Defect text recognition and condition assessment method of power equipment based on BERT-BiLSTM

The accumulated text of power equipment defects in operation are crucial for condition evaluation and maintenance. However, different structures of defect texts and their sensitivity to the experience of operation and maintenance personnel have led to difficulties in information mining, resulting in low efficiency of defect management. To address this issue, a defect text recognition method for power equipment is proposed based on Bidirectional Encoder Representations from Transformers-Bidirectional Long Short-Term Memory Neural Network (BERT-BiLSTM). First, the pre-training model of BERT is used to provide word vectors, which are fine-tuned by the BERT model based on the context of the defect text. Combined with the global semantic information, the vectors are inputted into the BiLSTM model. Then BiLSTM is used to encode and fuse features of the sequences to extract deep semantic features of defects. The obtained categories include defect severity, component, and type identification results. Based on algorithm comparison and case verification, the accuracy of the method is greater than 94% in identifying defect severity and component, proving the accurate recognition of defect text of power equipment. Finally, a defect condition assessment method is proposed to improve health management and condition maintenance.