Regression-Based Estimation of Silicone Rubber Hydrophobicity Via Deep Learning

This study proposes a new regression-based deep learning approach to quantify SiR insulator hydrophobicity by analyzing contact angles from water droplet images. A comprehensive dataset of 360 images was acquired by capturing droplets under varying corona discharge exposure and recovery conditions. Advanced data preprocessing techniques, including object detection, cropping, labeling, and augmentation, were employed. The state-of-the-art deep learning architectures, such as ResNet50, InceptionV3, VGG16, and EfficientNetB6, were fine-tuned for a multi-output regression task, simultaneously predicting the left and right contact angles. The VGG16 model with 100% frozen layers achieved the lowest mean absolute error (MAE) of 4.67. This demonstrates its robustness in leveraging pre-trained weights. The proposed approach enables a continuous and granular evaluation of hydrophobicity, facilitating timely interventions and optimized maintenance strategies for the SiR insulators.