Enhanced Graph Skeleton Trees with Adaptive Feature Learning for Transformer Condition Monitoring

Transformer condition monitoring and fault diagnosis are critical issues in ensuring the long-term effective operation of modern power system. However, existing graph neural network methods are unable to carry out thorough and comprehensive feature extraction, which will be exacerbated when dealing with the complex, strongly coupled power measurement data. To address this, this paper proposes the enhanced adaptive skeleton tree (EAST) model to realize comprehensive feature extraction and enhanced monitoring. First, an initial graph is constructed according to the oil chromatographic data of transformer. Then, the improved tree decomposition is employed to represent the graph as a skeleton with multiple hierarchical branches, each of them conducts feature learning and long-range information flow independently. Ultimately, the multi-level branches facilitate the adaptive feature learning and the comprehensive graph-level representation is derived for condition monitoring. Experimental results on real-world transformer system demonstrate the practicality and effectiveness of the proposed EAST method.