Enhanced High-Voltage power line insulator and contamination classification using score-level fusion

Abstrakt

High-voltage (HV) power line insulators are critical
for grid reliability, but their performance degrades due to
contamination (e.g., salt, soot, excrement). Traditional visual
inspection methods are subjective, risky, and time-consuming.
To overcome these challenges, this paper proposes an efficient
and accurate framework for classifying insulator materials (glass,
porcelain, composite) and contamination types. The framework
employs features extracted independently from three lightweight
CNNs: MobileNetV2, ShuffleNet, and EfficientNet-B0. These
features are then fed into base classifiers, and their outputs
(scores) are combined using various score-level fusion rules
(Majority Vote, Maximum, Average, Sum, Minimum, Product) to
enhance classification accuracy and robustness. The framework’s
effectiveness is validated on three datasets, including synthetic
contamination scenarios and real-world images from the Merged
Public Insulator Dataset (MPID). Results demonstrate that score
fusion significantly outperforms individual lightweight models,
achieving accuracies up to 98.49% for contamination classification,
98.59% for combined material/contamination classification,
and 99.26% for real-world material identification. Comparative
analysis demonstrates significant improvements over
existing methods, including VGG16 (97.00%) and custom CNNs
(98.00%), highlighting the efficacy of feature and score fusion.
The results validate the framework’s adaptability to diverse environments,
computational efficiency, and potential for deployment
in resource-constrained settings.