Wind turbine blade breakage detection based on environment-adapted contrastive learning

Blade breakages in wind turbines are serious threats to reliable power generation. However, blade breakage detection is still challenging owing to varying circumstances and the questionable assumption that distinct health conditions result in inherently separable measurements. To address these challenges, a novel method is proposed for blade breakage detection based on environment-adapted contrastive learning. Specifically, nonparametric regression models between operational and environmental variables are established, and the variable adjustment is conducted to suppress the influence of variational circumstances on SCADA measurements. In addition, an improved strategy of contrastive learning is developed to learn valid data representations, guaranteeing the assumed separability of different samples. Furthermore, normal models of healthy operations are constructed, and online monitoring is implemented based on likelihood estimation and control charts. Actual data from commercial wind farms have been employed to demonstrate the effectiveness and superiority of the proposed method. Experimental results confirm that the correlation degree between operational and environmental variables can be reduced by 28.7 %, and blade breakages can be detected more than 23 h in advance.