A Power Transformer Fault Diagnosis Framework Based on Data Augmentation and Multiscale Residual Convolutional Networks

As a critical component of the power system, power transformers fault diagnosis is essential for system security. Dissolved Gas Analysis (DGA) is a commonly used fault diagnosis method whose model performance is constrained by data imbalance and insufficient sample size. To address this issue, this study proposes a transformer fault diagnosis framework that integrates a data augmentation method based on the Support Vector Machines with Synthetic Minority Over-sampling Technique (SVMSMOTE) and IEC three-ratio encoding, together with a Multi-Scale Residual Convolutional Network (MSRCNN) that extracts hierarchical fault features through residual-enhanced multi-scale convolution. By optimizing hyperparameters through the Optuna framework, the MSRCNN model achieved a classification accuracy of 89.62% on the test set, outperforming other models, including TCN with 86.15%, Transformer with 82.31%, 1D CNN with 80.38%, LSTM with 78.85%, and IEC three-ratio with 76.15%. The experimental results demonstrate that the MSRCNN model exhibits excellent performance in DGA fault diagnosis tasks for oil-immersed transformers.