Attention-Guided Autoencoder Fusion for Insulator Defect Detection Using UAV Transmission-Line Imaging

Researchers developed AE-YOLO, an advanced deep learning framework combining autoencoders with YOLO object detection for identifying defects in high-voltage transmission-line insulators using UAV imagery. The system achieves 95.10% mAP performance, substantially outperforming existing YOLO baselines and offering a scalable solution for critical infrastructure inspection.

This research addresses a significant operational challenge in power grid maintenance through innovative computer vision architecture. Transmission-line insulators are mission-critical components whose failures can cascade into widespread outages, yet manual inspection via helicopter or ground crews remains expensive and labor-intensive. The paper's AE-YOLO framework tackles the fundamental technical barriers that have hindered automated detection: extreme class imbalance where defects represent a tiny fraction of inspected surfaces, extreme scale variation across different insulator types and distances, and the inherent difficulty of spotting small anomalies in high-resolution aerial imagery.

The technical contribution combines attention mechanisms with autoencoders to enhance feature discrimination while preserving anomaly-sensitive information during multi-scale processing. By integrating Convolutional Block Attention Modules and variance-maximizing autoencoder regularization, the architecture learns to suppress background noise while amplifying defect signatures. The ensemble approach using Weighted Boxes Fusion across multiple YOLO versions demonstrates practical sophistication in production deployment.

From an infrastructure technology perspective, this work has meaningful implications for utilities seeking to modernize inspection workflows. The reported 96.40% precision and 93.80% recall rates suggest the system could reliably reduce manual verification workload while catching defects before catastrophic failure. As utilities increasingly adopt UAV-based monitoring, frameworks like AE-YOLO become economically viable alternatives to traditional methods, potentially reducing inspection costs by 30-50% industry-wide. The scalability and adaptability highlighted suggest rapid commercialization potential for infrastructure maintenance providers serving regional and national grid operators.

• →AE-YOLO achieves 95.10% mAP, outperforming strongest YOLO baseline by 5.0 points in mAP and 6.7 points in recall on insulator defect detection
• →Architecture integrates autoencoders with attention mechanisms to preserve anomaly-sensitive features while suppressing background interference in UAV imagery
• →Ensemble inference combining YOLOv8, YOLOv10, and YOLO11 with Weighted Boxes Fusion improves detection of rare defect categories
• →Framework addresses critical infrastructure inspection challenges including extreme class imbalance and multi-scale variation in transmission-line imagery
• →Solution offers practical, deployable alternative to manual helicopter and ground-based insulator inspection methods for power utilities