VLM-Assisted Prognostics — Turbofan RUL on NASA C-MAPSS
- Vision-Language Models
- Deep Learning
- PHM
Built a two-branch prognostic framework that pairs eight numerical remaining-useful-life (RUL) models (XGBoost, BiLSTM, TCN, MS-TCN, Transformer and more) with Qwen2.5-VL-7B, which reads full-history plots of all 21 engine sensors and reports degradation trend, regime, a qualitative life band and figure legibility — or abstains when a plot is unreadable. A deterministic, bounded fusion rule keeps the final RUL under numerical control.
Evaluated on all four C-MAPSS subsets (707 test engines) across three random seeds. Visual evidence did not improve RUL accuracy, but it exposed a strong reliability signal: the VLM judged 13–20% of figures unreadable under a single operating condition versus 58–60% under six — pointing to VLMs as a qualitative screening and evidence-quality layer rather than a numerical predictor.
Status: Submitted for publication — Aerospace Science and Technology (Elsevier).
- Python
- PyTorch
- Qwen2.5-VL
- XGBoost
- Transformers
- Predictive Maintenance






