metadata
license: mit
tags:
- predictive-maintenance
- cmapss
- remaining-useful-life
- graph-neural-networks
- jepa
- anomaly-detection
library_name: pytorch
Diagnostic Graph PDM — Best C-MAPSS Checkpoints
Best observed checkpoints from the Diagnostic-Graph-PDM stack: Graph-JEPA pretrain → Energy Critic → TCN+Transformer RUL with energy fusion.
Files
| File | Role | Selection criterion |
|---|---|---|
checkpoints/jepa/jepa_epoch0003.pt |
Graph-JEPA encoder (FD123 + unlabeled FD004) | Epoch 3 val loss 0.0025 |
checkpoints/critic/vulcan_epoch0030.pt |
Energy Critic + encoder (JEPA-init) | Best saved FD1234 critic (~0.97 val F1) |
checkpoints/rul/rul_sequence.pt |
TCN+Transformer RUL (FD001) | Best val RMSE epoch 54 — test RMSE 12.485, NASA 294.6 |
See checkpoints.json for paths and metrics.
Usage
pip install torch typer omegaconf rich
git clone https://github.com/Sph3inz/Diagnostic-Graph-PDM.git
cd Diagnostic-Graph-PDM
pip install -e .
hf download Sph3inxz/Diagnostic-Graph-PDM --local-dir ./hf_ckpt
# Scan / localize (critic)
python -m vulcan scan data/cmapss_fd001.graphs.jsonl `
--checkpoint hf_ckpt/checkpoints/critic/vulcan_epoch0030.pt --format summary
# RUL eval
python eval_ckpt.py hf_ckpt/checkpoints/rul/rul_sequence.pt
Train critic from JEPA:
python -m vulcan train --config configs/critic_cmapss_fd1234_jepa.yaml `
--resume-from hf_ckpt/checkpoints/jepa/jepa_epoch0003.pt
Citation
If you use these weights, link the GitHub repo: https://github.com/Sph3inz/Diagnostic-Graph-PDM