Early-warning system for wind turbine faults: Improving its real-time performance using cyclic DBSMOTE and deep-learning algorithms

Wind turbine fault warnings can significantly improve the efficiency and management of wind farm operations and maintenance services. However, the volatility, imbalance, and low-value density of wind turbine operation data make accurate fault warnings challenging. In this study, a data-extraction and balancing method comprising a sliding window and cyclic DBSMOTE was developed to effectively improve the imbalance of datasets. Next, a hybrid model for wind turbine fault early warnings based on a deep-learning algorithm was established. Traditional convolutional neural and long short-term memory networks were used to extract spatial and temporal features from the balanced dataset, respectively, to mine deep information. A multihead self-attention mechanism was introduced to improve the early-warning model training and learning. Based on the test dataset obtained from the same wind farm and time series, the proposed fault early-warning model achieved early warning abilities up to 32 h through optimization of key hyperparameters: the number of convolution modules (2), LSTM layers and hidden units (2 layers with 512 and 2 units), and classification weight ratio (0.3:0.7), with key evaluation metrics at 32 h reaching 0.807 (MCC), 0.862 (F1-score), and 0.887 (G-means). The proposed fault early-warning model was verified through ablation, comparison, and cross-verification. Compared with CNN-LSTM, the model improved MCC by 11.2%, F1-score by 7.1%, and G-means by 5.5%; compared with SVM, it increased MCC by 33.8%, F1-score by 19.0%, and G-means by 14.7%. Ten-fold cross-validation showed high stability, with maximum differences in MCC, F1-score, and G-means of only 0.75%, 0.82%, and 0.45%, respectively. Therefore, the developed early-warning system demonstrated excellent fault early-warning performance in actual wind farm applications.