Towards Automated Frequency Response Analysis of Power Transformers with Deep Learning

Frequency response analysis (FRA) has emerged as one of the standard techniques for monitoring the integrity of the mechanical structure of power transformer windings. Interpreting FRA test results, though, is still largely dependent on expert identification of graphical features. Machine learning, however, presents an opportunity to automate and improve this feature identification process. In this study, FRA measurements were simulated and then image and series data representations were used to train three neural networks. The Xception network, trained with image magnitude data, obtained the best performance, with an F1 score of 98.6%. The ResNet and Fully Connected Neural Network, trained with series magnitude data, obtained F1 scores of 94.6% and 91.4%, respectively. Results revealed that networks trained using image-encoded FRA data outperformed those trained using series FRA data.