Shanghai Jiao Tong University
National Natural Science Foundation of China (61773260)
The icing detection of the wind turbine blade has very important practical significance for the safety, reliability and economy of wind turbines. Aimed at the problem of imbalanced and single-time-point non-sequentiality of wind turbine operating observation data, a method is proposed based on the oversampling and the time-dimensional upsampling convolutional neural network model. First, the Adaptive Synthetic algorithm is applied to original dataset to achieve the balance of the imbalanced dataset. Then, a time-dimensional upsampling convolutional neural network model is proposed and constructed. On one hand, the model can reconstruct and upsample the original single-time-point vector data into the two-dimensional grid data. On the other hand, it can automatically map the data into a sparse feature representation, to achieve an accurate icing detection of the wind turbine blade. Finally, the method is verified on a dataset collected from a real wind farm, and the experimental results show that the proposed icing detection method of the wind turbine blade is effective, stable, and feasible when the dataset is imbalanced and limited.