With the continuous expansion of power grid scale in recent years, a vast amount of unstructured equipment defect records has been collected and accumulated. This data contains entity information crucial for equipment condition assessment and operational decision-making. However, the prevalent nested entity structures in such data lead to increased entity boundary ambiguity and contextual semantic complexity, posing significant challenges to traditional named entity recognition methods. To achieve accurate identification of nested entities in defect records, this paper takes power transformers as a typical example and proposes a named entity recognition method for power transformer defect records that combines pre-training bidirectional encoder representations from transformers with permuted language model(PERT), bidirectional gated recurrent unit(BiGRU), and efficient global pointer(EGP). This method first employs the PERT model as a vector embedding layer for dynamic semantic encoding, leveraging its permuted pre-training characteristics to deeply capture contextual dependencies. Subsequently, a BiGRU network is introduced as the text encoding layer to comprehensively extract semantic features through its bidirectional gating mechanism. Finally, EGP is adopted as the decoding output layer to focus on entity spans and boundaries, enabling precise extraction of nested entities while avoiding the label conflict issues associated with traditional CRF decoding. Experimental results show that this entity recognition method effectively addresses the challenges of nested entities, achieving a comprehensive F1 score of 96.01%, which is 0.55% and 0.70% higher than those of the bidirectional encoder representations from transformers-biaffine attention(BERT-Biaffine) and bidirectional encoder representations from transformers-machine reading comprehension(BERT-MRC), respectively. It attained the highest F1 scores across all five entity label categories. Specifically, for defect equipment and defect component recognition, where nesting is most prominent, the F1 scores of the proposed method reached 100% and 94.74%, representing improvements of 0.57% and 0.13% over the best baseline models.