Abstract:Dissolved gas analysis (DGA) is a typical fault diagnosis method for oil filled power equipment, which is widely used for power transformer fault detection and state evaluation. However, due to the reliability of sample data and the effectiveness of diagnosis model, the accuracy of DGA diagnosis method is low. In this paper, a Box-plot-SA-BP model is proposed. First, the Box-plot data detection method is used to remove abnormal data to solve the problem of data quality. Then, the Self-attention mechanism (SA) is used to accurately capture the relationship between multi-parameter samples and extract more stable and reliable features. Finally, a BP network-based multi-classification model is designed to achieve transformer fault diagnosis result. The comparative experimental results prove that the Box-plot-SA-BP model has good performance and high application value.