Feature selection by joint manifold structures for transformer fault diagnostic

Fault features derived from dissolved gas analysis (DGA) in transformer oil provide critical prior information for fault diagnosis. However, a dual relationship exists between fault types and diagnostic features—wherein fault categories shape feature relevance, and feature distributions reflect underlying fault mechanisms. Conventional data-driven feature selection methods frequently ignore interactions, resulting in limited generalizability and poor performance on imbalanced datasets. To address these limitations, we proposed a feature selection framework based on joint feature–case manifold structures. By constructing similarity graphs over case–case, feature–feature and case–feature spaces, the framework captures the intrinsic correlations among diagnostic instances. These graphs are embedded with domain-informed initial weights and optimised using manifold learning to generate a structurally consistent and noise-resilient feature score distribution. The selected features were evaluated through a hierarchical decision tree network to quantify their effectiveness in transformer fault classification. The proposed method achieves an overall diagnostic accuracy of 85.85% and maintains 81.2% accuracy for low-frequency fault types, which demonstrates strong generalizability and robustness to data imbalance. This work establishes a generalisable and structure-informed paradigm for feature selection in condition-based diagnostics, empowering broad classification scenarios of industrial fault prediction.