Self-Supervised Multiple Faults Detection Method Based on Time–Frequency Feature Fusion With Unlabeled Wind Turbine Samples

Fault detection is an essential aspect of power generation in wind turbines (WTs). However, existing fault detection methods are developed specifically for identifying a single type of fault and rely on a sufficient amount of labeled data. These approaches tend to be less accurate when used to detect many types of faults with limited data. To solve those problems, this article proposes a self-supervised fault detection method based on a time-frequency feature fusion module (TF-FFM). First, a periodicity-based hybrid data augmentation is presented in order to expand the number and diversity of fault samples. Second, TF-FFM can fully extract the features of fault in both time and frequency domains. Third, a fault detection method based on self-supervised learning is proposed during the training process to reduce the cost of data collection and labeling. Meanwhile, the loss function is optimized based on the energy conservation theorem in the time-frequency domain. This optimization leads to an advanced accuracy of the fault detection method for WTs by establishing a power consistency relationship. Finally, this article evaluates the effectiveness of the proposed method through a comparative analysis with various fault diagnosis techniques, feature visualization, and ablation experiments. The accuracy of the proposed method achieves 95% in the context of multiple fault detection, which is 3% higher than the results of existing methods.