High-Frequency Transformers Fault Diagnosis under Multi-Physical Field Coupling Based on Feature Learning

High-frequency transformers (HFTs) are used in applications such as renewable energy, electric vehicles, and aerospace power distribution. However, the complex multi-physical interactions within HFTs, including electromagnetic, thermal, and mechanical couplings, make them highly susceptible to various fault conditions, such as harmonic distortions and DC biasing. In this paper, a fault diagnosis method leveraging multi-physics field coupling and feature learning-based image analysis is proposed. The deformation and thermal stress fields under different operating conditions—including no-load, harmonic excitation, and DC bias—are simulated and captured through high-frequency image data. A convolutional neural network (CNN) is employed to automatically extract hierarchical features from the temperature and stress distributions, while a complementary bag-of-visual-words (BoVW) approach further enhances feature representation. The proposed method integrates data augmentation techniques tailored for multi-modal inputs, improving the generalization capability of the model under high-noise and small-sample scenarios. Experimental results demonstrate that the developed approach achieves a fault classification accuracy exceeding 97%, significantly outperforming traditional single-modal diagnostic methods. The model's ability to differentiate subtle fault patterns and handle overlapping feature distributions is validated, highlighting its robustness in diverse and noisy environments.