A Novel Dual-Stage Semi-Supervised Learning Model for Fault Diagnosis in Transformers With Limited Imbalanced Labeled Data

Abstract

Dissolved gas analysis (DGA) methods, a key technique in transformer fault diagnosis, face a typical multicategory data-imbalance problem due to the extremely heterogeneous distribution of fault types across different transformers in their respective operating environments. In addition, the high cost and time requirements for fault-type labeling result in scarce relevant data. To address this issue, a model called the dual-stage semi-supervised learning model (DS-SSLM) is proposed for transformer fault diagnosis with a small amount of unbalanced labeled data. This model uses a two-phase training approach to mitigate the negative impact caused by dataset imbalance. In the pretraining phase, a cross-entropy loss function is used to minimize the difference between the predicted and actual values of the model. Simultaneously, a student model and a teacher model are introduced to process different augmented versions of the same unlabeled sample, aiming to minimize the discrepancy between their outputs, thereby enhancing the model’s robustness to data imbalance. In the downstream task, the classification head of the model is fine-tuned to improve classification reliability. Model performance tests and comparative experiments conducted on the IEC TC 10 dataset demonstrate that the proposed method is robust to data imbalance and achieves high-precision fault diagnosis.