<b>Propagation-Aware Sensor Evidence Mapping for Aircraft Engine Health Diagnosis</b>

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Aircraft engines generate continuous multivariate sensor time series from exhaust gas temperature, fuel flow, vibration, rotor speed, oil pressure, compressor pressure ratio, turbine temperature, and control actuator signals. These variables are tightly coupled through thermodynamic, mechanical, and control feedback relationships. A local abnormal change in one component may propagate to multiple downstream sensors, making it difficult to distinguish the initiating fault from secondary responses. This study proposes a causal propagation modeling method for fine-grained anomaly diagnosis in aircraft engine sensor time series. The method first constructs a time-lagged causal graph using conditional independence testing, physical dependency constraints, and operating-regime segmentation. A counterfactual signal reconstruction module then estimates expected sensor behavior after removing suspected causal drivers. Variable-level anomaly contribution is calculated by comparing observed sensor trajectories with causal counterfactual trajectories. Experiments are conducted on an aircraft engine monitoring dataset containing 286 engines, 92 sensor variables, and 1-second observations collected across 4,800 flight cycles. The dataset includes 318 million timestamped records and 1,640 maintenance-confirmed abnormal episodes, including compressor efficiency degradation, fuel-control instability, oil-pressure fluctuation, turbine temperature rise, and rotor vibration abnormality. The proposed method reduces median root-cause diagnosis delay from 31.6 minutes to 9.4 minutes compared with a temporal reconstruction baseline. The mean reciprocal rank for initiating-variable localization reaches 0.827. Causal path analysis assigns 1,210 abnormal episodes to interpretable fault-propagation chains, and unnecessary component inspection tickets decrease from 940 to 372. The model completes one full flight-cycle assessment in 3.8 minutes. The results show that causal propagation modeling can improve fine-grained anomaly diagnosis in highly coupled aircraft engine time series.