Spectral Aliasing Pretext: Self-Supervised Fault Diagnosis for Rotating Machinery
A novel self-supervised learning technique known as Spectral Aliasing Pretext (SAP) has been introduced for diagnosing faults in rotating machinery. This method, outlined in a paper on arXiv (2608.05705), tackles the issue of scarce labeled data in industrial environments by pretraining models using unlabeled vibration data. By intentionally undersampling signals, SAP generates a folded spectrum and subsequently trains a Transformer to reconstruct the original unfolded spectrum. This approach compels the model to grasp frequency-domain invariants linked to mechanical faults without relying on potentially harmful augmentations. Tests conducted on the CWRU dataset reveal that SAP develops stable and highly discriminative representations, achieving impressive classification performance with minimal labeled data and low variance, unlike full fine-tuning methods.
Key facts
- Spectral Aliasing Pretext (SAP) is a self-supervised learning method for fault diagnosis in rotating machinery.
- The method pretrains models on unlabeled vibration data by exploiting spectral aliasing.
- Signals are deliberately undersampled to create a folded spectrum, and a Transformer is trained to reconstruct the original unfolded spectrum.
- The pretext task forces the model to learn frequency-domain invariants characteristic of mechanical faults.
- Experiments were conducted on the CWRU dataset.
- SAP learns stable and highly discriminative representations.
- In a linear probing setting, SAP achieves high classification performance with only a small fraction of labeled data and low variance.
- Full fine-tuning, including fully supervised training, does not lead to more stable or better results.
Entities
Institutions
- arXiv