Label-Free Partial Discharge Severity Assessment in Gas-Insulated Switchgear via Semantic-Guided Unsupervised Domain Adaptation

As an early indicator of insulation degradation, partial discharge (PD) provides critical information for evaluating the insulation health of gas-insulated switchgear (GIS). However, existing PD severity assessment methods generally rely on sufficient labelled samples collected under controlled conditions, which limits their applicability to practical field environments characterized by domain distribution discrepancies, class imbalance, and limited labelled data. Moreover, the semantic knowledge embedded in inspection records, such as discharge types and defect locations, is seldom utilized in existing data-driven assessment frameworks. To address these challenges, this paper proposes a semantic-guided unsupervised domain adaptation network (UDAN) for label-free PD severity assessment in field-deployed GIS. The proposed framework jointly leverages unstructured textual knowledge and structured PD detection data to improve cross-domain assessment performance under unlabelled field conditions. First, a semantic analysis strategy is introduced to transform unstructured inspection records into structured semantic representations for automatic knowledge learning. Second, an improved Transformer-based feature extraction architecture is developed to simultaneously capture local discriminative characteristics and global contextual dependencies within PD signals. Third, a task-oriented domain adaptation framework is developed for PD severity assessment by jointly performing weighted marginal distribution alignment and adversarial conditional alignment. Unlike conventional domain adaptation approaches primarily designed for PD diagnosis or PD classification, the proposed framework is specifically tailored to address the unique challenges of PD severity assessment, including continuous degradation evolution, severe class imbalance, and heterogeneous information representation in practical field environments. Experimental results demonstrate that the proposed UDAN achieves an assessment accuracy of 96.00% for label-free PD severity assessment in field GIS scenarios. By extending unsupervised domain adaptation from conventional PD diagnosis to label-free PD severity assessment, the proposed framework provides an effective and practical solution for intelligent GIS insulation condition assessment under realistic operating conditions.