Detecting Ice on Wind Turbine Rotor Blades: Towards Deep Transfer Learning for Image Data

Abstract

AbstractWind turbines, in particular, their rotor blades, are not only subjected to specific structural loads but also harsh weather conditions. There exists a risk of ice forming on the leading edge of the rotor blade depending on the location and, notably, at lower temperatures and high humidity. Some of the effects include significant power decreases, turbine damages and shutdowns. Ice detection systems for wind turbines operating in cold climates thus become important. Therefore, this paper proposes a method of ice detection that uses RGB-images, taken from rotating nacelles under different conditions, and pre-trained models of MobileNet, VGG-19 and Xception. The output is an icing prediction that is performed within milliseconds. The novelty of this research lies in utilizing network-based deep transfer learning with unfrozen backbones and learning schedulers. Results showed that the MobileNetV2 obtained up to 99% accuracy. The method outperformed previous research on ice detection by 3% and was evaluated in two different data sets, including near and far views of rotor blades, variety of ice densities and day phases.KeywordsWind energyIce detectionDeep transfer learning