Turbofan Engine Degradation Dataset Data: Dataset Structure and Field Coverage

Four simulated fleets that differ on two axes: operating conditions (one versus six) and fault modes (one versus two - HPC degradation and fan degradation). Every row is one operational cycle of one engine unit across 26 whitespace-delimited columns: unit number, time in cycles, three operational settings (altitude, Mach number, throttle resolver angle) and 21 sensor channels such as fan inlet temperature, core speed, bypass duct pressure, fuel flow and oil temperature, with simulated sensor noise baked in.

Each subset arrives as three files: a training file of full run-to-failure trajectories, a test file truncated before failure, and a RUL file giving true remaining cycles for each test unit. That train/test split is what turns raw sensor text into a supervised benchmark - you predict cycles-to-failure and score against the RUL file.

train_FD001.txt alone carries 20,631 rows, which is why FD001 remains the standard starter configuration for a first remaining-useful-life model.

The four configurations scale difficulty along condition diversity and fault complexity, and the unit counts below come straight from the readme tables shipped inside the archive. Pick FD001 or FD003 to validate a pipeline under constant conditions; move to FD002 and FD004 once your model must cope with six operating regimes.

Whenever the failure mode you care about is not a jet engine. The NASA Prognostics Data Set Repository lists 21 datasets ranging from about 12 MB (C-MAPSS v1) to about 15.8 GB (C-MAPSS v2) - tens of gigabytes in total - spanning lithium-ion battery aging, bearing vibration and FEMTO bearing accelerated degradation, CNC milling machine and milling tool wear, IGBT accelerated aging, capacitor electrical stress, MOSFET thermal overstress and fatigue crack growth in aluminium lap joints.

A trained degradation model answers "when does this engine fail?"; the rest of this industry's pool answers the commercial questions around it.

The contrast between sensor archives and demand series runs through the head-to-head vs JMTBA Machine Tool Statistics (Japan) comparison: quality 8 sensor data versus quality 7 order statistics, answering different questions entirely.