Fault Diagnosis for Power Transformers through Semi-Supervised Transfer Learning.

Area of Science:

• Electrical Engineering
• Artificial Intelligence

Background:

• Power transformer fault diagnosis is complex due to heterogeneous multisource faults and limited historical data.
• Existing methods struggle with undetermined fault types and data imbalances.

Purpose of the Study:

• To develop an advanced fault diagnosis method for power transformers.
• To address limitations of existing methods using deep learning and semi-supervised transfer learning.

Main Methods:

• A novel semi-supervised transfer learning framework, adaptive reinforcement (AR), was developed, enhancing consistency regularization.
• Deep neural networks were trained on source domain data and transferred to target domains with unbalanced and undefined fault datasets.
• Experiments utilized real-world 110 kV power transformer grounding current data for four fault types (Phase A, B, C, and ABC to ground).

Main Results:

• The proposed AR framework achieved over 95% accuracy in classifying fault types.
• The model demonstrated superior performance compared to other popular neural networks.
• The AR framework required significantly fewer epochs (dozens) to adapt to target domain data than other semi-supervised techniques.

Conclusions:

• Combining deep neural networks with the AR framework provides a reliable and efficient solution for diagnosing faults in power transformers with limited historical data.
• The method significantly reduces training time while maintaining high diagnostic accuracy.
• This approach enhances the capability of diagnosing power transformers lacking extensive fault diagnosis knowledge.