Few-Shot power transformers fault diagnosis based on Gaussian prototype network

• Explored the performance of the few-shot algorithm and transformer dissolved gas analysis data after organic combination. • A Gaussian Prototype Network (GPN) based transformer fault diagnosis study under few-shot learning is proposed for improving the problem of accurate identification of transformer faults. • The classification results of the Gaussian prototype for dissolved gas analysis boundary samples are drawn. • The performance of transformer fault diagnosis based on Gaussian prototype is better than the four algorithms compared in the paper. Power transformer diagnostic methods based on traditional intelligent learning are affected by the scarcity of transformer fault data, which hinders their further application and prevents them from obtaining high diagnostic accuracy. To solve this problem, a few-shot method based on Gaussian Prototype Network (GPN) is proposed to achieve an effective and accurate diagnosis of power transformers using even a small number of fault samples. The method is an organic combination of embedding network and distance metric. The proposed approach is verified by datasets of dissolved gas and literature, which come from real power transformers and historical data. The results show that the method can achieve up to 96.7% accuracy, which is suitable for the field of power transformer fault diagnosis.