dummy033485/rul-transformer-cmapss-fd001

Industrial AI / PHM Suite — MechE × AI Portfolio

A single project suite covering five research frontiers of industrial AI — PHM, Transfer Learning & Domain Adaptation, Multimodal Machine Learning, Similarity-Based (few-shot) Learning, and Industrial Physical AI — every result below measured on held-out test data with real public datasets (NASA C-MAPSS and Case Western CWRU bearings). Built by Zain (Mechanical Engineering + AI), Fall 2027 graduate applicant.

Live interactive demo (all 5 modules): https://huggingface.co/spaces/dummy033485/rul-transformer-demo

Verified results

Frontier	Experiment	Result
PHM	RUL Transformer, C-MAPSS FD001 test (100 engines)	RMSE 16.74 / MAE 12.61 / NASA 470 — Random Forest baseline 17.15 / 13.18 / 656
Transfer Learning	FD001 → FD004 (6 operating conditions, 2 fault modes), 248 test engines	Zero-shot RMSE 52.41 / MAE 44.37 / NASA 149,187 → fine-tuned RMSE 29.22 / MAE 21.46 / NASA 9,431. Training from scratch on FD004 also gives RMSE 29.22 — with ~58k target windows the target data is sufficient, so pretraining shows no edge here; the adaptation advantage appears in the low-data regime (see few-shot row)
Multimodal ML	CWRU bearings (4 classes): signal 1D-CNN / spectrogram 2D-CNN / fusion	100% / 100% / 100% in-domain (1797 rpm)
Multimodal × shift	Same models on a new load (1730 rpm), no retraining	99.5% / 85.6% / 100% — the image modality degrades under shift, fusion stays robust
Similarity-Based ML	Prototypical few-shot on spectrogram embeddings	In-domain 5-shot 99.99%; on the new load, 5-shot adaptation recovers 99.2% (from the CNN's 85.6%) using only 5 labelled target examples/class
Industrial Physical AI	Monotonicity constraint (degradation is irreversible) on FD001 RUL trajectories	Physically-impossible RUL increases: 46.7% of steps → 0%; last-window RMSE 16.74 → 16.21

Files

• rul_transformer.pt — Transformer encoder (d_model 64, 2 layers, window 30) trained on FD001. Config dict inside the checkpoint.
• scaler.json — per-feature mean/std (train split only) + feature list for FD001.
• rul_transformer_fd004_ft.pt — FD001 model fine-tuned on FD004 (same architecture/config format).
• scaler_fd004.json — FD004 scaler stats.
• cwru_models.pt — CWRU signal 1D-CNN, spectrogram 2D-CNN and fusion head state dicts + input normalisation stats.
• transfer_report.json, cwru_report.json, cwru_crossload_report.json, physical_report.json — full measured metrics for every module.

Methods (short)

• PHM: piecewise-linear RUL labels (cap 125), sliding windows of 30 cycles, engine-wise train/val split (no leakage), scaler fit on train only. Baselines: Ridge, Random Forest.
• Transfer: zero-shot = FD001 weights + FD001 normalisation applied to FD004; fine-tuned = same weights, FD004 normalisation, lr 5e-4, early stopping on engine-wise validation split. Short test engines (<30 cycles, 11 of 248) left-padded with their first observation. A from-scratch FD004 model reaches the same test RMSE (29.22), i.e. the failure being demonstrated is the zero-shot distribution shift, and recovery needs target-domain data — a little (fine-tuning / few-shot) or a lot (full training).
• Multimodal: CWRU 12 kHz drive-end signals, fault size 0.007" (files 97/105/118/130 @1797 rpm; 100/108/121/133 @1730 rpm), 1024-sample segments, contiguous 70/30 split per file. Spectrograms: 128-point FFT, log magnitude, 64×64.
• Similarity: class prototypes = mean spectrogram-CNN embedding of k support examples; query classified by nearest prototype (Euclidean); 200–300 episodes, mean ± std reported.
• Physical AI: per-engine isotonic regression (decreasing) over predicted RUL trajectories enforces the physics constraint that degradation cannot reverse.

Limitations

C-MAPSS is simulated turbofan data; CWRU experiments use a single fault size (0.007") and seeded faults; few-shot uses spectrogram embeddings only. Numbers are from the author's own runs (seed 42) and are reproducible from the Space's bundled report files.