Dissolved Gas Analysis for Power Transformer Fault Diagnosis Combining Domain Knowledge and Capsule Network

As the types and concentration of dissolved gases in oil are different when a power transformer operates in different conditions, dissolved gas analysis (DGA) has become a basic means for power transformer fault diagnosis. To further improve diagnosis accuracy and reliability, a new approach combining domain knowledge and capsule network is proposed in this paper. Firstly, new knowledge features based on the principle of ratio methods are constructed and used as the input of the capsule network, in which the knowledge can offer restrictions and guidance for training the capsule network. Then the capsule network based diagnosis model is built, which uses capsule (vector) instead of node (scalar) as the fundamental unit and is capable of extracting intricate and subtle patterns in the input data. Combining the advantages of knowledge-driven and data-driven approaches, the proposed approach achieves diagnosis accuracies of 92.11% and 94.02% on the real-world DGA data and IEC TC10 database, respectively, which attains a more than 10 percent increase compared with the typical purely data-driven or knowledge-driven diagnosis approaches. Meanwhile, the diagnostic accuracy reaches 83.98% even when utilizing only 10% of the original training set for the capsule network training and it increases steadily with the increasing proportion, which validates the strong stability and reliability of the proposed approach.