A novel adaptive capsule network with dual-branch feature extraction for multi-source partial discharge diagnosis in gas-insulated switchgear

• A novel ACN is proposed for accurate multi-source GIS PD diagnosis in GIS. • A dual-branch feature extraction module is designed to capture local and global PD features. • An adaptive capsule is introduced to improve classification robustness via dynamic routing. • A multi-label classification module is incorporated to support concurrent PD type prediction. • The method achieves 96.68 % accuracy for multi-source PD in noisy, small-sample scenarios. Although state-of-the-art artificial intelligence models have achieved remarkable performance in partial discharge (PD) diagnosis for gas-insulated switchgear (GIS), accurately and robustly identifying multi-source PDs remains challenging due to the complex coupling of signal patterns and noise. To address these limitations, this paper proposes a novel adaptive capsule network (ACN) featuring a dual-branch feature extraction architecture for GIS multi-source PD diagnosis. First, a dual-branch module employing U-Net and U-Transformer in parallel is developed to capture both fine-grained local details and long-range global dependencies of PD signals, with the U-Net structure further enhancing noise robustness through effective suppression. Second, an adaptive capsule structure is then introduced, where a dynamic routing algorithm models interactions among primary capsules, computes coupling coefficients, and adaptively aggregates semantically similar information into higher-level capsules to improve feature discriminability. Finally, a multi-label classification module enables simultaneous prediction of single-source and multi-source PD types, achieving comprehensive diagnosis across both categories. The proposed ACN achieves a multi-source PD diagnostic accuracy of 96.68 % on representative datasets. Experimental results demonstrate its superiority over existing methods, showing not only high diagnostic accuracy but also enhanced noise tolerance and reduced dependence on large labeled datasets, highlighting its novelty and practical applicability in real-world noisy and small-sample scenarios.