Topological Data Analysis Applied to Wind Turbine Vibration Spectra for Blade Icing Detection

Ice buildup on wind turbine blades is a significant issue, leading to operational risks and reduced efficiency. Conventional ice detection methods, such as visual inspection, power curve analysis or specialised sensors, are often slow, inefficient, or costly. This paper proposes an approach using 0-dimensional persistence homology from topological data analysis (TDA) applied to tower and blade vibration spectra. This method extracts key features representing the lifespan of the sub-level sets of the spectra, allowing the formulation of a clearer supervised learning problem. The resulting persistence diagrams are embedded into persistence images and persistence rank functions. Persistence images are employed alongside convolutional neural networks (CNN) to distinguish asymmetrical ice distribution on one or two blades as well as symmetrical ice distribution across three blades from normal conditions. For the symmetrical ice distribution scenario, persistence rank functions with functional principal component analysis (FPCA) and support vector machines (SVM) offer a simpler classification. This approach not only improves ice detection accuracy but also reduces equipment costs and maintenance, promising enhanced wind turbine blade monitoring and maintenance efficiency.