A self-supervised learning method for fault detection of wind turbines

As promising solutions to condition-based maintenance of wind turbines, artificial intelligence-based techniques have drawn extensive attention in the era of industry 4.0. However, accurate fault detection is still challenging owing to volatile operating conditions in real-world settings. To handle this problem, a novel method is proposed for fault detection of wind turbines. Specifically, a data augmentation scheme is developed to simulate the effects of time-varying environments and noise. Then, a self-supervised proxy task of variant prediction is designed and conducted. In this way, valid data representations can be extracted to represent the health status of wind turbines. Additionally, the compactness of data representations is guaranteed by the directional evolution, which can relieve the confusion of health conditions. The effectiveness of the proposed method is verified with actual measurements. Using the proposed method, several faults can be detected more than 10 days earlier, and blade breakage can be identified more than 22 hours earlier. Furthermore, the developed method outperforms several benchmark approaches.