Mask-Superpoint Fusion for Contamination Visualization and Graded Assessment on Insulators

Transmission line insulator contamination detection is a key aspect of preventing contamination flashover, but existing methods have limitations in contamination visualization assessment and non-contact detection. In this context, this study presents an advanced methodology for detecting and visualizing contamination on transmission line insulator surfaces. Leveraging spectral imaging and computer vision technologies, a multi-view multispectral imaging device is developed, integrated with a non-coaxial spectral image preprocessing model that combines Mask R-CNN for instance segmentation and Superpoint for feature matching. This fusion enables precise spatial alignment and effective removal of background interference, facilitating the visualization of contamination degree and distribution. Additionally, a graded contamination diagnostic model is constructed using a 6 dimensional spectral feature vector and optimized via Bayesian hyperparameter tuning, achieving a testing accuracy of 95.3% with random forest classification. Transfer learning is applied to adapt the model to field conditions, enhancing its practicality for natural contamination scenarios with an accuracy improvement from 71.5% to 87.8%. Field tests on operational 110kV and 220kV transmission lines demonstrate 87.5% accuracy in contamination level assessment. The proposed system offers a robust technological solution for online insulator monitoring and pollution flashover prevention, validated through both laboratory experiments and real-world deployments.