Diffusion-generated images sharpen AI detection of faulty power grid

Power grid operators increasingly use drones to photograph insulators on high-voltage lines, but AI has struggled to learn from too few images of faulty equipment. A research team led by Stefano Frizzo Stefenon and others reports a cascaded deep-learning framework that uses conditional diffusion-generated images to correct the data imbalance while also making model reasoning visible. Published in Autonomous Intelligent Systems, the YOLO26-Swin pipeline first uses a YOLO26x detector to draw bounding boxes, then a Swin-V2-B classifier labels each insulator as good, broken, or damaged by flashover. The authors used Bayesian optimization for detector hyperparameters and generated synthetic fault images based on a dataset of 1,596 photos with 17,108 labeled instances.