Quantitative Detection Method for Insulator Contamination Based on Imbalanced Regression

The detection of insulator contamination is crucial for ensuring the reliable operation of power systems. Non-contact detection methods primarily rely on the integration of various optical technologies and deep learning models. However, the distribution of insulator contamination data is highly imbalanced, making insulator contamination detection a typical imbalanced regression task. This paper analyzes the distribution characteristics of an insulator contamination dataset and experimentally compares the effectiveness of label distribution smoothing algorithms, resampling algorithms, error-sensitive loss functions, and cost-sensitive reweighting methods in addressing data imbalance. Furthermore, it explores the performance of different method combinations. The research results indicate that combining error-sensitive loss functions, reweighting methods, and label distribution smoothing algorithms significantly enhances the regression performance of the model across high-, medium-, and low-density contamination levels, improving its fitting capability for data across different density ranges.