Power transformer defect texts contain information about equipment reliability

aiding intelligent operation

and lifecycle management. According to the bidirectional encoder representation from transformers (BERT) model

this study proposes a method that combines pre-training BERT with the permuted language model (PERT) and the efficient global pointer (EGP) for entity recognition in defect texts. The PERT model is initially developed by pre-training a vast Chinese corpus with a permuted language model. Then

PERT

as the semantic encoding layer

uncovers the text's internal semantic dependencies and complexities. EGP

as the information decoding layer

focuses on precisely locating key information and extracting the distribution features of entities within defect texts

accurately identifying defect entities. Finally

the PERT-EGP model identifies five entity types in defect texts: defective equipment

defective components

defective locations

defective phenomena

and defective severity. Experiments demonstrate that the proposed approach significantly improves handling complex composite entities and lengthy texts

significantly reducing model training time and providing enhanced text recognition performance.