Reference-based degradation screening of 150-kV insulators using GWO-optimized one-class SVM under limited puncture-test data

This study proposes a reference-based degradation screening framework for 150-kV ceramic, glass, and polymer insulators using a Grey Wolf Optimizer–One-Class Support Vector Machine (GWO-OCSVM) under limited puncture-test data. The dataset comprised 72 observations, including 36 replacement samples used exclusively as reference data and 36 existing samples reserved for screening and evaluation. Input features were constructed from normalized string position and voltage ratio relative to the material-specific reference mean. Model development employed nested leave-one-position-out validation, while the OCSVM parameters ν and γ were optimized by GWO using reference data only. REF-GWO-OCSVM achieved an AUC of 0.8299, improving over REF-OCSVM at 0.7535, while increasing the mean anomaly-score gap from 0.0560 to 0.1447. The final consensus parameters were ν=0.1338 and γ=1.8257. A conservative P95=0.1970 threshold reduced reference rejection to 2/36 while flagging 9/36 existing samples. The results demonstrate that the proposed method enables reference-based anomaly screening without using existing-condition data during training or hyperparameter optimization.