FaultFormer: Pretraining Transformers for Adaptable Bearing Fault Classification
The study proposes a self-supervised pretraining and fine-tuning framework using transformer models to classify bearing faults. By masking tokens in vibration signals and applying various tokenization and data augmentation strategies, the model learns robust representations from unlabeled data. This pretraining significantly improves performance on scarce, unseen training samples and enables rapid adaptation to new fault classes or datasets.