CruiseBench: новый бенчмарк для оценки ресурса авиадвигателей нейросетями

A group of researchers published CruiseBench on arXiv in July 2026 — a benchmark for predicting aircraft engine remaining useful life (RUL) that isolates only the cruise phase of flight from the N-CMAPSS dataset; in baseline tests, the TSMixer model showed the best accuracy with an average RMSE of 3.4.

What is CruiseBench

CruiseBench is a specialized benchmark for comparing neural networks that predict aircraft engine remaining useful life (RUL, Remaining Useful Life), built on top of the N-CMAPSS dataset. RUL prediction itself estimates how much longer an engine can safely operate, and underlies maintenance planning. The core of the benchmark is the CPM-N-CMAPSS artifact (Cruising-Period Mask for N-CMAPSS): a mask that stores per-cycle intervals of cruise flight, identified using the common-altitude method, for the nine available sub-datasets.

• The benchmark is derived from the N-CMAPSS dataset — an extension of the classic C-MAPSS
• CPM-N-CMAPSS is a mask of cruise intervals for 9 sub-datasets
• The protocol's input consists of scenario descriptors and real sensor readings
• Virtual sensors, health parameters, and metadata are excluded from the features
• Baseline models: LSTM, GRU, TCN, and TSMixer

Why only the cruise phase?

The cruise phase removes the noise of operating-mode changes that interferes with honestly measuring engine degradation. N-CMAPSS stores full time series within each flight, rather than per-cycle snapshots like the earlier C-MAPSS. According to the authors, this realism simultaneously increases the volume of data and "intertwines" wear features with mode variations — takeoff, climb, descent. As a result, preprocessing choices start to influence the final metrics, and direct model comparisons become unreliable. CruiseBench addresses this with a fixed protocol: it works only with "masked" cruise rows, preserves windows at native resolution, and applies separate RUL constraints for each of the nine sub-datasets.

Which model turned out more accurate?

TSMixer showed the best result. Under the CruiseBench-eta5-W256-S10 configuration, it achieved the lowest average RMSE of 3.4 ± 1.71 and a Saxena score of (2.50 ± 2.99) × 10⁴, outperforming the recurrent LSTM and GRU, as well as the convolutional TCN — all four models were run by the authors as baselines. The RMSE metric measures the average forecast error in cycles, while the Saxena score penalizes late failure warnings more heavily than early ones. The authors emphasize that the numbers are sensitive to three factors, as shown by ablation experiments: the choice of flight phase, the method of temporal downsampling, and the RUL constraint threshold.

"CruiseBench provides a reproducible sub-benchmark for controlled

comparison of RUL models, while CPM-N-CMAPSS provides a stage-specific data foundation for future research into transfer learning and domain adaptation," the paper's abstract on arXiv states.

What this means

Predictive maintenance for aircraft is getting a fairer yardstick for comparing models. By fixing a single flight phase, CruiseBench reduces the influence of preprocessing on the final numbers and improves reproducibility — a painful problem for tasks on "raw" industrial data. A separate value is the CPM-N-CMAPSS mask itself: the authors position it as a foundation for future work on transfer learning and domain adaptation, where a model trained on one mode or engine is transferred to another.

Frequently asked questions

What is RUL in the context of aircraft engines?

RUL (Remaining Useful Life) is an estimate of how much longer an engine can safely operate. According to the authors, it is a central quantity for maintenance planning.

How does N-CMAPSS differ from C-MAPSS?

N-CMAPSS extends C-MAPSS: it simulates engine trajectories to failure based on recorded real flight profiles and preserves full time series within a flight, rather than per-cycle snapshots. It is precisely this level of detail that makes the data more realistic, but also more difficult to process.