With the deep integration of power systems and information technology, the volume of power data is growing exponentially. However, there are issues with the large volume, chaotic content, and low credibility of knowledge in power equipment defect texts, and there is currently a lack of efficient methods for integrating and utilizing them. To address these issues, this paper analyzes the characteristics of power defect texts, introduces transfer learning into model design to improve data preprocessing efficiency, and solves the problem of difficult data acquisition. At the same time, a BERT-BAC-PCNN knowledge extraction model based on the field of power equipment defects is proposed to achieve entity extraction and relationship extraction of power equipment defect texts, thereby obtaining a triple dataset suitable for the construction of a power knowledge graph. Finally, a knowledge extraction experiment is conducted using the operation and maintenance records of a certain power company as an example, and the results show that the knowledge extraction model in this paper has a fast convergence speed, achieves good accuracy with fewer training times, and has strong generalization ability, which is of practical significance for improving the construction efficiency of the power knowledge graph.