TFCPrediction: A cross-domain self-supervised learning framework for bearing fault diagnosis in wind turbines with limited annotations

Abstract The scarcity of labeled bearing fault data and the prohibitive cost of manual annotation pose significant challenges for intelligent diagnosis in wind farms. To overcome these limitations, we present TFCPrediction, an integrated self-supervised learning framework. The framework extends existing time-frequency SSL approaches through cross-domain prediction, reducing dependence on manual labeling by mining deep features from unlabeled vibration signals. First, we design a time-frequency data augmentation method to enhance the model’s ability to extract latent features from raw signals. Second, we introduce a cross-domain sequence prediction task to capture temporal dependencies and learn holistic time-frequency representations. Finally, we evaluate the framework on three datasets, including data from a real wind farm. Experimental results demonstrate superior performance when fine-tuned with only 1% of labeled data, showing strong potential for industrial applications with limited annotations.