A Semi-Supervised Fault Diagnosis Method for Transformers Based on Discriminative Feature Enhancement and Adaptive Weight Adjustment

The deep learning algorithms have become the general trend in transformer fault diagnosis. The diagnostic accuracy is contingent upon the quantity of fault labeled samples. However, obtaining sufficient fault labeled samples remains a challenge and the collected samples in practice are unlabeled. Therefore, a semi-supervised fault diagnosis method is proposed for the transformer based on discriminative feature enhancement (DEF) and adaptive weight adjustment (AWA). First, the pseudolabels of unlabeled samples were generated through graph propagation, and the quality of pseudolabels is crucial for fault diagnosis. Second, the DEF was used to improve the quality of pseudolabels by optimizing diagnostic boundaries. Then, the weight of pseudolabels involved in training was adaptively adjusted using the truncated normal distribution function, based on the deviation between the confidence of pseudolabels and the mean of the normal distribution function. Finally, the proposed method was verified on the collected dataset. The experiment results demonstrated that the proposed method can guarantee the high quality and quantity of pseudolabels involved in training. The proposed method achieves a diagnostic accuracy of 94.1% for transformer faults.