DiffRUL
Diffusion-based data augmentation for remaining-useful-life estimation of aero-engines and bearings.
Run-to-failure data is scarce and heavily imbalanced, which limits remaining-useful-life (RUL) models. Following Wang et al. (RESS 2024), a denoising diffusion model generates realistic degradation sequences to augment training data. Building on the authors’ official code, I adapted the pipeline to two datasets. Research assistantship at Sharif University’s Center for Information Systems and Data Science, supervised by Dr. Babak Khalaj and Dr. Mohammad Hossein Rohban.
- Aero-engines (NASA C-MAPSS). Diffusion augmentation feeding an LSTM remaining-useful-life model.
- Bearings (XJTU-SY). The same augmentation on rolling-bearing vibration data, with CNN, recurrent and hybrid CNN–LSTM predictors.