Wind turbine blade icing diagnosis using RFECV-TSVM pseudo-sample processing

Abstract

Wind turbine blade icing seriously affects turbine power generation and fatigue life, and an accurate diagnosis of blade icing is beneficial for wind turbines to make in-time adjustments. However, the high dimensional and unbalanced original data recorded by Supervisory Control and Data Acquisition (SCADA) systems pose great challenges to the accurate diagnosis of blade icing. To effectively address the challenges of difficult feature extraction and small number of fault samples, we propose a data processing method based on pseudo-sample processing. Specifically, Recursive Feature Elimination and Cross-Validation (RFECV) is used to analyze the influence of various SCADA features on the diagnostic model and select the most compelling feature set. A Transductive Support Vector Machine (TSVM) is implemented to regenerate unlabelled samples. The labeled pseudo samples and ice data are combined to form the training set. The effectiveness of the proposed method is examined using the three most commonly used classifier algorithms, i.e., Random Forest (RF), Support Vector Machine (SVM), and XGBoost, for four utility-scale wind turbines. The results show that this method can effectively obtain the optimal selection, utilize unlabelled samples, and improve the diagnostic accuracy of the model, especially for small sample data, with an average accuracy improvement of 10.06%.