A Directional Defect Detector for Corona Ring Fault Diagnosis in Power Transmission Line

Fault diagnosis of corona rings in aerial imagery is a crucial task in the inspection of power transmission line. Detecting defects such as corona ring tilt and reverse installation faults requires precise directional judgment based on domain-specific knowledge. Currently, no optimal detection method addresses these challenges effectively. In this paper, we present a novel approach leveraging deep learning, termed the Directional Defect Detector. Our detector contains three parallel convolutional branches and a points group pooling layer to model the spatial positional correlations between corona rings and their surrounding components. This approach reformulates the diagnosis of tilt and reverse installation faults by representing and computing two distinct group embeddings. By integrating both regional and point-based predictions into a unified network, our method offers a streamlined end-to-end solution. Experimental results demonstrate that our detector achieves average precision (AP) scores of 0.87 and 0.93 for the two aforementioned defects on the power inspection dataset, surpassing the performance of all existing general object detectors commonly used in power transmission line inspection.