Propeller Damage Detection: Adapting Models to Diverse UAV Sizes

This manuscript introduces a transfer learning method for adapting propeller fault detection neural networks to different unmanned aerial vehicles (UAVs). After training a simple model for detecting if any propeller in a specific vehicle has a failure (in this case, a chipped tip), a domain adaptation based in the vehicles’ physics is performed in order to use the same model to detect failures in vehicles with different structures, weights, or motor-propeller sets. A key feature is that the detection model uses only inertial sensors that are standard in commercial UAVs, making it broadly applicable without the need for additional hardware.

ICUAS 2025