Real-Time Power Line Insulator Defect Detection via Improved PSO-Driven Hardware-Aware Neural Architecture Search

Insulator defect detection is essential for ensuring the safety and reliability of power transmission lines. Although deep learning models like YOLOv7 achieve high accuracy, their manually designed architectures tend to be computationally heavy and unsuitable for deployment on resource-limited edge devices. To tackle these issues, the authors propose EHANN-NAS, a neural architecture search framework optimized for real-time insulator defect detection on edge platforms. EHANN-NAS employs a Graph Convolutional Network (GCN)-based surrogate model to efficiently predict the performance of candidate architectures without full training, significantly reducing search time and computational cost. Furthermore, it integrates an enhanced Particle Swarm Optimization (PSO) approach featuring uncertainty-aware sampling and adaptive diversity mechanisms to better explore the architecture space and identify lightweight yet effective network designs. Experiments on a real-world dataset show that EHANN-NAS achieves a mean Average Precision (mAP0.5) of 95.1%, outperforming state-of-the-art methods.