Dissolved Gas Analysis for Transformer Fault Diagnosis Based on Capsule Network

To solve the limitations of current dissolved gas analysis based transformer fault diagnosis methods and further improve the diagnostic accuracy, this paper develops a novel transformer fault diagnosis approach based on the Capsule Network (CapsNet). By combining the unique capability of CapsNet to handle vectors and the characteristics of dissolved gases in transformer oil, our approach is able to accurately model the intricate nonlinear relationship between dissolved gases and fault types, realizing the automatic extraction of key features. The dynamic routing technique and backpropagation algorithm are used for training the diagnostic model. Case studies on the real DGA data from a power grid validate that our approach outperforms both SVM and DBN across three performance metrics, demonstrating its advantages in real-world transformer fault diagnosis applications.