Bearing Fault Classification with Vision Transformers
ViT-based condition monitoring pipeline for fine-grained bearing fault diagnosis from vibration signals.
This project turns 1D vibration signals from rotating machinery into rich 2D time–frequency representations and feeds them into a Vision Transformer-based classifier for fine-grained bearing fault diagnosis.
Approach
- Converted raw vibration traces into spectrogram-style images so ViT could exploit global spatial attention.
- Trained and tuned a ViT backbone to separate 13 bearing defect classes, including subtle fault modes that look similar in the raw signal.
Results
- Reached 98.8% accuracy across 13 classes, demonstrating that transformer-style global context significantly improves reliability over conventional CNN baselines.
- Stress-tested the model under noise and domain shifts to understand when predictions remain trustworthy for real-world predictive maintenance pipelines.