Hierarchical Deep Feature Fusion for High-Voltage Insulator Contamination Diagnosis

Abstract

The image-based identification of surface contamination on high-voltage insulators may support preventive inspection and maintenance. Existing deep learning approaches frequently employ single-stage classification, which can make it difficult to distinguish visually similar contamination types. This study proposes a hierarchical, two-stage framework that separates clean-versus-contaminated detection from contamination-type classification. The framework evaluates pairwise combinations of features extracted from pretrained convolutional and Transformer-based architectures, followed by principal component analysis and conventional machine learning classifiers. On a synthetic insulator image dataset, the best configurations achieved classification accuracies of 99.09% for Stage I and 99.54% for Stage II. Nonparametric statistical tests indicated significant improvements over the ResNet50 baseline. These results demonstrate promising classification performance under controlled synthetic conditions; validation on real-world images remains necessary.