Posters

Abstract

With the continuously increasing blade size of modern wind turbines, erosion protection requires new and stronger materials. The main method for qualification and validation of new erosion protection materials for use in the wind industry is accelerated rain erosion testing (RET). The output of these tests are large pools of images that require manual assessment to build the V/N curves for identification of when erosion occurs. Currently, the assessor does not capture all defects visible on the RET images due to time limitations. Instead, only erosion front (the furthest propagation of damage towards the low speed area of the sample) is recorded. The REQUIM project aims to use machine learning to detect defects on the images from the rain erosion tester, based on a structured data flow. This will optimise processing speed of the tests and allow users to create a more in-depth analysis of the test data. We will be able to gather more information about different failure mechanisms occurring on the test sample. This will enable us to improve the Leading Edge Protection (LEP) product in the design phase, as well as maintain the material better once installed on the blades. Another benefit will be the increased testing speed and lower time for certification and launch to market. Image recognition through machine learning requires a large set of training data to be efficient. Moreover, structured defect categories shall be defined for grouping the findings. The project group has developed a dedicated defect catalogue for findings on rain erosion tests. It describes damages by their type and affected layer. In order to make the process for creating training data faster and more efficient, we developed a user friendly annotation tool that significantly reduced RET assessment times for the assessing specialist. The most significant breakthrough of the project is taking advantage of the data structure from an RET. The images position is constant and the only variable is time. This allows us to compare the same area of the test blade in different image capture time. In this way we can detect a defect in an advanced stage through computer vision techniques and highlight the area of the blade, where more advanced machine learning algorithms can be utilised to detect the defect in its initial stage. That strategy can create a large pool of training data without the need for a specialist annotation. The learnings from the project can significantly reduce RET annotation times. Furthermore, they can direct our attention the improving inspection methodology of real turbines to enable efficient machine learning for defect detection on blades in operation.