Battery PHM — SOH Estimation, RUL Prediction with Conformal Uncertainty, Cross-Cell Transfer
Lithium-ion battery prognostics on the NASA PCoE Battery Data Set (cells B0005, B0006, B0007, B0018; discharge cycles only). Built to mirror the research lines of UConn's REIL lab (Prof. Chao Hu): battery health management, uncertainty quantification of ML models, and transfer learning for capacity estimation.
Protocol: leave-one-cell-out. Headline test cell B0018 (never seen in training). EOL = capacity below 1.4 Ah (70% of 2 Ah rated, first crossing). RUL metrics are computed on pre-EOL windows only.
Verified results (test cell B0018)
| Task | Result |
|---|---|
| SOH estimation from discharge-curve features (no capacity input) | RF RMSE 0.070 Ah · MLP 0.188 Ah (all 4 rotations: RF 0.029–0.093 Ah) |
| RUL, 20-cycle windows, zero-shot | Transformer ensemble (×5) RMSE 14.96 / MAE 12.97 · RF RMSE 17.46 (late-life RUL≤40: RF 13.66, ensemble 18.51) |
| Uncertainty (90% intervals) | Raw ensemble-spread intervals cover only 9% of truths → split-conformal: RF 89.7% coverage (width 57.8 cycles), ensemble 100% (width 76.4, conservative) |
| Transfer, adapt on first 30 cycles | RF refit 16.89 → 12.60 · ensemble head-only FT does not help (15.03 → 16.85) — the degradation knee is not yet visible |
| Transfer, adapt on first 60 cycles | Ensemble full FT 19.30 → 8.83 (−54%) on the near-EOL segment · RF refit plateaus (13.03 → 13.22) · RF conformal coverage recovers to 91.9%; ensemble in-sample calibration still undercovers (48.6%) |
Files
rul_ensemble.pt— 5-member Transformer ensemble (zero-shot) + normalization stats + conformal qrul_ensemble_ft60.pt— ensemble fine-tuned on B0018's first 60 cyclessoh_models.pkl— RF + MLP SOH estimators with scalerbattery_report.json— full measured results and protocol
Live demo
https://huggingface.co/spaces/dummy033485/battery-phm-demo
Limitations
Four cells from one NASA campaign at room temperature; impedance cycles unused; fine-tuned ensemble intervals are calibrated on the adaptation windows themselves (optimistic); conformal guarantees are marginal, not conditional on life stage — coverage degrades when calibration windows and prediction windows come from different life stages.
By Zain — Mechanical Engineering + AI, Fall 2027 PhD applicant.