Abstract: To address the shortcomings in time-series analysis and data imbalance research for operational monitoring data of wind turbine blades, this paper proposes a fault detection method for blade icing based on feature engineering and a HO-CNN-BiLSTM-Transformer framework. Firstly, feature engineering is employed by using the blade icing mechanism model to construct mechanism variables for blade icing. Secondly, a CNN-BiLSTM-Transformer detection model is developed to explore the temporal information in supervisory control and data acquisition (SCADA) system data. Finally, the Hippopotamus optimization (HO) algorithm is utilized to optimize the model’s hyperparameters, enhancing its diagnostic performance and generalization ability. Experimental results demonstrate that this detection method achieves precision, recall, and F₁ scores of 0.983 8, 0.990 2, and 0.987 0, respectively, outperforming other comparative models and optimization algorithms. This method provides valuable insights for optimizing maintenance strategies in wind farms, ensuring safe and efficient operation of wind turbines under cold conditions.

Key words: wind turbines, blade icing, fault diagnosis, feature engineering, imbalanced data, hippopotamus optimization algorithm, deep learning