Cloud-Edge-End Cooperative Detection of Wind Turbine Blade Surface Damage Based on Lightweight Deep Learning Network

Abstract

Blade health is directly related to the safety and efficiency of wind turbine (WT) operation. In this article, a cloud-edge-end collaborative detection method for WT blade surface damage is proposed based on lightweight deep learning network. The blade images are obtained by unmanned aerial vehicle. The YOLOv3 is optimized on the cloud server, including backbone network replacement, filter pruning, and knowledge distillation. After model training, the lightweight deep learning model YOLOv3-Mobilenet-PK is obtained and deployed on edge device to detect the surface damage of the WT blades, then the detection results can be viewed through the portable mobile device. The results show that the mean average precision (mAP) of the detection method proposed in this article is over 90%, the detection speed is about two times that of the YOLOv3-DarkNet53. This method has the advantages of fast detection speed, high accuracy, and less occupation of bandwidth.