Fault diagnosis of wind turbine gearbox under limited labeled data through temporal predictive and similarity contrast learning embedded with self-attention mechanism

Data-driven models for wind turbine (WT) gearbox health monitoring have garnered significant attention. However, these models usually depend on extensive manually labeled data, which is costly and time-consuming to obtain at industrial sites. The model performance is also heavily affected by volatile operating conditions of WT gearbox. To overcome these issues, this paper proposes a self-supervised fault diagnosis method based on temporal predictive and similarity contrast learning (TPSCL) embedded with self-attention mechanism. This method aims to extract latent fault features from unlabeled vibration signals, thereby improving diagnostic performance under limited labeled data and volatile working conditions. Firstly, a data augmentation combination strategy is proposed to improve the variety of input data and enhance the generalization. Furthermore, a temporal predictive and similarity contrastive learning model embedded with self-attention mechanism is proposed to extract representations related to WT gearbox health conditions. In this case, the intrinsic characteristics of unlabeled vibration signals are learned, and the operating environment disturbance can be eliminated. Finally, the fault separability is realized by fine-tuning the model with few labeled data, which can finally achieve accurate fault identification of WT gearbox. Compared with existing methods, the proposed method can fully utilize existing monitoring data and achieve better performance in WT gearbox status monitoring and diagnosis under limited labeled data and variable operating conditions, contributing to WT predictive maintenance and efficiency improvement.