Self-Supervised Learning Boosts Machinery Fault Diagnosis

Deep learning offers a powerful approach for machinery fault diagnosis, but its effectiveness is often hampered by the scarcity of labeled data in industrial environments. To overcome this, a new self-supervised learning method called Spectral Aliasing Pretext (SAP) has been introduced. SAP pretrains models using readily available unlabeled vibration data by ingeniously exploiting the phenomenon of spectral aliasing. The core idea behind SAP involves deliberately undersampling vibration signals to create a "folded spectrum." A Transformer model is then trained to reconstruct the original, unfolded spectrum from this aliased version. This pretext task compels the model to learn frequency-domain invariants that are characteristic of mechanical faults, without relying on potentially destructive data augmentations. Experiments conducted on the CWRU dataset demonstrated that SAP effectively learns stable and highly discriminative representations. When used in a linear probing setting, SAP achieved very high classification performance with only a small fraction of labeled data and exhibited low variance. Notably, full fine-tuning, including fully supervised training, did not yield more stable or superior results, suggesting that SAP combined with linear probing can be a more effective and reliable strategy for fault diagnosis when labeled data is limited.