ARCHIVED — Experimental project, not maintained.
Remaining useful life prediction on NASA C-MAPSS turbofan data with adversarial input validation for sensor integrity.
| Metric | Value |
|---|---|
| Dataset | NASA C-MAPSS FD001–FD004 |
| Input features | 21 sensor channels + 3 operational settings |
| Output | RUL prediction (cycles remaining) |
| Evaluation | RMSE + NASA scoring function |
| Adversarial check | Sensor-input perturbation bounds |
| Training samples | ~20,000 degradation trajectories |
| Status | Archived / Experimental |
┌──────────────────┐ ┌──────────────────┐ ┌─────────────────┐
│ Sensor Streams │────▶│ Input Validator │────▶│ RUL Predictor │
│ 21 channels │ │ Adversarial check│ │ Neural network │
└──────────────────┘ └──────────────────┘ └─────────────────┘
│ │ │
▼ ▼ ▼
C-MAPSS time series Reject out-of-bound RMSE + score
(multivariate) sensor readings benchmarked
Data Pipeline:
- Load C-MAPSS degradation trajectories (run-to-failure sequences)
- Normalize sensor channels per operational condition
- Apply sliding window to create fixed-length input sequences
- Validate inputs against learned sensor operational envelopes
- Predict RUL via neural network
- Evaluate against published benchmark metrics (RMSE, scoring function)
Datasets:
| Subset | Engines (train) | Engines (test) | Fault modes | Operating conditions |
|---|---|---|---|---|
| FD001 | 100 | 100 | 1 (HPC) | 1 |
| FD002 | 260 | 259 | 1 (HPC) | 6 |
| FD003 | 100 | 100 | 2 | 1 |
| FD004 | 249 | 248 | 2 | 6 |
The input validator checks sensor readings against physical plausibility bounds before inference. This addresses a practical deployment concern in safety-critical ML systems:
- Sensor spoofing — adversarial manipulation of input channels to trigger incorrect predictions
- Faulty sensors — hardware degradation producing out-of-range values
- Distribution shift — operational conditions outside the training envelope
The validator learns per-channel operational envelopes from training data and rejects inputs that violate these bounds, flagging potential adversarial manipulation or sensor failure before the prediction reaches downstream maintenance scheduling.
git clone https://github.com/poojakira/PulseNet-RUL-Forecasting.git && cd PulseNet-RUL-Forecasting
pip install -r requirements.txt
# Download NASA C-MAPSS data
python scripts/download_data.py
# Train and evaluate on FD001
python main.py --dataset FD001
# Run with adversarial input validation enabled
python main.py --dataset FD001 --validate-inputsSafety-critical ML systems (predictive maintenance, autonomous vehicles, medical devices) face a threat model where adversarial inputs have physical consequences. A manipulated RUL prediction could defer maintenance on failing equipment or trigger unnecessary shutdowns.
Input validation against operational envelopes is a minimal-cost defense that catches both adversarial perturbations and sensor faults. This project applies AI security thinking — input validation, bounds checking, anomaly rejection — to a domain where model failure has real-world safety implications beyond information security.
The approach demonstrates that AI security extends beyond LLMs and classifiers into any ML system where predictions drive physical-world decisions.
MIT