Machine learning evaluation of leakage current indicators for polymer-based insulating material degradation assessment

Reliable identification of degradation in polymer-based insulating material is essential to ensure the operational reliability of high-voltage transmission and distribution systems. Leakage current (LC) characteristics are widely used as diagnostic indicators to evaluate the surface condition of polymeric insulating materials. This study investigates the effectiveness of two LC indicators, namely total harmonic distortion (%THD) and root mean square current (Irms), in differentiating the degradation levels of field-aged EPDM (Ethylene Propylene Diene Monomer) polymer insulators using machine learning-based classification techniques. Three binary classifiers, namely Support Vector Machine (SVM), K-Nearest Neighbours (KNN), and Decision Tree (DT), were implemented to examine the consistency of degradation classification using different learning approaches. The classifiers were implemented using fixed algorithm configurations without hyperparameter optimization to ensure a consistent comparison and to focus on the discriminative capability of the leakage current features. The LC data were acquired during an Inclined-Plane Tracking (IPT) test conducted in accordance with the BS EN IEC 60587:2022 standard, while the classification performance was evaluated using Mean Absolute Error (MAE) at two predetermined classification accuracy reference levels (%CAref = 50% and %CAref = 100%). The results show that the %THD feature consistently produced lower average MAE values ranging from 3.13 to 5.00, whereas the Irms feature produced considerably higher values, ranging from 19.09 to 20.05. These results indicate that %THD provides classification outcomes that are more consistent with the predetermined reference levels than Irms. The consistent performance trend across multiple classifiers indicates that the effectiveness of degradation classification is strongly influenced by the discriminative capability of the extracted leakage current indicator rather than the specific classifier employed. The classification outcomes were further compared by hydrophobicity assessments based on the IEC TS 62073:2016 guideline, including contact angle measurement and wetted area analysis, which provided supporting physical evidence for the correspondence between LC indicators and surface condition of the insulators. Overall, within the observation-level evaluation employed in this study, the findings indicate that %THD provides more consistent classification outcomes than Irms and shows potential as an LC indicator for supporting machine learning-based assessment of polymer insulator degradation.