FlashDetR: A deep learning pipeline for early detection and time estimation of flashover in high-voltage insulators using infrared videos

Flashover in high-voltage insulators poses a significant risk to power system reliability, potentially leading to outages and safety hazards. This study introduces an innovative deep learning-based approach for early prediction of flashover events and time-to-flashover estimation by analyzing infrared videos of dry band arcing, a known precursor to flashover. In this work, we propose a pipeline named Flashover Detector and Time Estimator , which integrates a transformer-based model to accurately predict flashover occurrences, while a Three Dimensional Convolutional Neural Network-based model estimates the time to flashover. Flashover Detector and Time Estimator progressively samples video frames at multiple scales, enhancing prediction accuracy. Experimental results demonstrate that the models achieve up to 88.73% accuracy in predicting flashover events and a mean absolute error of 3.41 in time-to-flashover estimation. These findings substantially improve the ability to implement preventive measures. Flashover Detector and Time Estimator thus represents a significant advancement in proactively managing power system reliability, with demonstrated effectiveness and real-time application potential. • End-to-end DL model for early flashover prediction and precise time-to-flashover. • IR video dataset captured in controlled conditions, showing full DBA progression. • High-accuracy models for early flashover detection and low MAE for time-to-flashover prediction.