Abstract: To address the issues of scarce fault data for traction converters and the weak cross-condition generalization capability of traditional methods, a cross-domain transfer fault diagnosis method based on multi-dimensional features is proposed. Time-domain and frequency-domain features are extracted, and sensitive features are screened to construct a fused diagnostic model integrating support vector machine (SVM) and multi-layer perceptron (MLP). Experiments show that the proposed method achieves an accuracy rate of over 90% for simulated grounding faults under different load conditions. After transferring the diagnostic scheme to real-world overcurrent fault data and applying data augmentation, the model maintains an accuracy rate of 94%. The proposed method demonstrates good accuracy and generalization capability, providing a low-sample-dependent cross-domain solution for traction converter fault diagnosis.