803 150 Detection of Wind Turbine Icing J. A. L. McGowan DTU Compute, Technical University of Denmark Vattenfall Wind Power I NTRODUCTION Wind turbine icing is a phenomenon where snow and ice accretes on a wind turbine’s wings. This effect poses several issues for wind farms in cold climates. These include the potential safety hazard from falling ice, increased wing fatigue/wear due to increased wing weight, and decreased power production due to changes in the aerodynamic characteristics of the wings as well as increased down time. The research fields of de- and anti-icing focus on removing and avoiding ice accumulation respectively. In order for any such system to function efficiently it must be aware of the current state of icing. I.e. it should only be active when icing is occurring/has occurred. For this reason a robust ice detection method is of great value. Many different ice detection systems are currently available, but all with known issues making them close to useless as standalone systems. Furthermore to the knowledge of the author no system exists which detects ice based on image analysis. Using image analysis for ice detection is of interest for various reasons, one being the high quality of off-the-shelf cameras which could translate into sensor solutions which are far cheaper than custom built sensor arrays. M ETHODS A variety of image analysis and machine learning techniques are used to detect ice on the unsorted images provided for the project by Vattenfall. The main parts are: 1. Wing detection: Only about half of the images show any part of the wing. The images are therefore first sorted in wing/no wing using Basic Image Features(BIFs), and a custom clustering algorithm developed under the project. 2. Image segmentation: The images must then be segmented into wing/background segments in order to perform closer analysis on wing parts. This is done using manually extracted wing shape templates and background subtraction. 3. The final classification is performed using clustered BIF histograms from a manually annotated training data set. RESULTS AND CONCLUSION The final results have not been analyzed fully at time of writing but indicate classification accuracies for each of the three steps between 70 and 90% depending on the conditions under which the images were captured. Due to the extreme diversity and strong noise sources in the images in the data set this accuracy is taken as a proof of concept for image based ice detection systems which can function robustly in cold climates, and thereby help to increase the geographical areas in which wind farms are a plausible sustainable energy source. ENERGY FROM WIND, SUN AND WATER POSTER IDEA MASTER THESIS
Grøn Dyst 2016
To see the actual publication please follow the link above