Bearing-DiffRUL

Diffusion-based remaining-useful-life estimation for rolling bearings

Sharif Center for Information Systems and Data Science — 2025 Supervised by Dr. Babak Khalaj and Dr. Mohammad Hossein Rohban

A companion to DiffRUL-CMAPSS that carries the diffusion-based augmentation idea from aero-engines to rolling-bearing degradation on the XJTU-SY run-to-failure dataset. Vibration signals are noisy, and healthy-vs-failing samples are heavily imbalanced — a natural fit for generative augmentation.

Highlights

  • Denoising Diffusion Probabilistic Model (DDPM) to synthesize realistic bearing-vibration sequences
  • Augmented training data feeding sequence models for remaining-useful-life (RUL) prediction
  • Evaluated on the XJTU-SY accelerated-life-test benchmark