What is it about?
Power transformers often produce partial discharge (PD) signals when insulation defects begin to develop. Detecting and identifying these discharge types early is essential for preventing equipment failures and improving grid reliability. However, PD signals are highly complex, noisy, and non-stationary, making accurate diagnosis challenging. This study proposes a new fault diagnosis framework that combines an Improved Symplectic Geometric Mode Decomposition (ISGMD) method with a Time-Shifted Multi-Scale Fusion Entropy (TSMFE) feature extraction approach. ISGMD first separates complex PD signals into meaningful components while reducing mode mixing and redundant information. TSMFE then captures signal complexity from both time-domain and frequency-domain perspectives across multiple temporal scales. Finally, the extracted features are classified using a hybrid deep-learning model that combines Temporal Convolutional Networks (TCN), Bidirectional Gated Recurrent Units (BiGRU), and an Attention mechanism. The method is validated using four typical transformer discharge types: corona discharge, air-gap discharge, floating discharge, and surface discharge.
Featured Image
Photo by American Public Power Association on Unsplash
Why is it important?
Reliable identification of transformer insulation defects is critical for condition-based maintenance and preventing costly outages. Existing signal decomposition and entropy-based feature extraction methods often suffer from mode mixing, parameter sensitivity, or insufficient discrimination between different discharge types. The proposed ISGMD-TSMFE framework improves both signal representation and feature separability. Experimental results show that the method achieves a classification accuracy of 98.33%, outperforming several existing decomposition–entropy combinations. The approach also demonstrates strong robustness under imbalanced datasets, which is particularly important because real-world fault data are often unevenly distributed across fault categories. These results suggest that the method can provide more reliable transformer condition monitoring and support early fault detection in practical power system applications.
Perspectives
What makes this work particularly interesting is that it addresses two fundamental challenges in partial discharge diagnosis simultaneously: extracting meaningful information from highly non-stationary signals and distinguishing subtle differences among discharge mechanisms. Rather than relying solely on more complex deep-learning architectures, the study emphasizes improving signal quality and feature representation before classification. The combination of ISGMD and TSMFE provides a physically interpretable way to reveal discharge characteristics hidden in noisy measurements, while the deep-learning classifier further enhances recognition performance. This integration of advanced signal processing and intelligent diagnosis offers a practical pathway toward more reliable transformer health monitoring and predictive maintenance.
Chair, IEEE PES EICC Task Force on AI-Enabled Resilience of CPES|Clarivate HCR|AE: IEEE TSG/TSTE/TII Yang Li
Northeast Electric Power University
Read the Original
This page is a summary of: An improved symplectic geometric mode decomposition and time-shifted multi-scale fusion entropy approach for transformer partial discharge fault diagnosis, Chaos Solitons & Fractals, September 2026, Elsevier,
DOI: 10.1016/j.chaos.2026.118577.
You can read the full text:
Contributors
The following have contributed to this page
