Quantitative Detection Method for Insulator Contamination Based on Residual Structure Convolutional Neural Network

Non-contact detection of insulator contamination level helps to timely implement preventive measures and schedule insulator cleaning cycles, thereby ensuring the safe operation of transmission lines. However, traditional methods for detecting insulator contamination are cumbersome and prone to significant human error. To address these issues, this paper collects hyperspectral images of both artificially prepared and actual contamination samples. Through image correction, spectral preprocessing, and other methods, full-spectrum hyperspectral spectral lines are extracted to establish a hyperspectral spectral line dataset. Based on the salt-ash spectral line characteristics in the hyperspectral dataset, a contamination detection network for insulators is constructed using a residual convolutional neural network module, enabling the quantitative detection of insulator contamination. Comparative experiments conducted on XGBoost, traditional convolutional neural networks, and residual convolutional neural networks show that the residual convolutional neural network-based model for quantifying insulator contamination performs better in full-spectrum spectral modeling. It demonstrates good fitting ability on both artificially prepared and actual sample datasets, providing reliable contamination data for scheduling insulator cleaning cycles.