AeroPulse AI
Predictive maintenance that says when not to trust the prediction
- Status
- Live demo
- Year
- 2026
- Rooms
- ML & Impact, Products & Web
Problem
A remaining-useful-life number alone is not a maintenance decision. Planners need to know whether the prediction can be trusted and what is feasible with the people and parts they actually have.
Approach
TrustTwin checks the sensor data first, gates the model's output (valid, degraded or unavailable), attaches calibrated intervals and exact SHAP explanations, then plans maintenance with OR-Tools — with an officer approving every decision.
Engineering decisions
- When data quality fails, no RUL is produced at all — a verification task is created instead.
- Split-conformal 90% intervals with coverage measured on unseen engines.
- Prediction, recommendation, decision and work order are separate records in a hash-chained audit log.
What it does
- Sensor reliability checks
- Calibrated RUL intervals
- What-if digital twin
- CP-SAT maintenance planner
Stack
My part
Built for Smart India Hackathon 2026 (SIH26249); all 17 commits in the repository are mine.
Limits, stated plainly
- Advisory only: it never authorises flight or airworthiness and has not been validated on military aircraft.
- Trained on NASA's public C-MAPSS (simulated commercial turbofans); fleet, people and parts are synthetic and labelled so.
Evidence
- AeroPulse-AI README
- docs/model-training.md
Screenshots
AeroPulse command overview with fleet status and latest predictions
TrustTwin scenario comparing plans when sensors, parts or technicians drop out
The maintenance planner's proposed schedule across technicians