Wind turbine blade structural state evaluation by hybrid object detector relying on deep learning models

Abstract

Surveillance drones are remarkable devices for monitoring, as they have strong spatial and remote sensing capabilities. The prompt detection of peripheral damage to the blades of wind turbines is necessary to reduce downtime and prevent the potential failure of wind farms. Computer vision breakthroughs with deep learning have developed and been refined over time, mainly using convolution neural networks. From this perspective, we suggest a deep learning model for monitoring and diagnosing the blade health of wind turbines based on images captured by surveillance drones. The main limitations of standard monitoring devices are their poor detection accuracy and lack of real-time performance, making it complex to obtain the attributes of blades from aerial images. Based on the foregoing, this study introduces a method for increasing detection accuracy when carrying out operations in real time using You Only Look at Once version 3 (YOLOv3). We train and evaluate three deep learning models on the wind turbine image dataset. We find that many aerial images are unclear because of blurred motion. As avoiding such low-resolution images for training can affect accuracy, we use a super-resolution convolution neural network to reconstruct a blurred picture as a high-resolution one. The computational results demonstrate that YOLOv3 outperforms traditional models in terms of both accuracy and handling time.