Transformer fault diagnosis method based on Gramian angular field and graph convolution network

Power transformers, being pivotal in electrical power systems, necessitate vigilant monitoring for operational sustainability and reliability. The traditional approach to such monitoring, Dissolved Gas Analysis (DGA), has been grounded in gas volume fraction ratio analysis—a technique that, while effective, is constrained by static thresholds and simplistic assumptions that often do not encapsulate the intricacies of fault progression. In light of these limitations, the exploration of Artificial Intelligence (AI) methodologies presents a compelling alternative, offering a dynamic and data-driven avenue for fault detection and diagnosis. This paper proposes a transformer fault diagnosis method based on the Gramian Angular Field (GAF) transformation and Graph Convolutional Networks (GCN). Firstly, the one-dimensional DGA fault samples are converted into three-dimensional DGA feature image samples using the Gramian Angular Field transformation method. Then, a dataset expansion is performed on the image samples using a GAN-based data augmentation method to meet the data volume input requirements of the deep learning model. Subsequently, a graph convolutional network based on GCN layers, and a Self-Attention Graph Pooling (SAGPool) mechanism is constructed. This network aggregates neighbor nodes according to the node features and topological structure of the three-dimensional DGA graph signal to adaptively mine fault features and diagnose fault types. The experimental results show that the method proposed in this paper can serve as an effective transformer fault diagnosis method. It is more accurate compared to the traditional gas ratio method for diagnosis, and its advantages are more significant compared to other traditional fault diagnosis deep learning methods.