Pneuma: A Large-Scale Real-World Time Series Benchmark for Commercial Aircraft Predictive Maintenance
Even strong models struggled to detect subtle aircraft fault signatures in tests spanning more than 25 models.
Pneuma contains over 920 million multivariate time steps from bleed air systems across four aircraft types, covering nine verified fault scenarios tied to maintenance logs. The benchmark tests forecasting, supervised fault classification and unsupervised anomaly detection. Its data includes substantial class imbalance, subtle fault evolution and variation across aircraft configurations. Additional in-context examples and multivariate inference consistently improved results, while forecasting quality was strongly associated with anomaly detection performance.
ArXiv · AI/CL/LG's note
Pneuma contains over 920 million multivariate time steps from bleed air systems across four aircraft types, covering nine verified fault scenarios tied to maintenance logs. The benchmark tests forecasting, supervised fault classification and unsupervised anomaly detection. Its data includes substantial class imbalance, subtle fault evolution and variation across aircraft configurations. Additional in-context examples and multivariate inference consistently improved results, while forecasting quality was strongly associated with anomaly detection performance.
ArXiv · AI/CL/LG's note
score 5