Semi-Supervised Domain Adaptation with Feature Separation for Wind Turbine Anomaly Detection

Anomaly detection for the wind power generation process is important for the maintenance and early warning of wind turbines. Deep learning-based methods are widely used in wind turbine anomaly detection. However, the data distribution changes frequently due to the changing working conditions of wind turbines, and deep learning methods are difficult to directly apply to new working conditions with scarce labeled data. Therefore, this paper proposes a semi-supervised domain adaptation method with feature separation (SSDA-FS) for wind turbine anomaly detection. First, a feature separation domain adaptation strategy is designed to eliminate inter-domain differences. The feature space is decomposed into domain-related space and anomaly-related space through the devised adversarial training. Then a semi-supervised timing augment strategy is proposed, which can augment the unlabeled samples and a few labeled samples in the target domain. The proposed method is implemented and validated in a wind turbine anomaly detection dataset, and the performance is significantly improved compared with the comparison methods.