Objective For predictive maintenance of bearings in urban rail transit vehicles, existing fault diagnosis models exhibit limitations. On the one hand, constrained by specific operating conditions and single data distribution, the models’ performance degrades significantly when transferred across datasets. On the other hand, fault diagnosis datasets are generally small in scale and highly heterogeneous, with substantial differences in sampling frequency, channel configurations, and operating conditions, which poses challenges to the construction of universal model architectures and stable learning. To achieve early identification and reliable cross-scenario generalization, it is necessary to study fault diagnosis and prognostic models.

Method The research progress on fault diagnosis and time-series foundation models is reviewed, and the key problems to be addressed are clarified. Heterogeneous data, as well as the pre-training and few-shot fine-tuning processes are symbolically modeled, and a unified learning objective and training procedure are formulated. An overall framework for a fault diagnosis and prognostic foundation model is proposed, within which data preprocessing and data coordination modules are organized to support joint pre-training of the foundation model and downstream adaptation. Finally, according to the experimental setup and comparative evaluation system, a pre-training is completed on 10 bearing datasets, and the few-shot evaluations are performed on the XJTU-SY, UO, and KAIST datasets, with multiple baseline methods for comparison and training efficiency statistics.

Result & Conclusion  Through pre-training on multiple fault diagnosis datasets comprising a total of 4 billion data points, the proposed model achieves stable improvements under few-shot fine-tuning conditions and attains optimal or near-optimal performance across multiple dataset evaluations, verifying its effectiveness in cross-scenario fault diagnosis tasks.