Hydrophobicity Classification of Composite Insulators Using Multi-scale Water Droplet Detection and Feature Fusion

Hydrophobicity degradation is a critical factor leading to pollution flashovers in composite insulator systems. The International Electrotechnical Commission (IEC) 62073 visual inspection method suffers from ±1 level measurement uncertainty, while existing deep learning methods lack physical interpretability for maintenance decision-making. This paper presents a hierarchical framework employing a physically grounded intermediate representation. A five-level droplet taxonomy (C1-C5) based on contact angle physics decouples droplet morphology from surface assessment. An enhanced Faster Region-based Convolutional Neural Network (Faster R-CNN) with Squeeze-and-Excitation Residual Network-50 (SE-ResNet-50) backbone achieves 75.8% mean Average Precision on 16,535 labeled droplets. The fusion of 21 detection-derived and five image-processing features yields a 4.08% improvement over the naive combination. Evaluated on 1,160 independent images spanning four climate zones and 2-13 years of service life, the method achieves 77.09% Full Hit Accuracy (FHA, exact HC level match) and 99.13% Tolerant Accuracy (TA, predictions within ±1 level), compliant with IEC 62073 uncertainty tolerance. The hierarchical design enables verification of classification reasoning through inspectable C1-C5 intermediate outputs.