Partial Discharge Anomaly Detection and Severity Assessment for HV Bushings Based on Self-Supervised Adversarial Encoder Network and Gaussian Mixture Model

High voltage bushings are critical insulation components in power equipment, directly affecting the safe and stable operation of power systems. Partial discharge (PD) is a primary indicator of insulation degradation in bushings, but existing PD detection and assessment methods face severe limitations: (1) heavy dependence on large amounts of labeled data that cannot cover rare fault scenarios; (2) limited generalization across diverse equipment conditions; (3) lack of adaptive severity grading, relying on fixed thresholds that cannot reflect actual degradation processes. These limitations hinder timely identification of critical insulation states. To address these challenges, this paper proposes a PD anomaly detection and severity assessment method based on self-supervised adversarial encoder network and Gaussian mixture model (GMM). Unlike standard autoencoders that over-generalize, the Transformer-based adversarial encoder with dual competing decoders creates sharper anomaly boundaries, enabling high-sensitivity detection of weak PD signals without labeled data. A GMM with composite clustering evaluation automatically determines optimal severity levels by capturing the natural multimodal distribution of PD anomaly scores, where different degradation stages form distinct Gaussian components. Transparent bushing experiments establish quantitative mapping between anomaly scores and observable physical phenomena (gas generation, thermal decomposition), providing interpretable diagnostic criteria. Experimental results demonstrate effective PD detection across developmental stages, automatic severity identification, and quantitative condition assessment. Particularly when approaching critical states, the method captures significant anomaly score changes, enabling timely risk warnings. Compared with existing methods, this approach achieves high-precision detection without labeled data while providing adaptive severity assessment, offering significant practical value.