Pneuma: A Large-Scale Real-World Time-Series Benchmark for Predictive Maintenance of Commercial Aircraft

Problem

Predictive maintenance research in commercial aviation lacks large public real-world datasets, so most of it is done on private data, which makes reproduction and fair comparison difficult. Existing public datasets are based on simulated degradation, small general aviation aircraft, a single aircraft type, or artificially injected faults, so they do not reflect real failures in mainstream commercial aircraft types.

Method

The authors collected QAR records from 71,451 flights (about 921 million time steps) from the bleed air systems of the A320, A321, B787 and B777, gathered from December 2021 to June 2026. Using fault events confirmed in maintenance logs, they built nine subsets, one per aircraft type–fault combination. From each flight, four transition phases (taxi→takeoff, takeoff→climb, climb→cruise and cruise→descent) are cut to a fixed length. For labeling, the 100 flights immediately before a fault form the warning pool, and the remaining flights, excluding the 500 flights before and 50 flights after a fault, form the normal pool. There are three tasks: multivariate forecasting (using normal data only), supervised classification and unsupervised anomaly detection. More than 25 models from four families, including specialized forecasting, classification and anomaly detection models and time-series foundation models, were evaluated under a unified setup.

Results

• At about 921 million time steps, it is far larger than existing public aviation time-series datasets (GATS about 38.3 million, NGAFID about 46.9 million, DASHlink about 15.8 million time steps), and it contains nine fault scenarios based on actual maintenance records.
• The authors reported that even recent strong models struggle to capture subtle fault signs.
• The authors reported that adding more in-context examples or performing multivariate inference consistently improved performance.
• The authors reported that models with better forecasting performance on normal data tended to also have higher anomaly detection performance.

Limitations

The provided text was cut off before the experimental results, so per-model quantitative figures could not be checked. The data is limited to a single system, the bleed air system, and appears to come from a single airline's operations. The warning and normal labels approximate the truth using a rule of 100 flights before a fault, so the warning pool may include flights that show no actual fault signs. The authors describe this study as preliminary work.

Why it matters

It provides the first reproducible common evaluation standard in aviation maintenance, enabling fair comparison between methods. It exposes the difficulties of real industrial time series, such as class imbalance, subtle fault signs and differences between aircraft types, which also makes it a useful reference for predictive maintenance research on other equipment.

Key terms

Predictive maintenance

An approach that uses sensor data to find signs of failure early so that maintenance can be planned before a failure occurs.

QAR(Quick Access Recorder)

A device that continuously records aircraft system parameters such as pressure, temperature and valve states during flight.

Bleed air system (BAS)

A system that draws high-pressure air from the engine compressor, regulates its pressure and temperature, and supplies it for cabin pressurization, air conditioning and other uses.

Unsupervised anomaly detection

A method that learns normal patterns from normal data only, then flags segments that deviate from those patterns as anomalies.

Original

• arXiv abstract
• PDF
• HTML full text
• Hugging Face

Authors: Jiafeng Lin, Yutian Zhang, Huan Zhang, Yuxuan Wang, Jinyan Feng, Yuhui Liu, Zhongyi Pei, Jianmin Wang · Posted 2026-10-27 · License CC BY 4.0 · Based on the full text

This review is an AI-written summary of the paper. Please treat the original paper as the reference.