Abstract: At present, the wind turbine blades are usually de-iced using a heating system after ice coating, and the energy consumption is about 1%~4% of the annual power generation of the generator set. In this paper, the finite element model is adjusted by experimental modal results, and the equations of the corresponding relationship between the thickness of the ice coating and the natural frequency at different positions are obtained by the simulation modes of the finite element model. These equations are used to generate random samples to train the BP neural network model. The frequency is the input, and the thickness of the ice is the nonlinear relationship of the output to realize the detection of the ice status. The research shows that through the results of the modal test of the blade′s force hammer excitation, the parameters of the blade model are adjusted, and the error between the first three-order natural frequency and the test value of the optimized three-dimensional model of the bladed iced hollow solid is within 2%. Through BP neural network modeling and training, the average error rate of the results of the model detection ice thickness and the actual value is 8.83%, the error at the blade tip is the smallest, and the error at the blade root is the largest, the relative error rate decreased with the increase of the ice thickness. The trained BP neural network model can basically realize the detection of ice coating position and thickness information, and provide a theoretical basis for the precise heating position and heating time of the heating system and reducing energy consumption.

Key words: wind turbine blade, icing detection, modal parameters, natural frequency, composites