A Lightweight Convolutional Neural Network for Edge Computing-Oriented Transmission Line Defect Recognition
IEEE Access · 2026 · 14 · 43185-43199
DOI: 10.1109/access.2026.3672203 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von He Su und 2 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Power transmission line inspection is crucial for ensuring the reliability of electrical infrastructure. However, manual inspection methods pose significant safety risks and are often inefficient. This paper presents a systematically optimized lightweight convolutional neural network (CNN) architecture for edge-deployed transmission line defect detection. The proposed model leverages MobileNetV3-Small for robust feature extraction and incorporates a bidirectional feature pyramid network for effective multi-scale defect analysis. The CNN employs a decoupled detection head to enhance classification and localization performance. To achieve efficient computation, model compression strategies are applied, including knowledge distillation to improve model accuracy and post-training INT8 quantization to reduce computational load. These approaches substantially lower computational requirements while maintaining high precision in object detection tasks. Results demonstrated a mean average precision of 87.4% in transmission line defect classification at IoU-th = 0.5, evaluated on 1,268 test images. The model required only 2.1 million parameters and 5.8 GFLOPs, making it suitable for edge deployment. Practical validation on edge devices showed inference speeds of 89 FPS on an NVIDIA Jetson Xavier NX and 52 FPS on an NVIDIA Jetson Nano, with an energy efficiency of 12.4 frames per watt in power transmission inspection tasks.
Zitieren
APA
Su, H., Li, K., & Wang, Z. (2026). A Lightweight Convolutional Neural Network for Edge Computing-Oriented Transmission Line Defect Recognition. IEEE Access, 14, 43185-43199. https://doi.org/10.1109/access.2026.3672203
Vancouver
Su H, Li K, Wang Z. A Lightweight Convolutional Neural Network for Edge Computing-Oriented Transmission Line Defect Recognition. IEEE Access. 2026;14:43185-43199. doi:10.1109/access.2026.3672203
BibTeX
@article{Su2026,
author = {Su, He and Li, Kui and Wang, Zhenzhou},
title = {A Lightweight Convolutional Neural Network for Edge Computing-Oriented Transmission Line Defect Recognition},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {43185--43199},
doi = {10.1109/access.2026.3672203}
}
EUVIMED