Fault Diagnosis Method for Transformer Windings Based on Residual Neural Network and State Code

Abstract

Currently, methods for power transformer winding fault diagnosis are still less intelligent. And there are few studies on its fault localization. This paper proposes a state code suitable for transformer intelligent fault diagnosis, which is combined with residual neural network for fault localization of transformer windings. First, obtain transformer winding frequency response data of different states based on pspice simulation. Next, the Pearson correlation coefficient of the FR data and the fingerprint is calculated using windowing calculation method to obtain the feature sequence dataset. Then, map the value of the one-dimensional feature sequence to the track of the space filling curve to obtain the two-dimensional state code dataset. Finally, the dataset is used for transfer training and verification of the residual neural network constructed based on the diagnostic scenario in this paper. Finally, the accuracy of the method proposed in this paper has an average increase of 5.43% over the traditional machine learning methods, reaching 94.44%.