Transformer short-circuit fault diagnosis using convolutional neural networks

Transformer short-circuits can result in critical safety concerns as well as considerable economic impact. In this study, the electrical characteristics under transformer short-circuit conditions are investigated, and a diagnostic framework based on feature selection and convolutional neural networks (CNN) is proposed. Firstly, under fault conditions, electrical features—including disturbances, sequence components, differential currents, and unbalance—are investigated. Secondly, to address the issue that feature prioritization cannot be easily distinguished and feature selection often relies on empirical experience, a ReliefF-based feature selection method combined with the maximum information coefficient is proposed. ReliefF calculates the weight of each feature and performs an initial ranking. The maximum information coefficient identifies redundancy among features, and a secondary ranking is conducted based on the initial ranking. Thirdly, a three-level diagnostic framework for transformer short-circuits is established. An improved CNN model is employed to extract and classify the features. A CNN with training interference is proposed. Two strategies are used, one is variable Dropout, the other is micro training batch. The noise tolerance is improved by adding training interference. The results demonstrate that the proposed method achieves an accuracy exceeding 97 %, confirming the effectiveness of the CNN-based approach for transformer Short-circuit impedance testing. • A three-stage framework is proposed for transformer short-circuit fault diagnosis. • Fifty-two electrical features are constructed to capture temporal and spatial fault characteristics. • A MIC-ReliefF-based feature selection method improves diagnostic accuracy and reduces overfitting. • An MTI-DCNN model demonstrates strong robustness under severe noise conditions. • The proposed approach maintains over 90 % accuracy even at 0 dB SNR.