Prior Knowledge-Guided Active Learning Network for Multi-Source Partial Discharge Diagnosis in Gas-Insulated Switchgear

The advancement of artificial intelligence presents both opportunities and significant challenges for multi-source partial discharge (PD) diagnosis in gas-insulated switchgear (GIS), but challenges remain due to insufficient incorporation of underlying physical mechanisms, limited labeled data, noise interference, class imbalance, and difficulties in decoupling mixed PD sources. To address these issues, this paper proposes a prior knowledge–guided active learning network (PKGALN) for multi-source PD diagnosis in GIS, in which physics-related PD characteristics are explicitly embedded into the model design under a limited-label regime. Specifically, a prior knowledge–guided U-former is constructed by incorporating physically interpretable PD descriptors (e.g., time–frequency statistical features reflecting discharge behavior) as global guidance for feature learning, while a scale-guided convolutional–wavelet encoder enhances denoising and multi-scale representation capability. In addition, a multi-label decoupling recognition module is designed to explicitly handle mixed PD sources, and a distribution-balanced loss is adopted to alleviate severe class imbalance. To further reduce annotation cost, an adaptive active learning strategy is introduced, dynamically balancing uncertainty and diversity during sample selection. Experimental results on both laboratory and field datasets demonstrate that PKGALN achieves diagnostic accuracies of 96.39% and 93.26%, respectively, using only 10% of labeled samples. By leveraging limited labeled data, the proposed approach enables robust diagnosis of both single-source and multi-source PDs.