A Swin transformer enhanced Depthwise convolution with dynamic channel reweighting for partial discharge diagnosis in gas insulated switchgear

Partial discharge (PD) is a critical indicator of insulation degradation in gas-insulated switchgear (GIS). Accurate PD pattern recognition is essential for fault diagnosis and enhancing the reliability of power equipment. In recent years, deep learning has shown notable success in identifying discharge patterns from phase-resolved partial discharge (PRPD) spectra. However, convolutional networks relying on convolutional kernels limit their ability to capture long range dependencies. Although vision transformer introduces global modeling, their hard patch partitioning may disrupt the phase continuity in spectrum, compromising feature integrity. To address these challenges, this paper proposes convolutional squeeze-excitation with shifted-window transformer (CSEST) which integrates depthwise separable convolutions, squeeze-and-excitation attention mechanism, and a hierarchical Swin Transformer. CSEST adaptively emphasizes discharge-relevant channels and captures long range dependencies through shifted-window self-attention. Experimental results demonstrate that CSEST outperforms mainstream pattern recognition models. Visualization through raincloud plots and t-SNE further confirms the model's robustness and superior discriminative capability, exhibiting clear cross-class separation and same class clustering. These results highlight the CSEST's effectiveness for PD pattern recognition in GIS, enabling early fault detection and supporting the safe, stable operation of power systems.