A Real-Time Cascade Framework for UAV-Based Insulator Defect Detection with Attention-Guided Lightweight CNN

The physical condition of insulators on electrical transmission lines is critical for system reliability. Early fault detection such as broken discs or flashovers prevents outages and accidents. Traditional inspection methods are costly and difficult, driving demand for UAV-based autonomous systems. This article proposes a novel dataset enhancement method and a cascade deep learning model for detecting and classifying insulator disc conditions. Existing datasets were enriched by approximately 30% with wide-background, real-field images to improve model generalization. After comparative analyses, YOLO was selected for detection due to superior performance, while MobileNetV2 was chosen as the classifier for its processing speed advantage. CBAM attention module and Focal Loss function were integrated to boost classification performance and handle class imbalance. The resulting YOLO + MobileNetV2 model achieved 99.76% overall accuracy across "Normal", "Broken", and "Flashover" classes, with an F1-score of 0.99 for broken discs, operating with 185.60 ms latency, confirming its viability for real-time inspection.