Readable AI estimates simulated jet-engine life—and exposes diagnostic blind spots

Separating prediction from diagnosis

Hugo Henry and colleagues at the University of Cincinnati describe an interpretable model for estimating turbofan remaining useful life: the number of operating cycles left before failure. Its distinguishing feature is an intermediate layer intended to represent individual engine components, rather than internal signals that merely correlate with remaining life.

The study is an arXiv preprint posted on November 8, 2026. The supplied evidence does not establish peer-reviewed publication. Its results concern a simulation benchmark, not a validated airline maintenance system.

How the model works

The researchers used N-CMAPSS, which simulates engine degradation under recorded flight conditions. Their working dataset contained 99 simulated engines and 7,473 cycle-level records, covering isolated and combined component faults.

Each cycle becomes 13 sensor averages from cruise conditions. These measurements are corrected for operating conditions and smoothed across cycles. Averaging simplifies the input but discards transient behavior; the model does not process complete flight trajectories.

The architecture is a fuzzy tree. Fuzzy rules allow measurements to belong partly to categories such as low, medium and high, then blend the corresponding rule outputs. Here, small rule systems first estimate ten simulator-provided health modifiers describing component efficiency and flow capacity. Fixed mappings convert these estimates into component damage indices. Further layers combine components into the engine’s two shaft systems, or spools, and finally predict remaining life.

Crucially, training happens in stages. The health-estimation rules are fitted to their own physical targets and frozen. Only the downstream aggregation is optimized for remaining-life accuracy. This prevents that final objective from redefining the diagnostic signals, although it does not guarantee that those signals are physically correct.

The damage indices run from zero to one, using thresholds drawn from degrading training engines. They describe position within that reference population—not a measured percentage of physical damage.

Accuracy with a compact rule base

On the dataset’s native test split, root mean squared error was 8.44 cycles for the multi-fault subset and 11.51 cycles across all fault modes. This error measure summarizes prediction discrepancies while giving larger errors greater weight. With an entire fault mode withheld from training, average error was 12.20 cycles.

The deployed inference model has 333 free parameters and 411 expanded rules. Only 63 parameters are optimized against remaining life. Its parameter count is roughly 32 to 12,000 times smaller than the previously reported models compiled for comparison.

That is not evidence of superior accuracy or efficiency under matched conditions. The comparison studies used different populations, features, splits and target conventions, and were not rerun. This study also caps remaining-life targets, which lowers reported error relative to uncapped evaluation. The parameter count excludes preprocessing regressions and the many candidate models fitted during sensor selection.

Protecting the diagnostic layer carried an accuracy cost: replacing its fixed component mapping with one trained directly for remaining life reduced cross-validation error on the multi-fault subset from 8.07 to 6.78 cycles. The resulting internal states, however, no longer supported the same component-health interpretation.

What the sensors could—and could not—identify

For each health target, the team searched all 286 possible three-sensor combinations. Candidates had to track the target and avoid responding too strongly when degradation occurred on the opposite spool.

The strongest diagnostic results concerned both high-pressure compressor modifiers and low-pressure turbine efficiency. High-pressure turbine modifiers and low-pressure turbine flow were marginal. No tested triple met both acceptance thresholds for the fan or low-pressure compressor.

These limits appeared in prediction performance. When isolated fan degradation was withheld during training, error rose to 19.59 cycles. Estimated damage increased too little, leaving remaining-life predictions overly optimistic.

The diagnostic test distinguishes responses across spools; it does not establish discrimination between components sharing one spool. Sensor selection and reporting also used the same fleet, creating optimism that was not resolved through fully nested selection.

Why deployment remains unproven

Every simulated engine eventually fails, with monotonic degradation and no repairs or surviving, incomplete histories. Operational fleets differ. Cruise-only observations can also miss faults visible under other loads, and predictions have no uncertainty intervals.

One preprocessing choice uses the first fifth of each engine’s total lifetime to set a calibration window. That window length would not be known early in service. A fixed-window alternative was proposed but not evaluated.

The tree forces predicted life to zero when estimated damage saturates; this cannot protect against an actual fault its sensors fail to detect. Its finite rule structure may support formal verification, but no solver-based proofs were performed. No maintenance benefits or airworthiness decisions were tested. The contribution is an inspectable research model with explicitly measured blind spots, not demonstrated operational assurance.