A Power Transformer Fault Diagnosis Method Based on Hierarchical Classification and Ensemble Learning

Abstract

Since conventional machine learning methods often result in low diagnostic accuracy and non-negligible recognition difference among fault types with imbalanced class distribution among transformer fault types, a multi-level hierarchical power transformer fault diagnostic model is proposed based on hierarchical classification and ensemble learning, where classifiers are hierarchically constructed for level-by-level diagnosis according to the imbalance extent on each level. The Level I neural network classifier extracts 3 generalized feature labels of normal, discharge fault and thermal fault for feature fusion with original data input, to guide classification among 9 detailed operation status under DL/T 722-2014 standard. The Level II classifier adopts EasyEnsemble, generating balanced training subsets by undersampling majority classes and training sub-classifiers in parallel for parameter synthesis in ultimate classifier, to balance information between major fault types and minor ones. Experimental result shows that: compared to traditional methods, our proposed method improves the generalization ability on minority class faults and the overall accuracy by 7%.