An improved symplectic geometric mode decomposition and time-shifted multi-scale fusion entropy approach for transformer partial discharge fault diagnosis

Partial discharge (PD) signals in transformers are strongly non-stationary and contain complicated dynamic information, which makes reliable feature extraction difficult. In this study, a feature extraction scheme is developed by combining Improved Symplectic Geometric Mode Decomposition (ISGMD) with Time-Shifted Multi-Scale Fusion Entropy (TSMFE). The original PD signal is first decomposed into initial symplectic geometric components, after which hierarchical clustering is introduced to reorganize the decomposed components. Correlation coefficient (CC) analysis is then used to retain the components that carry the most representative information. On this basis, TSMFE is employed to describe signal complexity by jointly using time-domain and frequency-domain information, while features are extracted over multiple temporal scales. The distribution differences among discharge types are further examined through t-SNE, which provides a visual interpretation of the high-dimensional feature space. The extracted features are subsequently fed into a TCN-BiGRU-Attention network for PD type classification. The proposed method produces clearer class separation and achieves a classification accuracy of 98%. Compared with existing methods, it also shows better overall performance, indicating its practical value for transformer partial discharge diagnosis.