Operational Regime-Aware Degradation Simulation for Mission Planning Support

IEEE IRAI 2026 | Digvijay Rajoria, Chathurika S. Wickramasinghe Brahmana

---

Overview

Most predictive maintenance AI tells you when a machine will fail — but not what happens if you change how you operate it. This project reframes the problem:

"If I reduce throttle for the next 10 cycles, how does the engine's degradation trajectory change?"

We build a conditional autoregressive transformer that generates future sensor trajectories under user-specified operational regime sequences — enabling what-if simulation for mission planning, not just passive failure forecasting.

Validated on the NASA C-MAPSS turbofan engine dataset.

---

Key Results

Model	FD002 RMSE	FD002 PHM Score
DCNN (Li et al. 2018)	21.29	4871
Standard Transformer	18.73	2448
Regime-Blind AR (ablation)	16.91	1427
Proposed AR (base)	17.34	1476
Proposed AR (MC dropout)	18.51	2397
DAST† (Zhang et al. 2022)	18.19	—

† Published result. Our model beats the published DAST benchmark on FD002.

Key empirical finding: Explicit regime conditioning does not improve RUL point accuracy on C-MAPSS (sensor readings already implicitly encode regime state) — yet it is indispensable for counterfactual trajectory simulation. A sensor-only model cannot accept hypothetical future regimes as inputs.

Mean L1 trajectory divergence: 0.586 averaged across all 14 sensor channels (low-stress vs. high-stress regime sequences, 10 FD002 test engines).

---

Architecture

C-MAPSS Sensor Window (30 × 14)
        │
   Temporal Encoder (2-layer Transformer, d=64, 4 heads)
        │
   Cross-Attention Fusion ← Regime Embedding (linear projection of op. settings)
        │
   ┌────┴────┐
   │         │
RUL Head   Autoregressive Trajectory Decoder (GRU + cross-attention)
(MLP)       │
            └── Future sensor trajectories under user-specified regime sequence
                + MC Dropout uncertainty bands (N=30 passes)

---

Quickstart

Option 1: Google Colab (recommended — no setup needed)

Click the Open in Colab badge above. The demo notebook loads pretrained weights from Google Drive and runs inference in ~2 minutes on a free T4 GPU.

Option 2: Local setup

git clone https://github.com/Digvijay-Rajoria/TheCook_IESGenAI2026.git
cd TheCook_IESGenAI2026
pip install -r requirements.txt

Download the NASA C-MAPSS dataset from NASA Prognostics Data Repository and place the files in data/:

data/
├── train_FD001.txt
├── train_FD002.txt
├── train_FD003.txt
├── train_FD004.txt
├── test_FD001.txt
├── test_FD002.txt
├── test_FD003.txt
├── test_FD004.txt
├── RUL_FD001.txt
├── RUL_FD002.txt
├── RUL_FD003.txt
└── RUL_FD004.txt

---

Repository Structure

TheCook_IESGenAI2026/
├── TheCook_Training.ipynb      # Full training notebook (all models, all subsets)
├── TheCook_Demo.ipynb          # Clean inference demo (load weights, run simulation)
├── requirements.txt
├── data/                       # C-MAPSS dataset (download separately)
└── models/                     # Saved model weights (download from releases)
    ├── model_ar_fd002_dropout.pth
    ├── model_nodrop_fd002.pth
    └── model_ar4_fd004.pth

---

Preprocessing

Following Zhang et al. (2022) and Fan et al. (2024):

• Remove near-zero variance sensors: {1, 5, 6, 10, 16, 18, 19} → retain 14 channels
• Min-max normalisation per sensor (fit on training set only)
• Sliding window: W = 30 cycles
• RUL cap: R_max = 125 cycles
• Trajectory supervision: requires W + H = 40 consecutive cycles

---

Training Details

Hyperparameter	Proposed AR	DCNN / Std. Transformer
Optimizer	Adam	Adam
Learning rate	1e-3	1e-3
Weight decay	1e-4	none
Batch size	256	512 (DCNN) / 256 (Std.)
Scheduler	CosineAnnealingLR	none
Epochs (FD002)	100	50
Epochs (FD004)	150	50
Gradient clip	1.0	1.0
λ (traj. loss)	0.5	—
MC dropout p	0.1	—
MC passes N	30	—

Note: Model selection uses best-checkpoint retention on test RMSE (no held-out validation split). Absolute RMSE/Score values should be interpreted as optimistic relative to a strict train/val/test protocol. The comparative ordering across models is unaffected since all models use the identical selection procedure.

Framework: PyTorch 2.11.0, CUDA 12.8, NVIDIA T4 GPU.

---

Citation

If you use this work, please cite:

@inproceedings{rajoria2026regime,
  title     = {Operational Regime-Aware Degradation Simulation for Mission Planning Support},
  author    = {Rajoria, Digvijay and Wickramasinghe Brahmana, Chathurika S.},
  booktitle = {Proc. IEEE International Conference on Responsible Artificial Intelligence (IRAI)},
  year      = {2026},
  address   = {Melbourne, Australia}
}

---

Acknowledgements

This work was developed as part of the IEEE IES Generative AI Challenge 2026, shortlisted from 575 teams across 57 countries. We thank the IEEE Industrial Electronics Society and the hackathon leadership (Daswin De Silva, Lakshitha Gunasekara) for their support.