Abstract
High-voltage bushings are critical insulation components, yet conventional PRPD-based severity assessment methods that rely on global pattern morphologies such as "rabbit ears" and "tortoise shell" remain coarse, lack local sensitivity, and fail to track continuous degradation. This paper proposes a dynamic severity assessment method that shifts the focus from global contours to dense partial discharge (PD) clusters, defined as high-density aggregations of PD pulses in specific phase-magnitude regions of PRPD patterns. Each dense cluster is treated as the statistical projection of a physical discharge channel, and the evolution of its number, intensity, location, and shape provides a fine-scale description of defect development. A multi-level relative density and morphological image processing algorithm is used to extract dense clusters directly from PRPD histograms, followed by a 20-dimensional feature set and a five-index system describing discharge activity, development speed, complexity, instability, and evolution trend. A fuzzy comprehensive evaluation model further converts these indices into three severity levels with confidence measures. Long-term degradation tests on defective bushings demonstrate that the proposed method captures key turning points from dispersed multi-cluster patterns to a single dominant cluster and yields a stable, stage-consistent severity evaluation, offering a more sensitive and physically interpretable tool for condition monitoring and early warning of HV bushings. The method achieved a high evaluation confidence (average 60.1%), which rose to 100% at the critical failure stage. It successfully identified three distinct degradation stages (stable, accelerated, and critical) across the 49 test intervals. A quantitative comparison demonstrated significant advantages: 8.3% improvement in early warning (4 windows earlier than IEC 60270), 50.6% higher monotonicity, 125.2% better stability, and 45.9% wider dynamic range, while maintaining physical interpretability and requiring no training data.