Combining Small Sample Parameters and GRU to Perceive and Locate Latent Faults in Power Transformers

Abstract

To address the insufficient temporal modeling capability in latent fault diagnosis of power transformers under small-sample conditions, this paper proposes an intelligent diagnosis framework integrating Few-Shot Learning (FSL) and a Gated Recurrent Unit (GRU) network for electromagnetic field perception and fault localization. The method employs an unsupervised deep autoencoder to perform dimensionality reduction and extract discriminative features from high-dimensional monitoring data, while a multi-physical weighted fusion mechanism dynamically integrates current, voltage, and other electromagnetic operating parameters to enhance feature representation. K-means clustering is further introduced to improve the separability of small-sample characteristics. A dual-channel GRU jointly models historical and real-time monitoring sequences to capture temporal dependencies and accurately locate latent fault regions through a field-aware localization strategy. Experimental results demonstrate that the proposed approach achieves a test accuracy of 0.83–0.87 and an average positioning error of 0.58 across ten fault categories, outperforming conventional GCN and TCN models. Even when only 5% of training samples are available, the framework maintains reliable diagnostic performance with an accuracy of 0.75 and a positioning distance of 0.78. The proposed method provides an effective solution for intelligent perception and localization of latent transformer faults while offering valuable support for electromagnetic monitoring and condition assessment of modern power equipment operating in complex electromagnetic environments.