Abstract: To address sample scarcity and operating condition shift in fault prediction for transformer dissolved gas analysis (DGA) and high-voltage circuit breaker opening and closing coil current waveforms, this paper proposes a few-shot transfer learning method that integrates self-supervised contrastive learning, maximum mean discrepancy (MMD) feature alignment, and knowledge distillation. The method first uses a SimCLR variant to perform self-supervised contrastive learning pre-training on unlabeled data from the source domain and extract intrinsic feature representations of equipment states. It then uses MMD constraints to remove feature drift caused by different operating conditions between the source domain and the target domain. A combined strategy of parameter transfer and knowledge distillation is used to compress the model and adapt it to edge deployment requirements. Finally, fine-tuning is completed in the target domain with only five labeled samples for each class. Experimental results on a transformer DGA dataset and a high-voltage circuit breaker coil current dataset show that the proposed method achieves accuracies of 89.3% and 91.7%, respectively, outperforming existing comparison methods.