aero-rul

Predict how many flights a turbofan engine has left, with an interval whose coverage is measured on engines the model has never seen.

aero-rul works on NASA's N-CMAPSS dataset, which simulates full flights (climb, cruise, descent) of a fleet of engines until failure. It is a small, tested Python package with one command per step, and every number it reports can be reproduced from a config file.

What it does

Step	Command	Output
Read the raw 1 Hz recordings, one row per flight	aerorul prepare	cycles.parquet
Ridge and LightGBM baselines, leave-one-engine-out CV	aerorul evaluate	baselines.md
Prediction intervals and their measured coverage	aerorul uncertainty	intervals.md, band plots
Which sensors the model relies on	aerorul explain	explain.md, SHAP plot
ABC vs random search vs TPE, same budget	aerorul benchmark	benchmark.md, convergence plot

Design choices

• An engine is either in training or in evaluation, never both. Consecutive flights of one engine are nearly identical; splitting by rows would let the model memorise.
• Uncertainty is measured, not assumed. Conformal intervals come with a guarantee under an exchangeability assumption; we check whether it holds on the test engines.
• Comparisons are fair or not made. Optimisers get the same budget, space and CV; every result is a mean over seeds.
• No label leakage. The simulator's hidden health parameters and health-state flag, and the cycle counter, are never model inputs.

See Method for the reasoning behind each choice and Results for the numbers.