TF-ProFM: Prototype-Based Time–Frequency Foundation Model for Generalizable Industrial Bearing Fault Diagnosis

The development of generalizable foundation models for intelligent bearing fault diagnosis is critical for modern industrial systems, yet it is severely hindered by the challenge of handling multisource, heterogeneous data. Existing methods often struggle to adapt to varying operating conditions and diverse signal characteristics. To address this issue, we propose a novel prototype-based time–frequency foundation model (TF-ProFM) that leverages self-supervised pretraining and few-shot fine-tuning. TF-ProFM features a time–frequency embedding module that uses random patching and local fast Fourier transform to process diverse signals into unified spectral representations. For pretraining, it employs a prototype-based contrastive learning objective, which explicitly encourages the model to learn a feature space with a clear cluster structure that aligns with downstream fault classification tasks. Extensive experiments on nine public datasets validate the superiority of our approach. In rigorous cross-dataset evaluations, TF-ProFM achieves the state-of-the-art fault diagnosis performance, outperforming existing methods even when fine-tuned with 1% of the labeled data from the target domain.