A robotic vision system for inspection of soiling at CSP plants
Joe Coventry, Charles-Alexis Asselineau, Ehab Salahat, Md Arifur Raman, Robert Mahony
Abstract
This paper describes efforts to develop a method of determining soiling levels on heliostat mirrors from images taken from aerial drones. Lessons learned from a flight mission relate to drone positioning errors due to heat and GPS drift, and the challenges of handling lots of data. A method to determine waypoints for missions is suggested, to gather data in a random manner with a known confidence level. Methods of interpreting the images taken are described, including a processing pipeline for identification and segmentation of individual mirror facets. Finally, the paper focusses on efforts to develop and validate a method of determining reflectance from visual images, including development of a sky fitting model, examination of the impact of image resolution, and presentation of the estimated reflectance compared to reference results taken with a hand-held reflectometer. The initial results show promise, currently with accuracy within around 2-3% for soiling cases most typical in a commercial heliostat field, however further improvements is required to achieve the target accuracy of <1%.