Toward More Reliable UAVs: Interpretable Machine Learning for Understanding Propeller Fault Separability

This study investigates the fundamental limits of UAV propeller fault discrimination using an interpretable XGBoost–SHAP framework applied to the high-resolution DronePropA dataset. Three fault types—edge-cut, structural crack, and surface erosion—were evaluated across three severity levels under constrained flight conditions to isolate the physical signatures of degradation. A curated subset of 1 kHz IMU and ESC telemetry from Drone 1 (trajectory t1, speed SP1) enabled analysis of fault separability independent of kinematic variability. Despite deploying a moderately complex XGBoost classifier (300 estimators, depth = 6) and a comprehensive multi-domain feature set, the model achieved a performance ceiling of 38.04% accuracy and 38.27% macro-F1 which reflects the inherent aerodynamic similarity among fault classes rather than model limitations. Confusion analysis revealed structured misclassification, with 62% of total errors arising from overlap between edge-cut and surface-erosion states, while crack-major exhibited markedly higher recall (66.7%) due to its distinctive high-frequency dynamics. SHAP analysis further demonstrated that only a subset of features—most notably wx_Hs (0.082) and omega_fpk_mean3 (0.058)—dominate fault separability. These findings mean that the proposed physics-informed diagnostic framework indicates that certain propeller faults may be intrinsically indistinguishable using minimal onboard sensors alone.