An imbalanced semi-supervised wind turbine blade icing detection method based on contrastive learning

Wind power has emerged as a crucial renewable energy source, experiencing significant growth in recent years. However, blade icing remains a pressing challenge in the operation of wind turbines, potentially resulting in systems faults and component damage. Traditional approaches to blade icing detection often rely on domain expertise, incurring additional costs. While data-driven techniques have proven effective in detecting blade icing, they require substantial amounts of labeled data for model training, which can be time-consuming and prohibitively expensive. Furthermore, blade icing detection data is often highly imbalanced since wind turbines typically operate under normal conditions for extended periods. To address these issues, we propose a novel method based on unified imbalanced semi-supervised contrastive learning (UISSCL) that can simultaneously address class imbalance scenarios and semi-supervised scenarios. UISSCL integrates unsupervised and supervised contrastive learning into a unified framework capable of extracting discriminative features from both labeled and unlabeled imbalanced data. A linear classifier is then trained based on the representations learned from the contrastive learning approach. The results obtained from computational experiments on two wind turbine blade icing datasets demonstrate that our method outperforms state-of-the-art methods in both the supervised and semi-supervised settings integrating with class imbalance scenarios.