstrate that the proposed method out-performs others in multiple metrics, achieving an F1 score of 96%. The research findings can serve as a reference for achieving intelligent processing of power textual information.

Keywords

• Power defect text
• natural language processing
• deep learning
• BERT
• dual channel model